1.0	P15-1001	A00-1004	b'1 Introduction Neural machine translation (NMT) is a recently introduced approach to solving machine translation (Kalchbrenner and Blunsom, 2013; Bahdanau et al, 2015; Sutskever et al, 2014). In neural machine translation, one builds a single neural network that reads a source sentence and generates its translation. The whole neural network is jointly trained to maximize the conditional probability of a correct translation given a source sentence, using the bilingual corpus. The NMT models have shown to perform as well as the most widely used conventional translation systems (Sutskever et al, 2014; Bahdanau et al, 2015). Neural machine translation has a number of advantages over the existing statistical machine translation system, specifically, the phrase-based system (Koehn et al, 2003). First, NMT requires a minimal set of domain knowledge. For instance, all of the models proposed in (Sutskever et al, 2014), (Bahdanau et al, 2015) or (Kalchbrenner and Blunsom, 2013) do not assume any linguistic property in both source and target sentences except that they are sequences of words. Second, the whole system is jointly trained to maximize the translation performance, unlike the existing phrase-based system which consists of many separately trained features whose weights are then tuned jointly. Lastly, the memory footprint of the NMT model is often much smaller than the existing system which relies on maintaining large tables of phrase pairs. Despite these advantages and promising results, there is a major limitation in NMT compared to the existing phrase-based approach. That is, the number of target words must be limited. This is mainly because the complexity of training and using an NMT model increases as the number of target words increases. A usual practice is to construct a target vocabulary of the K most frequent words (a socalled shortlist), where K is often in the range of 30k (Bahdanau et al, 2015) to 80k (Sutskever et al., 2014). Any word not included in this vocabulary is mapped to a special token representing an unknown word [UNK]. This approach works well when there are only a few unknown words in the target sentence, but it has been observed 1 that the translation performance degrades rapidly as the number of unknown words increases (Cho et al, 2014a; Bahdanau et al, 2015). In this paper, we propose an approximate training algorithm based on (biased) importance sampling that allows us to train an NMT model with a much larger target vocabulary. The proposed algorithm effectively keeps the computational complexity during training at the level of using only a small subset of the full vocabulary. Once the model with a very large target vocabulary is trained, one can choose to use either all the target words or only a subset of them. We compare the proposed algorithm against the baseline shortlist-based approach in the tasks of English?French and English?German translation using the NMT model introduced in (Bahdanau et al, 2015). The empirical results demonstrate that we can potentially achieve better translation performance using larger vocabularies, and that our approach does not sacrifice too much speed for both training and decoding. Furthermore, we show that the modeltrained with this algorithm gets the best translation performance yet achieved by single NMT models on the WMT?14 English?French translation task. 2 Neural Machine Translation and Limited Vocabulary Problem In this section, we briefly describe an approach to neural machine translation proposed recently in (Bahdanau et al, 2015). Based on this description we explain the issue of limited vocabularies in neural machine translation. 2.1 Neural Machine Translation Neural machine translation is a recently proposed approach to machine translation, which uses a single neural network trained jointly to maximize the translation performance (Forcada and ? Neco, 1997; Kalchbrenner and Blunsom, 2013; Cho et al., 2014b; Sutskever et al, 2014; Bahdanau et al, 2015). Neural machine translation is often implemented as the encoder?decoder network. The encoder reads the source sentence x = (x 1 , . . . , x T ) and encodes it into a sequence of hidden states h = (h 1 , ? ? ? , h T ): h t = f (x t , h t?1 ) . (1) Then, the decoder, another recurrent neural network, generates a corresponding translation y = (y 1 , ? ? ? , y T ? ) based on the encoded sequence of hidden states h: p(y t | y <t , x) ? exp {q (y t?1 , z t , c t )} , (2) where z t = g (y t?1 , z t?1 , c t ) , (3) c t = r (z t?1 , h 1 , . . . , h T ) , (4) and y <t = (y 1 , . . . , y t?1 ). The whole model is jointly trained to maximize the conditional log-probability of the correct translation given a source sentence with respect to the parameters ? of the model: ? ? = argmax ? N ? n=1 T n ? t=1 log p(y n t | y n <t , x n ), where (x n , y n ) is the n-th training pair of sentences, and T n is the length of the n-th target sentence (y n ). 2.1.1 Detailed Description In this paper, we use a specific implementation of neural machine translation that uses an attention mechanism, as recently proposed in (Bahdanau et al., 2015). In (Bahdanau et al, 2015), the encoder in Eq. (1) is implemented by a bi-directional recurrent neural network such that h t = [ ?? h t ; ?? h t ] , where ?? h t = f ( x t , ?? h t+1 ) , ?? h t = f ( x t , ?? h t?1 ) . They used a gated recurrent unit for f (see, e.g., (Cho et al, 2014b)). The decoder, at each time, computes the context vector c t as a convex sum of the hidden states (h 1 , . . . , h T ) with the coefficients ? 1 , . . . , ? T computed by ? t = exp {a (h t , zt?1 )} ? k exp {a (h k , z t?1 )} , (5) where a is a feedforward neural network with a single hidden layer. A new hidden state z t of the decoder in Eq. (3) is computed based on the previous hidden state z t?1 , previous generated symbol y t?1 and the computed 2 context vector c t . The decoder also uses the gated recurrent unit, as the encoder does. The probability of the next target word in Eq. (2) is then computed by p(y t | y <t , x) = 1 Z exp { w > t ? (y t?1 , z t , c t ) + b t } , (6) where ? is an affine transformation followed by a nonlinear activation, and w t and b t are respectively the target word vector and the target word bias. Z is the normalization constant computed by Z = ? k:y k ?V exp { w > k ? (y t?1 , z t , c t ) + b k } , (7) where V is the set of all the target words. For the detailed description of the implementation, we refer the reader to the appendix of (Bahdanau et al, 2015). 2.2 Limited Vocabulary Issue and Conventional Solutions One of the main difficulties in training this neural machine translation model is the computational complexity involved in computing the target word probability (Eq. (6)). More specifically, we need to compute the dot product between the feature ? (y t?1 , z t , c t ) and the word vector w t as many times as there are words in a target vocabulary in order to compute the normalization constant (the denominator in Eq. (6)). This has to be done for, on average, 20?30 words per sentence, which easily becomes prohibitively expensive even with a moderate number of possible target words. Furthermore, the memory requirement grows linearly with respect to the number of target words. This has been a major hurdle for neural machine translation, compared to the existing non-parametric approaches such as phrase-based translation systems. Recently proposed neural machine translation models, hence, use a shortlist of 30k to 80k most frequent words (Bahdanau et al, 2015; Sutskever et al, 2014). This makes training more feasible, but comes with a number of problems. First of all, the performance of the model degrades heavily if the translation of a source sentence requires many words that are not included in the shortlist (Cho et al, 2014a). This also affects the performance evaluation of the system which is often measured by BLEU. Second, the first issue becomes more problematic with languages that have a rich set of words such as German or other highly inflected languages. There are two model-specific approaches to this issue of large target vocabulary. The first approach is to stochastically approximate the target word probability. This has been proposed recently in (Mnih and Kavukcuoglu, 2013; Mikolov et al, 2013) based on noise-contrastive estimation (Gutmann and Hyvarinen, 2010). In the second approach, thetarget words are clustered into multiple classes, or hierarchical classes, and the target probability p(y t |y <t , x) is factorized as a product of the class probability p(c t |y <t , x) and the intraclass word probability p(y t |c t , y <t , x). This reduces the number of required dot-products into the sum of the number of classes and the words in a class. These approaches mainly aim at reducing the computational complexity during training, but do not often result in speed-up when decoding a translation during test time. 1 Other than these model-specific approaches, there exist translation-specific approaches. A translation-specific approach exploits the properties of the rare target words. For instance, Luong et al proposed such an approach for neural machine translation (Luong et al, 2015). They replace rare words (the words that are not included in the shortlist) in both source and target sentences into corresponding ?OOV n ? tokens using the word alignment model. Once a source sentence is translated, each ?OOV n ? in the translation will be replaced based on the source word marked by the corresponding ?OOV n ?. It is important to note that the modelspecific approaches and the translation-specific approaches are often complementary and can be used together to further improve the translation performance and reduce the computational complexity. 3 Approximate Learning Approach to Very Large Target Vocabulary 3.1 Description In this paper, we propose a model-specific approach that allows us to train a neural machine translation model with a very large target vocabulary. With the proposed approach, the compu1 This is due to the fact that the beam search requires the conditional probability of every target word at each time step regardless of the parametrization of the output probability. 3 tational complexity of training becomes constant with respect to the size of the target vocabulary. Furthermore, the proposed approach allows us to efficiently use a fast computing device with limited memory, such as a GPU, to train a neural machine translation model with a much larger target vocabulary. As mentioned earlier, the computational inefficiency of training a neural machine translation model arises from the normalization constant in Eq. (6). In order to avoid the growing complexity of computing the normalization constant, we propose here to use only a small subset V ? of the target vocabulary at each update. The proposed approach is based on the earlier work of (Bengio and S?en?ecal, 2008). Let us consider the gradient of the logprobability of the output in Eq. (6). The gradient is composed of a positive and negative part: ? log p(y t | y <t , x) (8) =?E(y t )? ? k:y k ?V p(y k | y <t , x)?E(y k ), where we define the energy E as E(y j ) = w > j ? (y j?1 , z j , c j ) + b j . The second, or negative, term of the gradient is in essence the expected gradient of the energy: E P [?E(y)] , (9) where P denotes p(y| y <t , x). The main idea of the proposed approach is to approximate this expectation, or the negative term of the gradient, by importance sampling with a small number of samples. Given a predefined proposal distribution Q and a set V ? of samples from Q, we approximate the expectation in Eq. (9) with E P [?E(y)] ? ? k:y k ?V ? ? k ? k ? :y k ? ?V ? ? k ? ?E(y k ), (10) where ? k = exp {E(y k )? logQ(y k )} . (11) This approach allows us to compute the normalization constant during training using only a small subset of the target vocabulary, resulting in much lower computational complexity for each parameter update. Intuitively, at each parameter update, we update only the vectors associated with the correct word w t and with the sampled words in V ? . Once training is over, we can use the full target vocabulary to compute the output probability of each target word. Although the proposed approach naturally addresses the computational complexity, using this approach naively does not guarantee that the number of parameters being updated for each sentence pair, which includes multiple target words, is bounded nor can be controlled. This becomes problematic when training is done, for instance, on a GPU with limited memory. In practice, hence, we partition the training corpus and define a subset V ? of the target vocabulary for each partition prior to training. Before training begins, we sequentially examine each target sentence in the training corpus and accumulate unique target words until the number of unique target words reaches the predefined threshold ? . The accumulated vocabulary will be used for this partition of the corpus during training. We repeat this until the end of the training set is reached. Let us refer to the subset of target words used for the i-th partition by V ? i . This may be understood as having a separate proposal distribution Q i for each partition of the training corpus. The distribution Q i assigns equal probability mass to all the target words included in the subset V ? i , and zero probability mass to all the other words, i.e., Q i (y k ) = ? ? ? 1 | V ? i | if y t ? V ? i 0 otherwise. This choice of proposal distribution cancels out the correction term ? logQ(y k ) from the importance weight in Eqs. (10)?(11), which makes the proposed approach equivalent to approximating the exact output probability in Eq. (6) with p(y t | y <t , x) = exp { w > t ? (y t?1 , z t , c t ) + b t } ? k:y k ?V ? exp { w > k ? (y t?1 , z t , c t ) + b k } . It should be noted that this choice of Q makes the estimator biased. The proposed procedure results in speed up against usual importance sampling, asit exploits the advantage of modern computers in doing matrix-matrix vs matrix-vector multiplications. 4 3.1.1 Informal Discussion on Consequence The parametrization of the output probability in Eq. (6) can be understood as arranging the vectors associated with the target words such that the dot product between the most likely, or correct, target word?s vector and the current hidden state is maximized. The exponentiation followed by normalization is simply a process in which the dot products are converted into proper probabilities. As learning continues, therefore, the vectors of all the likely target words tend to align with each other but not with the others. This is achieved exactly by moving the vector of the correct word in the direction of ? (y t?1 , z t , c t ), while pushing all the other vectors away, which happens when the gradient of the logarithm of the exact output probability in Eq. (6) is maximized. Our approximate approach, instead, moves the word vectors of the correct words and of only a subset of sampled target words (those included in V ? ). 3.2 Decoding Once the model is trained using the proposed approximation, we can use the full target vocabulary when decoding a translation given a new source sentence. Although this is advantageous as it allows the trained model to utilize the whole vocabulary when generating a translation, doing so may be too computationally expensive, e.g., for realtime applications. Since training puts the target word vectors in the space so that they align well with the hidden state of the decoder only when they are likely to be a correct word, we can use only a subset of candidate target words during decoding. This is similar to what we do during training, except that at test time, we do not have access to a set of correct target words. The most na??ve way to select a subset of candidate target words is to take only the top-K most frequent target words, where K can be adjusted to meet the computational requirement. This, however, effectively cancels out the whole purpose of training a model with a very large target vocabulary. Instead, we can use an existing word alignment model to align the source and target words in the training corpus and build a dictionary. With the dictionary, for each source sentence, we construct a target word set consisting of the K-most frequent words (according to the estimated unigram probability) and, using the dictionary, at most K ? likely target words for each source word. K and K ? may be chosen either to meet the computational requirement or to maximize the translation performance on the development set. We call a subset constructed in either of these ways a candidate list. 3.3 Source Words for Unknown Words In the experiments, we evaluate the proposed approach with the neural machine translation model called RNNsearch (Bahdanau et al, 2015) (see Sec. 2.1.1). In this model, as a part of decoding process, we obtain the alignments between the target words and source locations via the alignment model inEq. (5). We can use this feature to infer the source word to which each target word was most aligned (indicated by the largest ? t in Eq. (5)). This is especially useful when the model generated an [UNK] token. Once a translation is generated given a source sentence, each [UNK] may be replaced using a translation-specific technique based on the aligned source word. For instance, in the experiment, we try replacing each [UNK] token with the aligned source word or its most likely translation determined by another word alignment model. Other techniques such as transliteration may also be used to further improve the performance (Koehn, 2010). 4 Experiments We evaluate the proposed approach in English?French and English?German translation tasks. We trained the neural machine translation models using only the bilingual, parallel corpora made available as a part of WMT?14. For each pair, the datasets we used are: ? English?French: 2 ? Common Crawl ? News Commentary ? Gigaword ? Europarl v7 ? UN ? English?German: ? Common Crawl ? News Commentary ? Europarl v7 2 The preprocessed data can be found and downloaded from http://www-lium.univ-lemans.fr/ ? schwenk/nnmt-shared-task/README. 5 English-French English-German Train Test Train Test 15k 93.5 90.8 88.5 83.8 30k 96.0 94.6 91.8 87.9 50k 97.3 96.3 93.7 90.4 500k 99.5 99.3 98.4 96.1 All 100.0 99.6 100.0 97.3 Table 1: Data coverage (in %) on target-side corpora for different vocabulary sizes. ?All? refers to all the tokens in the training set. To ensure fair comparison, the English?French corpus, which comprises approximately 12 million sentences, is identical to the one used in (Kalchbrenner and Blunsom, 2013; Bahdanau et al, 2015; Sutskever et al, 2014). As for English?German, the corpus was preprocessed, in a manner similar to (Peitz et al, 2014; Li et al, 2014), in order to remove many poorly translated sentences. We evaluate the models on the WMT?14 test set (news-test 2014), 3 while the concatenation of news-test-2012 and news-test-2013 is used for model selection (development set). Table 1 presents data coverage w.r.t. the vocabulary size, on the target side. Unless mentioned otherwise, all reported BLEU scores (Papineni et al, 2002) are computed with the multi-bleu.perl script 4 on the cased tokenized translations. 4.1 Settings As a baseline for English?French translation, we use the RNNsearch model proposed by (Bahdanau et al, 2015), with 30k source and target words. 5 Another RNNsearch model is trained for English?German translation with 50k source and target words. For each language pair, we train another set of RNNsearch models with much larger vocabularies of 500k source and target words, using the proposed approach. We call these models RNNsearch-LV. We vary the size of the shortlist used during training (? in Sec. 3.1). We tried 3 To compare with previous submissions, we use the filtered test sets. 4 https://github.com/moses-smt/ mosesdecoder/blob/master/scripts/ generic/multi-bleu.perl 5 The authors of (Bahdanau et al, 2015) gave us access to their trained models. We chose the best one on the validation set and resumed training. 15k and 30k for English?French, and 15k and 50k for English?German. We later report the results for thebest performance on the development set, with models generally evaluated every twelve hours. The training speed is approximately the same as for RNNsearch. Using a 780 Ti or Titan Black GPU, we could process 100k mini-batches of 80 sentences in about 29 and 39 hours respectively for ? = 15k and ? = 50k. For both language pairs, we also trained new models, with ? = 15k and ? = 50k, by reshuffling the dataset at the beginning of each epoch. While this causes a non-negligible amount of overhead, such a change allows words to be contrasted with different sets of other words each epoch. To stabilize parameters other than the word embeddings, at the end of the training stage, we freeze the word embeddings and tune only the other parameters for approximately two more days after the peak performance on the development set is observed. This helped increase BLEU scores on the development set. We use beam search to generate a translation given a source. During beam search, we keep a set of 12 hypotheses and normalize probabilities by the length of the candidate sentences, as in (Cho et al, 2014a). 6 The candidate list is chosen to maximize the performance on the development set, for K ? {15k, 30k, 50k} and K ? ? {10, 20}. As explained in Sec. 3.2, we test using a bilingual dictionary to accelerate decoding and to replace unknown words in translations. The bilingual dictionary is built using fast align (Dyer et al., 2013). We use the dictionary only if a word starts with a lowercase letter, and otherwise, we copy the source word directly. This led to better performance on the development sets. Note on ensembles For each language pair, we began training four models from each of which two points corresponding to the best and secondbest performance on the development set were collected. We continued training from each point, while keeping the word embeddings fixed, until the best development performance was reached, and took the model at this point as a single model in an ensemble. This procedure resulted in a total of eight models from which we averaged the length-normalized log-probabilities. Since much of training had been shared, the composition of 6 These experimental details differ from (Bahdanau et al, 2015). 6 RNNsearch RNNsearch-LV Google Phrase-based SMT Basic NMT 29.97 (26.58) 32.68 (28.76) 30.6 ? 33.3 ? 37.03 ? +Candidate List ? 33.36 (29.32) ? +UNK Replace 33.08 (29.08) 34.11 (29.98) 33.1 ? +Reshuffle (?=50k) ? 34.60 (30.53) ? +Ensemble ? 37.19 (31.98) 37.5 ? (a) English?French RNNsearch RNNsearch-LV Phrase-based SMT Basic NMT 16.46 (17.13) 16.95 (17.85) 20.67 \x05 +Candidate List ? 17.46 (18.00) +UNK Replace 18.97 (19.16) 18.89 (19.03) +Reshuffle ? 19.40 (19.37) +Ensemble ? 21.59 (21.06) (b) English?German Table 2: The translation performances in BLEU obtained by different models on (a) English?French and (b) English?German translation tasks. RNNsearch is the model proposed in (Bahdanau et al, 2015), RNNsearch-LV is the RNNsearch trained with the approach proposed in this paper, and Google is the LSTM-based model proposed in (Sutskever et al,2014). Unless mentioned otherwise, we report singlemodel RNNsearch-LV scores using ? = 30k (English?French) and ? = 50k (English?German). For the experiments we have run ourselves, we show the scores on the development set as well in the brackets. (?) (Sutskever et al, 2014), (?) (Luong et al, 2015), (?) (Durrani et al, 2014), (?) Standard Moses Setting (Cho et al, 2014b), (\x05) (Buck et al, 2014). such ensembles may be sub-optimal. This is supported by the fact that higher cross-model BLEU scores (Freitag et al, 2014) are observed for models that were partially trained together. 4.2 Translation Performance In Table 2, we present the results obtained by the trained models with very large target vocabularies, and alongside them, the previous results reported in (Sutskever et al, 2014), (Luong et al, 2015), (Buck et al, 2014) and (Durrani et al, 2014). Without translation-specific strategies, we can clearly see that the RNNsearch-LV outperforms the baseline RNNsearch. In the case of the English?French task, RNNsearch-LV approached the performance level of the previous best single neural machine translation (NMT) model, even without any translationspecific techniques (Sec. 3.2?3.3). With these, however, the RNNsearch-LV outperformed it. The performance of the RNNsearch-LV is also better than that of a standard phrase-based translation system (Cho et al, 2014b). Furthermore, by combining 8 models, we were able to achieve a translation performance comparable to the state of the art, measured in BLEU. For English?German, the RNNsearch-LV outperformed the baseline before unknown word replacement, but after doing so, the two systems performed similarly. We could reach higher largevocabulary single-model performance by reshuffling the dataset, but this step could potentially also help the baseline. In this case, we were able to surpass the previously reported best translation result on this task by building an ensemble of 8 models. With ? = 15k, the RNNsearch-LV performance worsened a little, with best BLEU scores, without reshuffling, of 33.76 and 18.59 respectively for English?French and English?German. The English?German ensemble described in this paper has also been used for the shared translation task of the 10 th Workshop on Statistical Machine Translation (WMT?15), where it was ranked first in terms of BLEU score. The translations by this ensemble can be found online. 7 4.3 Analysis 4.3.1 Decoding Speed In Table 3, we present the timing information of decoding for different models. Clearly, decoding from RNNsearch-LV with the full target vocab7 http://matrix.statmt.org/matrix/ output/1774?run_id=4079 7 CPU ? GPU ? RNNsearch 0.09 s 0.02 s RNNsearch-LV 0.80 s 0.25 s RNNsearch-LV 0.12 s 0.05 s +Candidate list Table 3: The average per-word decoding time. Decoding here does not include parameter loading and unknown word replacement. The baseline uses 30k words. The candidate list is built with K = 30k and K ? = 10. (?) i7-4820K (single thread), (?) GTX TITAN Black ulary is slowest. If we use a candidate list for decoding each translation, the speed of decoding substantially improves and becomes close to the baseline RNNsearch. A potential issue with using a candidate list is that for each source sentence, we must re-build a target vocabulary andsubsequently replace a part of the parameters, which may easily become timeconsuming. We can address this issue, for instance, by building a common candidate list for multiple source sentences. By doing so, we were able to match the decoding speed of the baseline RNNsearch model. 4.3.2 Decoding Target Vocabulary For English?French (? = 30k), we evaluate the influence of the target vocabulary when translating the test sentences by using the union of a fixed set of 30k common words and (at most) K ? likely candidates for each source word according to the dictionary. Results are presented in Figure 1. With K ? = 0 (not shown), the performance of the system is comparable to the baseline when not replacing the unknown words (30.12), but there is not as much improvement when doing so (31.14). As the large vocabulary model does not predict [UNK] as much during training, it is less likely to generate it when decoding, limiting the effectiveness of the post-processing step in this case. With K ? = 1, which limits the diversity of allowed uncommon words, BLEU is not as good as with moderately larger K ? , which indicates that our models can, to some degree, correctly choose between rare alternatives. If we rather use K = 50k, as we did for testing based on validation performance, the improvement over K ? = 1 is approximately 0.2 BLEU. When validating the choice of K, we found it to be correlated with the value of ? used during 10 0 10 1 10 2 10 3 K? 32.8 33.0 33.2 33.4 33.6 33.8 34.0 34.2 B L E U s c o r e With UNK replacement Without UNK replacement Figure 1: Single-model test BLEU scores (English?French) with respect to the number of dictionary entries K ? allowed for each source word. training. For example, on the English?French validation set, with ? = 15k (and K ? = 10), the BLEU score is 29.44 with K = 15k, but drops to 29.19 and 28.84 respectively for K = 30k and 50k. For ? = 30k, the score increases moderately from K = 15k to K = 50k. A similar effect was observed for English?German and on the test sets. As our implementation of importance sampling does not apply the usual correction to the gradient, it seems beneficial for the test vocabularies to resemble those used during training.5 Conclusion In this paper, we proposed a way to extend the size of the target vocabulary for neural machine translation. The proposed approach allows us to train a model with much larger target vocabulary without any substantial increase in computational complexity. It is based on the earlier work in (Bengio and S?en?ecal, 2008) which used importance sampling to reduce the complexity of computing the normalization constant of the output word probability in neural language models. On English?French and English?German translation tasks, we observed that the neural machine translation models trained using the proposed method performed as well as, or better than, those using only limited sets of target words, even when replacing unknown words. As performance of the RNNsearch-LV models increased when only a selected subset of the target vocabulary was used during decoding, this makes the proposed learning algorithm more practical. When measured by BLEU, our models showed translation performance comparable to the 8 state-of-the-art translation systems on both the English?French task and English?German task. On the English?French task, a model trained with the proposed approach outperformed the best single neural machine translation (NMT) model from (Luong et al, 2015) by approximately 1 BLEU point. The performance of the ensemble of multiple models, despite its relatively less diverse composition, is approximately 0.3 BLEU points away from the best system (Luong et al, 2015). On the English?German task, the best performance of 21.59 BLEU by our model is higher than that of the previous state of the art (20.67) reported in (Buck et al, 2014). Finally, we release the source code used in our experiments to encourage progress in neural machine translation. 8 Acknowledgments The authors would like to thank the developers of Theano (Bergstra et al, 2010; Bastien et al, 2012). We acknowledge the support of the following agencies for research funding and computing support: NSERC, Calcul Qu?ebec, Compute Canada, the Canada Research Chairs, CIFAR and Samsung.'	b'1 In t roduct ion   Parallel texts have been used in a number of studies  in computational linguistics. Brown et al (1993)  defined a series of probabilistic translation models  for MT purposes. While people may question the  effectiveness of using these models for a full-blown  MT system, the models are certainly valuable for de-  veloping translation assistance tools. For example,  we can use such a translation model to help com-  plete target ext being drafted by a human transla-  tor (Langlais et al, 2000).  Another utilization is in cross-language informa-  tion retrieval (CLIR) where queries have to be trans-  lated from one language to another language in  which the documents are written. In CLIR, the qual-  ity requirement for translation is relatively low. For  example, the syntactic aspect is irrelevant. Even if  the translated word is not a true translation but is  strongly related to the original query, it is still help-  ful. Therefore, CLIR is a suitable application for  such a translation model.  However, a major obstacle to this approach is the  lack of parallel corpora for model training. Only  a few such corpora exist, including the Hansard  English-French corpus and the HKUST English-  Chinese corpus (Wu, 1994). In this paper, we will  describe a method which automatically searches for  parallel texts on the Web. We will discuss the text  mining algorithm we adopted, some issues in trans-  lation model training using the generated parallel  corpus, and finally the translation model\'s perfor-  mance in CLIR.  2 Para l le l  Text  M in ing  A lgor i thm  The PTMiner system is an intelligent Web agent  that is designed to search for large amounts of paral-  lel text on the Web. The mining algorithm is largely  language independent. It can thus be adapted to  other language pairs with only minor modifications.  Taking advantage ofWeb search engines as much  as possible, PTMiner implements he following steps  (illustrated in Fig. 1):  1 Search for candidate sites - Using existing Web  search engines, search for the candidate sites  that may contain parallel pages;  2 File name fetching - For each candidate site,  fetch the URLs of Web pages that are indexed  by the search engines;  3 Host crawling - Starting from the URLs col-  lected in the previous tep, search through each  candidate site separately for more URLs;  4 Pair scan - From the obtained URLs of each  site, scan for possible parallel pairs;  5 Download and verifying - Download the parallel  pages, determine file size, language, and charac-  ter set of each page, and filter out non-parallel  pairs.  2.1 Search for candidate Sites  We take advantage of the huge number of Web sites  indexed by existing search engines in determining candidate sites. This is done by submitting some  particular equests to the search engines. The re-  quests are determined according to the following ob-  servations. In the sites where parallel text exists,  there are normally some pages in one language con-  taining links to the parallel version in the other lan-  guage. These are usually indicated by those links\'  anchor texts 1. For example, on some English page  there may be a link to its Chinese version with  the anchor text "Chinese Version" or "in Chinese".  1An anchor text  is a piece of text on a Web page which,  when clicked on, will take you to another linked page. To  be helpful, it usual ly  contains the key information about the  l inked page.  21  Figure 1: The workflow of the mining process.  The same phenomenon can be observed on Chinese  pages. Chances are that a site with parallel texts  will contain such links in some of its documents.  This fact is used as the criterion in searching for  candidate sites.  Therefore, to determine possible sites for English-  Chinese parallel texts, we can request an English  document containing the following anchor:  anchor : "engl ish version H \\["in english", ...\\].  Similar requests are sent for Chinese documents.  From the two sets of pages obtained by the above  queries we extract wo sets of Web sites. The union  of these two sets constitutes then the candidate sites.  That  is to say, a site is a candidate site when it  is found to have either an English page linking to  its Chinese version or a Chinese page linking to its  English version.  2.2 File Name Fetching  We now assume that a pair of parallel texts exists on  the same site. To search for parallel pairs on a site,  PTMiner first has to obtain all (or at least part of)  the HTML file names on the site. From these names  pairs are scanned. It is possible to use a Web crawler  to explore the candidate sites completely. However,  we can take advantage of the search engines again to  accelerate the process. As the first step, we submit  the following query to the search engines:  host : hostname  to fetch the Web pages that they indexed from this  site. If we only require a small amount of parallel  texts, this result may be sufficient. For our purpose,  however, we need to explore the sites more thor-  oughly using a host crawler. Therefore, we continue  our search for files with a host crawler which uses  the documents found by the search engines as the  starting point.  2.3 Host Crawling  A host crawler is slightly different from a Web  crawler. Web crawlers go through innumerable  pages and hosts on theWeb. A host crawler is a  Web crawler that crawls through documents on a  given host only. A breadth-first crawling algorithm  is applied in PTMiner as host crawler. The principle  is that when a link to an unexplored ocument on  the same site is found in a document, it is added to  a list that will be explored later. In this way, most  file names from the candidate sites are obtained.  2.4 Pair Scan  After collecting file names for each candidate site,  the next task is to determine the parallel pairs.  Again, we try to use some heuristic rules to guess  which files may be parallel texts before downloading  them. The rules are based on external features of  the documents. By external feature, we mean those  features which may be known without analyzing the  contents of the file, such as its URL, size, and date.  This is in contrast with the internal features, such as  language, character set, and HTML structure, which  cannot be known until we have downloaded the page  and analyzed its contents.  The heuristic criterion comes from the following  observation: We observe that parallel text pairs usu-  ally have similar name patterns. The difference be-  tween the names of two parailel pages usually lies  in a segment which indicates the language. For ex-  ample, "file-ch.html" (in Chinese) vs. "file-en.html"  (in English). The difference may also appear in the  path, such as ".../chinese/.../fi le.html" vs. ".../en-  glish/.../f i le.html\'. The name patterns described  above are commonly used by webmasters to help or-  ganize their sites. Hence, we can suppose that a  pair of pages with this kind of pattern are probably  parallel texts.  22 First, we establish four lists for English pre-  fixes, English suffixes, Chinese prefixes and Chi-  nese suffixes. For example: Engl ish P re f ix  =  {e, en, e_, en_, e - ,  en - ,  ...}. For each file in one lan-  guage, if a segment in its name corresponds to one  of the language affixes, several new names are gener-  ated by changing the segment to the possible corre-  sponding affixes of the other language. If a generated  name corresponds to an existing file, then the file is  considered as a candidate parallel document of the  original file.  2.5 Filtering  Next, we further examine the contents of the paired  files to determine if they are really parallel according  to various external and internal features. This may  further improve the pairing precision. The following  methods have been implemented in our system.  2.5.1 Text Length  Parallel files often have similar file lengths. One sim-  ple way to filter out incorrect pairs is to compare  the lengths of the two files. The only problem is to  set a reasonablethreshold that will not discard too  many good pairs, i.e. balance recall and precision.  The usual difference ratio depends on the language  pairs we are dealing with. For example, Chinese-  English parallel texts usually have a larger differ-  ence ratio than English-French parallel texts. The  filtering threshold had to be determined empirically,  from the actual observations. For Chinese-English,  a difference up to 50% is tolerated.  2.5.2 Language and  Character Set  It is also obvious that the two files of a pair have  to be in the two languages of interest. By auto-  matically identifying language and character set, we  can filter out the pairs that do not satisfy this basic  criterion. Some Web pages explicitly indicate the  language and the character set. More often such  information is omitted by authors. We need some  language identification tool for this task.  SILC is a language and encoding identification  system developed by the RALI laboratory at the  University of Montreal. It employs a probabilistic  model estimated on tri-grams. Using these mod-  els, the system is able to determine the most proba-  ble language and encoding of a text (Isabelle et al,  1997).  2.5.3 HTML Structure and Alignment  In the STRAND system (Resnik, 1998), the candi-  date pairs are evaluated by aligning them according  to their HTML structures and computing confidence  values. Pairs are assumed to be wrong if they have  too many mismatching markups or low confidence  values.  Comparing HTML structures seems to be a sound  way to evaluate candidate pairs since parallel pairs  usually have similar HTML structures. However, we  also noticed that parallel texts may have quite dif-  ferent HTML structures. One of the reasons is that  the two files may be created using two HTML ed-  itors. For example, one may be used for English  and another for Chinese, depending on the language  handling capability of the editors. Therefore, cau-  tion is required when measuring structure difference  numerically.  Parallel text alignment is still an experimental  area. Measuring the confidence values of an align-  ment is even more complicated. For example, the  alignment algorithm we used in the training of the  statistical translation model produces acceptable  alignment results but it does not provide a confi-  dence value that we can "confidently" use as an eval-  uation criterion. So, for the moment his criterion is  not used in candidate pair evaluation.  3 Generated  Corpus  and Trans la t ion   Mode l  Tra in ing   In this section, we describe the results of our parallel  text mining and translation model training.  3.1 The Corpus  Using the above approach for Chinese-English, 185  candidate sites were searched from the domain hk.  We limited the mining domain to hk because Hong  Kongis a bilingual English-Chinese city where high  quality parallel Web sites exist. Because of the small  number of candidate sites, the host crawler was used  to thoroughly explore each site. The resulting cor-  pus contains 14820 pairs of texts including 117.2Mb  Chinese texts and 136.5Mb English texts. The entire  mining process lasted about a week. Using length  comparison and language identification, we refined  the precision of the corpus to about 90%. The preci-  sion is estimated by examining 367 randomly picked  pairs.  3.2 Statistical Translation Model  Many approaches in computational linguistics try to  extract ranslation knowledge from previous trans-  lation examples. Most work of this kind establishes  probabilistic models from parallel corpora. Based  on one of the statistical models proposed by Brown  et al (1993), the basic principle of our translation  model is the following: given a corpus of aligned sen-  tences, if two words often co-occur in the source and  target sentences, there is a good likelihood that they  are translations of each other. In the simplest case  (model 1), the model earns the probability, p(tls), of  having a word t in the translation of a sentence con-  taining a word s. For an input sentence, the model  then calculates a sequence of words that are most  probable to appear in its translation. Using a sim-  ilar statistical model, Wu (1995) extracted a large-  scale English-Chinese l xicon from the HKUST cor-  23   <s id="00~">  <HTML> <HEAD>  <META HTrP-EQUIV="Content-type"  CONTENT="text/html; charset--iso-8859-1">  <META HTI\'P-EQUIV="Content-language"  CONTENT="Western">  </s>  <s id="0001">  <TITLE>Journal of Primary Education 1996,  VoI., No. l&2, pp. 19-27 </TITLE>  </HEAD>  </s>  <s id="0002">  <BODY BACKGROUND=".Jgif/pejbg.jpg"  TEXT="#000(3(O" BGCOLOR="#ffffff">  <CENTER>  </s>  <s id="0003">  <HI>Journal of Primary Education </HI>  </s>  <s id="0004">  <HR> <B>Volume 6, No l&2, pp. 19-27 (May,  1996) </B> <HR>  </s>  <s id="0005">  <H3>Principles for Redesigning Teacher  Education </H3> Alan TOM </CENTER>  </s>  <s id="0006">  <P> <B> <I> Abstract </I> </B>  </s>  <s id="0000">  <HTML> <HEAD>  <META H\'ITP-EQUW="Content-type"  CONTENT="text/html; charset=bigS">  <META HTTP-EQUIV="Content-language"  CONTENT="zh">  <Is>  <s id="0001">  <TITLE> Journal of Primary Education 1996,  Vol., No. l&2, Page 19-27 </TITLE>  </HEAD>  </s>  <s id="0002">  <BODY BACKGROUND=".Jgif/pejbg.jpg"  TEXT="#000000" BGCOLOR="#ffffff"> <A  HREF="/erdpej/b2g__pej.phtml?URL=%2fen%2fp  ej%2f0601%2f0601019c.htm">  <IMG SRC="/en/gif/kan.gif" ALT="~"   BORDER=0 ALIGN=R IGHT> </A> <CENTER>  </s>  <s id="0003">  <H2>~ ~ 11I ~ O.</H2>  </s>  <s id="0004">  <HR> (~:~h-fv-c?.JLJl) ~,-\\]\'?~..  </s>  <s id="0005">  ~ 19-27\\]~ <I-1R>  </s>  Figure 2: An alignment example using pure length-based method.  pus which is built manually. In our case, the prob-  abilistic translation model will be used for CLIR.  The requirement on our translation model may be  less demanding: itis not absolutely necessary that  a word t with high p(tls ) always be a true trans-  lation of s. It is still useful if t is strongly related  to s. For example, although "railway" is not a true  translation of "train" (in French), it is highly useful  to include "railway" in the translation of a query on  "train". This is one of the reasons why we think a  less controlled parallel corpus can be used to train a  translation model for CLIR.  3.3 Parallel Text Al ignment  Before the mined documents can be aligned into par-  allel sentences, the raw texts have to undergo a se-  ries of some preprocessing, which, to some extent, is  language dependent. For example, the major opera-  tions on the Chinese-English corpus include encod-  ing scheme transformation (for Chinese), sentence  level segmentation, parallel text alignment, Chinese  word segmentation (Nie et al, 1999) and English  expression extraction.  The parallel Web pages we collected from vari-  ous sites are not all of the same quality. Some are  highly parallel and easy to align while others can be  very noisy. Aligning English-Chinese parallel texts  is already very difficult because of the great differ-  ences in the syntactic structures and writing sys-  tems of the two languages. A number of alignment  techniques have been proposed, varying from statis-  tical methods (Brown et al, 1991; Gale and Church,  1991) to lexical methods (Kay and RSscheisen, 1993;  Chen, 1993). The method we adopted is that of  Simard et al (1992). Because it considers both  length similarity and cognateness as alignment cri-  teria, the method is more robust and better able  to deal with noise than pure length-based methods.  Cognates are identical sequences of characters in cor-  responding words in two languages. They are com-  monly found in English and French. In the case of  English-Chinese alignment, where there are no cog-  nates shared by the two languages, only the HTML  markup in both texts are taken as cognates. Be-  cause the HTML structures of parallel pages are nor-  mally similar, the markup was found to be helpful  for alignment.  To illustrate how markup can help with the align-  ment, we align the same pair with both the pure  length-based method of Gale & Church (Fig. 2),  and the method of Simard et al (Fig. 3). First of  all, we observe from the figures that the two texts are  24 <s id="0000">  <HTML> <HEAD>  <META HTTP-EQUIV="Content-type"  CONTENT="text/html; charset=iso-8859-1 ">  <META HTTP-EQUIV="Content-language"  CONTENT="Westem">  </s>  <s id="0001">  <TITLE>Journal of Primary Education 1996,  Vol., No. l&2, pp. 19-27 </TITLE>  </HEAD>  </s>  <s id="0002">  <BODY BACKGROUND=-". Jgif/pejbg.jpg"  TEXT="#000000" BGCOLOR="#ffffff">  <CENTER>  </s>  <s id="0003">  <H 1 >Journal of Primary Education </H 1>  <Is>  <s id="0004">  <HR> <B>Volume 6,No l&2, pp. 19-27 (May,  1996) </B> <HR>  </$>  <s id="0000">  <HTML> <HEAD>  <META HTrP-EQUIV="Content-type"  CONTENT="text/html; charset=big5">  <META H\'lTP-EQUIV="Content-language"  CONTENT="zh">  <Is>  <s id="0001">  :<TITLE> Journal of Primary Education 1996,  Vol., No. l&2, Page 19-27 </TITLE>  </HEAD>  </s>  <s id="0002">  <BODY BACKGROUND=-". Jgiffpejbg.jpg"  TEXT="#O00000" BGCOLOR="#fffffff> <A  HREF="/ergpej/b2g_pej.phtml?URL=%2fen%2fp  ej %2f0601%2 f0601019c.htm">  <IMG SRC="/erdgif/kan.gif" ALT="~k"  BORDER={) ALIGN=R IGHT> </A> <CEHTEIL~  </s>  <s id="0003">  <H2>~k ~ ~ ~\\[1.</H2>  </s>  <s id="0004">  <HR> (~t~-~?-#cJL.~) ,-~?~.  </s>  <s id="0005">  ~ $ ~  19-27 \\]~ <HR>  <\\]s>  <s id="0005"> <s id="0006">  <H3>Principles for Redesigning Teacher <H3>.~ k~4Vt ~\'~ ~ ~J </H3> Alan TOM  Education </H3> Alan TOM </CENTER> </CENTER>  <Is> <Is>  <s id="0006"> <s id="0007">  <P> <B> <I> Abstract </I> </B> <P> <I> <B> ~4\\[- </B> </I> <P>  </s> </s>  Figure 3: An alignment example considering cognates.  divided into sentences. The sentences are marked by  <s id="xxxx"> and </s>.  Note that we determine  sentences not only by periods, but also by means of  HTML markup.  We further notice that it is difficult to align sen-  tences 0002. The sentence in the Chinese page is  much longer than its counterpart in the English page  because some additional information (font) is added.  The length-based method thus tends to take sen-  tence 0002, 0003, and 0004 in the English page as  the translation of sentence 0002 in the Chinese page  (Fig. 2), which is wrong. This in turn provocated  the three following incorrect alignments. As we can  see in Fig. 3, the cognate method did not make the  same mistake because of the noise in sentence 0002.  Despite their large length difference, the two 0002  sentences are still aligned as a 1-1 pair, because the  sentences in the following 4 alignments (0003 - 0003;  0004 - 0004, 0005; 0005 - 0006; 0006 - 0007) have  rather similar HTML markups and are taken by the  program to be the most likely alignments.  Beside HTML markups, other criteria may also  be incorporated. For example, it would be helpful  to consider strong correspondence b tween certain  English and Chinese words, as in (Wu, 1994). We  hope to implement such correspondences in our fu-  ture research.  3.4 Lex icon  Eva luat ion   To evaluate the precision of the English-Chinese  translation model trained on the Web corpus, we  examined two sample lexicons of 200 words, one in  each direction. The 200 words for each lexicon were  randomly selected from the training source. We ex-  amined the most probable translation for each word.  The Chinese-English lexicon was found to have a  precision of 77%. The English-Chinese l xicon has a higher precision of 81.5%. Part of the lexicons  are shown in Fig. 4, where t / f  indicates whether a  translation is true or false.  These precisions seem to be reasonably high.  They are quite comparable to that obtained by Wu  (1994) using a manual Chinese-English parallel cor-  pus.  3.5 Effect  o f  S topwords   We also found that stop-lists have significant effect  on the translation model. Stop-list is a set of the  most frequent words that we remove from the train-  2fi  English word  a .n l .   access  adaptation  add  adopt  agent  agree  airline  amendment  , appliance  apply  attendance  auditor  - ,average  base_on  t/f  t  f  t  t  t  t  t  t  t  t  t  t  f  t  f  Translmion Probability Chinese word  ~\'~- 0.201472 ~t l :   ~"  0.071705 "~"   ~f~.,~ 0.179633 JllL~  0.317435  ~ 0.231637 ~.~  1~tA~ 0.224902 4J~\'~  0.36569  0.344001  0.367518  J~ 4~ 0.136319  i~.~I 0.19448 J~   ~\',1~ 0.171769 ,~- JJ~  *~ 0.15011 -~-~  ~- ~ 0.467646 * *~  0.107304  Figure 4: Part of the evaluation lexicons.  t/f  t  t  t  t  t  f  t  f  t  t  t  t  t  t  t  Translation Probability  office 0.375868  protection 0.343071  report 0.358592  prepare 0.189513  loca l  0.421837  follow 0.023685  standard 0.445453  adu l t  0.044959  inadequate 0.093012  part 0.313676  financial 0.16608  visit 0.309642  bill 0.401997  vehicle 0.467034  saving 0.176695  Figure 5: Effect of stop lists in C-E translation.  ing source. Because these words exist in most align-  ments, the statistical model cannot derive correct  translations for them. More importantly, their ex-  istence greatly affects the accuracy of other transla-  tions. They can be taken as translations for many  words.  A priori, it would seem that both the English and  Chinese stop-lists hould be applied to eliminate the  noise caused by them. Interestingly, from our ob-  servation and analysis we concluded that for better  precision, only the stop-list of the target language  should be applied in the model training.  We first explain why the stop-list of the target lan-  guage has to be applied. On the left side of Fig. 5,  if the Chinese word C exists in the same alignments  with the English word E more than any other Chi-  nese words, C will be the most probable translation  for E. Because of their frequent appearance, some  Chinese stopwords may have more chances tobe in  the same alignments with E. The probability of the  translation E --+ C is then reduced (maybe ven less  than those of the incorrect ones). This is the reason  why many English words are translated to "~ \'  (of)  by the translation model trained without using the  Chinese stop-list.  We also found that it is not necessary to remove  the stopwords of the source language. In fact, as il-  lustrated on the right side of Fig. 5, the existence of  the English stopwords has two effects on the proba-  bility of the translation E -~ C:  1 They may often be found together with the Chi-  nese word C. Owing to the Expectation Maxi-  mization algorithm, the probability of E -~ C  may therefore be reduced.  2 On the other hand, there is a greater likelihood  that English stopwords will be found together  with the most frequent Chinese words. Here,  we use the term "Chinese frequent words" in-  stead of "Chinese stopwords" because ven if a  stop-list is applied, there may still remain some  common words that have the same effect as the  stopwords. The coexistence ofEnglish and Chi-  nese frequent words reduces the probability that  the Chinese frequent words are the translations  of E, and thus raise the probability of E -+ C.  The second effect was found to be more signifi-  cant than the first, since the model trained without  the English stopwords has better precision than the  model trained with the English stopwords. For the  correct ranslations given by both models, the model  26 Mono-Lingual IR  Translation Model  Dictionary  C-E CLIR  0.3861  0.1504 (39.0%mono)  0.1530 (39.6%mono)  0.2583 (66.9%mono)  E-C CLIR  0.3976  0.1841 (46.3%mono)  0.1427 (35.9%mono)  0.2232 (56.1%mono)  Table 1: CLIR results.  trained without considering the English stopwords  gives higher probabilities.  4 Eng l i sh -Ch inese  CL IR  Resu l ts   Our final goal was to test the performance of the  translation models trained on the Web parallel cor-  pora in CLIR. We conducted CLIR experiments u -  ing the Smart IR system.  4.1 Results   The English test corpus (for C-E CLIR) was the  AP corpus used in TREC6 and TREC7. The short  English queries were translated manually into Chi-  nese and then translated back to English by the  translation model. The Chinese test corpus was the  one used in the TREC5 and TREC6 Chinese track.  It contains both Chinese queries and their English  translations.  Our experiments on these two corpora produced  the results hown in Tab. 1. The precision of mono-  lingual IR is given as benchmark. In both E-C and  C-E CLIR, the translation model achieved around  40% of monolingual precision. To comparewith the  dictionary-based approach, we employed a Chinese-  English dictionary, CEDICT (Denisowski, 1999),  and an English-Chinese online dictionary (Anony-  mous, 1999a) to translate queries. For each word  of the source query, all the possible translations  given by the dictionary are included in the translated  query. The Chinese-English dictionary has about  the same performace as the translation model, while  the English-Chinese dictionary has lower precision  than that of the translation model.  We also tried to combine the translations given by  the translation model and the dictionary. In both  C-E and E-C CLIR, significant improvements were  achieved (as shown in Tab. 1). The improvements  show that the translations given by the translation  model and the dictionary complement each other  well for IR purposes. The translation model may  give either exact ranslations orincorrect but related  words. Even though these words are not correct in  the sense of translation, they are very possibly re-  lated to the subject of the query and thus helpful  for IR purposes. The dictionary-based approach ex-  pands a query along another dimension. It gives  all the possible translations for each word including  those that are missed by the translation model.  4.2 Comparison Wi th  MT Systems  One advantage of a parallel text-based translation  model is that it is easier to build than an MT system.  Now that we have examined the CLIR performance  of the translation model, we will compare it with  two existing MT systems. Both systems were tested  in E-C CLIR.  4.2.1 Sunshine WebTran Server  Using the Sunshine WebTran server (Anonymous,  1999b), an online Engiish-Chinese MT system, to  translate the 54 English queries, we obtained an  average precision of 0.2001, which is 50.3% of the  mono-lingual precision. The precision is higher than  that obtained using the translation model (0.1804)  or the dictionary (0.1427) alone, but lower than the  precison obtained using them together (0.2232).  4.2.2 Transperfect  Kwok (1999) investigated the CLIR performance of an English-Chinese MT software called Transper-  fect, using the same TREC Chinese collection as we  used in this study. Using the MT software alone,  Kwok achieved 56% of monolingual precision. The  precision is improved to 62% by refining the trans-  lation with a dictionary. Kwok also adopted pre-  translation query expansion, which further improved  the precison to 70% of the monolingual results.  In our case, the best E-C CLIR precison using the  translation model (and dictionary) is 56.1%. It is  lower than what Kwok achieved using Transperfect,  however, the difference is not large.  4.3 Further  Problems  The Chinese-English translation model has a fax  lower CLIR performance than that of the English-  French model established using the same method  (Nie et al, 1999). The principal reason for this is the  fact that English and Chineseare much more differ-  ent than English and French. This problem surfaced  in many phases of this work, from text alignment to  query translation. Below, we list some further fac-  tors affecting CLIR precision.  ? The Web-collected corpus is noisy and it is dif-  ficult to align English-Chinese t xts. The align-  ment method we employed has performed more  poorly than on English-French alignment. This  in turn leads to poorer performance of the trans-  lation model. In general, we observe a higher  27  variability in Chinese-English translations than  in English-French translations.  ? For E-C CLIR, although queries in both lan-  guages were provided, the English queries were  not strictly translated from the original Chi-  nese ones. For example, A Jg ,~ (human right  situation) was translated into human right is-  sue. We cannot expect he translation model  to translate issue back to ~ (situation).  ? The training source and the CLIR collections  were from different domains. The Web cor-  pus are retrieved from the parallel sites in Hong  Kong while the Chinese collection is from Peo-  ple\'s Daily and Xinhua News Agency, which are  published in mainland China. As the result,  some important erms such as ~$ $ (most-  favored-nation) and --- I!! ~ ~ (one-nation-two-  systems) in the collection are not known by the  model.  5 Summary   The goal of this work was to investigate he feasibil-  ity of using a statistical translation model trained on  a Web-collected corpus to do English-Chinese CLIR.  In this paper, we have described the algorithm and  implementation we used for parallel text mining,  translation model training, and some results we ob-  tained in CLIR experiments. Although further work  remains to be done, we can conclude that it is pos-  sible to automatically construct a Chinese-English  parallel corpus from the Web. The current system  can be easily adapted to other language pairs. De-  spite the noisy nature of the corpus and the great  difference in the languages, the evaluation lexicons  generated by the translation model produced accept-  able precision. While the current CLIR results are  not as encouraging asthose of English-French CLIR,  they could be improved in various ways, such as im-  proving the alignment method by adapting cognate  definitions to HTML markup, incorporating a lexi-  con and/or removing some common function words  in translated queries.  We hope to be able to demonstrate in the near  future that a fine-tuned English-Chinese translation  model can provide query translations for CLIR with  the same quality produced by MT systems. '
0.0	P15-1001	I05-3012	b'1 Introduction Neural machine translation (NMT) is a recently introduced approach to solving machine translation (Kalchbrenner and Blunsom, 2013; Bahdanau et al, 2015; Sutskever et al, 2014). In neural machine translation, one builds a single neural network that reads a source sentence and generates its translation. The whole neural network is jointly trained to maximize the conditional probability of a correct translation given a source sentence, using the bilingual corpus. The NMT models have shown to perform as well as the most widely used conventional translation systems (Sutskever et al, 2014; Bahdanau et al, 2015). Neural machine translation has a number of advantages over the existing statistical machine translation system, specifically, the phrase-based system (Koehn et al, 2003). First, NMT requires a minimal set of domain knowledge. For instance, all of the models proposed in (Sutskever et al, 2014), (Bahdanau et al, 2015) or (Kalchbrenner and Blunsom, 2013) do not assume any linguistic property in both source and target sentences except that they are sequences of words. Second, the whole system is jointly trained to maximize the translation performance, unlike the existing phrase-based system which consists of many separately trained features whose weights are then tuned jointly. Lastly, the memory footprint of the NMT model is often much smaller than the existing system which relies on maintaining large tables of phrase pairs. Despite these advantages and promising results, there is a major limitation in NMT compared to the existing phrase-based approach. That is, the number of target words must be limited. This is mainly because the complexity of training and using an NMT model increases as the number of target words increases. A usual practice is to construct a target vocabulary of the K most frequent words (a socalled shortlist), where K is often in the range of 30k (Bahdanau et al, 2015) to 80k (Sutskever et al., 2014). Any word not included in this vocabulary is mapped to a special token representing an unknown word [UNK]. This approach works well when there are only a few unknown words in the target sentence, but it has been observed 1 that the translation performance degrades rapidly as the number of unknown words increases (Cho et al, 2014a; Bahdanau et al, 2015). In this paper, we propose an approximate training algorithm based on (biased) importance sampling that allows us to train an NMT model with a much larger target vocabulary. The proposed algorithm effectively keeps the computational complexity during training at the level of using only a small subset of the full vocabulary. Once the model with a very large target vocabulary is trained, one can choose to use either all the target words or only a subset of them. We compare the proposed algorithm against the baseline shortlist-based approach in the tasks of English?French and English?German translation using the NMT model introduced in (Bahdanau et al, 2015). The empirical results demonstrate that we can potentially achieve better translation performance using larger vocabularies, and that our approach does not sacrifice too much speed for both training and decoding. Furthermore, we show that the modeltrained with this algorithm gets the best translation performance yet achieved by single NMT models on the WMT?14 English?French translation task. 2 Neural Machine Translation and Limited Vocabulary Problem In this section, we briefly describe an approach to neural machine translation proposed recently in (Bahdanau et al, 2015). Based on this description we explain the issue of limited vocabularies in neural machine translation. 2.1 Neural Machine Translation Neural machine translation is a recently proposed approach to machine translation, which uses a single neural network trained jointly to maximize the translation performance (Forcada and ? Neco, 1997; Kalchbrenner and Blunsom, 2013; Cho et al., 2014b; Sutskever et al, 2014; Bahdanau et al, 2015). Neural machine translation is often implemented as the encoder?decoder network. The encoder reads the source sentence x = (x 1 , . . . , x T ) and encodes it into a sequence of hidden states h = (h 1 , ? ? ? , h T ): h t = f (x t , h t?1 ) . (1) Then, the decoder, another recurrent neural network, generates a corresponding translation y = (y 1 , ? ? ? , y T ? ) based on the encoded sequence of hidden states h: p(y t | y <t , x) ? exp {q (y t?1 , z t , c t )} , (2) where z t = g (y t?1 , z t?1 , c t ) , (3) c t = r (z t?1 , h 1 , . . . , h T ) , (4) and y <t = (y 1 , . . . , y t?1 ). The whole model is jointly trained to maximize the conditional log-probability of the correct translation given a source sentence with respect to the parameters ? of the model: ? ? = argmax ? N ? n=1 T n ? t=1 log p(y n t | y n <t , x n ), where (x n , y n ) is the n-th training pair of sentences, and T n is the length of the n-th target sentence (y n ). 2.1.1 Detailed Description In this paper, we use a specific implementation of neural machine translation that uses an attention mechanism, as recently proposed in (Bahdanau et al., 2015). In (Bahdanau et al, 2015), the encoder in Eq. (1) is implemented by a bi-directional recurrent neural network such that h t = [ ?? h t ; ?? h t ] , where ?? h t = f ( x t , ?? h t+1 ) , ?? h t = f ( x t , ?? h t?1 ) . They used a gated recurrent unit for f (see, e.g., (Cho et al, 2014b)). The decoder, at each time, computes the context vector c t as a convex sum of the hidden states (h 1 , . . . , h T ) with the coefficients ? 1 , . . . , ? T computed by ? t = exp {a (h t , zt?1 )} ? k exp {a (h k , z t?1 )} , (5) where a is a feedforward neural network with a single hidden layer. A new hidden state z t of the decoder in Eq. (3) is computed based on the previous hidden state z t?1 , previous generated symbol y t?1 and the computed 2 context vector c t . The decoder also uses the gated recurrent unit, as the encoder does. The probability of the next target word in Eq. (2) is then computed by p(y t | y <t , x) = 1 Z exp { w > t ? (y t?1 , z t , c t ) + b t } , (6) where ? is an affine transformation followed by a nonlinear activation, and w t and b t are respectively the target word vector and the target word bias. Z is the normalization constant computed by Z = ? k:y k ?V exp { w > k ? (y t?1 , z t , c t ) + b k } , (7) where V is the set of all the target words. For the detailed description of the implementation, we refer the reader to the appendix of (Bahdanau et al, 2015). 2.2 Limited Vocabulary Issue and Conventional Solutions One of the main difficulties in training this neural machine translation model is the computational complexity involved in computing the target word probability (Eq. (6)). More specifically, we need to compute the dot product between the feature ? (y t?1 , z t , c t ) and the word vector w t as many times as there are words in a target vocabulary in order to compute the normalization constant (the denominator in Eq. (6)). This has to be done for, on average, 20?30 words per sentence, which easily becomes prohibitively expensive even with a moderate number of possible target words. Furthermore, the memory requirement grows linearly with respect to the number of target words. This has been a major hurdle for neural machine translation, compared to the existing non-parametric approaches such as phrase-based translation systems. Recently proposed neural machine translation models, hence, use a shortlist of 30k to 80k most frequent words (Bahdanau et al, 2015; Sutskever et al, 2014). This makes training more feasible, but comes with a number of problems. First of all, the performance of the model degrades heavily if the translation of a source sentence requires many words that are not included in the shortlist (Cho et al, 2014a). This also affects the performance evaluation of the system which is often measured by BLEU. Second, the first issue becomes more problematic with languages that have a rich set of words such as German or other highly inflected languages. There are two model-specific approaches to this issue of large target vocabulary. The first approach is to stochastically approximate the target word probability. This has been proposed recently in (Mnih and Kavukcuoglu, 2013; Mikolov et al, 2013) based on noise-contrastive estimation (Gutmann and Hyvarinen, 2010). In the second approach, thetarget words are clustered into multiple classes, or hierarchical classes, and the target probability p(y t |y <t , x) is factorized as a product of the class probability p(c t |y <t , x) and the intraclass word probability p(y t |c t , y <t , x). This reduces the number of required dot-products into the sum of the number of classes and the words in a class. These approaches mainly aim at reducing the computational complexity during training, but do not often result in speed-up when decoding a translation during test time. 1 Other than these model-specific approaches, there exist translation-specific approaches. A translation-specific approach exploits the properties of the rare target words. For instance, Luong et al proposed such an approach for neural machine translation (Luong et al, 2015). They replace rare words (the words that are not included in the shortlist) in both source and target sentences into corresponding ?OOV n ? tokens using the word alignment model. Once a source sentence is translated, each ?OOV n ? in the translation will be replaced based on the source word marked by the corresponding ?OOV n ?. It is important to note that the modelspecific approaches and the translation-specific approaches are often complementary and can be used together to further improve the translation performance and reduce the computational complexity. 3 Approximate Learning Approach to Very Large Target Vocabulary 3.1 Description In this paper, we propose a model-specific approach that allows us to train a neural machine translation model with a very large target vocabulary. With the proposed approach, the compu1 This is due to the fact that the beam search requires the conditional probability of every target word at each time step regardless of the parametrization of the output probability. 3 tational complexity of training becomes constant with respect to the size of the target vocabulary. Furthermore, the proposed approach allows us to efficiently use a fast computing device with limited memory, such as a GPU, to train a neural machine translation model with a much larger target vocabulary. As mentioned earlier, the computational inefficiency of training a neural machine translation model arises from the normalization constant in Eq. (6). In order to avoid the growing complexity of computing the normalization constant, we propose here to use only a small subset V ? of the target vocabulary at each update. The proposed approach is based on the earlier work of (Bengio and S?en?ecal, 2008). Let us consider the gradient of the logprobability of the output in Eq. (6). The gradient is composed of a positive and negative part: ? log p(y t | y <t , x) (8) =?E(y t )? ? k:y k ?V p(y k | y <t , x)?E(y k ), where we define the energy E as E(y j ) = w > j ? (y j?1 , z j , c j ) + b j . The second, or negative, term of the gradient is in essence the expected gradient of the energy: E P [?E(y)] , (9) where P denotes p(y| y <t , x). The main idea of the proposed approach is to approximate this expectation, or the negative term of the gradient, by importance sampling with a small number of samples. Given a predefined proposal distribution Q and a set V ? of samples from Q, we approximate the expectation in Eq. (9) with E P [?E(y)] ? ? k:y k ?V ? ? k ? k ? :y k ? ?V ? ? k ? ?E(y k ), (10) where ? k = exp {E(y k )? logQ(y k )} . (11) This approach allows us to compute the normalization constant during training using only a small subset of the target vocabulary, resulting in much lower computational complexity for each parameter update. Intuitively, at each parameter update, we update only the vectors associated with the correct word w t and with the sampled words in V ? . Once training is over, we can use the full target vocabulary to compute the output probability of each target word. Although the proposed approach naturally addresses the computational complexity, using this approach naively does not guarantee that the number of parameters being updated for each sentence pair, which includes multiple target words, is bounded nor can be controlled. This becomes problematic when training is done, for instance, on a GPU with limited memory. In practice, hence, we partition the training corpus and define a subset V ? of the target vocabulary for each partition prior to training. Before training begins, we sequentially examine each target sentence in the training corpus and accumulate unique target words until the number of unique target words reaches the predefined threshold ? . The accumulated vocabulary will be used for this partition of the corpus during training. We repeat this until the end of the training set is reached. Let us refer to the subset of target words used for the i-th partition by V ? i . This may be understood as having a separate proposal distribution Q i for each partition of the training corpus. The distribution Q i assigns equal probability mass to all the target words included in the subset V ? i , and zero probability mass to all the other words, i.e., Q i (y k ) = ? ? ? 1 | V ? i | if y t ? V ? i 0 otherwise. This choice of proposal distribution cancels out the correction term ? logQ(y k ) from the importance weight in Eqs. (10)?(11), which makes the proposed approach equivalent to approximating the exact output probability in Eq. (6) with p(y t | y <t , x) = exp { w > t ? (y t?1 , z t , c t ) + b t } ? k:y k ?V ? exp { w > k ? (y t?1 , z t , c t ) + b k } . It should be noted that this choice of Q makes the estimator biased. The proposed procedure results in speed up against usual importance sampling, asit exploits the advantage of modern computers in doing matrix-matrix vs matrix-vector multiplications. 4 3.1.1 Informal Discussion on Consequence The parametrization of the output probability in Eq. (6) can be understood as arranging the vectors associated with the target words such that the dot product between the most likely, or correct, target word?s vector and the current hidden state is maximized. The exponentiation followed by normalization is simply a process in which the dot products are converted into proper probabilities. As learning continues, therefore, the vectors of all the likely target words tend to align with each other but not with the others. This is achieved exactly by moving the vector of the correct word in the direction of ? (y t?1 , z t , c t ), while pushing all the other vectors away, which happens when the gradient of the logarithm of the exact output probability in Eq. (6) is maximized. Our approximate approach, instead, moves the word vectors of the correct words and of only a subset of sampled target words (those included in V ? ). 3.2 Decoding Once the model is trained using the proposed approximation, we can use the full target vocabulary when decoding a translation given a new source sentence. Although this is advantageous as it allows the trained model to utilize the whole vocabulary when generating a translation, doing so may be too computationally expensive, e.g., for realtime applications. Since training puts the target word vectors in the space so that they align well with the hidden state of the decoder only when they are likely to be a correct word, we can use only a subset of candidate target words during decoding. This is similar to what we do during training, except that at test time, we do not have access to a set of correct target words. The most na??ve way to select a subset of candidate target words is to take only the top-K most frequent target words, where K can be adjusted to meet the computational requirement. This, however, effectively cancels out the whole purpose of training a model with a very large target vocabulary. Instead, we can use an existing word alignment model to align the source and target words in the training corpus and build a dictionary. With the dictionary, for each source sentence, we construct a target word set consisting of the K-most frequent words (according to the estimated unigram probability) and, using the dictionary, at most K ? likely target words for each source word. K and K ? may be chosen either to meet the computational requirement or to maximize the translation performance on the development set. We call a subset constructed in either of these ways a candidate list. 3.3 Source Words for Unknown Words In the experiments, we evaluate the proposed approach with the neural machine translation model called RNNsearch (Bahdanau et al, 2015) (see Sec. 2.1.1). In this model, as a part of decoding process, we obtain the alignments between the target words and source locations via the alignment model inEq. (5). We can use this feature to infer the source word to which each target word was most aligned (indicated by the largest ? t in Eq. (5)). This is especially useful when the model generated an [UNK] token. Once a translation is generated given a source sentence, each [UNK] may be replaced using a translation-specific technique based on the aligned source word. For instance, in the experiment, we try replacing each [UNK] token with the aligned source word or its most likely translation determined by another word alignment model. Other techniques such as transliteration may also be used to further improve the performance (Koehn, 2010). 4 Experiments We evaluate the proposed approach in English?French and English?German translation tasks. We trained the neural machine translation models using only the bilingual, parallel corpora made available as a part of WMT?14. For each pair, the datasets we used are: ? English?French: 2 ? Common Crawl ? News Commentary ? Gigaword ? Europarl v7 ? UN ? English?German: ? Common Crawl ? News Commentary ? Europarl v7 2 The preprocessed data can be found and downloaded from http://www-lium.univ-lemans.fr/ ? schwenk/nnmt-shared-task/README. 5 English-French English-German Train Test Train Test 15k 93.5 90.8 88.5 83.8 30k 96.0 94.6 91.8 87.9 50k 97.3 96.3 93.7 90.4 500k 99.5 99.3 98.4 96.1 All 100.0 99.6 100.0 97.3 Table 1: Data coverage (in %) on target-side corpora for different vocabulary sizes. ?All? refers to all the tokens in the training set. To ensure fair comparison, the English?French corpus, which comprises approximately 12 million sentences, is identical to the one used in (Kalchbrenner and Blunsom, 2013; Bahdanau et al, 2015; Sutskever et al, 2014). As for English?German, the corpus was preprocessed, in a manner similar to (Peitz et al, 2014; Li et al, 2014), in order to remove many poorly translated sentences. We evaluate the models on the WMT?14 test set (news-test 2014), 3 while the concatenation of news-test-2012 and news-test-2013 is used for model selection (development set). Table 1 presents data coverage w.r.t. the vocabulary size, on the target side. Unless mentioned otherwise, all reported BLEU scores (Papineni et al, 2002) are computed with the multi-bleu.perl script 4 on the cased tokenized translations. 4.1 Settings As a baseline for English?French translation, we use the RNNsearch model proposed by (Bahdanau et al, 2015), with 30k source and target words. 5 Another RNNsearch model is trained for English?German translation with 50k source and target words. For each language pair, we train another set of RNNsearch models with much larger vocabularies of 500k source and target words, using the proposed approach. We call these models RNNsearch-LV. We vary the size of the shortlist used during training (? in Sec. 3.1). We tried 3 To compare with previous submissions, we use the filtered test sets. 4 https://github.com/moses-smt/ mosesdecoder/blob/master/scripts/ generic/multi-bleu.perl 5 The authors of (Bahdanau et al, 2015) gave us access to their trained models. We chose the best one on the validation set and resumed training. 15k and 30k for English?French, and 15k and 50k for English?German. We later report the results for thebest performance on the development set, with models generally evaluated every twelve hours. The training speed is approximately the same as for RNNsearch. Using a 780 Ti or Titan Black GPU, we could process 100k mini-batches of 80 sentences in about 29 and 39 hours respectively for ? = 15k and ? = 50k. For both language pairs, we also trained new models, with ? = 15k and ? = 50k, by reshuffling the dataset at the beginning of each epoch. While this causes a non-negligible amount of overhead, such a change allows words to be contrasted with different sets of other words each epoch. To stabilize parameters other than the word embeddings, at the end of the training stage, we freeze the word embeddings and tune only the other parameters for approximately two more days after the peak performance on the development set is observed. This helped increase BLEU scores on the development set. We use beam search to generate a translation given a source. During beam search, we keep a set of 12 hypotheses and normalize probabilities by the length of the candidate sentences, as in (Cho et al, 2014a). 6 The candidate list is chosen to maximize the performance on the development set, for K ? {15k, 30k, 50k} and K ? ? {10, 20}. As explained in Sec. 3.2, we test using a bilingual dictionary to accelerate decoding and to replace unknown words in translations. The bilingual dictionary is built using fast align (Dyer et al., 2013). We use the dictionary only if a word starts with a lowercase letter, and otherwise, we copy the source word directly. This led to better performance on the development sets. Note on ensembles For each language pair, we began training four models from each of which two points corresponding to the best and secondbest performance on the development set were collected. We continued training from each point, while keeping the word embeddings fixed, until the best development performance was reached, and took the model at this point as a single model in an ensemble. This procedure resulted in a total of eight models from which we averaged the length-normalized log-probabilities. Since much of training had been shared, the composition of 6 These experimental details differ from (Bahdanau et al, 2015). 6 RNNsearch RNNsearch-LV Google Phrase-based SMT Basic NMT 29.97 (26.58) 32.68 (28.76) 30.6 ? 33.3 ? 37.03 ? +Candidate List ? 33.36 (29.32) ? +UNK Replace 33.08 (29.08) 34.11 (29.98) 33.1 ? +Reshuffle (?=50k) ? 34.60 (30.53) ? +Ensemble ? 37.19 (31.98) 37.5 ? (a) English?French RNNsearch RNNsearch-LV Phrase-based SMT Basic NMT 16.46 (17.13) 16.95 (17.85) 20.67 \x05 +Candidate List ? 17.46 (18.00) +UNK Replace 18.97 (19.16) 18.89 (19.03) +Reshuffle ? 19.40 (19.37) +Ensemble ? 21.59 (21.06) (b) English?German Table 2: The translation performances in BLEU obtained by different models on (a) English?French and (b) English?German translation tasks. RNNsearch is the model proposed in (Bahdanau et al, 2015), RNNsearch-LV is the RNNsearch trained with the approach proposed in this paper, and Google is the LSTM-based model proposed in (Sutskever et al,2014). Unless mentioned otherwise, we report singlemodel RNNsearch-LV scores using ? = 30k (English?French) and ? = 50k (English?German). For the experiments we have run ourselves, we show the scores on the development set as well in the brackets. (?) (Sutskever et al, 2014), (?) (Luong et al, 2015), (?) (Durrani et al, 2014), (?) Standard Moses Setting (Cho et al, 2014b), (\x05) (Buck et al, 2014). such ensembles may be sub-optimal. This is supported by the fact that higher cross-model BLEU scores (Freitag et al, 2014) are observed for models that were partially trained together. 4.2 Translation Performance In Table 2, we present the results obtained by the trained models with very large target vocabularies, and alongside them, the previous results reported in (Sutskever et al, 2014), (Luong et al, 2015), (Buck et al, 2014) and (Durrani et al, 2014). Without translation-specific strategies, we can clearly see that the RNNsearch-LV outperforms the baseline RNNsearch. In the case of the English?French task, RNNsearch-LV approached the performance level of the previous best single neural machine translation (NMT) model, even without any translationspecific techniques (Sec. 3.2?3.3). With these, however, the RNNsearch-LV outperformed it. The performance of the RNNsearch-LV is also better than that of a standard phrase-based translation system (Cho et al, 2014b). Furthermore, by combining 8 models, we were able to achieve a translation performance comparable to the state of the art, measured in BLEU. For English?German, the RNNsearch-LV outperformed the baseline before unknown word replacement, but after doing so, the two systems performed similarly. We could reach higher largevocabulary single-model performance by reshuffling the dataset, but this step could potentially also help the baseline. In this case, we were able to surpass the previously reported best translation result on this task by building an ensemble of 8 models. With ? = 15k, the RNNsearch-LV performance worsened a little, with best BLEU scores, without reshuffling, of 33.76 and 18.59 respectively for English?French and English?German. The English?German ensemble described in this paper has also been used for the shared translation task of the 10 th Workshop on Statistical Machine Translation (WMT?15), where it was ranked first in terms of BLEU score. The translations by this ensemble can be found online. 7 4.3 Analysis 4.3.1 Decoding Speed In Table 3, we present the timing information of decoding for different models. Clearly, decoding from RNNsearch-LV with the full target vocab7 http://matrix.statmt.org/matrix/ output/1774?run_id=4079 7 CPU ? GPU ? RNNsearch 0.09 s 0.02 s RNNsearch-LV 0.80 s 0.25 s RNNsearch-LV 0.12 s 0.05 s +Candidate list Table 3: The average per-word decoding time. Decoding here does not include parameter loading and unknown word replacement. The baseline uses 30k words. The candidate list is built with K = 30k and K ? = 10. (?) i7-4820K (single thread), (?) GTX TITAN Black ulary is slowest. If we use a candidate list for decoding each translation, the speed of decoding substantially improves and becomes close to the baseline RNNsearch. A potential issue with using a candidate list is that for each source sentence, we must re-build a target vocabulary andsubsequently replace a part of the parameters, which may easily become timeconsuming. We can address this issue, for instance, by building a common candidate list for multiple source sentences. By doing so, we were able to match the decoding speed of the baseline RNNsearch model. 4.3.2 Decoding Target Vocabulary For English?French (? = 30k), we evaluate the influence of the target vocabulary when translating the test sentences by using the union of a fixed set of 30k common words and (at most) K ? likely candidates for each source word according to the dictionary. Results are presented in Figure 1. With K ? = 0 (not shown), the performance of the system is comparable to the baseline when not replacing the unknown words (30.12), but there is not as much improvement when doing so (31.14). As the large vocabulary model does not predict [UNK] as much during training, it is less likely to generate it when decoding, limiting the effectiveness of the post-processing step in this case. With K ? = 1, which limits the diversity of allowed uncommon words, BLEU is not as good as with moderately larger K ? , which indicates that our models can, to some degree, correctly choose between rare alternatives. If we rather use K = 50k, as we did for testing based on validation performance, the improvement over K ? = 1 is approximately 0.2 BLEU. When validating the choice of K, we found it to be correlated with the value of ? used during 10 0 10 1 10 2 10 3 K? 32.8 33.0 33.2 33.4 33.6 33.8 34.0 34.2 B L E U s c o r e With UNK replacement Without UNK replacement Figure 1: Single-model test BLEU scores (English?French) with respect to the number of dictionary entries K ? allowed for each source word. training. For example, on the English?French validation set, with ? = 15k (and K ? = 10), the BLEU score is 29.44 with K = 15k, but drops to 29.19 and 28.84 respectively for K = 30k and 50k. For ? = 30k, the score increases moderately from K = 15k to K = 50k. A similar effect was observed for English?German and on the test sets. As our implementation of importance sampling does not apply the usual correction to the gradient, it seems beneficial for the test vocabularies to resemble those used during training.5 Conclusion In this paper, we proposed a way to extend the size of the target vocabulary for neural machine translation. The proposed approach allows us to train a model with much larger target vocabulary without any substantial increase in computational complexity. It is based on the earlier work in (Bengio and S?en?ecal, 2008) which used importance sampling to reduce the complexity of computing the normalization constant of the output word probability in neural language models. On English?French and English?German translation tasks, we observed that the neural machine translation models trained using the proposed method performed as well as, or better than, those using only limited sets of target words, even when replacing unknown words. As performance of the RNNsearch-LV models increased when only a selected subset of the target vocabulary was used during decoding, this makes the proposed learning algorithm more practical. When measured by BLEU, our models showed translation performance comparable to the 8 state-of-the-art translation systems on both the English?French task and English?German task. On the English?French task, a model trained with the proposed approach outperformed the best single neural machine translation (NMT) model from (Luong et al, 2015) by approximately 1 BLEU point. The performance of the ensemble of multiple models, despite its relatively less diverse composition, is approximately 0.3 BLEU points away from the best system (Luong et al, 2015). On the English?German task, the best performance of 21.59 BLEU by our model is higher than that of the previous state of the art (20.67) reported in (Buck et al, 2014). Finally, we release the source code used in our experiments to encourage progress in neural machine translation. 8 Acknowledgments The authors would like to thank the developers of Theano (Bergstra et al, 2010; Bastien et al, 2012). We acknowledge the support of the following agencies for research funding and computing support: NSERC, Calcul Qu?ebec, Compute Canada, the Canada Research Chairs, CIFAR and Samsung.'	b'1 Introduction  WSD tries to resolve lexical ambiguity which  refers to the fact that a word may have multiple  meanings such as the word ?walk? in  ?Walk or  Bike to school? and ?BBC Education Walk  Through Time?, or the Chinese word  ???? in   ??????(?local government?) and ????? ????(?He is also partly right?). WSD tries to  automatically assign an appropriate sense to an  occurrence of a word in a given context.   Various approaches have been proposed to deal  with the word sense disambiguation problem  including rule-based approaches, knowledge or  dictionary based approaches, corpus-based approaches, and hybrid approaches. Among these  approaches, the supervised corpus-based approach had been applied and discussed by many  researches ([2-8]). According to [1], the corpusbased supervised machine learning methods are  the most successful approaches to WSD where  contextual features have been used mainly to  distinguish ambiguous words in these methods.  However, word occurrences in the context are  too diverse to capture the right pattern, which  means that the dimension of contextual words  will be very large when all words in the training  samples are used for WSD [14]. Certain  uninformative features will weaken the discriminative power of a classifier resulting in a  lower precision rate. To narrow down the context, we propose to use collocations as contextual information as defined in Section 3.1.2. It is  generally understood that the sense of an ambiguous word is unique in a given collocation  [19]. For example, ???? means ?burden? but  not ?baggage? when it appears in the collocation  ?????? (? burden of thought?).  In this paper, we apply a classifier to combine  the local features of collocations which contain  the target word with other contextual features to  discriminate the ambiguous words. The intuition  is that when the target context captures a collocation, the influence of other dimensions of 87 contextual words can be reduced or even ignored. For example, in the expression ????? ?????? ? (?terrorists burned down the  gene laboratory?), the influence of contextual  word ???? (?gene?) should be reduced to work  on the target word ???? because ?????? is  a collocation whereas ???? and ???? are not  collocations even though they do co-occur. Our  intention is not to generally replace contextual  information by collocation only. Rather, we  would like to use collocation as an additional  feature in WSD. We still make use of other  contextual features because of the following reasons.  Firstly, contextual information is proven to be  effective for WSD in the previous research  works. Secondly, collocations may be independent on the training corpus and a sentence in consideration may not contain any collocation.  Thirdly, to fix the tie case such as ?????? ?????? (?terrorists? gene checking?),   ???? means ?human? when presented in  the collocation ??????, but ?particle?  in the collocation ??????.  The primary purpose of using collocation in WSD is to improve precision rate without any sacrifices in  recall rate. We also want to investigate whether  the use of collocation as an additional feature  can reduce the size of hand tagged sense corpus.   The rest of this paper is organized as follows.  Section 2 summarizes the existing Word Sense  Disambiguation techniques based on annotated  corpora. Section 3 describes the classifier and  the features in our proposed WSD approach.  Section 4 describes the experiments and the  analysis of our results. Section 5 is the conclusion.  2 Related Work  Automating word sense disambiguation tasks  based on annotated corpora have been proposed.  Examples of supervised learning methods for  WSD appear in [2-4], [7-8]. The learning algorithms applied including: decision tree, decisionlist [15], neural networks [7], na?ve Bayesian  learning ([5],[11]) and maximum entropy [10].  Among these leaning methods, the most important issue is what features will be used to construct the classifier. It is common in WSD to use  contextual information that can be found in the  neighborhood of the ambiguous word in training  data ([6], [16-18]). It is generally true that when  words are used in the same sense, they have  similar context and co-occurrence information  [13]. It is also generally true that the nearby context words of an ambiguous word give more effective patterns and features values than those  far from it [12]. The existing methods consider  features selection for context representation including both local and topic features where local  features refer to the information pertained only  to the given context and topical features are statistically obtained from a training corpus. Most  of the recent works for English corpus including  [7] and [8], which combine both local and topical information in order to improve their performance. An interesting study on feature  selection for Chinese [10] has considered topical  features as well as local collocational, syntactic,  and semantic features using the maximum entropy model. In Dang?s [10] work, collocational  features refer to the local PoS information and  bi-gram co-occurrences of words within 2 positions of the ambiguous word. A useful result  from this work based on (about one million  words) the tagged People?s Daily News shows  that adding more features from richer levels of  linguistic information such as PoS tagging  yielded no significant improvement (less than  1%) over using only the bi-gram co-occurrences  information. Another similar study for Chinese  [11] is based on the Naive Bayes classifier  model which has taken into consideration PoS  with position information and bi-gram templates  in the local context. The system has a reported  60.40% in both precision and recall based on the  SENSEVAL-3 Chinese training data. Even  though in both approaches, statistically significant bi-gram co-occurrence information is used,  they are not necessarily true collocations.  For  example, in the express???\x12?\x12??\x12??\x12? ?\x12????\x12?\x12?\x12??\x12???, the bi-grams in  their system are (??\x0f?\x0c,\x03\x0b?\x0f??\x0c, \x0b??\x0f? ?\x0c, \x0b??\x0f??\x0c, \x0b??\x0f???\x0c, \x0b???\x0f\x03?\x0c\x0f\x03 \x0b?\x0f?,\x0c? \x0b??\x0f??\x0c\x11Some bi-grams such as  \x0b??\x0f??\x0c\x03may have higher frequency but  may introduce noise when considering it as features in disambiguating the sense ?human|??  and ?symbol|??? like in the example case of  ?????????. In our system, we do not rely  on co-occurrence information. Instead, we utilize true collocation information (???, ??)  which fall in the window size of (-5, +5) as fea88 tures and the sense of ?human|?? can be decided clearly using this features. The collocation  information is a pre-prepared collocation list  obtained from a collocation extraction system  and verified with syntactic and semantic methods ([21], [24]).     Yarowsky [9] used the one sense per collocation  property as an essential ingredient for an unsupervised Word-Sense Disambiguation algorithm  to perform bootstrapping algorithm on a more  general high-recall disambiguation. A few recent research works have begun to pay attention  to collocation features on WSD. Domminic [19]  used three different methods called bilingual  method, collocation method and UMLS (Unified  Medical Language System) relation based  method to disambiguate unsupervised English  and German medical documents. As expected,  the collocation method achieved a good precision around 79% in English and 82% in German  but a very low recall which is 3% in English and  1% in German. The low recall is due to the nature of UMLS where many collocations would  almost never occur in natural text.  To avoid this  problem, we combine the contextual features in  the target context with the pre-prepared collocations list to build our classifier.   3 The Classifier With Topical Contextual and Local Collocation Features  3.1 The Feature Set  As stated early, an important issue is what features will be used to construct the classifier in  WSD. Early researches have proven that using  lexical statistical information, such as bi-gram  co-occurrences was sufficient to produce close  to the best results [10] for Chinese WSD. Instead of including bi-gram features as part of  discrimination features, in our system, we consider both topical contextual features as well as  local collocation features. These features are  extracted form the 60MB human sense-tagged  People?s Daily News with segmentation information.   3.1.1 Topical Contextual Features  Niu [11] proved in his experiments that Na?ve  Bayes classifier achieved best disambiguation  accuracy with small topical context window size  (< 10 words).  We follow their method and set  the contextual window size as 10 in our system.   Each of the Chinese words except the stop  words inside the window range will be considered as one topical feature. Their frequencies are  calculated over the entire corpus with respect to  each sense of an ambiguous word w.  The sense  definitions are obtained from HowNet.  3.1.2 Local Collocation Features  We chose collocations as the local features. A collocation is a recurrent and conventional fixed  expression of words which holds syntactic and  semantic relations [21]. Collocations can be  classified as fully fixed collocations, fixed collocations, strong collocations and loose collocations. Fixed collocations means the appearance  of one word implies the co-occurrence of another one such as ?????? (?burden of history?), while strong collocations allows very  limited substitution of the components, for example, ?????? (?local college?), or ? ??? ?? (?local university?). The sense of ambiguous  words can be uniquely determined in these two  types of collocations, therefore are the collocations applied in our system. The sources of the  collocations will be explained in Section 4.1.  In both Niu [11] and Dang?s [10] work, topical  features as well as the so called collocational  features were used. However, as discussed in  Section 2, they both used bi-gram cooccurrences as the additional local features.  However, bi-gram co-occurrences only indicate  statistical significance which may not actually  satisfy the conceptual definition of collocations.  Thus instead of using co-occurrences of bigrams, we take the true bi-gram collocations  extracted from our system and use this data to  compare with bi-gram co-occurrences to test the  usefulness of collocation for WSD. The local  features in our system make use of the collocations using the template (wi, w) within a window  size of ten (where i = ? 5). For example, ??? ??\x12\x03?\x12\x03??\x12\x03??\x12\x03??? (?Government  departments and local government commanded  that?) fits the bi-gram collocation template (w,  w1) with the value of (??\x0f\x03??). During the  training and the testing processes, the counting  of frequency value of the collocation feature will  be increased by 1 if a collocation containing the  ambiguous word occurs in a sentence. To have a  good analysis on collocation features, we have  also developed an algorithm using lonely  adjacent bi-gram as locals features(named Sys89 adjacent bi-gram as locals features(named System A)  and another using collocation as local  features(named System B).  3.2 The Collocation Classifier  We consider all the features in the features set F  = Ft ?Fl = {f1, f2,  ? , fm } as independent, where  Ft stands for the topical contextual features set,  and Fl stands for the local collocation features  set. For an ambiguous word w with n senses, let  Sw = {ws1, ws2,  ? , wsn } be the sense set. For  the contextual features, we directly apply the  Na?ve Bayes algorithm using Add-Lambda  Smoothing to handle unknown words:    )|(log)(log)(1 sij Ff sisi wfpwpwscore tj ? ? +=    (1)  For each sense siw of an ambiguous word w:   )( )()( wfreq wfreqwp sisi =                       (2)  For each contextual feature fj respects to each  sense siwof w :  ),( ),( )|( si Ff t sij sij wffreq wffreq wfp tt ? ? =   (3)  To integrate the local collocation feature fj ? Fl   with respect to each sense siw  of w, we use the  follows formula:  )()()( 21 sisisi wscorewscorewscore ?+= ?  (4)    where ? is tuned from experiments (Section 4.5),  score1( siw ) refers the score of the topical contextual features based on formula (1) and  score2( siw ) refers the score of collocation features with respect to the sense sjw  of w defined  below.  ? ? = lj Ff sjjsi wfwscore )|()(2 ?           (5)  where ?(fj| sjw ) = 1 for fj ? Fl if the collocation  occurs in the local context. Otherwise this term  is set as 0.  Finally, we choose the right skw so that  )(maxarg sks wscores k=        (6)  4 Experimental Results  We have designed a set of experiments to compare the classifier with and without the collocation features. In system A, the classifier is built  with local bi-gram features and topical contextual features. The classifier in system B is constructed from combining the local collocation  features with topical features.  4.1 Preparation the Data Set  We have selected 20 ambiguous words from  nouns and verbs with the sense number as 4 in  average. The sense definition is taken from  HowNet [22]. To show the effect of the algorithm, we try to choose words with high degree  of ambiguity, high frequency of use [23], and  high frequency of constructing collocations. The  selection of these 20 words is not completely  random although within each criterion class we  do try to pick word randomly.  Based on the 20 words, we extracted 28,000  sentences from the 60 MB People?s Daily News  with segmentation information as our training/test set which is then manually sense-tagged.   The collocation list is constructed from a  combination of a digital collocation dictionary, a  return result from a collocation automatic extraction system [21], and a hand collection from  the People?s Daily News. As we stated early, the  sense of ambiguous words in the fixed collocations and strong collocations can be decided  uniquely although they are not unique in loose  collocations. For example, the ambiguous word  ???? in the collocation ??????? may  have both the sense of ?appearance|??? or  ?reputation|???. Therefore, when labeling the  sense of collocations, we filter out the ones  which cannot uniquely determine the sense of  ambiguous words inside. However, this does not  mean that loose collocations have no contribution in WSD classification. We simply reduce its  weight when combining it with the contextual  features compared with the fixed and strong collocations. The sense and collocation distribution over the 20 words on the training examples can  be found in Table 1.  Table 1. Sense and Collocation Distribution of the 20 target words in the training corpus  Am.  W  T# S1  co#  S2  co#  S3  co#  S4  co#  S5  co#  S6  co#  90 ?? 31 1  1  30  10 NA      ??\x03 499 479  324  18   0 0 0  NA   ??\x03 944 908  129  1   1  17  10  18   0  0 NA  ??\x03 409 3  2  389   171  17  0 NA     ??\x03 110 3  0  101  36  6   9 NA     ??\x03 41 3  0  37   6  1   0 NA     ??\x03 4885 26  0  34   0  72   0  4492  1356  261  1 NA  ??\x03 3508 7  0  48   4  3194  1448  259  194  NA   ??\x03 348 312  117  22  11  14   4 NA     ??\x03 4438 3983 721  33   10  123   37  153  123  102  23  44  5  ??\x03 1987 1712 723  274  10 NA      ??\x03 83 36  14  47   4 00 NA     ??\x03 995 168  108  827  513 NA      ??\x03 31 11  3  20   11 NA      ??\x03 2725 227 1772  498  49  102  424  1898  201  NA   ??\x03 592 1  0  208  63  367  124 16 1  NA   ??\x03 1155 756  571  399  135 NA      ??\x03 2792 691  98  1765  113  336   29 0  NA   ??\x03 2460 82  63  36  11  1231  474  877  103  NA   ??\x03 125 11  0  64   0  15   3  32   4  3   0 NA  T#: total number of sentences contain the ambiguous word  s1- s6: sense no; co#: number of collocations in each sense  4.2 The Effect of Collocation Features  We recorded 6 trials with average precision over  six-fold validation for each word. Their average  precision for the six trials in the system A, and B  can be found inTable 2 and Table 3. From Table 3, regarding to precision, there are 16 words  have improved and 4 words remained the same  in the system B. The results from the both system confirmed that collocation features do improve the precision. Note that 4 words have the  same precision in the two systems, which fall  into two cases. In the first case, it can be seen  that these words already have very high precision in the system A (over 93%) which means  that one sense dominates all other senses. In this  case, the additional collation information is not  necessary. In fact, when we checked the intermediate outputs, the score of the candidate  senses of the ambiguous words contained in the  collocations get improved. Even though, it  would not change the result. Secondly, no collocation appeared in the sentences which are  tagged incorrectly in the system A. This is confirmed when we check the error files. For example, the word ???? with the sense as ????  (?closeness?) appeared in 4492 examples over  the total 4885 examples (91.9%). In the mean  time, 99% of collocation in its collocation list  has the same sense of ??? ? (?closeness?).  Only one collocation ????? has the sense of  ??? ? (?power?). Therefore, the collocation  features improved the score of sense ??? ?  which is already the highest one based on the  contextual features.   As can be seen from Table 3, the collocation  features work well for the sparse data. For example, the word ???? in the training corpus  has only one example with the sense ??? (?human?), the other 30 examples all have the sense  ???? (?management?). Under this situation,  the topical contextual features failed to identify  the right sense for the only appearance of the  sense ??? (?human?) in the training instance  ??????????????????. However, it can be correctly identified in the system  B because the appearance of the collocation ?? ??????.  To well show the effect of collocations on  the accuracy of classifier for the task of WSD,  we also tested both systems on SENSEVAL-3  data set, and the result is recorded in the Table 4.  From the difference in the relative improvement  of both data sets, we can see that collocation  features work well when the statistical model is  not sufficiently built up such as from a small  corpus like SENSEVAL-3. Actually, in this case,  the training examples appear in the corpus only  once or twice so that the parameters for such  sparse training examples may not be accurate to  forecast the test examples, which convinces us  that collocation features are effective on handling sparse training data even for unknown  words. Fig. 1 shows the precision comparison in  the system A, and B on SENVESAL-3.  Table 2.  Average Precision (5/6 training, 1/6 test) of  systemA on People?s Daily News  Amb.  W T1 T2 T3 T4 T5 T6  Ave.  Prec.  ??\x03 1.00 1.00 1.00 1.00 1.00 .83 .972  ??\x03 .90 .97 1.00 1.00 .97 .98 .972  ??\x03 .97 .96 .96 .92 .98 .96 .958  91 ??\x03 .94 .94 .97 .92 .97 .97 .951  ??\x03 1.00 1.00 .77 .94 .88 1.00 .932  ??\x03 .83 1.00 1.00 1.00 .83 .90 .927  ??\x03 .93 .95 .91 .92 .92 .92 .925  ??\x03 .93 .94 .89 .91 .89 .90 .91  ??\x03 .94 .93 .86 .93 .89 .87 .903  ??\x03 .83 .94 .89 .90 .88 .94 .897  ??\x03 .86 .88 .92 .84 .82 .87 .865  ??\x03 .92 .84 .92 .76 .84 .72 .833  ??\x03 .84 .83 .88 .82 .88 .71 .827  ??\x03 .80 .60 .80 .20 1.00 1.00 .733  ??\x03 .68 .72 .67 .77 .70 .68 .703  ??\x03 .51 .67 .47 .60 .68 .59 .586  ??\x03 .70 .63 .66 .64 .64 .64 .652  ?? .57 .74 .55 .64 .72 .67 .648  ??\x03 .65 .58 .66 .64 .54 .47 .58  ??\x03 .55 .50 .45 .45 .45 .64 .507  Total Average Precision 0.815  Table 3. Average Precision (5/6 training, 1/6 test) of  system B on People?s Daily News  Amb.  W T1 T2 T3 T4 T5 T6  Ave.  Prec.  ?? 1.00 1.00 1.00 1.00 1.00 1.00 1.00  ??\x03 .90 .97 1.00 1.00 .97 .98 .970  ??\x03 .96 .98 .97 .96 .98 .96 .968  ??\x03 .94 .94 .97 .94 .97 .98 .957  ??\x03 1.00 1.00 .77 .94 .88 1.00 .931  ??\x03 .83 1.00 1.00 1.00 .83 .90 .927  ??\x03 .93 .95 .91 .92 .92 .92 .925  ??\x03 .92 .95 .92 .92 .91 .91 .922  ??\x03 .94 .94 .86 .93 .91 .87 .908  ??\x03 .80 .95 .89 .93 .89 .94 .902  ??\x03 .87 .88 .92 .84 .83 .91 .875  ??\x03 .84 1.00 .92 .76 .84 .77 .855  ??\x03 .88 .86 .89 .84 .90 .74 .852  ??\x03 1.00 .80 .80 .20 1.00 1.00 .800  ??\x03 .69 .72 .68 .79 .75 .72 .725  ??\x03 .69 .76 .73 .74 .82 .79 .755  ??\x03 .58 .59 .70 .67 .64 .59 .628  ??\x03 .68 .67 .66 .63 .65 .63 .653  ??\x03 .65 .68 .71 .61 .70 .69 .673  ??\x03 .60 .55 .54 .54 .54 .64 .568  Total Average Precision 0.840  Table 4.  Average Precision of System A & B on  SENSEVAL-3 Data Set  Amb.  Word  Total  S  Ave. Prec. in  Sys A  Ave. Prec.  in Sys B  ??\x03 48 .207 .290  ??\x03 20 .742 .742  ?\x03 49 .165 .325  ?\x03 25 .325 .325  ??\x03 36 .260 .373  ??\x03 30 .167 .267  ??\x03 30 .192 .392  ??\x03 36 .635 .635  ?\x03 57 .238 .275  ??\x03 36 .327 .385  ??\x03 31.100 .322  ?\x03 40 .358 .442  ??\x03 40 .308 .308  ?\x03 76 .110 .123  ?\x03 28 .308 .475  ?667. 500. 30 ?  ?\x03 42 .165 .260  ?\x03 57 .037 .422  ??\x03 28 .833 .103  Total Ave.  Precision .276 .376  Fig. 1. The precision comparison in system A, and B based  on SENSEVAL-3    4.3 The Effect of Collocations on the Size  of Training Corpus Needed  Hwee [21] stated that a large-scale, human  sense-tagged corpus is critical for a supervised  learning approach to achieve broad coverage  and high accuracy WSD. He conducted a thorough study on the effect of training examples on  the accuracy of supervised corpus based WSD.  As the result showed, WSD accuracy continues  to climb as the number of training examples increases. Similarly, we have tested the system A,  and B with the different size of training corpus  based on the PDN corpus we prepared. Our experiment results shown in Fig 2 follow the same  fact.  The purpose we did the testing is that we  hope to disclose the effect of collocations on the  size of training corpus needed. From Fig 2, we  can see by using the collocation features, the  precision of the system B has increased slower  along with the growth of training examples than  the precision of the system A.  The result is reasonable because with collocation feature, the  statistical contextual information over the entire  corpus becomes side effect. Actually, as can be  seen from Fig. 2, after using collocation features  92 in the system B, even we use 1/6 corpus as training, the precision is still higher than we use 5/6  train corpus in the system A.  Fig. 2. The precision variation respect to the size of   training corpus in system A, and B based on PDN corpus    4.4 Investigation of Sense Distribution on  the Effect of Collocation Features  To investigate the sense distribution on the effect of collocation features, we selected the ambiguous words with the number of sense varied  from 2 to 6. In each level of the sense number,  the words are selected randomly. Table 5 shows  the effect of sense distribution on the effect of  collocation features. From the table, we can see  that the collocation features work well when the  sense distribution is even for a particular ambiguous word under which case the classifier  may get confused.  Table 5.  The Effect of Sense Distribution on the Effect of  collocation Features  Amb.  word  Prec.  Wihtout  coll  Prec.  With   coll  Sense  #  Sense  Distri.  ??\x03 .972 1 2 97% *  ??\x03 .97 .97 4 96% *  ??\x03 .957 .968 5 96% *  ??\x03 .951 .957 3 95% *  ??\x03.931 .931 3 92% *  ??\x03 .927 .927 3 90% *  ??\x03 .925 .925 5 92% *  ??\x03 .915 .922 4 91% *  ??\x03 .903 .908 3 90% *  ??\x03 .902 .902 6 90% *  ??\x03 .865 .875 2 86% o  ??\x03 .833 .855 3 ^  ??\x03 .823 .852 2 83% o  ??\x03 .733 .8 2 ^  ??\x03 .706 .725 4 ^  ?? .65 .653 4 ^  ??\x03 .618 .755 4 ^  ??\x03 .582 .628 2 ^  ??\x03 .563 .673 4 ^  ??\x03 .507 .568 5 ^       *: over 90% samples fall in one dominate sense       ^: Even distribution over all senses        o: 83% to 86% samples fall in one dominate sense  4.5 The Test of ?  We have conducted a set of experiments based  on both the PDN corpus and SENSEVLA-3 data  to set the best value of ? for the formula (4) described in Section 3.2. The best start value of ?  is tested based on the precision rate which is  shown in Fig. 3. It is shown from the experiment  that ? takes the start value of 0.5 for both corpuses.   Fig. 3. The best value of ? vs the precision rate    5 Conclusion and the Future Work  This paper reports a corpus-based Word Sense  Disambiguation approach for Chinese word using local collocation features and topical contextual features. Compared with the base-line  systems in which a Na?ve Bayes classifier is  constructed by combining the contextual features with the bi-gram features, the new system  achieves 3% precision improvement in average  in Peoples? Daily News corpus and 10% improvement in SENSEVAL-3 data set. Actually,  it works very well when disambiguating the  sense with sparse distribution over the entire  corpus under which the statistic calculation  prone to identify it incorrectly. In the same time,  because disambiguating using collocation fea93 tures does not need statistical calculation, it  makes contribution to reduce the size of human  tagged corpus needed which is critical and time  consuming in corpus based approach.   Because different types of collocations may  play different roles in classifying the sense of an  ambiguous word, we hope to extend this work  by integrating collocations with different weight  based on their types in the future, which may  need a pre-processing job to categorize the collocations automatically. 6 Acknowledgements  We would like to present our thanks to the IR  Laboratory in HIT University of China for sharing their sense number definition automatically  extracted from HowNet with us. '
1.0	W02-0208	A00-1007	b'1 Introduction There is a growing interest in annotating humanhuman dialogue. Annotated dialogues are useful for formulating and verifying theories of dialogue and for building statistical models. In addition to orthographic word transcription, one might want the following dialogue annotations. \x0f Annotation of the speech repairs. Speech repairs are a type of dis\ruency where speakers go back and change or repeat something they just said. \x0f Segmentation of the speech of each speaker into utterance units, with a single ordering of the utterances. We refer to this as linearizing the dialogue (Heeman and Allen, 1995a). \x0f Each utterance tagged with its speech act function. \x0f The utterances grouped into hierarchical discourse segments. There are tools that address subsets of the above tasks. However, we feel that doing dialogue annotation is very di?cult. Part of this di?culty is due to the interactions between the annotation tasks. An error at a lower level can have a large impact on the higher level annotations. For instance, there can be ambiguity as to whether an \\okay" is part of a speech repair; this will impact the segmentation of the speech into utterance units and the speech act coding. Sometimes, it is only by considering the higher level annotations that one can make sense of what is going on at the lower levels, especially when there is overlapping speech. Hence, a tool is needed that lets users examine and code the dialogue at all levels. The second reason why dialogue annotation is di?cult is because it is di?cult to follow what is occurring in the dialogue, especially for coding discourse structure. A dialogue annotation tool needs to help the user deal with this overload. In this paper, we describe our dialogue annotation tool, DialogueView. This tool displays the dialogue at three di\x0berent levels of abstraction. The word level shows the words time-aligned with the audio signal. The utterance level shows the dialogue as a sequence of utterance, as if it were a script for a play. It abstracts away from the exact timing of the words and even skips words that do not impact the progression of the dialogue. The block level shows the dialogue as a hierarchy of discourse segment purposes, and abstracts away from the exact utterances that were said. Annotations are done at the view that is most appropriate for what is being annotated. The tool allows the user to easily navigate between the three views and automatically updates the higher level views when changes are made in the lower level views. Because the higher levels abstract away lower level details, it is easier for the user to understand what is happening in the dialogue. 1 Yet, the user can easily refer to the lower level views to see what 1 This approach bears resemblance to the work of J\x7fonsson and Dahlb\x7fack (2000) in which they distill human-human dialogues by removing those parts that would not occur in human-computer dialogue. They do this to create training data for their spoken dialogue systems.        Philadelphia, July 2002, pp. 50-59. Association for Computational Linguistics.                   Proceedings of the Third SIGdial Workshop on Discourse and Dialogue, actually occurred when necessary. In the rest of this paper, we \x0crst discuss other annotation tools. We then describe the three levels of abstraction in our annotation tool. We then discuss audio playback. Next, we discuss how the tool can be customized for di\x0berent annotation schemes. Finally, we discuss the implementation of the tool. 2 Existing Tools There are a number of tools that let users annotate speech audio \x0cles. These include Emu (Cassidy and Harrington, 2001), SpeechView (Sutton et al, 1998) and Waves+ (Ent, 1993). These tools often allow multiple text annotation \x0cles to be associated with the waveform and allow users an easy facility to capture time alignments. For instance, for the ToBI annotation scheme (Pitrelli et al, 1994), one can have the word alignment in one text \x0cle, intonation features in a second, break indices in a third, and miscellaneous features in a fourth. The audio annotation tools often have powerful signal analysis packages for displaying such phenomena as spectrograms and voicing. These tools, however, do not have any built-in facility to group words into utterances nor group utterances into hierarchical discourse segments. Several tools allow users to annotate higher level structure in dialogue. The annotation tool DAT from the University of Rochester (Ferguson, 1998) allows users to annotate utterances or groups of utterances with a set of hard-coded annotation tags for the DAMSL annotation scheme (Allen and Core, 1997; Core and Allen, 1997). The tool Nb from MIT (Flammia, 1995; Flammia, 1998) allows users to impose a hierarchical grouping on a sequence of utterances, and hence annotate discourse structure. Both DAT and Nb take as input a linearization of the speaker utterances. Mistakes in this input cannot be \x0cxed. Whether there are errors or not, users cannot see the exact timing interactions between the speakers\' words or the length of pauses. This simpli\x0ccation can make it di?cult for the annotator to truly understand what is happening in the dialogue, especially for overlapping speech, where speakers \x0cght over the turn or make back-channels. The tools Transcriber (Barras et al, 2001) and Mate (McKelvie et al, 2001) allow multiple views of the data: a word view with words time-aligned to the audio signal and an utterance view. However, Transcriber is geared to single channel data and has a weak ability to handle overlapping speech and Mate only allows one view to be shown at a time. The author of a competing tool has remarked that \\speed and stability of [Mate] are both still problematic for real annotation. Also, the highly generic approach increases the initial e\x0bort to set up the tool since you basically have to write your own tool using the Mate style-sheet language" (Kipp, 2001). Hence, he developed a new tool Anvil that he claims is a better trade-o\x0b among generality, functionality, and complexity. This tool o\x0bers multi-modal annotation support, but like Transcriber, does notallow detailed annotation of dialogue phenomena, such as overlapping speech and abandoned speech, and has no abstraction mechanism. 3 Word View Our dialogue annotation tool, DialogueView, gives the user three views of the data. The lowest level view is called WordView and takes as input the words said by each speaker and their start and stop times and shows them time-aligned with the audio signal. Figure 1 shows an excerpt from the Trains corpus (Heeman and Allen, 1995b) in WordView. This view is ideal for viewing the exact timing of speech, especially overlapping speech. As will be discussed below, we use it for segmenting the speech into utterances, and annotating communicative status and speech repairs. 2 These annotations will allow us to build a simpler representation of what is happening in the speech for the utterance view, which is discussed in the next section. 3.1 Utterance Segmentation As can be seen in Figure 1, WordView shows the words segmented into utterances, which for our purpose is simply a grouping of consecutive words by a single speaker, in which there is minimal e\x0bect from the other speaker. Consider the exchange in Figure 1, where the upper speaker says \\and then take the remaining boxcar down to" followed by the lower speaker saying \\right, to du-", followed by the upper speaker saying \\Corning". Although one could consider the upper speaker\'s speech as one utterance, we preclude that due to the interaction from the lower speaker. A full definition of `utterance\' is beyond the scope of this paper, and is left to the users to specify in their annotation scheme. Utterance boundaries in WordView are only shown by their start times. The end of the utterance is either the start of the next utterance by the 2 There currently is no support for changing the word transcription. There are already a number of tools that do an excellent job at this task. Hence, adding this capability is a low priority for us. Figure 1: Utterance Segmentation in WordView same speaker or the end of the \x0cle. With WordView, the user can easily move, insert or delete utterance boundaries. The tool ensures that the boundaries never fall in the middle of a word by the speaker. The start times of the utterances are used to derive a single ordering of the utterances of the two speakers. This linearization of the dialogue captures how the speakers are sequentially adding to the dialogue. The linearization is used by the next view to give an abstraction away from the exact timing of the speech. 3.2 Communicative Status WordView allows features of the utterances to be annotated. In practice, however, we only annotate features related to its communicative status, based on the DAMSL annotation scheme (Allen and Core, 1997). Below, we give the utterance tags that we assign in WordView. Abandoned: The speaker abandoned what they were saying and it had no impact on the rest of the dialogue. This mainly happens where one speaker loses the \x0cght over who speaks next. Figure 2 gives an example where the upperspeaker tries to take the turn by saying \\takes," and than abandons this e\x0bort. Incomplete: The speaker did not make a complete utterance, either prosodically, syntactically, or semantically. Unlike the previous category, the partial utterance did impact the dialogue behavior. Figure 1 gives two examples. The \x0crst is where the upper speaker says \\and than take the remaining boxcar down to", and the second is where the lower speaker said \\to du-," which was than completed by the upper speaker. Overlap: The speech in the utterance overlaps with the speech from the prior utterance. For instance, in Figure 1, the lower speaker utters \\okay" during the middle of an utterance, perhaps to tell the upper speaker that they are understanding everything so far. However, a simple linearization would make it seem that the \\okay" is an acknowledgment of the entire utterance, which is not the case. Hence, we tag the \\okay" utterance with the overlap tag. The next view, UtteranceView, will take the overlap tag into account in displaying the utterances, as will be discussed in Section 4.3. Not all overlapping speech needs to be annotated with the overlap tag. In Figure 1, the second instance of \\Corning" overlaps slightly with the \x0crst instance of \\Corning". However, viewing it sequentially does not alter the analysis of the exchange. 3.3 Reordering Overlapping Speech Consider the example in Figure 2. Before the excerpt shown, the lower speaker had just \x0cnished an utterance and then paused for over a second. The upper speaker then acknowledged the utterance with \\okay", but this happened a fraction of a second after the bottom speaker started speaking again. A simple linearization of the dialogue would have the \\okay" following the wrong stretch of speech|\\and than th(at)- takes w- what three hours." A solution to this would be to annoFigure 2: Utterance Ordering and Speech Repairs in WordView tate the \\okay" with the overlap tag. However, this \\Okay" is more similar to the overlapping instances of \\Corning" in Figure 1. The fact that \\Okay" overlaps with the start of the next utterance is not critical to understanding what is occurring in the dialogue, as long as we linearize the \\Okay" to occur before the other utterance. WordView allows the user to change the linearization by moving the start times of utterances. This can be done provided that the speaker was silent in the time interval preceding where the other person started talking. In summary, overlapping speech can be handled in three ways. The utterance can be explicitly tagged as overlapping; the overlap can be ignored if it is not critical in understanding what is going on in the dialogue; or the start time of the utterance can be changed so that the overlap does not need to be tagged. 3.4 Speech Repairs WordView also allows users to annotate speech repairs. A speech repair is where a user goes back and repeats or changes something that was just said (Heeman and Allen, 1999). Below we give an example of a speech repair and show its principle components: reparandum, interruption point, andediting term. Example 1 why don\'t we take |{z} reparandum " ip um |{z} et take two boxcars The reparandum is the speech that is being replaced, the interruption point is the end of the reparandum, and the editing term consists of words such as \\um", \\uh", \\okay", \\let\'s see" that help signal the repair. To annotate a repair, the user highlights a sequence of words and then tags it as a reparandum or an editing term of a repair. The user can also specify the type of repair. Figure 2 shows how speech repairs are displayed in WordView. The words in the reparandum and editing term are underlined and displayed in a special color. 3.5 Speech Repairs and Utterances Some phenomena can be marked as either a speech repair, or could be marked using the utterance tags of incomplete or abandon. This is especially true for fresh starts (Heeman and Allen, 1999), where a speaker abandons the current utterance and starts over. To avoid having multiple ways of annotating the same phenomena, we impose the following restrictions in our annotation scheme. \x0f There cannot be an utterance boundary inside of a reparandum, inside of an editing term, at the interruption point, nor at the end of the editing term. Hence, something annotated as a reparandum cannot also be annotated as an abandoned utterance. \x0f Abandoned or incomplete utterances cannot be followed by an utterance by the same speaker. \x0f All word fragments must either be the last word of a reparandum or the last word of an utterance that is marked as abandoned or incomplete. Figure 3: UtteranceView: Segmented Utterances in UtteranceView \x0f Abandoned or incomplete utterances can end with an editing term, which would be marked as the editing term of an abridged repair. 3.6 Summary There are a number of reasons why we annotate utterance boundaries, speech repairs, and communicative status in WordView. Annotating utterance boundaries and overlapping speech requires the user to take into account the exact timing of the utterances, which is best done in this view. Speech repairs also require \x0cne tuned listening to the speech and have strong interactions with utterance boundaries. Furthermore, all three types of annotations can be used to build a simpler view of what is happening in the dialogue, as will be explained in the next section. 4 Utterance View The annotations from the word view are used to build the next view, which we refer to as UtteranceView. The dialogue excerpts from Figures 1 and 2 are shown in the utterance view in Figures 3 and 4, respectively. The utterance view abstracts away from the detailed timing information and individual words that were spoken. Instead, it focuses on the sequence of utterances between the speakers. By removing details that were annotated in the word view, we still preserve the important details that are needed to annotate speech act types for the utterances and to annotate discourse segments. Of course, if the user wants to see the exact timing of the words in the utterances, they can examine the word view,as it is Figure 4: UtteranceView: Reordered and Abandoned Utterances in UtteranceView displayed alongside the utterance view. There are also navigation buttons on each view that allow the user to easily reposition the portion of the dialogue in the other view. Furthermore, changes made in the word view are immediately propagated into the utterance view, and hence the user will immediately see the impact of their annotations. 4.1 Utterance Ordering Utterance ordering in the utterance view is determined by the start times of the utterances as speci\x0ced in WordView. As was explained earlier, altering the start time of an utterance can be used to simplify some cases of overlapping speech, where the overlap is not critical to understanding the role of the utterance in the dialogue. Figure 2 gave the word view of such an example. Rather than code it as an overlap, we moved the start time of the \\okay" utterance so that it precedes the overlapping speech by the other speaker. Figure 4 shows how this looks in the utterance view. Here, the annotator would view the \\okay" as an acknowledgment that occurred between the two utterances of the lower speaker. 4.2 Speech Repairs In the word view, the user annotates the reparandum and editing term of speech repairs. If the reparandum and editing term are removed, the resulting utterance re\rects what the speaker intended to say. Speech repairs do carry information. \x0f Their occurrence can signal a lack of certainty of the speaker. \x0f The reparandum of a repair can have an anaphoric reference, as in \\Peter was, well he was \x0cred." However, removing the reparandum and editing term of speech repairs from utterances in the utterance view leads to a simpler representation of what is occurring in the dialogue. Hence, in the utterance view, we clean up the speech repairs, as shown in Figures 2 and 4. Figure 2, which shows the word view, contains the utterance \\and then th(at) that takes w- what three hours"; whereas Figure 4, which shows the utterance view, contains \\and then that takes what three hours." Of course, a user can always refer to the word view when annotating in the utterance view if they want to see the exact speech that was said. In most cases, we feel that this will not be necessary for annotating speech act tags and discourse segments. 4.3 Communicative Status The communicative status coded in the word view is used in formatting the utterance view. Utterances tagged as overlapping are indented and displayed with `+\' on either side, as shown in Figure 3. Utterances tagged as abandoned are not shown, as can be seen in Figure 4, in which the abandoned utterance \\takes" made by the upper speaker is not included. Utterances tagged as incomplete are shown with a trailing \\..." as shown in Figure 3. 4.4 Annotating Utterance Tags In the utterance view, one can also annotate the utterances with various tags. For our work, we use a subset of the DAMSL tags corresponding to forward and backward functions (Allen and Core, 1997). Forward functionsinclude statement, request information, and suggestion. Backward functions include answer, acknowledgment, and agreement. Although these utterance tags could be annotated in the word view, doing it in the utterance view allows us to see more context, which is needed to give the utterance the proper tags. When necessary, the annotator can easily refer to the word view to see the exact local context. 4.5 Annotating Blocks of Utterances In the utterance view, the user can also annotate hierarchical groupings of utterances. 3 We use the utterance blocks to annotate discourse structure (Grosz and Sidner, 1986). This is similar to what Flammia\'s tool allows (Flammia, 1995). Rather than showing it with indentation and color, we draw boxes around segments. Figure 3 shows a dialogue excerpt with three utterance blocks inside of a larger block. To create a segment, the user highlights a sequence of utterances and then presses the \\make segment" button. The user can change the boundaries of the blocks by simply dragging either the top or bottom edge of the box. Blocks can also be deleted. The tool ensures that if two blocks have utterances in common then one block is entirely contained in the other. Tags can also be assigned to the blocks. We have just started using the tool for discourse segmentation, and so we are still re\x0cning these tags. In Grosz and Sidner\'s theory of discourse structure (1986), the speaker who initiates the block does so to accomplish a certain purpose. We have a tag for annotating this purpose. We also have tags to categorize the block as a greeting, specify goal, construct plan, summarize plan, verify plan, give info, request info, or other (Strayer and Heeman, 2001). The utterance view also allows the user to open or close a block. When a block is open (the default), all of its utterances and sub-blocks are dis3 We do not allow segments to be interleaved. It is unclear if such phenomena actually occur. played. When it is closed, its utterances and subblocks are replaced by the single line purpose. Opening and closing blocks is useful as it allows the user to control the level of detail that is shown. Consider the third embedded block shown in Figure 3, in which the conversants take seven utterances to jointly make a suggestion. After we have analyzed it, we can close it and just see the purpose. This will make it easier to determine the segmentation of the blocks that contain it. We are experimenting with a special type of dialogue block. Consider the example from the previous paragraph, in which the conversants take seven utterances to jointly make a suggestion. This is related to the shared turns of Schi\x0brin (1987), the co-operative completions of Linell (1998), and the grounding units of Traum and Nakatani (1999). We are experimenting with how to support the annotation of such phenomena. We have added a tag to indicate whether the utterances in the block are being used to build a single contribution. For these single contributions, we also supply a concise paraphrase of what wassaid. We have found that this paraphrase can be built from a sequential subset of the words in the utterances of the block. For instance, the paraphrase of our example block is \\and then take the remaining boxcar down to Corning." 5 Block View We are experimenting with a third view of the dialogue. This view, which we refer to as BlockView, abstracts away from the individual utterances, and shows the hierarchical structure of the discourse segments. This gives a very concise view of the dialogue. The block view is also convenient for it provides an index to the whole dialogue. This allows the user to quickly move around the dialogue. 6 Audio Playback Each view gives the user the ability to select a region of the dialogue and to play it. In the word view, the user can play each speaker channel individually or both combined. 4 This ability is especially useful for overlapping speech, where the annotator would want to listen to what each speaker said individually, as well as hear the timing between the speaker utterances. Just as each view provides a visual abstraction from the previous one, we also do the same with audio playback. In the word view, which has 4 In order to play each speaker individually, we require a separate audio channel for each speaker. wordViewUtt => atmostoneof abandoned incomplete overlap uttViewUtt => anyof forward backward comment uttViewUtt.forward => oneof statement question suggestion other uttViewUtt.backward => oneof agreement understanding answer other uttViewUtt.comment => other Figure 5: Sample Speci\x0ccation of Utterance Tags Figure 6: Sample Utterance Annotation Panel the speech repairs annotated, the user can play back the speech cleanly of either speaker, where the stretches of speech annotated as the reparandum or editing term of a repair are skipped. We have found this to be of great assistance in verifying if something should be annotated as a repair or not. It gives us an audio means to verify the speech repair annotations. If we have annotated the repair correctly, the edited audio signal should sound fairly natural. In formatting the utterance view, we take into account whether utterances have been marked as abandoned or overlapped. We provide a special playback in the utterance view that takes this into account. We build an audio \x0cle in which we skip over repairs, skip over abandoned speech, and shorten large silences. If there is overlap that is not marked as signi\x0ccant, we linearize it by concatenating the utterances together. If the overlap is marked as signi\x0ccant, we keep the overlap. We are \x0cnding that this gives us an audio means to ensure that our markings of abandonment, overlap and our linearization is correct. We are also experimenting with even further simplifying the audio output. For blocks that have a paraphrase, and the block is closed, we play the paraphrase by constructing it from the words said in the block. For blocks that are closed that do not have a paraphrase, we use the text-to-speech engine in the CSLU toolkit (Colton et al, 1996; Sutton et al, 1997) tosay the purpose, as if there was a narrator. 7 Customizations Some aspects of the tool are built in, such as the notion of utterances, speech repairs, and hierarchical grouping of utterances into blocks. However, the annotations of these phenomena and how they are displayed can be customized through a con\x0cguration \x0cle. This allows us to easily experiment as we revise our annotation scheme; to use domain speci\x0cc tags; and to make the tool useful for other researchers who might use di\x0berent tags. Speech repair tags, utterance tags, and block tags are speci\x0ced in the con\x0cguration \x0cle. Figure 5 gives a sample of how the annotation tags for an utterance are speci\x0ced. The two top level entries in the \x0cgure are \\wordViewUtt" and \\uttViewUtt", which specify the utterance annotation tags in WordView and UtteranceView, respectively. The decomposition can be of one of three types. atmostoneof: at most one of the attributes can be speci\x0ced oneof: exactly one of the attributes must be speci\x0ced anyof: there is no restriction on which attributes can be speci\x0ced The subcomponents can either be terminals as is the case for the decomposition of \\wordViewUtt", or can be non-terminals, as is the case for each of the three subcomponents of \\uttViewUtt". Hence, hierarchical tags are supported. Terminals are assumed to be of binary type, except for \\other", which is assumed to be a string. The con\x0cguration \x0cle determines how the annotation panel is generated. For the annotation scheme speci\x0ced in Figure 5, Figure 6 shows the annotation panel that would be automatically generated for the utterance view. As we explained earlier, some of the utterance tags a\x0bect how the word view and utterance view are formatted. Rather than hard code this functionality, it is speci\x0ced in the con\x0cguration \x0cle. We are still experimenting with the best way to code this functionality. Figure 7 gives an example of how we code the utterance tag functionality. The \x0crst line indicates that the utterance wordViewUtt.abandoned do wordView color red wordViewUtt.abandoned uttView ignore wordViewUtt.incomplete wordView color yellow wordViewUtt.incomplete uttView trailsoff wordViewUtt.overlap wordView color blue wordViewUtt.overlap uttView overlap Figure 7: Sample Utterance Display Speci\x0ccation tag of \\abandoned" coded in WordView should be displayed in red in WordView. The second line indicates that it should not be displayed in UtteranceView. 8 Implementation DialogueView is written in Tcl/Tk. We also use utilities from the CSLU Speech Toolkit (Colton et al, 1996; Sutton et al, 1997), including audio and wave handling and speech synthesis. We have rewritten the tool to use an object-oriented extension of Tcl called Tclpp, designed and developed by Stefan Simnige. This is allowing us to better manage the growing complexity of the tool as well as reuse pieces of the software in our annotation comparison tool (Yang et al, 2002). It should also help in expanding the tool so that it can handle any number of speakers.being built. The higher levels not only abstract away from the exact timing, but also can skip over words, whole utterances, and even simplify segments to a single line paraphrase. Along with the visual presentation, the audio can also be played at these di\x0berent levels of abstraction. We feel that these capabilities should help annotators better code dialogue. This tool is still under active development. In particular, we are currently re\x0cning how blocks are displayed, improving the ability to customize the tool for di\x0berent tagsets, and improving the audio playback facilities. As we develop this tool, we are also doing dialogue annotation, and re\x0cning our scheme for annotating dialogue in order to better capture the salient features of dialogue and improve the inter-coder reliability. 10 Acknowledgments The authors acknowledgment funding from the Intel Research Council.'	b"1 In t roduct ion   It has been known for quite some time now, that  the language used when interacting with a comput-  er is different from the one used in dialogues between  people, (c.f. JSnsson and Dahlb~ick (1988)). Given  that we know that the language will be different,  but not how it will be different, we need to base  our development of natural language dialogue sys-  tems on a relevant set of dialogue corpora. It is our  belief that we need to clarify a number of different  issues regarding the collection and use of corpora in  the development of speech-only and multimodal dia-  logue systems. Exchanging experiences and develop-  ing guidelines in this area are as important as, and in  some sense a necessary pre-requisite to, the develop-  ment of computational models of speech, language,  and dialogue/discourse. It is interesting to note the  difference in the state of art in the field of natu-  ral language dialogue systems with that of corpus  linguistics, where issues of the usefulness of different  samples, the necessary sampling size, representative-  ness in corpus design and other have been discussed  for quite some time (e.g. (Garside t al., 1997; Atkins  et al, 1992; Crowdy, 1993; Biber, 1993)). Also the  neighboring area of evaluation of NLP systems (for  an overview, see Sparck Jones and Galliers (1996))  seems to have advanced further.  Some work have been done in the area of natu-  ral language dialogue systems, e.g. on the design  of Wizard of Oz-studies (Dahlb~ck et al, 1998),  on measures for inter-rater eliability (Carletta,  1996), on frameworks for evaluating spoken dialogue  agents (Walker et al, 1998) and on the use of differ-  ent corpora in the development of a particular sys-  tem (The Carnegie-Mellon Communicator, Eskenazi  et al (1999)).  The question we are addressing in this paper is  how to collect and analyse relevant corpora. We be-  gin by describing what we consider to be the main  advantages and disadvantages of the two currently  used methods; studies of human dialogues and Wiz-  ard of Oz-dialogues, especially focusing on the eco-  logical validity of the methods. We then describe a  method called 'distilling dialogues', which can serve  as a supplement to the other two.  2 Natural and Wizard of  Oz-Dialogues  The advantage of using real dialogues between peo-  ple is that they will illustrate which tasks and needs  that people actually bring to a particular service  provider. Thus, on the level of the users' general  goals, such dialogues have a high validity. But there  are two drawbacks here. First; it is not self-evident  that users will have the same task expectations from  a computer system as they have with a person. Sec-  ond, the language used will differ from the language  usedwhen interacting with a computer.  These two disadvantages have been the major  force behind the development of Wizard of Oz-  methods. The advantage here is that the setting will  be human-computer interaction. But there are im-  portant disadvantages, too. First, on the practical  side, the task of setting up a high quality simulation  environment and training the operators ('wizards')  to use this is a resource consuming task (Dahlb~ck et  al., 1998). Second, and probably even more impor-  tant, is that we cannot hen observe real users using  a system for real life tasks, where they bring their  own needs, motivations, resources, and constraints  to bear. To some extent this problem can be over-  come using well-designed so called 'scenarios'. As  pointed out in Dahlb~ck (1991), on many levels of  analysis the artificiality of the situation will not af-  44  fect the language used. An example of this is the  pattern of pronoun-antecedent relations. But since  the tasks given to the users are often pre-described  by the researchers, this means that this is not a good  way of finding out which tasks the users actually  want to perform. Nor does it provide a clear enough  picture on how the users will act to find something  that satisfies their requirements. If e.g. the task is  one of finding a charter holiday trip or buying a TV-  set within a specified set of constraints (economical  and other), it is conceivable that people will stay  with the first item that matches the specification,  whereas in real life they would probably look for  alternatives. In our experience, this is primarily a  concern if the focus is on the users' goals and plans,  but is less a problem when the interest is on lower-  level aspects, such as, syntax or patterns of pronoun-  antecedent relationship (c.f. Dahlb~ick (1991)).  To summarize; real life dialogues will provide a  reasonably correct picture of the way users' ap-  proach their tasks, and what tasks they bring to  the service provider, but the language used will not  give a good approximation of what the system un-  der construction will need to handle. Wizard of Oz-  dialogues, on the other hand, will give a reasonable  approximation of some aspects of the language used,  but in an artificial context.  The usual approach has been to work in three  steps. First analyse real human dialogues, and based  on these, in the second phase, design one or more  Wizard of Oz-studies. The final step is to fine-tune  the system's performance on real users. A good ex-  ample of this method is presented in Eskenazi et al  (1999). But there are also possible problems with  this approach (though we are not claiming that this  was the case in their particular project). Eskenazi et  al.(1999) asked a human operator to act 'computer-  like' in their Wizard of Oz-phase. The advantage  is of course that the human operator will be able  to perform all the tasks that is usually provided by  this service. The disadvantage is that it puts a heavy  burden on the human operator to act as a comput-  er. Since we know that lay-persons' ideas of what  computers can and cannot do are in many respects  far removed from what is actually the case, we risk  introducing some systematic distortion here. And  since it is difficult to perform consistently in similar  situations, we also risk introducing non-systematic  distortion here, even in those cases when the 'wiz-  ard' is an NLP-professional.  Our suggestion is therefore to supplement he  above mentioned methods, and bridge the gap be-  tween them, by post-processing human dialogues to  give them a computer-like quality. The advantage,  compared to having people do the simulation on the  fly, is both that it can be done with more consis-  tency, and also that it can be done by researchers  that actually know what human-computer natural  language dialogues can look like. A possible dis-  advantage with using both Wizard of Oz-and real  computer dialogues, is that users will quickly adapt  to what the system can provide them with, and will  therefore not try to use it for tasks they know it  cannot perform. Consequently, we will not get a full  picture of the different services they would like the  system to provide.  A disadvantage with this method is, of course,  that post-processing takes some time compared to  using the natural dialogues as they are. There is al-  so a concern on the ecological validity of the results,  as discussed later.  3 Distilling dialogues  Distilling dialogues, i.e. re-writing human interac-  tions in order to have them reflect what a human-  computer interaction could look like involves a num-  ber of considerations. The main issue is that in cor-  pora of natural dialogues one of the interlocutors i not a dialogue system. The system's task is instead  performed by a human and the problem is how to  anticipate the behaviour of a system that does not  exist based on the performance of an agent with dif-  ferent performance characteristics. One important  aspect is how to deal with human features that are  not part of what the system is supposed to be able   to handle, for instance if the user talks about things  outside of the domain, such as discussing an episode  of a recent TV show. It also involves issues on how  to handle situations where one of the interlocuters  discusses with someone lse on a different opic, e.g.  discussing the up-coming Friday party with a friend  in the middle of an informationproviding dialogue  with a customer.  It is important for the distilling process to have at  least an outline of the dialogue system that is under  development: Will it for instance have the capacity  to recognise users' goals, even if not explicitly stat-  ed? Will it be able to reason about the discourse  domain? What services will it provide, and what  will be outside its capacity to handle?  In our case, we assume that the planned dialogue  system has the ability to reason on various aspects  of dialogue and properties of the application. In our  current work, and in the examples used for illustra-  tion in this paper, we assume a dialogue model that  can handle any relevant dialogue phenomenon and  also an interpreter and speech recogniser being able  to understand any user input that is relevant o the  task. There is is also a powerful domain reason-  ing module allowing for more or less any knowledge  reasoning on issues that can be accomplished with-  in the domain (Flycht-Eriksson, 1999). Our current  system does, however, not have an explicit user task  model, as opposed to a system task model (Dahlb~ick  45  and JSnsson, 1999), which is included, and thus, we  can not assume that the 'system' remembers utter-  ances where the user explains its task. Furthermore,  as our aim is system development we will not con-  sider interaction outside the systems capabilities as  relevant o include in the distilled dialogues.  The context of our work is the development a  multi-modal dialogue system. However, in our cur-  rent work with distilling dialogues, the abilities of  a multi-modal system were not fully accounted for.  The reason for this is that the dialogues would be  significantly affected, e.g. a telephone conversation  where the user always likes to have the next con-  nection, please will result in a table if multi-modal  output is possible and hence a fair amount of the di-  alogne is removed. We have therefore in this paper  analysed the corpus assuming a speech-only system,  since this is closer to the original telephone conversa-  tions, and hence needs fewer assumptions on system  performance when distilling the dialogues.  4 Dis t i l l a t ion  gu ide l ines   Distilling dialogues requires guidelines for how to  handle various types of utterances. In this section  we will present our guidelines for distilling a corpus  of telephone conversations between a human infor-  mation provider on local buses 1to be used for devel-  oping a multimodal dialogue system (Qvarfordt and  JSnsson, 1998; Flycht-Eriksson and JSnsson, 1998;  Dahlb~ick et al, 1999; Qvarfordt, 1998). Similar  guidelines are used within another project on devel-  oping Swedish Dialogue Systems where the domain  is travel bureau information.  We can distinguish three types ofcontributors:  'System' (i.e. a future systems) utterances, User ut-  terances, and other types, such as moves by other  speakers, and noise.  4.1 Modifying system utterances  The problem of modifying 'system' utterances can  be divided into two parts: how to change and when  to change. They are in some respects intertwined,  but as the how-part affects the when-part more we  will take this as a starting point.  ? The 'system' provides as much relevant infor-  mation as possible at once. This depends on  the capabilities of the systems output modal-  ities. If we have a screen or similar output  device we present as much as possible which  normally is all relevant information. If we, on  the other hand, only have spoken output the  amount of information that the hearer can inter-  pret in one utterance must be considered when  1The bus time table dialogues are collected at  LinkSping University and are available (in Swedish) on  http://www.ida.l iu.se/~arnjo/kfb/dialoger.html  distilling. The system might in such cases pro-  vide less information. The principle of provid-  ing all relevant information is based on the as-  sumption that a computer system often has ac-  cess to all relevant information when querying  the background system and can also present it  more conveniently, especially in a multimodal  system (Ahrenberg et al, 1996). A typical ex-  ample is the dialogue fragment in figure 1. In  this fragment he system provides information  on what train to take and how to change to a  bus. The result of distilling this fragment pro-  vides the revised fragment of figure 2. As seen in  the fragment of figure 2 we also remove a num-  ber of utterances typical for human interaction,  as discussed below.  * System utterances are made more computer-l ike  and do not include irrelevant information. The  latter is seen in $9 in the dialogue in figure 3  where the provided information is not relevant.  It could also be possible to remove $5 and re-  spond with $7 at once. This, however, depends  on if the information grounded in $5-U6 is need-  ed for the 'system' in order to know the arrival  time or if that could be concluded from U4.  This in turn depends on the system's capabili-  ties. If we assume that the dialogue system has  a model of user tasks, the information in $5-U6  could have been concluded from that. We will,  in this case, retain $5-U6 as we do not assume a user task model (Dahlb/ick and JSnsson, 1999)  and in order to stay as close to the original di-  alogue as possible.  The next problem concerns the case when 'system'  utterances are changed or removed.  ? Dialogue contributions provided by something or  someone other than the user or the 'system' are removed. These are regarded as not being part  of the interaction. This means that if some-  one interrupts the current interaction, say that  the telephone rings during a face-to-face inter-  action, the interrupting interaction is normally  removed from the corpus.  Furthermore, 'system' interruptions are re-  moved. A human can very well interrupt anoth-  er human interlocuter, but a computer system  will not do that.  However, this guideline could lead to problems,  for instance, when users follow up such interrup-  tions. If no information is provided or the in-  terrupted sequence does not affect the dialogue,  we have no problems removing the interruption.  The problem is what to do when information  from the 'system' is used in the continuing dia-  logue. For such cases we have no fixed strategy,  46  U4:  $5:  U6:  $7:  U8:  $9:  U10:  $11:  U12:  S13:  U14:  $15:  yes I wonder if you have any mm buses or (.) like express buses leaving from LinkSping  to Vadstena (.) on sunday  ja ville undra om ni hade ndgra 5h bussar eUer (.) typ expressbussar sore dkte frdn LinkSping  till Vadstena (.) pd sSnda  no the bus does not run on sundays  nej bussen g~r inte pd sSndagar  how can you (.) can you take the train and then change some way (.) because (.)  to MjSlby 'n' so  hur kan man (.) kan man ta tdg d sen byta p~ ndtt sStt (.) fSr de (.)  till mjSlby ~ sd  that you can do too yes  de kan du gSra ocksd ja  how (.) do you have any such suggestions  hut (.) har du n~ra n~gra s~na fSrslag  yes let's see (4s) a moment (15s) now let us see here (.) was it on the sunday you should travel  ja ska se h~ir (4s) eft 5gonblick (15s) nu ska vise hSr (.) va de p~ sSndagen du skulle dka pd  yes right afternoon preferably  ja just de eftermidda ggirna  afternoon preferable (.) you have train from LinkSping fourteen twenty nine  eftermidda gSrna (.) du hat t~g frdn LinkSping fjorton d tjugonie  mm  mm  and then you will change from MjSlby station six hundred sixty  sd byter du frdn MjSlby station sexhundrasexti  sixhundred sixty  sexhundrasexti  fifteen and ten  femton ~ tie  Figure 1: Dialogue fragment from a real interaction on bus time-table information  U4: I wonder if you have any buses or (.) like express buses going from LinkSping  to Vadstena (.) on sunday  S5: no the bus does not run on sundays  U6: how can you (.) can you take the train and then change some way (.) because (.)  to MjSlby and so  $7: you can take the train from LinkSping fourteen andtwenty nine and then you will  change at MjSlby station to bus six hundred sixty at fifteen and ten  Figure 2: A distilled version of the dialogue in figure 1  the dialogue needs to be rearranged epending  on how the information is to be used (c.f. the  discussion in the final section of this paper).  ? 'System' utterances which are no longer valid  are removed. Typical examples of this are the  utterances $7, $9, $11 and $13 in the dialogue  fragment of figure 1.  * Remove sequences of utterances where the 'sys-  tem' behaves in a way a computer would not do.  For instance jokes, irony, humor, commenting  on the other dialogue participant, or dropping  the telephone (or whatever is going on in $7  in figure 4). A common case of this is when  the 'system' is talking while looking for infor-  mation, $5 in the dialogue fragment of figure 4  is an example of this. Related to this is when  the system provides its own comments. If we  can assume that it has such capabilities they  are included, otherwise we remove them.  The system does not repeat information that has  already been provided unless explicitly asked to  do so. In human interaction it is not uncommon  to repeat what has been uttered for purposes  other than to provide grounding information or  feedback. This is for instance common during  47   U4: 'n' I must be at Resecentrum before fourteen and thirty five (.) 'cause we will going to the  interstate buses  ja ska va p~ rececentrum innan \\]jorton ~ trettifem (.) f5 vi ska till  l~ngf~irdsbussarna  $5: aha (.) 'n' then you must be there around twenty past two something then  jaha (.) ~ dd behhver du va here strax e~ter tjuge 5vet tvd n~nting d~  U6: yes around that  ja ungefgir  $7: let's see here ( l ls)  two hundred and fourteen Ryd end station leaves forty six (.) thirteen 'n'  forty six then you will be down fourteen oh seven (.)  d~ ska vise hSr (11s) tv~hundrafjorton Ryd 5ndh~llplatsen gdr ~5rtisex (.) tretton d  \\]Srtisex d~ dr du nere ~jorton noll sju 5)  U8: aha  jaha  $9: 'n' (.) the next one takes you there (.) fourteen thirty seven (.) but that is too late  (.) ndsta dr du nere 5) ~jorton d trettisju (.) men de 5 ju ~Sr sent  Figure 3: Dialogue fragment from a real interaction on bus time-table information  U2: Well, hi (.) I am going to Ugglegatan eighth  ja hej (.) ja ska till Ugglegatan dtta  $3: Yes  ja  U4: and (.) I wonder (.) it is somewhere in Tannefors  och (.) jag undrar (.) det ligger ndnstans i Tannefors  $5: Yes (.) I will see here one one I willlook exactly where it is one moment please  ja (.) jag ska se hhr eft eft jag ska titta exakt vat det ligger eft 6gonblick barn  U6: Oh Yeah  jar~  $7: (operator disconnects) (25s) mm (.) okey (hs) what the hell (2s)  (operator connects again) hello yes  ((Telefonisten kopplar ur sig)) (25s) iihh (.) okey (hs) de va sore \\]aan (2s)  ((Telefonisten kopplar in sig igen)) halld ja  U8: Yes hello  ja hej  $9: It is bus two hundred ten which runs on old tannefors road that you have to take and get off at  the bus stop at that bus stop named vetegatan  det ~i buss tv~hundratio sore g~r gamla tanne~orsvSgen som du ~r  ~ka ~ g~ av rid  den hdllplatsen rid den hdllplatsen sore heter vetegatan.  Figure 4: Dialogue fragment from a natural bus timetable interaction  search procedures as discussed above.  ? The system does not ask for information it has  already achieved. For instance asking again if it  is on Sunday as in $9 in figure 1. This is not un-  common in human interaction and such utter-  ances from the user are not removed. However,  we can assume that the dialogue system does  not forget what has been talked about before.  4.2 Mod i fy ing  user  u t te rances   The general rule is to change user utterances as lit-  tle as possible. The reason for this is that we do not  want to develop systems where the user needs to  restrict his/her behaviour to the capabilities of the  dialogue system. However, there are certain changes  made to user utterances, in most cases as a conse-  quence of changes of system utterances.  Utterances that are no longer valid are removed.  The most common cases are utterances whose  request has already been answered, as seen in  the distilled dialogue in figure 2 of the dialogue  in figure 1.  48  Sl1: sixteen fifty five  sexton \\]emti/em  U12: sixteen fifty five (.) aha  sexton femti/em (.) jaha  S13: bus line four hundred thirty five  linje \\]yrahundra tretti/em  Figure 5: Dialogue fragment from a natural bus  timetable interaction  ? Utterances are removed where the user discuss-  es things that are in the environment. For  instance commenting the 'systems' clothes or  hair. This also includes other types of commu-  nicative signals such as laughter based on things  outside the interaction, for instance, in the en-  vironment of the interlocuters.  ? User utterances can also be added in order to  make the dialogue continue. In the dialogue in  figure 5 there is nothing in the dialogue xplain-  ing why the system utters S13. In such cases  we need to add a user utterance, e.g. Which  bus is that?. However, it might turnout that  there are cues, such as intonation, found when  listening to the tapes. If  such detailed analyses  are carried out, we will, of course, not need to  add utterances. Furthermore, it is sometimes  the case that the telephone operator deliberate-  ly splits the information into chunks that can  be comprehended by the user, which then must  be considered in the distillation.  5 App ly ing  the  method  To illustrate the method we will in this section try to  characterise the results from our distillations. The  illustration is based on 39 distilled dialogues from  the previously mentioned corpus collected with a  telephone operator having information on local bus  time-tables and persons calling the information ser-  vice.  The distillation took about three hours for all 39  dialogues, i.e. it is reasonably fast. The distilled  dialogues are on the average 27% shorter. However,  this varies between the dialogues, at most 73% was  removed but there were also seven dialogues that  were not changed at all.  At the most 34 utterances where removed from  one single dialogue and that was from a dialogue  with discussions on where to find a parking lot, i.e.  discussions outside the capabilities of the applica-  tion. There was one more dialogue where more than  30 utterances were removed and that dialogue is a  typical example of dialogues where distillation actu-  ally is very useful and also indicates what is normal-  ly removed from the dialogues. This particular dia-  logue begins with the user asking for the telephone  number to 'the Lost property office' for a specific bus  operator. However, the operator starts a discussion  on what bus the traveller traveled on before provid-  ing the requested telephone number. The reason for  this discussion is probably that the operator knows  that different bus companies are utilised and would  like to make sure that the user really understands  his/her request. The interaction that follows can,  thus, in that respect be relevant, but for our pur-  pose of developing systems based on an overall goal  of providing information, not to understand human  interaction, our dialogue system will not able to han-  dle such phenomenon (JSnsson, 1996).  The dialogues can roughly be divided into five dif-  ferent categories based on the users task. The dis-  cussion in twenty five dialogues were on bus times  between various places, often one departure and one  arrival but five dialogues involved more places. In  five dialogues the discussion was one price and var-  ious types of discounts. Five users wanted to know  the telephone number to 'the Lost property office',  two discussed only bus stops and two discussed how  they could utilise their season ticket to travel out-  side the trafficking area of the bus company. It is interesting to note that there is no correspondence  between the task being performed uring the inter-  action and the amount of changes made to the dia-   logue. Thus, if we can assume that the amount of  distillation indicates omething about a user's inter-  action style, other factors than the task are impor-  tant when characterising user behaviour.  Looking at what is altered we find that the most  important distilling principle is that the 'system'  provides all relevant information at once, c.f. fig-  ures 1 and 2. This in turn removes utterances pro-  vided by both 'system' and user.  Most added utterances, both from the user and  the 'system', provide explicit requests for informa-  tion that is later provided in the dialogue, e.g. ut-  terance $3 in figure 6. We have added ten utterances  in all 39 dialogues, five 'system' utterances and five  user utterances. Note, however, that we utilised the  transcribed ialogues, without information on into-  nation. We would probably not have needed to add  this many utterances if we had utilised the tapes.  Our reason for not using information on intonation  is that we do not assume that our system's peech  recogniser can recognise intonation.  Finally, as discussed above, we did not utilise the  full potential of multi-modality when distilling the  dialogues. For instance, some dialogues could be  further distilled if we had assumed that the system  had presented a time-table. One reason for this is  that we wanted to capture as many interesting as-  pects intact as possible. The advantage is, thus, that  we have a better corpus for understanding human-  49  U2: Yees hi Anna Nilsson is my name and I would like to take the bus from Ryd center to Resecentrum  in LinkSping  jaa hej Anna Nilsson heter jag och jag rill ~ka buss ~r~n Ryds centrum till resecentrum  i LinkSping.  $3: mm When do you  want  to  leave?  mm N~ir r i l l  du  ?ka?  U4: 'n' I must be at Resecentrum before fourteen and thirty five (.) 'cause we will going to the  interstate buses  ja ska va p~ rececentrum innan fjorton d trettifem (.) f5 vi ska till  l~ngfiirdsbussarna  Figure 6: Distilled dialogue fragment with added utterance  computer interaction and can from that corpus do  a second distillation where we focus more on multi-  modal interaction.  6 Discuss ion  We have been presenting a method for distilling hu-  man dialogues to make them resemble human com-  puter interaction, in order to utilise such dialogues  as a knowledge source when developing dialogue sys-  tems. Our own main purpose has been to use them  for developing multimodal systems, however, as dis-  cussed above, we have in this paper rather assumed  a speech-only system. But webelieve that the basic  approach can be used also for multi-modal systems  and other kinds of natural language dialogue sys-  tems.  It is important o be aware of the limitations of  the method, and how 'realistic' the produced result  will be, compared to a dialogue with the final sys-  tem. Since we are changing the dialogue moves, by  for instance providing all required information in one  move, or never asking to be reminded of what the us-  er has previously requested, it is obvious that what  follows after the changed sequence would probably  be affected one way or another. A consequence of  this is that the resulting dialogue is less accurate as  a model of the entire dialogue. It is therefore not an  ideal candidate for trying out the systems over-all  performance during system development. But for  the smaller sub-segments or sub-dialogues, we be-  lieve that it creates a good approximation of what  will take place once the system is up and running.  Furthermore, we believe distilled dialogues in some  respects to be more realistic than Wizard of Oz-  dialogues collected with a wizard acting as a com-  puter.  Another issue, that has been discussed previously  in the description of the method, is that the distilling  is made based on a particular view of what a dialogue  with a computer will look like. While not necessari-  ly being a detailed and specific model, it is at least  an instance of a class of computer dialogue models.  One example of this is whether the system is meant  to acquire information on the user's underlying mo-  tivations or goals or not. In the examples presented,  we have not assumed such capabilities, but this as-  sumption is not an absolute necessity. We believe,  however, that the distilling process should be based  on one such model, not the least to ensure a con-  sistent treatment of similar recurring phenomena t  different places in the corpora.  The validity of the results based on analysing dis-  tilled dialogues depends part ly on how the distilla-  tion has been carried out. Even when using natural  dialogues we can have situations where the interac-  tion is somewhat mysterious, for instance, if some of  the dialogue participants behaves irrational such as  not providing feedback or being too elliptical. How-  ever, if careful considerations have been made to stay  as close to the original dialogues as possible, we be-  lieve that distilled dialogues will reflect what a hu-  man would consider to be a natural interaction.  Acknowledgments   This work results from a number of projects on de-  velopment of natural language interfaces upported  by The Swedish Transport & Communications Re-  search Board (KFB) and the joint Research Program  for Language Technology (HSFR/NUTEK) .  We are indebted to the participants of the Swedish Dialogue  Systems project, especially to Staffan Larsson, Lena  Santamarta, and Annika Flycht-Eriksson for inter-  esting discussions on this topic. "
0.0	W02-0208	W90-0119	b'1 Introduction There is a growing interest in annotating humanhuman dialogue. Annotated dialogues are useful for formulating and verifying theories of dialogue and for building statistical models. In addition to orthographic word transcription, one might want the following dialogue annotations. \x0f Annotation of the speech repairs. Speech repairs are a type of dis\ruency where speakers go back and change or repeat something they just said. \x0f Segmentation of the speech of each speaker into utterance units, with a single ordering of the utterances. We refer to this as linearizing the dialogue (Heeman and Allen, 1995a). \x0f Each utterance tagged with its speech act function. \x0f The utterances grouped into hierarchical discourse segments. There are tools that address subsets of the above tasks. However, we feel that doing dialogue annotation is very di?cult. Part of this di?culty is due to the interactions between the annotation tasks. An error at a lower level can have a large impact on the higher level annotations. For instance, there can be ambiguity as to whether an \\okay" is part of a speech repair; this will impact the segmentation of the speech into utterance units and the speech act coding. Sometimes, it is only by considering the higher level annotations that one can make sense of what is going on at the lower levels, especially when there is overlapping speech. Hence, a tool is needed that lets users examine and code the dialogue at all levels. The second reason why dialogue annotation is di?cult is because it is di?cult to follow what is occurring in the dialogue, especially for coding discourse structure. A dialogue annotation tool needs to help the user deal with this overload. In this paper, we describe our dialogue annotation tool, DialogueView. This tool displays the dialogue at three di\x0berent levels of abstraction. The word level shows the words time-aligned with the audio signal. The utterance level shows the dialogue as a sequence of utterance, as if it were a script for a play. It abstracts away from the exact timing of the words and even skips words that do not impact the progression of the dialogue. The block level shows the dialogue as a hierarchy of discourse segment purposes, and abstracts away from the exact utterances that were said. Annotations are done at the view that is most appropriate for what is being annotated. The tool allows the user to easily navigate between the three views and automatically updates the higher level views when changes are made in the lower level views. Because the higher levels abstract away lower level details, it is easier for the user to understand what is happening in the dialogue. 1 Yet, the user can easily refer to the lower level views to see what 1 This approach bears resemblance to the work of J\x7fonsson and Dahlb\x7fack (2000) in which they distill human-human dialogues by removing those parts that would not occur in human-computer dialogue. They do this to create training data for their spoken dialogue systems.        Philadelphia, July 2002, pp. 50-59. Association for Computational Linguistics.                   Proceedings of the Third SIGdial Workshop on Discourse and Dialogue, actually occurred when necessary. In the rest of this paper, we \x0crst discuss other annotation tools. We then describe the three levels of abstraction in our annotation tool. We then discuss audio playback. Next, we discuss how the tool can be customized for di\x0berent annotation schemes. Finally, we discuss the implementation of the tool. 2 Existing Tools There are a number of tools that let users annotate speech audio \x0cles. These include Emu (Cassidy and Harrington, 2001), SpeechView (Sutton et al, 1998) and Waves+ (Ent, 1993). These tools often allow multiple text annotation \x0cles to be associated with the waveform and allow users an easy facility to capture time alignments. For instance, for the ToBI annotation scheme (Pitrelli et al, 1994), one can have the word alignment in one text \x0cle, intonation features in a second, break indices in a third, and miscellaneous features in a fourth. The audio annotation tools often have powerful signal analysis packages for displaying such phenomena as spectrograms and voicing. These tools, however, do not have any built-in facility to group words into utterances nor group utterances into hierarchical discourse segments. Several tools allow users to annotate higher level structure in dialogue. The annotation tool DAT from the University of Rochester (Ferguson, 1998) allows users to annotate utterances or groups of utterances with a set of hard-coded annotation tags for the DAMSL annotation scheme (Allen and Core, 1997; Core and Allen, 1997). The tool Nb from MIT (Flammia, 1995; Flammia, 1998) allows users to impose a hierarchical grouping on a sequence of utterances, and hence annotate discourse structure. Both DAT and Nb take as input a linearization of the speaker utterances. Mistakes in this input cannot be \x0cxed. Whether there are errors or not, users cannot see the exact timing interactions between the speakers\' words or the length of pauses. This simpli\x0ccation can make it di?cult for the annotator to truly understand what is happening in the dialogue, especially for overlapping speech, where speakers \x0cght over the turn or make back-channels. The tools Transcriber (Barras et al, 2001) and Mate (McKelvie et al, 2001) allow multiple views of the data: a word view with words time-aligned to the audio signal and an utterance view. However, Transcriber is geared to single channel data and has a weak ability to handle overlapping speech and Mate only allows one view to be shown at a time. The author of a competing tool has remarked that \\speed and stability of [Mate] are both still problematic for real annotation. Also, the highly generic approach increases the initial e\x0bort to set up the tool since you basically have to write your own tool using the Mate style-sheet language" (Kipp, 2001). Hence, he developed a new tool Anvil that he claims is a better trade-o\x0b among generality, functionality, and complexity. This tool o\x0bers multi-modal annotation support, but like Transcriber, does notallow detailed annotation of dialogue phenomena, such as overlapping speech and abandoned speech, and has no abstraction mechanism. 3 Word View Our dialogue annotation tool, DialogueView, gives the user three views of the data. The lowest level view is called WordView and takes as input the words said by each speaker and their start and stop times and shows them time-aligned with the audio signal. Figure 1 shows an excerpt from the Trains corpus (Heeman and Allen, 1995b) in WordView. This view is ideal for viewing the exact timing of speech, especially overlapping speech. As will be discussed below, we use it for segmenting the speech into utterances, and annotating communicative status and speech repairs. 2 These annotations will allow us to build a simpler representation of what is happening in the speech for the utterance view, which is discussed in the next section. 3.1 Utterance Segmentation As can be seen in Figure 1, WordView shows the words segmented into utterances, which for our purpose is simply a grouping of consecutive words by a single speaker, in which there is minimal e\x0bect from the other speaker. Consider the exchange in Figure 1, where the upper speaker says \\and then take the remaining boxcar down to" followed by the lower speaker saying \\right, to du-", followed by the upper speaker saying \\Corning". Although one could consider the upper speaker\'s speech as one utterance, we preclude that due to the interaction from the lower speaker. A full definition of `utterance\' is beyond the scope of this paper, and is left to the users to specify in their annotation scheme. Utterance boundaries in WordView are only shown by their start times. The end of the utterance is either the start of the next utterance by the 2 There currently is no support for changing the word transcription. There are already a number of tools that do an excellent job at this task. Hence, adding this capability is a low priority for us. Figure 1: Utterance Segmentation in WordView same speaker or the end of the \x0cle. With WordView, the user can easily move, insert or delete utterance boundaries. The tool ensures that the boundaries never fall in the middle of a word by the speaker. The start times of the utterances are used to derive a single ordering of the utterances of the two speakers. This linearization of the dialogue captures how the speakers are sequentially adding to the dialogue. The linearization is used by the next view to give an abstraction away from the exact timing of the speech. 3.2 Communicative Status WordView allows features of the utterances to be annotated. In practice, however, we only annotate features related to its communicative status, based on the DAMSL annotation scheme (Allen and Core, 1997). Below, we give the utterance tags that we assign in WordView. Abandoned: The speaker abandoned what they were saying and it had no impact on the rest of the dialogue. This mainly happens where one speaker loses the \x0cght over who speaks next. Figure 2 gives an example where the upperspeaker tries to take the turn by saying \\takes," and than abandons this e\x0bort. Incomplete: The speaker did not make a complete utterance, either prosodically, syntactically, or semantically. Unlike the previous category, the partial utterance did impact the dialogue behavior. Figure 1 gives two examples. The \x0crst is where the upper speaker says \\and than take the remaining boxcar down to", and the second is where the lower speaker said \\to du-," which was than completed by the upper speaker. Overlap: The speech in the utterance overlaps with the speech from the prior utterance. For instance, in Figure 1, the lower speaker utters \\okay" during the middle of an utterance, perhaps to tell the upper speaker that they are understanding everything so far. However, a simple linearization would make it seem that the \\okay" is an acknowledgment of the entire utterance, which is not the case. Hence, we tag the \\okay" utterance with the overlap tag. The next view, UtteranceView, will take the overlap tag into account in displaying the utterances, as will be discussed in Section 4.3. Not all overlapping speech needs to be annotated with the overlap tag. In Figure 1, the second instance of \\Corning" overlaps slightly with the \x0crst instance of \\Corning". However, viewing it sequentially does not alter the analysis of the exchange. 3.3 Reordering Overlapping Speech Consider the example in Figure 2. Before the excerpt shown, the lower speaker had just \x0cnished an utterance and then paused for over a second. The upper speaker then acknowledged the utterance with \\okay", but this happened a fraction of a second after the bottom speaker started speaking again. A simple linearization of the dialogue would have the \\okay" following the wrong stretch of speech|\\and than th(at)- takes w- what three hours." A solution to this would be to annoFigure 2: Utterance Ordering and Speech Repairs in WordView tate the \\okay" with the overlap tag. However, this \\Okay" is more similar to the overlapping instances of \\Corning" in Figure 1. The fact that \\Okay" overlaps with the start of the next utterance is not critical to understanding what is occurring in the dialogue, as long as we linearize the \\Okay" to occur before the other utterance. WordView allows the user to change the linearization by moving the start times of utterances. This can be done provided that the speaker was silent in the time interval preceding where the other person started talking. In summary, overlapping speech can be handled in three ways. The utterance can be explicitly tagged as overlapping; the overlap can be ignored if it is not critical in understanding what is going on in the dialogue; or the start time of the utterance can be changed so that the overlap does not need to be tagged. 3.4 Speech Repairs WordView also allows users to annotate speech repairs. A speech repair is where a user goes back and repeats or changes something that was just said (Heeman and Allen, 1999). Below we give an example of a speech repair and show its principle components: reparandum, interruption point, andediting term. Example 1 why don\'t we take |{z} reparandum " ip um |{z} et take two boxcars The reparandum is the speech that is being replaced, the interruption point is the end of the reparandum, and the editing term consists of words such as \\um", \\uh", \\okay", \\let\'s see" that help signal the repair. To annotate a repair, the user highlights a sequence of words and then tags it as a reparandum or an editing term of a repair. The user can also specify the type of repair. Figure 2 shows how speech repairs are displayed in WordView. The words in the reparandum and editing term are underlined and displayed in a special color. 3.5 Speech Repairs and Utterances Some phenomena can be marked as either a speech repair, or could be marked using the utterance tags of incomplete or abandon. This is especially true for fresh starts (Heeman and Allen, 1999), where a speaker abandons the current utterance and starts over. To avoid having multiple ways of annotating the same phenomena, we impose the following restrictions in our annotation scheme. \x0f There cannot be an utterance boundary inside of a reparandum, inside of an editing term, at the interruption point, nor at the end of the editing term. Hence, something annotated as a reparandum cannot also be annotated as an abandoned utterance. \x0f Abandoned or incomplete utterances cannot be followed by an utterance by the same speaker. \x0f All word fragments must either be the last word of a reparandum or the last word of an utterance that is marked as abandoned or incomplete. Figure 3: UtteranceView: Segmented Utterances in UtteranceView \x0f Abandoned or incomplete utterances can end with an editing term, which would be marked as the editing term of an abridged repair. 3.6 Summary There are a number of reasons why we annotate utterance boundaries, speech repairs, and communicative status in WordView. Annotating utterance boundaries and overlapping speech requires the user to take into account the exact timing of the utterances, which is best done in this view. Speech repairs also require \x0cne tuned listening to the speech and have strong interactions with utterance boundaries. Furthermore, all three types of annotations can be used to build a simpler view of what is happening in the dialogue, as will be explained in the next section. 4 Utterance View The annotations from the word view are used to build the next view, which we refer to as UtteranceView. The dialogue excerpts from Figures 1 and 2 are shown in the utterance view in Figures 3 and 4, respectively. The utterance view abstracts away from the detailed timing information and individual words that were spoken. Instead, it focuses on the sequence of utterances between the speakers. By removing details that were annotated in the word view, we still preserve the important details that are needed to annotate speech act types for the utterances and to annotate discourse segments. Of course, if the user wants to see the exact timing of the words in the utterances, they can examine the word view,as it is Figure 4: UtteranceView: Reordered and Abandoned Utterances in UtteranceView displayed alongside the utterance view. There are also navigation buttons on each view that allow the user to easily reposition the portion of the dialogue in the other view. Furthermore, changes made in the word view are immediately propagated into the utterance view, and hence the user will immediately see the impact of their annotations. 4.1 Utterance Ordering Utterance ordering in the utterance view is determined by the start times of the utterances as speci\x0ced in WordView. As was explained earlier, altering the start time of an utterance can be used to simplify some cases of overlapping speech, where the overlap is not critical to understanding the role of the utterance in the dialogue. Figure 2 gave the word view of such an example. Rather than code it as an overlap, we moved the start time of the \\okay" utterance so that it precedes the overlapping speech by the other speaker. Figure 4 shows how this looks in the utterance view. Here, the annotator would view the \\okay" as an acknowledgment that occurred between the two utterances of the lower speaker. 4.2 Speech Repairs In the word view, the user annotates the reparandum and editing term of speech repairs. If the reparandum and editing term are removed, the resulting utterance re\rects what the speaker intended to say. Speech repairs do carry information. \x0f Their occurrence can signal a lack of certainty of the speaker. \x0f The reparandum of a repair can have an anaphoric reference, as in \\Peter was, well he was \x0cred." However, removing the reparandum and editing term of speech repairs from utterances in the utterance view leads to a simpler representation of what is occurring in the dialogue. Hence, in the utterance view, we clean up the speech repairs, as shown in Figures 2 and 4. Figure 2, which shows the word view, contains the utterance \\and then th(at) that takes w- what three hours"; whereas Figure 4, which shows the utterance view, contains \\and then that takes what three hours." Of course, a user can always refer to the word view when annotating in the utterance view if they want to see the exact speech that was said. In most cases, we feel that this will not be necessary for annotating speech act tags and discourse segments. 4.3 Communicative Status The communicative status coded in the word view is used in formatting the utterance view. Utterances tagged as overlapping are indented and displayed with `+\' on either side, as shown in Figure 3. Utterances tagged as abandoned are not shown, as can be seen in Figure 4, in which the abandoned utterance \\takes" made by the upper speaker is not included. Utterances tagged as incomplete are shown with a trailing \\..." as shown in Figure 3. 4.4 Annotating Utterance Tags In the utterance view, one can also annotate the utterances with various tags. For our work, we use a subset of the DAMSL tags corresponding to forward and backward functions (Allen and Core, 1997). Forward functionsinclude statement, request information, and suggestion. Backward functions include answer, acknowledgment, and agreement. Although these utterance tags could be annotated in the word view, doing it in the utterance view allows us to see more context, which is needed to give the utterance the proper tags. When necessary, the annotator can easily refer to the word view to see the exact local context. 4.5 Annotating Blocks of Utterances In the utterance view, the user can also annotate hierarchical groupings of utterances. 3 We use the utterance blocks to annotate discourse structure (Grosz and Sidner, 1986). This is similar to what Flammia\'s tool allows (Flammia, 1995). Rather than showing it with indentation and color, we draw boxes around segments. Figure 3 shows a dialogue excerpt with three utterance blocks inside of a larger block. To create a segment, the user highlights a sequence of utterances and then presses the \\make segment" button. The user can change the boundaries of the blocks by simply dragging either the top or bottom edge of the box. Blocks can also be deleted. The tool ensures that if two blocks have utterances in common then one block is entirely contained in the other. Tags can also be assigned to the blocks. We have just started using the tool for discourse segmentation, and so we are still re\x0cning these tags. In Grosz and Sidner\'s theory of discourse structure (1986), the speaker who initiates the block does so to accomplish a certain purpose. We have a tag for annotating this purpose. We also have tags to categorize the block as a greeting, specify goal, construct plan, summarize plan, verify plan, give info, request info, or other (Strayer and Heeman, 2001). The utterance view also allows the user to open or close a block. When a block is open (the default), all of its utterances and sub-blocks are dis3 We do not allow segments to be interleaved. It is unclear if such phenomena actually occur. played. When it is closed, its utterances and subblocks are replaced by the single line purpose. Opening and closing blocks is useful as it allows the user to control the level of detail that is shown. Consider the third embedded block shown in Figure 3, in which the conversants take seven utterances to jointly make a suggestion. After we have analyzed it, we can close it and just see the purpose. This will make it easier to determine the segmentation of the blocks that contain it. We are experimenting with a special type of dialogue block. Consider the example from the previous paragraph, in which the conversants take seven utterances to jointly make a suggestion. This is related to the shared turns of Schi\x0brin (1987), the co-operative completions of Linell (1998), and the grounding units of Traum and Nakatani (1999). We are experimenting with how to support the annotation of such phenomena. We have added a tag to indicate whether the utterances in the block are being used to build a single contribution. For these single contributions, we also supply a concise paraphrase of what wassaid. We have found that this paraphrase can be built from a sequential subset of the words in the utterances of the block. For instance, the paraphrase of our example block is \\and then take the remaining boxcar down to Corning." 5 Block View We are experimenting with a third view of the dialogue. This view, which we refer to as BlockView, abstracts away from the individual utterances, and shows the hierarchical structure of the discourse segments. This gives a very concise view of the dialogue. The block view is also convenient for it provides an index to the whole dialogue. This allows the user to quickly move around the dialogue. 6 Audio Playback Each view gives the user the ability to select a region of the dialogue and to play it. In the word view, the user can play each speaker channel individually or both combined. 4 This ability is especially useful for overlapping speech, where the annotator would want to listen to what each speaker said individually, as well as hear the timing between the speaker utterances. Just as each view provides a visual abstraction from the previous one, we also do the same with audio playback. In the word view, which has 4 In order to play each speaker individually, we require a separate audio channel for each speaker. wordViewUtt => atmostoneof abandoned incomplete overlap uttViewUtt => anyof forward backward comment uttViewUtt.forward => oneof statement question suggestion other uttViewUtt.backward => oneof agreement understanding answer other uttViewUtt.comment => other Figure 5: Sample Speci\x0ccation of Utterance Tags Figure 6: Sample Utterance Annotation Panel the speech repairs annotated, the user can play back the speech cleanly of either speaker, where the stretches of speech annotated as the reparandum or editing term of a repair are skipped. We have found this to be of great assistance in verifying if something should be annotated as a repair or not. It gives us an audio means to verify the speech repair annotations. If we have annotated the repair correctly, the edited audio signal should sound fairly natural. In formatting the utterance view, we take into account whether utterances have been marked as abandoned or overlapped. We provide a special playback in the utterance view that takes this into account. We build an audio \x0cle in which we skip over repairs, skip over abandoned speech, and shorten large silences. If there is overlap that is not marked as signi\x0ccant, we linearize it by concatenating the utterances together. If the overlap is marked as signi\x0ccant, we keep the overlap. We are \x0cnding that this gives us an audio means to ensure that our markings of abandonment, overlap and our linearization is correct. We are also experimenting with even further simplifying the audio output. For blocks that have a paraphrase, and the block is closed, we play the paraphrase by constructing it from the words said in the block. For blocks that are closed that do not have a paraphrase, we use the text-to-speech engine in the CSLU toolkit (Colton et al, 1996; Sutton et al, 1997) tosay the purpose, as if there was a narrator. 7 Customizations Some aspects of the tool are built in, such as the notion of utterances, speech repairs, and hierarchical grouping of utterances into blocks. However, the annotations of these phenomena and how they are displayed can be customized through a con\x0cguration \x0cle. This allows us to easily experiment as we revise our annotation scheme; to use domain speci\x0cc tags; and to make the tool useful for other researchers who might use di\x0berent tags. Speech repair tags, utterance tags, and block tags are speci\x0ced in the con\x0cguration \x0cle. Figure 5 gives a sample of how the annotation tags for an utterance are speci\x0ced. The two top level entries in the \x0cgure are \\wordViewUtt" and \\uttViewUtt", which specify the utterance annotation tags in WordView and UtteranceView, respectively. The decomposition can be of one of three types. atmostoneof: at most one of the attributes can be speci\x0ced oneof: exactly one of the attributes must be speci\x0ced anyof: there is no restriction on which attributes can be speci\x0ced The subcomponents can either be terminals as is the case for the decomposition of \\wordViewUtt", or can be non-terminals, as is the case for each of the three subcomponents of \\uttViewUtt". Hence, hierarchical tags are supported. Terminals are assumed to be of binary type, except for \\other", which is assumed to be a string. The con\x0cguration \x0cle determines how the annotation panel is generated. For the annotation scheme speci\x0ced in Figure 5, Figure 6 shows the annotation panel that would be automatically generated for the utterance view. As we explained earlier, some of the utterance tags a\x0bect how the word view and utterance view are formatted. Rather than hard code this functionality, it is speci\x0ced in the con\x0cguration \x0cle. We are still experimenting with the best way to code this functionality. Figure 7 gives an example of how we code the utterance tag functionality. The \x0crst line indicates that the utterance wordViewUtt.abandoned do wordView color red wordViewUtt.abandoned uttView ignore wordViewUtt.incomplete wordView color yellow wordViewUtt.incomplete uttView trailsoff wordViewUtt.overlap wordView color blue wordViewUtt.overlap uttView overlap Figure 7: Sample Utterance Display Speci\x0ccation tag of \\abandoned" coded in WordView should be displayed in red in WordView. The second line indicates that it should not be displayed in UtteranceView. 8 Implementation DialogueView is written in Tcl/Tk. We also use utilities from the CSLU Speech Toolkit (Colton et al, 1996; Sutton et al, 1997), including audio and wave handling and speech synthesis. We have rewritten the tool to use an object-oriented extension of Tcl called Tclpp, designed and developed by Stefan Simnige. This is allowing us to better manage the growing complexity of the tool as well as reuse pieces of the software in our annotation comparison tool (Yang et al, 2002). It should also help in expanding the tool so that it can handle any number of speakers.being built. The higher levels not only abstract away from the exact timing, but also can skip over words, whole utterances, and even simplify segments to a single line paraphrase. Along with the visual presentation, the audio can also be played at these di\x0berent levels of abstraction. We feel that these capabilities should help annotators better code dialogue. This tool is still under active development. In particular, we are currently re\x0cning how blocks are displayed, improving the ability to customize the tool for di\x0berent tagsets, and improving the audio playback facilities. As we develop this tool, we are also doing dialogue annotation, and re\x0cning our scheme for annotating dialogue in order to better capture the salient features of dialogue and improve the inter-coder reliability. 10 Acknowledgments The authors acknowledgment funding from the Intel Research Council.'	b'Introduction  Somewhat obviously, plan recognition is the process  of inferring an agent\'s plan from observation of the  agent\'s actions. The agent\'s actions can be physical  actions or speech actions. Four principal methods  for plan recognition have been proposed in the  literature. The methods are plausible inference  (Allen \\[1\\], Carberry \\[2\\], Litman \\[15\\], Sidner \\[25\\]),  parsing (Huff and Lesser \\[9\\]), circumscribing a  hierarchical representation f plans and using deduc-  tion (Kautz \\[12, 13\\]), and abduction (Charniak and  McDermott \\[6\\], Konolige and Pollack \\[14\\], Poole  \\[24\\]).  Our particular interest is in the use of plan recog-  nition in question-answering systems, where recog-  nizing the plan underlying a user\'s queries aids in the  generation of an appropriate response. Here, the  plans and goals of the user, once recognized, have  been used to: supply more information than is expli-  citly requested (Allen \\[1\\], Luria \\[16\\]), handle prag-  matically ill-formed queries and resolve some inter-  sentential ellipses (Carberry \\[2, 3, 4\\]), provide an  explanation from the appropriate perspective  (McKeown et hi. \\[17\\]), respond to queries that result  from an invalid plan (Pollack \\[20, 21, 22\\]), and avoid  misleading responses and produce user-specific  cooperative responses (Joshi et a\\]. \\[10, 11\\], van Beck  and Cohen \\[26, 27\\], Cohen et al \\[7\\]).  Example  1 (Joshi et al \\[11\\]). As an example of  a cooperative response consider the following  exchange between student and student-advisor sys-  tem. The plan of the student is to avoid failing the  course by dropping it.  User: Can I drop numerical analysis?  System: Yes, however you will still fail the course  since your mark will be recorded as with-  drawal while failing.  If the system just gives the direct answer, "Yes" the  student will remain unaware that the plan is faulty.  The more cooperative answer warns the student.  An important weakness of this work in response  generation, however, is the reliance on a plan recog-  nition component being able to uniquely determine  the plan of the user. This is clearly too strong an  assumption as the user\'s actions often will be con-  sistent with more than one plan, especially after only  one or a few utterances when there is insufficient  context to help decide the plan of the user. In  Example 1 there are many reasons why a student  may want to drop a course, such as resolving a  scheduling conflict, avoiding failing the course, or  finding the material uninteresting. There may be no  reason to prefer one alternative over the other, yet  we may still want to generate a response that does  more than just give a direct answer to the user\'s  query.  In this paper, we address the problem of what the  system should do when the user\'s actions are ambi-  guous as they are consistentwith more than one  plan. To the extent that this problem has been con-  sidered by researchers in plan recognition, it is gen-  erally assumed that the plan recognition system  overcomes the ambiguity problem by inferring the  most likely interpretation, given its assessment of the  context and dialogue so far and knowledge of typical  plans of action. Thus there is a dependence on sali-  ence heuristics to solve the ambiguity problem \\[e.g.  1, 2, 17, and see the final section\\]. Existing propo-  sals for resolving ambiguity beyond heuristics are  underspecified and what usually underlies these pro-  posals is the assumption that we always want to  determine one unique plan \\[2, 15, 19, 25\\].  We show how to relax the assumption that the  plan recognition component returns a single plan.  That is, given that the result of the plan recognition  phase will usually be a disjunction of possible plans,  we show how to design a response component o  generate cooperative responses given the disjunction.  We show that it is not always necessary to resolve  ambiguity, and provide a procedure that allows the  144  response component o estimate whether the ambi-  guity matters to the task of formulating a response.  If the ambiguity does not matter, the response com-  ponent can continue to answer the user\'s queries and  ignore the ambiguity in the underlying goals and  plan until further queries help clarify which plan the  user is pursuing. If the ambiguity does matter, the  system should take advantage of the interactive and  collaborative nature of dialogue in recognizing plans  and resolving ambiguity. A key contribution of this  work therefore is providing a clear criterion for  when to respond to a question with a question that  will differentiate between some of the possibilities.  We also propose a specific solution to what questions  should then be asked of the user. Moreover, these  questions are asked only to resolve the ambiguity to  the point where it no longer matters (this is not  necessarily to a unique plan).  Example  2. Here are two different examples to  give a flavor of what we are proposing. There are  two agents: a cook and an expert who is cooperative,  helpful, and adept at recognizing plans. The expert  observes the cook making marinara sauce and recog-  nizes the cook could be pursuing three possible  plans, all with the same top level goal of preparing a meal: make fettucini marinara or spaghetti marinara  (both a pasta dish) or chicken marinara (a meat  dish).  a. Suppose the cook then asks the expert: "Is a red  wine a good choice?" The expert has the criteria for  wine selection that red wine should be served if the  meal is chicken, fettucini marinara, or spaghetti  marinara and white if fettucini alfredo. Thereis  enough information for the expert to decide that red  wine should be bought and the ambiguity does not  need to be resolved to cooperatively answer the  question.  b. Now suppose the expert also knows that the  guest of honor is allergic to gluten and so would not  be able to eat if a pasta dish was served. Here the  ambiguity is important as the expert has recognized  that the cook\'s prepare-a-meal goal may conflict  with the cook\'s entertain-important-guest goal. The  expert will want to resolve the ambiguity enough to  he assured that the proposed meal does not include a  pasta dish and so clarifies this with the cook.  Est imat ing Whether  the  Ambigu i ty  Matters   Example 2, above, showed that sometimes it is  necessary to resolve ambiguity and sometimes it is  not. Here we give criteria for judging which is the  case. The result is a procedure that allows the  response component o estimate whether the ambi-  guity matters to the task of formulating a response.  Assuming we can answer the user\'s query, deciding  when we want to give more than just a direct  answer to the user\'s query depends on the plan of  the user. Previous work has identified several kinds  of faults in a plan that a cooperative response should  warn a user about. We generalize this work in iden-  tifying faults and call it plan critiquing. Our propo-  sal therefore is to first apply a plan critiquing pro-  cedure to determine possible faults.  Plan Critiquing  A plan may be labeled as faulty if it will fail to  achieve its high level goal (e.g. Allen \\[1\\]) or if there  are simply better alternatives for reaching that goal  (e.g. Pollack \\[20\\]). Joshi et al \\[10, 11\\] formalize the  above and identify additional cases (such as warning  the user that a certain plan is the only way to  achieve a goal).  In \\[26, 27\\], we make some extensions to Joshi et al  \\[10, 11\\] and give a procedure to determine faults in  plans and to address these faults through coopera-  tive responses. Faults include both plans which fail  and plans which can be replaced by better alterna-  tives. In \\[26, 27\\], we also include the critical exten-  sion of domain goals. To  explain, the system needs  to not only respond to goals inferred from the  current discourse but also to the domain goals a user  is likely to have or known to have even though they  are not stated in, or derivable from, the discourse.  Example 3.  User: I\'m not interested in the material and so  would like to drop the course. Can I?  The ideal response should say more than just "Yes",  but also warn the student that the domaingoal of  achieving a degree conflicts with the immediate goal  of dropping the course. This example shows how  competing oals may exist and need to be addressed  in a response.  To determine whether the ambiguity matters, we  propose to apply an extension of the procedure of  \\[26, 27\\], which will be sensitive to complementary  goals as well. The standard assumption in coopera-  tive response work is that the user is pursuing only a  single chain of goals; we allow actions to be used to  achieve more than one complementary goal. For  example, in the course-advising domain we can  assume that all users have the goal to avoid failing a  course. If a user then asks "I\'m not interested in the  course and so would like to drop it. Can I?" the  response should address not just the goal of avoiding  uninteresting courses but also the complementary  goal of avoiding failing courses. For example, the  response should warn the user if he will fail the  course because of withdrawal while failing (as in  Example I).  The algorithm to estimate whether the ambiguity  matters is below. The deciding criterion is whether  the critiques for all plans are the same. By same cri-  tique or same fault (Case 1. b. in the algorithm) we  mean, for example, same better way or same conflict  with competing domain goal.  145  Input: A set of possible plans (the output from a  plan recognition algorithm).  Output: The set of possible plans with critiques  attached to each plan.  Algorithm Critique:  begin  for each plan in the set of possible plans do  critique the plan and, if there is a fault, anno-  tate the plan with the fault  Input: A set of critiqued possible plans.  Output: "Yes", the ambiguity matters, or "No"   the ambiguity does not matter.   Algorithm Ambiguity_matters:  We are in one of the following two cases:  Case 1. "No", the ambiguity does not matter.  The critiques are the same for all the plans. That   is, either  a. every plan is faultless, or  b. every plan is annotated with the same fault.  Case 2. "Yes", the ambiguity does matter.  The critiques are different for some or all of the  plans. That  is, either  a. some, but not all, of the plans are faultless (the  faults may or may not be the same), or  b. every plan is annotated with a fault and the  faults are not all the same.  end  An  Example  to  I l lustrate the P roposa l   Suppose the user asks "Can I drop numerical  analysis?". First, a plan recognition algorithm is  called to determine the possible plans of the user.  They are found to be:  end ~resolve schedule conflict --~ drop course  end ~ avoid uninteresting prof --~ drop course  Second, algorithm Critique is called to critique the  plans. As a result, both plans are labeled with the  same fault that there is a better plan for achieving  the goal. Third, Mgorithm Ambiguity_matter8 is  called to determine whether the ambiguity regarding  the plan of the user matters to the task of formulat-  ing a response. It is found that the ambiguity does  not matter  as both plans are annotated with the  same fault (Case 1.b of the algorithm). Finally, the  critiqued plans are passed to a response generation  procedure. The answer then given in this example  is, "Yes, but a better way is to switch to another  section."  In general, in (Case 1.a) a response generation pro-  cedure can just give a direct answer to the user\'s  query, and in (Case 1.b) can give a direct answer  plus any warranted additional information, such as  telling the user about the fault.  In the above example it was found that the ambi-  guity did not matter as there was enough informa-  tion to generate a cooperative response. If instead it  was found that the ambiguity did matter (Case 2 of  the algorithm) we propose that we we enter into a  clarification dialogue with the user to resolve the  ambiguity to the point where it no longer does  matter. That  is, until we are in Case 1. A response  generation procedure would then be called.  Clar i f i ca t ion  D ia logues :   Quest ions  in Response  to  Quest ions   What should we ask the user when a clarification is  necessary? Clearly, we do not want to simply list  the set of possible plans and ask which is being pur-  sued. Below is an algorithm that determines what to  say. Our algorithm for estimating whether the  ambiguity matters is not dependent on the method  of plan recognition used. Now our proposal for cla-  rification dialogues is tied to a hierarchical plan  library in the style of Kautz \\[12\\]. The input to the  algorithm is a set of possible plans where the user\'s  action is related to top-level or end goals through  chMns of goals. Each plan in the set is annotated  with a a critique. The key idea is to ask about the  highest level possible, check whether the ambiguity  still needs to be further resolved, and if so, ask at  the next level down, iteratively, through the hierar-  chy of goals.  Input: A set of critiqued possible plans (the out-  put from Mgorithm Critique).  Output: The pruned set of possible plans such that  the ambiguity no longer matters.  Algorithm Clarify:  begin  initiMize the current levelto be the first branch  point from the top in the set of possible plans  while Ambiguity_matter8 = "Yes" do  separate out the distinct goals in the set of  remaining possible plans that are one level  below the current level and are annotated with  a fault  list the goals (perhaps with their accompanying  annotations as justification for why we are ask-  ing) and ask the user whether one of them is  part  of the plan being pursued  according to the answer, remove the plans that  are not being pursued from the set of possible  plans and update the current level in the hierar-  chy that is being looked at to be the next  branch point  end whi le   end  146  next query  . . . . .  ~ . . . . . . . .  - I  f" . . . .  1  I I I \'l\' H res?lve H4" generate \' i 1. generate 2. resolve i  i candidate ambiguity w. \' I I ambiguity response I  "i---\'i with user i plans heuristics i  I I  I Plan Recognition Component i i Response Generation Component i  L J L. J  Fig. 1. Major modules of our proposed query-answering system  Example  2b (Revisited). In our story in the  introduction about cooking for our allergic guest the  expert has recognized the following three plans:  1. end ~ prepare meal ~ make meat dish ~ make  chicken marinara ~ make marinara  2. end ~ prepare meal ~ make pasta dish ~ make  fettucini marinara ~ make marinara  3. end ~ prepare meal ~ make pasta dish ~ make  spaghetti marinara ~ make marinara  Using the algorithm of the previous section, the  three plans are critiqued and it is found that the  ambiguity matters. The plan involving a meat dish  is found to be faultless but the two plans involving a  pasta dish are found to conflict with another goal of  the cook: to entertain the guest. Using the algo-  rithm of this section, the question asked to resolve  the ambiguity would be "Are you making a pasta  dish (perhaps with justification of why we are ask-  ing)?" After either answer of yes or no we know  enough that the ambiguity no longer matters. Note  that if we just ask the more general "What are you  making?" this allows such uninformative responses as  "dinner" or just "you\'ll see".  When asking a question we propose to ask about  as high a level of goal as possible that still helps to  resolve the ambiguity and to work top down. A top  down approach is better as it provides a useful con-  text for any later queries and makes users give us an  answerat a higher level than they may otherwise do.  Moreover, the user may be mistaken about decompo-  sitions or have some other wrong view of the domain  and by stepping downward through the paths of pos-  sible plans these misconceptions may be detected.  Here is an example. Suppose the user asks "Can I  take numerical analysis?". The system recognizes  two plans.  end ~ get_degree ~ B.A. ~ electives ~ course  end ~ get_degree ~ B.Sc. ~ required ~ course  a. System: Are you pursuing a B.Sc.?  b. System: Are you trying to fulfill your elected or  required courses?  Question (a) is what our algorithm would ask. Ques-  tion (b) is what a procedure which uses a bottom up  approach would ask. But (b) allows potential confu-  sion to go undetected. The user could answer  "required", believing that the course is required for  a B.A., for example. Thus, we advocate Kautz style  plan recognition \\[12\\], as other plan recognition  methods \\[2, 15, 25\\] would stop after branch points  points and thus could only propose electives and  required as the two possible plans. Question (b) is  the only question this previous work could ask the  user.  Starting with the top most goals and working  down may sometimes give as many questions as bot-  tom up approaches. However, others have noted  that bottom up dialogues are difficult to assimilate  without misinterpretation \\[2, p. 54\\]. Therefore, we  maintain that the top down approach is more desir-  able. Moreover, some higher level questions, such as  question (a), above, can be eliminated using goals  known from the previous discourse or background  knowledge about the user.  Current extensions we are examining include  allowing the user to have multiple goals so that more  than one path from top level goals to action may be  correct. This requires resolving ambiguity in multi-  ple paths through the set of possible plans. This  could be done in a depth-first manner, using clue or  redirection words to guide the user when we return  to resolve the ambiguity in the other branches. We  are also investigating selecting the minimal subset of  the possible plans from those with faults and those  without faults (at the moment he algorithm always  takes the subset with faults). Discuss ion  and Conc lus ion   In this section we summarize our proposals and  defend our position that this straightforward way of  doing things is a good way. With reference to Fig.  1, we discuss the design of boxes 2, 3, and 4 and the  tradeoffs involved between boxes 2 and 3.  Box 2: Resolve the ambiguity with heuris-  tics. As mentioned earlier, many researchers have  proposed heuristics to prefer one plan over another  \\[1, 2, 17, 8, 18, 12, 23, 14\\]. Some of these heuristics  can be incompatible with cooperative response  147  generation. For example, Allen\'s \\[1\\] preference  heuristics are generally incompatible with recogniz-  ing and responding to faulty plans (such as the  response in Example 1). Because we are using plan  recognition for response generation, this should  affect the design of Box 2 and therefore what gets  passed to Box 3.  Box 3: Reso lve  the  ambigu i ty  w i th  the  user.   Previous work in response generation makes the  assumption that what gets passed to the RG com-  ponent is a single plan the PR \'component proposes  the user is pursuing. We argue that, unless we are  willing to sometimes arbitrarily commit to one plan  instead of another, there will be times when one plan  cannot be chosen over another and therefore there  will be ambiguity about which plan the user is pur-  suing. Result: we need a method to resolve the  ambiguity. In plan recognition in a discourse setting  (as opposed to key-hole recognition), the goals and  plan the user holds are knowable simply by asking  the user. But we do not want to always just ask if it  is not necessary so we need to know when to start a  clarification dialogue and what to say. And when we  do ask, we want to ask the minimal number of ques-  tions necessary to resolve the ambiguity until it no  longer matters. To this end, box 3 contains a pro-  cedure that estimates by plan critiquing whether the  ambiguity matters to the task of formulating a  response. If the ambiguity does not matter the  result is passed to box 4. If the ambiguity does  matter, a procedure is called that starts a clarifica-  tion dialogue, responding to the user\'s question with  questions that iteratively differentiate between the  possibilities.  Box 2 vs. Box  3: The  t radeof fs .  Much previ-  ous work in plan recognition makes the assumption  that we want the PR component to commit to and  return a single plan. Carberry and McKeown, for  example, use a strong heuristic to commit to a single  plan \\[2, 17\\]. However, this means the system will at  times commit to the wrong plan. Doing itthis way  requires the ability to handle natural language  debugging dialogues. Why we do not want to com-  mit to a single plan and then, if we are wrong, repair  using a debugging dialogue? Carberry \\[5, p.4\\] argues  that a system will appear "unintelligent, obtuse, and  uncooperative" if it engages in lengthy clarification  dialogues. However, a procedure to perform a  debugging dialogue is not specified and is, we specu-  late, a difficult problem. We argue for not commit-  ting early. Our hypothesis is that a clarification  dialogue is better than a debugging dialogue. The  questions in the clarification dialogues are simple to  answer, whereas determining that the system has  misunderstood your goals and plan requires users to  engage in plan recognition. That is, users must  recognize the plan the RG component is using from  its responses and note that it differs from their  plans. Moreover, the user may not recognize we are  wrong and be mislead. Finally, we argue that, if the  questions are carefully chosen, the clarification dialo-  gues need not be lengthy or too frequent. Note that  preference heuristics can still be used in our  approach. These would best be applied when too  many top level goals give an unwieldy clarification  question.  Box 4: Generate  the  response .  Once Box 3 has  estimated that any remaining ambiguity does not  matter to generating a cooperative response, the dis-  junction of possible plans is passed to Box 4. There  are two cases; both can now be handled as in previ-  ous work except that there is now the additional  complication that we allow one action to contribute  to more than one chain of goals. The response com-  ponent must then generate a conjunctive response  that addresses each of the goals.  1.Every plan is faultless, so we just give a direct  answer to the user\'s query but ignore the underly-  ing goals and plan until further queries help clarify  which plan the user is pursuing, and  2. Every plan has the same fault, so we give a direct  answer plus some additional information that  warns the user about the deficiency and perhaps  suggests ome alternatives ( ee \\[10\\], \\[26\\]).  Soap Box:  This paper offers an important con-  tribution to natural language generation. It  discusses a clear criterion for when to initiate a cla-  rification dialogue and proposes a specific solution to  what questions should be asked of the user to  achieve clarification. We believe that natural  language response generation systems should be  designed to involve the user more directly and this is  sometimes achieved quite successfully with our pro-  posals.  There may be tradeoffs between overcommitting  in the plan recognition process and engaging in  lengthy clarification dialogue, particularly with a  large setof complex candidate plans. This may sug-  gest applying pruning heuristics more actively in the  plan recognition process (Box 2) to reduce the  number of questions asked in the clarification dialo-  gue (Box 3). For future work, these tradeoffs will be  examined more closely as we test the algorithms  more extensively.  Acknowledgements .  We thank Fei Song and  Bruce Spencer for comments on an earlier version of  this paper and for many discussions about plan  recognition. Financial assistance was received from  the Natural Sciences and Engineering Research  Council of Canada (NSERC). '
1.0	W08-0123	A00-1008	b'1 Introduction Human computer dialogue is a wide research area in Artificial Intelligence. Computer dialogue is now used at production stage for applications such as tutorial dialogue ? that helps teaching students (Freedman, 2000) ? task-oriented dialogue ? that achieves a particular, limited task, such as booking a trip (Allen et al, 2000) ? and chatbot dialogue (Levy et al, 1997) ? that is used within entertainment and help systems. None of these approaches use persuasion as a mechanism to achieve dialogue goals. However, research towards the use of persuasion in Human Computer Interactions has spawned around the field of natural argumentation (Norman and Reed, 2003). Similarly research on Embodied Conversational Agents (ECA) (Bickmore and Picard, 2005) is also attempting to improve the persuasiveness of agents with persuasion techniques; however, it concentrates on the visual representation of the interlocutor rather than the dialogue management. Previous research on human computer dialogue has rarely focused on persuasive techniques (Guerini, Stock, and Zancanaro, 2004, initiated some research in that field). Our dialogue management system applies a novel method, taking advantage of persuasive and argumentation techniques to achieve persuasive dialogue. According to the cognitive dissonance theory (Festinger, 1957), people will try to minimise the discrepancy between their behaviour and their beliefs by integrating new beliefs or distorting existing ones. In this paper, we approach persuasion as a process shaping user?s beliefs to eventually change their behaviour. The presented dialogue management system has been developed to work on known limitations of current dialogue systems: The impression of lack of control is an issue when the user is interacting with a purely task-oriented dialogue system (Farzanfar et al, 2005). The system follows a plan to achieve the particular task, and the user?s dialogue moves are dictated by the planner and the plan operators. The lack of empathy of computers is also a problem in human-computer interaction for applications such as health-care, where persuasive dialogue could be applied (Bickmore and Giorgino, 2004). The system does not respond to the user?s personal and emotional state, which sometimes lowers the user?s implication in the dialogue. However, existing research (Klein, Moon, and Picard, 1999) shows that a system that gives appropriate response to the user?s emotion can lower frustration. In human-human communication, these limitations reduce the effectiveness of persuasion 138 (Stiff and Mongeau, 2002). Even if the response towards the computer is not always identical to the one to humans, it seems sensible to think that persuasive dialogue systems can be improved by applying known findings from human-human communication. The dialogue management architecture described in this paper (see Figure 1) addresses these dialogue management issues by using a novel layered approach to dialogue management, allowing the mixing of techniques from task-oriented dialogue management and chatbot techniques (see Section 4). Figure 1: Layered Management Architecture The use of a planner guarantees the consistency of the dialogue and the achievement of persuasive goals (see Section 4.2). Argumentative dialogue can be seen as a form of task-oriented dialogue where the system?s task is to persuade the user by presenting the arguments. Thus, the dialogue managerfirst uses a task-oriented dialogue methodology to create a dialogue plan that will determine the content of the dialogue. The planning component?s role is to guarantee the consistency of the dialogue and the achievement of the persuasive goals. In state-of-the-art task-oriented dialogue management systems, the planner provides instructions for a surface realizer (Green and Lehman, 2002), responsible of generating the utterance corresponding to the plan step. Our approach is different to allow more reactivity to the user and give a feeling of control over the dialogue. In this layered approach, the reactive component provides a direct reaction to the user input, generating one or more utterances for a given plan step, allowing for reactions to user?s counter arguments as well as backchannel and chitchat phases without cluttering the plan. Experimental results show that this layered approach allows the user to feel more comfortable in the dialogue while preserving the dialogue consistency provided by the planner. Eventually, this translates into a more persuasive dialogue (see Section 6). 2 Related Work Persuasion through dialogue is a novel field of Human Computer Interaction. Reiter, Robertson, and Osman (2003),Reed (1998) and Carenini and Moore (2000) apply persuasive communication principles to natural language generation, but only focus on monologue. The 3-tier planner for tutoring dialogue by Zinn, Moore, and Core (2002) provides a dialogue management technique close to our approach: a top-tier generates a dialogue plan, the middle-tier generates refinements to the plan and the bottom-tier generates utterances. Mazzotta, de Rosis, and Carofiglio (2007) also propose a planning framework for user-adapted persuasion where the plan operators are mapped to natural language (or ECA) generation. However, these planning approaches do not include a mechanism to react to user?s counter arguments that are difficult to plan beforehand. This paper propose a novel approach that could improve the user?s comfort in the dialogue as well as its persuasiveness. 3 Case Study Part of the problem in evaluating persuasive dialogue is using an effective evaluation framework. Moon (1998) uses the Desert Survival Scenario to evaluate the difference of persuasion and trust in interaction between humans when face-to-face or when mediated by a computer system (via an instant messaging platform). The Desert Survival Scenario (Lafferty, Eady, and Elmers, 1974) is a negotiation scenario used in team training. The team is put in a scenario where they are stranded in the desert after a plane crash. They have to negotiate a ranking of the most eligible items (knife, compass, map, etc.) that they should keep for their survival. For the evaluation of the dialogue system, a similar scenario is presented to the participants. The user has to choose an initial preferred ranking of items 139 and then engages in a discussion with the dialogue system that tries to persuade the user to change the ranking. At the end of the dialogue, the user has the opportunity to either change or keep the ranking. The architecture of the dialogue system is described throughout this paper using examples from the Desert Scenario. The full evaluation protocol is described in Section 5 and 6. 4 DialogueManagement Architecture The following sections provide a description of the dialogue management architecture introduced in Figure 1. 4.1 Argumentation Model The Argumentation model represents the different arguments (conclusions and premises) that can be proposed by the user or by the system. Figure 2 gives a simplified example of the Desert Scenario model. Figure 2: Argumentation Model Sample This model shows the different facts that are known by the system and the relations between them. Arrows represent the support relation between two facts. For example, rescue knows where you are is a support to the fact goal(signal) (the user goal is to signal presence to the rescue) as well as a support to goal(stay put) (the user goal is to stay close to the wreckage). This relational model is comparable to the argumentation framework proposed by Dung (1995), but stores more information about each argument for reasoning within the planning and reactive component (see Section 4.2). Each fact in this model represents a belief to be introduced to the user. For example, when the dialogue tries to achieve the goal reorder(flashlight > air map): the system wants the user to believe that the ?flashlight? item should be ranked higher than the ?air map? item. The argumentation model describes the argumentation process that is required to introduce this new belief: the system first has to make sure the user believes in rate lower(air map) and rate higher(flashlight). Lower level facts (see Figure 2) are the goal facts of the dialogue, the ones the system chooses as dialogue goals, according to known user beliefs and the system?s goal beliefs (e.g. according to the ranking the system is trying to defend). The facts in the middle of the hierarchy are intermediate facts that need to be asserted during the dialogue. The toplevel facts are world knowledge: facts that require minimum defense and can be easily grounded in the dialogue. 4.2 Planning Component The planning component?s task is to find a plan using the argumentation model to introduce the required facts in the user?s belief to support the persuasive goals. The plan is describes a path in the argumentation model beliefs hierarchy that translates to argumentation segments in the dialogue. In our current evaluation method, the goal of the dialogue is to change the user?s beliefs about the items so that the user eventually changes the ranking. At the beginning of the dialogue, the ranking of the system is chosen and persuasive goals are computed for the dialogue. These persuasive goals correspond to the lower level facts in the argumentation model ? like ?reorder(flashlight > air map)? in our previous example. The available planning operators are: use world(fact) describes a step in the dialogue that introduces a simple fact to the user. ground(fact) describes a step in the dialogue that grounds a fact in the user beliefs. Grounding a fact is a different task from the use world operator as it will need more support during the dialogue. do support([fact0, fact1, . . . ], fact2) describes a complex support operation. The system will initiate a dialoguesegment supporting fact2 with the facts fact1 and fact0, etc. that have previously been introduced in the user beliefs. The planning component can also use two non-argumentative operators, do greetings and 140 do farewells, that are placed respectively at the beginning and the end of the dialogue plan to open and close the session. Here is an example plan using the two arguments described in Figure 2 to support the goal reorder(flashlight > air map): Step 1 do greetings Step 2 use world(goal(be found)) ground(rescue knows where you are) ground(can(helpatnight, item(flashlight))) Step 3 do support([can(helpatnight, item(flashlight))], rate higher(item(flashlight))) do support( [rescue knows where you are, goal(be found)], goal(stay put)) Step 4 do support([goal(stay put)], rate lower(item(air map))) Step 5 do support(..., reorder(item(flashlight), item(air map))) Step 6 do farewells The plan is then interpreted by the reactive component that is responsible for realizing each step in a dialogue segment. 4.3 The Reactive Component The reactive component?s first task is to realize the operators chosen by the planning component into dialogue utterance(s). However, it should not be mistaken for a surface language realizer. The reactive component?s task, when realizing the operator, is to decide how to present the particular argumentation operator and its parameters to the user according to the dialogue context and the user?s reaction to the argument. This reactive process is described in the following sections. 4.3.1 Realization and Reaction Strategies Each step of the plan describes the general topic of a dialogue segment1. A dialogue segment is a set of utterances from the system and from 1i.e. it is not directly interpreted as an instruction to generate one unique utterance. the user that are related to a particular argument. For example, in the Desert Scenario, the operator ground(can(helpatnight, item(flashlight))) may result in the following set of utterances: S(ystem) I think the flashlight could be useful as it could help us at night, U(ser) How is that? We are not going to move during the night. S well, if we want to collect water, it will be best to do things at night and not under the burning sun. U I see. It could be useful then. In this example, the ground operator has been realized by the reactive component in two different utterances to react to the user?s interaction. The goal of the reactive component is to make the user feel that the system understands what has been said. It is also important to avoid replanning as it tries to defend the arguments chosen in the plan. As described in Section 4.2, the planner relies on the argumentation model to create a dialogue plan. Encoding all possible defenses and reactions to the user directly in this model will explode the search space of the planner and require careful authoring to avoid planning inconsistencies2 . In addition, predicting at the planning level what counter arguments a user is likely to make requires a prior knowledge of the user?s beliefs. At the beginning of a one-off dialogue, it is not possible to make prior assumptions on the user?s beliefs; the system has ashallow knowledge of the user?s beliefs and will discover them as the dialogue goes. Hence, it is more natural to author a reactive dialogue that will respond to the user?s counter arguments as they come and extends the user beliefs model as it goes. In our architecture if the user is disagreeing with an argument, the plan is not revised directly; if possible, the reactive component selects new, contextually appropriate, supporting facts for the current plan operator. It can do this multiple consecutive local repairs if the user needs more convincing and the domain model provides enough defenses. This allows for a simpler planning framework. 2a new plan could go against the previously used arguments. 141 In addition, when available, and even if the user agrees with the current argument, the reactive component can also choose from a set of ?dialogue smoothing? or backchannel utterances to make the dialogue feel more natural. Here is an example from the Desert Scenario: S We don?t have much water, we need to be rescued as soon as possible. (from plan step: user world( goal(be found))) U right S I am glad we agree.(backchannel) S There is a good chance that the rescue team already knows our whereabouts. We should be optimistic and plan accordingly, don?t you think? (from plan step: use world( rescue knows where you are)) 4.3.2 Detecting user reactions The reactive component needs to detect if the user is agreeing to its current argument or resisting the new fact that is presented. Because the dialogue management system was developed from the perspective of a system that could be easily ported to different domains, choice was made to use a domain independent and robust agreement/disagreement detection. The agreement/disagreement detection is based on an utterance classifier. The classifier is a cascade of binary Support Vector Machines (SVM) (Vapnik, 2000) trained on the ICSI Meeting corpus (Janin et al, 2003). The corpus contains 8135 spurts3 annotated with agreement/disagreement information Hillard, Ostendorf, and Shriberg (2003). A multi-class SVM classifier is trained on local features of the spurts such as a) the length of the spurt, b) the first word of the spurt, c) the bigrams of the spurts, and d) part of speech tags. The classification achieves an accuracy of 83.17% with an N-Fold 4 ways split cross validation. Additional results and comparison with state-of-the-art are available in Appendix A. During the dialogue, the classifier is applied on each of the user?s utterances, trying to determine if the user is agreeing or disagreeing with the system. 3speech utterances that have no pauses longer than .5 seconds. According to this labelling, the strategies described in section 4.3.1 and 4.3.3 are applied. 4.3.3 Revising the plan The reactive component will attempt local repairs to the plan by defending the argumentation move chosen by the planning component. However, there are cases when the user will still not accept an argument. In these cases, imposing the belief to the user is counter-productive and the current goal belief should be dropped from the plan. For each utterance chosen by the reactive component,the belief model of the user is updated to represent the system knowledge of the user?s beliefs. Every time the user agrees to an utterance from the system, the belief model is extended with a new belief; in the previous example, when the user says ?I see, it could be useful then.?, the system detects an agreement (see the Section 4.3.2) and extends the user?s beliefs model with the belief: can(helpatnight, item(flashlight)). The agreement is then followed by a local repair, since the user doesn?t disagree with the statement made, the system also extends the belief model with beliefs relevant to the content of the local repair, thus learning more about the user?s belief model. As a result of this process, when the system decides to revise the plan, the planning component does not start from the same beliefs state as previously. In effect, the system is able to learn user?s beliefs based on the agreement/disagreement with the user, it can therefore make a more effective use of the argumentation hierarchy to find a better plan to achieve the persuasive goals. Still, there are some cases when the planning component will be unable to find a new plan from the current belief state to the goal belief state ? this can happen when the planner has exhausted all its argumentative moves for a particular sub-goal. In these cases, the system has to make concessions and drop the persuasive goals that it cannot fulfil. By dropping goals, the system will lower the final persuasiveness, but guarantees not coercing the user. 4.3.4 Generation Utterance generation is made at the reactive component level. In the current version of the dialogue management system, the utterance generation 142 is based on an extended version of Alicebot AIML 4 . AIML is an XML language that provides a pattern/template generation model mainly used for chatbot systems. An AIML bot defines a set of categories that associate a topic, the context of the previous bot utterance (called that in the AIML terminology), a matching pattern that will match the last user utterance and a generation template. The topic, matching and that field define matching patterns that can contain * wildcards accepting any token(s) of the user utterance (e.g. HELLO * would match any utterance starting by ?Hello?). They are linked to a generation template that can reuse the tokens matched by the patterns wildcards to generate an utterance tailored to the user input and the dialogue context. For the purpose of layered dialogue management, the AIML language has been extended to include more features: 1) A new pattern slot has been introduced to link a set of categories to a particular argumentation operator; 2) Utterances generations are linked to the belief they are trying to introduce to the user and if an agreement is detected, this belief is added to the user belief model. For example, a set of matching categories for the Desert Scenario could be: Plan operator: use world(goal(survive)) Category 1 : Pattern * Template Surviving is our priority, do you want to hear about my desertsurvival insights? Category 2 : Pattern * insights That * survival insights Template I mean, I had a few ideas ...common knowledge I suppose. Category 3 : Pattern * That * survival insights Template Well, we are in this together. Let me tell you of what I think of desert survival, ok? 4http://www.alicebot.org/ These three categories can be used to match the user reaction during the dialogue segment corresponding to the plan operator: use world(goal(survive)). Category 1 is used as the initiative taking generation. It will be the first one to be used when the system comes from a previously finished step. Categories 2-3 are all ?defenses? that support Category 1. They will be used to react to the user if no agreement is detected from the last utterances. For example, if the user says ?what kind of survival insights??? as a reply to the generation from Category 1, a disagreement is detected and the reactive component will have a contextualised answer as given by category 2 whose that pattern matches the last utterance from the system, the pattern pattern matches the user utterance. The dialogue management system uses 187 categories tailored to the Desert Scenario as well as 3737 general categories coming from the Alice chatbot and used to generate the dialogue smoothing utterances. Developing domain specific reactions is a tedious and slow process that was iteratively achieved with Wizard of OZ experiments with real users. In these experiments, users were told they were going to have a dialogue with another human in the Desert Scenario context. The dialogue system manages the whole dialogue, except for the generation phase that is mediated by an expert that can either choose the reaction of the system from an existing set of utterances, or type a new one. 5 Persuasiveness Metric Evaluating a behavior change would require a longterm observation of the behavior that would be dependent to external elements (Bickmore and Picard, 2005). To evaluate our system, an evaluation protocol measuring the change in the beliefs underlying the behavior was chosen. As explained in Section 3, the Desert Scenario is used as a base for the evaluation. Each participant is told that he is stranded in the desert. The user gives a preferred initial ranking Ri of the items (knife, compass, map, etc.). The user then engages in a dialogue with the system. The system then attempts to change the user?s ranking to a different ranking Rs through persuasive dialogue. At the end of the dialogue, the user can change this 143 choice to arrive at a final ranking Rf . The persuasiveness of the dialogue can be measured as the evolution of the distance between the user ranking (Ri, Rf ) and the system ranking (Rs). The Kendall ? distance (Kendall, 1938) is used to compute the pairwise disagreement between two rankings. The change of the Kendall ? distance during the dialogue gives an evaluation of the persuasiveness of the dialogue: Persuasiveness = K?(Ri, Rs) ? K?(Rf , Rs). In the current evaluation protocol, the Rs is always the reverse of theRi, so K?(Ri, Rs) is always the maximum distance possible: n?(n?1)2 where n is the number of items to rank. The minimum Kendall tau distance is 0. If the system was persuasive enough to make the user invert the initial ranking, Persuasiveness of the system is maximum and equal to: n?(n?1)2 . If the system does not succeed in changing the user ranking, then Persuasiveness is zero. 6 Evaluation Results and Discussion 16 participants have been recruited from a variety of ages (from 20 to 59) and background. They were all told to use a web application that describes the Desert Scenario (see Section 3) and proposes to undertake two instant messaging chats with two human users5. However, both discussions are managed by different versions of the dialogue system, following a similar protocol: ? one version of the dialogue is managed by a limited version of the dialogue system, with no reactive component. This version is similar to a purely task-oriented system, planning and revising the plan directly on dialogue failures, ? the second version is the full dialogue system as described in this paper. Each participant went through one dialogue with each system, in a random order. This comparison shows that the dialogue flexibility provided by the reactive component allows a more persuasive dialogue. In addition, when faced with the second dialogue, the participant has formed more beliefs about the scenario and is more able to counter argue. 5The evaluation is available Online at http://www.cs.york.ac.uk/aig/eden Figure 3: Comparative Results. interpret, not coercive, perceived persuasion are on a scale of [0 ? 4] (see Appendix B). Persuasiveness is on a scale of [?10, 10]. Figure 3 reports the independent Persuasiveness metric results as well as interesting answers to a questionnaire that the participants filled after each dialogue (see the Appendix B for detailed results and questionnaire). Over all the dialogues, the full system is 18% more persuasive than the limited system. This is measured by the Persuasiveness metric introduced in Section 5. With the full system, the participants did an average of 1.33 swaps of items towards the system?s ranking. With the limited system, the participants did an average of 0.47 swaps of items away from the system?s ranking. However, the answers to the self evaluated perceived persuasion question show that the participants did not see any significant difference in the ability to persuade of the limited and the full systems. According to the question interpret, the participants found that the limited system understood better what they said. This last result might be explained by the behavior of the systems: the limited system drops an argument at every user disagreement, making the user believe that the disagreement was understood. The full system tries to defend the argument; if possible with a contextually tailored support, however, if this is not available, it may use a generic support, making the user believe he was not fully understood. Our interpretation of the fact that the discrepancy between user self evaluation of the interaction with the system and the measured persuasion is that, even if thefull system is more argumentative, the user 144 didn?t feel coerced6. These results show that a more persuasive dialogue can be achieved without deteriorating the user perception of the interaction.7 Conclusion Our dialogue management system introduces a novel approach to dialogue management by using a layered model mixing the advantages of state-ofthe-art dialogue management approaches. A planning component tailored to the task of argumentation and persuasion searches the ideal path in an argumentation model to persuade the user. To give a reactive and natural feel to the dialogue, this taskoriented layer is extended by a reactive component inspired from the chatbot dialogue management approach. The Desert Scenario evaluation, providing a simple and independent metric for the persuasiveness of the dialogue system provided a good protocol for the evaluation of the dialogue system. This one showed to be 18% more persuasive than a purely task-oriented system that was not able to react to the user interaction as smoothly. Our current research on the dialogue management system consists in developing another evaluation domain where a more complex utterance generation can be used. This will allow going further than the simple template based system, offering more diverse answers to the user and avoiding repetitions; it will also allow us to experiment textual persuasion tailored to other parameters of the user representation, such as the user personality.'	b'1 Introduction  The purpose of the Atlas project is to enlarge the  scope of student interaction in an intelligent  tutoring system (ITS) to include coherent  conversational sequences, including both written  text and GUI actions. A key component of Atlas  is APE, the Atlas Planning Engine, a "just-in-  time" planner specialized for easy construction  and quick generation of hierarchically organized  dialogues. APE is a domain- and task-independent  system. Although to date we have used APE as a  dialogue manager for intelligent tutoring systems,  APE could also be used to manage other types of  human-computer conversation, such as an advice-  giving system or an interactive help system.  Planning is an essential component of a  dialogue-based ITS. Although there are many  reasons for using natural anguage in an ITS, as  soon as the student gives an unexpected response  to a tutor question, the tutor needs to be able to  This research was supported by NSF grant number  9720359 to CIRCLE, the Center for Interdisciplinary  Research on Constructive Learning Environments atthe  University of Pittsburgh and Carnegie-Mellon  University.  plan in order to achieve its goals as well as  respond appropriately to the student\'s tatement.  Yet classical planning is inappropriate for  dialogue generation precisely because it assumes  an unchanging world. A more appropriate  approach is the "practical reason" approach  pioneered by Bratman (1987, 1990). According to  Bratman, human beings maintain plans and prefer  to follow them, but they are also capable of  changing the plans on the fly when needed.  Bratman\'s approach has been introduced into  computer science under the name of reactive  planning (Georgeff and Ingrand 1989, Wilkins et  al. 1995).  In this paper we discuss the rationale for the use  of reactive planning as well as the use of the  hierarchical task network (HTN) style of plan  operators. Then we describe APE (the Atlas  Planning Engine), a dialogue planner we have  implemented to embody the above concepts. We  demonstrate he use of APE by showing how we  have used it to add a dialogue capability to an  existing ITS, the Andes physics tutor. By showing  dialogues that Atlas-Andes can generate, we  demonstrate the advantages of this architecture  over the finite-state machine approach to dialogue  management.  2 Integrated planning and execution for  dialogue generation  2.1 \'Practical reason\' and the BDI model  For an ITS, planning is required in order to ensure  a coherent conversation as well as to accomplish  tutorial goals. But it is impossible to plan a whole  conversation in advance when the student can  respond freely at every turn, just as human beings  cannot plan their daily lives in advance because of  possible changes in conditions. Classical planning  algorithms are inappropriate because the tutor  must be able to change plans based onthe  52  student\'s responses.  For this reason we have adopted the ideas of the  philosopher Michael Bratman (1987, 1990).  Bratman uses the term "practical reason" to  describe his analysis since he is concerned with  how to reason about practical matters. For human  beings, planning is required in order to  accomplish one\'s goals. Bratman\'s key insight is  that human beings tend to follow a plan once they  have one, although they are capable of dropping  an intention or changing a partial plan when  necessary. In other words, human beings do not  decide what to do from scratch at each turn.  Bratman and others who have adopted his  approach use a tripartite mental model that  includes beliefs, desires and intentions (Bratman,  Israel and Pollack 1988, Pollack 1992, Georgeff  et al 1998), hence the name "BDI model."  Beliefs, which are uninstantiated plans in the  speaker\'s head, are reified by the plan library.  Desires are expressed as the agent\'s goals.  Intentions, or plan steps that the agent has  committed to but not yet acted on, are stored in an  agenda. Thus the agent\'s partial plan for  achieving a goal is a network of intentions. A plan  can be left in a partially expanded state until it is  necessary to refine it further.  2.2 Implementation via reactive planning  Bratman\'s approach has been elaborated in a  computer science context by subsequent  researchers (Bratman, Israel and Pollack 1988,  Pollack 1992, Georgeff et al 1998). Reactive  planning (Georgeff and Ingrand 1989, Wilkins et  al. 1995), originally known as "integrated  planning and execution," is one way of  implementing Bratman\'s model. Originally  developed for real-time control of the space  shuttle, reactive planning has since been used in a  variety of other domains. For the Atlas project we  have developed a reactive planner called APE  (Atlas Planning Engine) which uses these ideas to  conduct a conversation. After each student  response, the planner can choose to continue with  its previous intention or change something in the  plan to respond better to the student\'s utterance.  Like most reactive planners, APE is a  hierarchical task network (HTN) style planner  (Yang 1990, Erol, Hendler and Nau 1994).  Hierarchical decomposition asserts that each goal  can be achieved via a series of subgoals instead of  relying on means-end reasoning. Hierarchical  decomposition is more appropriate to dialogue  generation for a number of reasons. First,  decomposition is better suited to the type of large-  scale dialogue planning required in a real-world  tutoring system, as it is easier to establish what a  human speaker will say in a given situation than  to be able to understand why in sufficient detail  and generality to do means-end planning. Second,  Hierarchical decomposition minimizes search  time. Third, our dialogues are task-oriented and  havea hierarchical structure (Grosz and Sidner  1986). In such a case, matching the structure of  the domain simplifies operator development  because they can often be derived from transcripts  of human tutoring sessions. The hierarchy  information is also useful in determining  appropriate referring expressions. Fourth, inter-  leaved planning and execution is important for  dialogue generation because we cannot predict he  human user\'s future utterances. In an HTN-based  system, it is straightforward to implement  interleaved planning and execution because one  only needs to expand the portion of the plan that  is about to be executed. Finally, the conversation  is in a certain sense the trace of the plan. In other  words, we care much more about the actions  generated by the planner than the states involved,  whether implicitly or explicitly specified.  Hierarchical decomposition provides this trace  naturally.  3 Background: the Andes physics tutor  Andes (Gertner, Conati and VanLehn 1998) is an  intelligent tutoring system in the domain of first-  year college physics. Andes teaches via coached  problem solving (VanLehn 1996). In coached  problem solving, the tutoring system tracks the  student as the latter attempts to solve a problem.  If the student gets stuck or deviates too far from a  correct solution path, the tutoring system provides  hints and other assistance.  A sample Andes problem is shown in mid-  solution in Figure 1. A physics problem is given  in the upper-left corner with a picture below it.  Next to the picture the student has begun to  sketch the vectors involved using the GUI buttons  along the left-hand edge of the screen. As the  53  student draws vectors, Andes and the student  cooperatively fill in the variable definitions in the  upper-right corner. Later the student will use the  space below to write equations connecting the  variables.  In this example, the elevator is decelerating, so  the acceleration vector should face the opposite  direction from the velocity vector. (If the  acceleration vector went the same direction as the  velocity vector, the speed of the elevator would  increase and it would crash into the ground.) This  is an important issue in beginning physics; it  occurs in five Andes problems.  When such errors occur, Andes turns the  incorrect item red and provides hints to students  in the lower-left corner of the screen. A sample of  these hints, shown in the order a student would  encounter them, is shown in Fig. 2. But hints are  an output-only form of natural language; the  student can\'t take the initiative or ask a question.  In addition, there is no way for the system to ask  the student a question or lead the student through  a multi-step directed line of reasoning. Thus there  is no way to use some of the effective rhetorical  methodsused by skilled human tutors, such as  analogy and reductio ad absurdum. Current  psychological research suggests that active  methods, where students have to answer  questions, will improve the performance of  tutoring systems.  4 Structure of the Atlas Planning Engine  Figure3 shows a sample plan operator. For  legibility, the key elements have been rendered in  English instead of in Lisp. The hiercx slot  provides a way for the planner to be aware of the  context in which a decomposition is proposed.  Items in the hiercx slot are instantiated and added  to the transient database only so long as the  operator which spawned them is in the agenda.  To initiate a planning session, the user invokes  the planner with an initial goal. The system  searches the operator library to find all operators  whose goal field matches the next goal on the  agenda and whose filter conditions and precon-  An elevator slows to a stop from an initial downward velocity  of 10.0 m\\]s in 2.00 seconds. A passenger in the elevator is  holding a 3.00 kilogram package by a vertical string.  What is the tension in the string during the process?  i ........ ii ....... i i i   . I ?Y  ~TO e\',ev~o, at 10 m/s  elev~or at a stop  mass of p~:w\'.,I,,~  magnitude of the inst~?~taneous Velocity of pack, age ~ {rkne TO v._w  magnitude of the avelage Acceleratiorl of package ,dudng TO... a._x  v_v  a~  - - I   pkg  Figure I: Screen shot of the Andes physics tutor  54  S: (draws acceleration vector in same direction as velocity)  T: Wrong.  S: What\'s wrongwith that?  T: Think about he direction of the acceleration vector.  S: Please explain further.  T: Remember that the direction of acceleration is the direction of the change in velocity.  S: Please explain further.  T: The\'direction of the acceleration vector is straight up.  S: (draws acceleration vector correctly)  Figure 2: Andes hint sequence formatted as dialogue  ditions are satisfied. Goals are represented in  first-order logic without quantifiers and matched  via unification. Since APE is intended especially  for generation of hierarchically organized task-  oriented iscourse, each operator has a multi-step  recipe in the style of Wilkins (1988). When a  match is found, the matching goal is removed  from the agenda and is replaced by the steps in  the recipe. APE has two kinds of primitive  actions; one ends a turn and the other doesn\'t.  From the point of view of discourse generation,  the most important APE recipe items are those  allowing the planner to change the agenda when  necessary. These three types of recipe items make  APE more powerful than a classical planner.  ? Fact: Evaluate a condition. If false, skip the  rest of the recipe. Factis used to allow run-time  decision making by bypassing the rest of an  operator when circumstances change during its  execution. Fact can be used with retry-at to  implement a loop just as in Prolog.  ? Retry-at. The purpose of retry-at is to allow  the planner to back up to a choice point and make  a new decision. It removes goals sequentially  from the top of the agenda, a full operator at a  time, until the supplied argument is false. Then it  restores the parent goal of the last operator  removed, so that further planning can choose a  new way to achieve it. Retry-at implements a Prolog-like choice of alternatives, but it differs  from backtracking in that the new operator is  chosen based on conditions that apply when the  retry operation is executed, rather than on a list of  possible operators formed when the original  operator was chosen. For retry-at o be useful, the  author must provide multiple operators for the  same goal. Each operator must have a set of  preconditions enabling it to be chosen at the  appropriate ime.  ? Prune-replace: The intent of prune-replace is  (de f -operator  hand le -same-d i rec t ion   :goal  (...)  : f i l te r  ()  :p recond (...)  ; We have  asked  a quest ion  about  acce le ra t ion   ; ... and the s tudent  has g iven  an answer   ; ... f rom wh ich  we can deduce  that  s /he  th inks  accel ,  and ve loc i ty  go in  ; the same d i rec t ion   ; and  we have  not  g iven  the exp lanat ion  be low yet   : rec ipe  (...)  ; Te l l  the s tudent :  "But i f  the acce le ra t ion  went  the same  d i rec t ion  as the ve loc i ty ,  then the e levator  wou ld  be speed ing  up."   ; Mark  that  we are g iv ing  th is  exp lanat ion   ; Te l l  the s tudent  that  tu tor  is request ing  another  answer  ("Try aga in . " )   ; Ed i t  the agenda (us ing prune-replace) so that  respond ing  to another   answer  is at the top  of the agenda  :h ie rcx  ())  Figure 3: Sample plan operator  55  to allow the planner to remove goals from the  agenda based on a change in circumstances. It  removes goals sequentiallyfrom the top of the  agenda, one at a time, until the supplied argument  becomes false. Then it replaces the removed goals  with an optional ist of new goals. Prune-replace  allows a type of decision-making frequently used  in dialogue generation. When a conversation  partner does not give the expected response, one  would often like to remove the next goal from the  agenda and replace it with one or more  replacement goals. Prune-replace implements a generalized version of this concept.  APE is domain-independent a d communicates  with a host system via an API. As a partner in a  dialogue, it needs to obtain information from the  world as well as produce output turns.  Preconditions on plan operators can be used to  access information from external knowledge  sources. APE contains a recipe item type that can  be used to execute an external program such as a  call to a GUI interface. APE also has recipe items  allowing the user to assert and retract facts in a  knowledge base. Further details about the APE  planner can be found in (Freedman, 2000).  5 Implementation of At las-Andes  5.1 Architecture of Atlas-Andes  The first system we have implemented with APE  is a prototype Atlas-Andes system that replaces  the hints usually given for an incorrect  acceleration vector by a choice of generated  subdialogues. Figure 4 shows the architecture of  Atlas-Andes; any other system built with APE  would look similar. Robust natural language  understanding in Atlas-Andes is provided by  Ros6\'s CARMEL system (Ros6 2000); it uses the  spelling correction algorithm devised by Elmi and  Evens (1998).  5.2 Structure of human tutorial dialogues  In an earlier analysis (Kim, Freedman and Evens  1998) we showed that a significant portion of  human-human tutorial dialogues can be modeled  with the hierarchical structure of task-oriented  dialogues (Grosz and Sidner 1986). Furthermore,  a main building block of the discourse hierarchy,  corresponding to the transaction level in  Conversation Analysis (Sinclair and Coulthard  1975), matches the tutoring episode defined by  VanLehn et al (1998). A tutoring episode  consists of the turns necessary to help the student  make one correct entry on the interface.  NLU  (CARMEL) Plan Library  User APE  < Interface I  I I  GUI Transient  Interpreter Knowledge  (Andes) Base  Host  (Andes)  Figure 4: Interface between Atlas and host system  56  To obtain empirical data for the Atlas-Andes  plan operators, we analyzed portions of a corpus  of human tutors helping students olve similar  physics problems. Two experienced tutors were  used. Tutor A was a graduate student in computer  science who had majored in physics; tutor B was  a professional physics tutor.  The complete corpus contained solutions to five  physics problems by 41 students each. We  analyzed every tutoringepisode dealing with the  acceleration vector during deceleration, totaling  29 examples divided among 20 students and both  tutors. The tutors had very different styles.  Tutor A tended to provide encouragement rather  than content, making those transcripts less useful  for deriving an information-based approach.  Tutor B used an information-based approach, but  after one wrong answer tended to complete the  solution as a monologue. Largely following  tutor B\'s approach to sequence and content, we  isolated six ways of teaching the student about  direction of acceleration.  5.3 Sample output and evaluation  Figure 5 shows an example of text that can be  generated by the Atlas-Andes ystem, showing an  analogy-based approach to teaching this content.  The operator library used to generate this text  could generate a combinatorially arge number of  versions of this dialogue as well as selected  examples of other ways of teaching about  direction of acceleration.  This operator library used to generate this text  contained 1l 1 plan operators, divided as follows:  Tutoring schemata  Switching between schemata  API and GUI handling  Answer handling  Domain-dep. lex. insertion  Domain-indep. lex. insertion  TOTAL  4 4%  5 4%  33 30%  35 31%  24 22%  10 9%  111 100%  We are currently working on components hat will  allow us to increase the number of physics  concepts covered without a corresponding  increase in the number of operators. The schema  switching operators prevent the tutor from  repeating itself during a physics problem. They  could be reduced or eliminated by a general  discourse history component that tutoring schema  operators could refer to. Domain-dependent  lexical insertion refers to the choice of lexical  items such as car and east in the sample dialogue,  while domain-independent iexical insertion refers  to items such as OK and exactly. Both categories  could be eliminated, or at least severely reduced,  through the use of a text realization package.  Together that would provide a one-third reduction  in the number of operators needed. As the set of  API and GUI handling operators is fixed, that  would reduce by half the number of application  operators needed.  The largest remaining category of operators is  the answer handlers. These operators handle a  variety of answers for each of the five questions  that the system can ask. The answers we  recognize include categories such as "don\'t  know" as well as specific answers (e.g. a direction  perpendicular to the correct answer) which we  recognize because the tutor has specific replies  for them. In order to reduce the number of  S: (draws acceleration vector in same direction as velocity)  T: What is the definition of acceleration?  S: Don\'t know.  T: OK, let\'s try this. If a car was driving along east, which way would you have topush on it  to make it stop?  S: West.  T: Exactly. The opposite direction. So the net force goes the opposite direction, and so does  the acceleration. Try to draw the acceleration vector again now.  S: (draws acceleration vector correctly)  Figure 5: Example of generated ialogue  57  operators further, we must investigate more  general methods of handling student errors. In  particular, we plan to investigate error-classifying  predicates that apply to more than one question as  well as the use of intention-based predicates.  Since the system only covers one rule of physics,  albeit in a variety of ways, we plan to make some  of these efficiency improvements before adding  new rules of physics and testing it with users.  Preconditions for the operators in the plan  library utilize discourse or interaction history, the  current goal hierarchy, recent information such as  the tutor\'s current goal and the student\'s latest  response, shared information such as a model of  objects on the screen, and domain knowledge. As  an example of the latter, if the student draws an  acceleration vector which is incorrect but not  opposite to the velocity vector, a different  response will be generated.  5.4 Discussion  Many previous dialogue-based ITSs have been  implemented with finite-state machines, either  simple or augmented. In the most common finite  state mode\\[, each time the human user issues an  utterance, the processor educes it to one of a  small number of categories. These categories  represent the possible transitions between states.  Thus history can be stored, and context  considered, only by expanding the number of  states. This approach puts an arbitrary restriction  on the amount of context or depth of  conversational nesting that can be considered.  More importantly, it misses the significant  generalization that these types of dialogues are  hierarchical: larger units contain repeated  instances of the same smaller units in different  sequences and instantiated with different values.  Furthermore, the finite-state machine approach  does not allow the author to drop one line of  attack and replace it by another without hard-  coding every possible transition.  It is also clear that the dialogue-based approach  has many benefits over the hint-sequence  approach. In addition to providing a multi-step  teaching methods with new content, it can  respond flexibly to a variety of student answers at  each step and take context into account when  generating a reply.  6 Related work   Wenger (1987), still the chief textbook on ITSs,  states that using a global planner to control an ITS  is too inefficient to try. This is no longer true, if  indeed it ever was. Vassileva (1995) proposes a  system based on AND-OR graphs with a separate  set of rules for reacting to unexpected events.  Lehuen, Nicolle and Luzzati (1996) present a  methodof dialogue analysis that produces  schemata very similar to ours. Earlier dialogue-  based ITSs that use augmented finite-state  machines or equivalent include CIRCSIM-Tutor  (Woo et al 1991, Zhouet al 1999) and the  system described by Woolf (1984). Cook (1998)  uses levels of finite-state machines. None of these  systems provides for predicates with variables or  unification. 7 Conclusions  In this paper we described APE, an integrated  planner and execution system that we have  implemented as part of the Atlas dialogue  manager. APE uses HTN-style operators and is  based on reactive planning concepts. Although  APE is intended largely for use in domains with  hierarchical, multi-turn plans, it can be used to  implement any conversation-based system, where  turns in the \'conversation\' may include graphical  actions and/or text. We illustrated the use of APE  with an example from the Atlas-Andes physics  tutor. We showed that previous models based on  finite-state machines are insufficient to handle the  nested subdialogues and abandoned partial  subdialogues that occur in practical applications.  We showed how APE generated a sample  dialogue that earlier systems could not handle.  Acknowledgments  We thank Abigail Gertner for her generous  assistance with the Andes system, and Michael  Ringenberg for indispensible programming  support. Carolyn Ros6 built the CARMEL  natural language understanding component.  Mohammed EImi and Michael Glass of Illinois  Institute of Technology provided the spelling  correction code. We thank Pamela Jordan and the  referees for their comments.  B8 '
0.0	W08-0123	W11-0322	b'1 Introduction Human computer dialogue is a wide research area in Artificial Intelligence. Computer dialogue is now used at production stage for applications such as tutorial dialogue ? that helps teaching students (Freedman, 2000) ? task-oriented dialogue ? that achieves a particular, limited task, such as booking a trip (Allen et al, 2000) ? and chatbot dialogue (Levy et al, 1997) ? that is used within entertainment and help systems. None of these approaches use persuasion as a mechanism to achieve dialogue goals. However, research towards the use of persuasion in Human Computer Interactions has spawned around the field of natural argumentation (Norman and Reed, 2003). Similarly research on Embodied Conversational Agents (ECA) (Bickmore and Picard, 2005) is also attempting to improve the persuasiveness of agents with persuasion techniques; however, it concentrates on the visual representation of the interlocutor rather than the dialogue management. Previous research on human computer dialogue has rarely focused on persuasive techniques (Guerini, Stock, and Zancanaro, 2004, initiated some research in that field). Our dialogue management system applies a novel method, taking advantage of persuasive and argumentation techniques to achieve persuasive dialogue. According to the cognitive dissonance theory (Festinger, 1957), people will try to minimise the discrepancy between their behaviour and their beliefs by integrating new beliefs or distorting existing ones. In this paper, we approach persuasion as a process shaping user?s beliefs to eventually change their behaviour. The presented dialogue management system has been developed to work on known limitations of current dialogue systems: The impression of lack of control is an issue when the user is interacting with a purely task-oriented dialogue system (Farzanfar et al, 2005). The system follows a plan to achieve the particular task, and the user?s dialogue moves are dictated by the planner and the plan operators. The lack of empathy of computers is also a problem in human-computer interaction for applications such as health-care, where persuasive dialogue could be applied (Bickmore and Giorgino, 2004). The system does not respond to the user?s personal and emotional state, which sometimes lowers the user?s implication in the dialogue. However, existing research (Klein, Moon, and Picard, 1999) shows that a system that gives appropriate response to the user?s emotion can lower frustration. In human-human communication, these limitations reduce the effectiveness of persuasion 138 (Stiff and Mongeau, 2002). Even if the response towards the computer is not always identical to the one to humans, it seems sensible to think that persuasive dialogue systems can be improved by applying known findings from human-human communication. The dialogue management architecture described in this paper (see Figure 1) addresses these dialogue management issues by using a novel layered approach to dialogue management, allowing the mixing of techniques from task-oriented dialogue management and chatbot techniques (see Section 4). Figure 1: Layered Management Architecture The use of a planner guarantees the consistency of the dialogue and the achievement of persuasive goals (see Section 4.2). Argumentative dialogue can be seen as a form of task-oriented dialogue where the system?s task is to persuade the user by presenting the arguments. Thus, the dialogue managerfirst uses a task-oriented dialogue methodology to create a dialogue plan that will determine the content of the dialogue. The planning component?s role is to guarantee the consistency of the dialogue and the achievement of the persuasive goals. In state-of-the-art task-oriented dialogue management systems, the planner provides instructions for a surface realizer (Green and Lehman, 2002), responsible of generating the utterance corresponding to the plan step. Our approach is different to allow more reactivity to the user and give a feeling of control over the dialogue. In this layered approach, the reactive component provides a direct reaction to the user input, generating one or more utterances for a given plan step, allowing for reactions to user?s counter arguments as well as backchannel and chitchat phases without cluttering the plan. Experimental results show that this layered approach allows the user to feel more comfortable in the dialogue while preserving the dialogue consistency provided by the planner. Eventually, this translates into a more persuasive dialogue (see Section 6). 2 Related Work Persuasion through dialogue is a novel field of Human Computer Interaction. Reiter, Robertson, and Osman (2003),Reed (1998) and Carenini and Moore (2000) apply persuasive communication principles to natural language generation, but only focus on monologue. The 3-tier planner for tutoring dialogue by Zinn, Moore, and Core (2002) provides a dialogue management technique close to our approach: a top-tier generates a dialogue plan, the middle-tier generates refinements to the plan and the bottom-tier generates utterances. Mazzotta, de Rosis, and Carofiglio (2007) also propose a planning framework for user-adapted persuasion where the plan operators are mapped to natural language (or ECA) generation. However, these planning approaches do not include a mechanism to react to user?s counter arguments that are difficult to plan beforehand. This paper propose a novel approach that could improve the user?s comfort in the dialogue as well as its persuasiveness. 3 Case Study Part of the problem in evaluating persuasive dialogue is using an effective evaluation framework. Moon (1998) uses the Desert Survival Scenario to evaluate the difference of persuasion and trust in interaction between humans when face-to-face or when mediated by a computer system (via an instant messaging platform). The Desert Survival Scenario (Lafferty, Eady, and Elmers, 1974) is a negotiation scenario used in team training. The team is put in a scenario where they are stranded in the desert after a plane crash. They have to negotiate a ranking of the most eligible items (knife, compass, map, etc.) that they should keep for their survival. For the evaluation of the dialogue system, a similar scenario is presented to the participants. The user has to choose an initial preferred ranking of items 139 and then engages in a discussion with the dialogue system that tries to persuade the user to change the ranking. At the end of the dialogue, the user has the opportunity to either change or keep the ranking. The architecture of the dialogue system is described throughout this paper using examples from the Desert Scenario. The full evaluation protocol is described in Section 5 and 6. 4 DialogueManagement Architecture The following sections provide a description of the dialogue management architecture introduced in Figure 1. 4.1 Argumentation Model The Argumentation model represents the different arguments (conclusions and premises) that can be proposed by the user or by the system. Figure 2 gives a simplified example of the Desert Scenario model. Figure 2: Argumentation Model Sample This model shows the different facts that are known by the system and the relations between them. Arrows represent the support relation between two facts. For example, rescue knows where you are is a support to the fact goal(signal) (the user goal is to signal presence to the rescue) as well as a support to goal(stay put) (the user goal is to stay close to the wreckage). This relational model is comparable to the argumentation framework proposed by Dung (1995), but stores more information about each argument for reasoning within the planning and reactive component (see Section 4.2). Each fact in this model represents a belief to be introduced to the user. For example, when the dialogue tries to achieve the goal reorder(flashlight > air map): the system wants the user to believe that the ?flashlight? item should be ranked higher than the ?air map? item. The argumentation model describes the argumentation process that is required to introduce this new belief: the system first has to make sure the user believes in rate lower(air map) and rate higher(flashlight). Lower level facts (see Figure 2) are the goal facts of the dialogue, the ones the system chooses as dialogue goals, according to known user beliefs and the system?s goal beliefs (e.g. according to the ranking the system is trying to defend). The facts in the middle of the hierarchy are intermediate facts that need to be asserted during the dialogue. The toplevel facts are world knowledge: facts that require minimum defense and can be easily grounded in the dialogue. 4.2 Planning Component The planning component?s task is to find a plan using the argumentation model to introduce the required facts in the user?s belief to support the persuasive goals. The plan is describes a path in the argumentation model beliefs hierarchy that translates to argumentation segments in the dialogue. In our current evaluation method, the goal of the dialogue is to change the user?s beliefs about the items so that the user eventually changes the ranking. At the beginning of the dialogue, the ranking of the system is chosen and persuasive goals are computed for the dialogue. These persuasive goals correspond to the lower level facts in the argumentation model ? like ?reorder(flashlight > air map)? in our previous example. The available planning operators are: use world(fact) describes a step in the dialogue that introduces a simple fact to the user. ground(fact) describes a step in the dialogue that grounds a fact in the user beliefs. Grounding a fact is a different task from the use world operator as it will need more support during the dialogue. do support([fact0, fact1, . . . ], fact2) describes a complex support operation. The system will initiate a dialoguesegment supporting fact2 with the facts fact1 and fact0, etc. that have previously been introduced in the user beliefs. The planning component can also use two non-argumentative operators, do greetings and 140 do farewells, that are placed respectively at the beginning and the end of the dialogue plan to open and close the session. Here is an example plan using the two arguments described in Figure 2 to support the goal reorder(flashlight > air map): Step 1 do greetings Step 2 use world(goal(be found)) ground(rescue knows where you are) ground(can(helpatnight, item(flashlight))) Step 3 do support([can(helpatnight, item(flashlight))], rate higher(item(flashlight))) do support( [rescue knows where you are, goal(be found)], goal(stay put)) Step 4 do support([goal(stay put)], rate lower(item(air map))) Step 5 do support(..., reorder(item(flashlight), item(air map))) Step 6 do farewells The plan is then interpreted by the reactive component that is responsible for realizing each step in a dialogue segment. 4.3 The Reactive Component The reactive component?s first task is to realize the operators chosen by the planning component into dialogue utterance(s). However, it should not be mistaken for a surface language realizer. The reactive component?s task, when realizing the operator, is to decide how to present the particular argumentation operator and its parameters to the user according to the dialogue context and the user?s reaction to the argument. This reactive process is described in the following sections. 4.3.1 Realization and Reaction Strategies Each step of the plan describes the general topic of a dialogue segment1. A dialogue segment is a set of utterances from the system and from 1i.e. it is not directly interpreted as an instruction to generate one unique utterance. the user that are related to a particular argument. For example, in the Desert Scenario, the operator ground(can(helpatnight, item(flashlight))) may result in the following set of utterances: S(ystem) I think the flashlight could be useful as it could help us at night, U(ser) How is that? We are not going to move during the night. S well, if we want to collect water, it will be best to do things at night and not under the burning sun. U I see. It could be useful then. In this example, the ground operator has been realized by the reactive component in two different utterances to react to the user?s interaction. The goal of the reactive component is to make the user feel that the system understands what has been said. It is also important to avoid replanning as it tries to defend the arguments chosen in the plan. As described in Section 4.2, the planner relies on the argumentation model to create a dialogue plan. Encoding all possible defenses and reactions to the user directly in this model will explode the search space of the planner and require careful authoring to avoid planning inconsistencies2 . In addition, predicting at the planning level what counter arguments a user is likely to make requires a prior knowledge of the user?s beliefs. At the beginning of a one-off dialogue, it is not possible to make prior assumptions on the user?s beliefs; the system has ashallow knowledge of the user?s beliefs and will discover them as the dialogue goes. Hence, it is more natural to author a reactive dialogue that will respond to the user?s counter arguments as they come and extends the user beliefs model as it goes. In our architecture if the user is disagreeing with an argument, the plan is not revised directly; if possible, the reactive component selects new, contextually appropriate, supporting facts for the current plan operator. It can do this multiple consecutive local repairs if the user needs more convincing and the domain model provides enough defenses. This allows for a simpler planning framework. 2a new plan could go against the previously used arguments. 141 In addition, when available, and even if the user agrees with the current argument, the reactive component can also choose from a set of ?dialogue smoothing? or backchannel utterances to make the dialogue feel more natural. Here is an example from the Desert Scenario: S We don?t have much water, we need to be rescued as soon as possible. (from plan step: user world( goal(be found))) U right S I am glad we agree.(backchannel) S There is a good chance that the rescue team already knows our whereabouts. We should be optimistic and plan accordingly, don?t you think? (from plan step: use world( rescue knows where you are)) 4.3.2 Detecting user reactions The reactive component needs to detect if the user is agreeing to its current argument or resisting the new fact that is presented. Because the dialogue management system was developed from the perspective of a system that could be easily ported to different domains, choice was made to use a domain independent and robust agreement/disagreement detection. The agreement/disagreement detection is based on an utterance classifier. The classifier is a cascade of binary Support Vector Machines (SVM) (Vapnik, 2000) trained on the ICSI Meeting corpus (Janin et al, 2003). The corpus contains 8135 spurts3 annotated with agreement/disagreement information Hillard, Ostendorf, and Shriberg (2003). A multi-class SVM classifier is trained on local features of the spurts such as a) the length of the spurt, b) the first word of the spurt, c) the bigrams of the spurts, and d) part of speech tags. The classification achieves an accuracy of 83.17% with an N-Fold 4 ways split cross validation. Additional results and comparison with state-of-the-art are available in Appendix A. During the dialogue, the classifier is applied on each of the user?s utterances, trying to determine if the user is agreeing or disagreeing with the system. 3speech utterances that have no pauses longer than .5 seconds. According to this labelling, the strategies described in section 4.3.1 and 4.3.3 are applied. 4.3.3 Revising the plan The reactive component will attempt local repairs to the plan by defending the argumentation move chosen by the planning component. However, there are cases when the user will still not accept an argument. In these cases, imposing the belief to the user is counter-productive and the current goal belief should be dropped from the plan. For each utterance chosen by the reactive component,the belief model of the user is updated to represent the system knowledge of the user?s beliefs. Every time the user agrees to an utterance from the system, the belief model is extended with a new belief; in the previous example, when the user says ?I see, it could be useful then.?, the system detects an agreement (see the Section 4.3.2) and extends the user?s beliefs model with the belief: can(helpatnight, item(flashlight)). The agreement is then followed by a local repair, since the user doesn?t disagree with the statement made, the system also extends the belief model with beliefs relevant to the content of the local repair, thus learning more about the user?s belief model. As a result of this process, when the system decides to revise the plan, the planning component does not start from the same beliefs state as previously. In effect, the system is able to learn user?s beliefs based on the agreement/disagreement with the user, it can therefore make a more effective use of the argumentation hierarchy to find a better plan to achieve the persuasive goals. Still, there are some cases when the planning component will be unable to find a new plan from the current belief state to the goal belief state ? this can happen when the planner has exhausted all its argumentative moves for a particular sub-goal. In these cases, the system has to make concessions and drop the persuasive goals that it cannot fulfil. By dropping goals, the system will lower the final persuasiveness, but guarantees not coercing the user. 4.3.4 Generation Utterance generation is made at the reactive component level. In the current version of the dialogue management system, the utterance generation 142 is based on an extended version of Alicebot AIML 4 . AIML is an XML language that provides a pattern/template generation model mainly used for chatbot systems. An AIML bot defines a set of categories that associate a topic, the context of the previous bot utterance (called that in the AIML terminology), a matching pattern that will match the last user utterance and a generation template. The topic, matching and that field define matching patterns that can contain * wildcards accepting any token(s) of the user utterance (e.g. HELLO * would match any utterance starting by ?Hello?). They are linked to a generation template that can reuse the tokens matched by the patterns wildcards to generate an utterance tailored to the user input and the dialogue context. For the purpose of layered dialogue management, the AIML language has been extended to include more features: 1) A new pattern slot has been introduced to link a set of categories to a particular argumentation operator; 2) Utterances generations are linked to the belief they are trying to introduce to the user and if an agreement is detected, this belief is added to the user belief model. For example, a set of matching categories for the Desert Scenario could be: Plan operator: use world(goal(survive)) Category 1 : Pattern * Template Surviving is our priority, do you want to hear about my desertsurvival insights? Category 2 : Pattern * insights That * survival insights Template I mean, I had a few ideas ...common knowledge I suppose. Category 3 : Pattern * That * survival insights Template Well, we are in this together. Let me tell you of what I think of desert survival, ok? 4http://www.alicebot.org/ These three categories can be used to match the user reaction during the dialogue segment corresponding to the plan operator: use world(goal(survive)). Category 1 is used as the initiative taking generation. It will be the first one to be used when the system comes from a previously finished step. Categories 2-3 are all ?defenses? that support Category 1. They will be used to react to the user if no agreement is detected from the last utterances. For example, if the user says ?what kind of survival insights??? as a reply to the generation from Category 1, a disagreement is detected and the reactive component will have a contextualised answer as given by category 2 whose that pattern matches the last utterance from the system, the pattern pattern matches the user utterance. The dialogue management system uses 187 categories tailored to the Desert Scenario as well as 3737 general categories coming from the Alice chatbot and used to generate the dialogue smoothing utterances. Developing domain specific reactions is a tedious and slow process that was iteratively achieved with Wizard of OZ experiments with real users. In these experiments, users were told they were going to have a dialogue with another human in the Desert Scenario context. The dialogue system manages the whole dialogue, except for the generation phase that is mediated by an expert that can either choose the reaction of the system from an existing set of utterances, or type a new one. 5 Persuasiveness Metric Evaluating a behavior change would require a longterm observation of the behavior that would be dependent to external elements (Bickmore and Picard, 2005). To evaluate our system, an evaluation protocol measuring the change in the beliefs underlying the behavior was chosen. As explained in Section 3, the Desert Scenario is used as a base for the evaluation. Each participant is told that he is stranded in the desert. The user gives a preferred initial ranking Ri of the items (knife, compass, map, etc.). The user then engages in a dialogue with the system. The system then attempts to change the user?s ranking to a different ranking Rs through persuasive dialogue. At the end of the dialogue, the user can change this 143 choice to arrive at a final ranking Rf . The persuasiveness of the dialogue can be measured as the evolution of the distance between the user ranking (Ri, Rf ) and the system ranking (Rs). The Kendall ? distance (Kendall, 1938) is used to compute the pairwise disagreement between two rankings. The change of the Kendall ? distance during the dialogue gives an evaluation of the persuasiveness of the dialogue: Persuasiveness = K?(Ri, Rs) ? K?(Rf , Rs). In the current evaluation protocol, the Rs is always the reverse of theRi, so K?(Ri, Rs) is always the maximum distance possible: n?(n?1)2 where n is the number of items to rank. The minimum Kendall tau distance is 0. If the system was persuasive enough to make the user invert the initial ranking, Persuasiveness of the system is maximum and equal to: n?(n?1)2 . If the system does not succeed in changing the user ranking, then Persuasiveness is zero. 6 Evaluation Results and Discussion 16 participants have been recruited from a variety of ages (from 20 to 59) and background. They were all told to use a web application that describes the Desert Scenario (see Section 3) and proposes to undertake two instant messaging chats with two human users5. However, both discussions are managed by different versions of the dialogue system, following a similar protocol: ? one version of the dialogue is managed by a limited version of the dialogue system, with no reactive component. This version is similar to a purely task-oriented system, planning and revising the plan directly on dialogue failures, ? the second version is the full dialogue system as described in this paper. Each participant went through one dialogue with each system, in a random order. This comparison shows that the dialogue flexibility provided by the reactive component allows a more persuasive dialogue. In addition, when faced with the second dialogue, the participant has formed more beliefs about the scenario and is more able to counter argue. 5The evaluation is available Online at http://www.cs.york.ac.uk/aig/eden Figure 3: Comparative Results. interpret, not coercive, perceived persuasion are on a scale of [0 ? 4] (see Appendix B). Persuasiveness is on a scale of [?10, 10]. Figure 3 reports the independent Persuasiveness metric results as well as interesting answers to a questionnaire that the participants filled after each dialogue (see the Appendix B for detailed results and questionnaire). Over all the dialogues, the full system is 18% more persuasive than the limited system. This is measured by the Persuasiveness metric introduced in Section 5. With the full system, the participants did an average of 1.33 swaps of items towards the system?s ranking. With the limited system, the participants did an average of 0.47 swaps of items away from the system?s ranking. However, the answers to the self evaluated perceived persuasion question show that the participants did not see any significant difference in the ability to persuade of the limited and the full systems. According to the question interpret, the participants found that the limited system understood better what they said. This last result might be explained by the behavior of the systems: the limited system drops an argument at every user disagreement, making the user believe that the disagreement was understood. The full system tries to defend the argument; if possible with a contextually tailored support, however, if this is not available, it may use a generic support, making the user believe he was not fully understood. Our interpretation of the fact that the discrepancy between user self evaluation of the interaction with the system and the measured persuasion is that, even if thefull system is more argumentative, the user 144 didn?t feel coerced6. These results show that a more persuasive dialogue can be achieved without deteriorating the user perception of the interaction.7 Conclusion Our dialogue management system introduces a novel approach to dialogue management by using a layered model mixing the advantages of state-ofthe-art dialogue management approaches. A planning component tailored to the task of argumentation and persuasion searches the ideal path in an argumentation model to persuade the user. To give a reactive and natural feel to the dialogue, this taskoriented layer is extended by a reactive component inspired from the chatbot dialogue management approach. The Desert Scenario evaluation, providing a simple and independent metric for the persuasiveness of the dialogue system provided a good protocol for the evaluation of the dialogue system. This one showed to be 18% more persuasive than a purely task-oriented system that was not able to react to the user interaction as smoothly. Our current research on the dialogue management system consists in developing another evaluation domain where a more complex utterance generation can be used. This will allow going further than the simple template based system, offering more diverse answers to the user and avoiding repetitions; it will also allow us to experiment textual persuasion tailored to other parameters of the user representation, such as the user personality.'	b'1 Introduction A growing interest in using semantic relatedness in word sense disambiguation (WSD) tasks has spurred investigations into the limitations of the WordNet ontology (Fellbaum, 1998) for this purpose. Although WordNet comprises a rich set of semantic 1http://www.cs.ucf.edu/? seansz/sem links between word senses (or concepts), indicating semantic similarity through subsumptive hypernymic and hyponymic relations (among others), it lacks a general indication of semantic relatedness. We present a semantic network that is automatically acquired from lexical co-occurrence in Wikipedia, and indicates general semantic relatedness between noun senses in WordNet 3.0. In this work, the discovery of relatedness is a context-sparse affair that takes place in absentia of the semantic annotations of Wikipedia, such as inter-article links, entries in disambiguation pages, the title of the article from which a sentence is extracted, and so on. We released an earlier version of such a network that was limited by the fact that only relationships involving at least one monosemous noun had been included, and it was not evaluated on a WSD task (Szumlanski and Gomez, 2010). In contrast, the network we present here has relatedness data for over 4,500 polysemous noun targets and 3,000 monosemous noun targets, each of which are related to an average of 27.5 distinct noun senses. It consists of 208,832 undirected edges ? a 181% increase in size over the previous network. The result is a semantic network that has reached maturity and, as we will show, can be successfully applied to a WSD task. This paper proceeds as follows. In the next section (Section 2), we discuss related work. We then give an overview of the method we use to construct our network (Sections 3 and 4). The network is evaluated through its application to a WSD task (Sections 5?7), where we compare its performance to WordNet and another automatically acquired semantic network called WordNet++ (Ponzetto and Navigli, 2010). A discussion follows (Section 8), 190 and we present our conclusions in Section 9. 2 Related Work Our work bears strong relation to WordNet++ (henceforth WN++), which is constructed automatically from the semantic annotations in Wikipedia (Ponzetto and Navigli, 2010).2 Links in WN++ are established between words whose articles link to one another. For example, the article on astronomy in Wikipedia links to the article on celestial navigation, so we find an edge from astronomy#n#1 to celestial navigation#n#1 in WN++.3 The nouns related in WN++ are disambiguated automatically using further semantic annotation data from Wikipedia, including sense labels, the titles of other pages linked to by any two related nouns, and the folksonomic categories to which articles belong. These serve as context words that are compared with context words from various WordNet relations in order to map the nouns to their appropriate WordNet senses. The resulting resource contains 1,902,859 unique edges between noun senses. Augmenting the structure of Wikipedia itself has been the subject of research as well, and involves the discovery of relations between articles. Mihalcea and Csomai (2007), for example, added links between Wikipedia pages after automatically identifying keywords in each article and disambiguating those words to theirappropriate Wikipedia concepts (article titles), while Ponzetto and Navigli (2009) used graph theoretic approaches to augment the taxonomic organization of Wikipedia articles. In terms of automatically discovering semantic relations, many pattern-based approaches have been used to extract specific types of relations from large corpora, e.g., hyponymy, meronymy, and synonymy (Hearst, 1992; Pantel and Pennacchiotti, 2006). Approaches based on distributional similarity have been applied toward the same end (Harris, 1985; Gorman and Curran, 2006), and there are several approaches that rely on the underlying structure of WordNet or Wikipedia to measure the relatedness between two concepts or nouns quantitatively (Hughes and Ramage, 2007; Gabrilovich and 2http://lcl.uniroma1.it/wordnetplusplus 3The notation astronomy#n#1 refers to sense 1 (#1) of the noun (#n) ?astronomy? in WordNet. Other parts of speech are denoted by #v (verbs), #a (adjectives), or #r (adverbs). Markovitch, 2007; Zaragoza et al, 2007; Patwardhan and Pedersen, 2006; Strube and Ponzetto, 2006; Budanitsky and Hirst, 2006; Resnik, 1995). Other quantitative approaches have leveraged the large amounts of data available on the Web to discover relatedness. Notably, Agirre and de Lacalle (2004) employed web queries to associate WordNet synsets with representative context words, known as topic signatures. Cuadros and Rigau (2008) have used these data to construct four KnowNets, semantic knowledge bases derived by disambiguating the top 5, 10, 15, and 20 nouns, respectively, from the topic signatures of Agirre and de Lacalle. 3 Automatic Acquisition of the Semantic Network The semantic network is automatically acquired in three distinct stages (Szumlanski and Gomez, 2010): (1) quantitative measurement of relatedness between nouns that co-occur in a large corpus; (2) categorical determination of whether the quantitative measure indicates strong and mutual semantic relatedness between a given pair of nouns; and (3) unsupervised disambiguation of all the nouns that are found to be semantically related. We provide an overview of each of these steps below (Sections 3.1?3.3), and then discuss how we have expanded this methodology to create a more complete semantic network (Section 4). 3.1 Quantitatively measuring relatedness from lexical co-occurrence We first measure the semantic relatedness, or relational strength, of a target, t, to one of its cooccurring nouns, or co-targets, c, with the following asymmetric function: Srel(t, c) = P (t|c)P (c|t)log P (c|t) P (c) where P (c|t) is the relative frequency of c among all nouns co-occurring with t, and vice versa for P (t|c). P (c) is the relative frequency of c among all nouns occurring in the corpus. For these values, we rely on lexical co-occurrence data extracted from Wikipedia. Co-occurrence is considered intra-sententially (as opposed to co-occurrence in entire articles or paragraphs, or co-occurrence within variable-sized windows of context). 191 This function essentially measures the degree to which an occurrence of t in a sentence predicts the co-occurrence of c. It is an adaptation of Resnik?s (1999) selectional association measure. Table 1 shows the results of applying this function to the co-targets of ?yoga? and ?meditation.? Target (t): yoga Target (t): meditation Co-target (c) Srel Co-target (c) Srel hatha yoga .1801 yoga .0707 asana .0761 mindfulness .0415 meditation .0673 contemplation .0165 bhakti.0508 prayer .0139 raja .0410 practice .0068 tantra .0148 technique .0060 yogi .0132 mantra .0053 karma .0125 relaxation .0048 posture .0104 retreat .0047 aerobics .0093 enlightenment .0031 tai chi .0089 monk .0025 exercise .0036 posture .0024 practice .0032 breathing .0017 instructor .0031 - - - - - - - - - - - - - - - - - - - - - - - - - exercise .0015 guru .0027 teaching .0014 massage .0026 practitioner .0014 exercise .0019 ascetic .0014 ... ... ... ... Table 1: The most strongly related co-targets of ?yoga? and ?meditation,? sorted by decreasing value of relational strength (Srel). Nouns above dashed lines are the top 5% of the target?s most strongly related co-targets. 3.2 Establishing categorical relatedness We then use a mutual relatedness algorithm to ascertain whether two nouns are semantically related by determining whether the nouns under consideration reciprocate a high degree of relatedness to one another. It proceeds as follows: For some target noun of interest, t, let Cx(t) be the set of the top x% of t?s co-targets as sorted by Srel(t, c). For each c ? Cx(t), if we have t ? Cx(c), then we say that t and c are categorically related and add the noun pair (t, c) to our semantic network. We then increment x by one and repeat the process: for every c ? Cx+1(t) such that (t, c) is not already in our network, we look for t in Cx+1(c), and add (t, c) to our network if we find it. This process continues until we have incremented x some number of times without adding any new relations to the semantic network. We then take the symmetric closure of the network, so that if (t, c) is in the network, (c, t) is, as well. (That is, the relation is considered undirected.) Consider, for example, the nouns in Table 1. Given the target ?yoga,? we might first examine the top 5% of its most strongly related co-targets (an arbitrary initial threshold chosen simply for illustrative purposes). In this case, we have all the nouns above the dashed line: C5(yoga) = {hatha yoga, asana, meditation, bhakti, raja, tantra, yogi, karma, posture, aerobics, tai chi, exercise, practice, instructor}. The algorithm then searches C5(hatha yoga), C5(asana), and so on, for ?yoga,? adding a new relation to the network every time ?yoga? is found. Thus, we can see by the inclusion of ?yoga? in C5(meditation) (all nouns above the dashed line in the second column of Table 1), that the pair (yoga, meditation) will be included in the network. This reliance on mutual relatedness ensures that only noun pairs exhibiting strong semantic relatedness are admitted to the network. 3.3 Disambiguation Disambiguation of the resulting noun-noun pairs is the product of majority-rules voting by the following three algorithms. Subsumption. The most frequently occurring immediate hypernyms of all nouns related to our target are permitted to disambiguate the polysemous nouns. This is useful because of the semantic clustering that tends to occur among related nouns. (E.g., ?astronomer? is related to several terms categorizedas celestial bodies in WordNet, such as ?planet,? ?star,? ?minor planet,? and ?quasar.?) Glosses. Senses of polysemous co-targets with occurrences of monosemous co-targets in their glosses are preferentially taken as the intended meanings of the polysemous nouns. Monosemous co-targets are matched directly, or by suffix replacement. (E.g., ?biology? can be matched by the occurrence of ?biologist? in a gloss, ?engineering? by ?engineers,? and so on.) Selectional Preferences. This method associates a numerical score with all superordinate synsets from the WordNet noun ontology that categorize 192 the monosemous nouns related to a target. For example, the noun ?unicorn? strongly predicts related nouns categorized as monsters (monster#1)4 and mythical beings (mythical being#1) in WordNet. These selectional preferences are applied to polysemous co-targets in decreasing order of their relational strength to the target noun. A polysemous noun is disambiguated to the first sense or senses subsumed by one of these selectional preferences. For example, ?phoenix,? as it relates to ?unicorn,? is disambiguated to phoenix#3 in WordNet (the fiery bird that is reborn from its own ashes) by virtue of its subsumption by mythical being#1. 4 Creating a More Complete Network A shortcoming of our previously released network is that it lacked concept-level relations between pairs of polysemous nouns. When humans encounter a pair of ambiguous but closely related words, like bus?horn, we automatically disambiguate to the automobile and the car horn, as opposed to a computer?s front-side bus or a rhinoceros?s horn. The human ability to perform this disambiguation stems from the fact that human semantic memory relates not just individual words, but specific concepts denoted by those words. But if our goal is to establish such a link in our computational model of semantic relatedness, then we cannot rely on the link to perform that disambiguation for us; another approach is called for. One reasonable approach (the one taken in our previous work) is to go where the problem no longer exists ? to relationships that involve at least one monosemous noun. Monosemous-tomonosemous noun relationships require no disambiguation. Monosemous-to-polysemous noun relationships, on the other hand, require that only one noun be disambiguated. This ameliorates our problem tremendously, because the monosemous noun in the pair anchors the polysemous noun in an unambiguous context where disambiguation can more readily take place. That context includes all the nouns related to our monosemous noun, which, through their transitive relatedness to the polysemous noun in question, can assist in the act of disam4We sometimes drop the part of speech from our word sense notation for brevity, but only in the case of noun senses. biguation vis-a`-vis the algorithms described in Section 3.3. Consider, in contrast, the polysemous ?batter,? which can refer to the baseball player or the cake batter. The algorithm for discovering semantic relatedness yields several nouns related to each of these senses of ?batter? (see Table 2). If we wish to disambiguate the pair (batter, cake), we are left with the question: which of the nouns in Table 2 should we take as contextual anchors for the disambiguation? baking fastball inning strike ball flour outfielder strikeout baseglove pancake swing baseball hitter pitch tempura bat home plate pudding umpire cake home run runner waffle dugout infielder shortstop Table 2: An excerpt of some of the nouns related to ?batter? by the algorithm for automatic acquisition. In considering this question, it is important to note that although the ontological categories that subsume the nouns related to ?batter? exhibit greater entropy than we usually observe among the terms related to a monosemous noun, clear delineations still exist. For example, Figure 1 shows the clusters that form as we consider shared hypernymic relationships between all senses of the nouns related to ?batter? (gray nodes in the graph). We see that many of the nouns related to ?batter? have senses categorized by food#1, cake#3, pitch#2, ballplayer#1, or equipment#1 ? the heads of five distinct clusters by semantic similarity. It is worth noting that some nouns related to ?batter? (such as ?baking,? ?swing,? and ?umpire?) do not fall into any of these semantic clusters. In these cases, the WordNet glosses serve as our primary tool for disambiguation. (For example, the glosses of both swing#8 and umpire#1 include mention of ?baseball,? which is also related to ?batter.?) Conversely, some of the polysemous nouns in our example have senses that join semantic clusters unintendedly. For instance, cake#2 (?[a] small flat mass of chopped food,? according to WordNet) falls under the cluster headed by food#1. Although this is potentially problematic, cake#2 is discarded in this particular case in favor of cake#3 (the baked good), 193 which has a greater mass because of its subsumption of waffle#1 and pancake#1, and is indeed the intended meaning of ?cake? as it relates to ?batter.? Another example of unintended cluster membership comes from bat#4 (the cricket bat), which is categorized by sports equipment#1. In contrast, the baseball bat does not have its own entry in WordNet, and the most reasonable sense choice, bat#5 (?a club used for hitting a ball in various games?), is categorized as a stick (stick#1), and not as equipment, sports equipment, or game equipment. These unintended cluster memberships are bound to cause minor errors in our disambiguation efforts. However, our analysis reveals that we do not find such high entropy among the relatives of a polysemous noun that the semantic clustering effect (which is necessary for the success of the disambiguation algorithms described above in Section 3.3) is diminished. Thus, to construct our network, we apply the disambiguation algorithms described above, with the following modification: when confronted with a pair of semantically related polysemous nouns, we apply the disambiguation mechanism described above in both directions, and then fuse the results together. So, in one direction, the various baked goods related to ?batter? help us to properly disambiguate ?cake? to cake#3 in WordNet, yielding the pair (batter, cake#3). A similar scenario yields the pair (cake, batter#2) when disambiguating in the other direction, and we fuse the results together into the properly disambiguated pair (batter#2, cake#3). Using this method, we have automatically created a semantic network that has 208,832 pairs of related noun senses ? the most extensive semanticnetwork between WordNet noun senses to be derived automatically from a simple lexical co-occurrence measure. For the remainder of this paper, we will refer to our network as the Szumlanski-Gomez network (SGN). 5 Coarse-Grained WSD Experiments To evaluate our semantic network, and to provide fair comparison to related work, we take our cue from Ponzetto and Navigli (2010), who evaluated the performance of WN++ on the SemEval-2007 (Navigli et al, 2007) coarse-grained all-words WSD task using extended gloss overlaps (Banerjee and entity#1 food#1 cake#2 equipment#1 ballplayer#1 pitch#2 cake#3 dessert#1 tempura#1 game_equipment#1 sports_equipment#1 runner#4 fielder#1 hitter#1 fastball#1 strike#5 waffle#1 pancake#1 pudding#2 pudding#3 ball#1 infielder#1 outfielder#1 baseball#2 bat#4 base#3 glove#1 glove#3 shortstop#1 centerfielder#1 Figure 1: A partial view of the WordNet graph, showing senses of nouns related to ?batter? (gray nodes) and intermediary concepts (white nodes) that connect them to the root of the taxonomy through hypernymic relationships. Pedersen, 2003) and the graph-based degree centrality algorithm (Navigli and Lapata, 2010). In this particular SemEval task, we are presented with 237 sentences in which lemmatized target words have been flagged for disambiguation. In our experiments, we disambiguate nouns only (as did Ponzetto and Navigli), since both SGN (our network) and WN++ relate only concepts denoted by nouns, and no other parts of speech. In our experimental setup, each sentence is considered in isolation from the rest, and all lemmatized content words in a sentence are provided to the disambiguation algorithms; the verbs, adjectives, and adverbs, although we do not resolve their senses, lend additional context to the disambiguation algorithms. The coarse-grained nature of the SemEval-2007 task provides that there may be more than one acceptable sense assignment for many of the targets. In the coarse-grained setting, an algorithm?s sense assignment is considered correct when it appears in the list of acceptable senses for the given target word. The algorithms below both allow for multiple disambiguation results to be returned in the event of a tie. In these cases (although they are rare), we adopt the approach of Banerjee and Pedersen (2003), who award partial credit and discredit proportionally for all the senses returned by the algorithm. 194 6 Extended Gloss Overlaps (ExtLesk) The first disambiguation algorithm we employ is the extended gloss overlaps measure (henceforth ExtLesk) of Banerjee and Pedersen (2003), which is an extension of the Lesk (1986) gloss overlap measure. Loosely speaking, the algorithm disambiguates a target noun by maximizing the overlap (number of words in common) between the glosses of word senses related5 to the target?s noun senses and those related to all context words (all lemmatized targets in the sentence under consideration other than the target itself). The sense with the greatest overlap is selected as the intended meaning of a target noun. In the event of a tie, multiple senses may be selected. ExtLesk does not attempt to perform sense assignment if the score for every sense of a target noun is zero, except when dealing with a monosemous noun, in which case we default to the only sense possible. 6.1 Results We have run ExtLesk on the SemEval-2007 taskusing five combinations of semantic resources: WordNet only, SGN (our semantic network) only, SGN and WordNet combined (that is, the union of all links contained in both networks), WN++ only, and WN++ combined withWordNet. We include the traditional baselines of most frequent sense (MFS) assignment and random sense assignment for comparison, and measure precision (number of correct sense assignments divided by the number of attempted sense assignments), recall (number of correct sense assignments divided by the number of target nouns to be disambiguated), and the harmonic mean of the two, F1, defined as: F1 = 2 ? precision ? recall precision + recall We present our results in Table 3, and offer the following observations. Firstly, SGN as a standalone network rivals the performance of WordNet. This is particularly impressive given the fact that 5We use all relations available in WordNet, as well as a related-to relation derived from the links in our semantic network. Resource P R F1 WordNet 78.80 74.82 76.76 SGN 78.64 72.82 75.62 SGN and WordNet 82.35 78.11 80.18 WN++ 74.67 61.87 67.67 WN++ and WordNet 77.35 73.38 75.31 MFS Baseline 77.40 77.40 77.40 Random Baseline 63.50 63.50 63.50 Table 3: ExtLesk disambiguation results on the SemEval2007 all-words coarse-grained WSD task (nouns only). the edges in SGN were derived automatically from a simple lexical co-occurrence measure. Equally impressive is the ability of SGN and WordNet, when used in combination, to achieve results that exceed what either network is able to accomplish as a stand-alone knowledge source. When combined, we see improvements of 3.42% and 4.56% over WordNet and SGN as stand-alone resources, respectively. It is also only with these resources combined that we are able to outperform the MFS baseline, and we do so by 2.78%.6 In contrast, WN++ fails to perform as a standalone resource, falling behind the MFS baseline by 9.73%.7 Of all the resources tested, WN++ yields the lowest results. When combined with WordNet, WN++ actually diminishes the ability of WordNet to perform on this WSD task by 1.45%. We defer our discussion of factors impacting the performance of WN++ to Section 8 (Discussion). 7 WSD with Degree Centrality Degree centrality is a graph-based measure of semantic relatedness (Navigli and Lapata, 2010) in which we search through a semantic network for paths of length l ? maxLength between all sense nodes for all lemmas in our context. The edges along all such paths are added to a new graph, G?, and for each target noun to be disambiguated, the sense node with the greatest number of incident edges (highest vertex degree) in G? is taken as its intended sense. 6Other systems have obtained better results on the same dataset, but we focus only on SGN and WN++ because our aim is to compare the resources themselves. 7Ponzetto and Navigli (2010) report results of F1 = 68.3 and 72.0 for WN and WN++ as stand-alone resources. Space considerations prevent us from discussing this disparity in detail. 195 In these graphs, nodes represent synsets, as opposed to instantiating separate nodes for different members of the same synsetand allowing edges to be constructed between them. We include all lemmas from a sentence in our context, but only return disambiguation results for the nouns. With SGN and WN++, the implementation of this algorithm is straightforward. We initiate a breadthfirst search (BFS)8 at each target sense node in the network, and proceed through ?maxLength+12 ? iterations of spreading activation. Whenever the tendrils of this spreading activation from one target sense node in the graph connect to those of another,9 we add the path between the nodes to our new graph, G?, potentially incrementing the degree of the involved target sense nodes in G? as we do so. Because BFS is an admissible algorithm (guaranteed to find the shortest path from an initial state to a goal), it provides a computationally efficient approach to finding all paths between all target nodes. Also, because any node on a path of length l ? maxLength between two target nodes is at most ? l2? nodes removed from at least one of those target sense nodes, we only need to perform a BFS of depth ?maxLength+12 ? from every target sense node in order to guarantee that every such path between them will be discovered. Since the time complexity of BFS is exponential with respect to the depth of the search, cutting this depth in half (in comparison to performing a BFS of depth maxLength) greatly reduces the running time of our algorithm. We take the same approach in traversing the WordNet noun graph, using all possible sense relations as edges. In keeping with the approach of Navigli and Lapata (2010), an edge is also induced between synsets if the gloss of one synset contains a monosemous content word. For example, the gloss for leprechaun#n#1, ?a mischievous elf in Irish folklore,? contains the monosemous noun ?folklore;? thus, we have an edge between leprechaun#n#1 and 8This is in contrast to the DFS implementation of Navigli and Lapata (2010), so for the sake of posterity, we expound upon our approach in this section. 9When maxLength is odd, this requires an additional check to ensure that the intersection is not taking place at a node that is exactly ?maxLength+12 ? degrees removed from each of the two target nodes it is connecting, as this would result in a path with overall length maxLength + 1 between the target nodes. folklore#n#1 in the WordNet graph. Once we have our new graph, G?, constructed in this manner, the vertex degree is considered an indication of the semantic relatedness of a particular synset to all other lemmas in our context. For each target noun, we use its sense node with the highest degree in G? for sense assignment. 7.1 Results We have tested the degree centrality algorithm with the following combinations of semantic resources: WordNet, SGN, WN++, Refined WN++, SGN and WordNet combined, and Refined WN++ and WordNet combined. (Refined WN++ consists of 79,422 of WN++?s strongest relations, and was created in an unsupervised setting by Ponzetto and Navigli specifically for use with degree centrality when they discovered that WN++ hadtoo many weak relations to perform well with the algorithm.) We have observed that the performance of degree centrality rapidly levels off as maxLength increases. Ponzetto and Lapata (2010) also report this so-called ?plateau? effect, and employ a maxLength of 6 in their experiments, despite finding that results level off around maxLength = 4. We, too, find that performance levels off around maxLength = 4 in almost all cases, and so only continue up to maxLength = 5. We find that, in all cases tested, degree centrality is unable to outperform the MFS baseline (with respect to F1) (see Table 4). SGN and WN++ exhibit comparable performance with this algorithm, with maximum F1 values of 68.4% (maxLength = 2) and 67.3% (maxLength = 3?5), respectively. Neither achieves the performance of WordNet with degree centrality (F1 = 74.0%), which underperforms the MFS baseline (F1 = 77.4%) by 3.4%.10 Ponzetto and Navigli (2010) reported that only performing sense assignment when the max degree exceeded an empirically derived but non-disclosed threshold improved performance, but we have found that implementing such a threshold universally lowers results for all resources we tested with degree centrality. 10Although Ponzetto and Navigli (2010) reported similar results with WordNet (F1 = 74.5), we have been unable to reproduce their results using Refined WN++, either combined with WordNet (F1 = 79.4) or as a stand-alone resource (F1 = 57.4). 196 The lowest performance using degree centrality comes from Refined WN++ as a stand-alone resource. We attribute this to the fact that Refined WN++ is so semantically sparse. On average, noun senses in Refined WN++ are related to only 3.42 other noun senses, while those in WN++ and SGN relate to an average of 44.59 and 10.92 noun senses, respectively. Accordingly, the success of Refined WN++ and WordNet combined is attributable mostly to the success of WordNet as a stand-alone resource; as maxLength increases, the contributions made by the sparse Refined WN++ network rapidly become negligible in comparison to those provided by the WordNet ontology. l P R F1 P R F1 WordNet SGN 1 96.9 16.8 28.6 79.7 32.9 46.6 2 77.6 45.1 57.0 72.0 64.6 68.4 3 76.7 65.6 70.7 68.7 63.5 66.0 4 76.9 71.0 73.9 68.0 63.9 65.9 5 76.6 71.6 74.0 68.0 64.2 66.1 SGN & WN WN++ 1 77.4 52.4 62.5 87.2 23.5 37.1 2 74.7 70.7 72.7 71.6 60.2 65.4 3 70.3 67.1 68.7 70.7 64.3 67.3 4 70.5 67.4 68.9 70.4 64.5 67.3 5 70.1 67.0 68.5 70.4 64.5 67.3 WN++refined WN ++ refined& WN 1 98.3 15.3 26.5 83.3 31.2 45.4 2 91.4 23.4 37.3 77.5 66.6 71.6 3 88.7 29.9 44.7 77.6 73.6 75.5 4 83.7 32.3 46.7 74.7 71.4 73.0 5 80.2 35.3 49.0 74.7 71.4 73.0 MFS Baseline Random Baseline 77.4 77.4 77.4 63.5 63.5 63.5 Table 4: Degree centrality disambiguation results on the SemEval-2007 all-words coarse-grained WSD task (nouns only). l is maximum path length. 8 Discussion The fact that the performance of degree centrality quickly plateaus hints at the root cause of its weak performance compared to ExtLesk andthe MFS baseline. As the maximum path length is increased in a dense semantic network, all possible edges from our target sense nodes rapidly find themselves involved with paths to other target sense nodes. This is particularly true of WN++ (notice its rapid and stable convergence), where certain ?sticky? nodes form bridges between seemingly unrelated concepts. For example, the frequent appearance of ?United States? in Wikipedia articles, and its tendency to be linked to the United States Wikipage when it occurs, causes the term to serve as a bridge between such diverse concepts as automaton#2 and burrito#1, which one would typically expect to be far removed from one another in a model of semantic relatedness. Nonetheless, the degree centrality algorithm has no difficulty finding short paths between target sense nodes when traversing any of the semantic networks we tested. In fact, we have discovered that as the results of degree centrality converge, they approach the performance obtained by foregoing the algorithm altogether and simply disambiguating each noun to the sense with the most edges in the network (regardless of whether those edges ultimately connect two word senses from the disambiguation context). The expected values of convergence attained by defaulting to the most semantically wellconnected sense of each target noun in each network are F1 = 66.3%, 67.5%, and 74.6% for SGN,WN++, and WordNet, respectively ? remarkably close to the experimentally derived degree centrality results of F1 = 66.1%, 67.3%, and 74.0%.9 Conclusion We have constructed a semantic network of related noun senses automatically from intra-sentential lexical co-occurrence data, and shown that on a WSD task, it outperforms a similar resource, WN++, which is derived from the rich set of semantic annotations available in the Wikipedia corpus. Our network has also shown competitive performance with the WordNet ontology onWSD, and when combined with WordNet, improves disambiguation results in a coarse-grained setting using the ExtLesk disambiguation algorithm. Acknowledgments This research was supported in part by the NASA Engineering and Safety Center under Grant/Cooperative Agreement NNX08AJ98A. 197'
1.0	P12-2040	A00-1010	b'1 Introduction  Data driven applications have proliferated in Computational Linguistics during the last decade. Several factors, such as the availability of more powerful computers, an almost unlimited storage capacity, the availability of large volumes of data in  digital format, as well as the recent advances in  machine learning theory, have significantly contributed to such a proliferation.    Among the many applications that have benefited from this data-driven boom, probably the most  representative examples are: information retrieval  (Qin et al, 2008), machine translation (Brown et  al., 1993), question answering (Molla-Aliod and  Vicedo, 2010) and dialogue systems (Rieser and  Lemon, 2011).  In the specific case of dialogue systems, data  acquisition can impose some challenges depending  on the specific domain and task the dialogue system is targeted for. In some specific domains, in  which human-human dialogue applications already  exists, data collection is generally straight forward,  while in some other cases, data design and collection can constitute a complex problem (Williams  and Young, 2003; Zue, 2007; Misu et al, 2009).   Depending on the objective being pursued, dialogue systems can be grouped into two major categories: task-oriented and chat-oriented systems. In  the first case, the system is required to help the  user to accomplish a specific goal or objective  (Busemann et al, 1997; Stallard, 2000). In the second case, the system objective is mainly entertainment oriented. Systems in this category are required to play, chitchat or just accompany the user  (Weizenbaum, 1966; Wallis, 2010).   In this work, we focus our attention on dialogue  data which is suitable for training chat-oriented  dialogue systems. Different from task-oriented dialogue collections (Mann, 2003), instead of being  concentrated on a specific domain or area of  knowledge, the training dataset for a chat-oriented  dialogue system must cover a wide variety of domains, as well as be able to provide a fair representation of world-knowledge semantics and pragmatics (Bunt, 2000). To this end, we have collected dialogues from movie scripts aiming at  constructing a dialogue corpus which should provide a good sample of domains, styles and world  knowledge, as well as constitute a valuable resource for research and development purposes.  The rest of the paper is structured as follows.  Section 2 describes in detail the implemented collection process and the structure of the generated  database. Section 3 presents the main statistics, as  well as the main characteristics of the resulting corpus. Finally, section 4 presents our conclusions  and future work plans.   203 2 Collecting Dialogues from Movies   As already stated in the introduction, our presented  dialogue corpus has been extracted from movie  scripts. More specifically, scripts freely available  from The Internet Movie Script Data Collection  (http://www.imsdb.com/) have been used. In this  section we describe the implemented data collection process and the data structure finally used for  the generated corpus.  As a first step of the collection construction,  dialogues have to be identified and extracted from  the crawled html files. Three basic types of information elements are extracted from the scripts:  speakers, utterances and context.   The utterance and speaker information elements  contain what is said at each dialogue turn and the  corresponding character who says it, respectively.  Context information elements, on the other hand,  contain all additional information/texts appearing  in the scripts, which are typically of narrative  nature and explain what is happening in the scene.   Figure 1 depicts a browser snapshot illustrating  the typical layout of a movie script and the most  common spatial distribution of the aforementioned  information elements.  It is important to mention that a lot of different  variants to the format presented in Figure 1 can be  actually encountered in The Internet Movie Script  Data Collection. Because of this, our parsing algorithms had to be revised and adjusted several  times in order to achieve a reasonable level of  robustness that allowed for processing the largest  possible amount of movie scripts.   Another important problem was the identification of dialogue boundaries. Some heuristics were  implemented by taking into account the size and  number of context elements between speaker turns.   A post-processing step was also implemented to  either filter out or amend some of the most common parsing errors occurring during the extraction  phase. Some of these errors include: corrupted formats, turn continuations, notes inserted within the  turn, misspelling of speaker names, etc.  In addition to this, a semi-automatic process was  still necessary to filter out movie scripts exhibiting  extremely different layouts or invalid file formats.  Approximately, 17% of the movie scripts crawled  from The Internet Movie Script Data Collection  had to be discarded. From a total of 911 crawled  scripts, only 753 were successfully processed.       Figure 1: Typical layout of a movie script    The extracted information was finally organized  in dialogical units, in which the information regarding turn sequences inside each dialogue, as well  as dialogue sequences within each movie script  was preserved. Figure 2 illustrates an example of  the XML representation for one of the dialogues  extracted from Who Framed Roger Rabbit.     <dialogue id="47" n_utterances="4">     <speaker>VALIANT</speaker>        <context></context>       <utterance>You shot Roger.</utterance>     <speaker>JESSICA RABBIT</speaker>        <context>Jessica moves the box aside and  tugs on the rabbit ears. The rabbit head pops  off. Underneath is a Weasle. In his hand is the  Colt .45 Buntline.</context>        <utterance>That\'s not Roger. It\'s one of  Doom\'s men. He killed R.K. Maroon.</utterance>     <speaker>VALIANT</speaker>        <context></context>        <utterance>Lady, I guess I had you pegged  wrong.</utterance>     <speaker>JESSICA RABBIT</speaker>        <context>As they run down the  alley...</context>        <utterance>Don\'t worry, you\'re not the  first. We better get out of here.</utterance>  </dialogue>    Figure 2: An example of a dialogue unit   204 3 Movie Dialogue Corpus Statistics  In this section we present the main statistics of the  resulting dialogue corpus and study some of its  more important properties. The final dialogue collection was the result of successfully processing  753 movie scripts. Table 1 summarizes the main  statistics of the resulting dialogue collection.     Total number of scripts collected 911  Total number of scripts processed 753  Total number of  dialogues 132,229  Total number of  speaker turns 764,146  Average amount of dialogues per movie 175.60  Average amount of turns per movie 1,014.80  Average amount of turns per dialogue 5.78    Table 1: Main statistics of the collected movie  dialogue dataset    Movies were mainly crawled from the action,  crime, drama and thriller genres. However, as each  movie commonly belongs to more than one single  genre, much more genres are actually represented  in the dataset. Table 2 summarizes the distribution  of movies by genre (notice that, as most of the  movies belong to more than one genre, the total  summation of percentages exceeds 100%).    Genre Movies Percentage Action 258 34.26  Adventure 133 17.66  Animation 22 2.92  Comedy 149 19.79  Crime 163 21.65  Drama 456 60.56  Family 31 4.12  Fantasy 82 10.89  Horror 104 13.81  Musical 18 2.39  Mystery 95 12.62  Romance 123 16.33  Sci-Fi 129 17.13  Thriller 329 43.69  War 25 3.32  Western 11 1.46    Table 2: Distribution of movies per genre    The first characteristic of the corpus to be analyzed is the distribution of dialogues per movie.  This distribution is shown in Figure 3. As seen  from the figure, the distribution of dialogues per  movie is clearly symmetric around its mean value  of 175 dialogues per movie. For most of the movies in the collection, a number of dialogues ranging from about 100 to 250 were extracted.         Figure 3: Distribution of dialogues per movie     The second property of the corpus to be studied  is the distribution of turns per dialogue. This distribution is shown in Figure 4. As seen from the  figure, this distribution approximates a power law  behavior, with a large number of very short dialogues (about 50K two-turn dialogues) and a small  amount of long dialogues (only six dialogues with  more than 200 turns). The median of the distribution is 5.63 turns per dialogue.         Figure 4: Distribution of turns per dialogue     The third property of the corpus to be described  is the distribution of number of speakers per dia205 logue. This distribution is shown in Figure 5. As  seen from the bar-plot depicted in the figure, the  largest proportion of dialogues (around 60K) involves two speakers. The second largest proportion  of ?dialogues? (about 35K) involves only a single  speaker, which means that this subset of the data  collection is actually composed by monologues or  single speaker interventions. The third and fourth  larger proportions are those involving three and  four speakers, respectively.          Figure 5: Distribution of number of speakers per  dialogue    Finally, in Figure 6, we present a cross-plot between the number of dialogues and the number of  turns within each movie script.       Figure 6: Cross-plot between the number of  dialogues and turns within each movie script  As seen from the cross-plot, an average movie  has between 150 and 200 dialogues comprising  between 1000 and 1200 turns in total. The crossplot also reveals some interesting extreme cases in  the data collection.   For instance, movies with few dialogues but many turns are located towards the upper-left corner  of the figure. In this zone we can find movies as:  Happy Birthday Wanda June, Hannah and Her  Sisters and All About Eve. In the lower-left corner  of the figure we can find movies with few dialogues and few turns, as for instance: 1492 Conquest of Paradise and The Cooler.   In the right side of the figure we find the lots-ofdialogues region. There we can find movies with  lots of very short dialogues (lower-right corner),  such as Jimmy and Judy and Walking Tall; or movies with lots of dialogues and turns (upper-right  corner), such as The Curious Case of Benjamin  Button and Jennifer?s Body.  4 Conclusions and Future Work   In this paper, we have described Movie-DiC a  Movie Dialogue Corpus that has been collected for  research and development purposes. The data collection comprises 132,229 dialogues containing a  total of 764,146 turns/utterances that have been  extracted from 753 movies. Details on how the  data collection has been created and how the  corpus is structured wereprovided along with the  main statistics and characteristics of the corpus.  Although strictly speaking, and by its particular  nature, Movie-DiC does not constitute a corpus of  real human-to-human dialogues, it does constitute  an excellent dataset for studying the semantic and  pragmatic aspects of human communication within  a wide variety of contexts, scenarios, styles and  socio-cultural settings.   Specific technologies and applications that can  exploit a resource like this include, but are not restricted to: example-based chat bots (Banchs and Li,  2012), question answering systems, discourse and  pragmatics analysis, narrative vs. colloquial style  classification, genre classification, etc.   As future work, we intend to expand the current  size of the collection from 0.7K to 2K movies, as  well as to improve some of our parsing and postprocessing algorithms for reducing the amount of  noise still present in the collection and enhance the  quality of the current version of the dataset.  206 Acknowledgments  The author would like to thank the Institute for  Infocomm Research for its support and permission  to publish this work. '	b'Introduction  This paper describes Talk\'n\'Travel, a spoken  language dialogue system for making complex  air travel plans over the telephone.  Talk\'n\'Travel is a research prototype system  sponsored under the DARPA Communicator  program (MITRE, 1999). Some other systems  in the program are Ward and Pellom (1999),  Seneff and Polifroni (2000) and Rudnicky et al (1999). The common task of this program is a  mixed-initiative dialogue over the telephone, in  which the user plans a multi-city trip by air,  including all flights, hotels, and rental cars, all in  conversational English over the telephone.  The Communicator common task presents  special challenges. It is a complex task with  many subtasks, including the booking of each  flight, hotel, and car reservation. Because the  number of legs of the trip may be arbitrary, the  number of such subtasks is not known in  advance. Furthermore, the user has complete  freedom to say anything at any time. His  utterances can affect just the current subtask, or  multiple subtasks at once ("I want to go from  Denver to Chicago and then to San Diego"). He  can go back and change the specifications for  completed subtasks. And there are important  constraints, such as temporal relationships  between flights, that must be maintained for the  solution to the whole task to be coherent.  In order to meet this challenge, we have sought  to develop dialogue techniques for  Talk\'n\'Travel that go beyond the rigid system-  directed style of familiar IVR systems.  Talk\'n\'Travel is instead a mixed initiative  system that allows the user to specify constraints  on his travel plan in arbitrary order. At any  point in the dialogue, the user can supply  information other than what the system is  currently prompting for, change his mind about  information he has previously given and even  ask questions himself. The system also tries to  be helpful, eliciting constraints from the user  when necessary. Furthermore, if at any point the  constraints the user has specified cannot all be  met, the system steps in and offers a relaxation  of them in an attempt o negotiate a partial  solution with the user.  The next section gives a brief overview of the  system. Relevant components are discussed in  subsequent sections.  I System Overview  The system consists of the following modules:  speech recognizer, language understander,  dialogue manager, state manager, language  generator, and speech synthesizer. The modules  68  interact with each other via the central hub  module of the Communicator Common  Architecture.  The speech recognizer is the Byblos system  (Nguyen, 1995). It uses an acoustic model  trained from the Macrophone telephone corpus,  and a bigram/trigram language model trained  from -40K utterances derived from various  sources, including data collected under the  previous ATIS program (Dahl et al 1994).  The speechsynthesizer is Lucent\'s commercial  system. Synthesizer and recognizer both  interface to the telephone via Dialogics  telephony board. The database is currently a  frozen snapshot of actual flights between 40  different US cities (we are currently engaged in  interfacing to a commercial air travel website).  The various language components are written in  Java. The complete system runs on Windows  NT, and is compliant with the DARPA  Communicator Common architecture.  The present paper is concerned with the dialogue  and discourse management, language generation  and language understanding components. In the  remainder of the paper, we present more detailed  discussion of these components, beginning with  the language understander in Section 2. Section  3 discusses the discourse and dialogue  components, and Section 4, the language  generator.  2 Language Understanding  2.1 Meaning Representation  Semantic frames have proven useful as a  meaning representation for many applications.  Their simplicity and useful computational  properties have often been seen as more  important han their limitations in expressive  power, especially in simpler domains.  Even in such domains, however, flames still  have some shortcomings. While most naturally  representing equalities between slot and filler,  flames have a harder time with inequalities, uch  as \'the departure time is before 10 AM\', or \'the  airline is not Delta\'. These require the slot-filler  to be some sort of predicate, interval, or set  object, at a cost to simplicity uniformity. Other  problematic ases include n-ary relations (\'3  miles from Denver\'), and disjunctions of  properties on different slots.  In our Talk\'n\'Travel work, we have developed  a meaning representation formalism called path  constraints, which overcomes these problems,  while retaining the computational advantages  that made frames attractive in the first place. A  path constraint is an expression of the form :  (<path> <relation> <arguments>*)  The path is a compositional chain of one or more  attributes, and relations are 1-place or higher  predicates, whose first argument is implicitly the  path. The relation is followed by zero or more  other arguments. In the simplest case, path  constraints can be thought of as flattenings of a  tree of frames. The following represents the  constraint that the departure time of the first leg  of the itinerary is the city Boston :  LEGS.0.ORIG_CITY EQ BoSToN  Because this syntax generalizes to any relation,  however, the constraint "departing before 10  AM" can be represented in a syntactically  equivalent way:  LEGS.0.DEPART_TIME LT 1000  Because the number of arguments i arbitrary, it  is equally straightforward to represent a one-  place property like "x is nonstop" and a three  place predicate like "x is 10 miles from Denver".  Like flames, path constraints have a fixed  format that is indexed in a computationally  useful way, and are simpler than logical forms.  Unlike flames, however, pathconstraints can be  combined in arbitrary conjunctions, disjunctions,  and negations, even across different paths. Path  constraint meaning representations are also flat  lists of constraints rather than trees, making  matching rules, etc, easier to write for them.  69  2.2 The GEM Understanding System  Language understanding in Talk\'n\'Travel is  carried out using a system called GEM (for  Generative Extraction Model). GEM (Miller,  1998) is a probabilistic semantic grammar that is  an outgrowth of the work on the HUM system  (Miller, 1996), but uses hand-specified  knowledge in addition to probability. The hand-  specified knowledge is quite simple, and is  expressed by a two-level semantic dictionary. In  the first level, the entries map alternative word  strings to a single word class. For example, the  following entry maps several alternative forms  to the word class DEPART:  Leave, depart, get out of => DEPART  In the second level, entries map sequences of  word classes to constraints:  Name: DepartCity 1  Head: DEPART  Classes: \\[DEPART FROM CITY\\]  Meaning: (DEST_CITY EQ <CITY>)  The "head" feature allows the entry to pass one  of its constituent word classes up to a higher  level pattern, allowing the given pattern to be a  constituent of others.  The dictionary entries generate a probabilistic  recursive transition network (PRTN), whose  specific structure is determined by dictionary  entries. Paths through this network correspond  one-to-one with parse trees, so that given a path,  there is exactly one corresponding tree. The  probabilities for the arcs in this network can be  estimated from training data using the EM  (Expectation-Maximization) procedure.  GEM also includes a noise state to which  arbitrary input between patterns can be mapped,  making the system quite robust to ill-formed  input. There is no separate phase for handling  ungrammatical input, nor any distinction  between grammatical nd ungrammatical input.  3 Discourse and Dialogue Processing  A key feature of the Communicator task is that  the user can say anything at any time, adding or  changing information at will. He may add new  subtasks (e.g. trip legs) or modifying existing  ones. A conventional dialogue state network  approach would be therefore infeasible, as the  network would be almost unboundedly arge and  complex.  A signifigant additional problem is that changes  need not be monotonic. In particular, when  changing his mind, or correcting the system\'s  misinterpretations, the user may delete subtask  structures altogether, as in the subdialog:  S: What day are you returning to Chicago?  U: No, I don\'t want a return flight.  Because they take information away rather than  add it, scenarios like this one make it  problematic to view discourse processing as  producing a contextualized, or "thick frame",  version of the user\'s utterance. In our system,  therefore, we have chosen a somewhat different  approach. The discourse processor, called the state  manager, computes the most likely new task  state, based on the user\'s input and the current  task state. It also computes a discourse vent,  representing its interpretation f what happened  in the conversation as a result of the user\'s  utterance.  The dialogue manager is a separate module, as  has no state managing responsibilities at all.  Rather, it simply computes the next action to  take, based on its current goal agenda, the  discourse vent returned by the state manager,  and the new state. This design has the advantage  of making the dialogue manager considerably  simpler. The discourse event also becomes  available to convey to the user as confirmation.  We discuss these two modules in more detail  below.  70  3.1 State Manager  The state manager is responsible for computing  and maintaining the current ask state. The task  state is simply the set of path constraints which  currently constrain the user\'s itinerary. Also  included in the task state are the history of user  and system utterances, and the current subtask  and object in focus, if any.  The state manager takes the N-best list of  recognition hypotheses as input. It invokes the  understanding module on a hypothesis to obtain  a semantic interpretation. The semantic  interpretation so obtained is subjected to the  following steps:  1. Resolve ellipses if any  2. Match input meaning to subtask(s)  3. Expand local ambiguities  4. Apply inference and coherency rules  5. Compute database satisfiers  6. Relax constraints if neccesary  7. Determine the most likely alternative and  compute the discourse vent  At any of these steps, zero or more alternative  new states can result, and are fed to the next  step. If zero states result at any step, the new  meaning representation is rejected, and another  one requested from the understander. If no more  hypotheses are available, the entire utterance is rejected, and a DONT_UNDERSTAND event is  returned to the dialogue manager.  Step 1 resolves ellipses. Ellipses include both  short responses like "Boston" and yes/no  responses. In this step, a complete meaning  representation such as \' (ORIQCITY EQ  BOSTON)\' is generated based on the system\'s  prompt and the input meaning. The hypothesis  rejected if this cannot be done.  Step 2 matches the input meaning to one or more  of the subtasks of the problem. For the  Communicator p oblem, the subtasks are legs of  the user\'s itinerary, and matching is done based  on cities mentioned in the input meaning. The  default is the subtask currently in focus in the  dialogue.  A match to a subtask is represented by adding  the prefix for the subtask to the path of the  constraint. For example, "I want to arrive in  Denver by 4 PM" andthen continue on to  Chicago would be :  LEGS.0.DEST_CITY EQ DENVER  LEGS.0.ARRIVE_TIME LE 1600  LEGS. 1.ORIG_CITY EQ DENVER  LEGS. 1.DEST CITY EQ CHICAGO  In Step 3, local ambiguities are expanded into  their different possibilities. These include  partially specified times such as "2 o\'clock"  Step 4 applies inference and coherency rules.  These rules will vary from application to  application. They are written in the path  constraint formalism, augmented with variables  that can range over attributes and other values.  The following is an example, representing the  constraint a flight leg cannot be scheduled to  depart until after the preceding flight arrives:  LEGS.$N.ARRIVE  LT  LEGS. $N+ 1 .DEPART  States that violate coherency constraints are  discarded.  Step 5 computes the set of objects in the  database that satisfy the constraints on the  current subtask. This set will be empty when the  constraints are not all satisfiable, in which case  the relaxation of Step 6 is invoked. This  relaxation is a best-first search for the satisfiable  subset of the constraints that are deemed closest  to what the user originally wanted. Alternative  relaxations are scored according to a sum of  penalty scores for each relaxed constraint,  derived from earlier work by Stallard (1995).  The penalty score is the sum of two terms: one  for the relative importance of the attribute  concerned (e.g. relaxations of DEPART_DATE  are penalised more than relaxations of  AIRLINE) and the other for the nearness of the  satisfiers to the original constraint (relevant for  number-like attributes like departure time).  71  The latter allows the system to give credit to  solutions that are near fits to the user\'s goals,  even if they relax strongly desired constraints.  For example, suppose the user has expressed a  desire to fly on Delta and arrive by 3 PM, while  the system is only able to find a flight on Delta  that arrives at 3:15 PM. In this case, this flight,  which meets one constraint and almost meets the  other, may well satisfy the user more than a  flight on a different airline that happens to meet  the time constraint exactly.  In the final step, the alternative new states are  rank-ordered according to a pragmatic score, and  the highest-scoring alternative is chosen. The  pragmatic score is computed based on a number  of factors, including the plausibility of  disambiguated times and whether or not the state  interpreted the user as responding to the system  prompt.  The appropriate discourse event is then  deterministicaUy computed and returned. There  are several types of discourse vent. The most  common is UPDATE, which specifies the  constraints that have been added, removed, or  relaxed. Another type is REPEAT, which is  generated when the user has simply repeated  constraints the system alreadyknows. Other  types include QUESTION, TIMEOUT, and  DONT UNDERSTAND.  3.1 Dialogue Manager  Upon receiving the new discourse event from  the state manager, the dialogue manager  determines what next action to take. Actions  can be external, such as speaking to the user or  asking him a question, or internal, such as  querying the database or other elements of the  system state. The current action is determined by  consulting a stack-based agenda of goals and  actions.  The agenda stack is in turn determined by an  application-dependent library of plans. Plans are  tree structures whose root is the name of the goal  the plan is designed to solve, and whose leaves  are either other goal names or actions. An  example of a plan is the following:  Completeltinerary =>  (Prompt "How can I help you?")  (forall legs $n  GetRoutelnfo  GetSpecificFlight  GetHotelAndCar  GetNextLeg))  This is a plan for achieving the goal  Completeltinerary. It begins with a open-ended  prompt and then iterates over values of the  variable $N for which constraints on the prefix  LEGS.$N exist, working on high-level subgoals,  such as getting the route and booking a flight,  for each leg. The last goal determines whether  there is another leg to the itinerary, in which  case the itera  The system begins the interaction with the high-  level goal START on its stack. At each step, the  system examines the top of its goal stack and  either executes it if it is an action suitable for  execution, or replaces it on the stack with its  plan steps if it is a goal.  Actions are objects with success and relevancy  predicates and an execute method, somewhat  similar to the "handlers" of Rudnicky and Xu  (1999). An action has an underlying oal, such  as finding out the user\'s constraints on some  attribute. The action\'s success predicate will  return true if this underlying goal has been  achieved, and its relevancy predicate will return  true if it is still relevant to the current situation.  Before carrying out an action, the dialogue  manager first checks to see if its success  predicate returns false and its relevancy  predicate returns true. If either condition is not  met, the action is popped off the stack and  disposed of without being executed. Otherwise,  the action\'s execute method is invoked.  The system includes a set of actions that are  built in, and may be parameterized for each each  domain. For example, the action type ELICIT is  parameterized by an attribute A, a path prefix P,  and verbalization string S. Its success predicate  returns true if the path \'P.A\' is constrained in the  current state. Its execute method generates a meaning frame that is passed to the language  72  generator, ultimately prompting the userwith a  question such as "What city are you flying to?"  Once an action\'s execute method is invoked, it  remains on the stack for the next cycle, where it  is tested again for success and relevancy. In this  case, if the success condition is met - that is, if  the user did indeed reply with a specification of  his destination city - the action is popped off the  stack. If the system did not receive this  information, either because the user made a  stipulation about some different attribute, asked  a question, or simply was not understood, the  action remains on the stack to be executed again.  Of course, the user may have already specified  the destination city in a previous utterance. In  this case, the action is already satisfied, and is  not executed. In this way, the user has  flexibility in how he actually carries out the  dialogue.  In certain situations, other goals and actions may  be pushed onto the stack, temporarily  interrupting the execution of the current plan.  For example, the user himself may ask a  question. In this case, an action to answer the  question is created, and pushed onto the stack.  The dialogue manager then executes this action  to answer the user\'s question before continuing  on with the plan. Or the state manager may  generate a clarification question, which the  dialogue manager seeks to have the user answer.  Actions can also have a set of conditional  branchings that are tested after the action is  executed. If present, these determine the next  action to execute or goal to work on. For  example, the action that asks the user "Do you  want a return flight to X?" specifies the branch  to be taken when the user replies in the negative.  This branch includes an action that asks the user  "Is Y your final destination?", an action that is  executed if the user did not specify an additional  destination along with his negative reply.  Unlike the approach taken by Ward and Pellom  (1999), which seeks to avoid scripting entirely  by driving the dialogue off the current status of  the itinerary, the Talk\'n\'Travel dialogue  manager thus seeks to allow partially scripted  dialogue where appropriate to the situation.  4 Language Generation  The language generator takes a meaning frame  from the dialogue manager, and generates a text  string in English for it. It uses a set of pattern-  based rules that map constraints into alternative  syntactic realisations. For example, the  following rule allows a constraint on departure  time to be realized as "leave at 3 PM"  or "3 PM flight":  LEG.$N.DEPART_TIME EQ $X  =~ \\[leave at $X\\],  \\[nom-comp $X\\]  Different realization rules can be selected for  depending upon whether the constraint is to be realized as an assertion or as a description. The  generation algorithm assembles the selected  realizations for each constraint into a simplified  syntax tree, selecting appropriate inflections of  verb and noun heads as it does so. Terminal  values in constraints are realized as type-specific  nominals, such as "3 PM" or "Delta".  A crucial feature of the generation process is  that it adds to each prompt a paraphrase of the  most recent discourse event, corresponding to  what the system thinks the user just said. This  helps keep the conversation grounded in terms  of mutual understanding between the  participants.  5 Example Scenario  The following is an example dialog with the  system:  S: Welcome to Talk\'n\'Travel. How can I help  you?  U: I need a flight from Boston to Denver  tomorrow.  S: On Thursday, November 18, flying from  Boston to Denver. What time do you want to  leave?  U: I want to leave before 10 AM.  S: Leaving before 10 AM. Do you have an  airline preference?  U: Continental.  73  S: Continental. I \'m sorry, there are no flights on  Continental leaving before 10 AM. I can get  you a flight on Continental leaving at 10:10 AM  instead. Would that be OK?  U: Yes.  S: OK, I\'ve got you on Continental flight 429,  departing Boston at 10:10 AM, arriving at  Denver at 3:44 PM. Do you want to book a  return flight from Denver to Boston?  U: Yes, on Monday.  S: Alright, we\'ll add a flight on Monday,  November 22, from Boston to Denver to your  itinerary. I have a Continental flight at 7 AM.  Is that OK?  U. Anything later?  S: Leaving after 7 AM. There are no flights on  Continental after 7 AM. I can get you a flight  after 7 AM on Frontier, Northwest, and United  instead. Would that be OK?  U: How about United.  S: United. I have a United flight at 8:01 AM. Is  that flight OK?  U: Yes.  S: OK, I\'ve got you on United flight 1762,  departing Denver at 8:01 AM, arriving at Boston  at 1:42 PM. 6 Current Status and Conclusions  The Talk\'n\'Travel system described here was  successfully demonstrated at the DARPA  Communicator Compare and Contrast  Workshop in June 1999. We are currently  collecting data with test subjects and are using  the results to improve the system\'s performance  in all areas, in preparation for the forthcoming  common evaluation of Communicator systems  in June 2000.  8 of the subjects were successful. Of successful  sessions, the average duration was 387 seconds,  with a minimum of 272 and a maximum of 578.  The average number of user utterances was 25,  with a minimum of 18 and a maximum of 37.  The word error rate of the recognizer was  11.8%.  The primary cause of failure to complete the  scenario, as well as excessive time spent on  completing it, was corruption of the discourse  state due to recognition or interpretation errors.  While the system informs the user of the change  in state after every utterance, the user was not  always successful in correcting it when it made  errors, and sometimes the user did not even  notice when the system had made an error. If the  user is not attentive at the time, or happens not  to understand what the synthesizer said, there is  no implicit way for him to find out afterwards  what the system thinks his constraints are.  While preliminary, these results point to two  directions for future work. One is that the system  needs to be better able to recognize and deal  with problem situations in which the dialogue is  not advancing. The other is that the system  needs to be more communicative about its  current understanding of the user\'s goals, even  at points in the dialogue at which it might be  assumed that user and system were in  agreement. Acknowledgements  This work was sponsored by DARPA and  monitored by SPAWAR Systems Center under  Contract No. N66001-99-D-8615.  To determine the performance of the system, we  ran an informal experiment in which 11 different  subjects called into the system and attempted to  use it to solve a travel problem. None of the  subjects were system developers. Each subject  had a single session in which he was given a  three-city trip to plan, including dates of travel,  constraints on departure and arrival times, airline  preferences.  The author wishes to thank Scott Miller for the  use of his GEM system. '
0.0	P12-2040	W90-0125	b'1 Introduction  Data driven applications have proliferated in Computational Linguistics during the last decade. Several factors, such as the availability of more powerful computers, an almost unlimited storage capacity, the availability of large volumes of data in  digital format, as well as the recent advances in  machine learning theory, have significantly contributed to such a proliferation.    Among the many applications that have benefited from this data-driven boom, probably the most  representative examples are: information retrieval  (Qin et al, 2008), machine translation (Brown et  al., 1993), question answering (Molla-Aliod and  Vicedo, 2010) and dialogue systems (Rieser and  Lemon, 2011).  In the specific case of dialogue systems, data  acquisition can impose some challenges depending  on the specific domain and task the dialogue system is targeted for. In some specific domains, in  which human-human dialogue applications already  exists, data collection is generally straight forward,  while in some other cases, data design and collection can constitute a complex problem (Williams  and Young, 2003; Zue, 2007; Misu et al, 2009).   Depending on the objective being pursued, dialogue systems can be grouped into two major categories: task-oriented and chat-oriented systems. In  the first case, the system is required to help the  user to accomplish a specific goal or objective  (Busemann et al, 1997; Stallard, 2000). In the second case, the system objective is mainly entertainment oriented. Systems in this category are required to play, chitchat or just accompany the user  (Weizenbaum, 1966; Wallis, 2010).   In this work, we focus our attention on dialogue  data which is suitable for training chat-oriented  dialogue systems. Different from task-oriented dialogue collections (Mann, 2003), instead of being  concentrated on a specific domain or area of  knowledge, the training dataset for a chat-oriented  dialogue system must cover a wide variety of domains, as well as be able to provide a fair representation of world-knowledge semantics and pragmatics (Bunt, 2000). To this end, we have collected dialogues from movie scripts aiming at  constructing a dialogue corpus which should provide a good sample of domains, styles and world  knowledge, as well as constitute a valuable resource for research and development purposes.  The rest of the paper is structured as follows.  Section 2 describes in detail the implemented collection process and the structure of the generated  database. Section 3 presents the main statistics, as  well as the main characteristics of the resulting corpus. Finally, section 4 presents our conclusions  and future work plans.   203 2 Collecting Dialogues from Movies   As already stated in the introduction, our presented  dialogue corpus has been extracted from movie  scripts. More specifically, scripts freely available  from The Internet Movie Script Data Collection  (http://www.imsdb.com/) have been used. In this  section we describe the implemented data collection process and the data structure finally used for  the generated corpus.  As a first step of the collection construction,  dialogues have to be identified and extracted from  the crawled html files. Three basic types of information elements are extracted from the scripts:  speakers, utterances and context.   The utterance and speaker information elements  contain what is said at each dialogue turn and the  corresponding character who says it, respectively.  Context information elements, on the other hand,  contain all additional information/texts appearing  in the scripts, which are typically of narrative  nature and explain what is happening in the scene.   Figure 1 depicts a browser snapshot illustrating  the typical layout of a movie script and the most  common spatial distribution of the aforementioned  information elements.  It is important to mention that a lot of different  variants to the format presented in Figure 1 can be  actually encountered in The Internet Movie Script  Data Collection. Because of this, our parsing algorithms had to be revised and adjusted several  times in order to achieve a reasonable level of  robustness that allowed for processing the largest  possible amount of movie scripts.   Another important problem was the identification of dialogue boundaries. Some heuristics were  implemented by taking into account the size and  number of context elements between speaker turns.   A post-processing step was also implemented to  either filter out or amend some of the most common parsing errors occurring during the extraction  phase. Some of these errors include: corrupted formats, turn continuations, notes inserted within the  turn, misspelling of speaker names, etc.  In addition to this, a semi-automatic process was  still necessary to filter out movie scripts exhibiting  extremely different layouts or invalid file formats.  Approximately, 17% of the movie scripts crawled  from The Internet Movie Script Data Collection  had to be discarded. From a total of 911 crawled  scripts, only 753 were successfully processed.       Figure 1: Typical layout of a movie script    The extracted information was finally organized  in dialogical units, in which the information regarding turn sequences inside each dialogue, as well  as dialogue sequences within each movie script  was preserved. Figure 2 illustrates an example of  the XML representation for one of the dialogues  extracted from Who Framed Roger Rabbit.     <dialogue id="47" n_utterances="4">     <speaker>VALIANT</speaker>        <context></context>       <utterance>You shot Roger.</utterance>     <speaker>JESSICA RABBIT</speaker>        <context>Jessica moves the box aside and  tugs on the rabbit ears. The rabbit head pops  off. Underneath is a Weasle. In his hand is the  Colt .45 Buntline.</context>        <utterance>That\'s not Roger. It\'s one of  Doom\'s men. He killed R.K. Maroon.</utterance>     <speaker>VALIANT</speaker>        <context></context>        <utterance>Lady, I guess I had you pegged  wrong.</utterance>     <speaker>JESSICA RABBIT</speaker>        <context>As they run down the  alley...</context>        <utterance>Don\'t worry, you\'re not the  first. We better get out of here.</utterance>  </dialogue>    Figure 2: An example of a dialogue unit   204 3 Movie Dialogue Corpus Statistics  In this section we present the main statistics of the  resulting dialogue corpus and study some of its  more important properties. The final dialogue collection was the result of successfully processing  753 movie scripts. Table 1 summarizes the main  statistics of the resulting dialogue collection.     Total number of scripts collected 911  Total number of scripts processed 753  Total number of  dialogues 132,229  Total number of  speaker turns 764,146  Average amount of dialogues per movie 175.60  Average amount of turns per movie 1,014.80  Average amount of turns per dialogue 5.78    Table 1: Main statistics of the collected movie  dialogue dataset    Movies were mainly crawled from the action,  crime, drama and thriller genres. However, as each  movie commonly belongs to more than one single  genre, much more genres are actually represented  in the dataset. Table 2 summarizes the distribution  of movies by genre (notice that, as most of the  movies belong to more than one genre, the total  summation of percentages exceeds 100%).    Genre Movies Percentage Action 258 34.26  Adventure 133 17.66  Animation 22 2.92  Comedy 149 19.79  Crime 163 21.65  Drama 456 60.56  Family 31 4.12  Fantasy 82 10.89  Horror 104 13.81  Musical 18 2.39  Mystery 95 12.62  Romance 123 16.33  Sci-Fi 129 17.13  Thriller 329 43.69  War 25 3.32  Western 11 1.46    Table 2: Distribution of movies per genre    The first characteristic of the corpus to be analyzed is the distribution of dialogues per movie.  This distribution is shown in Figure 3. As seen  from the figure, the distribution of dialogues per  movie is clearly symmetric around its mean value  of 175 dialogues per movie. For most of the movies in the collection, a number of dialogues ranging from about 100 to 250 were extracted.         Figure 3: Distribution of dialogues per movie     The second property of the corpus to be studied  is the distribution of turns per dialogue. This distribution is shown in Figure 4. As seen from the  figure, this distribution approximates a power law  behavior, with a large number of very short dialogues (about 50K two-turn dialogues) and a small  amount of long dialogues (only six dialogues with  more than 200 turns). The median of the distribution is 5.63 turns per dialogue.         Figure 4: Distribution of turns per dialogue     The third property of the corpus to be described  is the distribution of number of speakers per dia205 logue. This distribution is shown in Figure 5. As  seen from the bar-plot depicted in the figure, the  largest proportion of dialogues (around 60K) involves two speakers. The second largest proportion  of ?dialogues? (about 35K) involves only a single  speaker, which means that this subset of the data  collection is actually composed by monologues or  single speaker interventions. The third and fourth  larger proportions are those involving three and  four speakers, respectively.          Figure 5: Distribution of number of speakers per  dialogue    Finally, in Figure 6, we present a cross-plot between the number of dialogues and the number of  turns within each movie script.       Figure 6: Cross-plot between the number of  dialogues and turns within each movie script  As seen from the cross-plot, an average movie  has between 150 and 200 dialogues comprising  between 1000 and 1200 turns in total. The crossplot also reveals some interesting extreme cases in  the data collection.   For instance, movies with few dialogues but many turns are located towards the upper-left corner  of the figure. In this zone we can find movies as:  Happy Birthday Wanda June, Hannah and Her  Sisters and All About Eve. In the lower-left corner  of the figure we can find movies with few dialogues and few turns, as for instance: 1492 Conquest of Paradise and The Cooler.   In the right side of the figure we find the lots-ofdialogues region. There we can find movies with  lots of very short dialogues (lower-right corner),  such as Jimmy and Judy and Walking Tall; or movies with lots of dialogues and turns (upper-right  corner), such as The Curious Case of Benjamin  Button and Jennifer?s Body.  4 Conclusions and Future Work   In this paper, we have described Movie-DiC a  Movie Dialogue Corpus that has been collected for  research and development purposes. The data collection comprises 132,229 dialogues containing a  total of 764,146 turns/utterances that have been  extracted from 753 movies. Details on how the  data collection has been created and how the  corpus is structured wereprovided along with the  main statistics and characteristics of the corpus.  Although strictly speaking, and by its particular  nature, Movie-DiC does not constitute a corpus of  real human-to-human dialogues, it does constitute  an excellent dataset for studying the semantic and  pragmatic aspects of human communication within  a wide variety of contexts, scenarios, styles and  socio-cultural settings.   Specific technologies and applications that can  exploit a resource like this include, but are not restricted to: example-based chat bots (Banchs and Li,  2012), question answering systems, discourse and  pragmatics analysis, narrative vs. colloquial style  classification, genre classification, etc.   As future work, we intend to expand the current  size of the collection from 0.7K to 2K movies, as  well as to improve some of our parsing and postprocessing algorithms for reducing the amount of  noise still present in the collection and enhance the  quality of the current version of the dataset.  206 Acknowledgments  The author would like to thank the Institute for  Infocomm Research for its support and permission  to publish this work. '	b'1 In t roduct ion   Our project rests on the belief that computer animation in the form of narrated animated  simulations can provide an engaging, effective and flexible medium for instructing agents of  varying capabilities to perform tasks that make varying demands in workplaces of varying layout.  To this end, we have been designing and implementing an integrated system which combines  ? animated agents which can demonstrate the behavior to be emulated;  ? automatic generation of appropriate Natural Language narration which can explain what  is being done and why.  To date, our primary concern with Natural Language has been as input to the system, in  line with the strong claim we make in \\[1\\] that moving task animation beyond direct graphical  manipulation forces one to Natural Language as the only instruction source accessible to other  users than the current community of manually skilled (or programming-wise) animators. (To  this end, we have been analysing constructions commonly found in NL instructions, in terms of  their representational requirements \\[3\\].  However here our point of discussion is NL Generation. What makes us such eager con-  sumers of advances and technology in this area is that animated simulations without narration  (ultimately, spoken narration) is only half the story. As researchers studying plan inference  have shown \\[2\\], it may be well-nigh impossible to infer an agent\'s intentions imply by observing  his or her actions alone. 2 And we know that the ability to perform an action effectively in a  range of environments requires understanding its intention, not just the physical motions used  in some performance. Thus, communicating intentions is as important o effective task instruc-  tion as demonstrating physical skills. Sharing the burden of communication between Natural  Language and graphics, as Feiner and McKeown have noted \\[4\\], takes advantage of the best of  both possible worlds.  While some parts of our system are further along than others, no work at all has yet been  done on generation. However, we have tried to take account of the needs of generation in  designing the system, so that we will not have painted ourselves in a hole from the start. We  clearly and hope to get further ideas and direction from this meeting. Basically, the system has  been designed so that the generator will receive information from three sources (see Figure 1.):  1 This research is partially supported by Lockheed Engineering and Management Services (NASA Johnson  Space Center), NASA Ames Grant NAG-2-426, FMC Corporation, Martin-Marietta Denver Aerospace, NSF  CISE Grant CDA88-22719, and ARO Grant DAAL03-89-C-0031 including participation by the U.S. Army Human  Engineering Laboratory.  Exaggerating behavior to make it more communicative may have the adverse ffect of making it less veridical,  a situation inversely turned advantageous by skilled cartoon animators \\[5\\].  189  * the partial global plan (an incrementally computed escription of what the animated agent  is meant o do and why); ? the basic animation commands (for particulars of what\'s happening "now")  ? the visualization plan (for what can the viewer see).  The resulting narrative is thus meant to satisfy the joint communicative goals of providing  an overall context in which to view the events on the "screen" and explaining the reasons for  particular events that are happening, thereby transcending the merely visible portion of any  event, to augment and reinforce observable behavior. For a more detailed description of the  system and further discussion of instructions and task performance, the reader is referred to \\[1\\]. '
1.0	C02-1042	A00-1023	b'1.   Introduction  The TREC evaluations of QA systems  (Voorhees, 1999) require answers to be drawn  from a given source corpus.  Early QA systems  used a simple filtering technique, question word  density within a fixed n-word window, to  pinpoint answers.  Robust though this may be,  the window method is not accurate enough.  In  response, factoid question answering systems  have evolved into two types:   ? Use-Knowledge: extract query words from  the input question, perform IR against the  source corpus, possibly segment resulting  documents, identify a set of segments  containing likely answers, apply a set of  heuristics that each consults a different  source of knowledge to score each  candidate, rank them, and select the best  (Harabagiu et al, 2001; Hovy et al, 2001;  Srihari and Li, 2000; Abney et al, 2000).   ? Use-the-Web: extract query words from the  question, perform IR against the web,  extract likely answer-bearing sentences,  canonicalize the results, and select the most  frequent answer(s).  Then, for justification,  locate examples of the answers in the source  corpus (Brill et al, 2001; Buchholz, 2001).   Of course, these techniques can be combined:  the popularity ratings from Use-the-Web can  also be applied as a filtering criterion (Clarke et  al., 2001), or the knowledge resource heuristics  can filter the web results.  However, simply  going to the web without using further  knowledge (Brill et al, 2001) may return the  web?s majority opinions on astrology, the killers  of JFK, the cancerous effects of microwave  ovens, etc.?fun but not altogether trustworthy.    In this paper we describe the range of  filtering techniques our system Webclopedia  applies, from simplest to most sophisticated, and  indicate their impact on the system.    2.   Webclopedia Architecture   As shown in Figure 1, Webclopedia adopts the  Use-Knowledge architecture. Its modules are  described in more detail in (Hovy et al, 2001;  Hovy et al, 1999):   ? Question parsing: Using BBN?s  IdentiFinder (Bikel et al, 1999), the  CONTEX parser (Hermjakob, 1997)  produces a syntactic-semantic analysis of  the question and determines the QA type.    ? Query formation: Single- and multi-word  units (content words) are extracted from the  analysis, and WordNet synsets (Fellbaum,  1998) are used for query expansion.  A  series of Boolean queries of decreasing  specificity is formed.   ? IR: The publicly available IR engine MG  (Witten et al, 1994) returns the top-ranked  N documents.    ? Selecting and ranking sentences: For each  document, the most promising K sentences  are located and scored using a formula that    rewards word and phrase overlap with the  question and its expanded query words.  Results  are ranked.    ? Parsing candidates: CONTEXparses the  top-ranked 300 sentences.    ? Pinpointing: As described in Section 3, a  number of knowledge resources are used to  perform filtering/pinpointing operations.    ? Ranking of answers: The candidate  answers? scores are compared and the  winner(s) are output.  3. Knowledge Used for Pinpointing  3.1   Type 1: Question Word Matching  Unlike (Prager et al, 1999), we do not first  annotate the source corpus, but perform IR  directly on the source text, using MG (Witten et  al., 1994).  To determine goodness, we assign an  initial base score to each retrieved sentence.  We  then compare the sentence to the question and  adapt this score as follows:   ? exact matches of proper names double the  base score.  ? matching an upper-cased term adds a 60%  bonus of the base score for multi-words  terms and 30% for single words (matching  ?United States? is better than just ?United?).   ? matching a WordNet synonym of a term  discounts by 10% (lower case) and 50%  (upper case).  (When ?Cage? matches  ?cage?, the former may be the last name of a  person and the latter an object; the case  mismatch signals less reliability.)   ? lower-case term matches after Porter  stemming are discounted 30%; upper-case  matches 70% (Porter stemming is more  aggressive than WordNet stemming).   ? Porter stemmer matches of both question  and sentence words with lower case are  discounted 60%; with upper case, 80%.   ? if CONTEX indicates a term as being  qsubsumed (see Section 3.9) the term is  discouned 90% (in ?Which country  manufactures weapons of mass  destruction??, ?country? will be marked as  qsubsumed).    The top-scoring 300 sentences are passed on for  further filtering.    3.2  Type 2: Qtargets, the QA Typology,  and the Semantic Ontology  We classify desired answers by their semantic  type, which have been taxonomized in the  Webclopedia QA Typology (Hovy et al, 2002),  Candidate answer parsing ? Steps: parse sentences ? Engines: CONTEX Matching ? Steps: match general constraint patterns against parse trees              match desired semantic type against parse tree elements              assign score to words in sliding window ? Engine: Matcher Ranking and answer extraction ? Steps: rank candidate answers              extract and format them ? Engine: Answer ranker/formatter QA typology ? QA types, categorized in taxonomy Constraint patterns ? Identify likely answers in relation to    other parts of the sentence Create query Retrieve documents Select & rank sentences Parse top sentences Parse question Input question Perform additional inference Rank and prepare answers Output answers Question parsing ?Steps: parse question              find desired semantic type ? Engines:  IdentiFinder  (BBN)                  CONTEX Match sentences against answers Query creation ?  Steps: extract, combine important words  expand query words using WordNet  create queries, order by specificity ?  Engines: Query creator IR ?  Steps: retrieve top 1000 documents ?  Engines: MG (RMIT Melbourne) Sentence selection and ranking ?  Steps: score each sentence in each document  rank sentences and pass top 300 along ?  Engines:Ranker Figure 1. Webclopedia architecture.   http://www.isi.edu/natural-language/projects/we  bclopedia/Taxonomy/taxonomy_toplevel.html).  The currently approx. 180 classes,  which we  call qtargets, were developed after an analysis of  over 17,000 questions (downloaded in 1999  from answers.com) and later enhancements to  Webclopedia.  They are of several types:   ? common semantic classes such as PROPERPERSON, EMAIL-ADDRESS, LOCATION,  PROPER-ORGANIZATION;   ? classes particular to QA such as YES:NO,  ABBREVIATION-EXPANSION, and WHYFAMOUS;   ? syntactic classes such as NP and NOUN,  when no semnatic type can be determined  (e.g., ?What does Peugeot manufacture??);   ? roles and slots, such as REASON and TITLEP respectively, to indicate a desired relation  with an anchoring concept.     Given a question, the CONTEX parser uses a  set of 276 hand-built rules to identify its most  likely qtarget(s), and records them in a backoff  scheme (allowing more general qtarget nodes to  apply when more specific ones fail to find a  match).  The generalizations are captured in a  typical concept ontology, a 10,000-node extract  of WordNet.    The recursive part of pattern matching is  driven mostly by interrogative phrases.  For  example, the rule that determines the  applicability of the qtarget WHY-FAMOUS  requires the question word ?who?, followed by  the copula, followed by a proper name.  When  there is no match at the current level, the system  examines any interrogative constituent, or words  in special relations to it.  For example, the  qtarget TEMPERATURE-QUANTITY (as in  ?What is the melting point of X?? requires as  syntactic object something that in the ontology is  subordinate to TEMP-QUANTIFIABLE-ABSTRACT with, as well, the word ?how? paired  with ?warm?, ?cold?, ?hot?, etc., or the phrase   ?how many degrees? and a TEMPERATUREUNIT (as defined in the ontology).    3.3 Type 3: Surface Pattern Matching  Often qtarget answers are expressed using rather  stereotypical words or phrases.  For example, the  year of birth of a person is typically expressed  using one of these phrases:   <name> was born in <birthyear>  <name> (<birthyear>?<deathyear>)  We have developed a method to learn such  patterns automatically from text on the web  (Ravichandran and Hovy, 2002).  Wehave  added into the QA Typology the patterns for  appropriate qtargets (qtargets with closed-list  answers, such as PLANETS, require no patterns).   Where some QA systems use such patterns  exclusively (Soubbotin and Soubbotin, 2001) or  partially (Wang et al, 2001; Lee et al, 2001),  we employ them as an additional source of  evidence for the answer.  Preliminary results on  for a range of qtargets, using the TREC-10  questions and the TREC corpus, are:   Question type  (qtarget)  Number of  questions  MRR on  TREC docs  BIRTHYEAR 8 0.47875  INVENTORS 6 0.16667  DISCOVERERS 4 0.1250  DEFINITIONS 102 0.3445  WHY-FAMOUS 3 0.6666  LOCATIONS 16 0.75  3.4  Type 4: Expected Numerical Ranges   Quantity-targeting questions are often  underspecified and rely on culturally shared   cooperativeness rules and/or world knowledge:  Q: How many people live in Chile?   S1: ?From our correspondent comes good  news about the nine people living in  Chile??  A1: nine   While certainly nine people do live in Chile,  we know what the questioner intends.  We have  hand-implemented a rule that provides default  range assumptions for POPULATION questions  and biases quantity questions accordingly.   3.5 Type 5: Abbreviation Expansion   Abbreviations often follow a pattern:  Q: What does NAFTA stand for?  S1: ?This range of topics includes the North  American Free Trade Agreement, NAFTA,  and the world trade agreement GATT.?   S2: ?The interview now changed to the subject  of trade and pending economic issues, such as  the issue of opening the rice market, NAFTA,  and the issue of Russia repaying economic  cooperation funds.?   After Webclopedia identifies the qtarget as  ABBREVIATION-EXPANSION, it extracts   possible answer candidates, including ?North  American Free Trade Agreement? from S1 and  ?the rice market? from S2.  Rules for acronym  matching easily prefer the former.   3.6 Type 6: Semantic Type Matching   Phone numbers, zip codes, email addresses,  URLs, and different types of quantities obey  lexicographic patterns that can be exploited for  matching, as in   Q: What is the zip code for Fremont, CA?   S1: ??from Everex Systems Inc., 48431  Milmont Drive, Fremont, CA 94538.?   and   Q: How hot is the core of the earth?   S1. ?The temperature of Earth?s inner core  may be as high as 9,000 degrees Fahrenheit  (5,000 degrees Celsius).?   Webclopedia identifies the qtargets respectively  as ZIP-CODE and TEMPERATURE-QUANTITY.   Approx. 30 heuristics (cascaded) apply to the  input before parsing to mark up numbers and  other orthographically recognizable units of all  kinds, including (likely) zip codes, quotations,  year ranges, phone numbers, dates, times,  scores, cardinal and ordinal numbers, etc.   Similar work is reported in (Kwok et al, 2001).  3.7 Type 7: Definitions from WordNet   We have found a 10% increase in accuracy in  answering definition questions by using external  glosses obtained from WordNet.  For   Q: What is the Milky Way?   Webclopedia identified two leading answer  candidates:    A1: outer regions   A2: the galaxy that contains the Earth   Comparing these with the WordNet gloss:   WordNet: ?Milky Way?the galaxy containing  the solar system?   allows Webclopedia to straightforwardly match  the candidate with the greater word overlap.    Curiously, the system also needs to use  WordNet to answer questions involving  common knowledge, as in:   Q: What is the capital of the United States?   because authors of the TREC collection do not  find it necessary to explain what Washington is:   Ex: ?Later in the day, the president returned to  Washington, the capital of the United States.?   While WordNet?s definition   Wordnet: ?Washington?the capital of the  United States?   directly provides the answer to the matcher, it  also allows the IR module to focus its search on  passages containing ?Washington?, ?capital?,  and ?United States?, and the matcher to pick a  good motivating passage in the source corpus.    Clearly, this capability can be extended to  include (definitional and other) information  provided by other sources, including  encyclopedias and the web (Lin 2002).  3.8 Type 8: Semantic Relation Matching   So far, we have considered individual words and  groups of words.  But often this is insufficient to  accurately score an answer.  As also noted in  (Buchholz, 2001), pinpointing can be improved  significantly by matching semantic relations  among constituents:   Q: Who killed Lee Harvey Oswald?   Qtargets: PROPER-PERSON & PROPER-NAME,  PROPER-ORGANIZATION   S1: ?Belli?s clients have included Jack Ruby,  who killed John F. Kennedy assassin Lee  Harvey Oswald, and Jim and Tammy Bakker.?   S2: ?On Nov. 22, 1963, the building gained  national notoriety when Lee Harvey Oswald  allegedly shot and killed President John F.  Kennedy from a sixth floor window as the  presidential motorcade passed.?   The CONTEX parser (Hermjakob, 1997;  2001) provides the semantic relations.  The  parser uses machine learning techniques to build  a robust grammar that produces semantically  annotated syntax parses of English (and Korean  and Chinese) sentences at approx. 90% accuracy  (Hermjakob, 1999).    The matcher compares the parse trees of S1  and S2 to that of the question.  Both S1 and S2  receive credit for matching question words ?Lee  Harvey Oswald? and ?kill? (underlined), as well  as for finding an answer (bold) of the proper  qtarget type (PROPER-PERSON).  However, is  the answer ?Jack Ruby? or ?President John F.  Kennedy??  The only way todetermine this is to  consider the semantic relationship between these   candidates and the verb ?kill? (parse trees  simplified, and only portions shown here):      [1] Who killed Lee Harvey Oswald?  [S-SNT]      (SUBJ) [2] Who  [S-INTERR-NP]          (PRED) [3] Who  [S-INTERR-PRON]      (PRED) [4] killed  [S-TR-VERB]      (OBJ) [5] Lee Harvey Oswald  [S-NP]          (PRED) [6] Lee?Oswald  [S-PROPER-NAME]              (MOD) [7] Lee  [S-PROPER-NAME]              (MOD) [8] Harvey  [S-PROPER-NAME]              (PRED) [9] Oswald  [S-PROPER-NAME]      (DUMMY) [10] ?  [D-QUESTION-MARK]    [1] Jack Ruby, who killed John F. Kennedy assassin     Lee Harvey Oswald  [S-NP]     (PRED) [2] <Jack Ruby>1  [S-NP]     (DUMMY) [6] ,  [D-COMMA]     (MOD) [7] who killed John F. Kennedy assassin                    Lee Harvey Oswald  [S-REL-CLAUSE]       (SUBJ) [8] who<1>  [S-INTERR-NP]       (PRED) [10] killed  [S-TR-VERB]       (OBJ) [11] JFK assassin?Oswald  [S-NP]           (PRED) [12] JFK?Oswald [S-PROP-NAME]               (MOD) [13] JFK  [S-PROPER-NAME]               (MOD) [19] assassin  [S-NOUN]               (PRED) [20] ?Oswald [S-PROPER-NAME]  Although the PREDs of both S1 and S2  match that of the question ?killed?, only S1  matches ?Lee Harvey Oswald? as the head of  the logical OBJect.  Thus for S1, the matcher  awards additional credit to node [2] (Jack Ruby)  for being the logical SUBJect of the killing  (using anaphora resolution). In S2, the parse tree  correctly records that node [13] (?John F.  Kennedy?) is not the object of the killing.  Thus  despite its being closer to ?killed?, the candidate  in S2 receives no extra credit from semantic  relation matching.    It is important to note that the matcher  awards extra credit for each matching semantic  relationship between two constituents, not only  when everything matches.  This granularity  improves robustness in the case of partial  matches.    Semantic relation matching applies not only  to logical subjects and objects, but also to all  otherroles such as location, time, reason, etc.  (for additional examples see http://www.isi.edu/  natural-language/projects/webclopedia/sem-relexamples.html).  It also applies at not only the  sentential level, but at all levels, such as postmodifying prepositional and pre-modifying  determiner phrases   Additionally, Webclopedia uses 10 lists of  word variations with a total of 4029 entries for  semantically related concepts such as ?to  invent?, ?invention? and ?inventor?, and rules  for handling them.  For example, via coercing  ?invention? to ?invent?, the system can give  ?Johan Vaaler? extra credit for being a likely  logical subject of ?invention?:   Q: Who invented the paper clip?   Qtargets: PROPER-PERSON & PROPER-NAME,  PROPER-ORGANIZATION   S1: ?The paper clip, weighing a desk-crushing  1,320 pounds, is a faithful copy of Norwegian  Johan Vaaler?s 1899 invention, said Per  Langaker of the Norwegian School of  Management.?   while ?David? actually loses points for being  outside of the clausal scope of the inventing:   S2: ??Like the guy who invented the safety pin,  or the guy who invented the paper clip,? David  added.?   3.9 Type 9: Word Window Scoring   Webclopedia also includes a typical windowbased scoring module that moves a window over  the text and assigns a score to each window  position depending on a variety of criteria (Hovy  et al, 1999).  Unlike (Clarke et al, 2001; Lee et  al., 2001; Chen et al, 2001), we have not  developed a very sophisticated scoring function,  preferring to focus on the modules that employ  information deeper than the word level.   This method is applied only when no other  method provides a sufficiently high-scoring  answer.  The window scoring function is   S  = (500/(500+w))*(1/r) * ?[(?I1.5*q*e*b*u)1.5]  Factors:  w: window width (modulated by gaps of  various lengths: ?white house? ? ?white car and  house?),  r: rank of qtarget in list returned by  CONTEX,  I: window word information content (inverse  log frequency score of each word), summed,   q: # different question words matched, plus  specific rewards (bonus q=3.0),   e: penalty if word matches one of question  word?s WordNet synset items (e=0.8),    b: bonus for matching main verb, proper  names, certain target words (b=2.0),   u: (value 0 or 1) indicates whether a word has  been qsubsumed (?subsumed? by the qtarget)  and should not contribute (again) to the score.   For example, ?In what year did Columbus  discover America?? the qsubsumed words are  ?what? and ?year?.  4. Performance Evaluation   In TREC-10?s QA track, Webclopedia received  an overall Mean Reciprocal Rank (MRR) score  of 0.435, which put it among the top 4  performers of the 68 entrants (the average MRR  score for the main QA task was about 0.234).   The pinpointing heuristics are fairly accurate:  whenWebclopedia finds answers, it usually  ranks them in the first place (1st place: 35.5%;  2nd: 8.94%; 3rd: 5.69%; 4th: 3.05%; 5th: 5.28%;  not found: 41.87%).   We determined the impact of each  knowledge source on system performance, using  the TREC-10 test corpus using the standard  MRR scoring.  We applied the system to the  questions of each knowledge type separately,  with and without its specific knowledge  source/algorithm.  Results are shown in Table 1,  columns A (without) and B (with).  To indicate  overall effect, we also show (in columns C and  D) the percentage of questions in TREC-10 and  -9 respecively of each knowledge type.   5. Conclusions  It is tempting to search for a single technique  that will solve the whole problem (for example,  Ittycheriah et al (2001) focus on the subset of  factoid questions answerable by NPs, and train a  statistical model to perform NP-oriented answer  pinpointing).  Our experience, however, is that  even factoid QA is varied enough to require  various special-purpose techniques and  knowledge.  The theoretical limits of the various  techniques are not known, though Light et al?s  (2001) interesting work begins to study this.    Column A: % questions of the knowledge type        answered correctly without using knowlege  Column B: % questions, now using knowledge  Column C: % questions of type in TREC-10   Column D: % questions of type in TREC-9    A B C D  Abbreviation exp. 20.0 70.0  1.0 2.3  Number ranges 50.0 50.0  1.2 1.8  WordNet (def Qs) 48.3 67.5 20.9 5.1  Semantic types      - locator types N/A N/A  0.0 0.4  - quantity types 22.5 48.7 10.8 5.5  - date/year types 45.0 57.3  9.2 10.2  Patterns       - definitions ? 34.4 20.9 5.1  - why-famous  ? 66.7 0.6 ?  - locations ? 75.0 3.2 ?  - birthyear ? 47.9 1.6 ?  Semantic relations 39.4 46.5 72.2 85.7  Table 1. Performance of knowledge sources.  Semantic relation scores measured only on  questions in which they could logically apply.    We conclude that factoid QA performance  can be significantly improved by the use of  knowledge attuned to specific question types  and specific information characteristics.  Most of  the techniques for exploiting this knowledge  require learning to ensure robustness.  To  improve performance beyond this, we believe a  combination of going to the web and turning to  deeper world knowledge and automated  inference (Harabagiu et al, 2001) to be the  answer.  It remains an open question how much  work these techniques would require, and what  their payoff limits are.   '	b'Introduction  With the explosion of information in Internet,  Natural language QA is recognized as a  capability with great potential. Traditionally,  QA has attracted many AI researchers, but most  QA systems developed are toy systems or games  confined to lab and a very restricted omain.  More recently, Text Retrieval Conference  (TREC-8) designed a QA track to stimulate the  research for real world application.  Due to little linguistic support from text  analysis, conventional IR systems or search  engines do not really perform the task of  information retrieval; they in fact aim at only  document retrieval. The following quote from the  QA Track Specifications (www.research.att.com/  -singhal/qa-track-spec.txt) in the TREC  community illustrates this point.  Current information retrieval systems allow  us to locate documents hat might contain the  pertinent information, but most of them leave  it to the user to extract he useful information  from a ranked list. This leaves the (often  unwilling) user with a relatively large  amount of text o consume. There is an urgent  need for tools that would reduce the amount  of text one might have to read in order to  obtain the desired information. This track  aims at doing exactly that for a special (and  popular) class of information seeking  behavior: QUESTION ANSWERING. People  have questions and they need answers, not  documents. Automatic question answering  will definitely be a significant advance in the  state-of-art information retrieval technology.  Kupiec (1993) presented a QA system  MURAX using an on-line encyclopedia. This  system used the technology of robust shallow  parsing but suffered from the lack of basic  information extraction support. In fact, the most  siginifcant IE advance, namely the NE (Named  Entity) technology, occured after Kupiec (1993),  thanks to the MUC program (MUC-7 1998).  High-level IE technology beyond NE has not  been in the stage of possible application until  recently.  AskJeeves launched a QA portal  (www.askjeeves.com). It is equipped with a  fairly sophisticated natural language question  parser, but it does not provide direct answers to  the asked questions. Instead, it directs the user to  the relevant web pages, just as the traditional  search engine does. In this sense, AskJeeves has  only done half of the job for QA.  We believe that QA is an ideal test bed for  demonstrating the power of IE. There is a natural  co-operation between IE and IR; we regard QA  as one major intelligence which IE can offer IR.  * This work was supported in part by the SBIR grants F30602-98-C-0043 and F30602-99-C-0102 from Air Force  Research Laboratory (AFRL)/IFED.  166  An important question then is, what type of  IE can support IR in QA and how well does it  support it? This forms the major topic of this  paper. We structure the remaining part of the paper as follows. In Section 1, we first give an  overview of the underlying IE technology which  our organization has been developing. Section 2  discusses the QA system. Section 3 describes the  limitation of the current system. Finally, in  Section 4, we propose a more sophisticated QA  system supported by three levels of IE.  1 Overview of Textract IE  The last decade has seen great advance and  interest in the area of IE. In the US, the DARPA  sponsored Tipster Text Program \\[Grishman  1997\\] and the Message Understanding  Conferences (MUC) \\[MUC-7 1998\\] have been  the driving force for developing this technology.  In fact, the MUC specifications for various IE  tasks have become de facto standards in the IE  research community. It is therefore necessary to  present our IE effort in the context of the MUC  program.  MUC divides IE into distinct tasks,  namely, NE (Named Entity), TE (Template  Element), TR (Template Relation), CO  (Co-reference), and ST (Scenario Templates)  \\[Chinchor & Marsh 1998\\]. Our proposal for  three levels of IE is modelled after the MUC  standards using MUC-style representation.  However, we have modified the MUC IE task  definitions in order to make them more useful  and more practical. More precisely, we propose a hierarchical, 3-level architecture for developing a  kernel IE system which is domain-independent  throughout.  The core of this system is a state-of-the-art  NE tagger \\[Srihari 1998\\], named Textract 1.0.  The Textract NE tagger has achieved speed and  accuracy comparable tothat of the few deployed  NE systems, such as NetOwl \\[Krupka &  Hausman 1998\\] and Nymble \\[Bikel et al1997\\].  It is to be noted that in our definition of NE,  we significantly expanded the type of  information to be extracted. In addition to all the  MUC defined NE types (person, organization,  location, time, date, money and percent), the  following types/sub-types of information are also  identified by the TextractNE module:  ? duration, frequency, age  ? number, fraction, decimal, ordinal, math  equation  ? weight, length, temperature, angle, area,  capacity, speed, rate  ? product, software  ? address, email, phone, fax, telex, www  ? name (default proper name)  Sub-type information like company,  government agency, school (belonging to the  type organization) and military person, religious  person (belonging to person) are also identified.  These new sub-types provide a better foundation  for defining multiple relationships between the  identified entities and for supporting question  answering functionality. For example, the key to  a question processor is to identify the asking  point (who, what, when, where, etc.). In many  cases, the asking point corresponds to an NE  beyond the MUC definition, e.g. the  how+adjective questions: how long (duration or  length), how far (length), how often (frequency),  how old (age), etc.  Level-2 IE, or CE(Correlated Entity), is  concerned with extracting pre-defined multiple  relationships between the entities. Consider the  person entity as an example; the TextractCE  prototype is capable of extracting the key  relationships uch as age, gender, affiliation,  position, birthtime, birth__place, spouse,  parents, children, where.from, address, phone,  fax, email, descriptors. As seen, the information  in the CE represents a mini-CV or profile of the  entity. In general, the CE template integrates and  greatly enriches the information contained in  MUC TE and TR.  The final goal of our IE effort is to further  extract open-ended general events (GE, or level 3  IE) for information like who did what (to whom)  when (or how often) and where. By general  events, we refer to argument structures centering  around verb notions plus the associated  information of time/frequency and location. We  show an example of our defined GE extracted  from the text below:  Julian Hill, a research chemist whose  accidental discovery of a tough, taffylike  compound revolutionized everyday life after  it proved its worth in warfare and courtship,  167  died on Sunday in Hockessin, Del.  \\[1\\] <GE_TEMPLATE> :=  PREDICATE: die  ARGUMENTI: Julian Hill  TIME: Sunday  LOCATION: Hockessin, Del  Figure 1 is the overall system architecture for  the IE system Textract hat our organization has  been developing.  Kernet IE Modutes  L|ngui_sti_cLM_odu!es  I . . . . . . . . . . . . .  I I . . . . . . . . . .  I  I I I I ,,l l I , ! I  I I I I  \'l J \' I I  ! I  I I  I I I I ,i I i \', I ,  I ! I I , l ? i  , ,  I I I I  I I  F L - -  - - - -~  . . . .  . L - -  - -  - -  . - -  - -  - -  | . . . .   Apptication Modutes  NE: NIiml~ EnlilyTitl~klll QA: Que~tlon Answering  CE: Come,led Entity ExtrmClkm BR: In~lllgenl ~ws lng   GE: Gcn~mI Evenl Ex~ct~on AS; Auio~ SUl len   co :  ce-  mfcmnc ~1 momial l  s ~ p~  Figure 1: Textract IE System Architecture  The core of the system consists of three  kernel IE modules and six linguistic modules.  The multi-level linguistic modules erve as an  underlying support system for different levels of  IE. The IE results are stored in a database which  is the basis for IE-related applications like QA,  BR (Browsing, threading and visualization) and  AS (Automatic Summarization). The approach  to IE taken here, consists of a unique blend of  machine learning andFST (finite state  transducer) rule-based system \\[Roche & Schabes  1997\\]. By combining machine learning with an  FST rule-based system, we are able to exploit he  best of both paradigms while overcoming their  respective weaknesses \\[Srihari 1998, Li & Srihari  2000\\].  2 NE-Supported QA  This section presents the QA system based on  Named Entity tagging. Out of the 200 questions  that comprised the TREC-8 QA track  competition, over 80% asked for an NE, e.g. who  (PERSON), when (T IME\\ [  DATE), where  (LOCATION), how far (LENGTH). Therefore,  the NE tagger has been proven to be very helpful.  Of course, the NE of the targeted type is only  necessary but not complete in answering such  questions because NE by nature only extracts  isolated individual entities from the text.  Nevertheless, using even crude methods like "the  nearest NE to the queried key words" or "the NE  and its related key words within the same line (or  same paragraph, etc.)", in most cases, the QA  system was able to extract ext portions which  contained answers in the top five list.  Figure 2 illustrates the system design of  TextractQA Prototype. There are two  components for the QA prototype: Question  Processor and Text Processor. The Text Matcher  module links the two processing results and tries  to find answers to the processed question.  Matching is based on keywords, plus the NE  type and their common location within a same  sentence.  Quest ion  Prc~:essor  i  : :eXt P r~_~ . . . . . . . . . . . .  ? ~   i i ~ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   i . . . .  i  Figure 2: Textract/QA 1.0 Prototype Architecture  The general algorithm for question  answering is as follows:  168  Process Question  Shallow parse question  Determine Asking Point  Question expansion (using word lists)  Process Documents  Tokenization, POS tagging, NE Indexing  Shallow Parsing (not yet utilized)  Text Matcher  Intersect search engine results with NE  rank answers  2.1 Question Processing  The Question Processing results are a list of  keywords plus the information for asking point.  For example, the question:  \\[2\\] Who won the 1998 Nobel Peace Prize?  contains the following keywords: won, 1998,  Nobel, Peace, Prize. The asking point Who refers  to the NE type person. The output before  question expansion is a simple 2-feature template  as shown below:  \\[3\\] asking_point: PERSON  key_word: { won, 1998, Nobel,  Peace, Prize }  The following is an example where the  asking point does not correspond to any type of  NE in ourdefinition.  \\[3\\] Why did David Koresh ask the FBI for a  word processor ?  The system then maps it to the following  question template :  \\[4\\] asking_point:  key_word:  REASON  { ask, David, Koresh,  FBI, word, processor } The question processor scans the question to  search for question words (wh-words) and maps  them into corresponding NE types/sub-types or  pre-defined notions like REASON.  We adopt wo sets of pattern matching rules  for this purpose: (i) structure based pattern  matching rules; (ii) simple key word based  pattern matching rules (regarded as default rules).  It is fairly easy to exhaust the second set of rules  as interrogative question words/phrases form a  closed set. In comparison, the development of the first set of rules are continuously being  fine-tuned and expanded. This strategy of using  two set of rules leads to the robustness of the  question processor.  The first set of rules are based on shallow  parsing results of the questions, using Cymfony  FST based Shallow Parser. This parser identifies  basic syntactic onstructions like BaseNP (Basic  Noun Phrase), BasePP (Basic Prepositional  Phrase) and VG (Verb Group).  The following is a sample of the first set of  rules:  \\[6\\] Name NP (city I country I company) -->  CITYICOUNTRYICOMPANY  \\[7\\] Name NP(person_w) --> PERSON  \\[8\\] Name NP(org_w) --> ORGANIZATION  \\[9\\] Name NP(NOT person_w, NOT org_w)  --> NAME  Rule \\[6\\] checks the head word of the NP. It  covers cases like VG\\[Name\\] NP\\[a country\\] that  VG\\[is developing\\] NP\\[a magnetic levitation  railway system\\]. Rule \\[7\\] works for cases like  VG\\[Name\\] NP\\[the first private citizen\\] VG\\[to  fly\\] PP\\[in space\\] as citizen belongs to the word  class person_w. Rule \\[9\\] is a catch-all rule: if the  NP is not of class person (person_w) or  organization (org_w), then the asking point is a  proper name (default NE), often realized in  English in capitalized string of words. Examples  include Name a film that has won the Golden  Bear in the Berlin Film Festival.  The word lists org_w and person_w are  currently manually maintained based on  inspection of large volumes of text. An effort is  underway to automate the learning of such word  lists by utilizing machine learning techniques.  We used the following pattern  transformations to expand our ruleset:  (Please) name NP\\[X\\]  --> what/which Aux(be) (the name of) NP\\[X\\]  --> NP(what/which...X)  In other words, the four rules are expanded to  12 rules. For example, Rule \\[10\\] below  corresponds to Rule \\[6\\]; Rule \\[11\\] is derived  169  from Rule \\[7\\].  \\[10\\] what/which Aux(be) NP (city \\[ country \\[  company) -->  CITY I COUNTRY \\[ COMPANY  \\[11\\] NP(what/which ... person_w) -->  PERSON  Rule \\[10\\] extracts the asking point from  cases like NP\\[What\\] Aux\\[is\\] NP\\[the largest  country\\] PP\\[inthe world\\]. Rule \\[11\\] covers the  following questions: NP\\[What costume  designer\\] VG\\[decided\\] that NP\\[Michael  Jacksonl VG\\[should only wear\\] NP\\[one glove\\],  NP\\[Which former Ku Klux Klan member\\]  VG\\[won\\] NP\\[an elected office\\] PP\\[in the U.S.\\],  NP\\[What Nobel laureate\\] VG\\[was expelled\\]  PP\\[from the Philippines\\] PP\\[before the  conference\\] PP\\[on East Timor\\], NP\\[What  famous communist leader\\] VG\\[died\\] PP\\[in  Mexico City\\], etc.  As seen, shallow parsing helps us to capture a variety of natural anguage question expressions.  However, there are cases where some simple key  word based pattern matching would be enough to  capture the asking point. That is our second set  of rules. These rules are used when the first set of  rules has failed to produce results. The following  is a sample of such rules:  \\[ 12\\] who/whom --> PERSON  \\[13\\] when --> TIME/DATE  \\[14\\] where/what place --> LOCATION  \\[15\\] what time (of day) --> TIME  \\[16\\] what day (of the week) --> DAY  \\[17\\] what/which month --> MONTH  \\[18\\] what age/how old --> AGE  \\[19\\] what brand --> PRODUCT  \\[20\\] what --> NAME  \\[21\\] how far/tall/high --> LENGTH  \\[22\\] how large/hig/small --> AREA  \\[23\\] how heavy --> WEIGHT  \\[24\\] how rich --> MONEY  \\[25\\] how often --> FREQUENCY  \\[26\\] how many --> NUMBER  \\[27\\] how long --> LENGTH/DURATION  \\[28\\] why/for what --> REASON  In the stage of question expansion, the  template in \\[4\\] would be expanded to the  template shown in \\[29\\]:  \\[29\\] asking_point:  key_word:  {because{because of\\]  due to{thanks to{since I  in order{to VB}  { asklaskslasked\\[asking,  David,Koresh,FBI,  word, processor}  The last item in the asking._point list attempts  to find an infinitive by checking the word to  followed by a verb (with the part-of-speech tag  VB). As we know, infinitive verb phrases are  often used in English to explain a reason for some  action.  2.2 Text Processing  On the text processing side, we first send the  question directly to a search engine in order to  narrow down the document pool to the first n, say  200, documents for IE processing. Currently,  this includes tokenization, POS tagging and NE  tagging. Future plans include several evels of  parsing as well; these are required to support CE  and GE extraction. It should be noted that all  these operations are extremely robust and fast,  features necessary for large volume text  indexing. Parsing is accomplished through  cascaded finite state transducer grammars.  2.3 Text Matching  The Text Matcher attempts to match the question  template with the processed ocuments for both  the asking point and the key words. There is a  preliminary ranking standard built-in the matcher  in order to find the most probable answers. The  primary rank is a count of how many unique  keywords are contained within a sentence.The  secondary ranking is based on the order that the  keywords appear in the sentence compared to  their order in the question. The third ranking is  based on whether there is an exact match or a  variant match for the key verb.  In the TREC-8 QA track competition,  Cymfony QA accuracy was 66.0%. Considering  we have only used NE technology to support QA  in this run, 66.0% is a very encouraging result.  3 Limitation  The first limitation comes from the types of  questions. Currently only wh-questions are  handled although it is planned that yes-no  questions will be handled once we introduce CE  170  and GE templates to support QA. Among the  wh-questions, the why-question and  how-question t are more challenging because the  asking point cannot be simply mapped to the NE  types/sub-types.  The second limitation is from the nature of  the questions. Questions like Where can l find  the homepage for Oscar winners or Where can I  find info on Shakespeare\'s works might be  answerable asily by a system based on a  well-maintained data base of home pages. Since  our system is based on the processing of the  underlying documents, no correct answer can be  provided if there is no such an answer (explicitly  expressed in English) in the processed  documents. In TREC-8 QA, this is not a problem  since every question is guaranteed to have at least  one answer in the given document pool.  However, in the real world scenario such as a QA  portal, it is conceived that the IE results based on  the processing of the documents hould be  complemented by other knowledge sources uch  as e-copy of yellow pages or other manually  maintained and updated ata bases.  The third limitation is the lack of linguistic  processing such as sentence-level parsing and  cross-sentential co-reference (CO). This problem  will be gradually solved when high-level IE  technology is introduced into the system.  4 Future Work: Multi-level IE Supported QA  A new QA architecture is under development; i   will exploit all levels of the IE system, including  CE and GE.  The first issue is how much CE can  contribute to a better support of QA. It is found  that there are some frequently seen questions  which can be better answered once the CE  information is provided. These questions are of  two types: (i) what/who questions about an NE;  (ii) relationship questions.  Questions of the following format require CE  templates as best answers: who/what is NE? For  example, Who is Julian Hill? Who is Bill  Clinton? What is Du Pont? What is Cymfony?  To answer these questions, the system can simply  1 For example, How did one make a chocolate cake?  How+Adjective questions (e.g. how long, how big,  howold, etc.) are handled fairly well.  retrieve the corresponding CE template to  provide an "assembled" answer, as shown below.  Q: Who is Julian Hill?  A: name: Julian Werner Hill  type: PERSON  age: 91  gender: MALE  position: research chemist  affiliation: Du Pont Co.  education: Washington University;  MIT  Q: What is Du Pont?  A: name: Du Pont Co,  type: COMPANY  staff: Julian Hill; Wallace Carothers.  Questions specifically about a CE  relationship include: For which company did  Julian Hill work? (affiliation relationship) Who  are employees of Du Pont Co.? (staff  relationship) What does Julian Hill do?  (position/profession relationship) Which  university did Julian Hill graduate from?  (education relationship), etc. 2  The next issue is the relationships between  GE and QA. It is our belief that the GE  technology will result in a breakthrough for QA.  In order to extract GE templates, the text  goes through a series of linguistic processing as  shown in Figure 1. It should be noted that the  question processing is designed to go through  parallel processes and share the same NLP  resources until the point of matching and ranking.  The merging of question templates and GE  templates in Template Matcher are fairly  straightforward. As they both undergo the same  NLP processing, the resulting semantic templates  are of the same form. Both question templates  and GE templates correspond to fairly  standard/predictable patterns (the PREDICATE  value is open-ended, but the structure remains  stable). More precisely, a user can ask questions  on general events themselves (did what) and/or  on the participants of the event (who, whom,  what) and/or the time, frequency and place of  events (when, how often, where). This addresses  2 An alpha version of TextractQA supported by both  NE and CE has been implemented and is being tested.  171  by far the most types of general questions of a  potential user.  For example, if a user is interested in  company acquisition events, he can ask questions  like: Which companies ware acquired by  Microsoft in 1999? Which companies did  Microsoft acquire in 1999? Our system will then  parse these questions into the templates as shown  below:  \\[31\\] <Q_TEMPLATE> :=  PREDICATE: acquire  ARGUMENT1: Microsoft  ARGUMENT2: WHAT(COMPANY)  TIME: 1999  If the user wants to know when some  acquisition happened, he can ask: When was  Netscape acquired? Our system will then  translate it into the pattern below:  \\[32\\] <QTEMPLATE> :=  PREDICATE: acquire  ARGUMENT1: WHO  ARGUMENT2: Netscape  TIME: WHEN  Note that WHO, WHAT, WHEN above are  variable to be instantiated. Such question  templates serve as search constraints o filter the  events in our extracted GE template database.  Because the question templates and the extracted  GE template share the same structure, asimple  merging operation would suffice. Nevertheless,  there are two important questions to be answered:  (i) what if a different verb with the same meaning  is used in the question from the one used in the  processed text? (ii) what if the question asks  about something beyond the GE (or CE)  information? These are issues that we are  currently researching. '
0.0	C02-1042	C94-1036	b'1.   Introduction  The TREC evaluations of QA systems  (Voorhees, 1999) require answers to be drawn  from a given source corpus.  Early QA systems  used a simple filtering technique, question word  density within a fixed n-word window, to  pinpoint answers.  Robust though this may be,  the window method is not accurate enough.  In  response, factoid question answering systems  have evolved into two types:   ? Use-Knowledge: extract query words from  the input question, perform IR against the  source corpus, possibly segment resulting  documents, identify a set of segments  containing likely answers, apply a set of  heuristics that each consults a different  source of knowledge to score each  candidate, rank them, and select the best  (Harabagiu et al, 2001; Hovy et al, 2001;  Srihari and Li, 2000; Abney et al, 2000).   ? Use-the-Web: extract query words from the  question, perform IR against the web,  extract likely answer-bearing sentences,  canonicalize the results, and select the most  frequent answer(s).  Then, for justification,  locate examples of the answers in the source  corpus (Brill et al, 2001; Buchholz, 2001).   Of course, these techniques can be combined:  the popularity ratings from Use-the-Web can  also be applied as a filtering criterion (Clarke et  al., 2001), or the knowledge resource heuristics  can filter the web results.  However, simply  going to the web without using further  knowledge (Brill et al, 2001) may return the  web?s majority opinions on astrology, the killers  of JFK, the cancerous effects of microwave  ovens, etc.?fun but not altogether trustworthy.    In this paper we describe the range of  filtering techniques our system Webclopedia  applies, from simplest to most sophisticated, and  indicate their impact on the system.    2.   Webclopedia Architecture   As shown in Figure 1, Webclopedia adopts the  Use-Knowledge architecture. Its modules are  described in more detail in (Hovy et al, 2001;  Hovy et al, 1999):   ? Question parsing: Using BBN?s  IdentiFinder (Bikel et al, 1999), the  CONTEX parser (Hermjakob, 1997)  produces a syntactic-semantic analysis of  the question and determines the QA type.    ? Query formation: Single- and multi-word  units (content words) are extracted from the  analysis, and WordNet synsets (Fellbaum,  1998) are used for query expansion.  A  series of Boolean queries of decreasing  specificity is formed.   ? IR: The publicly available IR engine MG  (Witten et al, 1994) returns the top-ranked  N documents.    ? Selecting and ranking sentences: For each  document, the most promising K sentences  are located and scored using a formula that    rewards word and phrase overlap with the  question and its expanded query words.  Results  are ranked.    ? Parsing candidates: CONTEXparses the  top-ranked 300 sentences.    ? Pinpointing: As described in Section 3, a  number of knowledge resources are used to  perform filtering/pinpointing operations.    ? Ranking of answers: The candidate  answers? scores are compared and the  winner(s) are output.  3. Knowledge Used for Pinpointing  3.1   Type 1: Question Word Matching  Unlike (Prager et al, 1999), we do not first  annotate the source corpus, but perform IR  directly on the source text, using MG (Witten et  al., 1994).  To determine goodness, we assign an  initial base score to each retrieved sentence.  We  then compare the sentence to the question and  adapt this score as follows:   ? exact matches of proper names double the  base score.  ? matching an upper-cased term adds a 60%  bonus of the base score for multi-words  terms and 30% for single words (matching  ?United States? is better than just ?United?).   ? matching a WordNet synonym of a term  discounts by 10% (lower case) and 50%  (upper case).  (When ?Cage? matches  ?cage?, the former may be the last name of a  person and the latter an object; the case  mismatch signals less reliability.)   ? lower-case term matches after Porter  stemming are discounted 30%; upper-case  matches 70% (Porter stemming is more  aggressive than WordNet stemming).   ? Porter stemmer matches of both question  and sentence words with lower case are  discounted 60%; with upper case, 80%.   ? if CONTEX indicates a term as being  qsubsumed (see Section 3.9) the term is  discouned 90% (in ?Which country  manufactures weapons of mass  destruction??, ?country? will be marked as  qsubsumed).    The top-scoring 300 sentences are passed on for  further filtering.    3.2  Type 2: Qtargets, the QA Typology,  and the Semantic Ontology  We classify desired answers by their semantic  type, which have been taxonomized in the  Webclopedia QA Typology (Hovy et al, 2002),  Candidate answer parsing ? Steps: parse sentences ? Engines: CONTEX Matching ? Steps: match general constraint patterns against parse trees              match desired semantic type against parse tree elements              assign score to words in sliding window ? Engine: Matcher Ranking and answer extraction ? Steps: rank candidate answers              extract and format them ? Engine: Answer ranker/formatter QA typology ? QA types, categorized in taxonomy Constraint patterns ? Identify likely answers in relation to    other parts of the sentence Create query Retrieve documents Select & rank sentences Parse top sentences Parse question Input question Perform additional inference Rank and prepare answers Output answers Question parsing ?Steps: parse question              find desired semantic type ? Engines:  IdentiFinder  (BBN)                  CONTEX Match sentences against answers Query creation ?  Steps: extract, combine important words  expand query words using WordNet  create queries, order by specificity ?  Engines: Query creator IR ?  Steps: retrieve top 1000 documents ?  Engines: MG (RMIT Melbourne) Sentence selection and ranking ?  Steps: score each sentence in each document  rank sentences and pass top 300 along ?  Engines:Ranker Figure 1. Webclopedia architecture.   http://www.isi.edu/natural-language/projects/we  bclopedia/Taxonomy/taxonomy_toplevel.html).  The currently approx. 180 classes,  which we  call qtargets, were developed after an analysis of  over 17,000 questions (downloaded in 1999  from answers.com) and later enhancements to  Webclopedia.  They are of several types:   ? common semantic classes such as PROPERPERSON, EMAIL-ADDRESS, LOCATION,  PROPER-ORGANIZATION;   ? classes particular to QA such as YES:NO,  ABBREVIATION-EXPANSION, and WHYFAMOUS;   ? syntactic classes such as NP and NOUN,  when no semnatic type can be determined  (e.g., ?What does Peugeot manufacture??);   ? roles and slots, such as REASON and TITLEP respectively, to indicate a desired relation  with an anchoring concept.     Given a question, the CONTEX parser uses a  set of 276 hand-built rules to identify its most  likely qtarget(s), and records them in a backoff  scheme (allowing more general qtarget nodes to  apply when more specific ones fail to find a  match).  The generalizations are captured in a  typical concept ontology, a 10,000-node extract  of WordNet.    The recursive part of pattern matching is  driven mostly by interrogative phrases.  For  example, the rule that determines the  applicability of the qtarget WHY-FAMOUS  requires the question word ?who?, followed by  the copula, followed by a proper name.  When  there is no match at the current level, the system  examines any interrogative constituent, or words  in special relations to it.  For example, the  qtarget TEMPERATURE-QUANTITY (as in  ?What is the melting point of X?? requires as  syntactic object something that in the ontology is  subordinate to TEMP-QUANTIFIABLE-ABSTRACT with, as well, the word ?how? paired  with ?warm?, ?cold?, ?hot?, etc., or the phrase   ?how many degrees? and a TEMPERATUREUNIT (as defined in the ontology).    3.3 Type 3: Surface Pattern Matching  Often qtarget answers are expressed using rather  stereotypical words or phrases.  For example, the  year of birth of a person is typically expressed  using one of these phrases:   <name> was born in <birthyear>  <name> (<birthyear>?<deathyear>)  We have developed a method to learn such  patterns automatically from text on the web  (Ravichandran and Hovy, 2002).  Wehave  added into the QA Typology the patterns for  appropriate qtargets (qtargets with closed-list  answers, such as PLANETS, require no patterns).   Where some QA systems use such patterns  exclusively (Soubbotin and Soubbotin, 2001) or  partially (Wang et al, 2001; Lee et al, 2001),  we employ them as an additional source of  evidence for the answer.  Preliminary results on  for a range of qtargets, using the TREC-10  questions and the TREC corpus, are:   Question type  (qtarget)  Number of  questions  MRR on  TREC docs  BIRTHYEAR 8 0.47875  INVENTORS 6 0.16667  DISCOVERERS 4 0.1250  DEFINITIONS 102 0.3445  WHY-FAMOUS 3 0.6666  LOCATIONS 16 0.75  3.4  Type 4: Expected Numerical Ranges   Quantity-targeting questions are often  underspecified and rely on culturally shared   cooperativeness rules and/or world knowledge:  Q: How many people live in Chile?   S1: ?From our correspondent comes good  news about the nine people living in  Chile??  A1: nine   While certainly nine people do live in Chile,  we know what the questioner intends.  We have  hand-implemented a rule that provides default  range assumptions for POPULATION questions  and biases quantity questions accordingly.   3.5 Type 5: Abbreviation Expansion   Abbreviations often follow a pattern:  Q: What does NAFTA stand for?  S1: ?This range of topics includes the North  American Free Trade Agreement, NAFTA,  and the world trade agreement GATT.?   S2: ?The interview now changed to the subject  of trade and pending economic issues, such as  the issue of opening the rice market, NAFTA,  and the issue of Russia repaying economic  cooperation funds.?   After Webclopedia identifies the qtarget as  ABBREVIATION-EXPANSION, it extracts   possible answer candidates, including ?North  American Free Trade Agreement? from S1 and  ?the rice market? from S2.  Rules for acronym  matching easily prefer the former.   3.6 Type 6: Semantic Type Matching   Phone numbers, zip codes, email addresses,  URLs, and different types of quantities obey  lexicographic patterns that can be exploited for  matching, as in   Q: What is the zip code for Fremont, CA?   S1: ??from Everex Systems Inc., 48431  Milmont Drive, Fremont, CA 94538.?   and   Q: How hot is the core of the earth?   S1. ?The temperature of Earth?s inner core  may be as high as 9,000 degrees Fahrenheit  (5,000 degrees Celsius).?   Webclopedia identifies the qtargets respectively  as ZIP-CODE and TEMPERATURE-QUANTITY.   Approx. 30 heuristics (cascaded) apply to the  input before parsing to mark up numbers and  other orthographically recognizable units of all  kinds, including (likely) zip codes, quotations,  year ranges, phone numbers, dates, times,  scores, cardinal and ordinal numbers, etc.   Similar work is reported in (Kwok et al, 2001).  3.7 Type 7: Definitions from WordNet   We have found a 10% increase in accuracy in  answering definition questions by using external  glosses obtained from WordNet.  For   Q: What is the Milky Way?   Webclopedia identified two leading answer  candidates:    A1: outer regions   A2: the galaxy that contains the Earth   Comparing these with the WordNet gloss:   WordNet: ?Milky Way?the galaxy containing  the solar system?   allows Webclopedia to straightforwardly match  the candidate with the greater word overlap.    Curiously, the system also needs to use  WordNet to answer questions involving  common knowledge, as in:   Q: What is the capital of the United States?   because authors of the TREC collection do not  find it necessary to explain what Washington is:   Ex: ?Later in the day, the president returned to  Washington, the capital of the United States.?   While WordNet?s definition   Wordnet: ?Washington?the capital of the  United States?   directly provides the answer to the matcher, it  also allows the IR module to focus its search on  passages containing ?Washington?, ?capital?,  and ?United States?, and the matcher to pick a  good motivating passage in the source corpus.    Clearly, this capability can be extended to  include (definitional and other) information  provided by other sources, including  encyclopedias and the web (Lin 2002).  3.8 Type 8: Semantic Relation Matching   So far, we have considered individual words and  groups of words.  But often this is insufficient to  accurately score an answer.  As also noted in  (Buchholz, 2001), pinpointing can be improved  significantly by matching semantic relations  among constituents:   Q: Who killed Lee Harvey Oswald?   Qtargets: PROPER-PERSON & PROPER-NAME,  PROPER-ORGANIZATION   S1: ?Belli?s clients have included Jack Ruby,  who killed John F. Kennedy assassin Lee  Harvey Oswald, and Jim and Tammy Bakker.?   S2: ?On Nov. 22, 1963, the building gained  national notoriety when Lee Harvey Oswald  allegedly shot and killed President John F.  Kennedy from a sixth floor window as the  presidential motorcade passed.?   The CONTEX parser (Hermjakob, 1997;  2001) provides the semantic relations.  The  parser uses machine learning techniques to build  a robust grammar that produces semantically  annotated syntax parses of English (and Korean  and Chinese) sentences at approx. 90% accuracy  (Hermjakob, 1999).    The matcher compares the parse trees of S1  and S2 to that of the question.  Both S1 and S2  receive credit for matching question words ?Lee  Harvey Oswald? and ?kill? (underlined), as well  as for finding an answer (bold) of the proper  qtarget type (PROPER-PERSON).  However, is  the answer ?Jack Ruby? or ?President John F.  Kennedy??  The only way todetermine this is to  consider the semantic relationship between these   candidates and the verb ?kill? (parse trees  simplified, and only portions shown here):      [1] Who killed Lee Harvey Oswald?  [S-SNT]      (SUBJ) [2] Who  [S-INTERR-NP]          (PRED) [3] Who  [S-INTERR-PRON]      (PRED) [4] killed  [S-TR-VERB]      (OBJ) [5] Lee Harvey Oswald  [S-NP]          (PRED) [6] Lee?Oswald  [S-PROPER-NAME]              (MOD) [7] Lee  [S-PROPER-NAME]              (MOD) [8] Harvey  [S-PROPER-NAME]              (PRED) [9] Oswald  [S-PROPER-NAME]      (DUMMY) [10] ?  [D-QUESTION-MARK]    [1] Jack Ruby, who killed John F. Kennedy assassin     Lee Harvey Oswald  [S-NP]     (PRED) [2] <Jack Ruby>1  [S-NP]     (DUMMY) [6] ,  [D-COMMA]     (MOD) [7] who killed John F. Kennedy assassin                    Lee Harvey Oswald  [S-REL-CLAUSE]       (SUBJ) [8] who<1>  [S-INTERR-NP]       (PRED) [10] killed  [S-TR-VERB]       (OBJ) [11] JFK assassin?Oswald  [S-NP]           (PRED) [12] JFK?Oswald [S-PROP-NAME]               (MOD) [13] JFK  [S-PROPER-NAME]               (MOD) [19] assassin  [S-NOUN]               (PRED) [20] ?Oswald [S-PROPER-NAME]  Although the PREDs of both S1 and S2  match that of the question ?killed?, only S1  matches ?Lee Harvey Oswald? as the head of  the logical OBJect.  Thus for S1, the matcher  awards additional credit to node [2] (Jack Ruby)  for being the logical SUBJect of the killing  (using anaphora resolution). In S2, the parse tree  correctly records that node [13] (?John F.  Kennedy?) is not the object of the killing.  Thus  despite its being closer to ?killed?, the candidate  in S2 receives no extra credit from semantic  relation matching.    It is important to note that the matcher  awards extra credit for each matching semantic  relationship between two constituents, not only  when everything matches.  This granularity  improves robustness in the case of partial  matches.    Semantic relation matching applies not only  to logical subjects and objects, but also to all  otherroles such as location, time, reason, etc.  (for additional examples see http://www.isi.edu/  natural-language/projects/webclopedia/sem-relexamples.html).  It also applies at not only the  sentential level, but at all levels, such as postmodifying prepositional and pre-modifying  determiner phrases   Additionally, Webclopedia uses 10 lists of  word variations with a total of 4029 entries for  semantically related concepts such as ?to  invent?, ?invention? and ?inventor?, and rules  for handling them.  For example, via coercing  ?invention? to ?invent?, the system can give  ?Johan Vaaler? extra credit for being a likely  logical subject of ?invention?:   Q: Who invented the paper clip?   Qtargets: PROPER-PERSON & PROPER-NAME,  PROPER-ORGANIZATION   S1: ?The paper clip, weighing a desk-crushing  1,320 pounds, is a faithful copy of Norwegian  Johan Vaaler?s 1899 invention, said Per  Langaker of the Norwegian School of  Management.?   while ?David? actually loses points for being  outside of the clausal scope of the inventing:   S2: ??Like the guy who invented the safety pin,  or the guy who invented the paper clip,? David  added.?   3.9 Type 9: Word Window Scoring   Webclopedia also includes a typical windowbased scoring module that moves a window over  the text and assigns a score to each window  position depending on a variety of criteria (Hovy  et al, 1999).  Unlike (Clarke et al, 2001; Lee et  al., 2001; Chen et al, 2001), we have not  developed a very sophisticated scoring function,  preferring to focus on the modules that employ  information deeper than the word level.   This method is applied only when no other  method provides a sufficiently high-scoring  answer.  The window scoring function is   S  = (500/(500+w))*(1/r) * ?[(?I1.5*q*e*b*u)1.5]  Factors:  w: window width (modulated by gaps of  various lengths: ?white house? ? ?white car and  house?),  r: rank of qtarget in list returned by  CONTEX,  I: window word information content (inverse  log frequency score of each word), summed,   q: # different question words matched, plus  specific rewards (bonus q=3.0),   e: penalty if word matches one of question  word?s WordNet synset items (e=0.8),    b: bonus for matching main verb, proper  names, certain target words (b=2.0),   u: (value 0 or 1) indicates whether a word has  been qsubsumed (?subsumed? by the qtarget)  and should not contribute (again) to the score.   For example, ?In what year did Columbus  discover America?? the qsubsumed words are  ?what? and ?year?.  4. Performance Evaluation   In TREC-10?s QA track, Webclopedia received  an overall Mean Reciprocal Rank (MRR) score  of 0.435, which put it among the top 4  performers of the 68 entrants (the average MRR  score for the main QA task was about 0.234).   The pinpointing heuristics are fairly accurate:  whenWebclopedia finds answers, it usually  ranks them in the first place (1st place: 35.5%;  2nd: 8.94%; 3rd: 5.69%; 4th: 3.05%; 5th: 5.28%;  not found: 41.87%).   We determined the impact of each  knowledge source on system performance, using  the TREC-10 test corpus using the standard  MRR scoring.  We applied the system to the  questions of each knowledge type separately,  with and without its specific knowledge  source/algorithm.  Results are shown in Table 1,  columns A (without) and B (with).  To indicate  overall effect, we also show (in columns C and  D) the percentage of questions in TREC-10 and  -9 respecively of each knowledge type.   5. Conclusions  It is tempting to search for a single technique  that will solve the whole problem (for example,  Ittycheriah et al (2001) focus on the subset of  factoid questions answerable by NPs, and train a  statistical model to perform NP-oriented answer  pinpointing).  Our experience, however, is that  even factoid QA is varied enough to require  various special-purpose techniques and  knowledge.  The theoretical limits of the various  techniques are not known, though Light et al?s  (2001) interesting work begins to study this.    Column A: % questions of the knowledge type        answered correctly without using knowlege  Column B: % questions, now using knowledge  Column C: % questions of type in TREC-10   Column D: % questions of type in TREC-9    A B C D  Abbreviation exp. 20.0 70.0  1.0 2.3  Number ranges 50.0 50.0  1.2 1.8  WordNet (def Qs) 48.3 67.5 20.9 5.1  Semantic types      - locator types N/A N/A  0.0 0.4  - quantity types 22.5 48.7 10.8 5.5  - date/year types 45.0 57.3  9.2 10.2  Patterns       - definitions ? 34.4 20.9 5.1  - why-famous  ? 66.7 0.6 ?  - locations ? 75.0 3.2 ?  - birthyear ? 47.9 1.6 ?  Semantic relations 39.4 46.5 72.2 85.7  Table 1. Performance of knowledge sources.  Semantic relation scores measured only on  questions in which they could logically apply.    We conclude that factoid QA performance  can be significantly improved by the use of  knowledge attuned to specific question types  and specific information characteristics.  Most of  the techniques for exploiting this knowledge  require learning to ensure robustness.  To  improve performance beyond this, we believe a  combination of going to the web and turning to  deeper world knowledge and automated  inference (Harabagiu et al, 2001) to be the  answer.  It remains an open question how much  work these techniques would require, and what  their payoff limits are.   '	b"6 CONCLUSION  This paper shows that this automatic segmenting sys-  tem MSS is quite efficient for segmentation of non-  separated language sentences. MSS do not use any  grammatical information to divide input sentences.  Instead, MSS uses MI l)etween morphenres included in  the input sentence to select the best segmentation(s)  frorn all the possibilities. According to the results  of the experiments, MSS can segment ahnost all the  sentences 'correctly'. This is such a remarkable result.  When it comes to the sentences which are not in the  corpus the ratio of selecting the right segmentation  as the best answer get a little bit lower, however, the  result is considerably good enough.  The result shows that using Mid between mor-  phemes is a very effective method of selecting 'correct'  sentences, aml this means a lot in NLP. "
1.0	W01-1206	A00-1023	b'1 Introduction Answer mining, a.k.a. textual Question/Answering (Q/A), represents the task of discovering the answer to an open-domain natural language question in large text collections. Answer mining became a topic of significant recent interest, partly due to the popularity of Internet Q/A services like AskJeeves and partly due to the recent evaluations of domain-independent Q/A systems organized in the context of the Text REtrieval Conference (TREC)1. The TREC 1The Text REtrieval Conference (TREC) is a series of evaluations of fully automatic Q/A systems specified two restrictions: (1) there is at least one document in the test collection that contains the answer to a test question; and (2) the answer length is either 50 contiguous bytes (short answers) or 250 contiguous bytes (long answers). These two requirements intentionally simplify the answer mining task, since the identification of the exact answer is left to the user. However, given that the expected information is recognized by inspecting text snippets of relatively small size, the TREC Q/A task took a step closer to information retrieval rather than document retrieval. Moreover, the techniques developed to extract text snippets where the answers might lie paved the way to a unified model for answer mining. To find the answer to a question several steps must be taken, as reported in (Abney et al, 2000) (Moldovan et al, 2000) (Srihari and Li, 2000): \x02 First, the question semantics needs to be captured. This translates into identifying (i) the expected answer type and (ii) the question keywords that can be used to retrieve text passages where the answer may be found. \x02 Secondly, the index of the document collection must be used to identify the text passages of interest. The retrieval method either employs special operators or simply modifies boolean or vector retrieval. Since the expected answer type is known at the time workshops organized by the National Institute of Standards and Technology (NIST), designed to advance the state-ofthe-art in information retrieval (IR) of the retrieval, the quality of the text passages is greatly improved by filtering out those passages where concepts of the same category as the answer type are not present. \x02 Thirdly, answer extraction takes place by combining several features that take into account the expected answer type. Since the expected answer type is the only information used in all the phases of textual Q/A, its recognition and usage is central to the performance of answer mining. For an open-domain Q/A system, establishing the possible answer types is a challenging problem. Currently, most of the systems recognize the answer type by associating the question stem (e.g. What, Who, Why or How) and one of the concepts from the question to a predefined general category, such as PERSON, ORGANIZATION, LOCATION, TIME, DATE, MONEY or NUMBER. Since many of these categories are represented in texts as named entities, their recognition as possible answers is enabled by state-of-the-art Named Entity (NE) recognizers, devised to work with high precision in Information Extraction (IE) tasks. To allow for NE-supported answer mining, a large number of semantic categories corresponding to various names must beconsidered, e.g. names of cars, names of diseases, names of dishes, names of boats, etc. Furthermore, a significant number of entities are not unique, therefore do not bear names, but are still potential answers to an opendomain question. Additionally, questions do not focus only on entities and their attributes; they also ask about events and their related entities. In this paper we introduce a model of answer types that accounts for answers to questions of various complexity. The model enables several different formats of the exact answer to opendomain questions and considers also the situation when the answer is produced from a number of different document sources. We define formally the answer types to open-domain questions and extend the recognition of answer types beyond the question processing phase, thus enabling several feed-back mechanisms derived from the processing of documents and answers. The main contribution of the paper is in providing a unified model of answer mining from large collections of on-line documents that accounts for the processing of open-domain natural language questions of varied complexity. The hope is that a coherent model of the textual answer discovery could help developing better text mining methods, capable of acquiring and rapidly prototyping knowledge from the vast amount of on-line texts. Additionally, such a model enables the development of intelligent conversational agents that operate on open-domain tasks. We first present a background of Q/A systems and then define several classes of question complexity. In Section 3 we present the formal answer type model whereas in Section 4 we show how to recognize the answer type of open-domain questions and use it to mine the answer. Section 5 presents the evaluation of the model and summarizes the conclusions. 2 Background Open-Domain Question/Answering To search in a large collection of on-line documents for the answer to a natural language question we need to know (1) what we are looking for, i.e. the expected answer type; and (2) where the answer might be located in the collection. Furthermore, knowing the answer type and recognizing a text passage where the answer might be found is not sufficient for extracting the exact answer. We also need to know the dependencies between the answer type and the other concepts from the question or the answer. For example, if the answer type of the TREC question QT: How many dogs pull a sled in the Iditarod? is known to be a number, we also need to be aware that this number must quantify the dogs harnessed to a sled in the Iditarod games and not the number of participants in the games. Capturing question or answer dependencies can be cast as a straightforward process of mapping syntactic trees to sets of binary headmodifier relationships, as first noted in (Collins, 1996). Given a parse tree, the head-child of each syntactic constituent can be identified based on a simple set of rules used to train syntactic parsers, cf. (Collins, 1996). Dependency relations are established between each leaf corresponding to the head child and the leaves of its constituent sibVP SQ WP PP What domost tourists visit in Question ET1: Reims RBSVBP NNS VB IN NNP NPNPWHNP SBARQParse: What visit Reimstouristsmost Question Dependecies (ET1): What do most tourists visit in Reims? (a) (b) Figure 1: Example of TREC test question lings that are not stop words, as illustrated by the mapping of Figure 1(a) into Figure 1(b). Unlike in IR systems, question stems are considered content words. When question dependencies are known (Harabagiu et al, 2000) proposed a technique of identifying the answer type based on the semantic category of the question stem and eventually of its most connected dependent concept. For example, in the case of question ET1, illustrated in Figure 1, the answer type is determined by the ambiguous question stem what and the verb visit. The answer type is the object of the verb visit, which is a place of attraction or entertainment, defined by the semantic category LANDMARK. The answer type replaces the question stem, generating the following dependency graph, that can be later unified with the answer dependency graph: mostLANDMARK tourists visit Reims However syntactic dependencies vary across question reformulations or equivalent answers made possible by the productive nature of natural language. For example, the dependency structure of ET2, a reformulation of question ET1 differs from the dependency structure of ET1: Due to the fact that verbs see and visit are synonyms (cf. WordNet (Miller, 1995)) and pronoun I can be read a possible visitor, the dependency Question ET2: ILANDMARK Reimssee What could I see in Reims? structures of ET1 and ET2 can be mapped one into another. The mapping is produced by unifying the two structures when lexical and semantic alternations are allowed. Possible lexical alternations are synonyms or morphological alternations. Semantic alternations consist of hypernyms, entailments or paraphrases. The unifying mapping of ET1 and ET2 shows that the two questions are equivalent only when I refers to a visitor; other readings of ET2 being possible when the referent is an investigator or a politician. In each of the other readings, the answer type of the question would be different. The unifying mapping of ET1 and ET2 is: ReimsI -->tourists see/visitLANDMARK Similarly, a pair of equivalent answers is recognized when lexical and semantic alternations of the concepts are allowed. This observation is crucial for answer mining because: 1. it establishes the dependency relations as the basic processing level for Q/A; and 2. it defines the search space based on alternations of the question and answer concepts. Consequently, lexical and semantic alternations are incorporated as feedback loops in the architecture of open-domain Q/A systems, as illustrated in Figure 2. To locate answers, text passages are retrieved based on keywords assembled from the question dependency structure. At the time of the query, it is unknown which keywords can be unified with answer dependencies. However, the relevance of the query is determined by the number of resulting passages. If too many passages are generated, the query was too broad, thus is needs a specialization by adding a new keyword. If too few passages were retrieved the query was too specific, thus one keywordneeds to be dropped. The relevance feedback based on the number of retrieved passages ends when no more keywords can be added or dropped. After this, the unifications of the question and answer dependencies Answer Dependencies KB KB Answer Type Keywords Index Relevance Feedback (# Passages) Lexical Alternations Semantic Unifications Abductive Justification Question On-line Documents Answer Dependencies Question Text Passages Semantic Alternations Answer Fusion Lexico Semantic Figure 2: A diagram of the feedbacks supporting Open-Domain Q/A is produced and the lexical alternations imposed by unifications are added to the list of keywords, making possible the retrieval of new, unseen text passages, as illustrated in Figure 2. The unification of dependency structures allows erroneous answers when the resulting mapping is a sparse graph. To justify the correctness of the answer an abductive proof backchaining from the answer to the question must be produced. Such abductive mechanisms are detailed in (Harabagiu et al, 2000). Moreover, the proof relies on lexico-semantic knowledge available from WordNet as well as rapidly formated knowledge bases generated by mechanisms described in (Chaudri et al, 2000). The justification process brings forward semantic alternations that are added to the list of keywords, the feedback destination of all loops represented in Figure 2. Mining the exact answer does not always end after extracting the answer type from a correct text snippet because often they result only in partial answers that need to be fused together. The fusion mechanisms are dictated by the answer type. Question Complexity Open-Domain natural language questions can also be of different complexity levels. Generally, the test questions used in the TREC evaluations were qualified as fact-based questions (cf. (Voorhees and Tice, 2000)) as they mainly were short inquiries about attributes or definitions of some entity or event. Table 1 lists a sample of TREC test questions. The TREC test set did not include any question Where is Romania located? Europe Who wrote ?Dubliners?? James Joyce What is the wingspan of a condor? 9 feet What is the population of Japan? 120 million What king signed the Magna Carta? King John Name a flying mammal. bat Table 1: TREC test questions and their exact answers (boldfaced) that can be modeled as Information Extraction (IE) task. Typically, IE templates model queries regarding who did an event of interest, what was produced by that event, when and where and eventually why. The event of interest is a complex event, like terrorism in Latin America, joint ventures or management successions. An example of template-modeled question is: What management successions occurred at IBM in 1999? In addition, questions may also ask about developments of events or trends that are usually answered by a text summary. Since data producing these summaries can be sourced in different documents, summary fusion techniques as proposed in (Radev and McKeown, 1998) can be employed. Template-based questions and summaryasking inquiries cover most of the classes of question complexity proposed in (Moldovan et al, 2000). Although the topic of natural language open-domain question complexity needs further study, we consider herein the following classes of questions: \x02 Class 1: Questions inquiringabout entities, events, entity attributes (including number), event themes, event manners, event conditions and event consequences. \x02 Class 2: Questions modeled by templates, including questions that focus only on one of the template slots (e.g. ?What managers were promoted last year at Microsoft??). \x02 Class 3: Questions asking for a summary that is produced by fusing templatebased information from different sources (e.g. ?What happened after the Titanic sunk??). Since (Radev and McKeown, 1998) describes the summary fusion mechanisms, Class 3 of questions can be reduced in this paper to Class 2, which deals with the processing of the template. 3 A Model of Answer Types This section describes a knowledge-based model of open-domain natural language answer types (ATs). In particular we formally define the answer type through a quadruple \x00\x02\x01\x04\x03\x06\x05 CATEGORY, DEPENDENCY, NUMBER, FORMAT \x07 . The CATEGORY is defined as one of the following possibilities: 1. one of the tops of a predefined ANSWER TAXONOMY or one of its nodes; 2. DEFINITION; 3. TEMPLATE; or 4. SUMMARY. For expert Q/A systems, this list of categories can be extended. The DEPENDENCY is defined as the question dependency structure when the CATEGORY belongs to the ANSWER TAXONOMY or is a DEFINITION. Otherwise it is a template automatically generated. The NUMBER is a flag indicating whether the answer should contain a single datum or a list of elements. The FORMAT defines the text span of the exact answer. For example, if the CATEGORY is DIMENSION, the FORMAT is \x08 Number   \x08 Measuring Unit   . The ANSWER TAXONOMY was created in three steps: Step 1 We devise a set of top categories modeled after the semantic domains encoded in the WordNet database, which contains 25 noun categories and 15 verb categories. The top of each WordNet hierarchy corresponding to every semantic category was manually inspected to select the most representative nodes and add them to the tops of he ANSWER TAXONOMY. Furthermore we have added open semantic categories corresponding to named entities. For example Table 2 lists the named entity categories we have considered in our experiments. Many of the tops of the ANSWER TAXONOMY are further categorized, as illustrated in Figure 3. In total, we have considered 33 concepts as tops of the taxonomy. D ATER URATIOND ERCENTAGEP OUNTCEGREED VALUEUMERICALN OCATIONL COUNTRY IMENSION TOWN PROVINCE O THER LOCATION Figure 3: Two examples of top answer hierarchies. Step 2 The additional categorization of the top ANSWER TAXONOMY generates a many-tomany mapping of the Named Entity categories in the tops of the ANSWER TAXONOMY. Figure 4 illustrates some of the mappings. date time organization city product price country money human disease phone number continent percent province other location plant mammal alphabet airport code game bird reptile university dog breed number quantity landmark dish Table 2: Named Entity Categories. ERSON MONEY PEEDS DURATION MOUNTA A NSWER TYPE ENTYTY P number N AMED CATEGORY human money price quantity Figure 4: Mappings of answer types in named entity categories. Step 3: Each leaf from the top of the ANSWER TAXONOMY is connected to one or several WordNALUEV dew point temperature body temperature zero absolute perimeter, girth size largeness bigness circumference distance, length wingspread wingspan, light time altitude duration, UMERICAL length N longevity longness ATIONALITY of the trip from...? What is the duration of an active volcano get? EGREED TEMPERATURE DURATION COUNT SPEED DIMENSION OCATIONL What is the wingspan of a condor? in diameter? How big is our galaxyHow hot does the inside Figure 5: Fragment of the ANSWER TAXONOMY. Net subherarchies. Figure 5 illustrates a fragment of the ANSWER TAXONOMY comprising several WordNet subhierarchies. 4 Answer Recognition and Extraction In this section we show how, given a question and its dependency structure, we can recognize its answer type and consequently extract the exact answer. Here we describe four representative cases. Case 1: The CATEGORY of the answer type is DEFINITION when the question can be matched by one of the following patterns: (Q-P1):What \x00 is \x01 are \x02\x04\x03 phrase to define \x05 ? (Q-P2):What is the definition of \x03 phrase to define \x05 ? (Q-P3):Who \x00 is \x01was \x01 are \x01were \x02\x06\x03 person name(s) \x05 ? The format of the DEFINITION answers is similarly dependent on a set of patterns, determined as the head of the \x08 Answer phrase   : (A-P1): \x07 \x03 phrase to define \x05 \x00 is \x01 are \x02 \x08 \x03 Answer phrase \x05 (A-P2): \x07 \x03 phrase to define \x05 , \x00 a \x01 the \x01 an \x03 Answer phrase \x05\x04\x02 \x08 (A-P3): \x07 \x03 phrase to define \x05 ? \x08 \x03 Answer phrase \x05 Case 2: The dependency structure of the question indicates that a special instance of a concept is sought. The cues are given either by the presence of words kind, type, name or by the question stems what or which connected to the object of a verb. Table 3 lists a set of such questions and their corresponding answers. In this case the answer type is given by the subhierarchy defined by the node from the dependency structure whose adjunct is either kind, type, name or the question stem. In this situation the CATEGORY does not belong to the top of the ANSWER TAXONOMY, but it is rather dynamically created by the interpretation of the dependency graph. For example, the dynamic CATEGORY bridge, generated for Q204 from Table 3, contains 14 member instances, including viaduct, rope bridge and suspension bridge. Similarly, question Q581 generates a dynamic CATEGORY flower, with 470 member instances, comprising orchid, petunia and sunflower. For dynamic categories all member instances are searched in the retrieved passages during answer extraction to detect candidate answers. Case 3: In all other cases, the concept related to the question stem in the question dependency graph is searched through the ANSWER TAXONOMY, returning the answer type as the top of it hierarchy. Figure 5 illustrates several questions and their answer type CATEGORY. Case 4: Whenever the semantic dependencies of several correct answers can be mapped one into another, we change the CATEGORY of the answer type into TEMPLATE. The slots of the actual template are determined by a three step procedure, that weillustrate with a walk-through example corresponding to the question What management successions occurred at IBM in 1999?: Step 1: For each pair of extracted candidate Q204: What type of bridge is the Golden Gate Bridge? Answer: the Seto Ohashi Bridge, consisting of six suspension bridges in the style of Golden Gate Bridge. Q267: What is the name for clouds that produce rain? Answer: Acid rain in Cheju Island and the Taean peninsula is carried by rain clouds from China. Q503: What kind of sports team is the Buffalo Sabres? Answer: Alexander Mogilny hopes to continue his hockey career with the NHL?s Buffalo Sabres. Q581: What flower did Vincent Van Gogh paint? Answer: In March 1987, van Gogh?s ?Sunflowers? sold for $39.9 million at Christie?s in London Table 3: TREC test questions and their answers. The exact answer is emphasized. nominate/assignOrganization (b) PositionPerson Organizationresign/leave PositionPerson (a) Position Person1 Person2 Organization Positionreplace/succeed Person1 Person2 Organization Figure 6: Dependencies that generate templates. answers unify the dependency graphs and find common generalizations whenever possible. Figure 6(a) illustrates some of the mappings. Step 2: Identify across mappings the common categories and the trigger-words that were used as keywords. In Figure 6(a) the trigger words are boldfaced. Step 3: Collect all common categories in a template and use their names as slots. Figure 6(b) illustrates the resulting template. This procedure is a reverse-engineering of the mechanisms used generally in Information Extraction (IE), where given a template, linguistic patterns are acquired to identify the text fragments having relevant information. In the case of answer mining, the relevant text passages are known. The dependency graphs help finding the linguistic rules and are generalized in a template. To be able to generate the template we also need to have a way of extracting the text where the answer dependencies are detected. For this purpose we have designed a method that employs a simple machine learning mechanism: the perceptron. For each text passage retrieved by the keyword-based query we define the following seven features: \x00\x02\x01\x04\x03\x06\x05\x08\x07\x04  the number of question words matched in the same phrase as the answer type CATEGORY; \x00 \x01\x04\x03\x06\x05\x08\x07\x0b\x07 the number of question words matched in the same sentence as the answer type CATEGORY; \x00\x0c\x01\x04\x03\x06\x05\x0e\r\x0f  : a flag set to 1 if the answer type CATEGORY is followed by a punctuation sign, and set to 0 otherwise; \x00\x10\x01\x04\x03\x06\x05\x08\x11\x13\x12\x15\x14\x17\x16 : the number of question words matches separated from the answer type CATEGORY by at most three words and one comma; \x00\x18\x01\x04\x03\x06\x05\x08\x07\x0b\x16\x19\x07 : the number of question words occurring in the same order in the answer text as in the question; \x00\x1a\x01\x04\x03\x06\x05\x08fffi\x14\x17\x16 : the average distance from the answer type CATEGORY to any of the question word matches; \x00fl\x01\x04\x03\x06\x05\x08ffi \x1f \x16 : the number of question words matched in the answer text. To train the perceptron we annotated the correct answers of 200 of the TREC test questions. Given a pair of answers, in which one of the answers is correct, we compute a relative comparison score using the formula: !#"%$ &(\'*) +-,/.fi0\x061204365\x02!#"%$70\x0612098 .fi:\x0b; 365\x02!#"%$7:\x0b; 8<.fi=?>\x04@A1B365\x02!#"%$7=?>\x04@A1\x1a8 .fi0\x06;C365\x02!#"%$70\x06; 8<.fi0\x060\x0c365\x02!#"%$70\x06098 .fiDFE 1G3H5\x02!#"%$7DFE 18<.fiIJ@A1B365\x02!#"%$7IJ@A1K8CLNMO!#"QP*MOR#$ S The perceptron learns the seven weights as well as the value of the threshold used for future tests on the remaining 693 TREC questions. Whenever the relative score is larger than the threshold, a passage is extracted as a candidate answer. In our experiments, the performance of the perceptron surpassed the performance of decision trees for answer extraction.5 Evaluations and Conclusion To evaluate our answer type model we used 693 TREC test questions on which we did not train the perceptron. Table 4 lists the breakdown of the answer type CATEGORIES recognized by our model as well as the coverage and precision of the recognition. Currently our ANSWER TAXONOMY encodes 8707 concepts from 129 WordNet hierarchies, covering only 81% of the expected answer types. This shows that we have to continue encoding more top concepts in the taxonomy and link them to more WordNet concepts. The recognition mechanism had better precision than coverage in our experiments. Moreover a relationship between the coverage of answer type recognition and the overall performance of answer mining, as illustrated in Table 4. Some of the test questions are listed in Tables 1 and 3. The experiments were conducted by using 736,794 on-line documents from Los Angeles Times, Foreign Broadcast Information Service, Financial Times AP Newswire, Wall Street Journal and San Jose Mercury News. CATEGORY (# Questions) Precision Coverage DEFINITION (64) 91% 84% Top ANSWER TAXONOMY (439) 79% 74% Dynamic answer category (17) 86% 79% TEMPLATE (14) 93% 65% # ANSWER Answer Type Q/A Precision Taxonomy Coverage Tops 8 44% 42% 22 56% 55% 33 83% 78% Table 4: Evaluation results. The experiments show that open-domain natural language questions of varied degrees of complexity can be answered consistently from vast amounts of on-line texts. One of the applications of a unified model of answer mining is the development of intelligent conversational agents (Harabagiu et al, 2001).Acknowledgement This research was supported in part by the Advanced Research and Development Activity (ARDA) grant 2001*H238400*000 and by the National Science Foundation CAREER grant CCR-9983600.'	b'Introduction  With the explosion of information in Internet,  Natural language QA is recognized as a  capability with great potential. Traditionally,  QA has attracted many AI researchers, but most  QA systems developed are toy systems or games  confined to lab and a very restricted omain.  More recently, Text Retrieval Conference  (TREC-8) designed a QA track to stimulate the  research for real world application.  Due to little linguistic support from text  analysis, conventional IR systems or search  engines do not really perform the task of  information retrieval; they in fact aim at only  document retrieval. The following quote from the  QA Track Specifications (www.research.att.com/  -singhal/qa-track-spec.txt) in the TREC  community illustrates this point.  Current information retrieval systems allow  us to locate documents hat might contain the  pertinent information, but most of them leave  it to the user to extract he useful information  from a ranked list. This leaves the (often  unwilling) user with a relatively large  amount of text o consume. There is an urgent  need for tools that would reduce the amount  of text one might have to read in order to  obtain the desired information. This track  aims at doing exactly that for a special (and  popular) class of information seeking  behavior: QUESTION ANSWERING. People  have questions and they need answers, not  documents. Automatic question answering  will definitely be a significant advance in the  state-of-art information retrieval technology.  Kupiec (1993) presented a QA system  MURAX using an on-line encyclopedia. This  system used the technology of robust shallow  parsing but suffered from the lack of basic  information extraction support. In fact, the most  siginifcant IE advance, namely the NE (Named  Entity) technology, occured after Kupiec (1993),  thanks to the MUC program (MUC-7 1998).  High-level IE technology beyond NE has not  been in the stage of possible application until  recently.  AskJeeves launched a QA portal  (www.askjeeves.com). It is equipped with a  fairly sophisticated natural language question  parser, but it does not provide direct answers to  the asked questions. Instead, it directs the user to  the relevant web pages, just as the traditional  search engine does. In this sense, AskJeeves has  only done half of the job for QA.  We believe that QA is an ideal test bed for  demonstrating the power of IE. There is a natural  co-operation between IE and IR; we regard QA  as one major intelligence which IE can offer IR.  * This work was supported in part by the SBIR grants F30602-98-C-0043 and F30602-99-C-0102 from Air Force  Research Laboratory (AFRL)/IFED.  166  An important question then is, what type of  IE can support IR in QA and how well does it  support it? This forms the major topic of this  paper. We structure the remaining part of the paper as follows. In Section 1, we first give an  overview of the underlying IE technology which  our organization has been developing. Section 2  discusses the QA system. Section 3 describes the  limitation of the current system. Finally, in  Section 4, we propose a more sophisticated QA  system supported by three levels of IE.  1 Overview of Textract IE  The last decade has seen great advance and  interest in the area of IE. In the US, the DARPA  sponsored Tipster Text Program \\[Grishman  1997\\] and the Message Understanding  Conferences (MUC) \\[MUC-7 1998\\] have been  the driving force for developing this technology.  In fact, the MUC specifications for various IE  tasks have become de facto standards in the IE  research community. It is therefore necessary to  present our IE effort in the context of the MUC  program.  MUC divides IE into distinct tasks,  namely, NE (Named Entity), TE (Template  Element), TR (Template Relation), CO  (Co-reference), and ST (Scenario Templates)  \\[Chinchor & Marsh 1998\\]. Our proposal for  three levels of IE is modelled after the MUC  standards using MUC-style representation.  However, we have modified the MUC IE task  definitions in order to make them more useful  and more practical. More precisely, we propose a hierarchical, 3-level architecture for developing a  kernel IE system which is domain-independent  throughout.  The core of this system is a state-of-the-art  NE tagger \\[Srihari 1998\\], named Textract 1.0.  The Textract NE tagger has achieved speed and  accuracy comparable tothat of the few deployed  NE systems, such as NetOwl \\[Krupka &  Hausman 1998\\] and Nymble \\[Bikel et al1997\\].  It is to be noted that in our definition of NE,  we significantly expanded the type of  information to be extracted. In addition to all the  MUC defined NE types (person, organization,  location, time, date, money and percent), the  following types/sub-types of information are also  identified by the TextractNE module:  ? duration, frequency, age  ? number, fraction, decimal, ordinal, math  equation  ? weight, length, temperature, angle, area,  capacity, speed, rate  ? product, software  ? address, email, phone, fax, telex, www  ? name (default proper name)  Sub-type information like company,  government agency, school (belonging to the  type organization) and military person, religious  person (belonging to person) are also identified.  These new sub-types provide a better foundation  for defining multiple relationships between the  identified entities and for supporting question  answering functionality. For example, the key to  a question processor is to identify the asking  point (who, what, when, where, etc.). In many  cases, the asking point corresponds to an NE  beyond the MUC definition, e.g. the  how+adjective questions: how long (duration or  length), how far (length), how often (frequency),  how old (age), etc.  Level-2 IE, or CE(Correlated Entity), is  concerned with extracting pre-defined multiple  relationships between the entities. Consider the  person entity as an example; the TextractCE  prototype is capable of extracting the key  relationships uch as age, gender, affiliation,  position, birthtime, birth__place, spouse,  parents, children, where.from, address, phone,  fax, email, descriptors. As seen, the information  in the CE represents a mini-CV or profile of the  entity. In general, the CE template integrates and  greatly enriches the information contained in  MUC TE and TR.  The final goal of our IE effort is to further  extract open-ended general events (GE, or level 3  IE) for information like who did what (to whom)  when (or how often) and where. By general  events, we refer to argument structures centering  around verb notions plus the associated  information of time/frequency and location. We  show an example of our defined GE extracted  from the text below:  Julian Hill, a research chemist whose  accidental discovery of a tough, taffylike  compound revolutionized everyday life after  it proved its worth in warfare and courtship,  167  died on Sunday in Hockessin, Del.  \\[1\\] <GE_TEMPLATE> :=  PREDICATE: die  ARGUMENTI: Julian Hill  TIME: Sunday  LOCATION: Hockessin, Del  Figure 1 is the overall system architecture for  the IE system Textract hat our organization has  been developing.  Kernet IE Modutes  L|ngui_sti_cLM_odu!es  I . . . . . . . . . . . . .  I I . . . . . . . . . .  I  I I I I ,,l l I , ! I  I I I I  \'l J \' I I  ! I  I I  I I I I ,i I i \', I ,  I ! I I , l ? i  , ,  I I I I  I I  F L - -  - - - -~  . . . .  . L - -  - -  - -  . - -  - -  - -  | . . . .   Apptication Modutes  NE: NIiml~ EnlilyTitl~klll QA: Que~tlon Answering  CE: Come,led Entity ExtrmClkm BR: In~lllgenl ~ws lng   GE: Gcn~mI Evenl Ex~ct~on AS; Auio~ SUl len   co :  ce-  mfcmnc ~1 momial l  s ~ p~  Figure 1: Textract IE System Architecture  The core of the system consists of three  kernel IE modules and six linguistic modules.  The multi-level linguistic modules erve as an  underlying support system for different levels of  IE. The IE results are stored in a database which  is the basis for IE-related applications like QA,  BR (Browsing, threading and visualization) and  AS (Automatic Summarization). The approach  to IE taken here, consists of a unique blend of  machine learning andFST (finite state  transducer) rule-based system \\[Roche & Schabes  1997\\]. By combining machine learning with an  FST rule-based system, we are able to exploit he  best of both paradigms while overcoming their  respective weaknesses \\[Srihari 1998, Li & Srihari  2000\\].  2 NE-Supported QA  This section presents the QA system based on  Named Entity tagging. Out of the 200 questions  that comprised the TREC-8 QA track  competition, over 80% asked for an NE, e.g. who  (PERSON), when (T IME\\ [  DATE), where  (LOCATION), how far (LENGTH). Therefore,  the NE tagger has been proven to be very helpful.  Of course, the NE of the targeted type is only  necessary but not complete in answering such  questions because NE by nature only extracts  isolated individual entities from the text.  Nevertheless, using even crude methods like "the  nearest NE to the queried key words" or "the NE  and its related key words within the same line (or  same paragraph, etc.)", in most cases, the QA  system was able to extract ext portions which  contained answers in the top five list.  Figure 2 illustrates the system design of  TextractQA Prototype. There are two  components for the QA prototype: Question  Processor and Text Processor. The Text Matcher  module links the two processing results and tries  to find answers to the processed question.  Matching is based on keywords, plus the NE  type and their common location within a same  sentence.  Quest ion  Prc~:essor  i  : :eXt P r~_~ . . . . . . . . . . . .  ? ~   i i ~ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   i . . . .  i  Figure 2: Textract/QA 1.0 Prototype Architecture  The general algorithm for question  answering is as follows:  168  Process Question  Shallow parse question  Determine Asking Point  Question expansion (using word lists)  Process Documents  Tokenization, POS tagging, NE Indexing  Shallow Parsing (not yet utilized)  Text Matcher  Intersect search engine results with NE  rank answers  2.1 Question Processing  The Question Processing results are a list of  keywords plus the information for asking point.  For example, the question:  \\[2\\] Who won the 1998 Nobel Peace Prize?  contains the following keywords: won, 1998,  Nobel, Peace, Prize. The asking point Who refers  to the NE type person. The output before  question expansion is a simple 2-feature template  as shown below:  \\[3\\] asking_point: PERSON  key_word: { won, 1998, Nobel,  Peace, Prize }  The following is an example where the  asking point does not correspond to any type of  NE in ourdefinition.  \\[3\\] Why did David Koresh ask the FBI for a  word processor ?  The system then maps it to the following  question template :  \\[4\\] asking_point:  key_word:  REASON  { ask, David, Koresh,  FBI, word, processor } The question processor scans the question to  search for question words (wh-words) and maps  them into corresponding NE types/sub-types or  pre-defined notions like REASON.  We adopt wo sets of pattern matching rules  for this purpose: (i) structure based pattern  matching rules; (ii) simple key word based  pattern matching rules (regarded as default rules).  It is fairly easy to exhaust the second set of rules  as interrogative question words/phrases form a  closed set. In comparison, the development of the first set of rules are continuously being  fine-tuned and expanded. This strategy of using  two set of rules leads to the robustness of the  question processor.  The first set of rules are based on shallow  parsing results of the questions, using Cymfony  FST based Shallow Parser. This parser identifies  basic syntactic onstructions like BaseNP (Basic  Noun Phrase), BasePP (Basic Prepositional  Phrase) and VG (Verb Group).  The following is a sample of the first set of  rules:  \\[6\\] Name NP (city I country I company) -->  CITYICOUNTRYICOMPANY  \\[7\\] Name NP(person_w) --> PERSON  \\[8\\] Name NP(org_w) --> ORGANIZATION  \\[9\\] Name NP(NOT person_w, NOT org_w)  --> NAME  Rule \\[6\\] checks the head word of the NP. It  covers cases like VG\\[Name\\] NP\\[a country\\] that  VG\\[is developing\\] NP\\[a magnetic levitation  railway system\\]. Rule \\[7\\] works for cases like  VG\\[Name\\] NP\\[the first private citizen\\] VG\\[to  fly\\] PP\\[in space\\] as citizen belongs to the word  class person_w. Rule \\[9\\] is a catch-all rule: if the  NP is not of class person (person_w) or  organization (org_w), then the asking point is a  proper name (default NE), often realized in  English in capitalized string of words. Examples  include Name a film that has won the Golden  Bear in the Berlin Film Festival.  The word lists org_w and person_w are  currently manually maintained based on  inspection of large volumes of text. An effort is  underway to automate the learning of such word  lists by utilizing machine learning techniques.  We used the following pattern  transformations to expand our ruleset:  (Please) name NP\\[X\\]  --> what/which Aux(be) (the name of) NP\\[X\\]  --> NP(what/which...X)  In other words, the four rules are expanded to  12 rules. For example, Rule \\[10\\] below  corresponds to Rule \\[6\\]; Rule \\[11\\] is derived  169  from Rule \\[7\\].  \\[10\\] what/which Aux(be) NP (city \\[ country \\[  company) -->  CITY I COUNTRY \\[ COMPANY  \\[11\\] NP(what/which ... person_w) -->  PERSON  Rule \\[10\\] extracts the asking point from  cases like NP\\[What\\] Aux\\[is\\] NP\\[the largest  country\\] PP\\[inthe world\\]. Rule \\[11\\] covers the  following questions: NP\\[What costume  designer\\] VG\\[decided\\] that NP\\[Michael  Jacksonl VG\\[should only wear\\] NP\\[one glove\\],  NP\\[Which former Ku Klux Klan member\\]  VG\\[won\\] NP\\[an elected office\\] PP\\[in the U.S.\\],  NP\\[What Nobel laureate\\] VG\\[was expelled\\]  PP\\[from the Philippines\\] PP\\[before the  conference\\] PP\\[on East Timor\\], NP\\[What  famous communist leader\\] VG\\[died\\] PP\\[in  Mexico City\\], etc.  As seen, shallow parsing helps us to capture a variety of natural anguage question expressions.  However, there are cases where some simple key  word based pattern matching would be enough to  capture the asking point. That is our second set  of rules. These rules are used when the first set of  rules has failed to produce results. The following  is a sample of such rules:  \\[ 12\\] who/whom --> PERSON  \\[13\\] when --> TIME/DATE  \\[14\\] where/what place --> LOCATION  \\[15\\] what time (of day) --> TIME  \\[16\\] what day (of the week) --> DAY  \\[17\\] what/which month --> MONTH  \\[18\\] what age/how old --> AGE  \\[19\\] what brand --> PRODUCT  \\[20\\] what --> NAME  \\[21\\] how far/tall/high --> LENGTH  \\[22\\] how large/hig/small --> AREA  \\[23\\] how heavy --> WEIGHT  \\[24\\] how rich --> MONEY  \\[25\\] how often --> FREQUENCY  \\[26\\] how many --> NUMBER  \\[27\\] how long --> LENGTH/DURATION  \\[28\\] why/for what --> REASON  In the stage of question expansion, the  template in \\[4\\] would be expanded to the  template shown in \\[29\\]:  \\[29\\] asking_point:  key_word:  {because{because of\\]  due to{thanks to{since I  in order{to VB}  { asklaskslasked\\[asking,  David,Koresh,FBI,  word, processor}  The last item in the asking._point list attempts  to find an infinitive by checking the word to  followed by a verb (with the part-of-speech tag  VB). As we know, infinitive verb phrases are  often used in English to explain a reason for some  action.  2.2 Text Processing  On the text processing side, we first send the  question directly to a search engine in order to  narrow down the document pool to the first n, say  200, documents for IE processing. Currently,  this includes tokenization, POS tagging and NE  tagging. Future plans include several evels of  parsing as well; these are required to support CE  and GE extraction. It should be noted that all  these operations are extremely robust and fast,  features necessary for large volume text  indexing. Parsing is accomplished through  cascaded finite state transducer grammars.  2.3 Text Matching  The Text Matcher attempts to match the question  template with the processed ocuments for both  the asking point and the key words. There is a  preliminary ranking standard built-in the matcher  in order to find the most probable answers. The  primary rank is a count of how many unique  keywords are contained within a sentence.The  secondary ranking is based on the order that the  keywords appear in the sentence compared to  their order in the question. The third ranking is  based on whether there is an exact match or a  variant match for the key verb.  In the TREC-8 QA track competition,  Cymfony QA accuracy was 66.0%. Considering  we have only used NE technology to support QA  in this run, 66.0% is a very encouraging result.  3 Limitation  The first limitation comes from the types of  questions. Currently only wh-questions are  handled although it is planned that yes-no  questions will be handled once we introduce CE  170  and GE templates to support QA. Among the  wh-questions, the why-question and  how-question t are more challenging because the  asking point cannot be simply mapped to the NE  types/sub-types.  The second limitation is from the nature of  the questions. Questions like Where can l find  the homepage for Oscar winners or Where can I  find info on Shakespeare\'s works might be  answerable asily by a system based on a  well-maintained data base of home pages. Since  our system is based on the processing of the  underlying documents, no correct answer can be  provided if there is no such an answer (explicitly  expressed in English) in the processed  documents. In TREC-8 QA, this is not a problem  since every question is guaranteed to have at least  one answer in the given document pool.  However, in the real world scenario such as a QA  portal, it is conceived that the IE results based on  the processing of the documents hould be  complemented by other knowledge sources uch  as e-copy of yellow pages or other manually  maintained and updated ata bases.  The third limitation is the lack of linguistic  processing such as sentence-level parsing and  cross-sentential co-reference (CO). This problem  will be gradually solved when high-level IE  technology is introduced into the system.  4 Future Work: Multi-level IE Supported QA  A new QA architecture is under development; i   will exploit all levels of the IE system, including  CE and GE.  The first issue is how much CE can  contribute to a better support of QA. It is found  that there are some frequently seen questions  which can be better answered once the CE  information is provided. These questions are of  two types: (i) what/who questions about an NE;  (ii) relationship questions.  Questions of the following format require CE  templates as best answers: who/what is NE? For  example, Who is Julian Hill? Who is Bill  Clinton? What is Du Pont? What is Cymfony?  To answer these questions, the system can simply  1 For example, How did one make a chocolate cake?  How+Adjective questions (e.g. how long, how big,  howold, etc.) are handled fairly well.  retrieve the corresponding CE template to  provide an "assembled" answer, as shown below.  Q: Who is Julian Hill?  A: name: Julian Werner Hill  type: PERSON  age: 91  gender: MALE  position: research chemist  affiliation: Du Pont Co.  education: Washington University;  MIT  Q: What is Du Pont?  A: name: Du Pont Co,  type: COMPANY  staff: Julian Hill; Wallace Carothers.  Questions specifically about a CE  relationship include: For which company did  Julian Hill work? (affiliation relationship) Who  are employees of Du Pont Co.? (staff  relationship) What does Julian Hill do?  (position/profession relationship) Which  university did Julian Hill graduate from?  (education relationship), etc. 2  The next issue is the relationships between  GE and QA. It is our belief that the GE  technology will result in a breakthrough for QA.  In order to extract GE templates, the text  goes through a series of linguistic processing as  shown in Figure 1. It should be noted that the  question processing is designed to go through  parallel processes and share the same NLP  resources until the point of matching and ranking.  The merging of question templates and GE  templates in Template Matcher are fairly  straightforward. As they both undergo the same  NLP processing, the resulting semantic templates  are of the same form. Both question templates  and GE templates correspond to fairly  standard/predictable patterns (the PREDICATE  value is open-ended, but the structure remains  stable). More precisely, a user can ask questions  on general events themselves (did what) and/or  on the participants of the event (who, whom,  what) and/or the time, frequency and place of  events (when, how often, where). This addresses  2 An alpha version of TextractQA supported by both  NE and CE has been implemented and is being tested.  171  by far the most types of general questions of a  potential user.  For example, if a user is interested in  company acquisition events, he can ask questions  like: Which companies ware acquired by  Microsoft in 1999? Which companies did  Microsoft acquire in 1999? Our system will then  parse these questions into the templates as shown  below:  \\[31\\] <Q_TEMPLATE> :=  PREDICATE: acquire  ARGUMENT1: Microsoft  ARGUMENT2: WHAT(COMPANY)  TIME: 1999  If the user wants to know when some  acquisition happened, he can ask: When was  Netscape acquired? Our system will then  translate it into the pattern below:  \\[32\\] <QTEMPLATE> :=  PREDICATE: acquire  ARGUMENT1: WHO  ARGUMENT2: Netscape  TIME: WHEN  Note that WHO, WHAT, WHEN above are  variable to be instantiated. Such question  templates serve as search constraints o filter the  events in our extracted GE template database.  Because the question templates and the extracted  GE template share the same structure, asimple  merging operation would suffice. Nevertheless,  there are two important questions to be answered:  (i) what if a different verb with the same meaning  is used in the question from the one used in the  processed text? (ii) what if the question asks  about something beyond the GE (or CE)  information? These are issues that we are  currently researching. '
1.0	D12-1028	A00-1031	b'1 Introduction The true nature of the notion of dependency (after removing sedimentary deposits of rules imposed only by more or less arbitrary conventions) remains still somewhat vague and elusive. This holds in spite of a seemingly strong background intuition and even after a decade of formalized large-scale dependencybased resources being available to the research community. It is undeniable that a huge progress has been reached in the field of supervised dependency parsing, especially due to the CoNLL shared task series. However, when it comes to unsupervised parsing, there are surprisingly few clues we could rely on. As mentioned e.g. by Ku?bler et al2009), one of the traditional linguistic criteria for recognizing dependency relations (including their head-dependent orientation) is that a head H of a construction C determines the syntactic category of C and can often replace C. Or, in words of Dependency Analysis by Reduction (Lopatkova? et al2005), stepwise deletion of dependent elements within a sentence preserves its syntactic correctness. A similar idea of dependency analysis by splitting a sentence into all possible acceptable fragments is used by Gerdes and Kahane (2011). Of course, all the above works had to respond to the notorious fact that there are many language phenomena precluding the ideal (word by word) sentence reducibility (e.g. in the case of prepositional groups, or in the case of subjects in English finite clauses). However, we disregard their solutions tentatively and borrow only the very core of the reducibility idea: if a word can be removed from a sentence without damaging it, then it is likely to be dependent on some other (still present) word. As it is usual with dichotomies in natural languages, it seems more adequate to use a continuous scale instead of the reducible-irreducible opposition. That is why we introduce a simple reducibility measure based on n-gram corpus statistics. We employ this reducibility measure as the main feature in our unsupervised parsing procedure. The procedure is based on a commonly used Bayesian inference technique called Gibbs sampling (Gilks et al1996). In our sampler, the more reducible a given token is, the more likely it is to be sampled as a dependant and not as a head. After certain number of sampling iterations, for each sentence a final dependency tree is created (one token per node, including punctuation) that maximizes the product of edge probabilities gathered along the sampling history. Our approach allows to utilize information from 297 very large corpora. While the computationally demanding sampling procedure can be applied only on limited data, the unrepeated precomputation of statistics for reducibility estimates can easily exploit much larger data. We are not aware of any other published work on unsupervised parsing employing reducibility or a similar idea. Dominating approaches in unsupervised parsing are typically based on repeated patterns, and not on the possibility of a deletion inside a pattern. It seems that the two views of dependency (frequent co-occurrence of head-dependant pair, versus reducibility of the dependant) are rather complementary, so fruitful combinations can be hopefully expected in future. The remainder of this paper is structured as follows.Section 2 briefly outlines the state of the art in unsupervised dependency parsing. Our measure of reducibility based on a large monolingual corpus is presented in Section 3. Section 4 shows our models which serve for generating probability estimates for edge sampling described in Section 5. Experimental parsing results for languages included in CoNLL shared task treebanks are summarized in Section 6. Section 7 concludes this article. 2 Related Work The most popular approach in unsupervised dependency parsing of the recent years is to employ Dependency Model with Valence (DMV), which was introduced by Klein and Manning (2004). The inference algorithm was further improved by Smith (2007) and Cohen et al2008). Headden, Johnson and McClosky (2009) introduced the Extended Valence Grammar (EVG) and added lexicalization and smoothing. Blunsom and Cohn (2010) use tree substitution grammars, which allow learning larger dependency fragments. Unfortunately, many of these works show results only for English.1 However, the main feature of unsupervised methods should be their applicability across a wide range of languages. Such experiments were done by Spitkovsky (2011b; 2011c), where the parsing algorithm was evaluated on all 19 languages included in CoNLL 2006 (Buchholz and 1The state-of-the-art unsupervised parsers achieve more than 50% of attachment score measured on the Penn Treebank. Marsi, 2006) and 2007 (Nivre et al2007) shared tasks. The fully unsupervised linguistic analysis (Spitkovsky et al2011a) shows that the unsupervised part-of-speech tags may be more useful for this task than the supervised ones. Another possibility for obtaining dependency structures for languages without any linguistically annotated resources can be the projection using a parallel treebank with a resource-rich language (typically English). McDonald et al2011) showed that such projection produce better structures than the current unsupervised parsers do. However, our task is different. We would like to produce structures that are not burdened by any linguistic conventions. In this paper, we describe a novel approach to unsupervised dependency parsing. Our model differs from DMV, since we employ the reducibility feature and use fertility of nodes instead of generating STOP signs. We use Gibbs sampling procedure for inference instead of Variational Bayes, which has been more common for induction of linguistic strucures. Gibbs sampling algorithm for grammar induction was used also by Marec?ek and Z?abokrtsky? (2011). However, their sampling algorithm produces generally nonprojective trees. Our sampler, which is described in Section 5, introduces a completely different smallchange operator that guarantees projective edges. 3 Computing Reducibility scores We call a word (or a sequence of words) in a sentence reducible, if the sentence after removing the word remains grammatically correct. Although we cannot automatically recognize grammaticality of such newly created sentence, we can search for it in a large corpus. If we find it, we assume the word was reducible in the original sentence. Since the number of such reducible word sequences found in any corpus will be low, we determine the reducibility scores from their individual types (part-of-speech tags). This then implicitly allows some sharing of the scores between different word sequences. The necessity to search for the whole sentences in the corpus and not onlyfor some smaller context (considering, for example, just left and right neigh298 bor), which would lead to lower sparsity, is rationalized by the following example: Their children went to school. I took their children to school. The verb ?went? would be reducible in the context ?their children went to school?, because the sequence ?their children to school? occurs in the second sentence. One could find such examples frequently even for large contexts. For instance, verbs in free wordorder languages can be placed almost at any position in a sentence; therefore, without the full sentence context, they would have to be considered as reducible. To prevent this, we decided to work exclusively with the full sentence context instead of shorter contexts. Other way that would lead to lower sparsity would be searching for sequences of part-of-speech tags instead of sequences of word forms. However, this also does not bring desired results. For instance, the two following sentence patterns DT NNS VBD IN DT NN . DT NNS VBD DT NN . are quite frequent in English and we can deduce from them that the preposition IN is reducible. But this is of course a wrong deduction, since the preposition cannot be removed from the prepositional phrase. Using part-of-speech tags instead of word forms is thus not suitable for computing reducibility scores. Although we search for reducible sequences of word forms in the corpus, we compute reducibility scores for sequences of part-of-speech tags. This requires to have the corpus morphologically disambiguated. A sequences of part-of-speech tags will be denoted as ?PoS n-gram? in the following text. Assume a PoS n-gram g = [t1, . . . , tn]. We go through the corpus and search for all its occurrences. For each such occurrence, we remove the respective words from the current sentence and check in the corpus whether the rest of the sentence occurs at least once elsewhere in the corpus.2 If so, then such occurrence of PoS n-gram is reducible, otherwise it is not. We denote the number of such reducible oc2We do not take into account sentences with less then 10 words, because they could be nominal (without any verb) and might influence the reducibility scores of verbs. unigrams R bigrams R trigrams R VB 0.04 VBN IN 0.00 IN DT JJ 0.00 TO 0.07 IN DT 0.02 JJ NN IN 0.00 IN 0.11 NN IN 0.04 NN IN NNP 0.00 VBD 0.12 NNS IN 0.05 VBN IN DT 0.00 CC 0.13 JJ NNS 0.07 JJ NN . 0.00 VBZ 0.16 NN . 0.08 DT JJ NN 0.04 NN 0.22 DT NNP 0.09 DT NNP NNP 0.05 VBN 0.24 DT NN 0.09 NNS IN DT 0.14 . 0.32 NN , 0.11 NNP NNP . 0.15 NNS 0.38 DT JJ 0.13 NN IN DT 0.23 DT 0.43 JJ NN 0.14 NNP NNP , 0.46 NNP 0.78 NNP . 0.15 IN DT NNP 0.55 JJ 0.84 NN NN 0.22 DT NN IN 0.59 RB 2.07 IN NN 0.67 NNP NNP NNP 0.64 , 3.77 NNP NNP 0.76 IN DT NN 0.80 CD 55.6 INNNP 1.81 IN NNP NNP 4.27 Table 1: Reducibility scores of the most frequent English n-grams. (V* are verbs, N* are nouns, DET are determiners, IN are prepositions, JJ are adjectives, RB are adverbs, CD are numerals, and CC are coordinating conjunctions) currences of PoS n-gram g by r(g). The number of all its occurrences is c(g). The relative reducibility R(g) of a PoS n-gram g is then computed as R(g) = 1 N r(g) + ?1 c(g) + ?2 , (1) where the normalization constant N , which expresses relative reducibility over all the PoS n-grams (denoted by G), causes the scores are concentrated around the value 1. N = ? g?G(r(g) + ?1) ? g?G(c(g) + ?2) (2) Smoothing constants ?1 and ?2, which prevent reducibility scores from being equal to zero, are set to ?1 = ? g?G r(g) ? g?G c(g) , ?2 = 1 (3) This setting causes that even if a given PoS n-gram is not reducible anywhere in the corpus, its reducibility score is 1/(c(g) + 1). Tables 1, 2, and 3 show reducibility scores of the most frequent PoS n-grams of three selected languages: English, German, and Czech. If we consider only unigrams, we can see that the scores for verbs are often among the lowest. Verbs are followed by prepositions and nouns, and the scores for adjectives 299 unigrams R bigrams R trigrams R VVPP 0.00 NN APPR 0.00 NN APPR NN 0.01 APPR 0.27 APPR ART 0.00 ADJA NN APPR 0.01 VVFIN 0.28 ART ADJA 0.00 APPR ART ADJA 0.01 APPRART 0.32 NN VVPP 0.00 NN KON NN 0.01 VAFIN 0.37 NN $( 0.01 ADJA NN $. 0.01 KON 0.37 NN NN 0.01 NN ART NN 0.32 NN 0.43 NN ART 0.21 ART NN ART 0.49 ART 0.49 ADJA NN 0.28 NN ART ADJA 0.90 $( 0.57 NN $, 0.67 ADJA NN ART 0.95 $. 1.01 NN VAFIN 0.85 NN APPR ART 0.95 NE 1.14 NN VVFIN 0.89 NN VVPP $. 1.01 CARD 1.38 NN $. 0.95 ART NN APPR 1.35 ADJA 2.38 ART NN 1.07 ART ADJA NN 1.58 $, 2.94 NN KON 2.41 APPR ART NN 2.60 ADJD 3.54 APPR NN 2.65 APPR ADJA NN 2.65 ADV 7.69 APPRART NN 3.06 ART NN VVFIN 9.51 Table 2: Reducibility scores of the most frequent German n-grams. (V* are verbs, N* are nouns, ART are articles, APPR* are prepositions, ADJ* are adjectives, ADV are adverbs, CARD are numerals, and KON are conjunctions) and adverbs are very high for all three examined languages. That is desired, because the reducible unigrams will more likely become leaves in dependency trees. Considering bigrams, the couples [determiner ? noun], [adjective ? noun], and [preposition ? noun] obtained reasonably high scores. However, there are also n-grams such as the German trigram [determiner ? noun ? preposition] (ART-NN-APPR) whose reducibility score is undesirably high.3 4 Models We introduce a new generative model that is different from the widely used Dependency Model with Valence (DMV). In DMV (Klein and Manning, 2004) and in the extended model EVG (Headden III etal2009), there is a STOP sign indicating that no more dependents in a given direction will be generated. Given a certain head, all its dependents in left direction are generated first, then the STOP sign in that direction, then all its right dependents and then STOP in the other direction. This process continues recursively for all generated dependents. Our model introduces fertility of a node, which substitutes the STOP sign. For a given head, we first generate the number of its left and right children 3The high reducibility score of ART-NN-APPR was probably caused by German particles, which have the same PoS tag as prepositions. unigrams R bigrams R trigrams R P4 0.00 RR AA 0.00 RR NN Z: 0.00 RV 0.00 Z: J, 0.00 NN RR AA 0.00 Vp 0.06 Vp NN 0.00 NN AA NN 0.16 Vf 0.06 VB NN 0.12 AA NN RR 0.23 P7 0.16 NN Vp 0.13 NN RR NN 0.46 J, 0.24 NN VB 0.18 NN J? NN 0.46 RR 0.28 NN RR 0.22 AA NN NN 0.47 VB 0.33 NN AA 0.23 NN Z: Z: 0.48 NN 0.72 NN J? 0.62 NN Z: NN 0.52 J? 1.72 AA NN 0.62 NN NN NN 0.70 C= 1.85 NN NN 0.70 AA AA NN 0.72 PD 2.06 NN Z: 0.97 AA NN Z: 0.86 AA 2.22 Z: NN 1.72 NN NN Z: 1.38 Dg 3.21 Z: Z: 1.97 RR NN NN 2.26 Z: 4.01 J? NN 2.05 RR AA NN 2.65 Db 4.62 RR NN 2.20 Z: NN Z: 8.32 Table 3: Reducibility scores of the most frequent Czech n-grams. (V* are verbs, N* are nouns, P* are pronouns, R* are prepositions, A* are adjectives, D* are adverbs, C* are numerals, J* are conjunctions, and Z* is punctuation) (fertility model) and then we fill these positions by generating its individual dependents (edge model). If a zero fertility is generated in both the directions, the head becomes a leaf. Besides the fertility model and the edge model, we use two more models (subtree model and distance model), which force the generated trees to have more desired shape.4 4.1 Fertility Model We express a fertility of a node by a pair of numbers: the number of its left dependents and the number of its right dependents. For example, fertility ?1-3? means that the node has one left and three right dependents, fertility ?0-0? indicates that it is a leaf. Fertility is conditioned by part-of-speech tag of the node and it is computed following the Chinese restaurant process. This means that if a specific fertility has been frequent for a given PoS tag in the past, it is more likely to be generated again. The formula for computing probability of fertility fi of a word on the position i in the corpus is as follows: Pf (fi|ti) = c?i(?ti, fi?) + ?P0(fi) c?i(?ti?) + ? , (4) 4In fact, the subtree model and the distance model disrupt a bit the generative story, because the probabilites do not sum up to one when they are used. However, they proved to helpwith inducing better linguistic structures. 300 where ti is part-of-speech tag of the word on the position i, c?i(?ti, fi?) stands for the count of words with PoS tag ti and fertility fi in the history, and P0 is a prior probability for the given fertility which depends on the total number of node dependents denoted by |fi| (the sum of numbers of left and right dependents): P0(fi) = 1 2|fi|+1 (5) This prior probability has a nice property: for a given number of nodes, the product of fertility probabilities over all the nodes is equal for all possible dependency trees. This ensures the stability of this model during the inference. Besides the basic fertility model, we introduce also an extended fertility model, which uses frequency of a given word form for generating number of children. We assume that the most frequent words are mostly function words (e.g. determiners, prepositions, auxiliary verbs, conjunctions). Such words tend to have a stable number of children, for example (i) some function words are exclusively leaves, (ii) prepositions have just one child, and (iii) attachment of auxiliary verbs depends on the annotation style, but number of their children is also not very variable. The higher the frequency of a word form, the higher probability mass is concentrated on one specific number of children and the lower Dirichlet hyperparameter ? in Equation 4 is needed. The extended fertility is described by equation P ?f (fi|ti, wi) = c?i(?ti, fi?) + ?eF (wi)P0(fi) c?i(?ti?) + ?eF (wi) , (6) where F (wi) is a frequency of the word wi, which is computed as a number of words wi in our corpus divided by number of all words. 4.2 Edge Model After the fertility (number of left and right dependents) is generated, the individual slots are filled using the edge model. A part-of-speech tag of each dependent is conditioned by part-of-speech tag of the head and the edge direction (position of the dependent related to the head).5 5For the edge model purposes, the PoS tag of the technical root is set to ?<root>? and it is in the zero-th position in the Similarly as for the fertility model, we employ Chinese restaurant process to assign probabilities of individual dependent. Pe(tj |ti, dj) = c?i(?ti, tj , dj?) + ? c?i(?ti, dj?) + ?|T | , (7) where ti and tj are the part-of-speech tags of the head and the generated dependent respectively; dj is a direction of edge between the words i and j, which can have two values: left and right. c?i(?ti, tj , dj?) stands for the count of edges ti ? tj with the direction dj in the history, |T | is a number of unique tags in the corpus and ? is a Dirichlet hyperparameter. 4.3 Distance Model Distance model is an auxiliary model that prevents the resulting trees from being too flat. Ideally, it would not be needed, but experiments showed that it helps to infer better trees. This simple model says that shorter edges are more probable than longer ones. We define probabilityof a distance between a word and its parent as its inverse value,6 which is then normalized by the normalization constant \x0fd. Pd(i, j) = 1 \x0fd ( 1 |i? j| )? (8) The hyperparameter ? determines the weight of this model. 4.4 Subtree Model The subtree model uses the reducibility measure. It plays an important role since it forces the reducible words to be leaves and reducible n-grams to be subtrees. Words with low reducibility are forced towards the root of the tree. We define desc(i) as a sequence of tags [tl, . . . , tr] that corresponds to all the descendants of the word wi including wi, i.e. the whole subtree of wi. The probability of such subtree is proportional to its reducibility R(desc(i)). The hyperparameter ? determines the weight of the model; \x0fs is a normalization constant. Ps(i) = 1 \x0fs R(desc(i))? (9) sentence, so the head word of the sentence is always its right dependent. 6Distance between any word and the technical root of the dependency tree was set to 10. Since each technical root has only one dependent, this value does not affect the model. 301 4.5 Probability of the Whole Treebank We want to maximize the probability of the whole generated treebank, which is computed as follows: Ptreebank = n? i=1 (P ?f (fi|ti, wi) (10) Pe(ti|tpi(i), di) (11) Pd(i, pi(i)) (12) Ps(i)), (13) where pi(i) denotes the parent of the word on the position i. We multiply the probabilities of fertility, edge, distance from parent, and subtree over all words (nodes) in the corpus. The extended fertility model P ?f can be substituted by its basic variant Pf . 5 Sampling Algorithm For stochastic searching for the most probable dependency trees, we employ Gibbs sampling, a standard Markov Chain Monte Carlo technique (Gilks et al., 1996). In each iteration, we loop over all words in the corpus in a random order and change the dependencies in their neighborhood (a small change described in Section 5.2). In the end, ?average? trees based on the whole sampling are built. 5.1 Initialization Before the sampling starts, we initialize the projective trees randomly. For doing so, we tried the following two initializers: ? For each sentence, we choose randomly one word as the head and attach all other words to it. ? We are picking one word after another in a random order and we attach it to the nearest left (or right) neighbor that has not been attached yet. The left-right choice is made by a coin flip. If it is not possible to attach a word to one side, we attach it to the other side. The last unattached word becomes the head of the sentence. While the first method generates only flat trees, the second one can generate all possible projective trees. However, the sampler converges to similar results for both the initializations. Therefore we conclude that the choice of the initialization mechanism The   dog   was   in   the   park  . (((The) dog) was (in((the) park)) (.)) Figure 1: Arrow and bracketing notation of a projective dependency tree.TTThe dogwadosinootpoTTre doki.(dooT)dd TTThe dogwadosinootpoTTre doki.(dooT)ddTTThe dogwadosinootpoTTre doki.(dooT)ddTTThe dogwadosinootpoTTre doki.(dooT)ddTTd d TTThe dogwadosinootpoTTre doki.(dooT)ddTTThe dogwadosinootpoTTre doki.(dooT)ddT Td d Figure 2: An example of small change in a projective tree. The bracket (in the park) is removed and there are five possibilities how to replace it. is not so important here and we choose the first one due to its simplicity. 5.2 Small Change Operator We use the bracketing notation for illustrating the small change operator. Each projective dependency tree consisting of n words can be expressed by n pairs of brackets. Each bracket pair belongs to one node and delimits its descendants from the rest of the sentence. Furthermore, each bracketed segment contains just one word that is not embedded deeper; this node is the segment head. An example of this notation is in Figure 1. The small change is then very simple. We remove one pair of brackets and add another, so that the conditions defined above are not violated. An example of such change is in Figure 2. From the perspective of dependency structures, the small change can be described as follows: 1. Pick a random non-root word w (the word in in our example) and find its parent p (the word was). 2. Find all other children of w and p (the words dog, park, and .) and denote this set by C. 3. Choose the new head out of w and p. Mark the new head as g and the second candidate as d. Attach d to g. 302 4. Select a neighborhoodD adjacent to the word d as a continuous subset ofC and attach all words from D to d. D may be also empty. 5. Attach the remaining words from C that were not in D to the new head g. 5.3 Building ?Average? Trees The ?burn-in? period is set to 10 iterations. After this period, we begin to count how many times an edge occurs at a particular location in the corpus. These counts are collected over the whole corpus with the collection-rate 0.01.7 When the samling is finished, we build final dependency trees based on the edge counts obtained during the sampling. We employ the maximum spanning tree (MST) algorithm (Chu and Liu, 1965) to find them; the weights of edges for computing MST correspond to the number of times they were present during the sampling. This averaging method was used also by Marec?ek and Z?abokrtsky? (2011). Other possibilities for obtaining final dependency trees would be using Eisner?s projective algorithm (Eisner, 1996) or using annealing method (favoring more likely changes) at the end of the sampling. However, the general non-projective MST algorithm enable non-projective edges, which are by no means negligible in treebanks (Havelka, 2007). 6 Experiments and Evaluation We evaluate our parser on 20 treebanks (18 languages) included in CoNLL shared tasks 2006 (Buchholz and Marsi, 2006) and 2007 (Nivre et al2007). Similarly to some previous papers on unsupervised parsing (Gillenwater et al2011; Spitkovsky et al2011b), the tuning experiments wereperformed on English only. We used English for checking functionality of the individual models and for optimizing hyperparameter values. The best configuration of the parser achieved on English development data was then used for parsing all other languages. This simulates the situation in which we have only one treebank (English) on which we can tune our parser and we want to parse other languages for which we have no manually annotated treebanks. 7After each small change is made, the edges from the whole corpus are collected with a probability 0.01. language tokens (mil.) language tokens (mil.) Arabic 19.7 Greek 20.9 Basque 14.1 Hungarian 26.3 Bulgarian 18.8 Italian 39.7 Catalan 27.0 Japanese 2.6 Czech 20.3 Portuguese 31.7 Danish 15.9 Slovenian 13.7 Dutch 27.1 Spanish 53.4 English 85.0 Swedish 19.2 German 56.9 Turkish 16.5 Table 4: Wikipedia texts statistics 6.1 Data We need two kinds of data for our experiments: a smaller treebank, which is used for sampling and for evaluation, and a large corpus, from which we compute n-gram reducibility scores. The treebanks are taken from the CoNLL shared task 2006 and 2007. The experiments are performed for all languages except for Chinese.8 We use only the testing parts of the treebanks (the files test.conll) for the dependency tree induction. As a source of the part-of-speech tags, we use the fine-grained gold PoS tags, which are in the fifth column in the CoNLL format. For obtaining reducibility scores, we used the W2C corpus9 of Wikipedia articles, which was downloaded by Majlis? and Z?abokrtsky? (2012). Their statistics across languages are shown in Table 4. To make them useful, the necessary preprocessing steps must have been done. The texts were first automatically segmented and tokenized10 and then they were part-of-speech tagged by TnT tagger (Brants, 2000), which was trained on the respective CoNLL training data (the files train.conll). The quality of such tagging is not very high, since we do not use any lexicons11 or pretrained models. However, it is sufficient for obtaining good reducibility scores. 8We do not have appropriate Chinese segmenter that would segment Chinese texts in the same way as in CoNLL. 9http://ufal.mff.cuni.cz/?majlis/w2c/ 10The segmentation to sentences and tokenization was performed using the TectoMT framework (Popel and Z?abokrtsky?, 2010). 11Using lexicons or another pretrained models for tagging means using other sources of human annotated data, which is not allowed if we want to compare our results with others. 303 6.2 Setting the Hyperparameters The applicability of individual models and their parameters were tested on development data set of English (the file en/dtest.conll in CoNLL shared task 2007). After several experiments, we have observed that the extended fertility model provides better results than the basic fertility model; the parser using the basic fertility model achieved 44.1% attachment score for English, whereas the extended fertility model increased the score to 46.8%. The four hyperparameters ?e (extended fertility model), ? (edge model), ? (distance model), and ? (subtree model), were set by a grid search algorithm,12 which found the following optimal values: ?e = 0.01, ? = 1, ? = 1.5, ? = 1 Ininformal experiments, parameters were tuned also for other treebanks and we found out that they vary across languages. Therefore, adjusting the hyperparameters on another language would probably change the scores significantly. 6.3 Evaluation The best setting from the experiments on English is now used for evaluating our parser on all CoNLL languages. To be able to compare our parser attachment score to previously published results, the following steps must be done: ? We take the testing part of each treebank (the file test.conll) and remove all the punctuation marks. If the punctuation node is not a leaf, its children are attached to the parent of the removed node. ? Some previous papers report results on up-to10-words sentences only. Therefore we extract such sentences from the test data and evaluate on this subsets as well. 12Here we make use of manually annotated trees. However, we use only English treebank an we are setting only four numbers out of several previously given values (e.g ?e out of 0.01, 0.1, 1, 10). These numbers could be tuned also by inspecting the outputs. So we believe this method can be treated as unsupervised. CoNLL ? 10 tokens all sentences language year gil11 our spi11 our Arabic 06 ? 40.5 16.6 26.5 Arabic 07 ? 48.0 49.5 27.9 Basque 07 ? 30.8 24.0 26.8 Bulgarian 06 58.3 53.2 43.9 46.0 Catalan 07 ? 63.5 59.8 47.0 Czech 06 53.2 58.9 27.7 49.5 Czech 07 ? 63.7 28.4 48.0 Danish 06 45.9 49.5 38.3 38.6 Dutch 06 33.5 48.8 27.8 44.2 English 07 ? 64.1 45.2 49.2 German 06 46.7 60.8 30.4 44.8 Greek 07 ? 30.2 13.2 20.2 Hungarian 07 ? 61.8 34.7 51.8 Italian 07 ? 50.5 52.3 43.3 Japanese 06 57.7 65.4 50.2 50.8 Portuguese 06 54.0 62.3 36.7 50.6 Slovenian 06 50.9 21.0 32.2 18.1 Spanish 06 57.9 67.3 50.6 51.9 Swedish 06 45.0 60.5 50.0 48.2 Turkish 07 ? 13.0 35.9 15.7 Average: 50.3? 54.7? 37.4 40.0 Table 5: Comparison of directed attachment scores with previously reported results on CoNLL treebanks. The column ?gil11? contains results reported by Gillenwater et al011) (see the best configuration in Table 7 in their paper). They provided only results on sentences of up to 10 tokens from CoNLL 2006 treebanks. Results in the column ?spi11? are taken from Spitkovsky et al011b), best configuration in Table 6 in their paper. The average score in the last line is computed across all comparable results, i.e. for comparison with ?gil11? only the CoNLL?06 results are averaged (?). Our parser was not evaluated on Turkish CoNLL?06 data and Chinese data, because we have not them available. The resulting scores are given in Table 5. We compare our results with results previously reported by Gillenwater (2011) and Spitkovsky (2011b), who used the CoNLL data for evaluation too. Since they provide results for several configurations of their parsers, we choose only the best one from each the paper. We define the best configuration as the one 304 with the highest average attachment score across all the tested languages. We can see that ourparser outperforms the previously published ones. In one case, it is better for 8 out of 10 data sets, in the other case, it is better for 14 out of 20 data sets. The average attachment scores, which are computed only from the results present for both compared parsers, also confirm the improvement. However, it is important to note that we used an additional source of information, namely large unannotated corpora for computing reducibility scores, while the others used only the CoNLL data. 6.4 Error Analysis Our main motivation for developing an unsupervised dependency parser was that we wanted to be able to parse any language. However, the experiments show that our parser fails for some languages. In this section, we try to analyze and explain some of the most substantial types of errors. Auxiliary verbs in Slovenian ? In the Slovenian treebank, many verbs are composed of two words: main verb (marked as Verb-main) and auxiliary verb (Verb-copula). Our parser choose the auxiliary verb as the head and the main verb and all its dependants become its children. That is why the attachment score is so poor (only 18.1%). In fact, the induced structure is not so bad. The main verb is switched with the auxiliary one which causes also the wrong attachment of all its dependants. Articles in German ? Attachment of about one half of German articles is wrong. Instead of the article being attached below the appropriate noun, the noun is attached below the article. It is a similar problem as the aforementioned Slovenian auxiliary verbs. The dependency between content and function word is switched and the dependants of the content word are attached to the function word. Klein and Manning (2004) observed a similar behavior in their experiments with DMV. Noun phrases in English ? The structure of phrases that consist of more nouns are often induced badly. This is caused probably by ignoring word forms. For example, the structure of the sequence ?NN NN NN? can be hardly recognized by our parser. fert. edge dist. subtr. en de cs (random baseline) 19.8 18.4 26.7 X 8.71 13.7 14.9 X 18.9 20.2 26.5 X 23.6 19.5 25.3 X 28.2 23.7 33.5 X X 21.2 22.9 23.5 X X 19.9 19.7 25.5 X X 7.8 17.5 22.7 X X 24.1 19.5 27.1 X X 25.5 27.5 40.7 X X 31.2 25.2 33.1 X X X 30.7 26.2 22.0 X X X 14.1 18.1 34.6 X X X 36.1 32.2 38.9 X X X 34.8 26.7 42.4 X X X X 46.8 36.5 47.2 Table 6: Ablation analysis. Unlabeled attachment scores for different combinations of model components (fertility model, edge model, distance model and subtree model). The scores are computed on all sentences of the development data. Punctuation is included into the evaluation. 6.5 Ablation Analysis To investigate the impact of individual components of the model, we run the parser for all possible component combinations. We choose three languages along the scale of word order freedom: English (very rigid word order), Czech (relatively free word order), and German (somewherein the middle). The attachment scores are shown in Table 6. If no model is used for the inference and the sampling algorithm samples completely random trees, we get the random baseline score, which is 19.8% for English13. From the perspective of the subtree model, which implements the reducibility feature, we can see that it is the most useful model here. Alone, it improves the score for English to 28.2%. If we do not use it, the score decreases from 46.8% (when all models are used) to 30.7%. Very important is also the distance model which eliminates the possibility of attaching all words to one head word. If we omit 13This relatively high baseline scores are caused by the MST algorithm, which chooses the most frequent edges from random trees i.e. the shortest ones. 305 it, the score for English falls drastically to 14.1%. Some combinations of models have their scores far below the baseline. This is caused by the fact that some regularities have been found but the structures are induced differently and thus all attachments are wrong. 6.6 Induction without Wikipedia Corpus We have performed also experiments using exclusively the CoNLL data. However, the numbers of reducible words in CoNLL training set were very low (50 words at maximum in CoNLL 2006 training data and 10 words at maximum in CoNLL 2007 training data). This led to completely unreliable reducibility scores and the consequent poor results. 7 Conclusions and Future Work We have shown that employing the reducibility feature is useful in unsupervised dependency parsing task. We extracted the n-gram reducibility scores from a large corpus, and then made the computationally demanding inference on smaller data using only these scores. We evaluated our parser on 18 languages included in CoNLL and for 14 of them, we achieved higher attachment scores than previously published results. The most errors were caused by function words, which sometimes take over the dependents of adjacent content words. This can be caused by the fact that the reducibility cannot handle function words correctly, because they must be reduced together with a content word, not one after another. In future work, we would like to estimate the hyperparameters automatically. Furthermore, we would like to get rid of manually designed PoS tags and use some kind of unsupervised clusters in order to have all the annotation process completely unsupervised. We would also like to employ lexicalized models that should help in situations in which the PoS tags are too coarse. Finally, we would like to move towards deeper syntactic structures, where the tree would be formed only by content words and the function words would be treated in a different way. Software The source code of our unsupervised dependency parser including the script for computing reducibility scores from large corpora is available at http://ufal.mff.cuni.cz/?marecek/udp.Acknowledgement This research was supported by the grants GAUK 116310, GA201/09/H057 (Res Informatica), LM2010013, MSM0021620838, and by the European Commission?s 7th Framework Program (FP7) under grant agreement n? 247762 (FAUST). We thank anonymous reviewers for their valuable comments and suggestions.'	b'1 In t roduct ion   A large number of current language processing sys-  tems use a part-of-speech tagger for pre-processing.  The tagger assigns a (unique or ambiguous) part-of-  speech tag to each token in the input and passes its  output o the next processing level, usually a parser.  Furthermore, there is a large interest in part-of-  speech tagging for corpus annotation projects, who  create valuable linguistic resources by a combination  of automatic processing and human correction.  For both applications, a tagger with the highest  possible accuracy is required. The debate about  which paradigm solves the part-of-speech tagging  problem best is not finished. Recent comparisons  of approaches that can be trained on corpora (van  Halteren et al, 1998; Volk and Schneider, 1998) have  shown that in most cases statistical aproaches (Cut-  ting et al, 1992; Schmid, 1995; Ratnaparkhi, 1996)  yield better results than finite-state, rule-based, or  memory-based taggers (Brill, 1993; Daelemans etal.,  1996). They are only surpassed by combinations of different systems, forming a "voting tagger".  Among the statistical approaches, the Maximum  Entropy framework has a very strong position. Nev-  ertheless, a recent independent comparison of 7 tag-  gets (Zavrel and Daelemans, 1999) has shown that  another approach even works better: Markov mod-  els combined with a good smoothing technique and  with handling of unknown words. This tagger, TnT,  not only yielded the highest accuracy, it also was the  fastest both in training and tagging.  The tagger comparison was organized as a "black-  box test": set the same task to every tagger and  compare the outcomes. This paper describes the  models and techniques used by TnT together with  the implementation.  The reader will be surprised how simple the under-  lying model is. The result of the tagger comparison  seems to support he maxime "the simplest is the  best". However, in this paper we clarify a number  of details that are omitted in major previous pub-  lications concerning tagging with Markov models.  As two examples, (Rabiner, 1989) and (Charniak  et al, 1993) give good overviews of the techniques  and equations used for Markov models and part-of-  speech tagging, but they are not very explicit in the  details that are needed for their application. We ar-  gue that it is not only the choice of the general model  that determines the result of the tagger but also the  various "small" decisions on alternatives.  The aim of this paper is to give a detailed ac-  count of the techniques used in TnT. Additionally,  we present results of the tagger on the NEGRA cor-  pus (Brants et al, 1999) and the Penn Treebank  (Marcus et al, 1993). The Penn Treebank results  reported here for the Markov model approach are  at least equivalent to those reported for the Maxi-  mum Entropyapproach in (Ratnaparkhi, 1996). For  a comparison to other taggers, the reader is referred  to (Zavrel and Daelemans, 1999).  2 Arch i tec ture   2.1 The Under ly ing Model   TnT uses second order Markov models for part-of-  speech tagging. The states of the model represent  tags, outputs represent the words. Transition prob-  abilities depend on the states, thus pairs of tags.  Output probabilities only depend on the most re-  cent category. To be explicit, we calculate  argmax P(tilti-1, ti-2)P(wilti P(tr+l ItT)  ti...iT  (i)  for a given sequence of words w I . . .  W T of length T.  tl... tT are elements of the tagset, the additional  224 tags t - l ,  to, and tT+l are beginning-of-sequence  and end-of-sequence markers. Using these additional  tags, even if they stem from rudimentary process-  ing of punctuation marks, slightly improves tagging  results. This is different from formulas presented  in other publications, which just stop with a "loose  end" at the last word. If sentence boundaries are  not marked in the input, TnT adds these tags if it  encounters one of \\[.!?;\\] as a token.  Transition and output probabilities are estimated  from a tagged corpus. As a first step, we use the  maximum likelihood probabilities /5 which are de-  rived from the relative frequencies:  Unigrams: /5(t3) = f(t3) (2)  N  f(t2, t3) (3)  Bigrams: P(t31t~)= f(t2)  f(ta, t2, t3)  Trigrams: /5(t3ltx,t~) - f ( t l , t2)  (4)  Lexical: /5(w3 It3) - /(w3, t3) (5)  f(t3)  for all tl, t2, t3 in the tagset and w3 in the lexi-  con. N is the total number of tokens in the training  corpus. We define a maximum likelihood probabil-  ity to be zero if the corresponding nominators and  denominators are zero. As a second step, contex-  tual frequencies are smoothed and lexical frequences  are completed by handling words that are not in the  lexicon (see below).  2.2 Smooth ing   Trigram probabilities generated from a corpus usu-  ally cannot directly be used because of the sparse-  data problem. This means that there are not enough  instances for each trigram to reliably estimate the  probability. Furthermore, setting a probability to  zero because the corresponding trigram never oc-  cured in the corpus has an undesired effect. It causes  the probability of a complete sequence to be set to  zero if its use is necessary for a new text sequence,  thus makes it impossible to rank different sequences  containing a zero probability.  The smoothing paradigm that delivers the best  results in TnT is linear interpolation of unigrams,  bigrams, and trigrams. Therefore, we estimate a  trigram probability as follows:  P(t3ltl ,  t2) = AlP(t3) + Ag_/5(t31t2) + A3/5(t3\\[t1, t2)  (6) /5 are maximum likelihood estimates of the proba-  bilities, and A1 + A2 + A3 = 1, so P again represent  probability distributions.  We use the context-independent variant of linear  interpolation, i.e., the values of the As do not depend  on the particular trigram. Contrary to intuition,  this yields better esults than the context-dependent  variant. Due to sparse-data problems, one cannot es-  timate a different set of As for each trigram. There-  fore, it is common practice to group trigrams by fre-  quency and estimate tied sets of As. However, we  are not aware of any publication that has investi-  gated frequency groupings for linear interpolation i part-of-speech tagging. All groupings that we have  tested yielded at most equivalent results to context-  independent linear interpolation. Some groupings  even yielded worse results. The tested groupings  included a) one set of As for each frequency value  and b) two classes (low and high frequency) on the  two ends of the scale, as well as several groupings  in between and several settings for partitioning the  classes.  The values of Ax, A2, and A3 are estimated by  deleted interpolation. This technique successively  removes each trigram from the training corpus and  estimates best values for the As from all other n-  grams in the corpus. Given the frequency counts  for uni-, bi-, and trigrams, the weights can be very  efficiently determined with a processing time linear  in the number of different rigrams. The algorithm  is given in figure 1. Note that subtracting 1 means  taking unseen data into account. Without this sub-  traction the model would overfit the training data  and would generally ield worse results.  2.3 Handling of Unknown Words  Currently, the method of handling unknown words  that seems to work best for inflected languages is  a suffix analysis as proposed in (Samuelsson, 1993).  Tag probabilities are set according to the word\'s end-  ing. The suffix is a strong predictor for word classes,  e.g., words in the Wall Street Journal part of the  Penn Treebank ending in able are adjectives (JJ) in  98% of the cases (e.g. fashionable, variable), the rest  of 2% are nouns (e.g. cable, variable).  The probability distribution for a particular suf-  fix is generated from all words in the training set  that share the same suffix of some predefined max-  imum length. The term suffix as used here means  "final sequence of characters of a word" which is not  necessarily a linguistically meaningful suffix.  Probabilities are smoothed by successive abstrac-  tion. This calculates the probability of a tag t  given the last m letters li of an n letter word:  P(t l ln - r ,+l , . . . ln) .  The sequence of increasingly  more general contexts omits more and more char-  acters ofthe suffix, such that P(t l ln_m+2,. . . , ln) ,   P(t \\ [ ln-m+3,. . . , l~) ,  . . . ,  P(t) are used for smooth-  ing. The recursiou formula is  P(t l l ,_ i+l,  . . . ln)  = P( t l ln - i+ l , . . .  In) + O iP ( t l l~- , . . . ,  ln) (7)  1 +0~  225 set  )%1 ---- )%2 = )%3 = 0  fo reach  t r ig ram tl,t2,t3 with  f(t l ,t2,t3) > 0  depending on the maximum of the fo l low ing  three  va lues :   f(tl ,t2,ts)-- 1 case f (h,t2)- I  " increment )%3 by f ( t l , t2 , t3)   f(t2,t3)-I case f(t2)- I  " increment )%2 by f ( t l , t2 , t s )   f ( t3 ) - - i   case N-1  " increment )%1 by f(t l ,t2,t3)  end  end  normalize )%1, )%2, )%3  Figure 1: Algorithm for calculting the weights for context-independent li ear interpolation )%1, )%2, )%3 when  the n-gram frequencies are known. N is the size of the corpus? If the denominator in one of the expressions  is 0, we define the result of that expression to be 0.  for i = m. . .  0, using the maximum likelihood esti-  mates/5  from frequencies in the lexicon, weights Oi  and the initialization  P(t) =/5(t) .  (8)  The maximum likelihood estimate for a suffix of  length i is derived from corpus frequencies by  P(t i /~- i+l ,  .. l~) = f (t ,  1~-~+1,... l~) ? (9 )   For the Markov model, we need the inverse condi-  tional probabilities P ( / , - i+ l , . . .  lnlt) which are ob-  tained by Bayesian inversion.  A theoretical motivated argumentation uses the  standard eviation of the maximum likelihood prob-  abilities for the weights 0i (Samuelsson, 1993)?  This leaves room for interpretation.  1) One has to identify a good value for m, the  longest suffix used? The approach taken for TnT is  the following: m depends on the word in question.  We use the longest suffix that we can find in the  training set (i.e., for which the frequency is greater  than or equal to 1), but at most 10 characters. This  is an empirically determined choice?  2) We use a context-independent approach for 0i,  as we did for the contextual weights )%i. It turned  out to be a good choice to set al 0i to the standard  deviation of the unconditioned maximum likelihood  probabilities of the tags in the training corpus, i.e.,  weset  1  $  Oi = ~--~(/5(tj) - 15)2 (10)  s 1 j= l   for all i = 0 . . .  m - 1, using a tagset of s tags and  the average  $  /5 = 1 ~/5(t j )  (11)  8  j----I  This usually yields values in the range 0.03 .. .  0.10.  3) We use different estimates for uppercase and  lowercase words, i.e., we maintain two different suffix  tries depending on the capitalization of the word.  This information improves the tagging results?  4) Another freedom concerns the choice of the  words in the lexicon that should be used for suf-  fix handling. Should we use all words, or are some  of them better suited than others? Accepting that  unknown words are most probably infrequent, one  can argue that using suffixes of infrequent words in  the lexicon is a better approximation for unknown  words than using suffixes of frequent words. There-  fore, we restrict the procedure of suffix handling to  words with a frequency smaller than or equal to some  threshold value. Empirically, 10 turned out to be a  good choice for this threshold.  2.4 Cap i ta l i za t ion   Additional information that turned out to be use-  ful for the disambiguation process for several cor-  pora and tagsets is capitalization information. Tags  are usually not informative about capitalization, but  probability distributions of tags around capitalized  words are different from those not capitalized. The  effect is larger for English, which only capitalizes  proper names, and smaller for German, which capi-  talizes all nouns.  We use flags ci that are true if wi is a capitalized  word and false otherwise? These flags are added to  the contextual probability distributions. Instead of  P(tsItl,t2) (12)  we use  P(t3, c3\\[tl , cl, t2, c2) (13)  and equations (3) to (5) are updated accordingly.  This is equivalent o doubling the size of the tagset  and using different ags depending on capitalization.  226 2.5 Beam Search  The processing time of the Viterbi algorithm (Ra-  biner, 1989) can be reduced by introducing a beam  search. Each state that receives a 5 value smaller  than the largest 5 divided by some threshold value  is excluded from further processing. While the  Viterbi algorithm is guaranteed to find the sequence  of states with the highest probability, this is no  longer true when beam search is added. Neverthe-  less, for practical purposes and the right choice of  0, there is virtually no difference between the algo-  rithm with and without a beam. Empirically, a value  of 0 = 1000 turned out to approximately double the  speed of the tagger without affecting the accuracy.  The tagger currently tags between 30~000 and  60,000 tokensper second (including file I/O) on a  Pentium 500 running Linux. The speed mainly de-  pends on the percentage of unknown words and on  the average ambiguity rate.  3 Eva luat ion   We evaluate the tagger\'s performance under several  aspects. First of all, we determine the tagging ac-  curacy averaged over ten iterations. The overall ac-  curacy, as well as separate accuracies for known and  unknown words are measured.  Second, learning curves are presented, that indi-  cate the performance when using training corpora of  different sizes, starting with as few as 1,000 tokens  and ranging to the size of the entire corpus (minus  the test set).  An important characteristic of statistical taggers  is that they not only assign tags to words but also  probabilities in order to rank different assignments.  We distinguish reliable from unreliable assignments  by the quotient of the best and second best assign-  ments 1. All assignments for which this quotient is  larger than some threshold are regarded as reliable,  the others as unreliable. As we will see below, accu-  racies for reliable assignments are much higher.  The tests are performed on partitions of the cor-  pora that use 90% as training set and 10% as test  set, so that the test data is guaranteed to be unseen  during training. Each result is obtained by repeat-  ing the experiment 10 times with different partitions  and averaging the single outcomes.  In all experiments, contiguous test sets are used.  The alternative is a round-robin procedure that puts  every 10th sentence into the test set. We argue that  contiguous test sets yield more realistic results be-  cause completely unseen articles are tagged. Using  the round-robin procedure, parts of an article are al-  ready seen, which significantly reduces the percent-  age of unknown words. Therefore, we expect even  1 By definition, this quotient is co if there is only one pos-  sible tag for a given word.  higher results when testing on every 10th sentence  instead of a contiguous et of 10%.  In the following, accuracy denotes the number of  correctly assigned tags divided by the number of to-  kens in the corpus processed. The tagger is allowed  to assign exactly one tag to each token.  We distinguish the overall accuracy, taking into  account all tokens in the test corpus, and separate  accuracies for known and unknown tokens. The lat-  ter are interesting, since usually unknown tokens are  much more difficult to process than known tokens,  for which a list of valid tags can be found in the  lexicon.  3.1 Tagging the  NEGRA corpus   The German NEGRA corpus consists of 20,000 sen-  tences (355,000 tokens) of newspaper texts (Frank-  furter Rundschau) that are annotated with parts-of-  speech and predicate-argument structures (Skut et al., 1997). It was developed at the Saarland Univer-  sity in Saarbrficken 2. Part of it was tagged at the  IMS Stuttgart. This evaluation only uses the part-  of-speech annotation and ignores structural annota-  tions.  Tagging accuracies for the NEGRA corpus are  shown in table 2.  Figure 3 shows the learning curve of the tagger,  i.e., the accuracy depending on the amount of train-  ing data. Training length is the nmnber of tokens  used for training. Each training length was tested  ten times, training and test sets were randomly cho-  sen and disjoint, results were averaged. The training  length is given on a logarithmic scale.  It is remarkable that tagging accuracy for known  words is very high even for very small training cot-  pora. This means that we have a good chance of  getting the right tag if a word is seen at least once  during training. Average percentages of unknown  tokens are shown in the bottom line of each diagram.  We exploit the fact that the tagger not only de-  termines tags, but also assigns probabilities. If there  is an alternative that has a probability "close to"  that of the best assignment, his alternative can be  viewed as almost equally well suited. The notion of  "close to" is expressed by the distance of probabil-  ities, and this in turn is expressed by the quotient  of probabilities. So, the distance of the probabili-  ties of a best tag tbest and an alternative tag tart  is expressed by P(tbest)/p(tau), which is some value  greater or equal to 1 since the best tag assignment  has the highest probability.  Figure 4 shows the accuracy when separating as-  signments with quotients larger and smaller than  the threshold (hence reliable and unreliable assign-  ments). As expected, we find that accuracies for  2For availability, please check  h~tp ://w~. col i. uni-sb, de/s fb378/negra-corpus  777 227 Table 2: Part-of-speech tagging accuracy for the NEGRA corpus, averaged over 10 test runs, training and  test set are disjoint. The table shows the percentage of unknown tokens, separate accuracies and standard  deviations for known and unknown tokens, as well as the overall accuracy.  percentage  unknowns  NEGRA corpus 11.9%  known  ace .   97.7% 0.23  unknown  acc. (x  89.0% 0.72  overall  aCE. o"  96.7% 0.29  NEGRA Corpus :  POS Learn ing  Curve   100  9O  80  70 /S   6O  50 , i  1 2 5  50.8 46.4 41.4  I i i I I I i  10 20 50 100 200 320 500  36.0 30.7 23.0 18.3 14.3 11.9 11).3  Overall  min =78.1%  max=96.7%  e Known  rain =95.7%  max=97.7%  - - -a- - -  Unknown  rain =61.2% max=89.0%  I  1000 x 1000 Training Length  St  avg. percentage unknown  Figure 3: Learning curve for tagging the NEGRA corpus. The training sets of variable sizes as well as test  sets of 30,000 tokens were randomly chosen. Training and test sets were disjoint, the procedure was repeated  10 times and results were averaged. Percentages of unknowns for 500k and 1000k training are determined  from an untagged extension.  NEGRA Corpus: Accuracy of  re l iab le  assignments  100  99  98  97  A  f " / :. Reliable rain =96.7% max=99.4%  96 I i i i i i i i i i i  2 5 10 20 50 100 500 2000 10000 threshold 0  100 97.9 95.1 92.7 90.3 86.8 84.1 81.0 76.1 71.9 68.3 64.1 62.0 % cases reliable  - 53.5 62.9 69.6 74.5 79.8 82.7 85.2 88.0 89.6 90.8 91.8 92.2 acc. of complement  Figure 4: Tagging accuracy for the NEGRA corpus when separating reliable and unreliable assignments. The  curve shows accuracies for reliable assignments. The numbers at the bottom line indicate the percentage of  reliable assignments and the accuracy of the complement set (i.e., unreliable assignments).  228 Table 5: Part-of-speech tagging accuracy for the Penn Treebank. The table shows the percentage of unknown  tokens, separate accuracies and standard deviations for known and unknown tokens, as well as the overall  accuracy.  I percentage known  unknowns acc. a  Penn Treebank 2.9% 97.0% 0.15  unknown  aCC. O"  85.5% 0.69  overall  aCE. O"  96.7% 0.15  100  9O  80  70 <  60  Penn Treebank: POS Learn ing  Curve   /  50 I I ~ i I I I i I  1 2 5 10 20 50 100 200 500  50.3 42.8 33.4 26.8 20.2 13.2 9.8 7.0 4.4  Overall  rain =78.6%  max=96.7%  Known  rain =95.2%  max=97.0%  Unknown  min =62.2%  max=85.5%  I  1000 ? 1000 Training Length  2.9 avg. percentage unknown  Figure 6: Learning curve for tagging the Penn Treebank. The training sets of variable sizes as well as test sets  of 100,000 tokens were randomly chosen. Training and test sets were disjoint, the procedure was repeated  10 times and results were averaged.  Penn Treebank: Accuracy of reliable assignments  100  99  98  97  Overall  rain =96.6%  max=99.4%  96 i i i i t i i i i i i i  2 5 10 20 50 100 500 2000 10000 threshold 0  100 97.7 94.6 92.2 89.8 86.3 83.5 80.4 76.6 73.8 71.0 67.2 64.5 % cases reliable  - 53.5 62.8 68.9 73.9 79.3 82.6 85.2 87.5 88.8 89.8 91.0 91.6 acc. of complement  Figure 7: Tagging accuracy for the Penn Treebank when separating reliable and unreliable assignments. The  curve shows accuraciesfor reliable assignments. The numbers at the bottom line indicate the percentage of  reliable assignments and the accuracy of the complement set.  229  reliable assignments are much higher than for unre-  liable assignments. This distinction is, e.g., useful  for annotation projects during the cleaning process,  or during pre-processing, sothe tagger can emit mul-  tiple tags if the best tag is classified as unreliable.  3.2 Tagging the Penn Treebank   We use the Wall Street Journal as contained in the  Penn Treebank for our experiments. The annotation  consists of four parts: 1) a context-free structure  augmented with traces to mark movement and dis-  continuous constituents, 2) phrasal categories that  are annotated as node labels, 3) a small set of gram-  matical functions that are annotated as extensions to  the node labels, and 4) part-of-speech tags (Marcus  et al, 1993). This evaluation only uses the part-of-  speech annotation.  The Wall Street Journal part of the Penn Tree-  bank consists of approx. 50,000 sentences (1.2 mil-  lion tokens).  Tagging accuracies for the Penn Treebank are  shown in table 5. Figure 6 shows the learning curve  of the tagger, i.e., the accuracy depending on the  amount of training data. Training length is the num-  ber of tokens used for training. Each training length  was tested ten times. Training and test sets were  disjoint, results are averaged. The training length is  given on a logarithmic scale. As for the NEGRA cor-  pus, tagging accuracy is very high for known tokens  even with small amounts of training data.  We exploit the fact that the tagger not only de-  termines tags, but also assigns probabilities. Figure  7 shows the accuracy when separating assignments  with quotients larger and smaller than the threshold  (hence reliable and unreliable assignments). Again,  we find that accuracies for reliable assignments are  much higher than for unreliable assignments.  3.3 Summary of Part-of-Speech Tagging  Results  Average part-of-speech tagging accuracy is between  96% and 97%, depending on language and tagset,  which is at least on a par with state-of-the-art re-  sults found in the literature, possibly better. For  the Penn Treebank, (Ratnaparkhi, 1996) reports an  accuracy of 96.6% using the Maximum Entropy ap-  proach, our much simpler and therefore faster HMM  approach delivers 96.7%. This comparison eeds to  be re-examined, since we use a ten-fold crossvalida-  tion and averaging of results while Ratnaparkhi only  makes one test run.  The accuracy for known tokens is significantly  higher than for unknown tokens. For the German  newspaper data, results are 8.7% better when the  word was seen before and therefore is in the lexicon,  than when it was not seen before (97.7% vs. 89.0%).  Accuracy for known tokens is high even with very  small amounts of training data. As few as 1000 to- kens are sufficient o achieve 95%-96% accuracy for  them. It is important for the tagger to have seen a  word at least once during training.  Stochastic taggers assign probabilities to tags. We  exploit the probabilities to determine reliability of  assignments. For a subset hat is determined uring  processing by the tagger we achieve accuracy rates  of over 99%. The accuracy of the complement set is  much lower. This information can, e.g., be exploited  in an annotation project to give an additional treat-  ment to the unreliable assignments, or to pass se-  lected ambiguities to a subsequent processing step. 4 Conc lus ion   We have shown that a tagger based on Markov mod-  els yields state-of-the-art results, despite contrary  claims found in the literature. For example, the  Markov model tagger used in the comparison of (van  Halteren et al, 1998) yielded worse results than all  other taggers. In our opinion, a reason for the wrong  claim is that the basic algorithms leave several deci-  sions to the implementor. The rather large amount  of freedom was not handled in detail in previous pub-  lications: handling of start- and end-of-sequence, the  exact smoothing technique, how to determine the  weights for context probabilities, details on handling  unknown words, and how to determine the weights  for unknown words. Note that the decisions we made  yield good results for both the German and the En-  glish Corpus. They do so for several other corpora  as well. The architecture remains applicable to a  large variety of languages.  According to current tagger comparisons (van  Halteren et al, 1998; Zavrel and Daelemans, 1999),  and according to a comparsion of the results pre-  sented here with those in (Ratnaparkhi, 1996), the  Maximum Entropy framework seems to be the only  other approach yielding comparable results to the  one presented here. It is a very interesting future  research topic to determine the advantages of either  of these approaches, to find the reason for their high  accuracies, and to find a good combination of both.  TnT is freely available to universities and re-  lated organizations for research purposes (see  http ://www. coli. uni-sb, de/-thorsten/tnt). Acknowledgements   Many thanks go to Hans Uszkoreit for his sup-  port during the development of TnT. Most of the  work on TnT was carried out while the author  received a grant of the Deutsche Forschungsge-  meinschaft in the Graduiertenkolleg Kognitionswis-  senschaft Saarbriicken. Large annotated corpora re  the pre-requisite for developing and testing part-of-  speech taggers, and they enable the generation of  high-quality language models. Therefore, I would  230  like to thank all the people who took the effort  to annotate the Penn Treebank, the Susanne Cor-  pus, the Stuttgarter Referenzkorpus, the NEGRA  Corpus, the Verbmobil Corpora, and several others.  And, last but not least, I would like to thank the  users of TnT who provided me with bug reports and  valuable suggestions for improvements. '
0.0	D12-1028	W10-1301	b'1 Introduction The true nature of the notion of dependency (after removing sedimentary deposits of rules imposed only by more or less arbitrary conventions) remains still somewhat vague and elusive. This holds in spite of a seemingly strong background intuition and even after a decade of formalized large-scale dependencybased resources being available to the research community. It is undeniable that a huge progress has been reached in the field of supervised dependency parsing, especially due to the CoNLL shared task series. However, when it comes to unsupervised parsing, there are surprisingly few clues we could rely on. As mentioned e.g. by Ku?bler et al2009), one of the traditional linguistic criteria for recognizing dependency relations (including their head-dependent orientation) is that a head H of a construction C determines the syntactic category of C and can often replace C. Or, in words of Dependency Analysis by Reduction (Lopatkova? et al2005), stepwise deletion of dependent elements within a sentence preserves its syntactic correctness. A similar idea of dependency analysis by splitting a sentence into all possible acceptable fragments is used by Gerdes and Kahane (2011). Of course, all the above works had to respond to the notorious fact that there are many language phenomena precluding the ideal (word by word) sentence reducibility (e.g. in the case of prepositional groups, or in the case of subjects in English finite clauses). However, we disregard their solutions tentatively and borrow only the very core of the reducibility idea: if a word can be removed from a sentence without damaging it, then it is likely to be dependent on some other (still present) word. As it is usual with dichotomies in natural languages, it seems more adequate to use a continuous scale instead of the reducible-irreducible opposition. That is why we introduce a simple reducibility measure based on n-gram corpus statistics. We employ this reducibility measure as the main feature in our unsupervised parsing procedure. The procedure is based on a commonly used Bayesian inference technique called Gibbs sampling (Gilks et al1996). In our sampler, the more reducible a given token is, the more likely it is to be sampled as a dependant and not as a head. After certain number of sampling iterations, for each sentence a final dependency tree is created (one token per node, including punctuation) that maximizes the product of edge probabilities gathered along the sampling history. Our approach allows to utilize information from 297 very large corpora. While the computationally demanding sampling procedure can be applied only on limited data, the unrepeated precomputation of statistics for reducibility estimates can easily exploit much larger data. We are not aware of any other published work on unsupervised parsing employing reducibility or a similar idea. Dominating approaches in unsupervised parsing are typically based on repeated patterns, and not on the possibility of a deletion inside a pattern. It seems that the two views of dependency (frequent co-occurrence of head-dependant pair, versus reducibility of the dependant) are rather complementary, so fruitful combinations can be hopefully expected in future. The remainder of this paper is structured as follows.Section 2 briefly outlines the state of the art in unsupervised dependency parsing. Our measure of reducibility based on a large monolingual corpus is presented in Section 3. Section 4 shows our models which serve for generating probability estimates for edge sampling described in Section 5. Experimental parsing results for languages included in CoNLL shared task treebanks are summarized in Section 6. Section 7 concludes this article. 2 Related Work The most popular approach in unsupervised dependency parsing of the recent years is to employ Dependency Model with Valence (DMV), which was introduced by Klein and Manning (2004). The inference algorithm was further improved by Smith (2007) and Cohen et al2008). Headden, Johnson and McClosky (2009) introduced the Extended Valence Grammar (EVG) and added lexicalization and smoothing. Blunsom and Cohn (2010) use tree substitution grammars, which allow learning larger dependency fragments. Unfortunately, many of these works show results only for English.1 However, the main feature of unsupervised methods should be their applicability across a wide range of languages. Such experiments were done by Spitkovsky (2011b; 2011c), where the parsing algorithm was evaluated on all 19 languages included in CoNLL 2006 (Buchholz and 1The state-of-the-art unsupervised parsers achieve more than 50% of attachment score measured on the Penn Treebank. Marsi, 2006) and 2007 (Nivre et al2007) shared tasks. The fully unsupervised linguistic analysis (Spitkovsky et al2011a) shows that the unsupervised part-of-speech tags may be more useful for this task than the supervised ones. Another possibility for obtaining dependency structures for languages without any linguistically annotated resources can be the projection using a parallel treebank with a resource-rich language (typically English). McDonald et al2011) showed that such projection produce better structures than the current unsupervised parsers do. However, our task is different. We would like to produce structures that are not burdened by any linguistic conventions. In this paper, we describe a novel approach to unsupervised dependency parsing. Our model differs from DMV, since we employ the reducibility feature and use fertility of nodes instead of generating STOP signs. We use Gibbs sampling procedure for inference instead of Variational Bayes, which has been more common for induction of linguistic strucures. Gibbs sampling algorithm for grammar induction was used also by Marec?ek and Z?abokrtsky? (2011). However, their sampling algorithm produces generally nonprojective trees. Our sampler, which is described in Section 5, introduces a completely different smallchange operator that guarantees projective edges. 3 Computing Reducibility scores We call a word (or a sequence of words) in a sentence reducible, if the sentence after removing the word remains grammatically correct. Although we cannot automatically recognize grammaticality of such newly created sentence, we can search for it in a large corpus. If we find it, we assume the word was reducible in the original sentence. Since the number of such reducible word sequences found in any corpus will be low, we determine the reducibility scores from their individual types (part-of-speech tags). This then implicitly allows some sharing of the scores between different word sequences. The necessity to search for the whole sentences in the corpus and not onlyfor some smaller context (considering, for example, just left and right neigh298 bor), which would lead to lower sparsity, is rationalized by the following example: Their children went to school. I took their children to school. The verb ?went? would be reducible in the context ?their children went to school?, because the sequence ?their children to school? occurs in the second sentence. One could find such examples frequently even for large contexts. For instance, verbs in free wordorder languages can be placed almost at any position in a sentence; therefore, without the full sentence context, they would have to be considered as reducible. To prevent this, we decided to work exclusively with the full sentence context instead of shorter contexts. Other way that would lead to lower sparsity would be searching for sequences of part-of-speech tags instead of sequences of word forms. However, this also does not bring desired results. For instance, the two following sentence patterns DT NNS VBD IN DT NN . DT NNS VBD DT NN . are quite frequent in English and we can deduce from them that the preposition IN is reducible. But this is of course a wrong deduction, since the preposition cannot be removed from the prepositional phrase. Using part-of-speech tags instead of word forms is thus not suitable for computing reducibility scores. Although we search for reducible sequences of word forms in the corpus, we compute reducibility scores for sequences of part-of-speech tags. This requires to have the corpus morphologically disambiguated. A sequences of part-of-speech tags will be denoted as ?PoS n-gram? in the following text. Assume a PoS n-gram g = [t1, . . . , tn]. We go through the corpus and search for all its occurrences. For each such occurrence, we remove the respective words from the current sentence and check in the corpus whether the rest of the sentence occurs at least once elsewhere in the corpus.2 If so, then such occurrence of PoS n-gram is reducible, otherwise it is not. We denote the number of such reducible oc2We do not take into account sentences with less then 10 words, because they could be nominal (without any verb) and might influence the reducibility scores of verbs. unigrams R bigrams R trigrams R VB 0.04 VBN IN 0.00 IN DT JJ 0.00 TO 0.07 IN DT 0.02 JJ NN IN 0.00 IN 0.11 NN IN 0.04 NN IN NNP 0.00 VBD 0.12 NNS IN 0.05 VBN IN DT 0.00 CC 0.13 JJ NNS 0.07 JJ NN . 0.00 VBZ 0.16 NN . 0.08 DT JJ NN 0.04 NN 0.22 DT NNP 0.09 DT NNP NNP 0.05 VBN 0.24 DT NN 0.09 NNS IN DT 0.14 . 0.32 NN , 0.11 NNP NNP . 0.15 NNS 0.38 DT JJ 0.13 NN IN DT 0.23 DT 0.43 JJ NN 0.14 NNP NNP , 0.46 NNP 0.78 NNP . 0.15 IN DT NNP 0.55 JJ 0.84 NN NN 0.22 DT NN IN 0.59 RB 2.07 IN NN 0.67 NNP NNP NNP 0.64 , 3.77 NNP NNP 0.76 IN DT NN 0.80 CD 55.6 INNNP 1.81 IN NNP NNP 4.27 Table 1: Reducibility scores of the most frequent English n-grams. (V* are verbs, N* are nouns, DET are determiners, IN are prepositions, JJ are adjectives, RB are adverbs, CD are numerals, and CC are coordinating conjunctions) currences of PoS n-gram g by r(g). The number of all its occurrences is c(g). The relative reducibility R(g) of a PoS n-gram g is then computed as R(g) = 1 N r(g) + ?1 c(g) + ?2 , (1) where the normalization constant N , which expresses relative reducibility over all the PoS n-grams (denoted by G), causes the scores are concentrated around the value 1. N = ? g?G(r(g) + ?1) ? g?G(c(g) + ?2) (2) Smoothing constants ?1 and ?2, which prevent reducibility scores from being equal to zero, are set to ?1 = ? g?G r(g) ? g?G c(g) , ?2 = 1 (3) This setting causes that even if a given PoS n-gram is not reducible anywhere in the corpus, its reducibility score is 1/(c(g) + 1). Tables 1, 2, and 3 show reducibility scores of the most frequent PoS n-grams of three selected languages: English, German, and Czech. If we consider only unigrams, we can see that the scores for verbs are often among the lowest. Verbs are followed by prepositions and nouns, and the scores for adjectives 299 unigrams R bigrams R trigrams R VVPP 0.00 NN APPR 0.00 NN APPR NN 0.01 APPR 0.27 APPR ART 0.00 ADJA NN APPR 0.01 VVFIN 0.28 ART ADJA 0.00 APPR ART ADJA 0.01 APPRART 0.32 NN VVPP 0.00 NN KON NN 0.01 VAFIN 0.37 NN $( 0.01 ADJA NN $. 0.01 KON 0.37 NN NN 0.01 NN ART NN 0.32 NN 0.43 NN ART 0.21 ART NN ART 0.49 ART 0.49 ADJA NN 0.28 NN ART ADJA 0.90 $( 0.57 NN $, 0.67 ADJA NN ART 0.95 $. 1.01 NN VAFIN 0.85 NN APPR ART 0.95 NE 1.14 NN VVFIN 0.89 NN VVPP $. 1.01 CARD 1.38 NN $. 0.95 ART NN APPR 1.35 ADJA 2.38 ART NN 1.07 ART ADJA NN 1.58 $, 2.94 NN KON 2.41 APPR ART NN 2.60 ADJD 3.54 APPR NN 2.65 APPR ADJA NN 2.65 ADV 7.69 APPRART NN 3.06 ART NN VVFIN 9.51 Table 2: Reducibility scores of the most frequent German n-grams. (V* are verbs, N* are nouns, ART are articles, APPR* are prepositions, ADJ* are adjectives, ADV are adverbs, CARD are numerals, and KON are conjunctions) and adverbs are very high for all three examined languages. That is desired, because the reducible unigrams will more likely become leaves in dependency trees. Considering bigrams, the couples [determiner ? noun], [adjective ? noun], and [preposition ? noun] obtained reasonably high scores. However, there are also n-grams such as the German trigram [determiner ? noun ? preposition] (ART-NN-APPR) whose reducibility score is undesirably high.3 4 Models We introduce a new generative model that is different from the widely used Dependency Model with Valence (DMV). In DMV (Klein and Manning, 2004) and in the extended model EVG (Headden III etal2009), there is a STOP sign indicating that no more dependents in a given direction will be generated. Given a certain head, all its dependents in left direction are generated first, then the STOP sign in that direction, then all its right dependents and then STOP in the other direction. This process continues recursively for all generated dependents. Our model introduces fertility of a node, which substitutes the STOP sign. For a given head, we first generate the number of its left and right children 3The high reducibility score of ART-NN-APPR was probably caused by German particles, which have the same PoS tag as prepositions. unigrams R bigrams R trigrams R P4 0.00 RR AA 0.00 RR NN Z: 0.00 RV 0.00 Z: J, 0.00 NN RR AA 0.00 Vp 0.06 Vp NN 0.00 NN AA NN 0.16 Vf 0.06 VB NN 0.12 AA NN RR 0.23 P7 0.16 NN Vp 0.13 NN RR NN 0.46 J, 0.24 NN VB 0.18 NN J? NN 0.46 RR 0.28 NN RR 0.22 AA NN NN 0.47 VB 0.33 NN AA 0.23 NN Z: Z: 0.48 NN 0.72 NN J? 0.62 NN Z: NN 0.52 J? 1.72 AA NN 0.62 NN NN NN 0.70 C= 1.85 NN NN 0.70 AA AA NN 0.72 PD 2.06 NN Z: 0.97 AA NN Z: 0.86 AA 2.22 Z: NN 1.72 NN NN Z: 1.38 Dg 3.21 Z: Z: 1.97 RR NN NN 2.26 Z: 4.01 J? NN 2.05 RR AA NN 2.65 Db 4.62 RR NN 2.20 Z: NN Z: 8.32 Table 3: Reducibility scores of the most frequent Czech n-grams. (V* are verbs, N* are nouns, P* are pronouns, R* are prepositions, A* are adjectives, D* are adverbs, C* are numerals, J* are conjunctions, and Z* is punctuation) (fertility model) and then we fill these positions by generating its individual dependents (edge model). If a zero fertility is generated in both the directions, the head becomes a leaf. Besides the fertility model and the edge model, we use two more models (subtree model and distance model), which force the generated trees to have more desired shape.4 4.1 Fertility Model We express a fertility of a node by a pair of numbers: the number of its left dependents and the number of its right dependents. For example, fertility ?1-3? means that the node has one left and three right dependents, fertility ?0-0? indicates that it is a leaf. Fertility is conditioned by part-of-speech tag of the node and it is computed following the Chinese restaurant process. This means that if a specific fertility has been frequent for a given PoS tag in the past, it is more likely to be generated again. The formula for computing probability of fertility fi of a word on the position i in the corpus is as follows: Pf (fi|ti) = c?i(?ti, fi?) + ?P0(fi) c?i(?ti?) + ? , (4) 4In fact, the subtree model and the distance model disrupt a bit the generative story, because the probabilites do not sum up to one when they are used. However, they proved to helpwith inducing better linguistic structures. 300 where ti is part-of-speech tag of the word on the position i, c?i(?ti, fi?) stands for the count of words with PoS tag ti and fertility fi in the history, and P0 is a prior probability for the given fertility which depends on the total number of node dependents denoted by |fi| (the sum of numbers of left and right dependents): P0(fi) = 1 2|fi|+1 (5) This prior probability has a nice property: for a given number of nodes, the product of fertility probabilities over all the nodes is equal for all possible dependency trees. This ensures the stability of this model during the inference. Besides the basic fertility model, we introduce also an extended fertility model, which uses frequency of a given word form for generating number of children. We assume that the most frequent words are mostly function words (e.g. determiners, prepositions, auxiliary verbs, conjunctions). Such words tend to have a stable number of children, for example (i) some function words are exclusively leaves, (ii) prepositions have just one child, and (iii) attachment of auxiliary verbs depends on the annotation style, but number of their children is also not very variable. The higher the frequency of a word form, the higher probability mass is concentrated on one specific number of children and the lower Dirichlet hyperparameter ? in Equation 4 is needed. The extended fertility is described by equation P ?f (fi|ti, wi) = c?i(?ti, fi?) + ?eF (wi)P0(fi) c?i(?ti?) + ?eF (wi) , (6) where F (wi) is a frequency of the word wi, which is computed as a number of words wi in our corpus divided by number of all words. 4.2 Edge Model After the fertility (number of left and right dependents) is generated, the individual slots are filled using the edge model. A part-of-speech tag of each dependent is conditioned by part-of-speech tag of the head and the edge direction (position of the dependent related to the head).5 5For the edge model purposes, the PoS tag of the technical root is set to ?<root>? and it is in the zero-th position in the Similarly as for the fertility model, we employ Chinese restaurant process to assign probabilities of individual dependent. Pe(tj |ti, dj) = c?i(?ti, tj , dj?) + ? c?i(?ti, dj?) + ?|T | , (7) where ti and tj are the part-of-speech tags of the head and the generated dependent respectively; dj is a direction of edge between the words i and j, which can have two values: left and right. c?i(?ti, tj , dj?) stands for the count of edges ti ? tj with the direction dj in the history, |T | is a number of unique tags in the corpus and ? is a Dirichlet hyperparameter. 4.3 Distance Model Distance model is an auxiliary model that prevents the resulting trees from being too flat. Ideally, it would not be needed, but experiments showed that it helps to infer better trees. This simple model says that shorter edges are more probable than longer ones. We define probabilityof a distance between a word and its parent as its inverse value,6 which is then normalized by the normalization constant \x0fd. Pd(i, j) = 1 \x0fd ( 1 |i? j| )? (8) The hyperparameter ? determines the weight of this model. 4.4 Subtree Model The subtree model uses the reducibility measure. It plays an important role since it forces the reducible words to be leaves and reducible n-grams to be subtrees. Words with low reducibility are forced towards the root of the tree. We define desc(i) as a sequence of tags [tl, . . . , tr] that corresponds to all the descendants of the word wi including wi, i.e. the whole subtree of wi. The probability of such subtree is proportional to its reducibility R(desc(i)). The hyperparameter ? determines the weight of the model; \x0fs is a normalization constant. Ps(i) = 1 \x0fs R(desc(i))? (9) sentence, so the head word of the sentence is always its right dependent. 6Distance between any word and the technical root of the dependency tree was set to 10. Since each technical root has only one dependent, this value does not affect the model. 301 4.5 Probability of the Whole Treebank We want to maximize the probability of the whole generated treebank, which is computed as follows: Ptreebank = n? i=1 (P ?f (fi|ti, wi) (10) Pe(ti|tpi(i), di) (11) Pd(i, pi(i)) (12) Ps(i)), (13) where pi(i) denotes the parent of the word on the position i. We multiply the probabilities of fertility, edge, distance from parent, and subtree over all words (nodes) in the corpus. The extended fertility model P ?f can be substituted by its basic variant Pf . 5 Sampling Algorithm For stochastic searching for the most probable dependency trees, we employ Gibbs sampling, a standard Markov Chain Monte Carlo technique (Gilks et al., 1996). In each iteration, we loop over all words in the corpus in a random order and change the dependencies in their neighborhood (a small change described in Section 5.2). In the end, ?average? trees based on the whole sampling are built. 5.1 Initialization Before the sampling starts, we initialize the projective trees randomly. For doing so, we tried the following two initializers: ? For each sentence, we choose randomly one word as the head and attach all other words to it. ? We are picking one word after another in a random order and we attach it to the nearest left (or right) neighbor that has not been attached yet. The left-right choice is made by a coin flip. If it is not possible to attach a word to one side, we attach it to the other side. The last unattached word becomes the head of the sentence. While the first method generates only flat trees, the second one can generate all possible projective trees. However, the sampler converges to similar results for both the initializations. Therefore we conclude that the choice of the initialization mechanism The   dog   was   in   the   park  . (((The) dog) was (in((the) park)) (.)) Figure 1: Arrow and bracketing notation of a projective dependency tree.TTThe dogwadosinootpoTTre doki.(dooT)dd TTThe dogwadosinootpoTTre doki.(dooT)ddTTThe dogwadosinootpoTTre doki.(dooT)ddTTThe dogwadosinootpoTTre doki.(dooT)ddTTd d TTThe dogwadosinootpoTTre doki.(dooT)ddTTThe dogwadosinootpoTTre doki.(dooT)ddT Td d Figure 2: An example of small change in a projective tree. The bracket (in the park) is removed and there are five possibilities how to replace it. is not so important here and we choose the first one due to its simplicity. 5.2 Small Change Operator We use the bracketing notation for illustrating the small change operator. Each projective dependency tree consisting of n words can be expressed by n pairs of brackets. Each bracket pair belongs to one node and delimits its descendants from the rest of the sentence. Furthermore, each bracketed segment contains just one word that is not embedded deeper; this node is the segment head. An example of this notation is in Figure 1. The small change is then very simple. We remove one pair of brackets and add another, so that the conditions defined above are not violated. An example of such change is in Figure 2. From the perspective of dependency structures, the small change can be described as follows: 1. Pick a random non-root word w (the word in in our example) and find its parent p (the word was). 2. Find all other children of w and p (the words dog, park, and .) and denote this set by C. 3. Choose the new head out of w and p. Mark the new head as g and the second candidate as d. Attach d to g. 302 4. Select a neighborhoodD adjacent to the word d as a continuous subset ofC and attach all words from D to d. D may be also empty. 5. Attach the remaining words from C that were not in D to the new head g. 5.3 Building ?Average? Trees The ?burn-in? period is set to 10 iterations. After this period, we begin to count how many times an edge occurs at a particular location in the corpus. These counts are collected over the whole corpus with the collection-rate 0.01.7 When the samling is finished, we build final dependency trees based on the edge counts obtained during the sampling. We employ the maximum spanning tree (MST) algorithm (Chu and Liu, 1965) to find them; the weights of edges for computing MST correspond to the number of times they were present during the sampling. This averaging method was used also by Marec?ek and Z?abokrtsky? (2011). Other possibilities for obtaining final dependency trees would be using Eisner?s projective algorithm (Eisner, 1996) or using annealing method (favoring more likely changes) at the end of the sampling. However, the general non-projective MST algorithm enable non-projective edges, which are by no means negligible in treebanks (Havelka, 2007). 6 Experiments and Evaluation We evaluate our parser on 20 treebanks (18 languages) included in CoNLL shared tasks 2006 (Buchholz and Marsi, 2006) and 2007 (Nivre et al2007). Similarly to some previous papers on unsupervised parsing (Gillenwater et al2011; Spitkovsky et al2011b), the tuning experiments wereperformed on English only. We used English for checking functionality of the individual models and for optimizing hyperparameter values. The best configuration of the parser achieved on English development data was then used for parsing all other languages. This simulates the situation in which we have only one treebank (English) on which we can tune our parser and we want to parse other languages for which we have no manually annotated treebanks. 7After each small change is made, the edges from the whole corpus are collected with a probability 0.01. language tokens (mil.) language tokens (mil.) Arabic 19.7 Greek 20.9 Basque 14.1 Hungarian 26.3 Bulgarian 18.8 Italian 39.7 Catalan 27.0 Japanese 2.6 Czech 20.3 Portuguese 31.7 Danish 15.9 Slovenian 13.7 Dutch 27.1 Spanish 53.4 English 85.0 Swedish 19.2 German 56.9 Turkish 16.5 Table 4: Wikipedia texts statistics 6.1 Data We need two kinds of data for our experiments: a smaller treebank, which is used for sampling and for evaluation, and a large corpus, from which we compute n-gram reducibility scores. The treebanks are taken from the CoNLL shared task 2006 and 2007. The experiments are performed for all languages except for Chinese.8 We use only the testing parts of the treebanks (the files test.conll) for the dependency tree induction. As a source of the part-of-speech tags, we use the fine-grained gold PoS tags, which are in the fifth column in the CoNLL format. For obtaining reducibility scores, we used the W2C corpus9 of Wikipedia articles, which was downloaded by Majlis? and Z?abokrtsky? (2012). Their statistics across languages are shown in Table 4. To make them useful, the necessary preprocessing steps must have been done. The texts were first automatically segmented and tokenized10 and then they were part-of-speech tagged by TnT tagger (Brants, 2000), which was trained on the respective CoNLL training data (the files train.conll). The quality of such tagging is not very high, since we do not use any lexicons11 or pretrained models. However, it is sufficient for obtaining good reducibility scores. 8We do not have appropriate Chinese segmenter that would segment Chinese texts in the same way as in CoNLL. 9http://ufal.mff.cuni.cz/?majlis/w2c/ 10The segmentation to sentences and tokenization was performed using the TectoMT framework (Popel and Z?abokrtsky?, 2010). 11Using lexicons or another pretrained models for tagging means using other sources of human annotated data, which is not allowed if we want to compare our results with others. 303 6.2 Setting the Hyperparameters The applicability of individual models and their parameters were tested on development data set of English (the file en/dtest.conll in CoNLL shared task 2007). After several experiments, we have observed that the extended fertility model provides better results than the basic fertility model; the parser using the basic fertility model achieved 44.1% attachment score for English, whereas the extended fertility model increased the score to 46.8%. The four hyperparameters ?e (extended fertility model), ? (edge model), ? (distance model), and ? (subtree model), were set by a grid search algorithm,12 which found the following optimal values: ?e = 0.01, ? = 1, ? = 1.5, ? = 1 Ininformal experiments, parameters were tuned also for other treebanks and we found out that they vary across languages. Therefore, adjusting the hyperparameters on another language would probably change the scores significantly. 6.3 Evaluation The best setting from the experiments on English is now used for evaluating our parser on all CoNLL languages. To be able to compare our parser attachment score to previously published results, the following steps must be done: ? We take the testing part of each treebank (the file test.conll) and remove all the punctuation marks. If the punctuation node is not a leaf, its children are attached to the parent of the removed node. ? Some previous papers report results on up-to10-words sentences only. Therefore we extract such sentences from the test data and evaluate on this subsets as well. 12Here we make use of manually annotated trees. However, we use only English treebank an we are setting only four numbers out of several previously given values (e.g ?e out of 0.01, 0.1, 1, 10). These numbers could be tuned also by inspecting the outputs. So we believe this method can be treated as unsupervised. CoNLL ? 10 tokens all sentences language year gil11 our spi11 our Arabic 06 ? 40.5 16.6 26.5 Arabic 07 ? 48.0 49.5 27.9 Basque 07 ? 30.8 24.0 26.8 Bulgarian 06 58.3 53.2 43.9 46.0 Catalan 07 ? 63.5 59.8 47.0 Czech 06 53.2 58.9 27.7 49.5 Czech 07 ? 63.7 28.4 48.0 Danish 06 45.9 49.5 38.3 38.6 Dutch 06 33.5 48.8 27.8 44.2 English 07 ? 64.1 45.2 49.2 German 06 46.7 60.8 30.4 44.8 Greek 07 ? 30.2 13.2 20.2 Hungarian 07 ? 61.8 34.7 51.8 Italian 07 ? 50.5 52.3 43.3 Japanese 06 57.7 65.4 50.2 50.8 Portuguese 06 54.0 62.3 36.7 50.6 Slovenian 06 50.9 21.0 32.2 18.1 Spanish 06 57.9 67.3 50.6 51.9 Swedish 06 45.0 60.5 50.0 48.2 Turkish 07 ? 13.0 35.9 15.7 Average: 50.3? 54.7? 37.4 40.0 Table 5: Comparison of directed attachment scores with previously reported results on CoNLL treebanks. The column ?gil11? contains results reported by Gillenwater et al011) (see the best configuration in Table 7 in their paper). They provided only results on sentences of up to 10 tokens from CoNLL 2006 treebanks. Results in the column ?spi11? are taken from Spitkovsky et al011b), best configuration in Table 6 in their paper. The average score in the last line is computed across all comparable results, i.e. for comparison with ?gil11? only the CoNLL?06 results are averaged (?). Our parser was not evaluated on Turkish CoNLL?06 data and Chinese data, because we have not them available. The resulting scores are given in Table 5. We compare our results with results previously reported by Gillenwater (2011) and Spitkovsky (2011b), who used the CoNLL data for evaluation too. Since they provide results for several configurations of their parsers, we choose only the best one from each the paper. We define the best configuration as the one 304 with the highest average attachment score across all the tested languages. We can see that ourparser outperforms the previously published ones. In one case, it is better for 8 out of 10 data sets, in the other case, it is better for 14 out of 20 data sets. The average attachment scores, which are computed only from the results present for both compared parsers, also confirm the improvement. However, it is important to note that we used an additional source of information, namely large unannotated corpora for computing reducibility scores, while the others used only the CoNLL data. 6.4 Error Analysis Our main motivation for developing an unsupervised dependency parser was that we wanted to be able to parse any language. However, the experiments show that our parser fails for some languages. In this section, we try to analyze and explain some of the most substantial types of errors. Auxiliary verbs in Slovenian ? In the Slovenian treebank, many verbs are composed of two words: main verb (marked as Verb-main) and auxiliary verb (Verb-copula). Our parser choose the auxiliary verb as the head and the main verb and all its dependants become its children. That is why the attachment score is so poor (only 18.1%). In fact, the induced structure is not so bad. The main verb is switched with the auxiliary one which causes also the wrong attachment of all its dependants. Articles in German ? Attachment of about one half of German articles is wrong. Instead of the article being attached below the appropriate noun, the noun is attached below the article. It is a similar problem as the aforementioned Slovenian auxiliary verbs. The dependency between content and function word is switched and the dependants of the content word are attached to the function word. Klein and Manning (2004) observed a similar behavior in their experiments with DMV. Noun phrases in English ? The structure of phrases that consist of more nouns are often induced badly. This is caused probably by ignoring word forms. For example, the structure of the sequence ?NN NN NN? can be hardly recognized by our parser. fert. edge dist. subtr. en de cs (random baseline) 19.8 18.4 26.7 X 8.71 13.7 14.9 X 18.9 20.2 26.5 X 23.6 19.5 25.3 X 28.2 23.7 33.5 X X 21.2 22.9 23.5 X X 19.9 19.7 25.5 X X 7.8 17.5 22.7 X X 24.1 19.5 27.1 X X 25.5 27.5 40.7 X X 31.2 25.2 33.1 X X X 30.7 26.2 22.0 X X X 14.1 18.1 34.6 X X X 36.1 32.2 38.9 X X X 34.8 26.7 42.4 X X X X 46.8 36.5 47.2 Table 6: Ablation analysis. Unlabeled attachment scores for different combinations of model components (fertility model, edge model, distance model and subtree model). The scores are computed on all sentences of the development data. Punctuation is included into the evaluation. 6.5 Ablation Analysis To investigate the impact of individual components of the model, we run the parser for all possible component combinations. We choose three languages along the scale of word order freedom: English (very rigid word order), Czech (relatively free word order), and German (somewherein the middle). The attachment scores are shown in Table 6. If no model is used for the inference and the sampling algorithm samples completely random trees, we get the random baseline score, which is 19.8% for English13. From the perspective of the subtree model, which implements the reducibility feature, we can see that it is the most useful model here. Alone, it improves the score for English to 28.2%. If we do not use it, the score decreases from 46.8% (when all models are used) to 30.7%. Very important is also the distance model which eliminates the possibility of attaching all words to one head word. If we omit 13This relatively high baseline scores are caused by the MST algorithm, which chooses the most frequent edges from random trees i.e. the shortest ones. 305 it, the score for English falls drastically to 14.1%. Some combinations of models have their scores far below the baseline. This is caused by the fact that some regularities have been found but the structures are induced differently and thus all attachments are wrong. 6.6 Induction without Wikipedia Corpus We have performed also experiments using exclusively the CoNLL data. However, the numbers of reducible words in CoNLL training set were very low (50 words at maximum in CoNLL 2006 training data and 10 words at maximum in CoNLL 2007 training data). This led to completely unreliable reducibility scores and the consequent poor results. 7 Conclusions and Future Work We have shown that employing the reducibility feature is useful in unsupervised dependency parsing task. We extracted the n-gram reducibility scores from a large corpus, and then made the computationally demanding inference on smaller data using only these scores. We evaluated our parser on 18 languages included in CoNLL and for 14 of them, we achieved higher attachment scores than previously published results. The most errors were caused by function words, which sometimes take over the dependents of adjacent content words. This can be caused by the fact that the reducibility cannot handle function words correctly, because they must be reduced together with a content word, not one after another. In future work, we would like to estimate the hyperparameters automatically. Furthermore, we would like to get rid of manually designed PoS tags and use some kind of unsupervised clusters in order to have all the annotation process completely unsupervised. We would also like to employ lexicalized models that should help in situations in which the PoS tags are too coarse. Finally, we would like to move towards deeper syntactic structures, where the tree would be formed only by content words and the function words would be treated in a different way. Software The source code of our unsupervised dependency parser including the script for computing reducibility scores from large corpora is available at http://ufal.mff.cuni.cz/?marecek/udp.Acknowledgement This research was supported by the grants GAUK 116310, GA201/09/H057 (Res Informatica), LM2010013, MSM0021620838, and by the European Commission?s 7th Framework Program (FP7) under grant agreement n? 247762 (FAUST). We thank anonymous reviewers for their valuable comments and suggestions.'	b'1 Introduction  Many tools have been developed to help children and  adults who cannot speak (or who have limited speech)  communicate better.  However, most of these tools have  focused on supporting communication for practical  goals, such as ?I am thirsty.? But human communication is also used for social goals; we develop friendships  and other inter-personal relationships via social interaction and communication. The bulk of conversation is  characterized by free narrative (Cheepen 1988). One of  the most important types of conversational narrative is  personal narrative: someone telling a story about what  happened to him or her.   People with limited or no functional speech do tell  stories, but these tend to be in monologue form, or in a  sequence of pre-stored utterances on voice output communication aids (Waller 2006). Individuals who use                                                             1  The term Complex Communication Needs (CCN) describes  individuals who, due to motor, language, cognitive, and/or  sensory perceptual impairments (e.g., as a result of cerebral  palsy), do not develop speech and language skills as expected.  This heterogeneous group typically experiences restricted  access to the environment, limited interactions with their  communication partners, and few opportunities for communication (Light and Drager 2007).  Augmentative and Alternative Communication (AAC)  tools tend to be passive, responding to questions with  single words or short sentences (e.g. Soto, Hartmann et  al. 2006) and if able to initiate and maintain extended  conversations tend to relate experience word for word  each time they tell a story, even though much of conversation is reused (Clarke and Clarke 1977). This is time  consuming and physically exhausting ? typical rates  range from 8 to 10 words per minute up to 12 to 15 per  minute when techniques such as word prediction are  used (Higginbotham, Shane et al 2007), with the result  that people seldom engage in storytelling. Despite the  importance of narrative, little work has been done on  specific tools to help language-impaired individuals  engage in personal storytelling. In this paper, we describe our work in progress on building a tool that uses  Natural Language Generation (NLG) technology to help  children tell stories about their day at school, describing  both the work we have done to date, and the challenges  that we face in further developing this concept.  2 Background  2.1 AAC  Technology underpins much of Augmentative and Alternative Communication (AAC), a field that attempts to  augment natural speech and provides alternative ways to  communicate for people with limited or no speech. At  the simplest level, people withComplex Communication Needs (CCN) can cause a pre-stored message to be  spoken by activating a single switch. At the most sophisticated level, literate users can generate novel text  using input methods ranging from a single switch to a  full keyboard.   Despite advances in AAC, there are still many individuals for whom communication remains problematic.  Although some individuals with CCN become effective  communicators, most do not ? they tend to be passive  communicators, responding mainly to questions or  prompts at a one or two word level. Conversational  1 skills such as initiation, elaboration and storytelling are  seldom observed (Waller 2006).  One reason for the reduced levels of communicative  ability is the cognitive demands of AAC interfaces. Current AAC technology provides the user with a purely  physical link to speech output. The user is required to  have sufficient cognitive abilities and physical stamina  to translate what they want to say into the sequence of  operations needed to produce the desired output. Mnemonic codes and dynamic displays (Beukelman and  Mirenda 2005) provide some help in the retrieval  process, but users still have to master complex retrieval  and production strategies.   A second reason for the impoverished quality of  conversation is the focus of AAC devices on transactional communication; conversation which expresses  needs wants and information transfer, for example, ?I  am thirsty?, ?I use a straw for drinking?. Instead, interactive conversation is characterized by free narrative  and phatic conversation, for example, ?Guess what  happened this morning??, ?Hello?, and ?How are  you?? Without easy access to extended interactive  communication, it is difficult to develop the skills  needed to initiate new topics and engage in storytelling.   2.2 Importance of Narrative   Conversational narratives (oral stories told during  interactive conversations) are crucial to social engagement. Narratives provide a means for people to relate  and share experiences, develop organizational skills,  work through problems, develop self image, express  personality, give form and meaning to life, and allow  people to be interesting entertainers (Waller 2006).    Narrative skills develop experientially with children  being able to engage in storytelling even before they are  verbal (Bruner 1975). Early personal experience stories  consist of a high point, for example, ?Mummy fall!?  with adults scaffolding the full story, eliciting the ?who?,  ?what?, ?when? and ?where?. However, not all experiences make good stories. An experience becomes a  story if the storyteller has an emotional connection to  the event (Labov, 1972), or if the event is unusual (Quasthoff & Nikolaus, 1982).   Parents of typically developing children encourage  development of narrative skills by eliciting stories from  their children (Peterson and McCabe 1983), but the development of narrative skills is problematic for people  with CCN. We recall a study where disabled children  were told different stories more often than typically  developing peers who were read the samestory night  after night (Light, Binger et al 1994). In doing so, the  disabled children did not have the chance to learn the  sequence of stories, or the structure commonly used in  narrative such as beginning, middle and end. As such,  initially children should use the same story template  consistently until they are ready to progress to another  one.   It is difficult to provide access to event information  which may become a story, and few AAC systems provide support for interactive story narration. However  NLG gives us a possibility to change the underlying  paradigm of AAC. Instead of placing the entire cognitive load on the user, AAC devices can be designed to  support the retrieval of story events and the scaffolding  of story narration for individuals with CCN.  2.3 NLG, Data-to-text  NLG systems generate texts in English (or other human  languages) from non-linguistic data (Reiter and Dale  2000).  Our vision is to use an NLG system to generate  a draft story, which the child can edit.  The nonlinguistic input to our story-generator is sensor data  about the child?s activities, including location data  (where the child was) and interaction data (what people  and objects the child interacted with).  We also want to  allow teachers and school staff to enter information  about the child?s activities (such as voice messages).   A number of data-to-text systems (Reiter 2007) have  been developed in recent years, which generate English  summaries of sensor and other numerical data.  The  most popular application area has been weather forecasting (generating textual weather forecasts from the  results of a numerical atmosphere simulation model),  and indeed several weather forecast generators have  been fielded and used operationally (Goldberg, Driedger  et al 1994; Reiter, Sripada et al 2005). A number of  data-to-text systems have also been developed in the  medical community, such as BabyTalk (Gatt, Portet et  al. 2009), which generates summaries of clinical data  from a neonatal intensive care unit, and the commercial  Narrative Engine (Harris 2008) which summarizes data  acquired during a doctor/patient encounter.   Most previous research in data-to-text has focused  on summarizing technical data for expert users, with the  goal of effectively communicating key information.  In  our work, in contrast, the focus is on summarizing data  about everyday events, with the goal of having something interesting to talk about.  There has been considerable work in the computational creativity community on  generating interesting stories (P?r?z and Sharples 2004),  but it has focused on fictional written stories, where the  computer system can say whatever it wishes, without  the constraint of describing real events.   Most previous work in NLG has focused on computer systems which generate texts without human input.  However, in our case we want children to be able  to annotate (evaluate) and edit stories, asfar as their  abilities permit.  There has been some research on human post-editing of NLG texts (Sripada, Reiter et al  2005), but this has focused on editing at the text level.   2 Since editing at the text level is very laborious for AAC  users, we need a higher-level interface that lets children  edit content and structure without needing to type  words.  We also want children to be able to control how  a story is narrated, perhaps in response to a listener?s  questions or body language.  For example children may  wish to add comments such as ?It was awesome!?, or  tell events out of sequence.      In short, we need to develop interfaces and interaction techniques that allow our users to control the NLG  system.  Unfortunately there has been very little previous work on this topic, indeed almost nothing is  known about Human-Computer Interaction aspects of  NLG systems.  Developing a better understanding of  these aspects is one of the main research challenges we  face from an NLG perspective.  3 Current and ongoing work  3.1 ?How was School today???  We developed an initial version of ?How was School  today??? in 2009; see Reiter et al(2009) for more details about this system.        Fig. 1: Participating pupil with support worker:   The prototype system is mounted on the wheelchair, and  the pupil has access to the system via head switch controlled row/column scanning.     This system used Radio Frequency Identification  (RFID), an emerging application in AAC to identify or  give access to relevant vocabulary (Bart, Riny et al  2008; DeRuyter and Fried-Oken 2010). Sensors were  used to track both location (by putting tags on doors,  which were automatically sensed by a long-range RFID  reader) and interaction (by asking staff to manually  swipe RFID cards in a short-range reader when the child  interacted with a person or object). Staff could also  record spoken messages about interesting events during  the day (see Fig. 1).  The software analyzed this data to remove sensor noise,  and then compared it to a timetable which specified  where children were supposed to be, what they were  supposed to be doing and which teacher was supposed  to be taking the class throughout the day.  This allowed  the software to both fill in missing information, and to  identify divergences from the schedule. The result of  this process was a series of events (which corresponded  to classes, for example, maths class), each of which had  a set of associated messages (interactions during the  event, divergence from schedule, etc.).   After the data analysis was completed, an NLG system identified the events most interesting (to the child),  using a heuristic that took into consideration both how  inherent interesting an event was(for example, lunch  was regarded as an inherently interesting event that  children were likely to want to talk about) and also  whether an event was unusual or not.  The latter is based  on the observation that most personal narratives focus  on unexpected or unusual events.  Unusual events were  identified by the presence of recorded voice messages  and by divergence from the timetable, e.g. a different  teacher present or a different location. The system selected the five most ?interesting? events and displayed  them to the child in a simple visual editing interface.  In  this interface the child could delete events he/she did  not wish to talk about, and also annotate events with  simple opinions (evaluations), such as I liked it, using  the evaluation buttons on the interface, generating appropriate utterances according to the last narrated event  or message.(see Fig. 2).        When editing was finished, the NLG system generated texts describing the events and messages, which the  child communicated using a simple narration interface  (which was similar to the editing interface). Emphasis  was placed on providing quick access to messages to  minimize the length of pauses between utterances due to  the physical accessing difficulties of the users. The  narrative model is based on the Labov social narrative  model (Labov 1972) which emphasizes the highpoint  and evaluation. The dialogue model from beginning  through to highpoint to the end with the user being able  to add evaluations at any point of the narration. Stories  are initially chronological order but interactively under  the control of the user. This control of narration differs  significantly from current AAC interventions where  narrative tends to be output in a monologue format.    From an NLG perspective, the system was fairly  straightforward. The most challenging microplanning  tasks were choosing connectives, time phrases, lexical  variety in embellishments, and pronouns based on discourse context. Connectives and time phrases were necessary since children could narrate events in different  orders (for example, ?I went to maths.  Then I went to  lunch? versus ?I went to lunch.  Before lunch, I went to  maths?).  Document structuring  was simple because we  Short range RFID reader  and microphone for voice  message recording  Visual interface  Access switch in head  rest  Long range RFID reader  3 assumed that the children would choose their own order  in which to narrate events. In fact some children are not  able to do this; such children would need to be supported by a more sophisticated document planner that  had a model of appropriate text structures in this domain.   We asked two children to use the system for one  week for a qualitative formative evaluation.  Researchers supplied ongoing support during this, primarily trial  observing how the children used the system, anddiscussing it with teachers, therapists and parents.  Generally it worked well for one child, Julie2, who had severe  motor impairment (no independent means of mobility  and interacted with a computer using a head switch with  row/column scanning, see Fig. 1). Her expressive abilities were limited but her comprehension skills were  comparable to her non-disabled peers with some developmental delay. The other child, Jessica, had more  cognitive impairment, and found the interface too difficult.          Fig. 2: Example screenshot from interface  1: Navigation: Day and date of story, maximum of five  story events, exit; 2: Event messages, numbers vary for  each event. Here: 2 computer-generated messages, 3  recorded messages, 1 user added evaluation;  3: Sequential message navigation: previous, repeat,  next; 4: Evaluation: delete event, negative evaluation,  positive evaluation;         In a second evaluation, a third child, Eric, joined and  all three children used the system over two weeks each.  In this evaluation, we asked teachers and other staff to  use the system without on-site support from the researchers. This highlighted many practical usability  issues, such as delays caused by starting the system in  the morning, and problems caused by limited battery  life. We eliminated the long-range RFID sensor because  of its difficult setup; instead we asked staff to swipe                                                             2  The names of the children mentioned in this paper are  changed to ensure anonymity.  door cards when children entered rooms.  However, this  strategy was not successful, as it was difficult for staff  to remember to swipe both interactions and location  changes.   The participants took the system home for use with  their parents who gave positive feedback but also reported issues with system usability (e.g. lack of access  to stories from previous days) or suitability (too complicated interface for Jessica).     Eric?s timetable was different from Julie and Jessica?s, because he visited college one morning a day, and  we could not collect data during this period. Since some  of the most exciting events in a school day happen outside the school building (sports and school trips as well  as college), in the long term we do need to see if we can  collect data outside as well as inside the school.   3.2 HWST example       Julie used the system on her DynaVox? Vmax?  Voice Output Communication Aid (VOCA) via head  switch using row/column scanning. The above transcript  shows an extract ofa conversation Julie had with her  Speech and Language Therapist (SLT) on day three  about her experiences during day two. The researcher  (RA) had been present all day for technical support. The  conversation extract starts with Julie reporting about her  morning break.        In this example Julie is able to quickly reply to context related questions from her communication partner  using the evaluations (?So what happened?? ? ?It was  fun!?). Compared to conversations usually observed  between aided and unaided partners Julie is able to control the conversation when starting a new topic after  talking about the morning break, inviting her communication partner to prompt for more detailed information.  Julie provides this with her next generated phrase.  When she is asked about the event she replies with an  evaluation the system has generated in relation to its  previously generated message ?A visitor was there.?.  We note that the system is able to refer to the correct  gender of the visitor.    1 Julie {next} [I had break.]  2 Julie {next} [Lesley was there.]  3 SLT Lesley was there?   4 Julie ((Opens mouth in agreement, then turns back  to screen))  5 SLT Ok mhmh. So what happened?  6 Julie {positive evaluation} [It was fun.]  7 SLT Oh good! ((laughs)) I?m glad to hear it!  8 RA We like Lesley.  9 SLT ((nods in direction of RA))  10 Julie ((smiles))  11 Julie:  {next} [Then I went to Junior Primary instead of Reading Class.]  1  2  3  4  4 12 SLT: Right, you went to Junior Primary? I wonder  why that was?   13 Julie: {next} [A visitor was there.]  14 SLT:  Oh, a visitor, right. Wonder what the visitor  was doing?   15 Julie: {next} [?The dental hygienist came to give a  talk.?]  16 SLT: Oh, dental hygienist.  17 Julie:  {previous} [A visitor was there.]  18 SLT: That was the visitor, okay. That?s why you  went to junior primary, uhm, what did you  think of the talk?  19 Julie:  {positive evaluation} [She was nice.]  20 SLT: She was nice, that was good! ((laughs))    Notation:  - Switch selected button by Julie: {curly brackets}  - Natural speech: standard text.   - Computer generated language accessed using one  button: [standard text in square brackets].   - Recorded messages accessed using one button:  [?quoted standard text in square brackets?].  - Paralinguistic behaviors:   ((standard text in double brackets)).     3.3  ?How was School Today? ? in the Wild  We have now started a new project to further develop  our work, called ?How was School today??? ? in the  Wild (?in the wild? indicates that the focus is on how the  technology works ina real school environment). The  basic goal is to improve the system sufficiently so that it  can be tried out over a period of several months, with  children with varying levels and types of impairments;  we will also work with several schools in the initial  phase, although for practical reasons the evaluations  may be at just one school.       During this project we will do some work on the  issues described in Section 4; in particular we will try to  make the system usable by children with different impairments and ability levels (Section 4.1). This means  having a very simple interface for children with considerable cognitive impairments (such as Jessica); but  also giving children with more cognitive abilities the  opportunity to exert more control over the story (during  both editing and narration), for example by supporting a  richer range of annotations, and by making it easier to  describe events and messages in any order.       Another intermediate goal is to improve the integration of voice messages entered by staff with the computer-generated messages. This could be done by some  combination of training staff to enter messages in a specific way (referring to the child in the first person); asking staff to annotate the messages so the computer  knows something about their content; and/or using  speech recognition to analyze the voice messages. In  general there is a lot of interesting information that can  only come from staff, and we need to think about the  best way to help staff enter information in a way that is  easy for them and useful for our system.   Now that a complete system is built, we are also able  to thoroughly and formally evaluate the system. Multiple baseline single case study methodology will be  used (Schlosser 2003) to evaluate the use and impact of  our system. We intend to have up to four children (with  varying ability levels) use the system for a period of 3  months. This will give us a chance to observe the impact of the system on the users and their environment  such as the children?s interaction with the system and  how staff at the school envisage using this new tool.  The observations will be supported by semi-structured  interviews with the children, their classroom teacher,  their speech and language therapist and a parent.   We will look at the children?s conversations (with  and without using our system) about interesting, staged  events with different partners, analyzing them conversational characteristics such as narrative initiation, structure, length and evaluation. Analysis methods will  include the Revised Edinburgh Functional Communication Profile (REFCP) (Wirz, Skinner et al 1990).       However, much of our focus will be on addressing  the practical issues that make it difficult to use our current research prototype over a period of months. We  have identified many such issues, both from our previous evaluations (Section 3.1) and also from a questionnaire that was distributed to school staff during an  in-service day.        Location tracking ? There are problems with both  of the techniques we have tried to date (automatically  reading RFID tags on doors, and asking staff to swipe  location information).  In this project we intend to try  tracking the location of a child using Wi-Fi location  tracking, which seems to be rapidly gaining popularity  in the commercial world (Liu, Sen et al 2008).   Data entry, 2D bar codes ? We need to allow staff  to easily enter and update information about the children  (for example, their timetables) and sensor tags (e.g., if a  new tag is given to a visitor).  For the latter, we want to  investigate 2D bar codes, which could allow encoding  of alphanumeric input data without reference to a central database.   Portability, battery life ? The current system runs on  a tablet PC (8?-12? touch screen, generic or VOCA  hardware). During the evaluation, late powering up, rundown batteries or simply forgetting a component caused  significant data loss and usability issues. A future prototype should favor an ?always-on? system, such as a mobile phone, allowing for easy portability and extended  battery life.   Story generation ? The prototype system was only  able to create a story towards the end of a day and gave  5 only access to stories generated on that day. However,  often the user desired to tell stories that had occurred on  previous days, or to, say, tell a story at lunch that occurred in the morning.  When data was insufficient for  the system to create a story, the only output was an error  message ?Can?t generate story right now.? This frustrated users, so future systems should be able to deliver  a story with incomplete data.    Voice messages ? as mentioned above, we want to  handle these in a more sophisticated way.  From a more  practical perspective, we also want to make it easier for  staff to listen to and change previously recorded messages.  We also want to allow parents to record messages about events at home.  4 Long-term vision and issues  4.1 Supporting children with different levels  and types of impairment  A key issue in AAC is of course the diversity of AAC  users. Children with CCN differ enormously in terms of  cognitive ability, motor ability, and social ability. This  was clear even in our initial evaluation where we  worked just with two children, and discovered that our  interface worked well for Julie but not Jessica.   Julie has little functional speech and severe physical  impairments, and accesses her VOCA using a head  switch through the slow process ofscanning the interface. Her VOCA interface consists of a grid of 15 to 30  buttons per page, with more than 20 pages of vocabulary. However, her cognitive skills were sufficient for  her to master the interface on the second day. She used  the system quite successfully, as shown in the example  in Section 3.2.    Jessica also has severe physical impairments but  does not use technology to support her communication  (she has functional speech).  She has cognitive impairments, which (amongst other things) affect her ability to  remember and place events correctly in time. She had  more difficulty mastering the interface than Julie. We  simplified the interface for her (no editing, minimal  control of narration), and then she displayed pragmatics  known from typical language development in children,  by telling her story with no room for interaction of her  communication partner.    We also need to keep in mind that abilities are not  static, but are likely to progress with age (see also Section 4.2) and (hopefully) with the assistance of communication aids. For example, the WriteTalk project  showed how pupils were both able to initiate and control communication more effectively with Talk:About?  and how their formal writing skills improved over time  (Waller et al, 1999).   In summary, some children may need a very simple  interface because of cognitive impairments, but this  should grow with them.  For example, the best narration  tool for Jessica at her current stage of development is  probably a single button that advances sequentially  through the computer-generated story. The challenge is  to provide an interface that Jessica can initially use via  repeatedly pressing an ?Advance? button, but which  gives her the possibility of exerting more control as her  skills and abilities develop.   Other children (such as Julie) may have motor difficulties that restrict the way in which they can interact  with computer systems, and thus may require simple  controls although they have reasonable cognitive skills.  Restricted motor skills make certain tasks, such as entering an arbitrary word, quite difficult and time consuming; hence the interface must avoid such tasks, and  instead endeavor to give the child as much control as  possible with a minimum amount of data entry.  Once  these users master a basic story telling structure, it may  help them develop their conversation skills if they use a  wide variety of conversation patterns.  For this purpose,  it may be worthwhile for the system to randomly vary  the structure and language used in the narratives.   Still other children, for example on the autistic spectrum, may have good cognitive and motor abilities, but  not have the experience of expressive communication  necessary to develop interactive skills. These children  are more likely to benefit from a system that supports  the pragmatics of language in general and personal narrative inparticular.  For example, children on the high  functioning end of autism may be comfortable with rather advanced software, which can help them adapt their  storytelling according to the intended listener.  Indeed,  giving these children more complex controls, if done  correctly, can make the software fun and challenging in  a positive way.       In the long term, as we broaden the range of children  we work with, there may be overlaps between our work  and research on tools to help typically developing children create stories, such as Robertson and Good (2005),  and also between our work and research on tools to help  adults with CCN tell personal narratives, such as Dempster (2008).  Ideally it would be very nice to combine  these efforts and create a story telling tool that could be  used across the age and impairment spectrum.  4.2 Narrative across the lifespan  We would like our tool to be able to support children  over time, as their abilities grow and as their experiences accumulate. From the perspective of changing  abilities, the challenge is to offer children an interface  which is not only appropriate for their current stage of  development (Sect 4.1), but also allows and indeed en6 courages them to exert more control over story content,  language, and narration as their abilities grow.   We would also like our tool to become a repository  of a child?s personal stories. The ability to relate relevant stories can influence the quality of life, as well as  social development and successful transitions. The life  stories of people who use AAC are often held by parents  and siblings (e.g., stories relating to health care  (Hemsley, Balandin et al 2007)), and there is the inevitable concern that these stories and others are lost as  parents age and siblings move away.    Technology has the potential both to support the  acquisition of conversational skills for people who use  AAC and to provide a repository for life stories. In the  context of our work, it is essential that we provide ways  of enabling children to develop their narrative skills so  that they are more able to manage their own story repository. In terms of development, young children will  narrate recent stories regardless of conversational context. By enabling the child to develop story structure by  scaffolding interaction and enabling children to easily  annotate stories, the child will begin to anticipate and  control conversation.   Conversational narratives have traditionally not been  supported by AAC tools partly due to the fact that they  are so nebulous; they emerge during interactive conversation (to date, events have to be manually input into a  system and it is difficult to predict what events will become a story); ?new? stories are repeated often (to date it  is difficult to save conversation online); as stories age  theyare repeated in context (retrieval is often contextual  e.g. topic based) and they grow longer having more embellishments added to them. The technology we are developing provides an opportunity for children to access  information about personal events over time, which they  can communicate and narrate during a conversation.  They can also evaluate (annotate) their stories, thereby  embellishing and lengthening the stories.  However this  will only be possible if the children can easily access  previously experienced, generated and saved stories.    We can provide fast access to recent stories while  anticipating the use of older stories such as for example  those which closely match the current conversation topic. In a research prototype called PROSE (Waller and  Newell 1997), stories had to be physically tagged; there  is now the potential to automate topic matching by recognizing topic words spoken by a listener and parsing  stored information for appropriate stories. Over a lifetime, some stories may fall into disuse, while others will  be weighted more strongly depending on frequency of  use and relevance.  4.3 True dialogue in narration  The ultimate goal of our research is to enable children to  tell stories in the context of a social dialogue; for example, we want children to be able to chat to their parents  and other interested parties about what they did during  the day.    Our current system incorporates a simple model of a  conversation, where children are restricted at any point  to choosing from a small number of options. The child  chooses an event to talk about, and then goes through  the sequence of messages associated with that event.   The child has the freedom to switch to a different event,  hence controlling the conversation, and to add annotations/evaluations (for example ?it was fun!?).       This is adequate in many cases, but in the long term  we would like to support more complex conversations;  for example interrupting a discussion about today?s  events to talk about what happened yesterday, or to discuss a particular teacher instead of an event.  We would  also like children to easily be able to add conversational  phrases, such as ?Guess what happened today at  school?.   Because our children have motor and cognitive impairments, we cannot present them with a large number  of options for conversational moves. Ideally, the system  would detect what the conversational partner wishes to  talk about, and from this present the child with a small  number of appropriate choices. For example, if the conversational partner asks the child what happened over  the past week, our system would detect this and then  give the child the option of talking about any individual  weekday or the week in general.   One way of detecting what the conversational partner intends is to use speech recognition and Natural  Language Processing(NLP) technology to analyze what  he or she says. Speech and NLP technology tend to  work best when it is possible to train the system to the  user?s voice, and also (in essence) train the user to understand what the speech/NLP system can and cannot  do. This should be possible in our context, at least for  people (such as parents) with whom the child regularly  interacts.   Another possibility is to create a graphical user interface for the conversational partner, perhaps on the  same device that the child uses, which the partner could  use to indicate what he/she wants to talk about.  This is  probably technically easier, but does move away from  the goal of having as natural a dialogue as possible.  4.4 Pragmatics of interacting with others  Currently, ?How was School today??? supports storytelling between language-impaired children and adults  who are the children?s parents, carers, teachers, and  therapists.  But of course for normally developing children, many of their most important social interactions are  with other children.   An interesting example here is the STANDUP system, which was developed to help children who use  7 AAC create and tell novel punning riddles. The study  results suggested that children saved the jokes so that  they could retell them to friends and family (Waller,  Black et al 2009). Whilst the evidence is anecdotal,  there did also appear to be a marked increase in joke  telling by participants, both amongst their peers and  with adults in the home environment. Hence STANDUP  succeeded in supporting interaction with other children  as well as with adults.   One of the key challenges in interacting with other  children, and indeed with adults who are not formally  involved in the care or teaching of the child, is to adapt  the story to the interests of the recipients. In other  words, a child?s parents and teachers will not insist on  stories that are interesting to them, but other conversational partners will.  These conversational partners may  also need additional information.  For example ?Jane  came to take me to the OT room? makes more sense if  the recipient knows that Jane is the occupational therapist; parents and teachers already know this, but other  people may need to be told this. Also if the conversational partner was present at an event, this should be  acknowledged and indeed used in the story. For example, "Did you really enjoy maths? I thought it was boring!?   In short, telling stories to peers and adults who do  not know the child well requires adapting the story to  the interests, knowledge, and involvement of the partner; this is part of learning pragmatics. This is not something we are looking at currently, but it is something  that we hope to look at in the future.  4.5 Security and privacy issues  We needto ensure that data about the children is private  and secure.  Taken to its logical conclusion, our project  would result in an intimate record of the child\'s life at  school, home and beyond.  It is important that both the  raw data and the generated content are under the control  of the child and his/her guardians, with the child exercising as much control as possible. This is especially  important since children with learning difficulties are  very vulnerable; there is potential for great harm if data  about a child?s activities got into the hands of a malicious outsider.   In a study on the software tool TalksBac, which  supports personal narrative (Waller, Dennis et al 1997),  privacy issues were coded along with stories. This allowed the NLG process to decide the appropriateness of  telling a story to a specific communication partner.  Children in general do not care who they tell their stories to. Only when older children learn to distinguish  which story is appropriate for a conversation partner.  This process could be embedded into the prediction  algorithm that presents stories for narration. Currently  prediction on AAC devices only support character, word  or phrase selection.     Another concern is information that is embarrassing  or otherwise puts the child in a negative light; for example, imagine a staff member entered the voice message  "I refused to eat my lunch today".  We believe that the  child should be free to delete such messages; she should  never be forced to include material in a story that she  does not want to include.   A related issue is whether we should allow stories  generated for one child to use information acquired  about another child.  In principle this is very valuable,  for example it allows messages such as ?Jane didn?t eat  her lunch today?. But is this acceptable from the perspective of ensuring the privacy of data about Jane?s activities? On the other hand, this is exactly the sort of  thing that a normally developed child would say about a  classmate. 5 Conclusion  In addition to having difficulty in communicating desires and needs, language-impaired children also find it  hard to participate in social linguistic interaction that  would help create and build up friendships and other  interpersonal relationships.  We believe that we can help  these children participate in such interactions by giving  them a tool that helps them tell a story about their day at  school, by using an NLG system that has access to sensor and other data about the child?s activities. We are  still at an early stage in this work, but our initial prototype system has shown great potential to improve the  quality of life of children with limited speech. Our current work plans to explore this potential further while  evaluating the efficacy of the system for four children  with varying ability levels.  Acknowledgements  We would like to express our thanks to the children,  their parents and staff and the special school where this  project had its base. Without their valuable contributions and feedback this research would not have been  possible. We would also like to thank DynaVox Systems Ltd for supplying the communication devices to  run our system on.   This research was supported by the UK Engineering  and Physical Sciences Research Council under grants  EP/F067151/1, EP/F066880/1, EP/E011764/1,  EP/H022376/1, and EP/H022570/1.  8'
1.0	E06-2024	A00-1031	b'1 Introduction The purpose of this paper is to present LX-Suite, a set of tools for the shallow processing of Portuguese, developed under the TagShare1 project by the NLX Group.2 The tools included in this suite are a sentence chunker; a tokenizer; a POS tagger; a nominal featurizer; a nominal lemmatizer; and a verbal featurizer and lemmatizer. These tools were implemented as autonomous modules. This option allows to easily replace any of the modules by an updated version or even by a third-party tool. It also allows to use any of these tools separately, outside the pipeline of the suite. The evaluation results mentioned in the next sections have been obtained using an accurately hand-tagged 280, 000 token corpus composed of newspaper articles and short novels. 2 Sentence chunker The sentence chunker is a finite state automaton (FSA), where the state transitions are triggered by specified character sequences in the input, and the emitted symbols correspond to sentence (<s>) and paragraph (<p>) boundaries. Within this setup, a transition rule could define, for example, 1http://tagshare.di.fc.ul.pt 2NLX?Natural Language and Speech Group, at the Department of Informatics of the University of Lisbon, Faculty of Sciences: http://nlx.di.fc.ul.pt that a period, when followed by a space and a capital letter, marks a sentence boundary: ?? ? ?. A? ? ?? ? ?? ? ?.</s><s>A? ? ?? Being a rule-based chunker, it was tailored to handle orthographic conventions that are specific to Portuguese, in particular those governing dialog excerpts. This allowed the tool to reach a very good performance, with values of 99.95% for recall and 99.92% for precision.3 3 Tokenizer Tokenization is, for the most part, a simple task, as the whitespace character is used to mark most token boundaries. Most of other cases are also rather simple: Punctuation symbols are separated from words, contracted forms are expanded and clitics in enclisis or mesoclisis position are detached from verbs. It is worth noting that the first element of an expanded contraction is marked with a symbol (+) indicating that, originally, that token occurred as part of a contraction:4 um, dois ?|um|,|dois| da ?|de+|a| viu-o ?|viu|-o| In what concerns Portuguese, the non-trivial aspects of tokenization are found in the handling of ambiguous strings that, depending on their POS tag, may or may not be considered a contraction. For example, the word deste can be tokenized as the single token |deste| if it occurs as a verb (Eng.: [you] gave) or as the two tokens |de+|este| if it occurs as a contraction (Eng.: of this). 3For more details, see (Branco and Silva, 2004). 4In these examples the | symbol will be used to mark token boundaries more clearly. 179 It is worth noting that this problem is not a minor issue, as these strings amount to 2% of the corpus that was used and any tokenization error will have a considerable negative influence on the subsequent steps of processing, such as POS tagging. To resolve the issue of ambiguous strings, a two-stage tokenization strategy is used, where the ambiguous strings are not immediately tokenized. Instead, the decision counts onthe contribution of the POS tagger: The tagger must first be trained on a version of the corpus where the ambiguous strings are not tokenized, and are tagged with a composite tag when occurring as a contraction (for example P+DEM for a contraction of a preposition and a demonstrative). The tagger then runs over the text and assigns a simple or a composite tag to the ambiguous strings. A second pass with the tokenizer then looks for occurrences of tokens with a composite tag and splits them: deste/V?|deste/V| deste/P+DEM?|de+/P|este/DEM| This approach allowed us to successfully resolve 99.4% of the ambiguous strings. This is a much better value than the baseline 78.20% obtained by always considering that the ambiguous strings are a contraction.5 4 POS tagger For the POS tagging task we used Brant?s TnT tagger (Brants, 2000), a very efficient statistical tagger based on Hidden Markov Models. For training, we used 90% of a 280, 000 token corpus, accurately hand-tagged with a tagset of ca. 60 tags, with inflectional feature values left aside. Evaluation showed an accuracy of 96.87% for this tool, obtained by averaging 10 test runs over different 10% contiguous portions of the corpus that were not used for training. The POS tagger we developed is currently the fastest tagger for the Portuguese language, and it is in line with state-of-the-art taggers for other languages, as discussed in (Branco and Silva, 2004). 5 Nominal featurizer This tool assigns feature value tags for inflection (Gender and Number) and degree (Diminutive, Superlative and Comparative) to words from nominal morphosyntactic categories. 5For further details see (Branco and Silva, 2003). Such tagging is typically done by a POS tagger, by using a tagset where the base POS tags have been extended with feature values. However, this increase in the number of tags leads to a lower tagging accuracy due to the data-sparseness problem. With our tool, we explored what could be gained by having a dedicated tool for the task of nominal featurization. We tried several approaches to nominal featurization. Here we report on the rule-based approach which is the one that better highlights the difficulties in this task. For this tool, we built on morphological regularities and used a set of rules that, depending on the word termination, assign default feature values to words. Naturally, these rules were supplemented by a list of exceptions, which was collected by using an machine readable dictionary (MRD) that allowed us to search words by termination. Nevertheless, this procedure is still not enough to assign a feature value to every token. The most direct reason is due to the so-called invariant words, which are lexically ambiguous with respect to feature values. For example, the Common Noun ermita (Eng.: hermit) can be masculine or feminine, depending on the occurrence. By simply using termination rules supplemented with exceptions, such words will always be tagged with underspecified feature values:6 ermita/?S To handle such cases the featurizer makes use of feature propagation. With this mechanism, words from closed classes, for which we know their feature values, propagate their values to thewords from open classes following them. These words, in turn, propagate those features to other words: o/MS ermita/MS humilde/MS Eng.: the-MS humble-MS hermit-MS but a/FS ermita/FS humilde/FS Eng.: the-FS humble-FS hermit-FS Special care must be taken to avoid that feature propagation reaches outside NP boundaries. For this purpose, some sequences of POS categories block feature propagation. In the example below, a PP inside an NP context, azul (an ?invariant? 6Values: M:masculine, F:feminine, S:singular, P:plural and ?:undefined. 180 adjective) might agree with faca or with the preceding word, ac?o. To prevent mistakes, propagation from ac?o to azul should be blocked. faca/FS de ac?o/MS azul/FS Eng.: blue (steel knife) or faca/FS de ac?o/MS azul/MS Eng.: (blue steel) knife For the sake of comparability with other possible similar tools, we evaluated the featurizer only over Adjectives and Common Nouns: It has 95.05% recall (leaving ca. 5% of the tokens with underspecified tags) and 99.05% precision.7 6 Nominal lemmatizer Nominal lemmatization consists in assigning to Adjectives and Common Nouns a normalized form, typically the masculine singular if available. Our approach uses a list of transformation rules that helps changing the termination of the words. For example, one states that any word ending in ta should have that ending transformed into to: gata ([female] cat) ? gato ([male] cat) There are, however, exceptions that must be accounted for. The word porta, for example, is a feminine common noun, and its lemma is porta: porta (door, feminine common noun) ? porta Relevant exceptions like the one above were collected by resorting to a MRD that allowed to search words on the basis of their termination. Being that dictionaries only list lemmas (and not inflected forms), it is possible to search for words with terminations matching the termination of inflected words (for example, words ending in ta). Any word found by the search can thus be considered as an exception. A major difficulty in this task lies in the listing of exceptions when non-inflectional affixes are taken into account. As an example, lets consider again the word porta. This word is an exception to the rule that transforms ta into to. As expected, this word can occur prefixed, as in superporta. Therefore, this derived word 7For a much more extensive analysis, including a comparison with other approaches, see (Branco and Silva, 2005a). should also appear in the list of exceptions to prevent it from being lemmatized into superporto by the rule. However, proceeding like this for every possible prefix leads to an explosion in the number of exceptions. To avoid this, a mechanism was used that progressively strips prefixes from words while checking the resulting word forms against the list of exceptions: supergata -----gata (apply rule) ? supergato but superporta -----porta (exception) ? superporta A similar problem arises when tackling words with suffixes. For instance, the suffix -zinho and its inflected forms (-zinha, -zinhos and -zinhas) are used as diminutives. These suffixes should be removed by the lemmatization process. However, there are exceptions, such as the word vizinho (Eng.: neighbor) which is not a diminutive. This word has tobe listed as an exception, together with its inflected forms (vizinha, vizinhos and vizinhas), which again leads to a great increase in the number of exceptions. To avoid this, only vizinho is explicitly listed as an exception and the inflected forms of the diminutive are progressively undone while looking for an exception: vizinhas (feminine plural) vizinha (feminine singular) vizinho (exception) ? vizinho To ensure that exceptions will not be overlooked, when both these mechanisms work in parallel one must follow all possible paths of affix removal. An heuristic chooses the lemma as being the result found in the least number of steps.8 To illustrate this, consider the word antena (Eng.: antenna). Figure 1 shows the paths followed by the lemmatization algorithm when it is faced with antenazinha (Eng.: [small] antenna). Both ante- and -zinha are possible affixes. In a first step, two search branches are opened, the first where ante- is removed and the second where -zinha is transformed into 8This can be seen as following a rationale similar to Karlsson?s (1990) local disambiguation procedure. 181 antenazinha nazinha antenazinho nazinho nazinho antena na no na no 0 1 2 3 4 Figure 1: Lemmatization of antenazinha -zinho. The search proceeds under each branch until no transformation is possible, or an exception has been found. The end result is the ?leaf node? with the shortest depth which, in this example, is antena (an exception). This branching might seem to lead to a great performance penalty, but only a few words have affixes, and most of them have only one, in which case there is no branching at all. This tool evaluates to an accuracy of 94.75%.9 7 Verbal featurizer and lemmatizer To each verbal token, this tool assigns the corresponding lemma and tag with feature values for Mood, Tense, Person and Number. The tool uses a list of rules that, depending on the termination of the word, assign all possible lemma-feature pairs. The word diria, for example, is assigned the following lemma-feature pairs: diria ? ?dizer,Cond-1ps? ? ?dizer,Cond-3ps? ? ?diriar,PresInd-3ps? ? ?diriar,ImpAfirm-2ps? Currently, this tool does not attempt to disambiguate among the proposed lemma-feature pairs. So, each verbal token will be tagged with all its possible lemma-feature pairs. The tool was evaluated over a list with ca. 800, 000 verbal forms. It achieves 100% precision, but at 50% recall, as half of those forms are ambiguous and receive more than one lemmafeature pair. 9For further details, see (Branco and Silva, 2005b). 8 Final Remarks So far, LX-Suite has mostly been used in-house for projects being developed by the NLX Group. It is being used in the GramaXing project, where a computational core grammar for deep linguistic processing of Portuguese is being developed under the Delphin initiative.10 In collaboration with CLUL,11 and under the TagShare project, LX-Suite is being used to help in the building of a corpus of 1 million accurately hand-tagged tokens, by providing an initial, highquality tagging which is then manually corrected. It is also used for the QueXting project, whose aim is to make available a question answering system onthe Portuguese Web. There is an on-line demo of LX-Suite located at http://lxsuite.di.fc.ul.pt. This on-line version of the suite is a partial demo, as it currently only includes the modules up to the POS tagger. By the end of the TagShare project (mid-2006), all the other modules described in this paper are planned to have been included. Additionally, the verbal featurizer and lemmatizer can be tested as a standalone tool at http://lxlemmatizer.di.fc.ul.pt. Future work will be focused on extending the suite with new tools, such as a named-entity recognizer and a phrase chunker.'	b'1 In t roduct ion   A large number of current language processing sys-  tems use a part-of-speech tagger for pre-processing.  The tagger assigns a (unique or ambiguous) part-of-  speech tag to each token in the input and passes its  output o the next processing level, usually a parser.  Furthermore, there is a large interest in part-of-  speech tagging for corpus annotation projects, who  create valuable linguistic resources by a combination  of automatic processing and human correction.  For both applications, a tagger with the highest  possible accuracy is required. The debate about  which paradigm solves the part-of-speech tagging  problem best is not finished. Recent comparisons  of approaches that can be trained on corpora (van  Halteren et al, 1998; Volk and Schneider, 1998) have  shown that in most cases statistical aproaches (Cut-  ting et al, 1992; Schmid, 1995; Ratnaparkhi, 1996)  yield better results than finite-state, rule-based, or  memory-based taggers (Brill, 1993; Daelemans etal.,  1996). They are only surpassed by combinations of different systems, forming a "voting tagger".  Among the statistical approaches, the Maximum  Entropy framework has a very strong position. Nev-  ertheless, a recent independent comparison of 7 tag-  gets (Zavrel and Daelemans, 1999) has shown that  another approach even works better: Markov mod-  els combined with a good smoothing technique and  with handling of unknown words. This tagger, TnT,  not only yielded the highest accuracy, it also was the  fastest both in training and tagging.  The tagger comparison was organized as a "black-  box test": set the same task to every tagger and  compare the outcomes. This paper describes the  models and techniques used by TnT together with  the implementation.  The reader will be surprised how simple the under-  lying model is. The result of the tagger comparison  seems to support he maxime "the simplest is the  best". However, in this paper we clarify a number  of details that are omitted in major previous pub-  lications concerning tagging with Markov models.  As two examples, (Rabiner, 1989) and (Charniak  et al, 1993) give good overviews of the techniques  and equations used for Markov models and part-of-  speech tagging, but they are not very explicit in the  details that are needed for their application. We ar-  gue that it is not only the choice of the general model  that determines the result of the tagger but also the  various "small" decisions on alternatives.  The aim of this paper is to give a detailed ac-  count of the techniques used in TnT. Additionally,  we present results of the tagger on the NEGRA cor-  pus (Brants et al, 1999) and the Penn Treebank  (Marcus et al, 1993). The Penn Treebank results  reported here for the Markov model approach are  at least equivalent to those reported for the Maxi-  mum Entropyapproach in (Ratnaparkhi, 1996). For  a comparison to other taggers, the reader is referred  to (Zavrel and Daelemans, 1999).  2 Arch i tec ture   2.1 The Under ly ing Model   TnT uses second order Markov models for part-of-  speech tagging. The states of the model represent  tags, outputs represent the words. Transition prob-  abilities depend on the states, thus pairs of tags.  Output probabilities only depend on the most re-  cent category. To be explicit, we calculate  argmax P(tilti-1, ti-2)P(wilti P(tr+l ItT)  ti...iT  (i)  for a given sequence of words w I . . .  W T of length T.  tl... tT are elements of the tagset, the additional  224 tags t - l ,  to, and tT+l are beginning-of-sequence  and end-of-sequence markers. Using these additional  tags, even if they stem from rudimentary process-  ing of punctuation marks, slightly improves tagging  results. This is different from formulas presented  in other publications, which just stop with a "loose  end" at the last word. If sentence boundaries are  not marked in the input, TnT adds these tags if it  encounters one of \\[.!?;\\] as a token.  Transition and output probabilities are estimated  from a tagged corpus. As a first step, we use the  maximum likelihood probabilities /5 which are de-  rived from the relative frequencies:  Unigrams: /5(t3) = f(t3) (2)  N  f(t2, t3) (3)  Bigrams: P(t31t~)= f(t2)  f(ta, t2, t3)  Trigrams: /5(t3ltx,t~) - f ( t l , t2)  (4)  Lexical: /5(w3 It3) - /(w3, t3) (5)  f(t3)  for all tl, t2, t3 in the tagset and w3 in the lexi-  con. N is the total number of tokens in the training  corpus. We define a maximum likelihood probabil-  ity to be zero if the corresponding nominators and  denominators are zero. As a second step, contex-  tual frequencies are smoothed and lexical frequences  are completed by handling words that are not in the  lexicon (see below).  2.2 Smooth ing   Trigram probabilities generated from a corpus usu-  ally cannot directly be used because of the sparse-  data problem. This means that there are not enough  instances for each trigram to reliably estimate the  probability. Furthermore, setting a probability to  zero because the corresponding trigram never oc-  cured in the corpus has an undesired effect. It causes  the probability of a complete sequence to be set to  zero if its use is necessary for a new text sequence,  thus makes it impossible to rank different sequences  containing a zero probability.  The smoothing paradigm that delivers the best  results in TnT is linear interpolation of unigrams,  bigrams, and trigrams. Therefore, we estimate a  trigram probability as follows:  P(t3ltl ,  t2) = AlP(t3) + Ag_/5(t31t2) + A3/5(t3\\[t1, t2)  (6) /5 are maximum likelihood estimates of the proba-  bilities, and A1 + A2 + A3 = 1, so P again represent  probability distributions.  We use the context-independent variant of linear  interpolation, i.e., the values of the As do not depend  on the particular trigram. Contrary to intuition,  this yields better esults than the context-dependent  variant. Due to sparse-data problems, one cannot es-  timate a different set of As for each trigram. There-  fore, it is common practice to group trigrams by fre-  quency and estimate tied sets of As. However, we  are not aware of any publication that has investi-  gated frequency groupings for linear interpolation i part-of-speech tagging. All groupings that we have  tested yielded at most equivalent results to context-  independent linear interpolation. Some groupings  even yielded worse results. The tested groupings  included a) one set of As for each frequency value  and b) two classes (low and high frequency) on the  two ends of the scale, as well as several groupings  in between and several settings for partitioning the  classes.  The values of Ax, A2, and A3 are estimated by  deleted interpolation. This technique successively  removes each trigram from the training corpus and  estimates best values for the As from all other n-  grams in the corpus. Given the frequency counts  for uni-, bi-, and trigrams, the weights can be very  efficiently determined with a processing time linear  in the number of different rigrams. The algorithm  is given in figure 1. Note that subtracting 1 means  taking unseen data into account. Without this sub-  traction the model would overfit the training data  and would generally ield worse results.  2.3 Handling of Unknown Words  Currently, the method of handling unknown words  that seems to work best for inflected languages is  a suffix analysis as proposed in (Samuelsson, 1993).  Tag probabilities are set according to the word\'s end-  ing. The suffix is a strong predictor for word classes,  e.g., words in the Wall Street Journal part of the  Penn Treebank ending in able are adjectives (JJ) in  98% of the cases (e.g. fashionable, variable), the rest  of 2% are nouns (e.g. cable, variable).  The probability distribution for a particular suf-  fix is generated from all words in the training set  that share the same suffix of some predefined max-  imum length. The term suffix as used here means  "final sequence of characters of a word" which is not  necessarily a linguistically meaningful suffix.  Probabilities are smoothed by successive abstrac-  tion. This calculates the probability of a tag t  given the last m letters li of an n letter word:  P(t l ln - r ,+l , . . . ln) .  The sequence of increasingly  more general contexts omits more and more char-  acters ofthe suffix, such that P(t l ln_m+2,. . . , ln) ,   P(t \\ [ ln-m+3,. . . , l~) ,  . . . ,  P(t) are used for smooth-  ing. The recursiou formula is  P(t l l ,_ i+l,  . . . ln)  = P( t l ln - i+ l , . . .  In) + O iP ( t l l~- , . . . ,  ln) (7)  1 +0~  225 set  )%1 ---- )%2 = )%3 = 0  fo reach  t r ig ram tl,t2,t3 with  f(t l ,t2,t3) > 0  depending on the maximum of the fo l low ing  three  va lues :   f(tl ,t2,ts)-- 1 case f (h,t2)- I  " increment )%3 by f ( t l , t2 , t3)   f(t2,t3)-I case f(t2)- I  " increment )%2 by f ( t l , t2 , t s )   f ( t3 ) - - i   case N-1  " increment )%1 by f(t l ,t2,t3)  end  end  normalize )%1, )%2, )%3  Figure 1: Algorithm for calculting the weights for context-independent li ear interpolation )%1, )%2, )%3 when  the n-gram frequencies are known. N is the size of the corpus? If the denominator in one of the expressions  is 0, we define the result of that expression to be 0.  for i = m. . .  0, using the maximum likelihood esti-  mates/5  from frequencies in the lexicon, weights Oi  and the initialization  P(t) =/5(t) .  (8)  The maximum likelihood estimate for a suffix of  length i is derived from corpus frequencies by  P(t i /~- i+l ,  .. l~) = f (t ,  1~-~+1,... l~) ? (9 )   For the Markov model, we need the inverse condi-  tional probabilities P ( / , - i+ l , . . .  lnlt) which are ob-  tained by Bayesian inversion.  A theoretical motivated argumentation uses the  standard eviation of the maximum likelihood prob-  abilities for the weights 0i (Samuelsson, 1993)?  This leaves room for interpretation.  1) One has to identify a good value for m, the  longest suffix used? The approach taken for TnT is  the following: m depends on the word in question.  We use the longest suffix that we can find in the  training set (i.e., for which the frequency is greater  than or equal to 1), but at most 10 characters. This  is an empirically determined choice?  2) We use a context-independent approach for 0i,  as we did for the contextual weights )%i. It turned  out to be a good choice to set al 0i to the standard  deviation of the unconditioned maximum likelihood  probabilities of the tags in the training corpus, i.e.,  weset  1  $  Oi = ~--~(/5(tj) - 15)2 (10)  s 1 j= l   for all i = 0 . . .  m - 1, using a tagset of s tags and  the average  $  /5 = 1 ~/5(t j )  (11)  8  j----I  This usually yields values in the range 0.03 .. .  0.10.  3) We use different estimates for uppercase and  lowercase words, i.e., we maintain two different suffix  tries depending on the capitalization of the word.  This information improves the tagging results?  4) Another freedom concerns the choice of the  words in the lexicon that should be used for suf-  fix handling. Should we use all words, or are some  of them better suited than others? Accepting that  unknown words are most probably infrequent, one  can argue that using suffixes of infrequent words in  the lexicon is a better approximation for unknown  words than using suffixes of frequent words. There-  fore, we restrict the procedure of suffix handling to  words with a frequency smaller than or equal to some  threshold value. Empirically, 10 turned out to be a  good choice for this threshold.  2.4 Cap i ta l i za t ion   Additional information that turned out to be use-  ful for the disambiguation process for several cor-  pora and tagsets is capitalization information. Tags  are usually not informative about capitalization, but  probability distributions of tags around capitalized  words are different from those not capitalized. The  effect is larger for English, which only capitalizes  proper names, and smaller for German, which capi-  talizes all nouns.  We use flags ci that are true if wi is a capitalized  word and false otherwise? These flags are added to  the contextual probability distributions. Instead of  P(tsItl,t2) (12)  we use  P(t3, c3\\[tl , cl, t2, c2) (13)  and equations (3) to (5) are updated accordingly.  This is equivalent o doubling the size of the tagset  and using different ags depending on capitalization.  226 2.5 Beam Search  The processing time of the Viterbi algorithm (Ra-  biner, 1989) can be reduced by introducing a beam  search. Each state that receives a 5 value smaller  than the largest 5 divided by some threshold value  is excluded from further processing. While the  Viterbi algorithm is guaranteed to find the sequence  of states with the highest probability, this is no  longer true when beam search is added. Neverthe-  less, for practical purposes and the right choice of  0, there is virtually no difference between the algo-  rithm with and without a beam. Empirically, a value  of 0 = 1000 turned out to approximately double the  speed of the tagger without affecting the accuracy.  The tagger currently tags between 30~000 and  60,000 tokensper second (including file I/O) on a  Pentium 500 running Linux. The speed mainly de-  pends on the percentage of unknown words and on  the average ambiguity rate.  3 Eva luat ion   We evaluate the tagger\'s performance under several  aspects. First of all, we determine the tagging ac-  curacy averaged over ten iterations. The overall ac-  curacy, as well as separate accuracies for known and  unknown words are measured.  Second, learning curves are presented, that indi-  cate the performance when using training corpora of  different sizes, starting with as few as 1,000 tokens  and ranging to the size of the entire corpus (minus  the test set).  An important characteristic of statistical taggers  is that they not only assign tags to words but also  probabilities in order to rank different assignments.  We distinguish reliable from unreliable assignments  by the quotient of the best and second best assign-  ments 1. All assignments for which this quotient is  larger than some threshold are regarded as reliable,  the others as unreliable. As we will see below, accu-  racies for reliable assignments are much higher.  The tests are performed on partitions of the cor-  pora that use 90% as training set and 10% as test  set, so that the test data is guaranteed to be unseen  during training. Each result is obtained by repeat-  ing the experiment 10 times with different partitions  and averaging the single outcomes.  In all experiments, contiguous test sets are used.  The alternative is a round-robin procedure that puts  every 10th sentence into the test set. We argue that  contiguous test sets yield more realistic results be-  cause completely unseen articles are tagged. Using  the round-robin procedure, parts of an article are al-  ready seen, which significantly reduces the percent-  age of unknown words. Therefore, we expect even  1 By definition, this quotient is co if there is only one pos-  sible tag for a given word.  higher results when testing on every 10th sentence  instead of a contiguous et of 10%.  In the following, accuracy denotes the number of  correctly assigned tags divided by the number of to-  kens in the corpus processed. The tagger is allowed  to assign exactly one tag to each token.  We distinguish the overall accuracy, taking into  account all tokens in the test corpus, and separate  accuracies for known and unknown tokens. The lat-  ter are interesting, since usually unknown tokens are  much more difficult to process than known tokens,  for which a list of valid tags can be found in the  lexicon.  3.1 Tagging the  NEGRA corpus   The German NEGRA corpus consists of 20,000 sen-  tences (355,000 tokens) of newspaper texts (Frank-  furter Rundschau) that are annotated with parts-of-  speech and predicate-argument structures (Skut et al., 1997). It was developed at the Saarland Univer-  sity in Saarbrficken 2. Part of it was tagged at the  IMS Stuttgart. This evaluation only uses the part-  of-speech annotation and ignores structural annota-  tions.  Tagging accuracies for the NEGRA corpus are  shown in table 2.  Figure 3 shows the learning curve of the tagger,  i.e., the accuracy depending on the amount of train-  ing data. Training length is the nmnber of tokens  used for training. Each training length was tested  ten times, training and test sets were randomly cho-  sen and disjoint, results were averaged. The training  length is given on a logarithmic scale.  It is remarkable that tagging accuracy for known  words is very high even for very small training cot-  pora. This means that we have a good chance of  getting the right tag if a word is seen at least once  during training. Average percentages of unknown  tokens are shown in the bottom line of each diagram.  We exploit the fact that the tagger not only de-  termines tags, but also assigns probabilities. If there  is an alternative that has a probability "close to"  that of the best assignment, his alternative can be  viewed as almost equally well suited. The notion of  "close to" is expressed by the distance of probabil-  ities, and this in turn is expressed by the quotient  of probabilities. So, the distance of the probabili-  ties of a best tag tbest and an alternative tag tart  is expressed by P(tbest)/p(tau), which is some value  greater or equal to 1 since the best tag assignment  has the highest probability.  Figure 4 shows the accuracy when separating as-  signments with quotients larger and smaller than  the threshold (hence reliable and unreliable assign-  ments). As expected, we find that accuracies for  2For availability, please check  h~tp ://w~. col i. uni-sb, de/s fb378/negra-corpus  777 227 Table 2: Part-of-speech tagging accuracy for the NEGRA corpus, averaged over 10 test runs, training and  test set are disjoint. The table shows the percentage of unknown tokens, separate accuracies and standard  deviations for known and unknown tokens, as well as the overall accuracy.  percentage  unknowns  NEGRA corpus 11.9%  known  ace .   97.7% 0.23  unknown  acc. (x  89.0% 0.72  overall  aCE. o"  96.7% 0.29  NEGRA Corpus :  POS Learn ing  Curve   100  9O  80  70 /S   6O  50 , i  1 2 5  50.8 46.4 41.4  I i i I I I i  10 20 50 100 200 320 500  36.0 30.7 23.0 18.3 14.3 11.9 11).3  Overall  min =78.1%  max=96.7%  e Known  rain =95.7%  max=97.7%  - - -a- - -  Unknown  rain =61.2% max=89.0%  I  1000 x 1000 Training Length  St  avg. percentage unknown  Figure 3: Learning curve for tagging the NEGRA corpus. The training sets of variable sizes as well as test  sets of 30,000 tokens were randomly chosen. Training and test sets were disjoint, the procedure was repeated  10 times and results were averaged. Percentages of unknowns for 500k and 1000k training are determined  from an untagged extension.  NEGRA Corpus: Accuracy of  re l iab le  assignments  100  99  98  97  A  f " / :. Reliable rain =96.7% max=99.4%  96 I i i i i i i i i i i  2 5 10 20 50 100 500 2000 10000 threshold 0  100 97.9 95.1 92.7 90.3 86.8 84.1 81.0 76.1 71.9 68.3 64.1 62.0 % cases reliable  - 53.5 62.9 69.6 74.5 79.8 82.7 85.2 88.0 89.6 90.8 91.8 92.2 acc. of complement  Figure 4: Tagging accuracy for the NEGRA corpus when separating reliable and unreliable assignments. The  curve shows accuracies for reliable assignments. The numbers at the bottom line indicate the percentage of  reliable assignments and the accuracy of the complement set (i.e., unreliable assignments).  228 Table 5: Part-of-speech tagging accuracy for the Penn Treebank. The table shows the percentage of unknown  tokens, separate accuracies and standard deviations for known and unknown tokens, as well as the overall  accuracy.  I percentage known  unknowns acc. a  Penn Treebank 2.9% 97.0% 0.15  unknown  aCC. O"  85.5% 0.69  overall  aCE. O"  96.7% 0.15  100  9O  80  70 <  60  Penn Treebank: POS Learn ing  Curve   /  50 I I ~ i I I I i I  1 2 5 10 20 50 100 200 500  50.3 42.8 33.4 26.8 20.2 13.2 9.8 7.0 4.4  Overall  rain =78.6%  max=96.7%  Known  rain =95.2%  max=97.0%  Unknown  min =62.2%  max=85.5%  I  1000 ? 1000 Training Length  2.9 avg. percentage unknown  Figure 6: Learning curve for tagging the Penn Treebank. The training sets of variable sizes as well as test sets  of 100,000 tokens were randomly chosen. Training and test sets were disjoint, the procedure was repeated  10 times and results were averaged.  Penn Treebank: Accuracy of reliable assignments  100  99  98  97  Overall  rain =96.6%  max=99.4%  96 i i i i t i i i i i i i  2 5 10 20 50 100 500 2000 10000 threshold 0  100 97.7 94.6 92.2 89.8 86.3 83.5 80.4 76.6 73.8 71.0 67.2 64.5 % cases reliable  - 53.5 62.8 68.9 73.9 79.3 82.6 85.2 87.5 88.8 89.8 91.0 91.6 acc. of complement  Figure 7: Tagging accuracy for the Penn Treebank when separating reliable and unreliable assignments. The  curve shows accuraciesfor reliable assignments. The numbers at the bottom line indicate the percentage of  reliable assignments and the accuracy of the complement set.  229  reliable assignments are much higher than for unre-  liable assignments. This distinction is, e.g., useful  for annotation projects during the cleaning process,  or during pre-processing, sothe tagger can emit mul-  tiple tags if the best tag is classified as unreliable.  3.2 Tagging the Penn Treebank   We use the Wall Street Journal as contained in the  Penn Treebank for our experiments. The annotation  consists of four parts: 1) a context-free structure  augmented with traces to mark movement and dis-  continuous constituents, 2) phrasal categories that  are annotated as node labels, 3) a small set of gram-  matical functions that are annotated as extensions to  the node labels, and 4) part-of-speech tags (Marcus  et al, 1993). This evaluation only uses the part-of-  speech annotation.  The Wall Street Journal part of the Penn Tree-  bank consists of approx. 50,000 sentences (1.2 mil-  lion tokens).  Tagging accuracies for the Penn Treebank are  shown in table 5. Figure 6 shows the learning curve  of the tagger, i.e., the accuracy depending on the  amount of training data. Training length is the num-  ber of tokens used for training. Each training length  was tested ten times. Training and test sets were  disjoint, results are averaged. The training length is  given on a logarithmic scale. As for the NEGRA cor-  pus, tagging accuracy is very high for known tokens  even with small amounts of training data.  We exploit the fact that the tagger not only de-  termines tags, but also assigns probabilities. Figure  7 shows the accuracy when separating assignments  with quotients larger and smaller than the threshold  (hence reliable and unreliable assignments). Again,  we find that accuracies for reliable assignments are  much higher than for unreliable assignments.  3.3 Summary of Part-of-Speech Tagging  Results  Average part-of-speech tagging accuracy is between  96% and 97%, depending on language and tagset,  which is at least on a par with state-of-the-art re-  sults found in the literature, possibly better. For  the Penn Treebank, (Ratnaparkhi, 1996) reports an  accuracy of 96.6% using the Maximum Entropy ap-  proach, our much simpler and therefore faster HMM  approach delivers 96.7%. This comparison eeds to  be re-examined, since we use a ten-fold crossvalida-  tion and averaging of results while Ratnaparkhi only  makes one test run.  The accuracy for known tokens is significantly  higher than for unknown tokens. For the German  newspaper data, results are 8.7% better when the  word was seen before and therefore is in the lexicon,  than when it was not seen before (97.7% vs. 89.0%).  Accuracy for known tokens is high even with very  small amounts of training data. As few as 1000 to- kens are sufficient o achieve 95%-96% accuracy for  them. It is important for the tagger to have seen a  word at least once during training.  Stochastic taggers assign probabilities to tags. We  exploit the probabilities to determine reliability of  assignments. For a subset hat is determined uring  processing by the tagger we achieve accuracy rates  of over 99%. The accuracy of the complement set is  much lower. This information can, e.g., be exploited  in an annotation project to give an additional treat-  ment to the unreliable assignments, or to pass se-  lected ambiguities to a subsequent processing step. 4 Conc lus ion   We have shown that a tagger based on Markov mod-  els yields state-of-the-art results, despite contrary  claims found in the literature. For example, the  Markov model tagger used in the comparison of (van  Halteren et al, 1998) yielded worse results than all  other taggers. In our opinion, a reason for the wrong  claim is that the basic algorithms leave several deci-  sions to the implementor. The rather large amount  of freedom was not handled in detail in previous pub-  lications: handling of start- and end-of-sequence, the  exact smoothing technique, how to determine the  weights for context probabilities, details on handling  unknown words, and how to determine the weights  for unknown words. Note that the decisions we made  yield good results for both the German and the En-  glish Corpus. They do so for several other corpora  as well. The architecture remains applicable to a  large variety of languages.  According to current tagger comparisons (van  Halteren et al, 1998; Zavrel and Daelemans, 1999),  and according to a comparsion of the results pre-  sented here with those in (Ratnaparkhi, 1996), the  Maximum Entropy framework seems to be the only  other approach yielding comparable results to the  one presented here. It is a very interesting future  research topic to determine the advantages of either  of these approaches, to find the reason for their high  accuracies, and to find a good combination of both.  TnT is freely available to universities and re-  lated organizations for research purposes (see  http ://www. coli. uni-sb, de/-thorsten/tnt). Acknowledgements   Many thanks go to Hans Uszkoreit for his sup-  port during the development of TnT. Most of the  work on TnT was carried out while the author  received a grant of the Deutsche Forschungsge-  meinschaft in the Graduiertenkolleg Kognitionswis-  senschaft Saarbriicken. Large annotated corpora re  the pre-requisite for developing and testing part-of-  speech taggers, and they enable the generation of  high-quality language models. Therefore, I would  230  like to thank all the people who took the effort  to annotate the Penn Treebank, the Susanne Cor-  pus, the Stuttgarter Referenzkorpus, the NEGRA  Corpus, the Verbmobil Corpora, and several others.  And, last but not least, I would like to thank the  users of TnT who provided me with bug reports and  valuable suggestions for improvements. '
0.0	E06-2024	D11-1133	b'1 Introduction The purpose of this paper is to present LX-Suite, a set of tools for the shallow processing of Portuguese, developed under the TagShare1 project by the NLX Group.2 The tools included in this suite are a sentence chunker; a tokenizer; a POS tagger; a nominal featurizer; a nominal lemmatizer; and a verbal featurizer and lemmatizer. These tools were implemented as autonomous modules. This option allows to easily replace any of the modules by an updated version or even by a third-party tool. It also allows to use any of these tools separately, outside the pipeline of the suite. The evaluation results mentioned in the next sections have been obtained using an accurately hand-tagged 280, 000 token corpus composed of newspaper articles and short novels. 2 Sentence chunker The sentence chunker is a finite state automaton (FSA), where the state transitions are triggered by specified character sequences in the input, and the emitted symbols correspond to sentence (<s>) and paragraph (<p>) boundaries. Within this setup, a transition rule could define, for example, 1http://tagshare.di.fc.ul.pt 2NLX?Natural Language and Speech Group, at the Department of Informatics of the University of Lisbon, Faculty of Sciences: http://nlx.di.fc.ul.pt that a period, when followed by a space and a capital letter, marks a sentence boundary: ?? ? ?. A? ? ?? ? ?? ? ?.</s><s>A? ? ?? Being a rule-based chunker, it was tailored to handle orthographic conventions that are specific to Portuguese, in particular those governing dialog excerpts. This allowed the tool to reach a very good performance, with values of 99.95% for recall and 99.92% for precision.3 3 Tokenizer Tokenization is, for the most part, a simple task, as the whitespace character is used to mark most token boundaries. Most of other cases are also rather simple: Punctuation symbols are separated from words, contracted forms are expanded and clitics in enclisis or mesoclisis position are detached from verbs. It is worth noting that the first element of an expanded contraction is marked with a symbol (+) indicating that, originally, that token occurred as part of a contraction:4 um, dois ?|um|,|dois| da ?|de+|a| viu-o ?|viu|-o| In what concerns Portuguese, the non-trivial aspects of tokenization are found in the handling of ambiguous strings that, depending on their POS tag, may or may not be considered a contraction. For example, the word deste can be tokenized as the single token |deste| if it occurs as a verb (Eng.: [you] gave) or as the two tokens |de+|este| if it occurs as a contraction (Eng.: of this). 3For more details, see (Branco and Silva, 2004). 4In these examples the | symbol will be used to mark token boundaries more clearly. 179 It is worth noting that this problem is not a minor issue, as these strings amount to 2% of the corpus that was used and any tokenization error will have a considerable negative influence on the subsequent steps of processing, such as POS tagging. To resolve the issue of ambiguous strings, a two-stage tokenization strategy is used, where the ambiguous strings are not immediately tokenized. Instead, the decision counts onthe contribution of the POS tagger: The tagger must first be trained on a version of the corpus where the ambiguous strings are not tokenized, and are tagged with a composite tag when occurring as a contraction (for example P+DEM for a contraction of a preposition and a demonstrative). The tagger then runs over the text and assigns a simple or a composite tag to the ambiguous strings. A second pass with the tokenizer then looks for occurrences of tokens with a composite tag and splits them: deste/V?|deste/V| deste/P+DEM?|de+/P|este/DEM| This approach allowed us to successfully resolve 99.4% of the ambiguous strings. This is a much better value than the baseline 78.20% obtained by always considering that the ambiguous strings are a contraction.5 4 POS tagger For the POS tagging task we used Brant?s TnT tagger (Brants, 2000), a very efficient statistical tagger based on Hidden Markov Models. For training, we used 90% of a 280, 000 token corpus, accurately hand-tagged with a tagset of ca. 60 tags, with inflectional feature values left aside. Evaluation showed an accuracy of 96.87% for this tool, obtained by averaging 10 test runs over different 10% contiguous portions of the corpus that were not used for training. The POS tagger we developed is currently the fastest tagger for the Portuguese language, and it is in line with state-of-the-art taggers for other languages, as discussed in (Branco and Silva, 2004). 5 Nominal featurizer This tool assigns feature value tags for inflection (Gender and Number) and degree (Diminutive, Superlative and Comparative) to words from nominal morphosyntactic categories. 5For further details see (Branco and Silva, 2003). Such tagging is typically done by a POS tagger, by using a tagset where the base POS tags have been extended with feature values. However, this increase in the number of tags leads to a lower tagging accuracy due to the data-sparseness problem. With our tool, we explored what could be gained by having a dedicated tool for the task of nominal featurization. We tried several approaches to nominal featurization. Here we report on the rule-based approach which is the one that better highlights the difficulties in this task. For this tool, we built on morphological regularities and used a set of rules that, depending on the word termination, assign default feature values to words. Naturally, these rules were supplemented by a list of exceptions, which was collected by using an machine readable dictionary (MRD) that allowed us to search words by termination. Nevertheless, this procedure is still not enough to assign a feature value to every token. The most direct reason is due to the so-called invariant words, which are lexically ambiguous with respect to feature values. For example, the Common Noun ermita (Eng.: hermit) can be masculine or feminine, depending on the occurrence. By simply using termination rules supplemented with exceptions, such words will always be tagged with underspecified feature values:6 ermita/?S To handle such cases the featurizer makes use of feature propagation. With this mechanism, words from closed classes, for which we know their feature values, propagate their values to thewords from open classes following them. These words, in turn, propagate those features to other words: o/MS ermita/MS humilde/MS Eng.: the-MS humble-MS hermit-MS but a/FS ermita/FS humilde/FS Eng.: the-FS humble-FS hermit-FS Special care must be taken to avoid that feature propagation reaches outside NP boundaries. For this purpose, some sequences of POS categories block feature propagation. In the example below, a PP inside an NP context, azul (an ?invariant? 6Values: M:masculine, F:feminine, S:singular, P:plural and ?:undefined. 180 adjective) might agree with faca or with the preceding word, ac?o. To prevent mistakes, propagation from ac?o to azul should be blocked. faca/FS de ac?o/MS azul/FS Eng.: blue (steel knife) or faca/FS de ac?o/MS azul/MS Eng.: (blue steel) knife For the sake of comparability with other possible similar tools, we evaluated the featurizer only over Adjectives and Common Nouns: It has 95.05% recall (leaving ca. 5% of the tokens with underspecified tags) and 99.05% precision.7 6 Nominal lemmatizer Nominal lemmatization consists in assigning to Adjectives and Common Nouns a normalized form, typically the masculine singular if available. Our approach uses a list of transformation rules that helps changing the termination of the words. For example, one states that any word ending in ta should have that ending transformed into to: gata ([female] cat) ? gato ([male] cat) There are, however, exceptions that must be accounted for. The word porta, for example, is a feminine common noun, and its lemma is porta: porta (door, feminine common noun) ? porta Relevant exceptions like the one above were collected by resorting to a MRD that allowed to search words on the basis of their termination. Being that dictionaries only list lemmas (and not inflected forms), it is possible to search for words with terminations matching the termination of inflected words (for example, words ending in ta). Any word found by the search can thus be considered as an exception. A major difficulty in this task lies in the listing of exceptions when non-inflectional affixes are taken into account. As an example, lets consider again the word porta. This word is an exception to the rule that transforms ta into to. As expected, this word can occur prefixed, as in superporta. Therefore, this derived word 7For a much more extensive analysis, including a comparison with other approaches, see (Branco and Silva, 2005a). should also appear in the list of exceptions to prevent it from being lemmatized into superporto by the rule. However, proceeding like this for every possible prefix leads to an explosion in the number of exceptions. To avoid this, a mechanism was used that progressively strips prefixes from words while checking the resulting word forms against the list of exceptions: supergata -----gata (apply rule) ? supergato but superporta -----porta (exception) ? superporta A similar problem arises when tackling words with suffixes. For instance, the suffix -zinho and its inflected forms (-zinha, -zinhos and -zinhas) are used as diminutives. These suffixes should be removed by the lemmatization process. However, there are exceptions, such as the word vizinho (Eng.: neighbor) which is not a diminutive. This word has tobe listed as an exception, together with its inflected forms (vizinha, vizinhos and vizinhas), which again leads to a great increase in the number of exceptions. To avoid this, only vizinho is explicitly listed as an exception and the inflected forms of the diminutive are progressively undone while looking for an exception: vizinhas (feminine plural) vizinha (feminine singular) vizinho (exception) ? vizinho To ensure that exceptions will not be overlooked, when both these mechanisms work in parallel one must follow all possible paths of affix removal. An heuristic chooses the lemma as being the result found in the least number of steps.8 To illustrate this, consider the word antena (Eng.: antenna). Figure 1 shows the paths followed by the lemmatization algorithm when it is faced with antenazinha (Eng.: [small] antenna). Both ante- and -zinha are possible affixes. In a first step, two search branches are opened, the first where ante- is removed and the second where -zinha is transformed into 8This can be seen as following a rationale similar to Karlsson?s (1990) local disambiguation procedure. 181 antenazinha nazinha antenazinho nazinho nazinho antena na no na no 0 1 2 3 4 Figure 1: Lemmatization of antenazinha -zinho. The search proceeds under each branch until no transformation is possible, or an exception has been found. The end result is the ?leaf node? with the shortest depth which, in this example, is antena (an exception). This branching might seem to lead to a great performance penalty, but only a few words have affixes, and most of them have only one, in which case there is no branching at all. This tool evaluates to an accuracy of 94.75%.9 7 Verbal featurizer and lemmatizer To each verbal token, this tool assigns the corresponding lemma and tag with feature values for Mood, Tense, Person and Number. The tool uses a list of rules that, depending on the termination of the word, assign all possible lemma-feature pairs. The word diria, for example, is assigned the following lemma-feature pairs: diria ? ?dizer,Cond-1ps? ? ?dizer,Cond-3ps? ? ?diriar,PresInd-3ps? ? ?diriar,ImpAfirm-2ps? Currently, this tool does not attempt to disambiguate among the proposed lemma-feature pairs. So, each verbal token will be tagged with all its possible lemma-feature pairs. The tool was evaluated over a list with ca. 800, 000 verbal forms. It achieves 100% precision, but at 50% recall, as half of those forms are ambiguous and receive more than one lemmafeature pair. 9For further details, see (Branco and Silva, 2005b). 8 Final Remarks So far, LX-Suite has mostly been used in-house for projects being developed by the NLX Group. It is being used in the GramaXing project, where a computational core grammar for deep linguistic processing of Portuguese is being developed under the Delphin initiative.10 In collaboration with CLUL,11 and under the TagShare project, LX-Suite is being used to help in the building of a corpus of 1 million accurately hand-tagged tokens, by providing an initial, highquality tagging which is then manually corrected. It is also used for the QueXting project, whose aim is to make available a question answering system onthe Portuguese Web. There is an on-line demo of LX-Suite located at http://lxsuite.di.fc.ul.pt. This on-line version of the suite is a partial demo, as it currently only includes the modules up to the POS tagger. By the end of the TagShare project (mid-2006), all the other modules described in this paper are planned to have been included. Additionally, the verbal featurizer and lemmatizer can be tested as a standalone tool at http://lxlemmatizer.di.fc.ul.pt. Future work will be focused on extending the suite with new tools, such as a named-entity recognizer and a phrase chunker.'	b"1 Introduction  Throughout the Automatic Content Extraction 1  (ACE) evaluations and the Message Understanding  Conferences2 (MUC), teams typically had a year or  more from release of the target to submitting system results. One exception was MUC-6 (Grishman  & Sundheim, 1996), in which scenario templates  for changing positions were extracted given only  one month. Our goal was to confine development  to a calendar week, in fact, <50 person hours. This                                                             1 http://www.nist.gov/speech/tests/ace/  2 http://www-nlpir.nist.gov/related_projects/muc/  is significant in two ways: the less effort it takes to  bring up a new domain, (1) the more broadly applicable the technology is and (2) the less effort  required to run a diagnostic research experiment.  Our second goal concerned minimizing training  data. Rather than approximately 250k words of  entity and relation annotation as in ACE, only ~20  example pairs per relation-type were provided as  training. Reducing the training requirements has  the same two desirable outcomes: demonstrating  that the technology can be broadly applicable and  reducing the overhead for running experiments.  The system achieved recall of 0.49 and precision  of 0.53 (for an F1 of 0.51) on a blind test set of 60  queries of the form Ri(arg1, arg2), where Ri is one  of the 5 new relations and exactly one of arg1 or  arg2 is a free variable for each query.  Key to this achievement was a hybrid of:    ? a variant of (Miller, et al, 2004) to learn two  new classes of entities via automatically induced  word classes and active learning (6 hours)  ? bootstrap relation learning (Freedman et al  2010) to learn 5 new relation classes (2.5 hours),   ? handwritten patterns over predicate-argument  structure (5 hours), and  ? coreference (20 hours)  Our bootstrap learner is initialized with relation  tuples (not annotated text) and uses LDC?s Gigaword and Wikipedia as a background corpus to  learn patterns for relation detection that are based  on normalized predicate argument structure as well  as surface strings.   These early empirical results suggest the following: (1) It is possible to specify a domain, adapt  our system, and complete manual scoring, includ1437 ing human performance, within a month. Experiments in machine reading (and in extraction) can  be performed much more quickly and cheaply than  ever before. (2) Through machine learning and  limited human pattern writing (6 hours), we  adapted a machine reading system within a week  (using less than 50 person hours), achieving question answering performance with an F1 of 0.5 and  withrecall 11% higher (relative) to a human reader. (3) Unfortunately, machine learning, though  achieving 80% precision,3 significantly lags behind  a gifted human pattern writer in recall. Thus, bootstrap learning with much higher recall at minimal  sacrifice in precision is highly desirable.  2 Related Work  This effort is evaluated extrinsically via formal  questions expressed as a binary relation with one  free variable. This contrasts with TREC Question  Answering, 4  where the questions are in natural  language, and not restricted to a single binary relation. Like the ?list? queries of TREC QA, the requirement is to find all answers, not just one.  Though question interpretation is not required in  our work, interpretation of the text corpus is.  The goal of rapid adaptation has been tested in  other contexts. In 2003, a series of experiments in  adapting to a new language in less than month  tested system performance on Cebuano and Hindi.  The primary goal was to adapt to a new language,  rather than a new domain. The extraction participants focused on named-entity recognition, not  relation extraction (May, et al 2003; Sekine &  Grishman, 2003; Li & McCallum, 2003; Maynard  et al 2003). The scenario templates of MUC-6  (Grishman & Sundheim, 1996) are more similar to  our relation extraction task, although the domain is  quite different. Our experiment allowed for 1 week  of development time, while MUC-6 allowed a  month. The core entities in the MUC-6 task  (people and organizations) had been worked on  previously. In contrast all of our relations included  at least one novel class. While MUC-6 systems  tended to use finite-state patterns, they did not incorporate bootstrapping or patterns based on the  output of a statistical parser.                                                               3 Handwritten patterns achieved 52% precision.  4 http://trec.nist.gov/data/qamain.html  For learning entity classes, we follow Miller, et  al., (2004), using word clustering and active learning to train a perceptron model, but unlike that  work we apply the technique not just to names but  also to descriptions. An alternative approach to  learning classes, applying structural patterns to  bootstrap description recognition without active  learning, is seen in Riloff (1996) and Kozareva et  al., (2008)    Much research (e.g. Ramshaw 2001) has focused on learning relation extractors using large  amounts of supervised training, as in ACE. The  obvious weakness of such approaches is the resulting reliance on manually annotated examples,  which are expensive and time-consuming to create.   Others have explored bootstrap relation learning from seed examples. Agichtein & Gravano (2000) and Ravichandran & Hovy (2002) reported  results for generating surface patterns for relation  identification; others have explored similar approaches (e.g. Pantel & Pennacchiotti, 2006). Mitchell et al (2009) showed that for macro-reading,  precision and recall can be improved by learning a  large set of interconnected relations and concepts  simultaneously. None use coreference to find training examples; all use surface (word) patterns.  Freedman et. al (2010) report improved performance from using predicate structure for bootstrappped relation learning.   Most approaches to automatic pattern generation have focused on precision, e.g., Ravichandran  and Hovy (2002) report results in TREC QA,  where extracting one instance of a relation can be  sufficient, rather than detecting all instances. Mitchell et al (2009), while demonstrating high precision, do not measure recall.  By contrast, our work emphasizes recall, not  just precision. Our question answering task asks  list-like questions that require multiple answers.   We also include the results of a secondary, extraction evaluation which requires that the system  identify every mention of the relations in a small  set of documents. This evaluation is loosely based  on the relation mention detection task in ACE.   3 Task Set-Up and Evaluation  Our effort was divided into four phases. During the  first phase, a third party produced an ontology and  the resources, which included: brief (~1 paragraph)  guidelines for each relation and class in the ontolo1438 gy; ~20 examples for each relation in the ontology;  2K documents that are rich in domain relations.  Table 1 lists the 5 new relations and number of examples provided for each. Arguments in italics  were known by the system prior to the evaluation.   Relation Ex.  possibleTreatment(Substance, Condition) 23  expectedDateOnMarket(Substance, Date) 11  responsibleForTreatment(Substance, Agent) 19  studiesDisease(Agent, Condition) 16  hasSideEffect(Substance, Condition) 27  Table 1: New Relations and Number of Examples  In phase two, we spent one week extending our  extraction system for the new ontology. During the  third phase, we ran our system over 10K documents to extract all instances of domain relations  from those documents. In the fourth phase, our  question answering system used the extracted information to answer queries.   4 Approach to Domain Specialization  Our approach to extracting domain relations integrated novel relation and class detectors into an  existing extraction system, designed primarily  around the ACE tasks. The existing system uses a  discriminatively trained classifier to detect the entity and value types of ACE. It also produces a syntactic parse for each sentence; normalizes these  parses to find logical predicate argument structure;  and detects and coreferences pronominal, nominal,  and name mentions for each of the 7 ACE entity  types (Person, Organization, Geopolitical Entity,  Location, Facility, Weapon, and Vehicle).5   The extraction system has three components that  allow for rapid adaptation to a new domain:   ? Class detectors trained using word classes derived fromunsupervised clustering and sentenceselected training data.  ? A bootstrap relation learner which given a few  seed examples learns patterns that indicate the  presence of relations.   ? An expressive pattern language which allows a  developer to express rules for relation extraction  in a simple, but fast manner.       Component Approach Effort  Class Recognizer Active Learning 6 hrs                                                             5 The extraction system detects relations and events in the  ACE ontology, but these were not used in the current work.     Class Recognizer Web-Mined List 1 hrs  Relation Recognizer  Semi-supervised  Bootstrapping  8.5 hrs  Relation Recognizer Manual Patterns 5 hrs  Coreference Heuristics 20 hrs  Table 2: Effort and Approach for New Domain  4.1 Class Extraction   Each of the relations in the new domain included at  least one argument that was new. While question  answering requires the system to identify the  classes only when they appear in a relation, knowledge of when a class is present provides important  information for relation extraction. For example in  our ontology, Y is a treatment for X only if Y is a  substance. Thus, ?Group counseling sessions are  effective treatments for depression? does not contain an instance of possibleTreatment(), while  ?SSRIs are effective treatments for depression?  does. The bootstrap learner allows constraints  based on argument type. To use this capability, we  trained the recognizer at the beginning of the week  of domain adaptation and used the predicted  classes during learning.   We annotated 1064 sentences (~31K words) using active learning combined with unsupervised  word clusters (Miller,et al, 2004) for the following  classes: Substance-Name, Substance-Description,  Condition-Name, and Condition-Description. Generic noun-phrases like new drugs, the illness, etc  were labeled as descriptors. Because of the time  frame, we did not develop extensive guidelines nor  measure inter-annotator agreement. Annotation  took 6 hours. We supplemented our annotation  with lists of substances and treatments from the  web, which took 1 hour.   4.2 Coreference  Providing a name reference is generally preferable  to a non-specific string (e.g. the drugs), but not  always feasible; for instance, reports of new research may appear without a name for the drug.  Our existing system?s coreference algorithms operate only on mentions of ACE entity types (persons, organizations, GPEs, other locations,  facilities, vehicles, and weapons). During the week  of domain adaption we developed new heuristics  for coreference over non-ACE types.  Most of our  heuristics are domain independent (e.g. linking the  parts of anappositive). Our decision to annotate  names and descriptions separately was driven par1439 tially by the need to select the best reference (i.e.  name) for co-referent clusters. Adding coreference  heuristics for the two new entity types was the single most time-consuming activity, taking 20 of the  total 43 hours.  4.3 Relation Extraction   For relation extraction, we used both pattern  learning and handwritten patterns. We initialized  our bootstrap relation learner with the example  instances provided with the domain ontology; Table 3 includes examples of the instances provided  to the system as training. Our bootstrap relation  learner finds instances of the relation argument  pairs in text and then proposes both predicateargument structure and word-based connections  between the arguments as possible new patterns for  the relation. The learner automatically prunes potential patterns using information about the number  of known-to-be true and novel instances matched  by a proposed pattern. By running the pattern extractor over a large corpus, the proposed patterns  generate new seeds which are in turn are used to  propose new patterns. For this experiment, we incorporated a small amount of supervision during  the bootstrapping process (roughly 1 hour total per  relation); we also performed ~30 minutes total in  pruning domain patterns at the end of learning.    Relation Arg-1 Arg-2  possTreatmnt AZT AIDS  studyDisease Dr Henri Joyeux cancer  studyDisease Samir Khleif cancer  Table 3: Sample Instances for Initializing Learner  We also used a small amount of human effort  creating rules for detecting the relations. The pattern writer was given the guidelines, the examples,  and a 2K document background corpus and spent 1  hour per relation writing rules.   The learned patterns use a subset of the full pattern language used by the pattern-writer. The language operates over surface-strings as well as  predicate-argument structure. Figure 1 illustrates  learned and handwritten patterns for the possibleTreatmentRelation(). The patterns in rectangles  match surface-string patterns; the tree-like patterns  match normalized predicate argument structure.   The ?WORD- token indicates a wild card of 1-3  words.  The blue rectangles at the root of the trees  in the handwritten patterns are sets of predicates  that can be matched by the pattern.  5 Evaluation   Our question answering evaluation was inspired  by the evaluation in DARPA?s machine reading  program, which requires systems to map the information in text into a formal ontology and answer questions based on that ontology. Unlike  ACE, this allows evaluators to measure performance without exhaustively annotating documents,  allows for balance between rare and common relations, and implicitly measures coreference without  requiring explicit annotation of answer keys for  coreference. However because the evaluation only  measures performance on the set of queries, many  relation instances will be unscored. Furthermore,  the system is not rewarded for finding the same  relation multiple times; finding 100 instances of  isPossibleTreatment(Penicillin,Strep Throat) is  the same as finding 1 (or 10) instances.     Figure 1: Sample Patterns for possibleTreatment()    The evaluation included only queries of the type  Find all instances for which the relation P(X, Z) is  true where one of X or Z is constant. For example,  Find possible treatments for diabetes; or What is  expected date to market for Abilify? There were 60  queries in the evaluation set to be answered from a  10K document corpus. To produce a preliminary  answer key, annotators were given the queries and  corpus indexed by Google Desktop. Annotators  were given 1 hour to find potential answers to each  query. If no answers were found after 1 hour, the  annotators were given a second hour to look for  answers. For two queries, both of the form Find  treatments with an expected date to market of MMYYYY, even after two hours of searching the annotators were unable to find any answers.6   Annotator answers served as the initial goldstandard. Given this initial answer key, annotators  reviewed system answers and aligned them with  gold-standard answers. System output not aligned  with the initial gold standard was assessed as correct or incorrect. We assume that the final goldstandard constitutes a complete answer key, and                                                             6 Evaluators wanted some queries with no answers.  1440 are thus able to calculate recall for our system and  for humans7. Because we had only one annotator  for each query and because we assumed that any  answer found by an annotator was correct, we  could not estimate human precision on this task.   Answers can be specific named concepts (e.g.  Penicillin) or generic descriptions (e.g. drug, illness). Given the sentence, ACME produces a wide  range of drugs including treatments for malaria  and athletes foot,? our reading system would extract the relations responsibleForTreatment(drugs,  ACME), possibleTreatment(drugs, malaria), possibleTreatment(drugs, athletes foot). When a name  was available in the document, annotators marked  the answer as correct, but underspecified. We calculated precision and recall treating underspecified  answers as incorrect and separately calculated precision and recall counting underspecified answers  as correct. When treated as correct, there was less  than a 0.05 absolute increase in both precision and  recall. Unless otherwise specified, all scores reported here use the stricter condition which treats  underspecified answers as incorrect.   We also evaluated extracting all information in a  small document collection (here human search of  the 10k documents does not play a role in finding  answers). Individuals were asked to annotateevery  instance of the 5 relations in a set of 102 documents. Recall, Precision, and F were calculated by  aligning system responses to the answer key. System answers that aligned are correct; those that did  not are incorrect; and answers in the key that were  not found by the system are misses. Unlike the  question answering evaluation, this evaluation  measures the ability to find every instance of a  fact. If the gold standard includes 100 instances of  isPossibleTreatment(Penicillin, Strep Throat), recall will decrease for each instance missed. The  ?extraction? evaluation does not penalize systems  for missing coreference.   6 Results  6.1 Class Detection                                                             7 The answer key may contain some answers that were found  neither by the annotator nor by the systems described here,  since the answer key includes answers pooled from other systems not reported in this paper. The system reported here was  the highest performing of all those participating in the experiment. Furthermore, if a system answer is marked as correct,  but underspecified, the specific  answer is put in the key.  The recall, precision, and F1 for class detection  using 10-fold cross validation of the ~1K annotated sentences appear in the 3-5th columns of Table  4. Given the amount of training, our results are  lower than in Miller et al(2004) (an F1 of 90 with  less than 25K words of training). Several factors  could explain this: Finding boundaries and types  for descriptions is more complex than for names in  English. 8  Our classes, pharmaceutical substances  and physiological conditions, may have been more  difficult to learn. Our classes are less common in  news reporting; as such, both word-class clusters  and active learning may have been less effective.  Finally, our evaluation was done on a 10-fold split  of the active-learning selected data; bias in selecting the data could explain at least a part of our  lower performance.   Type  # in  GS  Without Lists With Lists  R P F R P F  Subst-D 789 77 85 80.8 78 85 81.3  Subst-N 410 70 82 75.5 77 81 78.9  Cond-D 427 72 78 74.9 72 77 74.4  Cond-N 963 80 87 83.4 84 83 83.5  Table 4: Cross Validation:  Condition & Substance  We noticed that the system frequently reported  country names to be substance-names. Surprisingly, we found that our well-trained name finder  made the opposite mistake, occasionally reporting  drugs as geo-political entities.   We incorporated lists of known substances and  conditions to improverecall. Performance on the  same cross-validation split is shown in the final  three columns of Table 4. Incorporating the lists led  to recall gains for both substance-name and condition-name. Because a false-alarm in class recognition only leads to an incorrect relation extraction if  it appears in a context indicating a domain relation,  false alarms of classes may be less important in the  question answering and extraction evaluations.    6.2 Question Answering and Extraction  Figure 2 and Table 6 show system performance  using only handwritten rules (HW), only learned  patterns (L), and combining both (C). Figure 2   includes scores calculated with all of the systems?  answers (in the dotted boxes), and with just those  answers that were deemed useful (discussed be                                                           8 English names are capitalized; person names have a typical  form and are frequently signaled by titles; organization names  frequently have clear signal words, such as Corp.  1441 low). We include annotator recall. Handwritten  patterns outperform learned patterns consistently  with much higher recall. Encouragingly, however,  1. The combined system?s recall and F-Score  are noticeably higher for 3 of the relations.   2. The learned patterns generate answers not  found by handwritten patterns.   3. The learned patterns have high precision.9  There is variation across the different relations.  The two best performing relations possibleTreatment() and studiesDisease() have F1 more than  twice as high as the two worst performing relations, expectedDateToMarket() and hasSideEffect(). This is primarily due to differences in recall.     Figure 2: Overall Q/A Performance: All answers in   dotted boxes; 'Useful Answers' unboxed  The combined system?s recall (0.49), while low,  is higher than that of the annotators (0.44). While  hardly surprising that a machine can process information much more quickly than a person, it is  encouraging that higher recall is achieved even  with only one week?s effort. In the context of our  pooled answer-key, the relatively low recall of  both the system and the annotator suggests that  there was little overlap between the answers found  by the annotator and those found by the system.   As already described, the system answers can  include both specific references (e.g. Prozac) and  more generic references (the drug). When a more  specific answer is present in the document, generic  references have been treated as incorrect. However, sometimes there is not a more specific reference; for example an article written before a drug  has been released may never name the drug. Scores  reported thus far treat such answers as correct. These answers would be useful when answering  more complex queries. For example, given the sen                                                           9 The learned patterns' high precision is to be expected for two  reasons. First, a few bad patterns were manually removed for  each relation. More importantly, the learning algorithm strongly favors high precision patterns because it needs to maintain a  seed set with low noise in order to learn effectively.   tence ?ACME spent 5 years developing a pill to  treat the flu which it will release next week,? extracting relations involving ?the pill?  would allow  a system to answer questions that use multiple relations in the ontology to for example ask about  organizations developing treatments for the flu, or  the expected date of release for ACME?s drugs.  However, in our simple question answering  framework such generic answers never convey  novel information and thus were probably ignored  by human annotators.    To measure the impact of treating these generic  references as correct,10 we did additional annotation on the correct answers, marking answers as  ?useful? (specific) and ?not-useful? (generic). The  unboxed bars in Figure 2 show performance when  ?not-useful? answers are removed from the answerkey and the responses. For the four relations where  there was a change Table 5 provides the relative  change performance when only ?useful? answers  are considered. The annotator?s recall increases  noticeably while the combined system?s drops.  This results in the overall recall of annotators surpassing that of the combined system.    Relation  Recall Precision  A C H L C H L  possTreat 12 10 10 14 -10 -11 -3  respTreat 9 0 -5 8 -4 -4 -1  studyDis 12 -6 -9 13 -11 -13 0  hasSidEff 3 4 4 4 0 0 0  Total 11 -2 -4 6 -9 -10 -2  Table 5: Relative Change in Recall and Precision When  Non-Useful Answers are Removed  Table 7 shows the total number of answers produced by annotators and by each system, as well as  the percentage of queries with at least one correct  answer for each system. For one relation expectedDateOnMarket(), the learned system did not  find any answers. This relation had far fewer answers found by annotators and occurred far more  rarely in the fully annotated extraction set (see Table 8). Anecdotally, extracting this relation frequently required co-referencing ?it? (e.g. ?It will be  released in March 2011?). Our heuristics for coreference of the new classes did not account for pronouns. Learning from such examples would  require coreference during bootstrapping. Most  likely, thelearned system was unable to generate  enough novel instances to continue bootstrapping                                                             10 Generic answers were treated as correct only if a more specific reference was not available in the document.   1442 and was thus unable to learn the relation.   Relation Type  (# Queries; # Correct Ans.)  Recall Precision F  A C HW L C HW L C HW L  possTreatment (10;247) 0.27 0.63 0.50 0.34 0.51 0.47 0.83 0.56 0.48 0.48  respForTreat (15;134) 0.73 0.33 0.24 0.22 0.66 0.78 0.73 0.44 0.37 0.33  expectDateMarkt (11;60) 0.90 0.17 0.17 0.00 0.77 0.83 0.00 0.27 0.28 0  studiesDisease (13;292) 0.23 0.67 0.59 0.09 0.51 0.50 0.79 0.58 0.54 0.16  hasSideEffect (11;104) 0.80 0.10 0.13 0.02 0.83 0.70 1.00 0.17 0.23 0.04  Total (60;837) 0.44 0.49 0.42 0.17 0.53 0.52 0.80 0.51 0.46 0.28  Table 6: Question Answering Results by Relation Type  Relation Type    Total Number of Answers % Queries with At Least 1 Corr. Ans  A C HW L A C HW L  possTreatment  66 303 261 100 100.0% 90.0% 90.0% 90.0%  respForTreat  98 67 41 40 100.0% 66.7% 60.0% 60.0%  expectDateMarkt  54 13 12 0 72.7% 45.5% 45.5% 0.0%  studiesDisease  68 379 347 33 100.0% 61.5% 46.2% 46.2%  hasSideEffect  83 12 20 2 72.7% 36.4% 45.5% 18.2%  Total  369 774 681 175 90.0% 60.0% 56.7% 43.3%  Table 7: Number of Answers and Number of Queries Answered  Overall, the system did better on relations having more correct answers. Bootstrap learning has  an easier time discovering new instances and new  patterns when there are more examples to work  with. Even a human pattern writer will have more  examples to generalize from for common relations.   While possibleTreatment() and hasSideEffect()  have similar F-scores, their performance is very  different at the query level. The system was able to  find at least one correct answer to every possibleTreatment() query; however only 72.7% of the studiesDisease() queries were answered.   Table 8 presents results from the extraction  evaluation where a set of ~100 documents were  annotated for all mentions of the 5 relations. Because every mention in the document set must be  found, the system cannot rely on finding the easiest  answers for common relations. The results in Table  8 are significantly lower than for the question answering tasks; yet some of the same trends are  present. Handwritten rules outperform learned patterns. For at least some relations, the combination  of the two improves performance. The three relations forwhich the learned system has the lowest  performance on the question-answering task have  the fewest instances annotated in the document set.  Fewer instance in the large corpus make bootstrapping more difficult?the learner is less able to generate novel instances to expand its pattern set.   7 Discussion  7.1 Sources of Error  The most common source of error is pattern coverage. In the following figure, the system identified  responsibleForTreatment(Janssen Pharmaceutical,  Sporanox), but missed the corresponding relation  between Novartis and Lamisil.             Relation Type # Relations Found Recall Precision F  GS C HW L C HW L C HW L C HW L  possibleTreatment 518 225 187 68 0.15 0.10 0.09 0.34 0.28 0.66 0.21 0.15 0.15  respForTreatment 387 101 77 36 0.10 0.08 0.05 0.41 0.40 0.50 0.17 0.13 0.08  expDateOnMarket 66 13 13 0 0.06 0.06 0.00 0.31 0.31 0.00 0.10 0.10 0.00  studiesDisease 136 95 91 4 0.08 0.09 0.00 0.12 0.13 0.00 0.10 0.11 0.00  hasSideEffect 256 26 25 2 0.04 0.04 0.00 0.39 0.40 0.50 0.07 0.07 0.01  Table 8: Extraction Results on the 102 Document Test Set Annotated for All Instances of the Relations  Sporanox is made by Janssen Pharmaceutica Inc.,  of Titusville, N.J. Lamisil is a product of Novartis  Pharmaceuticals of East Hanover, N.J.      1443 Missed class instances contribute to errors, sometimes originating in errors in tokenization (e.g. not  removing the ?_? in each drug name in a bulleted  list of the form ?_Trovan, an antibiotic...; etc.)  However, many drug-names are simply missed:    The system correctly identifies Rebif and Aricept  as drugs, but misses Pregabalin and Serono. In  both misses, the immediately preceding and following words provide little evidence that the word  refers to a drug rather than some other product.  Substance detection might be better served with a  web-scale, list-learning approach like the doubly  anchored patterns described in (Kozareva et al,  2008). Alternatively, our approach may need to be  extended to include a larger context window.  7.2 Learned Patterns   One of the ways in which learned patterns supplement handwritten ones is learning highly specific  surface-string patterns that are insensitive to errors  in parsing. Figure 3 illustrates two examples of  what appear to be easy cases of possibleTreatment(). Because the handwritten patterns are not  exhaustive and make extensive use of syntactic  structure, parse errors prevented the system based  on handwritten rules from firing. Learned surfacestring patterns were able to find these relations.   Even when the syntactic structure is correct,  learned patterns capture expressions not common  enough to have been noticed by the rule writer. For  example, while the handwritten patterns included  ?withdrew? as a predicate indicating a company  was responsible for a drug, they did not include  ?pulled.? By including ?pulled?,learned patterns  extracted responsibleForTreatment() from ?American Home Products pulled Duract, a painkiller.?  Similarly, the learned patterns include an explicit  pattern ?CONDITION drug called SUBSTANCE?,  and thus extracted a possibleTreatment() relation  from ?newly approved narcolepsy drug called  modafinil? without relying on the coreference  component to link drug to modafinil.   Handwritten Patterns  Despite the examples above of successfully learned  patterns, handwritten patterns perform significantly  better. In the active-learning context used for these  experiments, the handwritten rules also required  less manual effort. This comparison is not entirely  fair-- while learned patterns required more hours,  supervising the bootstrapping algorithm requires  no training. The handwritten patterns, in contrast,  require a trained expert.     Figure 3: Extractions Missed by Handwritten Rules &  the Erroneous Parses that Hid them  While handwritten rules and learned patterns use  the same language, they make use of it differently.  The handwritten patterns group similar concepts  together. A human pattern writer adds relevant  synonyms, as well as words that are not synonymous but in the pattern context can be used interchangeably. In Figure 4, the handwritten patterns  include three word-sets: (patient*, people, participant*); (given, taken, took, using); and (report*,  experience*, develop*, suffer*). The ?*? serves as a  wild-card to further generalize a pattern. The wordsets in Figure 4 illustrate challenges for a learned  system: the words are not synonyms, but rather are  words that can be used to imply the relation.   A human pattern writer frequently generates  new classes not in the domain ontology. In Figure  4, the circled patterns form a class of ?people taking a substance.? The handwritten patterns for studiesDisease() include classes targeting scientists  and researchers. These classes are not necessarily  triggered by nouns. Such classes allow the pattern  writer to include complex patterns as in Figure 4  and to write relatively precise, but open-ended patterns such as: if there is a single named-drug and a  named, non-side-effect disease in the same sentence, the drug is a treatment for the disease.   Pfizer also hopes to introduce Pregabalin next  year for treatment of neuropathic pain, epilepsy  and anxiety?Other deals include co-promoting  Rebif for multiple sclerosis with its discoverer,  Serono, and marketing Aricept for Alzheimer's  disease with its developer, Eisai Co.  1444   Figure 4: Learned and Handwritten Patterns for   hasSideEffect()  A final difference between handwritten and  learned patterns is the level of predicate-argument  complexity used. In general, handwritten patterns  account for larger spans of predicate argument  structure while learned patterns tend to limit themselves to the connections between the arguments of  the relation with minor extensions.   8 Conclusions and Lessons Learned  First, it is encouraging that the synthesis of learning algorithms and handwritten algorithms can  achieve an F1 of 0.51 in a new domain in a week  (<50 hoursof effort). Second, it is exciting that so  little training data is required: ~20 relation pairs  out of context (~2.5 hours of effort) and ~6 hours  of active learning for the new classes.   Third, the effectiveness of learning algorithms is  still not competitive with handwritten patterns  based on predicate-argument structure (~5 hours of  effort on top of active learning for entities).  Though the learned patterns have high precision  (0.80 on average), recall is low (0.17) and varied  greatly across the relations. Though the dominant  factor in missing relations is pattern coverage,  missing instances of classes contributed to low recall. Comparing learned patterns to manually written patterns, (1) synonyms or other lexical  alternatives that a human pattern writer would include, (2) the creation of subclasses for argument  types, and (3) the scope of patterns11 are each major sources of the disparity in coverage. Research  on learning approaches to raise recall without significant sacrifice in precision seems essential.   Fourth, despite the disparity in performance of  learned versus manual patterns, and despite the low                                                             11 Learned patterns tend to focus on the structure that appears  between the two arguments, rather than structure surrounding  the left and right arguments.  recall of learned patterns, the combined system?s  recall and F-Score are higher for three of the relations because the learned patterns generated answers not found by handwritten patterns. We found  examples where highly specific, learned, surfacelevel patterns (lexical patterns) occasionally found  information missed by handwritten patterns due to  parsing errors or general low coverage.  Fifth, the effort for coreference was the most  time-consuming, given that every new relation  contained at least one of the new argument types.  While we included this in our estimate of domain  adaptation, the infrastructure we built is domain  generic. Improving generic coreference will reduce  domain specific effort in future.   Perhaps most significant of all, running a complete experiment from definition of the domain  through creation of training data and measurement  of end-to-end performance of the system can be  completed in a month. The ability to rapidly,  cheaply, and empirically measure the impact of  extraction research could prove a significant spur  to research across the board.  These experiments suggest three possible directions for improving the ability to quickly develop  information extraction technology for a new set of  relations: (1) reducing the amount of supervision  provided to the bootstrap-learner; (2) improving  the bootstrapping approach to reach the level of  recall achieved by the human pattern writer eliminatingthe need for a trained expert during domain  adaptation; and (3) focusing improvements to the  bootstrapping approach on techniques that allow it  to find more of the instances missed by the pattern  writer, thus improving the accuracy of the hybrid  system.    Acknowledgments  This work was supported, in part, by DARPA under AFRL Contract FA8750-09-C-179. Distribution Statement ?A? (Approved for Public Release,  Distribution Unlimited) Thank you to the reviewers for your insightful comments and to Michelle  Franchini for coordinating the assessment effort. "
1.0	I05-2038	A00-1031	b'1 Introduction  Research and development for information extraction from biomedical literature (biotextmining) has been rapidly advancing due to  demands caused by information overload in the  genome-related field. Natural language processing (NLP) techniques have been regarded as  useful for this purpose. Now that focus of information extraction is shifting from extraction  of ?nominal? information such as named entity  to ?verbal? information such as relations of entities including events and functions, syntactic  analysis is an important issue of NLP application in biomedical domain. In extraction of relation, the roles of entities participating in the  relation must be identified along with the verb  that represents the relation itself. In text analysis,  this corresponds to identifying the subjects, objects, and other arguments of the verb.  Though rule-based relation information extraction systems using surface pattern matching  and/or shallow parsing can achieve highprecision (e.g. Koike et al, 2004) in a particular  target domain, they tend to suffer from low recall due to the wide variation of the surface expression that describe a relation between a verb  and its arguments. In addition,  the portability of  such systems is low because the system has to  be re-equipped with different set of rules when  different kind of relation is to be extracted. One  solution to this problem is using deep parsers  which can abstract the syntactic variation of a  relation between a verb and its arguments represented in the text, and constructing extraction  rule on the abstract predicate-argument structure.  To do so, wide-coverage and high-precision  parsers are required.  While basic NLP techniques are relatively  general and portable from domain to domain,  customization and tuning are inevitable, especially in order to apply the techniques effectively to highly specialized literatures such as  research papers and abstracts. As recent advances in NLP technology depend on machinelearning techniques, annotated corpora from  which system can acquire rules (including  grammar rules, lexicon, etc.) are indispensable  220 resources for customizing general-purpose NLP  tools. In bio-textmining, for example, training  on part-of-speech (POS)-annotated GENIA corpus was reported to improve the accuracy of  JunK tagger (English POS tagger) (Kazama et  al., 2001) from  83.5% to 98.1% on MEDLINE  abstracts (Tateisi and Tsujii, 2004), and the  FraMed corpus (Wermter and Hahn, 2004) was  used to train TnT tagger on German (Brants,  2000) to improve its accuracy from 95.7% to  98% on clinical reports and other biomedical  texts. Corpus annotated for syntactic structures  is expected to play a similar role in tuning parsers to biomedical domain, i.e., similar improvement on the performance of parsers is expected  by using domain-specific treebank as a resource  for learning. For this purpose, we construct  GENA Treebank (GTB), a treebank on research  abstracts in biomedical domain.  2 Outline of the Corpus  The base text of GTB is that of the GENIA corpus constructed at University of Tokyo (Kim et al., 2003), which is a collection of research abstracts selected from the search results of  MEDLINE database with keywords (MeSH  terms) human, blood cells and transcription factors. In the GENIA corpus, the abstracts are encoded in an XML scheme where each abstract is  numbered with MEDLINE UID and contains  title and abstract. The text  of title and abstract is  segmented into sentences in which biological  terms are annotated with their semantic classes.  The GENIA corpus is also annotated for part-ofspeech (POS) (Tateisi and Tsujii, 2004), and  coreference is also annotated in a part of the  GENIA corpus by MedCo project at Institute for  Infocomm Research, Singapore (Yang et al  2004).   GTB is the addition of syntactic information  to the GENIA corpus. By annotating various  linguistic information on a same set of text, the  GENIA corpus will be a resource not only for  individual purpose such as named entity extraction or training parsers but also for integrated  systems such as information extraction using  deep linguistic analysis. Similar attempt of constructing integrated corpora is being done in  University of Pennsylvania, where a corpus of  MEDLINE abstracts in CYP450 and oncology  domains where annotated for named entities,  POS, and tree structure of sentences (Kulick et  al, 2004).   2.1 Annotation Scheme  The annotation scheme basically follows the  Penn Treebank II (PTB) scheme (Beis et al  1995), encoded in XML. A non-null constituent  is marked as an element, with its syntactic category (which may be combined with its function  tags indicating grammatical roles such as  -SBJ,  -PRD, and -ADV) used as tags. A null constituent is marked as a childless element whose tag  corresponds to its categories. Other function tags  are encoded as attributes. Figure 1 shows an example of annotated sentence in XML, and the  corresponding PTB notation. The label ?S?  means ?sentence?, ?NP? noun phrase, ?PP?  prepositional phrase, and ?VP? verb phrase.   The label ?NP-SBJ? means that the element is  an NP that serves as the subject of the sentence.  A null element, the trace of the object of ?studied? moved by passivization, is denoted by  ? <NP NULL="NONE" ref="i55"/>? in XML  and ?*-55? in PTB notation. The number ?55?  which refers to the identifier of the moved element, is denoted by ?id? and ?ref? attributes in  XML, and is denoted as a part of a label in PTB.  In addition to changing the encoding, we  made some modifications to the scheme. First,  analysis within the noun phrase is simplified.  Second, semantic division of adverbial phrases  such as ??TMP? (time) and ??MNR? (manner)  are not used: adverbial constituents other than  ?ADVP? (adverbial phrases) or ?PP? used adverbially are marked with ?ADV tags but not  with semantic tags. Third, a coordination structure is explicitly marked with the attribute  SYN=?COOD?whereas in the original PTB  scheme it is not marked as such.     In our GTB scheme, ?NX? (head of a complex noun phrase) and ?NAC? (a certain kind of  nominal modifier within a noun phrase) of the  PTB scheme are not used. A noun phrase is generally left unstructured. This is mainly in order  to simplify the process of annotation. In case of  biomedical abstracts, long noun phrases often  involve multi-word technical terms whose syntactic structure is difficult to determine without  deep domain knowledge. However, the structure  of noun phrases are usually independent of the  structure outside the phrase, so that it would be  221 easier to analyze the phrases involving such  terms independently (e.g. by biologists) and  later merge the two analysis together. Thus we  have decided that we leave noun phrases unstructured in GTB annotation unless their analysis is necessary for determining the structure  outside the phrase. One of the exception is the  cases that involves coordination where it is necessary to explicitly mark up the coordinated  constituents.  In addition, we have added special attributes  ?TXTERR?, ?UNSURE?,  and ?COMMENT?  for later inspection. The ?TXTERR? is used  when the annotator suspects that there is a  grammatical error in the original text; the  ?UNSURE? attribute is used when the annotator  is not confident; and the ?COMMENT? is used  for free comments (e.g. reason of using  ?UNSURE?) by the annotator.   2.2   Annotation Process  The sentences in the titles and abstracts of the  base text of GENIA corpus are annotated manually using an XML editor used for the Global  Document Annotation project (Hasida 2000).  Although the sentence boundaries were adopted  from the corpus, the tree structure annotation  was done independently of POS- and term- annotation already done on the GENIA corpus.  The annotator was a Japanese non-biologist who  has previously involved in the POS annotation  of the GENIA corpus and accustomed to the  style of research abstracts in English. Manually  annotated abstracts are automatically converted  to the PTB format, merged with the POS annotation of the GENIA corpus (version 3.02).  3 Annotation Results  So far, 500 abstracts are annotated and converted to the merged PTB format. In the merging process, we found several annotation errors.  The 500 abstracts with correction of these errors  are made publicly available as ?The GENIA  Treebank Beta Version? (GTB-beta).    For further clean-up, we also tried to parse  the corpus by the Enju parser (Miyao and Tsujii  2004), and identify the error of the corpus by  investigating into the parse errors. Enju is an  HPSG parser that can be trained with PTB-type  corpora which is reported to have 87% accuracy  on Wall Street Journal portion of Penn Treebank  corpus. Currently the accuracy of the parser  drops down to82% on GTB-beta, and although  proper quantitative analysis is yet to be done, it  was found that the mismatches between labels of  the treebank and the GENIA POS corpus (e.g.  an ?ing form labeled as noun in the POS corpus  and as the head of a verb phrase in the tree corpus) are a major source of parse error. The correction is complicated because several errors in  the GENIA POS corpus were found in this  cleaning-up process. When the cleaning-up  process is done, we will make the corpus publicly available as the proper release.  <S><PP>In <NP>the present paper </NP></PP>,  <NP-SBJ id="i55"><NP>the binding  </NP><PP>of <NP>a [125I]-labeled aldosterone  derivative </NP></PP><PP>to <NP><NP>plasma  membrane rich fractions </NP><PP>of HML  </PP></NP></PP></NP-SBJ><VP>was  <VP>studied <NP NULL="NONE"  ref="i55"/></VP>  </VP>.</S>    4 Inter-Annotator Agreement  We have also checked inter-annotator agreement.  Although the PTB scheme is popular among  natural language processing society, applicability of the scheme to highly specialized text such  as research abstract is yet to be discussed. Especially, when the annotation is done by linguists,  lack of domain knowledge might decrease the  stability and accuracy of annotation.  A small part of the base text set (10 abstracts) was annotated by another annotator. The  10 abstracts were chosen randomly, had 6 to 17  sentences per abstract (total 108 sentences). The  new annotator had a similar background as the  first annotator that she is a Japanese nonbiologist who has experiences in translation of  (S (PP In (NP the present paper)), (NP-SBJ-55 (NP  the binding) (PP of (NP a [125I]-labeled aldosterone  derivative)) (PP to (NP (NP plasma membrane rich  fractions) (PP of HML)))) (VP was (VP studied *55)).)  Figure 1. The sentence ?In the present paper, the binding of  a [125I]-labeled aldosterone derivative to plasma membrane rich fractions of HML was studied? annotated in  XML and PTB formats.   222 technical documents in English and in corpus  annotation of  English texts.  The two results were examined manually,  and there were 131 disagreements. Almost every  sentence had at least one disagreement. We have  made the ?gold standard? from the two sets of  abstracts by resolving the disagreements, and the  accuracy of the annotators against this gold  standard were 96.7% for the first annotator and  97.4% for the second annotator.   Of the disagreement, the most prominent  were the cases involving coordination, especially the ones with ellipsis. For example, one  annotator annotated the phrase ?IL-1- and IL-18mediated function? as in Figure 2a, the other  annotated as Figure 2b.    Such problem is addressed in the PTB  guideline and both formats are allowed as alternatives. As coordination with ellipsis occurs  rather frequently in research abstracts, this kind  of phenomena has higher effect on decrease of  the agreement rate than in Penn Treebank. Of  the 131 disagreements, 25 wereon this type of  coordination.  Another source of disagreement is the attachment of modifiers such as prepositional  phrases and pronominal adjectives. However,  most are ?benign ambiguity? where the difference of the structure does not affect on interpretation, such as ?high expression of STAT in  monocytes? where the prepositional phrase ?in  monocytes? can attach to ?expression? or  ?STAT? without much difference in meaning,  and ?is augmented when the sensitizing tumor is  a genetically modified variant? where the whclause can attach to ?is augmented? or ?augmented? without changing the meaning. The  PTB guideline states that the modifier should be  attached at the higher level in the former case  and at the lower case in the latter. In the annotation results, one annotator consistently attached  the modifiers in both cases at the higher level,  and the other consistently at the lower level, indicating that the problem is in understanding the  scheme rather than understanding the sentence.  Only 15 cases were true ambiguities that needed  knowledge of biology to solve, in which 5 involved coordination (e.g., the scope of ?various?  in ?various T cell lines and peripheral blood  cells?) .    Although the number was small, there were  disagreements on how to annotate a mathematical formula such as ?n=2? embedded in the sentence, since mathematical formulae were outside  the scope of the original PTB scheme. One annotator annotated this kind of phrase consistently as a phrase with ?=? as an adjective, the  other annotated as phrase with ?=? as a verb.  There were 6 such cases. Another disagreement  particular to abstracts is a treatment of labeled  sentences. There were 8 sentences in two abstracts where there is a label like ?Background:?.   One annotator included the colon (?:?) in the label, while the other did not. Yet another is that  one regarded the phrase ?Author et al as coordination, and the other regarded ?et al as a  modifier.    <NP SYN="COOD">  <NP><ADJP>IL-1- <ADJP NULL="QSTN"/></ADJP>           <NP NULL="RNR" ref="i20"/></NP>  and   <NP>IL-18-mediated <NP NULL="RNR" ref="i20"/></NP>  <NP id="i20">function </NP>   Other disagreements are more general type  such as regarding ?-ed? form of a verb as an adjective or a participle, miscellaneous errors such  as omission of a subtype of label (such as ?PRD? or ?-SBJ) or the position of <PRN> tags  <NP>  <ADJP SYN="COOD">    <ADJP>IL-1- <ADJP NULL="QSTN"/></ADJP>    and     <ADJP>IL-18-mediated </ADJP></ADJP>  function   </NP>      NP               ADJP   Function           ADJP     and ADJP               IL-1  *   IL-18 mediated   Figure 2a. Annotation of a coordinated phrase by the first  annotator. A* denotes a null constituent.  </NP>          NP               NP And NP                  ADJP  *20 IL-18 meidiated NP              IL-1 *      function20 Figure 2b. Annotation of the same phrase as in Figure 2a  by the second annotator.  A * denotes a null constituent  and ?20? denotes coindexing.  223 with regards to ?,? for the inserted phrase, or the  errors which look like just ?careless?. Such disagreements and mistakes are at least partially  eliminated when reliable taggers and parsers are  available for preprocessing  5 Discussion  The result of the inter-annotator agreement  test indicates that the writing style rather than  the contents of the research abstracts is the  source of the difficulty in tree annotation. Contrary to the expectation that the lack of domain  knowledge causes a problem in annotation on  attachments of modifiers, the number of cases  where annotation of modifier attachment needs  domain knowledge is small. This indicates that  linguists can annotate most of syntactic structure  without an expert level of domain knowledge.   A major source of difficulty is coordination,  especially the ones involving ellipsis. Coordination is reported to be difficult phenomena in annotation of different levels in the GENIA corpus  (Tateisi and Tsujii, 2004), (Kim et al, 2003). In  addition to the fact that this is the major source  of inter-annotator agreement, the annotator often  commented the coordinated structure as ?unsure?.  The problem of coordination can be divided into  two with different nature: one is that the annotation policy is still not well-established for the  coordination involving ellipsis, and the other is  an ambiguity when the coordinated phrase has  modifiers.   Syntax annotation of coordination with ellipsis is difficult in general but the more so in annotation of abstracts than in the case of general  texts, because in abstracts authors tend to pack  information in limited number of words. The  PTB guideline dedicates a long section for this  phenomena and allows alternatives in annotation,  but there are still cases which are not wellcovered by the scheme. For example, in addition  to the disagreement, the phrase illustrated in  Figure 2a and Figure 2b shows another problem  of the annotation scheme. Both annotators fail to  indicate that it is ?mediated? that was to be after  ?IL-1? because there is no mechanism of  coindexing a null element with a part of a token.   This problem of ellipsis can frequently occur  in research abstracts, and it can be argued that the tokenization criteria must be changed for  texts in biomedical domain (Yamamoto and Satou, 2004) so that such fragment as ?IL-18? and  ?mediated? in ?IL-18-ediated? should be regarede  as separate tokens. The Pennsylvania biology  corpus (Kulick et al, 2004) partially solves this  problem by separating a token where two or  more subtokens are connected with hyphens, but  in the cases where a shared part of the word is  not separated by a hyphen (e.g. ?metric? of ?stereo- and isometric alleles?) the word including  the part is left uncut. The current GTB follows  the GENIA corpus that it retains the tokenization criteria of the original Penn Treebank, but  this must be reconsidered in future.   For analysis of coordination with ellipsis, if  the information on full forms is available, one  strategy would be to leave the inside structure of  coordination unannotated in the treebank corpus  (and in the phase of text analysis the structure is  not established in the phase of parsing but with a  different mechanism) and later merge it with the  coordination structure annotation. The GENIA  term corpus annotates the full form of a technical term whose part is omitted in the surface as  an attribute of the ?<cons>? element indicating a  technical term (Kim et al, 2003). In the abovementioned Pennsylvania corpus, a similar  mechanism (?chaining?) is used for recovering  the full form of named entities. However, in  both corpora, no such information is available  outside the terms/entities.   The cases where scope of modification in  coordinated phrases is problematic are few but  they are more difficult in abstracts than in general texts because the resolution of ambiguity  needs domain knowledge. If term/entity annotation is already done, that information can help  resolve this type of ambiguity, but again the  problem is that outside the terms/entities such  information is not available. It would be practical to have the structure flat but specially  marked when the tree annotators are unsure and  have a domain expert resolve the ambiguity, as  the sentences that needs such intervention seems  few. Some cases of ambiguity in modifier attachment (which do not involve coordination)  can be solved with similar process.  We believe that other type of disagreements  can be solved with supplementing criteria for  linguistic phenomena not well-covered by the  scheme, and annotator training. Automatic preprocessing by POS taggers and parsers can also  help increase the consistent annotation.  2246 Conclusion  A subset of the GENIA corpus is annotated  for syntactic (tree) structure. Inter-annotator  agreement test indicated that the annotation can  be done stably by linguists without much  knowledge in biology, provided that proper  guideline is established for linguistic phenomena  particular to scientific research abstracts. We  have made the 500-abstract corpus in both XML  and PTB formats and made it publicly available  as ?the GENIA Treebank beta version? (GTBbeta). We are in further cleaning up process of  the 500-abstract set, and at the same time, initial  annotation of the remaining abstracts is being  done, so that the full GENIA set of 2000 abstracts will be annotated with tree structure.   For parsers to be useful for information extraction, they have to establish a map between  syntactic structure and more semantic predicateargument structure, and between the linguistic  predicate-argument structures to the factual relation to be extracted. Annotation of various information on a same set of text can help  establish these maps. For the factual relations,  we are annotating relations between proteins and  genes in cooperation with a group of biologists.  For predicate-argument annotation, we are investigating the use of the parse results of the  Enju parser.  Acknowledgments  The authors are grateful to annotators and colleagues that helped the construction of the corpus. This work is partially supported by Grantin-Aid for Scientific Research on Priority Area  C ?Genome Information Science? from the Ministry of Education, Culture, Sports, Science and  Technology of Japan. '	b'1 In t roduct ion   A large number of current language processing sys-  tems use a part-of-speech tagger for pre-processing.  The tagger assigns a (unique or ambiguous) part-of-  speech tag to each token in the input and passes its  output o the next processing level, usually a parser.  Furthermore, there is a large interest in part-of-  speech tagging for corpus annotation projects, who  create valuable linguistic resources by a combination  of automatic processing and human correction.  For both applications, a tagger with the highest  possible accuracy is required. The debate about  which paradigm solves the part-of-speech tagging  problem best is not finished. Recent comparisons  of approaches that can be trained on corpora (van  Halteren et al, 1998; Volk and Schneider, 1998) have  shown that in most cases statistical aproaches (Cut-  ting et al, 1992; Schmid, 1995; Ratnaparkhi, 1996)  yield better results than finite-state, rule-based, or  memory-based taggers (Brill, 1993; Daelemans etal.,  1996). They are only surpassed by combinations of different systems, forming a "voting tagger".  Among the statistical approaches, the Maximum  Entropy framework has a very strong position. Nev-  ertheless, a recent independent comparison of 7 tag-  gets (Zavrel and Daelemans, 1999) has shown that  another approach even works better: Markov mod-  els combined with a good smoothing technique and  with handling of unknown words. This tagger, TnT,  not only yielded the highest accuracy, it also was the  fastest both in training and tagging.  The tagger comparison was organized as a "black-  box test": set the same task to every tagger and  compare the outcomes. This paper describes the  models and techniques used by TnT together with  the implementation.  The reader will be surprised how simple the under-  lying model is. The result of the tagger comparison  seems to support he maxime "the simplest is the  best". However, in this paper we clarify a number  of details that are omitted in major previous pub-  lications concerning tagging with Markov models.  As two examples, (Rabiner, 1989) and (Charniak  et al, 1993) give good overviews of the techniques  and equations used for Markov models and part-of-  speech tagging, but they are not very explicit in the  details that are needed for their application. We ar-  gue that it is not only the choice of the general model  that determines the result of the tagger but also the  various "small" decisions on alternatives.  The aim of this paper is to give a detailed ac-  count of the techniques used in TnT. Additionally,  we present results of the tagger on the NEGRA cor-  pus (Brants et al, 1999) and the Penn Treebank  (Marcus et al, 1993). The Penn Treebank results  reported here for the Markov model approach are  at least equivalent to those reported for the Maxi-  mum Entropyapproach in (Ratnaparkhi, 1996). For  a comparison to other taggers, the reader is referred  to (Zavrel and Daelemans, 1999).  2 Arch i tec ture   2.1 The Under ly ing Model   TnT uses second order Markov models for part-of-  speech tagging. The states of the model represent  tags, outputs represent the words. Transition prob-  abilities depend on the states, thus pairs of tags.  Output probabilities only depend on the most re-  cent category. To be explicit, we calculate  argmax P(tilti-1, ti-2)P(wilti P(tr+l ItT)  ti...iT  (i)  for a given sequence of words w I . . .  W T of length T.  tl... tT are elements of the tagset, the additional  224 tags t - l ,  to, and tT+l are beginning-of-sequence  and end-of-sequence markers. Using these additional  tags, even if they stem from rudimentary process-  ing of punctuation marks, slightly improves tagging  results. This is different from formulas presented  in other publications, which just stop with a "loose  end" at the last word. If sentence boundaries are  not marked in the input, TnT adds these tags if it  encounters one of \\[.!?;\\] as a token.  Transition and output probabilities are estimated  from a tagged corpus. As a first step, we use the  maximum likelihood probabilities /5 which are de-  rived from the relative frequencies:  Unigrams: /5(t3) = f(t3) (2)  N  f(t2, t3) (3)  Bigrams: P(t31t~)= f(t2)  f(ta, t2, t3)  Trigrams: /5(t3ltx,t~) - f ( t l , t2)  (4)  Lexical: /5(w3 It3) - /(w3, t3) (5)  f(t3)  for all tl, t2, t3 in the tagset and w3 in the lexi-  con. N is the total number of tokens in the training  corpus. We define a maximum likelihood probabil-  ity to be zero if the corresponding nominators and  denominators are zero. As a second step, contex-  tual frequencies are smoothed and lexical frequences  are completed by handling words that are not in the  lexicon (see below).  2.2 Smooth ing   Trigram probabilities generated from a corpus usu-  ally cannot directly be used because of the sparse-  data problem. This means that there are not enough  instances for each trigram to reliably estimate the  probability. Furthermore, setting a probability to  zero because the corresponding trigram never oc-  cured in the corpus has an undesired effect. It causes  the probability of a complete sequence to be set to  zero if its use is necessary for a new text sequence,  thus makes it impossible to rank different sequences  containing a zero probability.  The smoothing paradigm that delivers the best  results in TnT is linear interpolation of unigrams,  bigrams, and trigrams. Therefore, we estimate a  trigram probability as follows:  P(t3ltl ,  t2) = AlP(t3) + Ag_/5(t31t2) + A3/5(t3\\[t1, t2)  (6) /5 are maximum likelihood estimates of the proba-  bilities, and A1 + A2 + A3 = 1, so P again represent  probability distributions.  We use the context-independent variant of linear  interpolation, i.e., the values of the As do not depend  on the particular trigram. Contrary to intuition,  this yields better esults than the context-dependent  variant. Due to sparse-data problems, one cannot es-  timate a different set of As for each trigram. There-  fore, it is common practice to group trigrams by fre-  quency and estimate tied sets of As. However, we  are not aware of any publication that has investi-  gated frequency groupings for linear interpolation i part-of-speech tagging. All groupings that we have  tested yielded at most equivalent results to context-  independent linear interpolation. Some groupings  even yielded worse results. The tested groupings  included a) one set of As for each frequency value  and b) two classes (low and high frequency) on the  two ends of the scale, as well as several groupings  in between and several settings for partitioning the  classes.  The values of Ax, A2, and A3 are estimated by  deleted interpolation. This technique successively  removes each trigram from the training corpus and  estimates best values for the As from all other n-  grams in the corpus. Given the frequency counts  for uni-, bi-, and trigrams, the weights can be very  efficiently determined with a processing time linear  in the number of different rigrams. The algorithm  is given in figure 1. Note that subtracting 1 means  taking unseen data into account. Without this sub-  traction the model would overfit the training data  and would generally ield worse results.  2.3 Handling of Unknown Words  Currently, the method of handling unknown words  that seems to work best for inflected languages is  a suffix analysis as proposed in (Samuelsson, 1993).  Tag probabilities are set according to the word\'s end-  ing. The suffix is a strong predictor for word classes,  e.g., words in the Wall Street Journal part of the  Penn Treebank ending in able are adjectives (JJ) in  98% of the cases (e.g. fashionable, variable), the rest  of 2% are nouns (e.g. cable, variable).  The probability distribution for a particular suf-  fix is generated from all words in the training set  that share the same suffix of some predefined max-  imum length. The term suffix as used here means  "final sequence of characters of a word" which is not  necessarily a linguistically meaningful suffix.  Probabilities are smoothed by successive abstrac-  tion. This calculates the probability of a tag t  given the last m letters li of an n letter word:  P(t l ln - r ,+l , . . . ln) .  The sequence of increasingly  more general contexts omits more and more char-  acters ofthe suffix, such that P(t l ln_m+2,. . . , ln) ,   P(t \\ [ ln-m+3,. . . , l~) ,  . . . ,  P(t) are used for smooth-  ing. The recursiou formula is  P(t l l ,_ i+l,  . . . ln)  = P( t l ln - i+ l , . . .  In) + O iP ( t l l~- , . . . ,  ln) (7)  1 +0~  225 set  )%1 ---- )%2 = )%3 = 0  fo reach  t r ig ram tl,t2,t3 with  f(t l ,t2,t3) > 0  depending on the maximum of the fo l low ing  three  va lues :   f(tl ,t2,ts)-- 1 case f (h,t2)- I  " increment )%3 by f ( t l , t2 , t3)   f(t2,t3)-I case f(t2)- I  " increment )%2 by f ( t l , t2 , t s )   f ( t3 ) - - i   case N-1  " increment )%1 by f(t l ,t2,t3)  end  end  normalize )%1, )%2, )%3  Figure 1: Algorithm for calculting the weights for context-independent li ear interpolation )%1, )%2, )%3 when  the n-gram frequencies are known. N is the size of the corpus? If the denominator in one of the expressions  is 0, we define the result of that expression to be 0.  for i = m. . .  0, using the maximum likelihood esti-  mates/5  from frequencies in the lexicon, weights Oi  and the initialization  P(t) =/5(t) .  (8)  The maximum likelihood estimate for a suffix of  length i is derived from corpus frequencies by  P(t i /~- i+l ,  .. l~) = f (t ,  1~-~+1,... l~) ? (9 )   For the Markov model, we need the inverse condi-  tional probabilities P ( / , - i+ l , . . .  lnlt) which are ob-  tained by Bayesian inversion.  A theoretical motivated argumentation uses the  standard eviation of the maximum likelihood prob-  abilities for the weights 0i (Samuelsson, 1993)?  This leaves room for interpretation.  1) One has to identify a good value for m, the  longest suffix used? The approach taken for TnT is  the following: m depends on the word in question.  We use the longest suffix that we can find in the  training set (i.e., for which the frequency is greater  than or equal to 1), but at most 10 characters. This  is an empirically determined choice?  2) We use a context-independent approach for 0i,  as we did for the contextual weights )%i. It turned  out to be a good choice to set al 0i to the standard  deviation of the unconditioned maximum likelihood  probabilities of the tags in the training corpus, i.e.,  weset  1  $  Oi = ~--~(/5(tj) - 15)2 (10)  s 1 j= l   for all i = 0 . . .  m - 1, using a tagset of s tags and  the average  $  /5 = 1 ~/5(t j )  (11)  8  j----I  This usually yields values in the range 0.03 .. .  0.10.  3) We use different estimates for uppercase and  lowercase words, i.e., we maintain two different suffix  tries depending on the capitalization of the word.  This information improves the tagging results?  4) Another freedom concerns the choice of the  words in the lexicon that should be used for suf-  fix handling. Should we use all words, or are some  of them better suited than others? Accepting that  unknown words are most probably infrequent, one  can argue that using suffixes of infrequent words in  the lexicon is a better approximation for unknown  words than using suffixes of frequent words. There-  fore, we restrict the procedure of suffix handling to  words with a frequency smaller than or equal to some  threshold value. Empirically, 10 turned out to be a  good choice for this threshold.  2.4 Cap i ta l i za t ion   Additional information that turned out to be use-  ful for the disambiguation process for several cor-  pora and tagsets is capitalization information. Tags  are usually not informative about capitalization, but  probability distributions of tags around capitalized  words are different from those not capitalized. The  effect is larger for English, which only capitalizes  proper names, and smaller for German, which capi-  talizes all nouns.  We use flags ci that are true if wi is a capitalized  word and false otherwise? These flags are added to  the contextual probability distributions. Instead of  P(tsItl,t2) (12)  we use  P(t3, c3\\[tl , cl, t2, c2) (13)  and equations (3) to (5) are updated accordingly.  This is equivalent o doubling the size of the tagset  and using different ags depending on capitalization.  226 2.5 Beam Search  The processing time of the Viterbi algorithm (Ra-  biner, 1989) can be reduced by introducing a beam  search. Each state that receives a 5 value smaller  than the largest 5 divided by some threshold value  is excluded from further processing. While the  Viterbi algorithm is guaranteed to find the sequence  of states with the highest probability, this is no  longer true when beam search is added. Neverthe-  less, for practical purposes and the right choice of  0, there is virtually no difference between the algo-  rithm with and without a beam. Empirically, a value  of 0 = 1000 turned out to approximately double the  speed of the tagger without affecting the accuracy.  The tagger currently tags between 30~000 and  60,000 tokensper second (including file I/O) on a  Pentium 500 running Linux. The speed mainly de-  pends on the percentage of unknown words and on  the average ambiguity rate.  3 Eva luat ion   We evaluate the tagger\'s performance under several  aspects. First of all, we determine the tagging ac-  curacy averaged over ten iterations. The overall ac-  curacy, as well as separate accuracies for known and  unknown words are measured.  Second, learning curves are presented, that indi-  cate the performance when using training corpora of  different sizes, starting with as few as 1,000 tokens  and ranging to the size of the entire corpus (minus  the test set).  An important characteristic of statistical taggers  is that they not only assign tags to words but also  probabilities in order to rank different assignments.  We distinguish reliable from unreliable assignments  by the quotient of the best and second best assign-  ments 1. All assignments for which this quotient is  larger than some threshold are regarded as reliable,  the others as unreliable. As we will see below, accu-  racies for reliable assignments are much higher.  The tests are performed on partitions of the cor-  pora that use 90% as training set and 10% as test  set, so that the test data is guaranteed to be unseen  during training. Each result is obtained by repeat-  ing the experiment 10 times with different partitions  and averaging the single outcomes.  In all experiments, contiguous test sets are used.  The alternative is a round-robin procedure that puts  every 10th sentence into the test set. We argue that  contiguous test sets yield more realistic results be-  cause completely unseen articles are tagged. Using  the round-robin procedure, parts of an article are al-  ready seen, which significantly reduces the percent-  age of unknown words. Therefore, we expect even  1 By definition, this quotient is co if there is only one pos-  sible tag for a given word.  higher results when testing on every 10th sentence  instead of a contiguous et of 10%.  In the following, accuracy denotes the number of  correctly assigned tags divided by the number of to-  kens in the corpus processed. The tagger is allowed  to assign exactly one tag to each token.  We distinguish the overall accuracy, taking into  account all tokens in the test corpus, and separate  accuracies for known and unknown tokens. The lat-  ter are interesting, since usually unknown tokens are  much more difficult to process than known tokens,  for which a list of valid tags can be found in the  lexicon.  3.1 Tagging the  NEGRA corpus   The German NEGRA corpus consists of 20,000 sen-  tences (355,000 tokens) of newspaper texts (Frank-  furter Rundschau) that are annotated with parts-of-  speech and predicate-argument structures (Skut et al., 1997). It was developed at the Saarland Univer-  sity in Saarbrficken 2. Part of it was tagged at the  IMS Stuttgart. This evaluation only uses the part-  of-speech annotation and ignores structural annota-  tions.  Tagging accuracies for the NEGRA corpus are  shown in table 2.  Figure 3 shows the learning curve of the tagger,  i.e., the accuracy depending on the amount of train-  ing data. Training length is the nmnber of tokens  used for training. Each training length was tested  ten times, training and test sets were randomly cho-  sen and disjoint, results were averaged. The training  length is given on a logarithmic scale.  It is remarkable that tagging accuracy for known  words is very high even for very small training cot-  pora. This means that we have a good chance of  getting the right tag if a word is seen at least once  during training. Average percentages of unknown  tokens are shown in the bottom line of each diagram.  We exploit the fact that the tagger not only de-  termines tags, but also assigns probabilities. If there  is an alternative that has a probability "close to"  that of the best assignment, his alternative can be  viewed as almost equally well suited. The notion of  "close to" is expressed by the distance of probabil-  ities, and this in turn is expressed by the quotient  of probabilities. So, the distance of the probabili-  ties of a best tag tbest and an alternative tag tart  is expressed by P(tbest)/p(tau), which is some value  greater or equal to 1 since the best tag assignment  has the highest probability.  Figure 4 shows the accuracy when separating as-  signments with quotients larger and smaller than  the threshold (hence reliable and unreliable assign-  ments). As expected, we find that accuracies for  2For availability, please check  h~tp ://w~. col i. uni-sb, de/s fb378/negra-corpus  777 227 Table 2: Part-of-speech tagging accuracy for the NEGRA corpus, averaged over 10 test runs, training and  test set are disjoint. The table shows the percentage of unknown tokens, separate accuracies and standard  deviations for known and unknown tokens, as well as the overall accuracy.  percentage  unknowns  NEGRA corpus 11.9%  known  ace .   97.7% 0.23  unknown  acc. (x  89.0% 0.72  overall  aCE. o"  96.7% 0.29  NEGRA Corpus :  POS Learn ing  Curve   100  9O  80  70 /S   6O  50 , i  1 2 5  50.8 46.4 41.4  I i i I I I i  10 20 50 100 200 320 500  36.0 30.7 23.0 18.3 14.3 11.9 11).3  Overall  min =78.1%  max=96.7%  e Known  rain =95.7%  max=97.7%  - - -a- - -  Unknown  rain =61.2% max=89.0%  I  1000 x 1000 Training Length  St  avg. percentage unknown  Figure 3: Learning curve for tagging the NEGRA corpus. The training sets of variable sizes as well as test  sets of 30,000 tokens were randomly chosen. Training and test sets were disjoint, the procedure was repeated  10 times and results were averaged. Percentages of unknowns for 500k and 1000k training are determined  from an untagged extension.  NEGRA Corpus: Accuracy of  re l iab le  assignments  100  99  98  97  A  f " / :. Reliable rain =96.7% max=99.4%  96 I i i i i i i i i i i  2 5 10 20 50 100 500 2000 10000 threshold 0  100 97.9 95.1 92.7 90.3 86.8 84.1 81.0 76.1 71.9 68.3 64.1 62.0 % cases reliable  - 53.5 62.9 69.6 74.5 79.8 82.7 85.2 88.0 89.6 90.8 91.8 92.2 acc. of complement  Figure 4: Tagging accuracy for the NEGRA corpus when separating reliable and unreliable assignments. The  curve shows accuracies for reliable assignments. The numbers at the bottom line indicate the percentage of  reliable assignments and the accuracy of the complement set (i.e., unreliable assignments).  228 Table 5: Part-of-speech tagging accuracy for the Penn Treebank. The table shows the percentage of unknown  tokens, separate accuracies and standard deviations for known and unknown tokens, as well as the overall  accuracy.  I percentage known  unknowns acc. a  Penn Treebank 2.9% 97.0% 0.15  unknown  aCC. O"  85.5% 0.69  overall  aCE. O"  96.7% 0.15  100  9O  80  70 <  60  Penn Treebank: POS Learn ing  Curve   /  50 I I ~ i I I I i I  1 2 5 10 20 50 100 200 500  50.3 42.8 33.4 26.8 20.2 13.2 9.8 7.0 4.4  Overall  rain =78.6%  max=96.7%  Known  rain =95.2%  max=97.0%  Unknown  min =62.2%  max=85.5%  I  1000 ? 1000 Training Length  2.9 avg. percentage unknown  Figure 6: Learning curve for tagging the Penn Treebank. The training sets of variable sizes as well as test sets  of 100,000 tokens were randomly chosen. Training and test sets were disjoint, the procedure was repeated  10 times and results were averaged.  Penn Treebank: Accuracy of reliable assignments  100  99  98  97  Overall  rain =96.6%  max=99.4%  96 i i i i t i i i i i i i  2 5 10 20 50 100 500 2000 10000 threshold 0  100 97.7 94.6 92.2 89.8 86.3 83.5 80.4 76.6 73.8 71.0 67.2 64.5 % cases reliable  - 53.5 62.8 68.9 73.9 79.3 82.6 85.2 87.5 88.8 89.8 91.0 91.6 acc. of complement  Figure 7: Tagging accuracy for the Penn Treebank when separating reliable and unreliable assignments. The  curve shows accuraciesfor reliable assignments. The numbers at the bottom line indicate the percentage of  reliable assignments and the accuracy of the complement set.  229  reliable assignments are much higher than for unre-  liable assignments. This distinction is, e.g., useful  for annotation projects during the cleaning process,  or during pre-processing, sothe tagger can emit mul-  tiple tags if the best tag is classified as unreliable.  3.2 Tagging the Penn Treebank   We use the Wall Street Journal as contained in the  Penn Treebank for our experiments. The annotation  consists of four parts: 1) a context-free structure  augmented with traces to mark movement and dis-  continuous constituents, 2) phrasal categories that  are annotated as node labels, 3) a small set of gram-  matical functions that are annotated as extensions to  the node labels, and 4) part-of-speech tags (Marcus  et al, 1993). This evaluation only uses the part-of-  speech annotation.  The Wall Street Journal part of the Penn Tree-  bank consists of approx. 50,000 sentences (1.2 mil-  lion tokens).  Tagging accuracies for the Penn Treebank are  shown in table 5. Figure 6 shows the learning curve  of the tagger, i.e., the accuracy depending on the  amount of training data. Training length is the num-  ber of tokens used for training. Each training length  was tested ten times. Training and test sets were  disjoint, results are averaged. The training length is  given on a logarithmic scale. As for the NEGRA cor-  pus, tagging accuracy is very high for known tokens  even with small amounts of training data.  We exploit the fact that the tagger not only de-  termines tags, but also assigns probabilities. Figure  7 shows the accuracy when separating assignments  with quotients larger and smaller than the threshold  (hence reliable and unreliable assignments). Again,  we find that accuracies for reliable assignments are  much higher than for unreliable assignments.  3.3 Summary of Part-of-Speech Tagging  Results  Average part-of-speech tagging accuracy is between  96% and 97%, depending on language and tagset,  which is at least on a par with state-of-the-art re-  sults found in the literature, possibly better. For  the Penn Treebank, (Ratnaparkhi, 1996) reports an  accuracy of 96.6% using the Maximum Entropy ap-  proach, our much simpler and therefore faster HMM  approach delivers 96.7%. This comparison eeds to  be re-examined, since we use a ten-fold crossvalida-  tion and averaging of results while Ratnaparkhi only  makes one test run.  The accuracy for known tokens is significantly  higher than for unknown tokens. For the German  newspaper data, results are 8.7% better when the  word was seen before and therefore is in the lexicon,  than when it was not seen before (97.7% vs. 89.0%).  Accuracy for known tokens is high even with very  small amounts of training data. As few as 1000 to- kens are sufficient o achieve 95%-96% accuracy for  them. It is important for the tagger to have seen a  word at least once during training.  Stochastic taggers assign probabilities to tags. We  exploit the probabilities to determine reliability of  assignments. For a subset hat is determined uring  processing by the tagger we achieve accuracy rates  of over 99%. The accuracy of the complement set is  much lower. This information can, e.g., be exploited  in an annotation project to give an additional treat-  ment to the unreliable assignments, or to pass se-  lected ambiguities to a subsequent processing step. 4 Conc lus ion   We have shown that a tagger based on Markov mod-  els yields state-of-the-art results, despite contrary  claims found in the literature. For example, the  Markov model tagger used in the comparison of (van  Halteren et al, 1998) yielded worse results than all  other taggers. In our opinion, a reason for the wrong  claim is that the basic algorithms leave several deci-  sions to the implementor. The rather large amount  of freedom was not handled in detail in previous pub-  lications: handling of start- and end-of-sequence, the  exact smoothing technique, how to determine the  weights for context probabilities, details on handling  unknown words, and how to determine the weights  for unknown words. Note that the decisions we made  yield good results for both the German and the En-  glish Corpus. They do so for several other corpora  as well. The architecture remains applicable to a  large variety of languages.  According to current tagger comparisons (van  Halteren et al, 1998; Zavrel and Daelemans, 1999),  and according to a comparsion of the results pre-  sented here with those in (Ratnaparkhi, 1996), the  Maximum Entropy framework seems to be the only  other approach yielding comparable results to the  one presented here. It is a very interesting future  research topic to determine the advantages of either  of these approaches, to find the reason for their high  accuracies, and to find a good combination of both.  TnT is freely available to universities and re-  lated organizations for research purposes (see  http ://www. coli. uni-sb, de/-thorsten/tnt). Acknowledgements   Many thanks go to Hans Uszkoreit for his sup-  port during the development of TnT. Most of the  work on TnT was carried out while the author  received a grant of the Deutsche Forschungsge-  meinschaft in the Graduiertenkolleg Kognitionswis-  senschaft Saarbriicken. Large annotated corpora re  the pre-requisite for developing and testing part-of-  speech taggers, and they enable the generation of  high-quality language models. Therefore, I would  230  like to thank all the people who took the effort  to annotate the Penn Treebank, the Susanne Cor-  pus, the Stuttgarter Referenzkorpus, the NEGRA  Corpus, the Verbmobil Corpora, and several others.  And, last but not least, I would like to thank the  users of TnT who provided me with bug reports and  valuable suggestions for improvements. '
0.0	I05-2038	C02-2011	b'1 Introduction  Research and development for information extraction from biomedical literature (biotextmining) has been rapidly advancing due to  demands caused by information overload in the  genome-related field. Natural language processing (NLP) techniques have been regarded as  useful for this purpose. Now that focus of information extraction is shifting from extraction  of ?nominal? information such as named entity  to ?verbal? information such as relations of entities including events and functions, syntactic  analysis is an important issue of NLP application in biomedical domain. In extraction of relation, the roles of entities participating in the  relation must be identified along with the verb  that represents the relation itself. In text analysis,  this corresponds to identifying the subjects, objects, and other arguments of the verb.  Though rule-based relation information extraction systems using surface pattern matching  and/or shallow parsing can achieve highprecision (e.g. Koike et al, 2004) in a particular  target domain, they tend to suffer from low recall due to the wide variation of the surface expression that describe a relation between a verb  and its arguments. In addition,  the portability of  such systems is low because the system has to  be re-equipped with different set of rules when  different kind of relation is to be extracted. One  solution to this problem is using deep parsers  which can abstract the syntactic variation of a  relation between a verb and its arguments represented in the text, and constructing extraction  rule on the abstract predicate-argument structure.  To do so, wide-coverage and high-precision  parsers are required.  While basic NLP techniques are relatively  general and portable from domain to domain,  customization and tuning are inevitable, especially in order to apply the techniques effectively to highly specialized literatures such as  research papers and abstracts. As recent advances in NLP technology depend on machinelearning techniques, annotated corpora from  which system can acquire rules (including  grammar rules, lexicon, etc.) are indispensable  220 resources for customizing general-purpose NLP  tools. In bio-textmining, for example, training  on part-of-speech (POS)-annotated GENIA corpus was reported to improve the accuracy of  JunK tagger (English POS tagger) (Kazama et  al., 2001) from  83.5% to 98.1% on MEDLINE  abstracts (Tateisi and Tsujii, 2004), and the  FraMed corpus (Wermter and Hahn, 2004) was  used to train TnT tagger on German (Brants,  2000) to improve its accuracy from 95.7% to  98% on clinical reports and other biomedical  texts. Corpus annotated for syntactic structures  is expected to play a similar role in tuning parsers to biomedical domain, i.e., similar improvement on the performance of parsers is expected  by using domain-specific treebank as a resource  for learning. For this purpose, we construct  GENA Treebank (GTB), a treebank on research  abstracts in biomedical domain.  2 Outline of the Corpus  The base text of GTB is that of the GENIA corpus constructed at University of Tokyo (Kim et al., 2003), which is a collection of research abstracts selected from the search results of  MEDLINE database with keywords (MeSH  terms) human, blood cells and transcription factors. In the GENIA corpus, the abstracts are encoded in an XML scheme where each abstract is  numbered with MEDLINE UID and contains  title and abstract. The text  of title and abstract is  segmented into sentences in which biological  terms are annotated with their semantic classes.  The GENIA corpus is also annotated for part-ofspeech (POS) (Tateisi and Tsujii, 2004), and  coreference is also annotated in a part of the  GENIA corpus by MedCo project at Institute for  Infocomm Research, Singapore (Yang et al  2004).   GTB is the addition of syntactic information  to the GENIA corpus. By annotating various  linguistic information on a same set of text, the  GENIA corpus will be a resource not only for  individual purpose such as named entity extraction or training parsers but also for integrated  systems such as information extraction using  deep linguistic analysis. Similar attempt of constructing integrated corpora is being done in  University of Pennsylvania, where a corpus of  MEDLINE abstracts in CYP450 and oncology  domains where annotated for named entities,  POS, and tree structure of sentences (Kulick et  al, 2004).   2.1 Annotation Scheme  The annotation scheme basically follows the  Penn Treebank II (PTB) scheme (Beis et al  1995), encoded in XML. A non-null constituent  is marked as an element, with its syntactic category (which may be combined with its function  tags indicating grammatical roles such as  -SBJ,  -PRD, and -ADV) used as tags. A null constituent is marked as a childless element whose tag  corresponds to its categories. Other function tags  are encoded as attributes. Figure 1 shows an example of annotated sentence in XML, and the  corresponding PTB notation. The label ?S?  means ?sentence?, ?NP? noun phrase, ?PP?  prepositional phrase, and ?VP? verb phrase.   The label ?NP-SBJ? means that the element is  an NP that serves as the subject of the sentence.  A null element, the trace of the object of ?studied? moved by passivization, is denoted by  ? <NP NULL="NONE" ref="i55"/>? in XML  and ?*-55? in PTB notation. The number ?55?  which refers to the identifier of the moved element, is denoted by ?id? and ?ref? attributes in  XML, and is denoted as a part of a label in PTB.  In addition to changing the encoding, we  made some modifications to the scheme. First,  analysis within the noun phrase is simplified.  Second, semantic division of adverbial phrases  such as ??TMP? (time) and ??MNR? (manner)  are not used: adverbial constituents other than  ?ADVP? (adverbial phrases) or ?PP? used adverbially are marked with ?ADV tags but not  with semantic tags. Third, a coordination structure is explicitly marked with the attribute  SYN=?COOD?whereas in the original PTB  scheme it is not marked as such.     In our GTB scheme, ?NX? (head of a complex noun phrase) and ?NAC? (a certain kind of  nominal modifier within a noun phrase) of the  PTB scheme are not used. A noun phrase is generally left unstructured. This is mainly in order  to simplify the process of annotation. In case of  biomedical abstracts, long noun phrases often  involve multi-word technical terms whose syntactic structure is difficult to determine without  deep domain knowledge. However, the structure  of noun phrases are usually independent of the  structure outside the phrase, so that it would be  221 easier to analyze the phrases involving such  terms independently (e.g. by biologists) and  later merge the two analysis together. Thus we  have decided that we leave noun phrases unstructured in GTB annotation unless their analysis is necessary for determining the structure  outside the phrase. One of the exception is the  cases that involves coordination where it is necessary to explicitly mark up the coordinated  constituents.  In addition, we have added special attributes  ?TXTERR?, ?UNSURE?,  and ?COMMENT?  for later inspection. The ?TXTERR? is used  when the annotator suspects that there is a  grammatical error in the original text; the  ?UNSURE? attribute is used when the annotator  is not confident; and the ?COMMENT? is used  for free comments (e.g. reason of using  ?UNSURE?) by the annotator.   2.2   Annotation Process  The sentences in the titles and abstracts of the  base text of GENIA corpus are annotated manually using an XML editor used for the Global  Document Annotation project (Hasida 2000).  Although the sentence boundaries were adopted  from the corpus, the tree structure annotation  was done independently of POS- and term- annotation already done on the GENIA corpus.  The annotator was a Japanese non-biologist who  has previously involved in the POS annotation  of the GENIA corpus and accustomed to the  style of research abstracts in English. Manually  annotated abstracts are automatically converted  to the PTB format, merged with the POS annotation of the GENIA corpus (version 3.02).  3 Annotation Results  So far, 500 abstracts are annotated and converted to the merged PTB format. In the merging process, we found several annotation errors.  The 500 abstracts with correction of these errors  are made publicly available as ?The GENIA  Treebank Beta Version? (GTB-beta).    For further clean-up, we also tried to parse  the corpus by the Enju parser (Miyao and Tsujii  2004), and identify the error of the corpus by  investigating into the parse errors. Enju is an  HPSG parser that can be trained with PTB-type  corpora which is reported to have 87% accuracy  on Wall Street Journal portion of Penn Treebank  corpus. Currently the accuracy of the parser  drops down to82% on GTB-beta, and although  proper quantitative analysis is yet to be done, it  was found that the mismatches between labels of  the treebank and the GENIA POS corpus (e.g.  an ?ing form labeled as noun in the POS corpus  and as the head of a verb phrase in the tree corpus) are a major source of parse error. The correction is complicated because several errors in  the GENIA POS corpus were found in this  cleaning-up process. When the cleaning-up  process is done, we will make the corpus publicly available as the proper release.  <S><PP>In <NP>the present paper </NP></PP>,  <NP-SBJ id="i55"><NP>the binding  </NP><PP>of <NP>a [125I]-labeled aldosterone  derivative </NP></PP><PP>to <NP><NP>plasma  membrane rich fractions </NP><PP>of HML  </PP></NP></PP></NP-SBJ><VP>was  <VP>studied <NP NULL="NONE"  ref="i55"/></VP>  </VP>.</S>    4 Inter-Annotator Agreement  We have also checked inter-annotator agreement.  Although the PTB scheme is popular among  natural language processing society, applicability of the scheme to highly specialized text such  as research abstract is yet to be discussed. Especially, when the annotation is done by linguists,  lack of domain knowledge might decrease the  stability and accuracy of annotation.  A small part of the base text set (10 abstracts) was annotated by another annotator. The  10 abstracts were chosen randomly, had 6 to 17  sentences per abstract (total 108 sentences). The  new annotator had a similar background as the  first annotator that she is a Japanese nonbiologist who has experiences in translation of  (S (PP In (NP the present paper)), (NP-SBJ-55 (NP  the binding) (PP of (NP a [125I]-labeled aldosterone  derivative)) (PP to (NP (NP plasma membrane rich  fractions) (PP of HML)))) (VP was (VP studied *55)).)  Figure 1. The sentence ?In the present paper, the binding of  a [125I]-labeled aldosterone derivative to plasma membrane rich fractions of HML was studied? annotated in  XML and PTB formats.   222 technical documents in English and in corpus  annotation of  English texts.  The two results were examined manually,  and there were 131 disagreements. Almost every  sentence had at least one disagreement. We have  made the ?gold standard? from the two sets of  abstracts by resolving the disagreements, and the  accuracy of the annotators against this gold  standard were 96.7% for the first annotator and  97.4% for the second annotator.   Of the disagreement, the most prominent  were the cases involving coordination, especially the ones with ellipsis. For example, one  annotator annotated the phrase ?IL-1- and IL-18mediated function? as in Figure 2a, the other  annotated as Figure 2b.    Such problem is addressed in the PTB  guideline and both formats are allowed as alternatives. As coordination with ellipsis occurs  rather frequently in research abstracts, this kind  of phenomena has higher effect on decrease of  the agreement rate than in Penn Treebank. Of  the 131 disagreements, 25 wereon this type of  coordination.  Another source of disagreement is the attachment of modifiers such as prepositional  phrases and pronominal adjectives. However,  most are ?benign ambiguity? where the difference of the structure does not affect on interpretation, such as ?high expression of STAT in  monocytes? where the prepositional phrase ?in  monocytes? can attach to ?expression? or  ?STAT? without much difference in meaning,  and ?is augmented when the sensitizing tumor is  a genetically modified variant? where the whclause can attach to ?is augmented? or ?augmented? without changing the meaning. The  PTB guideline states that the modifier should be  attached at the higher level in the former case  and at the lower case in the latter. In the annotation results, one annotator consistently attached  the modifiers in both cases at the higher level,  and the other consistently at the lower level, indicating that the problem is in understanding the  scheme rather than understanding the sentence.  Only 15 cases were true ambiguities that needed  knowledge of biology to solve, in which 5 involved coordination (e.g., the scope of ?various?  in ?various T cell lines and peripheral blood  cells?) .    Although the number was small, there were  disagreements on how to annotate a mathematical formula such as ?n=2? embedded in the sentence, since mathematical formulae were outside  the scope of the original PTB scheme. One annotator annotated this kind of phrase consistently as a phrase with ?=? as an adjective, the  other annotated as phrase with ?=? as a verb.  There were 6 such cases. Another disagreement  particular to abstracts is a treatment of labeled  sentences. There were 8 sentences in two abstracts where there is a label like ?Background:?.   One annotator included the colon (?:?) in the label, while the other did not. Yet another is that  one regarded the phrase ?Author et al as coordination, and the other regarded ?et al as a  modifier.    <NP SYN="COOD">  <NP><ADJP>IL-1- <ADJP NULL="QSTN"/></ADJP>           <NP NULL="RNR" ref="i20"/></NP>  and   <NP>IL-18-mediated <NP NULL="RNR" ref="i20"/></NP>  <NP id="i20">function </NP>   Other disagreements are more general type  such as regarding ?-ed? form of a verb as an adjective or a participle, miscellaneous errors such  as omission of a subtype of label (such as ?PRD? or ?-SBJ) or the position of <PRN> tags  <NP>  <ADJP SYN="COOD">    <ADJP>IL-1- <ADJP NULL="QSTN"/></ADJP>    and     <ADJP>IL-18-mediated </ADJP></ADJP>  function   </NP>      NP               ADJP   Function           ADJP     and ADJP               IL-1  *   IL-18 mediated   Figure 2a. Annotation of a coordinated phrase by the first  annotator. A* denotes a null constituent.  </NP>          NP               NP And NP                  ADJP  *20 IL-18 meidiated NP              IL-1 *      function20 Figure 2b. Annotation of the same phrase as in Figure 2a  by the second annotator.  A * denotes a null constituent  and ?20? denotes coindexing.  223 with regards to ?,? for the inserted phrase, or the  errors which look like just ?careless?. Such disagreements and mistakes are at least partially  eliminated when reliable taggers and parsers are  available for preprocessing  5 Discussion  The result of the inter-annotator agreement  test indicates that the writing style rather than  the contents of the research abstracts is the  source of the difficulty in tree annotation. Contrary to the expectation that the lack of domain  knowledge causes a problem in annotation on  attachments of modifiers, the number of cases  where annotation of modifier attachment needs  domain knowledge is small. This indicates that  linguists can annotate most of syntactic structure  without an expert level of domain knowledge.   A major source of difficulty is coordination,  especially the ones involving ellipsis. Coordination is reported to be difficult phenomena in annotation of different levels in the GENIA corpus  (Tateisi and Tsujii, 2004), (Kim et al, 2003). In  addition to the fact that this is the major source  of inter-annotator agreement, the annotator often  commented the coordinated structure as ?unsure?.  The problem of coordination can be divided into  two with different nature: one is that the annotation policy is still not well-established for the  coordination involving ellipsis, and the other is  an ambiguity when the coordinated phrase has  modifiers.   Syntax annotation of coordination with ellipsis is difficult in general but the more so in annotation of abstracts than in the case of general  texts, because in abstracts authors tend to pack  information in limited number of words. The  PTB guideline dedicates a long section for this  phenomena and allows alternatives in annotation,  but there are still cases which are not wellcovered by the scheme. For example, in addition  to the disagreement, the phrase illustrated in  Figure 2a and Figure 2b shows another problem  of the annotation scheme. Both annotators fail to  indicate that it is ?mediated? that was to be after  ?IL-1? because there is no mechanism of  coindexing a null element with a part of a token.   This problem of ellipsis can frequently occur  in research abstracts, and it can be argued that the tokenization criteria must be changed for  texts in biomedical domain (Yamamoto and Satou, 2004) so that such fragment as ?IL-18? and  ?mediated? in ?IL-18-ediated? should be regarede  as separate tokens. The Pennsylvania biology  corpus (Kulick et al, 2004) partially solves this  problem by separating a token where two or  more subtokens are connected with hyphens, but  in the cases where a shared part of the word is  not separated by a hyphen (e.g. ?metric? of ?stereo- and isometric alleles?) the word including  the part is left uncut. The current GTB follows  the GENIA corpus that it retains the tokenization criteria of the original Penn Treebank, but  this must be reconsidered in future.   For analysis of coordination with ellipsis, if  the information on full forms is available, one  strategy would be to leave the inside structure of  coordination unannotated in the treebank corpus  (and in the phase of text analysis the structure is  not established in the phase of parsing but with a  different mechanism) and later merge it with the  coordination structure annotation. The GENIA  term corpus annotates the full form of a technical term whose part is omitted in the surface as  an attribute of the ?<cons>? element indicating a  technical term (Kim et al, 2003). In the abovementioned Pennsylvania corpus, a similar  mechanism (?chaining?) is used for recovering  the full form of named entities. However, in  both corpora, no such information is available  outside the terms/entities.   The cases where scope of modification in  coordinated phrases is problematic are few but  they are more difficult in abstracts than in general texts because the resolution of ambiguity  needs domain knowledge. If term/entity annotation is already done, that information can help  resolve this type of ambiguity, but again the  problem is that outside the terms/entities such  information is not available. It would be practical to have the structure flat but specially  marked when the tree annotators are unsure and  have a domain expert resolve the ambiguity, as  the sentences that needs such intervention seems  few. Some cases of ambiguity in modifier attachment (which do not involve coordination)  can be solved with similar process.  We believe that other type of disagreements  can be solved with supplementing criteria for  linguistic phenomena not well-covered by the  scheme, and annotator training. Automatic preprocessing by POS taggers and parsers can also  help increase the consistent annotation.  2246 Conclusion  A subset of the GENIA corpus is annotated  for syntactic (tree) structure. Inter-annotator  agreement test indicated that the annotation can  be done stably by linguists without much  knowledge in biology, provided that proper  guideline is established for linguistic phenomena  particular to scientific research abstracts. We  have made the 500-abstract corpus in both XML  and PTB formats and made it publicly available  as ?the GENIA Treebank beta version? (GTBbeta). We are in further cleaning up process of  the 500-abstract set, and at the same time, initial  annotation of the remaining abstracts is being  done, so that the full GENIA set of 2000 abstracts will be annotated with tree structure.   For parsers to be useful for information extraction, they have to establish a map between  syntactic structure and more semantic predicateargument structure, and between the linguistic  predicate-argument structures to the factual relation to be extracted. Annotation of various information on a same set of text can help  establish these maps. For the factual relations,  we are annotating relations between proteins and  genes in cooperation with a group of biologists.  For predicate-argument annotation, we are investigating the use of the parse results of the  Enju parser.  Acknowledgments  The authors are grateful to annotators and colleagues that helped the construction of the corpus. This work is partially supported by Grantin-Aid for Scientific Research on Priority Area  C ?Genome Information Science? from the Ministry of Education, Culture, Sports, Science and  Technology of Japan. '	b"Introduction  Information extraction (IE) from texts currently  receives a large research interest. Traditionally,  it has been associated with the ? often verbatim  ? extraction of domain-specific information  from free text (Riloff & Lorenzen 1999). Input  documents are scanned for very specific relevant  information elements on a particular topic,  which are used to fill out empty slots in a  predefined frame. Other types of systems try to  acquire this knowledge automatically by  detecting reoccurring lexical and syntactic  information from manually annotated example  texts (e.g. Soderland 1999).  Most of these techniques are inherently limited  because they exclude natural language semantics  as much as possible. This is understandable for  reasons of efficiency and genericity but it  restricts the algorithms' possibilities and it  disregards the fact that ? at least in free text ? IE  has much to do with identifying semantic roles.  In most of these systems, case role detection as a  goal in itself has been treated in a rather trivial  way. Our research will try to provide a  systematic approach to case role detection as an  independent extraction task. Using notions from  systemic-functional grammar and presupposing  a possible mapping between morphosyntactic  properties and functional role patterns, we will  develop a general model for case role extraction.  The idea is to learn domain-independent case  role patterns from a tagged corpus, which are  then (automatically) specialized to particular  domain-dependent case role sets and which can  be reassigned to previously unseen text. In this  paper, we will focus on the first part of this task.   For IE, an accurate and speedy detection of  functional case roles is of major importance,  since they describe events (or states) and  participants to these events and thus allow for  identifying real-world entities, their properties  and interactions between them.  1 Theoretical setting  One of the earliest and most notable accounts on  case roles is without any doubt Charles  Fillmore's groundbreaking article (Fillmore  1968). His most fundamental argument is that  the notion of case is not so much connected to  morphosyntactic surface realisation as to  syntactico-semantic categories in the deep  structure of a language. Particular constellations  of case roles determine distinctive functional  patterns, a considerable part of which (according  to Fillmore) is likely to be universally valid.  This deep-structure case system can be realized  in the surface structure by means of a set of  language-dependent transformation rules (see  Fillmore 1968). As a consequence there has to  be a regular mapping between the case system  and its surface realization ? which includes case  markers, word order, grammatical roles, etc.  For  our research, we will disregard the  transformational dimension in Fillmore's theory  but we will nevertheless assume that there is at  least some degree of correspondence between the case role system underlying a language and  its (1) morphosyntax, (2) relative word order and  (3) lexicon.  In Halliday's systemic-functional grammar  (Halliday 1994; Halliday & Matthiessen 1999),  functional patterns that are part of the language's  deep structure are organized as figures, i.e.  configurations of case roles which consist of:  1. A nuclear process, which is typically realized  by a verb phrase. Processes express an event  or state as it is distinctly perceived by the  language user.  2. A limited number of participants, which are  inherent to the process and are typically  realized by noun phrases. They represent  entities or abstractions that participate in the  process.  3. An in theory unlimited number of  circumstantial elements. Circumstances are  in most cases optional and are typically  realized by prepositional or adverbial  phrases. They allocate the process and its  participants in a temporal, spatial, causal, ?  context.   Processes are classified into types and subtypes,  each having its particular participant  combinations. We discern four main process  types: Material, Mental, Verbal and Relational  (Halliday 1994). Figure 1 is an example of a  verbal process, the Sayer being the participant  'doing' the process and the Receiver the one to  whom the (implicit) verbal message is directed.    Invesco in merger talks with AIM Management  Sayer Verbal Process Receiver  Figure 1 ? Example of a verbal process    Since these main types (and some secondary  ones) correspond to universal experiential modi,  it is to be expected that they will have a certain  universal validity, i.e. that they are in some way  or another present in all languages of the world.  For our preliminary experiments, we use a  reduced version of the case role model proposed  by Halliday (1994, p. 106-175), as it is a  consistent, well-developed and relatively simple  system, which makes it very suitable for testing  the validity of our assumptions. For actual  applications, we will replace it by a more  elaborate variant, most likely Bateman's  Generalized Upper Model (Bateman 1990;  Bateman et al in progress). Bateman's model is  finer-grained than Halliday's; it is to a large  extent language-independent; and it has been  specifically developed for implementation into  NLP systems (see Bateman et al in progress).     2 Our approach  Given the framework outlined above, we  consider case role detection to be a standard  classification task. In pattern classification one  attempts to learn certain patterns or rules from  classified examples and to use them for  classifying previously unseen instances (Hand  1997). In our case, a class is a concatenation of  case roles that constitute one particular process  (i.e. the deep structure figure) and the pattern  itself is to be derived from the morphosyntactic  and lexical properties corresponding tothat  process (its surface realisation).  Taking that point-of-view, individual  realisations of figures ? roughly corresponding  to stripped-down clauses ? are translated into  fixed-length sets of lexical and morphosyntactic  features (word order is implicitly encoded) and a  functional classification is manually assigned to  them. For each verb the classification algorithm  then attempts to match all functional patterns to  one or a few relevant sets of distinctive features.  The latter are translated into patterns that can be  used to match an occurrence in a text to a  particular constellation of case roles.  The entire learning process consists of five main  steps:  1. Preprocessing  2. Annotation  3. Feature selection  4. Training of the classifier  5. Translation into rules  In the preprocessing phase, the input text is  tagged, lemmatised and chunked. The output is  standardized and passed to the annotation tool,  in which the user is asked to assign case role  patterns to individual clauses. For now, we will  only take into account processes, participants  and circumstantial elements of Extent and  Location.  In a next step, individual training examples ?  each example corresponding to one figure ? are  converted to a fixed-length feature vector. For  each phrase, the lexical and morphosyntactic  features of the head and of the left and right  context boundaries (i.e. the first and the last  token of the strings pre- and postmodifying the  head) are automatically extracted from the  tagged text and added to the vector. This enables  us to align corresponding features quite  accurately without having to resort to any  complex form of phrasal analysis. Although this  reduction of the context of the head word may  seem to be counter-intuitive from a grammatical  point-of-view, our initial tests indicate that it  does capture most constructions that are relevant  to the extraction task.  Feature selection is necessary for two main  reasons. Firstly, it is impossible to take into  account all lexical and morphosyntactic features,  since that would boost the time-complexity,  incorporate many irrelevant features and bring  down accuracy when a limited set of training  examples is available. Secondly, natural  language utterances have the uncanny habit of  being of variable length. The latter aspect is  problematic not only because classification  algorithms usually expect a clearly delineated  set of features, but also because it is crucial to  align examples in order to compare  correspondent features.  In our test setting, we will constrain the maximal  number of case roles per figure to four. Since  each case role is transformed into a set of 10  features, a figure will be translated into a 40dimensional feature vector (see Figure 2).  As a result, a particular constellation of case  roles is treated as one pattern in which each role  and each of itsrelevant features has a fixed  position. We expect this vector representation to  be relevant in most languages apart from free  word order languages. Currently, our model  focuses on English.  In the fourth step, the classifier is trained to  discriminate features that are distinctive for each  process type associated with a particular verb.  These features are again translated into rules that  can be used for reassigning case roles that have  been learned to previously unseen text.  This is necessary because the variable length of  figures and ? within figures ? of phrases is  bound to cause difficulties when applying the  patterns that were learned to new sentences.  Rules have the advantage over feature vectors in  that they allow us to use head-centred  stretching: when figures are assigned to  previously unseen sentences and no pattern can  immediately be matched, the nearest equivalent  according to the head of the figure will be  assigned; the rest of the pattern will be allocated  by shifting the left and right context of the head  towards the left and right sentence boundaries. A  similar approach will be used for matching  individual roles to phrases.  3 An experiment  Before engaging in the laboursome task of  building a set of tools and tagging an entire  corpus, we decided to test the practical validity  of our ideas on a small scale on the verb be. We  manually constructed a limited set of training  examples (76 occurrences) from the new Reuters  corpus (CD-rom, Reuters Corpus. Volume 1:  English Language, 1996-08-20 to 1997-08-19)  and processed it with the C4.5 classification  algorithm (Quinlan 1993).  Figure 3 gives an overview of the process. The  tagged text1 (step 1) is translated into a set of                                                         1 For our first experiment we used TnT  (http://www.coli.uni-sb.de/~thorsten/tnt/).  In our    Figure 2 ? The feature set    features (step 2). A functional pattern is used as  the class corresponding to the feature set (the  last entry in step 2). The classifier extracts one  or more distinctive features (step 3), which are  in turn transposed into a rule (step 4) that is used  in case role assignment.      Figure 3 ? Schematic illustration of the experiment    Initial results are encouraging. The evaluation  component of C4.5 revealed an error rate of  9.2% when reapplying its rule extractions on the  training data. Given the limited amount of data,  these results are reasonable. Manual application  of therules (from step 4) to new text confirmed  their natural look-and-feel. We are currently  testing the approach with larger amounts of  training and testing data. Most of the current  errors are caused by the limited amount of  training data in our experiment: in a number of  cases there was only one instance of a particular  figure.  4 Discussion and future improvements  Although most shortcomings that arose in our  present set-up can be settled relatively easily, a  number of issues still remains to be resolved.  From a theoretical angle, the most urgent  problem is the underspecification of the material  domain (or the disagreement on exactly how  material processes ought to be subclassified).  Unfortunately, most verb meanings are material                                                                                   present experiment, it is replaced by LT POS  (http://www.ltg.ed.ac.uk/~mikheev/software.html).  We manually lemmatised the tokens, but we are  currently using a lemmatizer based on WordNet.  and distinctions in the material domain tend to  be rather crucial in most IE applications.  Two major implementational difficulties are  related to circumstantial elements. Since  circumstances are normally not inherent to the  process, they do not tend to have a fixed position  in a figure. In addition, no formal parameters  exist to distinguish obligatory circumstances  from optional ones. Since it would be absurd to  encode all variation in separate patterns, it is  tempting just to add empty slots at the most  predictable positions where circumstances could  appear, but that would still not tell apart optional  and obligatory circumstances and it would be a  rather ad hoc solution. We are currently  investigating whether both problems might be  dealt with by encoding the relative position of  case roles explicitly.  In our current information society, it will  become increasingly important to extract  information on well-specified events or entities  from documents. Case role detection will  provide a way to do this by integrating 'real'  semantics into the systems without  overburdening the algorithms. For instance, in  our example analysis (Figure 1) we can  immediately identify two entities involved in a  communicative action, one that does the talking  ('Invesco') and one that is being talked to ('AIM  management'). An immediate application of case  role detection is straightforward IE, which  typically attempts to extract specific information  from a text. However, the algorithm could also  be used for optimising information retrieval  applications, inthe construction of knowledge  bases, in questioning-answering systems or in  case-based reasoning. Actually, for real natural  language understanding a highly accurate model  for interpreting case roles in some form will be  unavoidable.  A major advantage of our approach is that the  pattern base resulting from it will contain  semantic information and yet be fully domainindependent. In a next stage of our research, we  will try to specialize the generic case roles  automatically to domain-dependent ones. At first  sight, this two-step approach might appear  cumbersome, but it will enable us to easily  expand the pattern base while reusing the hardwon patterns.  5 Related research  Historically, case role detection has its roots in  frame-based approaches to IE (e.g. Schank &  Abelson 1977). The main problem here is that to  build case frames one needs prior knowledge on  which information exactly one wants to extract.  In recent years, different solutions have been  offered to automatically generate those frames  from annotated examples (e.g. Riloff &  Schmelzenbach 1998, Soderland 1999) or by  using added knowledge (e.g. Harabagiu &  Maiorano 2000). Many of those approaches  were very successful but most of them have a  tendency to blend syntactic and semantic  concepts and they still have to be trained on  individual domains.  Some very interesting research on case frame  detection has been done by Gildea (Gildea 2000,  Gildea 2001). He uses statistical methods to  learn case frames from parsed examples from  FrameNet (Johnson et al 2001). Conclusion  There is a definite need for case role analysis in  IE and in natural language processing in general.  In this article, we have tried to argue that generic  case role detection is possible by using shallow  text analysis methods. We outlined our  functional framework and presented a model  that considers case role pattern extraction to be a  standard classification task. Our main focus for  the near future will be on automating as many  aspects of the annotation process as possible and  on the construction of the case role assignment  algorithm. In these tasks, the emphasis will be  on genericity and reusability. Acknowledgements  We would like to thank the Institute for the  Promotion of Innovation by Science and  Technology in Flanders (IWT-Flanders) for the  research funding. "
1.0	P02-1063	A00-1031	b'1 Introduction Recently, corpus-based approaches have been widely studied in many natural language processing tasks, such as part-of-speech (POS) tagging, syntactic analysis, text categorization and word sense disambiguation. In corpus-based natural language processing, one important issue is to decide which learning model to use. Various learning models have been studied such as Hidden Markov models (HMMs) (Rabiner and Juang, 1993), decision trees (Breiman et al., 1984) and maximum entropy models (Berger et al, 1996). Recently, Support Vector Machines (SVMs) (Vapnik, 1998; Cortes and Vapnik, 1995) are getting to be used, which are supervised machine learning algorithm for binary classification. SVMs have good generalization performance and can handle a large number of features, and are applied to some tasks ? Presently with Oki Electric Industry successfully (Joachims, 1998; Kudoh and Matsumoto, 2000). However, their computational cost is large and is a weakness of SVMs. In general, a trade-off between capacity and computational cost of learning models exists. For example, SVMs have relatively high generalization capacity, but have high computational cost. On the other hand, HMMs have lower computational cost, but have lower capacity and difficulty in handling data with a large number of features. Learning models with higher capacity may not be of practical use because of their prohibitive computational cost. This problem becomes more serious when a large amount of data is used. To solve this problem, we propose a revision learning method which combines a model with high generalization capacity and a model with small computational cost to achieve high performance with small computational cost. This method is based on the idea that processing the entire target task using a model with higher capacity is wasteful and costly, that is, if a large portion of the task can be processed easily using a model with small computational cost, it should be processed by such a model, and only difficult portion should be processed by the model with higher capacity. Revision learning can handle a general multiclass classification problem, which includes POS tagging, text categorization and many other tasks in natural language processing. We apply this method to English POS tagging and Japanese morphological analysis. This paper is organized as follows: Section 2 describes the general multi-class classification                 Computational Linguistics (ACL), Philadelphia, July 2002, pp. 497-504.                          Proceedings of the 40th Annual Meeting of the Association for problem and the one-versus-rest method which is known as one of the solutions for the problem. Section 3 introduces revision learning, and discusses how to combine learning models. Section 4 describes one way to conduct Japanese morphological analysis with revision learning. Section 5 shows experimental results of English POS tagging and Japanese morphological analysis with revision learning. Section 6 discusses related works, and Section 7 gives conclusion. 2 Multi-Class Classification Problems and the One-versus-Rest Method Let us consider the problem to decide the class ofan example x among multiple classes. Such a problem is called multi-class classification problem. Many tasks in natural language processing such as POS tagging are regarded as a multiclass classification problem. When we only have binary (positive or negative) classification algorithm at hand, we have to reformulate a multiclass classification problem into a binary classification problem. We assume a binary classifier f(x) that returns positive or negative real value for the class of x, where the absolute value |f(x)| reflects the confidence of the classification. The one-versus-rest method is known as one of such methods (Allwein et al, 2000). For one training example of a multi-class problem, this method creates a positive training example for the true class and negative training examples for the other classes. As a result, positive and negative examples for each class are generated. Suppose we have five candidate classes A, B, C, D and E , and the true class of x is B. Figure 1 (left) shows the created training examples. Note that there are only two labels (positive and negative) in contrast with the original problem. Then a binary classifier for each class is trained using the examples, and five classifiers are created for this problem. Given a test example x?, all the classifiers classify the example whether it belongs to a specific class or not. Its class is decided by the classifier that gives the largest value of f(x?). The algorithm is shown in Figure 2 in a pseudo-code. x A : B : C : D : E :  Training Data OX X X X A E B C D A : B :  Training Data OX 123 Rank A E B C D45 xxxxx x xx-X-O -X -X -X -X -O OX Label : Positive: Negative Class Class Figure 1: One-versus-Rest Method (left) and Revision Learning (right) # Training Procedure of One-versus-Rest # This procedure is given training examples # {(xi, yi)}, and creates classifiers. # C = {c0, . . . , ck?1}: the set of classes, # xi: the ith training example, # yi ? C: the class of xi, # k: the number of classes, # l: the number of training examples, # fc(?): the binary classifier for the class c # (see the text). procedure TrainOVR({(x0, y0), . . . , (xl?1, yl?1)}) begin # Create the training data with binary label for i := 0 to l ? 1 begin for j := 0 to k ? 1 begin if cj 6= yi then Add xi to the training data for the class cj as a negative example. else Add xi to the training data for the class cj as a positive example. end end # Train the binary classifiers for j := 0 to k ? 1 Train the classifier fcj (?) using the training data. end # Test Function of One-versus-Rest # This function is given a test example and # returns the predicted class of it. # C = {c0, . . . , ck?1}: the set of classes, # x: thetest example, # k: the number of classes, # fc(?): binary classifier trained with the # algorithm above. function TestOVR(x) begin for j := 0 to k ? 1 confidencej := fcj (x) return cargmaxj confidencej end Figure 2: Algorithm of One-versus-Rest However, this method has the problem of being computationally costly in training, because the negative examples are created for all the classes other than the true class, and the total number of the training examples becomes large (which is equal to the number of original training examples multiplied by the number of classes). The computational cost in testing is also large, because all the classifiers have to work on each test example. 3 Revision Learning As discussed in the previous section, the oneversus-rest method has the problem of computational cost. This problem become more serious when costly binary classifiers are used or when a large amount of data is used. To cope with this problem, let us consider the task of POS tagging. Most portions of POS tagging is not so difficult and a simple POS-based HMMs learning 1 achieves more than 95% accuracy simply using the POS context (Brants, 2000). This means that the low capacity model is enough to do most portions of the task, and we need not use a high accuracy but costly algorithm in every portion of the task. This is the base motivation of the revision model we are proposing here. Revision learning uses a binary classifier with higher capacity to revise the errors made by the stochastic model with lower capacity as follows: During the training phase, a ranking is assigned to each class by the stochastic model for a training example, that is, the candidate classes are sorted in descending order of its conditional probability given the example. Then, the classes are checked in their ranking order to create binary classifiers as follows. If the class is incorrect (i.e. it is not equal to the true class for the example), the example is added to the training data for that class as a negative example, and the next ranked class is checked. If the class is correct, the example is added to the training data for that class as a positive exam1HMMs can be applied to either of unsupervised or supervised learning. In this paper, we use the latter case, i.e., visible Markov Models, where POS-tagged data is used for training. ple, and the remaining ranked classes are not taken into consideration (Figure 1, right). Using these training data, binary classifiers are created. Note that each classifier is a pure binary classifier regardless with the number of classes in the original problem. The binary classifier is trained just for answering whether the output from the stochastic model is correct or not. During the test phase, first the ranking of the candidate classes for a given example is assigned by the stochastic model as in the training. Then the binary classifier classifies the example according to the ranking. If the classifier answers the example as incorrect, the next highest ranked class becomes thenext candidate for checking. But if the example is classified as correct, the class of the classifier is returned as the answer for the example. The algorithm is shown in Figure 3. The amount of training data generated in the revision learning can be much smaller than that in one-versus-rest. Since, in revision learning, negative examples are created only when the stochastic model fails to assign the highest probability to the correct POS tag, whereas negative examples are created for all but one class in the one-versus-rest method. Moreover, testing time of the revision learning is shorter, because only one classifier is called as far as it answers as correct, but all the classifiers are called in the oneversus-rest method. 4 Morphological Analysis with Revision Learning We introduced revision learning for multi-class classification in the previous section. However, Japanese morphological analysis cannot be regarded as a simple multi-class classification problem, because words in a sentence are not separated by spaces in Japanese and the morphological analyzer has to segment the sentence into words as well as to decide the POS tag of the words. So in this section, we describe how to apply revision learning to Japanese morphological analysis. For a given sentence, a lattice consisting of all possible morphemes can be built using a mor# Training Procedure of Revision Learning # This procedure is given training examples # {(xi, yi)}, and creates classifiers. # C = {c0, . . . , ck?1}: the set of classes, # xi: the ith training example, # yi ? C: the class of xi, # k: the number of classes, # l: the number of training examples, # ni: the ordered indexes of C # (see the following code), # fc(?): the binary classifier for the class c # (see the text). procedure TrainRL({(x0, y0), . . . , (xl?1, yl?1)}) begin # Create the training data with binary label for i := 0 to l ? 1 begin Call the stochastic model to obtain the ordered indexes {n0, . . . , nk?1} such that P (cn0 |xi) ? ? ? ? ? P (cnk?1 |xi).for j := 0 to k ? 1 begin if cnj 6= yi then Add xi to the training data for the class cnj as a negative example. else begin Add xi to the training data for the class cnj as a positive example. break end end end # Train the binary classifiers for j := 0 to k ? 1 Train the classifier fcj (?) using the training data. end # Test Function of Revision Learning # This function is given a test example and # returns the predicted class of it. # C = {c0, . . . , ck?1}: the set of classes, # x: the test example, # k: the number of classes, # ni: the ordered indexes of C # (see the following code), # fc(?): binary classifier trained with the # algorithm above. function TestRL(x) begin Call the stochastic model to obtain the ordered indexes {n0, . . . , nk?1} such thatP (cn0 |x) ? ? ? ? ? P (cnk?1 |x).for j := 0 to k ? 1 if fcnj (x) > 0 then return cnj return undecidable end Figure 3: Algorithm of Revision Learning pheme dictionary as in Figure 4. Morphological analysis is conducted by choosing the most likely path on it. We adopt HMMs as the stochastic model and SVMs as the binary classifier. For any sub-paths from the beginning of the sentence (BOS) in the lattice, its generative probability can be calculated using HMMs (Nagata, 1999). We first pick up the end node of the sentence as the current state node, and repeat the following revision learning process backward until the beginning of the sentence. Rankings are calculated by HMMs to all the nodes connected to the current state node, and the best of these nodes is identified based on the SVMs classifiers. The selected node then becomes the current state node in the next round. This can be seen as SVMs deciding whether two adjoining nodes in the lattice are connected or not. In Japanese morphological analysis, for any given morpheme ?, we use the following features for the SVMs: 1. the POS tags, the lexical forms and the inflection forms of the two morphemes preceding ?; 2. the POS tags and the lexical forms of the two morphemes following ?; 3. the lexical form and the inflection form of ?. The preceding morphemes are unknown because the processing is conducted from the end of the sentence, but HMMs can predict the most likely preceding morphemes, and we use them as the features for the SVMs. English POS tagging is regarded as a special case of morphological analysis where the segmentation is done in advance, and can be conducted in the same way. In English POS tagging, given a word w, we use the following features for the SVMs: 1. the POS tags and the lexical forms of the two words preceding w, which are given by HMMs; 2. the POS tags and the lexical forms of the two words following w; 3. the lexical form of w and the prefixes and suffixes of up to four characters, the exisBOS EOS kinou (yesterday) [noun] ki (tree) [noun] nou (brain) [noun] ki (come) [verb] no  [particle] u [auxiliary] gakkou (school) [noun] sentence:  ni (to) [particle] ni (resemble) [verb] it (went) [verb] ta [auxiliary] kinou gakkou it ki ki noun verb noun verb noun... ... Dictionary: Lattice: "kinougakkouniitta (I went to school yesterday)" Figure 4: Example of Lattice for Japanese Morphological Analysis tence of numerals, capital letters and hyphens in w. 5 Experiments This section gives experimental results of English POS tagging and Japanese morphological analysis with revision learning. 5.1 Experiments of English Part-of-Speech Tagging Experiments of English POS tagging with revision learning (RL) are performed on the Penn Treebank WSJ corpus. The corpus is randomly separated into training data of 41,342 sentences and test data of 11,771 sentences. The dictionary for HMMs is constructed from all the words in the training data. T3 of ICOPOST release 0.9.0(Schro?der, 2001) is used as the stochastic model for ranking stage. This is equivalent to POS-based second order HMMs. SVMs with second order polynomial kernel are used as the binary classifier. The results are compared with TnT (Brants, 2000) based on second order HMMs, and with POS tagger using SVMs with one-versus-rest (1v-r) (Nakagawa et al, 2001). The accuracies of those systems for known words, unknown words and all the words are shown in Table 1. The accuracies for both known words and unknown words are improved through revision learning. However, revision learning could not surpass the one-versus-rest. The main difference in the accuracies stems from those for unknown words. The reason for that seems to be that the dictionary of HMMs for POS tagging is obtained from the training data, as a result, virtually no unknown words exist in the training data, and the HMMs never make mistakes for unknown words during the training. So no example of unknown words is available in the training data for the SVM reviser. This is problematic: Though the HMMs handles unknown words with an exceptional method, SVMs cannot learn about errors made by the unknown word processing in the HMMs. To cope with this problem, we force the HMMs to make mistakes by eliminating low frequent words from the dictionary. We eliminated the words appearing only once in the training data so as to make SVMs to learn about unknown words. The results are shown in Table 1 (row ?cutoff-1?). Such procedure improves the accuracies for unknown words. One advantage of revision learning is its small computational cost. We compare the computation time with the HMMs and the one-versusrest. We also use SVMs with linear kernel function that has lower capacity but lower computational cost compared to the second order polynomial kernel SVMs. The experiments are performed on an Alpha 21164A 500MHz processor. Table 2 shows the total number of training examples, training time, testing time and accuracy for each of the five systems. The training time and the testing time of revision learning are considerably smaller than those of the oneversus-rest. Using linear kernel, the accuracy decreases a little, but the computational cost is much lower than the second order polynomial kernel. Accuracy (Known Words / Unknown Words) Number of Errors T3 Original 96.59% (96.90% / 82.74%) 9720 with RL 96.93% (97.23% / 83.55%) 8734 with RL (cutoff-1) 96.98% (97.25% / 85.11%) 8588 TnT 96.62% (96.90% / 84.19%) 9626 SVMs 1-v-r 97.11% (97.34% / 86.80%) 8245 Table 1: Result of English POS Tagging Total Number of Training Time Testing Time Accuracy Examples for SVMs (hour) (second) T3 Original ? 0.004 89 96.59% with RL (polynomial kernel, cutoff-1) 1027840 16 2089 96.98% with RL (linear kernel, cutoff-1) 1027840 2 129 96.94% TnT ? 0.002 4 96.62% SVMs 1-v-r 999984?50 625 55239 97.11% Table 2: Computational Cost of English POS Tagging 5.2 Experiments of Japanese Morphological Analysis We use the RWCP corpus and some additional spoken language data for the experiments of Japanese morphological analysis. The corpus is randomly separated into training data of 33,831sentences and test data of 3,758 sentences. As the dictionary for HMMs, we use IPADIC version 2.4.4 with 366,878 morphemes (Matsumoto and Asahara, 2001) which is originally constructed for the Japanese morphological analyzer ChaSen (Matsumoto et al, 2001). A POS bigram model and ChaSen version 2.2.8 based on variable length HMMs are used as the stochastic models for the ranking stage, and SVMs with the second order polynomial kernel are used as the binary classifier. We use the following values to evaluate Japanese morphological analysis: recall = ?# of correct morphemes in system?s output??# of morphemes in test data? , precision = ?# of correct morphemes in system?s output??# of morphemes in system?s output? , F-measure = 2? recall? precisionrecall + precision . The results of the original systems and those with revision learning are shown in Table 3, which provides the recalls, precisions and Fmeasures for two cases, namely segmentation (i.e. segmentation of the sentences into morphemes) and tagging (i.e. segmentation and POS tagging). The one-versus-rest method is not used because it is not applicable to morphological analysis of non-segmented languages directly. When revision learning is used, all the measures are improved for both POS bigram and ChaSen. Improvement is particularly clear for the tagging task. The numbers of correct morphemes for each POS category tag in the output of ChaSen with and without revision learning are shown in Table 4. Many particles are correctly revised by revision learning. The reason is that the POS tags for particles are often affected by the following words in Japanese, and SVMs can revise such particles because it uses the lexical forms of the following words as the features. This is the advantage of our method compared to simple HMMs, because HMMs have difficulty in handling a lot of features such as the lexical forms of words. 6 Related Works Our proposal is to revise the outputs of a stochastic model using binary classifiers. Brill studied transformation-based error-driven learning (TBL) (Brill, 1995), which conducts POS tagging by applying the transformation rules to the POS tags of a given sentence, and has a resemblance to revision learning in that the second model revises the output of the first model. Word Segmentation Tagging Training Testing Time Time Recall Precision F-measure Recall Precision F-measure (hour) (second) POS Original 98.06% 98.77% 98.42% 95.61% 96.30% 95.96% 0.02 8 bigram with RL 99.06% 99.27% 99.16% 98.13% 98.33% 98.23% 11 184 ChaSen Original 99.06% 99.20% 99.13% 97.67% 97.81% 97.74% 0.05 15 with RL 99.22% 99.34% 99.28% 98.26% 98.37% 98.32% 6 573 Table 3: Result of Morphological Analysis Part-of-Speech # in Test Data Original with RL Difference Noun 41512 40355 40556 +201 Prefix 817 781 784 +3 Verb 8205 8076 8115 +39 Adjective 678 632 655 +23 Adverb 779 735 750 +15 Adnominal 378 373 373 0 Conjunction 258 243 243 0 Particle 20298 19686 19942 +256 Auxiliary 4419 4333 4336 +3 Interjection 94 90 91 +1 Symbol 15665 15647 15651 +4 Others 1 1 1 0 Filler 43 36 36 0 Table 4: The Number of Correctly Tagged Morphemes for EachPOS Category Tag However, our method differs from TBL in two ways. First, our revision learner simply answers whether a given pattern is correct or not, and any types of binary classifiers are applicable. Second, in our model, the second learner is applied to the output of the first learner only once. In contrast, rewriting rules are applied repeatedly in the TBL. Recently, combinations of multiple learners have been studied to achieve high performance (Alpaydm, 1998). Such methodologies to combine multiple learners can be distinguished into two approaches: one is the multi-expert method and the other is the multi-stage method. In the former, each learner is trained and answers independently, and the final decision is made based on those answers. In the latter, the multiple learners are ordered in series, and each learner is trained and answers only if the previous learner rejects the examples. Revision learning belongs to the latter approach. In POS tagging, some studies using the multi-expert method were conducted (van Halteren et al, 2001; Ma`rquez et al., 1999), and Brill and Wu (1998) combined maximum entropy models, TBL, unigram and trigram, and achieved higher accuracy than any of the four learners (97.2% for WSJ corpus). Regarding the multi-stage methods, cascading (Alpaydin and Kaynak, 1998) is well known, and Even-Zohar and Roth (2001) proposed the sequential learning model and applied it to POS tagging. Their methods differ from revision learning in that each learner behaves in the same way and more than one learner is used in their methods, but in revision learning the stochastic model assigns rankings to candidates and the binary classifier selects the output. Furthermore, mistakes made by a former learner are fatal in their methods, but is not so in revision learning because the binary classifier works to revise them. The advantage of the multi-expert method is that each learner can help each other even if it has some weakness, and generalization errors can be decreased. On the other hand, the computational cost becomes large because each learner is trained using every training data and answers for every test data. In contrast, multi-stage methods can decrease the computational cost, and seem to be effective when a large amount of data is used or when a learner with high computational cost such as SVMs is used.7 Conclusion In this paper, we proposed the revision learning method which combines a stochastic model and a binary classifier to achieve higher performance with lower computational cost. We applied it to English POS tagging and Japanese morphological analysis, and showed improvement of accuracy with small computational cost. Compared to the conventional one-versus-rest method, revision learning has much lower computational cost with almost comparable accuracy. Furthermore, it can be applied not only to a simple multi-class classification task but also to a wider variety of problems such as Japanese morphological analysis. Acknowledgments We would like to thank Ingo Schro?der for making ICOPOST publicly available.'	b'1 In t roduct ion   A large number of current language processing sys-  tems use a part-of-speech tagger for pre-processing.  The tagger assigns a (unique or ambiguous) part-of-  speech tag to each token in the input and passes its  output o the next processing level, usually a parser.  Furthermore, there is a large interest in part-of-  speech tagging for corpus annotation projects, who  create valuable linguistic resources by a combination  of automatic processing and human correction.  For both applications, a tagger with the highest  possible accuracy is required. The debate about  which paradigm solves the part-of-speech tagging  problem best is not finished. Recent comparisons  of approaches that can be trained on corpora (van  Halteren et al, 1998; Volk and Schneider, 1998) have  shown that in most cases statistical aproaches (Cut-  ting et al, 1992; Schmid, 1995; Ratnaparkhi, 1996)  yield better results than finite-state, rule-based, or  memory-based taggers (Brill, 1993; Daelemans etal.,  1996). They are only surpassed by combinations of different systems, forming a "voting tagger".  Among the statistical approaches, the Maximum  Entropy framework has a very strong position. Nev-  ertheless, a recent independent comparison of 7 tag-  gets (Zavrel and Daelemans, 1999) has shown that  another approach even works better: Markov mod-  els combined with a good smoothing technique and  with handling of unknown words. This tagger, TnT,  not only yielded the highest accuracy, it also was the  fastest both in training and tagging.  The tagger comparison was organized as a "black-  box test": set the same task to every tagger and  compare the outcomes. This paper describes the  models and techniques used by TnT together with  the implementation.  The reader will be surprised how simple the under-  lying model is. The result of the tagger comparison  seems to support he maxime "the simplest is the  best". However, in this paper we clarify a number  of details that are omitted in major previous pub-  lications concerning tagging with Markov models.  As two examples, (Rabiner, 1989) and (Charniak  et al, 1993) give good overviews of the techniques  and equations used for Markov models and part-of-  speech tagging, but they are not very explicit in the  details that are needed for their application. We ar-  gue that it is not only the choice of the general model  that determines the result of the tagger but also the  various "small" decisions on alternatives.  The aim of this paper is to give a detailed ac-  count of the techniques used in TnT. Additionally,  we present results of the tagger on the NEGRA cor-  pus (Brants et al, 1999) and the Penn Treebank  (Marcus et al, 1993). The Penn Treebank results  reported here for the Markov model approach are  at least equivalent to those reported for the Maxi-  mum Entropyapproach in (Ratnaparkhi, 1996). For  a comparison to other taggers, the reader is referred  to (Zavrel and Daelemans, 1999).  2 Arch i tec ture   2.1 The Under ly ing Model   TnT uses second order Markov models for part-of-  speech tagging. The states of the model represent  tags, outputs represent the words. Transition prob-  abilities depend on the states, thus pairs of tags.  Output probabilities only depend on the most re-  cent category. To be explicit, we calculate  argmax P(tilti-1, ti-2)P(wilti P(tr+l ItT)  ti...iT  (i)  for a given sequence of words w I . . .  W T of length T.  tl... tT are elements of the tagset, the additional  224 tags t - l ,  to, and tT+l are beginning-of-sequence  and end-of-sequence markers. Using these additional  tags, even if they stem from rudimentary process-  ing of punctuation marks, slightly improves tagging  results. This is different from formulas presented  in other publications, which just stop with a "loose  end" at the last word. If sentence boundaries are  not marked in the input, TnT adds these tags if it  encounters one of \\[.!?;\\] as a token.  Transition and output probabilities are estimated  from a tagged corpus. As a first step, we use the  maximum likelihood probabilities /5 which are de-  rived from the relative frequencies:  Unigrams: /5(t3) = f(t3) (2)  N  f(t2, t3) (3)  Bigrams: P(t31t~)= f(t2)  f(ta, t2, t3)  Trigrams: /5(t3ltx,t~) - f ( t l , t2)  (4)  Lexical: /5(w3 It3) - /(w3, t3) (5)  f(t3)  for all tl, t2, t3 in the tagset and w3 in the lexi-  con. N is the total number of tokens in the training  corpus. We define a maximum likelihood probabil-  ity to be zero if the corresponding nominators and  denominators are zero. As a second step, contex-  tual frequencies are smoothed and lexical frequences  are completed by handling words that are not in the  lexicon (see below).  2.2 Smooth ing   Trigram probabilities generated from a corpus usu-  ally cannot directly be used because of the sparse-  data problem. This means that there are not enough  instances for each trigram to reliably estimate the  probability. Furthermore, setting a probability to  zero because the corresponding trigram never oc-  cured in the corpus has an undesired effect. It causes  the probability of a complete sequence to be set to  zero if its use is necessary for a new text sequence,  thus makes it impossible to rank different sequences  containing a zero probability.  The smoothing paradigm that delivers the best  results in TnT is linear interpolation of unigrams,  bigrams, and trigrams. Therefore, we estimate a  trigram probability as follows:  P(t3ltl ,  t2) = AlP(t3) + Ag_/5(t31t2) + A3/5(t3\\[t1, t2)  (6) /5 are maximum likelihood estimates of the proba-  bilities, and A1 + A2 + A3 = 1, so P again represent  probability distributions.  We use the context-independent variant of linear  interpolation, i.e., the values of the As do not depend  on the particular trigram. Contrary to intuition,  this yields better esults than the context-dependent  variant. Due to sparse-data problems, one cannot es-  timate a different set of As for each trigram. There-  fore, it is common practice to group trigrams by fre-  quency and estimate tied sets of As. However, we  are not aware of any publication that has investi-  gated frequency groupings for linear interpolation i part-of-speech tagging. All groupings that we have  tested yielded at most equivalent results to context-  independent linear interpolation. Some groupings  even yielded worse results. The tested groupings  included a) one set of As for each frequency value  and b) two classes (low and high frequency) on the  two ends of the scale, as well as several groupings  in between and several settings for partitioning the  classes.  The values of Ax, A2, and A3 are estimated by  deleted interpolation. This technique successively  removes each trigram from the training corpus and  estimates best values for the As from all other n-  grams in the corpus. Given the frequency counts  for uni-, bi-, and trigrams, the weights can be very  efficiently determined with a processing time linear  in the number of different rigrams. The algorithm  is given in figure 1. Note that subtracting 1 means  taking unseen data into account. Without this sub-  traction the model would overfit the training data  and would generally ield worse results.  2.3 Handling of Unknown Words  Currently, the method of handling unknown words  that seems to work best for inflected languages is  a suffix analysis as proposed in (Samuelsson, 1993).  Tag probabilities are set according to the word\'s end-  ing. The suffix is a strong predictor for word classes,  e.g., words in the Wall Street Journal part of the  Penn Treebank ending in able are adjectives (JJ) in  98% of the cases (e.g. fashionable, variable), the rest  of 2% are nouns (e.g. cable, variable).  The probability distribution for a particular suf-  fix is generated from all words in the training set  that share the same suffix of some predefined max-  imum length. The term suffix as used here means  "final sequence of characters of a word" which is not  necessarily a linguistically meaningful suffix.  Probabilities are smoothed by successive abstrac-  tion. This calculates the probability of a tag t  given the last m letters li of an n letter word:  P(t l ln - r ,+l , . . . ln) .  The sequence of increasingly  more general contexts omits more and more char-  acters ofthe suffix, such that P(t l ln_m+2,. . . , ln) ,   P(t \\ [ ln-m+3,. . . , l~) ,  . . . ,  P(t) are used for smooth-  ing. The recursiou formula is  P(t l l ,_ i+l,  . . . ln)  = P( t l ln - i+ l , . . .  In) + O iP ( t l l~- , . . . ,  ln) (7)  1 +0~  225 set  )%1 ---- )%2 = )%3 = 0  fo reach  t r ig ram tl,t2,t3 with  f(t l ,t2,t3) > 0  depending on the maximum of the fo l low ing  three  va lues :   f(tl ,t2,ts)-- 1 case f (h,t2)- I  " increment )%3 by f ( t l , t2 , t3)   f(t2,t3)-I case f(t2)- I  " increment )%2 by f ( t l , t2 , t s )   f ( t3 ) - - i   case N-1  " increment )%1 by f(t l ,t2,t3)  end  end  normalize )%1, )%2, )%3  Figure 1: Algorithm for calculting the weights for context-independent li ear interpolation )%1, )%2, )%3 when  the n-gram frequencies are known. N is the size of the corpus? If the denominator in one of the expressions  is 0, we define the result of that expression to be 0.  for i = m. . .  0, using the maximum likelihood esti-  mates/5  from frequencies in the lexicon, weights Oi  and the initialization  P(t) =/5(t) .  (8)  The maximum likelihood estimate for a suffix of  length i is derived from corpus frequencies by  P(t i /~- i+l ,  .. l~) = f (t ,  1~-~+1,... l~) ? (9 )   For the Markov model, we need the inverse condi-  tional probabilities P ( / , - i+ l , . . .  lnlt) which are ob-  tained by Bayesian inversion.  A theoretical motivated argumentation uses the  standard eviation of the maximum likelihood prob-  abilities for the weights 0i (Samuelsson, 1993)?  This leaves room for interpretation.  1) One has to identify a good value for m, the  longest suffix used? The approach taken for TnT is  the following: m depends on the word in question.  We use the longest suffix that we can find in the  training set (i.e., for which the frequency is greater  than or equal to 1), but at most 10 characters. This  is an empirically determined choice?  2) We use a context-independent approach for 0i,  as we did for the contextual weights )%i. It turned  out to be a good choice to set al 0i to the standard  deviation of the unconditioned maximum likelihood  probabilities of the tags in the training corpus, i.e.,  weset  1  $  Oi = ~--~(/5(tj) - 15)2 (10)  s 1 j= l   for all i = 0 . . .  m - 1, using a tagset of s tags and  the average  $  /5 = 1 ~/5(t j )  (11)  8  j----I  This usually yields values in the range 0.03 .. .  0.10.  3) We use different estimates for uppercase and  lowercase words, i.e., we maintain two different suffix  tries depending on the capitalization of the word.  This information improves the tagging results?  4) Another freedom concerns the choice of the  words in the lexicon that should be used for suf-  fix handling. Should we use all words, or are some  of them better suited than others? Accepting that  unknown words are most probably infrequent, one  can argue that using suffixes of infrequent words in  the lexicon is a better approximation for unknown  words than using suffixes of frequent words. There-  fore, we restrict the procedure of suffix handling to  words with a frequency smaller than or equal to some  threshold value. Empirically, 10 turned out to be a  good choice for this threshold.  2.4 Cap i ta l i za t ion   Additional information that turned out to be use-  ful for the disambiguation process for several cor-  pora and tagsets is capitalization information. Tags  are usually not informative about capitalization, but  probability distributions of tags around capitalized  words are different from those not capitalized. The  effect is larger for English, which only capitalizes  proper names, and smaller for German, which capi-  talizes all nouns.  We use flags ci that are true if wi is a capitalized  word and false otherwise? These flags are added to  the contextual probability distributions. Instead of  P(tsItl,t2) (12)  we use  P(t3, c3\\[tl , cl, t2, c2) (13)  and equations (3) to (5) are updated accordingly.  This is equivalent o doubling the size of the tagset  and using different ags depending on capitalization.  226 2.5 Beam Search  The processing time of the Viterbi algorithm (Ra-  biner, 1989) can be reduced by introducing a beam  search. Each state that receives a 5 value smaller  than the largest 5 divided by some threshold value  is excluded from further processing. While the  Viterbi algorithm is guaranteed to find the sequence  of states with the highest probability, this is no  longer true when beam search is added. Neverthe-  less, for practical purposes and the right choice of  0, there is virtually no difference between the algo-  rithm with and without a beam. Empirically, a value  of 0 = 1000 turned out to approximately double the  speed of the tagger without affecting the accuracy.  The tagger currently tags between 30~000 and  60,000 tokensper second (including file I/O) on a  Pentium 500 running Linux. The speed mainly de-  pends on the percentage of unknown words and on  the average ambiguity rate.  3 Eva luat ion   We evaluate the tagger\'s performance under several  aspects. First of all, we determine the tagging ac-  curacy averaged over ten iterations. The overall ac-  curacy, as well as separate accuracies for known and  unknown words are measured.  Second, learning curves are presented, that indi-  cate the performance when using training corpora of  different sizes, starting with as few as 1,000 tokens  and ranging to the size of the entire corpus (minus  the test set).  An important characteristic of statistical taggers  is that they not only assign tags to words but also  probabilities in order to rank different assignments.  We distinguish reliable from unreliable assignments  by the quotient of the best and second best assign-  ments 1. All assignments for which this quotient is  larger than some threshold are regarded as reliable,  the others as unreliable. As we will see below, accu-  racies for reliable assignments are much higher.  The tests are performed on partitions of the cor-  pora that use 90% as training set and 10% as test  set, so that the test data is guaranteed to be unseen  during training. Each result is obtained by repeat-  ing the experiment 10 times with different partitions  and averaging the single outcomes.  In all experiments, contiguous test sets are used.  The alternative is a round-robin procedure that puts  every 10th sentence into the test set. We argue that  contiguous test sets yield more realistic results be-  cause completely unseen articles are tagged. Using  the round-robin procedure, parts of an article are al-  ready seen, which significantly reduces the percent-  age of unknown words. Therefore, we expect even  1 By definition, this quotient is co if there is only one pos-  sible tag for a given word.  higher results when testing on every 10th sentence  instead of a contiguous et of 10%.  In the following, accuracy denotes the number of  correctly assigned tags divided by the number of to-  kens in the corpus processed. The tagger is allowed  to assign exactly one tag to each token.  We distinguish the overall accuracy, taking into  account all tokens in the test corpus, and separate  accuracies for known and unknown tokens. The lat-  ter are interesting, since usually unknown tokens are  much more difficult to process than known tokens,  for which a list of valid tags can be found in the  lexicon.  3.1 Tagging the  NEGRA corpus   The German NEGRA corpus consists of 20,000 sen-  tences (355,000 tokens) of newspaper texts (Frank-  furter Rundschau) that are annotated with parts-of-  speech and predicate-argument structures (Skut et al., 1997). It was developed at the Saarland Univer-  sity in Saarbrficken 2. Part of it was tagged at the  IMS Stuttgart. This evaluation only uses the part-  of-speech annotation and ignores structural annota-  tions.  Tagging accuracies for the NEGRA corpus are  shown in table 2.  Figure 3 shows the learning curve of the tagger,  i.e., the accuracy depending on the amount of train-  ing data. Training length is the nmnber of tokens  used for training. Each training length was tested  ten times, training and test sets were randomly cho-  sen and disjoint, results were averaged. The training  length is given on a logarithmic scale.  It is remarkable that tagging accuracy for known  words is very high even for very small training cot-  pora. This means that we have a good chance of  getting the right tag if a word is seen at least once  during training. Average percentages of unknown  tokens are shown in the bottom line of each diagram.  We exploit the fact that the tagger not only de-  termines tags, but also assigns probabilities. If there  is an alternative that has a probability "close to"  that of the best assignment, his alternative can be  viewed as almost equally well suited. The notion of  "close to" is expressed by the distance of probabil-  ities, and this in turn is expressed by the quotient  of probabilities. So, the distance of the probabili-  ties of a best tag tbest and an alternative tag tart  is expressed by P(tbest)/p(tau), which is some value  greater or equal to 1 since the best tag assignment  has the highest probability.  Figure 4 shows the accuracy when separating as-  signments with quotients larger and smaller than  the threshold (hence reliable and unreliable assign-  ments). As expected, we find that accuracies for  2For availability, please check  h~tp ://w~. col i. uni-sb, de/s fb378/negra-corpus  777 227 Table 2: Part-of-speech tagging accuracy for the NEGRA corpus, averaged over 10 test runs, training and  test set are disjoint. The table shows the percentage of unknown tokens, separate accuracies and standard  deviations for known and unknown tokens, as well as the overall accuracy.  percentage  unknowns  NEGRA corpus 11.9%  known  ace .   97.7% 0.23  unknown  acc. (x  89.0% 0.72  overall  aCE. o"  96.7% 0.29  NEGRA Corpus :  POS Learn ing  Curve   100  9O  80  70 /S   6O  50 , i  1 2 5  50.8 46.4 41.4  I i i I I I i  10 20 50 100 200 320 500  36.0 30.7 23.0 18.3 14.3 11.9 11).3  Overall  min =78.1%  max=96.7%  e Known  rain =95.7%  max=97.7%  - - -a- - -  Unknown  rain =61.2% max=89.0%  I  1000 x 1000 Training Length  St  avg. percentage unknown  Figure 3: Learning curve for tagging the NEGRA corpus. The training sets of variable sizes as well as test  sets of 30,000 tokens were randomly chosen. Training and test sets were disjoint, the procedure was repeated  10 times and results were averaged. Percentages of unknowns for 500k and 1000k training are determined  from an untagged extension.  NEGRA Corpus: Accuracy of  re l iab le  assignments  100  99  98  97  A  f " / :. Reliable rain =96.7% max=99.4%  96 I i i i i i i i i i i  2 5 10 20 50 100 500 2000 10000 threshold 0  100 97.9 95.1 92.7 90.3 86.8 84.1 81.0 76.1 71.9 68.3 64.1 62.0 % cases reliable  - 53.5 62.9 69.6 74.5 79.8 82.7 85.2 88.0 89.6 90.8 91.8 92.2 acc. of complement  Figure 4: Tagging accuracy for the NEGRA corpus when separating reliable and unreliable assignments. The  curve shows accuracies for reliable assignments. The numbers at the bottom line indicate the percentage of  reliable assignments and the accuracy of the complement set (i.e., unreliable assignments).  228 Table 5: Part-of-speech tagging accuracy for the Penn Treebank. The table shows the percentage of unknown  tokens, separate accuracies and standard deviations for known and unknown tokens, as well as the overall  accuracy.  I percentage known  unknowns acc. a  Penn Treebank 2.9% 97.0% 0.15  unknown  aCC. O"  85.5% 0.69  overall  aCE. O"  96.7% 0.15  100  9O  80  70 <  60  Penn Treebank: POS Learn ing  Curve   /  50 I I ~ i I I I i I  1 2 5 10 20 50 100 200 500  50.3 42.8 33.4 26.8 20.2 13.2 9.8 7.0 4.4  Overall  rain =78.6%  max=96.7%  Known  rain =95.2%  max=97.0%  Unknown  min =62.2%  max=85.5%  I  1000 ? 1000 Training Length  2.9 avg. percentage unknown  Figure 6: Learning curve for tagging the Penn Treebank. The training sets of variable sizes as well as test sets  of 100,000 tokens were randomly chosen. Training and test sets were disjoint, the procedure was repeated  10 times and results were averaged.  Penn Treebank: Accuracy of reliable assignments  100  99  98  97  Overall  rain =96.6%  max=99.4%  96 i i i i t i i i i i i i  2 5 10 20 50 100 500 2000 10000 threshold 0  100 97.7 94.6 92.2 89.8 86.3 83.5 80.4 76.6 73.8 71.0 67.2 64.5 % cases reliable  - 53.5 62.8 68.9 73.9 79.3 82.6 85.2 87.5 88.8 89.8 91.0 91.6 acc. of complement  Figure 7: Tagging accuracy for the Penn Treebank when separating reliable and unreliable assignments. The  curve shows accuraciesfor reliable assignments. The numbers at the bottom line indicate the percentage of  reliable assignments and the accuracy of the complement set.  229  reliable assignments are much higher than for unre-  liable assignments. This distinction is, e.g., useful  for annotation projects during the cleaning process,  or during pre-processing, sothe tagger can emit mul-  tiple tags if the best tag is classified as unreliable.  3.2 Tagging the Penn Treebank   We use the Wall Street Journal as contained in the  Penn Treebank for our experiments. The annotation  consists of four parts: 1) a context-free structure  augmented with traces to mark movement and dis-  continuous constituents, 2) phrasal categories that  are annotated as node labels, 3) a small set of gram-  matical functions that are annotated as extensions to  the node labels, and 4) part-of-speech tags (Marcus  et al, 1993). This evaluation only uses the part-of-  speech annotation.  The Wall Street Journal part of the Penn Tree-  bank consists of approx. 50,000 sentences (1.2 mil-  lion tokens).  Tagging accuracies for the Penn Treebank are  shown in table 5. Figure 6 shows the learning curve  of the tagger, i.e., the accuracy depending on the  amount of training data. Training length is the num-  ber of tokens used for training. Each training length  was tested ten times. Training and test sets were  disjoint, results are averaged. The training length is  given on a logarithmic scale. As for the NEGRA cor-  pus, tagging accuracy is very high for known tokens  even with small amounts of training data.  We exploit the fact that the tagger not only de-  termines tags, but also assigns probabilities. Figure  7 shows the accuracy when separating assignments  with quotients larger and smaller than the threshold  (hence reliable and unreliable assignments). Again,  we find that accuracies for reliable assignments are  much higher than for unreliable assignments.  3.3 Summary of Part-of-Speech Tagging  Results  Average part-of-speech tagging accuracy is between  96% and 97%, depending on language and tagset,  which is at least on a par with state-of-the-art re-  sults found in the literature, possibly better. For  the Penn Treebank, (Ratnaparkhi, 1996) reports an  accuracy of 96.6% using the Maximum Entropy ap-  proach, our much simpler and therefore faster HMM  approach delivers 96.7%. This comparison eeds to  be re-examined, since we use a ten-fold crossvalida-  tion and averaging of results while Ratnaparkhi only  makes one test run.  The accuracy for known tokens is significantly  higher than for unknown tokens. For the German  newspaper data, results are 8.7% better when the  word was seen before and therefore is in the lexicon,  than when it was not seen before (97.7% vs. 89.0%).  Accuracy for known tokens is high even with very  small amounts of training data. As few as 1000 to- kens are sufficient o achieve 95%-96% accuracy for  them. It is important for the tagger to have seen a  word at least once during training.  Stochastic taggers assign probabilities to tags. We  exploit the probabilities to determine reliability of  assignments. For a subset hat is determined uring  processing by the tagger we achieve accuracy rates  of over 99%. The accuracy of the complement set is  much lower. This information can, e.g., be exploited  in an annotation project to give an additional treat-  ment to the unreliable assignments, or to pass se-  lected ambiguities to a subsequent processing step. 4 Conc lus ion   We have shown that a tagger based on Markov mod-  els yields state-of-the-art results, despite contrary  claims found in the literature. For example, the  Markov model tagger used in the comparison of (van  Halteren et al, 1998) yielded worse results than all  other taggers. In our opinion, a reason for the wrong  claim is that the basic algorithms leave several deci-  sions to the implementor. The rather large amount  of freedom was not handled in detail in previous pub-  lications: handling of start- and end-of-sequence, the  exact smoothing technique, how to determine the  weights for context probabilities, details on handling  unknown words, and how to determine the weights  for unknown words. Note that the decisions we made  yield good results for both the German and the En-  glish Corpus. They do so for several other corpora  as well. The architecture remains applicable to a  large variety of languages.  According to current tagger comparisons (van  Halteren et al, 1998; Zavrel and Daelemans, 1999),  and according to a comparsion of the results pre-  sented here with those in (Ratnaparkhi, 1996), the  Maximum Entropy framework seems to be the only  other approach yielding comparable results to the  one presented here. It is a very interesting future  research topic to determine the advantages of either  of these approaches, to find the reason for their high  accuracies, and to find a good combination of both.  TnT is freely available to universities and re-  lated organizations for research purposes (see  http ://www. coli. uni-sb, de/-thorsten/tnt). Acknowledgements   Many thanks go to Hans Uszkoreit for his sup-  port during the development of TnT. Most of the  work on TnT was carried out while the author  received a grant of the Deutsche Forschungsge-  meinschaft in the Graduiertenkolleg Kognitionswis-  senschaft Saarbriicken. Large annotated corpora re  the pre-requisite for developing and testing part-of-  speech taggers, and they enable the generation of  high-quality language models. Therefore, I would  230  like to thank all the people who took the effort  to annotate the Penn Treebank, the Susanne Cor-  pus, the Stuttgarter Referenzkorpus, the NEGRA  Corpus, the Verbmobil Corpora, and several others.  And, last but not least, I would like to thank the  users of TnT who provided me with bug reports and  valuable suggestions for improvements. '
0.0	P02-1063	C96-1053	b'1 Introduction Recently, corpus-based approaches have been widely studied in many natural language processing tasks, such as part-of-speech (POS) tagging, syntactic analysis, text categorization and word sense disambiguation. In corpus-based natural language processing, one important issue is to decide which learning model to use. Various learning models have been studied such as Hidden Markov models (HMMs) (Rabiner and Juang, 1993), decision trees (Breiman et al., 1984) and maximum entropy models (Berger et al, 1996). Recently, Support Vector Machines (SVMs) (Vapnik, 1998; Cortes and Vapnik, 1995) are getting to be used, which are supervised machine learning algorithm for binary classification. SVMs have good generalization performance and can handle a large number of features, and are applied to some tasks ? Presently with Oki Electric Industry successfully (Joachims, 1998; Kudoh and Matsumoto, 2000). However, their computational cost is large and is a weakness of SVMs. In general, a trade-off between capacity and computational cost of learning models exists. For example, SVMs have relatively high generalization capacity, but have high computational cost. On the other hand, HMMs have lower computational cost, but have lower capacity and difficulty in handling data with a large number of features. Learning models with higher capacity may not be of practical use because of their prohibitive computational cost. This problem becomes more serious when a large amount of data is used. To solve this problem, we propose a revision learning method which combines a model with high generalization capacity and a model with small computational cost to achieve high performance with small computational cost. This method is based on the idea that processing the entire target task using a model with higher capacity is wasteful and costly, that is, if a large portion of the task can be processed easily using a model with small computational cost, it should be processed by such a model, and only difficult portion should be processed by the model with higher capacity. Revision learning can handle a general multiclass classification problem, which includes POS tagging, text categorization and many other tasks in natural language processing. We apply this method to English POS tagging and Japanese morphological analysis. This paper is organized as follows: Section 2 describes the general multi-class classification                 Computational Linguistics (ACL), Philadelphia, July 2002, pp. 497-504.                          Proceedings of the 40th Annual Meeting of the Association for problem and the one-versus-rest method which is known as one of the solutions for the problem. Section 3 introduces revision learning, and discusses how to combine learning models. Section 4 describes one way to conduct Japanese morphological analysis with revision learning. Section 5 shows experimental results of English POS tagging and Japanese morphological analysis with revision learning. Section 6 discusses related works, and Section 7 gives conclusion. 2 Multi-Class Classification Problems and the One-versus-Rest Method Let us consider the problem to decide the class ofan example x among multiple classes. Such a problem is called multi-class classification problem. Many tasks in natural language processing such as POS tagging are regarded as a multiclass classification problem. When we only have binary (positive or negative) classification algorithm at hand, we have to reformulate a multiclass classification problem into a binary classification problem. We assume a binary classifier f(x) that returns positive or negative real value for the class of x, where the absolute value |f(x)| reflects the confidence of the classification. The one-versus-rest method is known as one of such methods (Allwein et al, 2000). For one training example of a multi-class problem, this method creates a positive training example for the true class and negative training examples for the other classes. As a result, positive and negative examples for each class are generated. Suppose we have five candidate classes A, B, C, D and E , and the true class of x is B. Figure 1 (left) shows the created training examples. Note that there are only two labels (positive and negative) in contrast with the original problem. Then a binary classifier for each class is trained using the examples, and five classifiers are created for this problem. Given a test example x?, all the classifiers classify the example whether it belongs to a specific class or not. Its class is decided by the classifier that gives the largest value of f(x?). The algorithm is shown in Figure 2 in a pseudo-code. x A : B : C : D : E :  Training Data OX X X X A E B C D A : B :  Training Data OX 123 Rank A E B C D45 xxxxx x xx-X-O -X -X -X -X -O OX Label : Positive: Negative Class Class Figure 1: One-versus-Rest Method (left) and Revision Learning (right) # Training Procedure of One-versus-Rest # This procedure is given training examples # {(xi, yi)}, and creates classifiers. # C = {c0, . . . , ck?1}: the set of classes, # xi: the ith training example, # yi ? C: the class of xi, # k: the number of classes, # l: the number of training examples, # fc(?): the binary classifier for the class c # (see the text). procedure TrainOVR({(x0, y0), . . . , (xl?1, yl?1)}) begin # Create the training data with binary label for i := 0 to l ? 1 begin for j := 0 to k ? 1 begin if cj 6= yi then Add xi to the training data for the class cj as a negative example. else Add xi to the training data for the class cj as a positive example. end end # Train the binary classifiers for j := 0 to k ? 1 Train the classifier fcj (?) using the training data. end # Test Function of One-versus-Rest # This function is given a test example and # returns the predicted class of it. # C = {c0, . . . , ck?1}: the set of classes, # x: thetest example, # k: the number of classes, # fc(?): binary classifier trained with the # algorithm above. function TestOVR(x) begin for j := 0 to k ? 1 confidencej := fcj (x) return cargmaxj confidencej end Figure 2: Algorithm of One-versus-Rest However, this method has the problem of being computationally costly in training, because the negative examples are created for all the classes other than the true class, and the total number of the training examples becomes large (which is equal to the number of original training examples multiplied by the number of classes). The computational cost in testing is also large, because all the classifiers have to work on each test example. 3 Revision Learning As discussed in the previous section, the oneversus-rest method has the problem of computational cost. This problem become more serious when costly binary classifiers are used or when a large amount of data is used. To cope with this problem, let us consider the task of POS tagging. Most portions of POS tagging is not so difficult and a simple POS-based HMMs learning 1 achieves more than 95% accuracy simply using the POS context (Brants, 2000). This means that the low capacity model is enough to do most portions of the task, and we need not use a high accuracy but costly algorithm in every portion of the task. This is the base motivation of the revision model we are proposing here. Revision learning uses a binary classifier with higher capacity to revise the errors made by the stochastic model with lower capacity as follows: During the training phase, a ranking is assigned to each class by the stochastic model for a training example, that is, the candidate classes are sorted in descending order of its conditional probability given the example. Then, the classes are checked in their ranking order to create binary classifiers as follows. If the class is incorrect (i.e. it is not equal to the true class for the example), the example is added to the training data for that class as a negative example, and the next ranked class is checked. If the class is correct, the example is added to the training data for that class as a positive exam1HMMs can be applied to either of unsupervised or supervised learning. In this paper, we use the latter case, i.e., visible Markov Models, where POS-tagged data is used for training. ple, and the remaining ranked classes are not taken into consideration (Figure 1, right). Using these training data, binary classifiers are created. Note that each classifier is a pure binary classifier regardless with the number of classes in the original problem. The binary classifier is trained just for answering whether the output from the stochastic model is correct or not. During the test phase, first the ranking of the candidate classes for a given example is assigned by the stochastic model as in the training. Then the binary classifier classifies the example according to the ranking. If the classifier answers the example as incorrect, the next highest ranked class becomes thenext candidate for checking. But if the example is classified as correct, the class of the classifier is returned as the answer for the example. The algorithm is shown in Figure 3. The amount of training data generated in the revision learning can be much smaller than that in one-versus-rest. Since, in revision learning, negative examples are created only when the stochastic model fails to assign the highest probability to the correct POS tag, whereas negative examples are created for all but one class in the one-versus-rest method. Moreover, testing time of the revision learning is shorter, because only one classifier is called as far as it answers as correct, but all the classifiers are called in the oneversus-rest method. 4 Morphological Analysis with Revision Learning We introduced revision learning for multi-class classification in the previous section. However, Japanese morphological analysis cannot be regarded as a simple multi-class classification problem, because words in a sentence are not separated by spaces in Japanese and the morphological analyzer has to segment the sentence into words as well as to decide the POS tag of the words. So in this section, we describe how to apply revision learning to Japanese morphological analysis. For a given sentence, a lattice consisting of all possible morphemes can be built using a mor# Training Procedure of Revision Learning # This procedure is given training examples # {(xi, yi)}, and creates classifiers. # C = {c0, . . . , ck?1}: the set of classes, # xi: the ith training example, # yi ? C: the class of xi, # k: the number of classes, # l: the number of training examples, # ni: the ordered indexes of C # (see the following code), # fc(?): the binary classifier for the class c # (see the text). procedure TrainRL({(x0, y0), . . . , (xl?1, yl?1)}) begin # Create the training data with binary label for i := 0 to l ? 1 begin Call the stochastic model to obtain the ordered indexes {n0, . . . , nk?1} such that P (cn0 |xi) ? ? ? ? ? P (cnk?1 |xi).for j := 0 to k ? 1 begin if cnj 6= yi then Add xi to the training data for the class cnj as a negative example. else begin Add xi to the training data for the class cnj as a positive example. break end end end # Train the binary classifiers for j := 0 to k ? 1 Train the classifier fcj (?) using the training data. end # Test Function of Revision Learning # This function is given a test example and # returns the predicted class of it. # C = {c0, . . . , ck?1}: the set of classes, # x: the test example, # k: the number of classes, # ni: the ordered indexes of C # (see the following code), # fc(?): binary classifier trained with the # algorithm above. function TestRL(x) begin Call the stochastic model to obtain the ordered indexes {n0, . . . , nk?1} such thatP (cn0 |x) ? ? ? ? ? P (cnk?1 |x).for j := 0 to k ? 1 if fcnj (x) > 0 then return cnj return undecidable end Figure 3: Algorithm of Revision Learning pheme dictionary as in Figure 4. Morphological analysis is conducted by choosing the most likely path on it. We adopt HMMs as the stochastic model and SVMs as the binary classifier. For any sub-paths from the beginning of the sentence (BOS) in the lattice, its generative probability can be calculated using HMMs (Nagata, 1999). We first pick up the end node of the sentence as the current state node, and repeat the following revision learning process backward until the beginning of the sentence. Rankings are calculated by HMMs to all the nodes connected to the current state node, and the best of these nodes is identified based on the SVMs classifiers. The selected node then becomes the current state node in the next round. This can be seen as SVMs deciding whether two adjoining nodes in the lattice are connected or not. In Japanese morphological analysis, for any given morpheme ?, we use the following features for the SVMs: 1. the POS tags, the lexical forms and the inflection forms of the two morphemes preceding ?; 2. the POS tags and the lexical forms of the two morphemes following ?; 3. the lexical form and the inflection form of ?. The preceding morphemes are unknown because the processing is conducted from the end of the sentence, but HMMs can predict the most likely preceding morphemes, and we use them as the features for the SVMs. English POS tagging is regarded as a special case of morphological analysis where the segmentation is done in advance, and can be conducted in the same way. In English POS tagging, given a word w, we use the following features for the SVMs: 1. the POS tags and the lexical forms of the two words preceding w, which are given by HMMs; 2. the POS tags and the lexical forms of the two words following w; 3. the lexical form of w and the prefixes and suffixes of up to four characters, the exisBOS EOS kinou (yesterday) [noun] ki (tree) [noun] nou (brain) [noun] ki (come) [verb] no  [particle] u [auxiliary] gakkou (school) [noun] sentence:  ni (to) [particle] ni (resemble) [verb] it (went) [verb] ta [auxiliary] kinou gakkou it ki ki noun verb noun verb noun... ... Dictionary: Lattice: "kinougakkouniitta (I went to school yesterday)" Figure 4: Example of Lattice for Japanese Morphological Analysis tence of numerals, capital letters and hyphens in w. 5 Experiments This section gives experimental results of English POS tagging and Japanese morphological analysis with revision learning. 5.1 Experiments of English Part-of-Speech Tagging Experiments of English POS tagging with revision learning (RL) are performed on the Penn Treebank WSJ corpus. The corpus is randomly separated into training data of 41,342 sentences and test data of 11,771 sentences. The dictionary for HMMs is constructed from all the words in the training data. T3 of ICOPOST release 0.9.0(Schro?der, 2001) is used as the stochastic model for ranking stage. This is equivalent to POS-based second order HMMs. SVMs with second order polynomial kernel are used as the binary classifier. The results are compared with TnT (Brants, 2000) based on second order HMMs, and with POS tagger using SVMs with one-versus-rest (1v-r) (Nakagawa et al, 2001). The accuracies of those systems for known words, unknown words and all the words are shown in Table 1. The accuracies for both known words and unknown words are improved through revision learning. However, revision learning could not surpass the one-versus-rest. The main difference in the accuracies stems from those for unknown words. The reason for that seems to be that the dictionary of HMMs for POS tagging is obtained from the training data, as a result, virtually no unknown words exist in the training data, and the HMMs never make mistakes for unknown words during the training. So no example of unknown words is available in the training data for the SVM reviser. This is problematic: Though the HMMs handles unknown words with an exceptional method, SVMs cannot learn about errors made by the unknown word processing in the HMMs. To cope with this problem, we force the HMMs to make mistakes by eliminating low frequent words from the dictionary. We eliminated the words appearing only once in the training data so as to make SVMs to learn about unknown words. The results are shown in Table 1 (row ?cutoff-1?). Such procedure improves the accuracies for unknown words. One advantage of revision learning is its small computational cost. We compare the computation time with the HMMs and the one-versusrest. We also use SVMs with linear kernel function that has lower capacity but lower computational cost compared to the second order polynomial kernel SVMs. The experiments are performed on an Alpha 21164A 500MHz processor. Table 2 shows the total number of training examples, training time, testing time and accuracy for each of the five systems. The training time and the testing time of revision learning are considerably smaller than those of the oneversus-rest. Using linear kernel, the accuracy decreases a little, but the computational cost is much lower than the second order polynomial kernel. Accuracy (Known Words / Unknown Words) Number of Errors T3 Original 96.59% (96.90% / 82.74%) 9720 with RL 96.93% (97.23% / 83.55%) 8734 with RL (cutoff-1) 96.98% (97.25% / 85.11%) 8588 TnT 96.62% (96.90% / 84.19%) 9626 SVMs 1-v-r 97.11% (97.34% / 86.80%) 8245 Table 1: Result of English POS Tagging Total Number of Training Time Testing Time Accuracy Examples for SVMs (hour) (second) T3 Original ? 0.004 89 96.59% with RL (polynomial kernel, cutoff-1) 1027840 16 2089 96.98% with RL (linear kernel, cutoff-1) 1027840 2 129 96.94% TnT ? 0.002 4 96.62% SVMs 1-v-r 999984?50 625 55239 97.11% Table 2: Computational Cost of English POS Tagging 5.2 Experiments of Japanese Morphological Analysis We use the RWCP corpus and some additional spoken language data for the experiments of Japanese morphological analysis. The corpus is randomly separated into training data of 33,831sentences and test data of 3,758 sentences. As the dictionary for HMMs, we use IPADIC version 2.4.4 with 366,878 morphemes (Matsumoto and Asahara, 2001) which is originally constructed for the Japanese morphological analyzer ChaSen (Matsumoto et al, 2001). A POS bigram model and ChaSen version 2.2.8 based on variable length HMMs are used as the stochastic models for the ranking stage, and SVMs with the second order polynomial kernel are used as the binary classifier. We use the following values to evaluate Japanese morphological analysis: recall = ?# of correct morphemes in system?s output??# of morphemes in test data? , precision = ?# of correct morphemes in system?s output??# of morphemes in system?s output? , F-measure = 2? recall? precisionrecall + precision . The results of the original systems and those with revision learning are shown in Table 3, which provides the recalls, precisions and Fmeasures for two cases, namely segmentation (i.e. segmentation of the sentences into morphemes) and tagging (i.e. segmentation and POS tagging). The one-versus-rest method is not used because it is not applicable to morphological analysis of non-segmented languages directly. When revision learning is used, all the measures are improved for both POS bigram and ChaSen. Improvement is particularly clear for the tagging task. The numbers of correct morphemes for each POS category tag in the output of ChaSen with and without revision learning are shown in Table 4. Many particles are correctly revised by revision learning. The reason is that the POS tags for particles are often affected by the following words in Japanese, and SVMs can revise such particles because it uses the lexical forms of the following words as the features. This is the advantage of our method compared to simple HMMs, because HMMs have difficulty in handling a lot of features such as the lexical forms of words. 6 Related Works Our proposal is to revise the outputs of a stochastic model using binary classifiers. Brill studied transformation-based error-driven learning (TBL) (Brill, 1995), which conducts POS tagging by applying the transformation rules to the POS tags of a given sentence, and has a resemblance to revision learning in that the second model revises the output of the first model. Word Segmentation Tagging Training Testing Time Time Recall Precision F-measure Recall Precision F-measure (hour) (second) POS Original 98.06% 98.77% 98.42% 95.61% 96.30% 95.96% 0.02 8 bigram with RL 99.06% 99.27% 99.16% 98.13% 98.33% 98.23% 11 184 ChaSen Original 99.06% 99.20% 99.13% 97.67% 97.81% 97.74% 0.05 15 with RL 99.22% 99.34% 99.28% 98.26% 98.37% 98.32% 6 573 Table 3: Result of Morphological Analysis Part-of-Speech # in Test Data Original with RL Difference Noun 41512 40355 40556 +201 Prefix 817 781 784 +3 Verb 8205 8076 8115 +39 Adjective 678 632 655 +23 Adverb 779 735 750 +15 Adnominal 378 373 373 0 Conjunction 258 243 243 0 Particle 20298 19686 19942 +256 Auxiliary 4419 4333 4336 +3 Interjection 94 90 91 +1 Symbol 15665 15647 15651 +4 Others 1 1 1 0 Filler 43 36 36 0 Table 4: The Number of Correctly Tagged Morphemes for EachPOS Category Tag However, our method differs from TBL in two ways. First, our revision learner simply answers whether a given pattern is correct or not, and any types of binary classifiers are applicable. Second, in our model, the second learner is applied to the output of the first learner only once. In contrast, rewriting rules are applied repeatedly in the TBL. Recently, combinations of multiple learners have been studied to achieve high performance (Alpaydm, 1998). Such methodologies to combine multiple learners can be distinguished into two approaches: one is the multi-expert method and the other is the multi-stage method. In the former, each learner is trained and answers independently, and the final decision is made based on those answers. In the latter, the multiple learners are ordered in series, and each learner is trained and answers only if the previous learner rejects the examples. Revision learning belongs to the latter approach. In POS tagging, some studies using the multi-expert method were conducted (van Halteren et al, 2001; Ma`rquez et al., 1999), and Brill and Wu (1998) combined maximum entropy models, TBL, unigram and trigram, and achieved higher accuracy than any of the four learners (97.2% for WSJ corpus). Regarding the multi-stage methods, cascading (Alpaydin and Kaynak, 1998) is well known, and Even-Zohar and Roth (2001) proposed the sequential learning model and applied it to POS tagging. Their methods differ from revision learning in that each learner behaves in the same way and more than one learner is used in their methods, but in revision learning the stochastic model assigns rankings to candidates and the binary classifier selects the output. Furthermore, mistakes made by a former learner are fatal in their methods, but is not so in revision learning because the binary classifier works to revise them. The advantage of the multi-expert method is that each learner can help each other even if it has some weakness, and generalization errors can be decreased. On the other hand, the computational cost becomes large because each learner is trained using every training data and answers for every test data. In contrast, multi-stage methods can decrease the computational cost, and seem to be effective when a large amount of data is used or when a learner with high computational cost such as SVMs is used.7 Conclusion In this paper, we proposed the revision learning method which combines a stochastic model and a binary classifier to achieve higher performance with lower computational cost. We applied it to English POS tagging and Japanese morphological analysis, and showed improvement of accuracy with small computational cost. Compared to the conventional one-versus-rest method, revision learning has much lower computational cost with almost comparable accuracy. Furthermore, it can be applied not only to a simple multi-class classification task but also to a wider variety of problems such as Japanese morphological analysis. Acknowledgments We would like to thank Ingo Schro?der for making ICOPOST publicly available.'	b'1 In t roduct ion   When analyzing long sentences with two or more pred-  icates (i.e. compound and complex sentences), it is dif-  ficult to grasp the proper structure of sentences hav-  ing a large nmnber of possible dependency (modifier-  modifee relation) structures. This difficulty is more  marked in Japanese than in English, since there are  more syntactically ambiguous structures in Japanese.  Tile Japanese language has few syntactic indicators for  dividing sentences into phrases or clauses, unlike En-  glish with its relative pronouns and subordinate con-  junctions. One of the most critical features of Japanese  is that the difference between a phra~se and a clause is  not cleat\'. Even subjects or other obligatory elements  of clauses are omitted very often when they aye indi-  cated by contexts. In addition, the Japanese language  does not have any parts of speech to clearly indicate  either the beginning or end of a phrase or a clause.  Another critical feature is that the Japanese language  is an almost pure Head-final language, i.e., predicates  and function words to signify the sentence structure ap-  pear at the end of the clause or sentence. This means  that it is syntactically possible for all phrases or clauses  that can modit) predicates to modify all other phrases  or clauses that appear in the latter part of long sen-  tences.  These syntactic haracteristics of the Japanese lan-  guage make it difficult to determine the dependency  (modification) structure of hmg sentences. Simple  parsing of Japanese long sentences inevitably produces  a huge number of possible modification structures. A conventional bottom-up arsing method can reduce  ambiguity in modification by local information in tim  surface structure. However, this inclines toward an  improper output, since the locally highest likelihood is  sometimes low on the whole.  To overcome this problem, several methods to pre-  dict the global structure of long sentences have been  proposed. One is a top-down parsing method by  matching the input sentence and the domain-specific  patterns (Furuse et al, 1992). Improvements made by  other researchers enabled this method to parse irregu-  lar, incomplete and multiplex patterns, by describing  the domain-dependent pa terns in the form of gram-  mar (Doi, Muraki, et M., 1993).  Another method employs global structure presump-  tion to divide a sentence into clauses by utilizing  general exical information. It predicts the sentence  structure prior to syntactic analysis only by utilizing  domain-independent lexical information such as con-  junctive particles, parallel expressions, theme transi-  tion, etc. (Mizuno et al, 1990; Kurohashi et al, 1992).  Lexical Discourse Grammar (LDG) is one of the  approaches with which a global structure of a long  sentence is presumed by focusing on function words  (Kamei et al 1986; Doi et al, 1991). LDG assumes  that Japanese function words, such as conjunctive par- tides (postpositions) located at the end of each clause,  convey modMity, or propositional ttitude, and suggest  global structures of Japanese long sentences in cooper-  ation with modality in predicates, especially within  310  auxiliary verbs. LDG can presume the inter-clausal de-  i)endency within Japanese sentences prior to syntactic  and semantic alLalyses by utilizing tilt; difDrences of  the encapsulating powers each ,\\]apaiLese function word  has, and by utilizing modification preference-between  function words and predicates that reflects consistency  (if lnodality reading or propositional attitude inte, rpre-  tation.  LDG is effective in reducing the syntactic ambigu-  ities, and it him alre~My been applied to a machine  translation system. Ilowever, it has not claritied the  level of tile encapsulation powers of Japanese flint:-  tion words or tile relation between modality and level.  IIence we refined the concept of LDG, lmrticularly tile  conjunction level of function words, and explain the  outline of the refilled LDG in tilts paper. First, we  present he encapsulation power of Japanese function  words, which are classified into six levels. Second, we  state moditication l)refi3renee of ,}apanese conjunctiw\'.  particles that reflects modality within them. Filmlly,  we present evidence of tile h;vels of aaplnLese func-  tion words. We think that tile h;vels of clauses pro-  duce prosodic infbrmation, especially tlw location and  h;ngth of pauses, which are influenced by tile sentence  51obal struchn\'e. We atnalyzed the speech utterances  of a professional new,~ announcer (male) and fonnd a  correla,tion between a particle\'s encapsulating power  and tile pause hmgt;h inserted ai\'ter dm clause with a  conjunctiw; i~ar ich;.  2 Lexical D iscourse Grammar   2.1 Levels of Conjunctive Particles in  J apanese   \\]Jr ,Jap;~n(;se Olnl)lex or compound SelLtellCes~ Sll})()r-  dinate clauses have several dependency levels relative  to the main clause. Conjunctive particles, which are  located at dm end of clauses and which link them, are  classified according to the elemeni;s that the clause  can contain, or to the correh~tion between clauses.  Se, e tl~e fbllowing examples with conjunctive pard-  cles ",,ode"(? "?) ~t,,(t "nagara"( ~,~ 7>" 6) (* is added to  meaningless sentences).  m\')   ,. oao6She)-.e.,(s ECT)  t,~suke(Help)--ta(PAST) -node(Because)  seikou (succeed) -shit a(PAST).  Ite succeeded because she helped him.  K,.re(HJ--w, ( OeIC)  kaet (Rot.m) - t ~   He retunmd while lie was talking.  3)* ~ ~,to~ i~r, " b*~ ~ ~\')i~}o ?:,,  ka lojo(She)-ga(Sm3.U C\'r)  hanashi(Ta110-,,~gara(While )   a(mSa\').  He returned while she was talking.  {  warai(Smile)-nagara(While)  tazune(Ask)- t~(PAST)- node(Becavse) }  wa tad(1)--wa (TOPIC)  k ot~e( Answe~9-ta(PAST ).  i answered as lie asked while tie was smiling.  {  4SU13JEC \')  famine(Ask) t~(l\'AST)- node(ibca,,se) }  wa,\'ai(S,*li 4,,ag, r,(Wm,O  watasi(.l) w,ffTOPKO kotae(Answel9 ta(l"AST).  I answered while i was smiling as lie asked.  A clatuse withthe conjunctive p~Lrticle to express  ~reason\' "node"too\'el can contaii, ~ subjective noun  phrase and an auxiliary verb of past tense "t~d\'(7:),  while a clanse with the particle indicating attendant  action %agara"(\'&~/d 6) cannot, as shown in 1) 3).  Sentence 4) in comparison with 5) shows that a clause  with "node"((/)(?) can subordinate atclause, with "na   g~md\'()7\')~6), but the reverse is impossible. In these  two selt~,011(;eS, brackets { } show subordinaU: clauses.  Consequently, "nagara"() \' /2 6) is ranked at a lower  h,vel than "nodr"( O -0").  in I,DG, conjunction levels of clauses are divided  into six classiiic~tions according to the elements the  clause clm contain, a,s listed in Table 1. These levels  construct ~t hierarchy, i.e., at lower level clause caltlctofi  subordinate a higher level one. The levels also repre-  sent the encapsulating powers of each Japanese. fltIlC-  tion words located at tile end of the clauses. Besides  em0unctiw\' partMes, Japanese conjunction ouns (u:  relative nouns are also classiiied avd assigned it level.  ltere, Japanese conjunction ILtltllLS, such as "toki" (lt~/:  : :when) are nouns that can often be nsed just like COL>  juncdve l~articles when alley are attached at the end  of clause. Japanese relative notln8, such as "lnae" (\\]ii\'j  ::before) are another type of nouns that play roles sim-  il;~r to those of conjunctions in English when they are  moditied by predicative phrases or clauses.  2.2 Modality in Conjunctive Particles  and Modification Preferences  The conjunction lewis we introduced above reduce the  syntactic ambiguities of long sentences. However, in  order to select the most reliable struchtre of sentences,  we use another important discourse feature tile con-  junctive pt~rticles have, i.e., modMity.  LDG assumes Japanese function words have moda\\]-  ity or ~proposidonal attitudes\' ~md suggest global  ,~tructures of Japanese hmg sentences in cooperation  with modality within ~mxiliary verbs. Wc mssllIne du~t  the same kind of modality in a conjunctive particle anti  311  Table 1 Conjunction Level in LDG  levels  LEV.0  LEV.1  LEV.2  LEV.3  LEV.4  LEV.5  Definition of level Example of function words  can contain every element  cannot contain  sentential particles  cannot contain  conclusive modal  cannot contain  probable modal  "to"( ~ =that(quotation))  "k~ra"(~ 6 =because), "node"(69 ~ =because),  "keredomo"( ~:~  E" % =but)  "nara"( t~? 6 =if), "hoka"(~N =besides), "to"( a =when),  "baai"(t~@ =in case of), "toki"("# =when)  "mae"(~ =before), "ato"(~ a =after)  cannot contain "totomoni"( ~& ~ K- =as), "tame"(?: ~5 =because)  tense expressions "todoujini"( ~ ~1~# ~-=at the same time)  cannot contain "nagara"( re ~" 6 =while), "tsu ts u"(o o =while),  particle "ga"(;O ~\') "kotonaku"( t ~ t? < =without)  higher  t  lower  a predicate (or an auxiliary verb) correspond to each  other. From the parsing viewpoint, thissuggests that  each conjunctive particle has modification preference  with certain predicates or auxiliary verbs.  From the viewpoint of modality, there are four pred-  icate types in Japanese; (1) Auxiliary verbs of the first-  type modality (conjecture tc.), (2) Auxiliary verbs of  the second-type modality (necessity etc.), (3) Copula,  and (4) Plain (present and past tense) forms of Verbs.  Here, first-type modality includes conjecture, such as  "darou"(f2"7~ ")) which corresponds to \'may,\' \'can,\'  \'maybe,\' and \'possibly\' in English auxiliary verbs, ad-  verbs, and adjectives. Second-type modality includes  necessity, such as "nakereba-naranai" (re t~ fc t:ft3 6  to? u,) and "ta-hou-ga-yoi"(t: t! ") ~3: u~) which corre-  spond to \'have to\' or \'must,\' and \'had better,\' \'should,\'  or \'preferably\' in English. The Japanese Copula "da"  (?Z) or "desu" (~?\'Y) means definition or speaker\'s  judgment with confidence. Phdn forms of verbs are  the present or past tense forms of verbs without any  modal auxiliary verbs. These forms do not have any  modal morpheme, but when they which appear at the  end of the sentence and are followed by a period they  CAN convey modality, that is, attitudes or intentions  of the subject or speaker. Plain forms of verbs in a  relative clause which modify a nominal phrase do not  have such modMity.  LDG assumes that each conjunctive particle has a  preference in modifying predicates or auxiliary verbs  with consistent modMity. There are six levels of modal-  ity in conjunctive particles, and there are four types of  modality in predicates, as mentioned above. A sub-  ordinate clause with modality modifies a consistent  modality predicate type. The following figure illus-  trates the modality consistency between particles and  auxiliary verbs in Japanese sentences.  {Clause + particle}  Modality  {Clause + auxiliary verb}  Modality  Take the Japanese conjunction oun "toki" (1~ =  when, if) for example. This word corresponds to ei-  ther the English conjunction \'when\' with neutral read-  ing or \'if\' with conjecture modality. When the word  "toki" is used as the \'if\' reading, this word modifies  a clause in which the modality is expressed. In most  cases, auxiliary verbs such as "darou" (?Z,5 ") = may,  maybe) or "ta-hou-ga-yoi"(t~ It 5 7~3: ~ = had better,  should, preferably) express the modality of the modifee  clause. The Japanese language has some words that  indicate or emphasize the fact that the word "toki" is  being used as the "if" reading. One of them is the ad-  verb "moshi" (~o L) that indicates a supposition read-  ing is applicable. This adverb is never used by itself  and always modifies conjunctive forms such as "toki,"  "nara," "to," and so on, and selects or emphasizes the  supposition reading of the conjunctive forms. Another  such word is the particle "w?\' (~:t) , which is usually  used asa topic marker for a sentence. When "wa" is  attached to "toki," that is, in the form of "toki-wa,"  the supposition reading is enhanced. This tendency  is strengthened by the use of comma after the phrase  "toki-wa." The phrase "toki-w?\' tends to be used to  modify phrases with auxiliary verbs of modality.  When this phrase with modality modifies a plain  form of a verb with a period at the end of the sentence,  the readers recognize that the plain form of the verb  contains a kind of modality, such as the subject\'s or  speaker\'s intention. In other words, modality informa-  tion of the subordinate clauses is attached to the plain  form of the main verb. The following figure illustrates  this interpreting mechanism.  312  \\[ Input Sentence Reader  \\[ Morphological Analysis  \\[ Discourse Structure Reference  \\[ Discourse Structure Assumption  <=> ~ict ionary I Discourse Form Preservation \\[ Discourse Structure Analysis  \\[ Synta.ctic & Sem~mtic Analysis  Figure 1 Analysis based on LDG  (Modality}--  1  {Clause + particle} {Clause (Phdn form).}  Modality  In contrast, when a subordinate clause does not have  modality explicitly and modifies a clause with modal-  ity, the readers interpret he subordinate clause as that  with a kind of modality such as conjecture. The fol-  lowing figure illustrates this situation.  ? (Modality)  {Clause + particle}  1  {Clause + auxiliary verb}  Modality  The modality coincidence described in this section  is the base for analyzing Japanese long sentences. The  .Japanese language has few syntactic indicators to show  the segments of sentences, but is rich in semantic indi-  cators which suggest sentence structure. The sem,~n-  tic indicators are the modalities that a wide range of  parts of speech have. Conjunctive particles, adverbs,  and even plain forms of verbs can have modMity in the  Japanese btnguage. The modality structure is the key  to comprehending Japanese long sentences.  2.3 Japanese Sentence Structure Pre-  sumption  We assume that the modality structure (:an mMnly be  detected by lexical intbrmation. Based on this assump-  tion, LDG presumes the sentence structure before syn-  tactic and semantic analyses on the ba~sis of previously  collected lexical information that characterizes the lex-  ical discourse.  Figure 1 shows the configuration of our Japanese  long sentence analyzer based on LDG. Input sentences  are first analyzed morphologically. The part \'Discourse  Structure Analysis\' in Fig. 1 then presumes the sen-  tence structure, before syntactic and semantic analysis.  ttere, \'Discourse\' means an inner-sentence ongruence  in Japanese long sentences thttt contain two or more  predicates.  In order to reduce the huge number of syntactic  structures of Japanese hmg sentences and give priori-  ties to each possible structure, the analyzing method  based on LDG uses global modality structure focusing  on lexic~tl information.  First, theDiscourse Structure Reference module re-  duces the number of possible syntactic structures, us-  ing the level of conjunctive particles described in the  previous ection. After that, the Discourse Structure  Assumption module gives priorities to each possible  syntactic structure, using the modification prefl;rence  based on modality.  3 An Appl icat ion of LDG  3.1 Pause Control with LDG  The level of conjunctive particles, which indicates the  structure of the Japanese long sentences, is the most  important feature of LDG. In this section we apply  the level to another linguistic phenomenml in order to  (\'onfirm the validity of this model.  The sentence structure influences a wide range of  linguistic phenomena. One example is prosodic infor-  mation (Dorffner et al, 1990; Iwata et al, 1990; Kaiki  et al, 1990; Sakai et al, 1990). If the correct sentence  structure is acquired for each input sentence, prosodic  information can be accurately calculated. As yet,  even the most up-to-date, advanced systems have not  achieved the analysis in the deep structure, therefore  sentence structure presumption in the surface struc-  ture is essentiM for a robust prosodic control system.  LDG meets this requirement since it presumes the sen-  tence structure by means of function words occurring  on the surface (Doi, Kamei, et al, 1993). Hereafter,  we propose the prosodic ontrol system based on LDG.  The presumption function ibr sentence structure  (lexical discourse) by LDG is applied to pre-processing  313  Table 2 Pause length data   Conjunction Levels  With a comma Without a comma  Number of cases  LEV. 1 ( "gW\', "ka,\'a", %ode",etc. )  Number of cases  (w i th /w i thout  a pause)  LEV\'. 0 ("~o", "~ te",etc.) 0 -  461.6 11 (11/0)  LEV:2(  "ha\', "~o\', "tara",ete.)  LEV.a("a*o",ete.)  LEV.4( "tame", temo", %odo\',etc. )  LEV. 5 ( "tsu tsu ", "naga~\'a", "zuni",etc.)  Verbs in adverbial form  Verbs in adverbial form + "te"(\'Q  Adjectives in advei\'bikl form  Adjectives in adverbial fbrm + "re"  Predicative auxiliary verb "da"(?2")  Average pause  length \\[msec\\] .. 1 ( 1/ 0)  4 (4 /0 )   11 (11/ 0) 437.0  2 ( 2/ 0) 277.5  5 ( 5/ 0! 421.5  "4 ( 4/ 0) 293.8  40 (40/ 0) 468.8  14 (12/2) 331,8  3 (3 /o )  542.5  410.0  603.1  (w i th /w i thout  a pauae)  7 (\'1/6)  Avel\'age pause  length \\[msee\\]  7.9  3 ( 3/ 0) 563.3  13 (11/ 2) 243.8  1 (0 /1 )  0.0  8 ( 6/ 2) 201.6  5 (2 /a )  12o.o  6 (G/o) 252.1  30 (15/15) 127.4  .... i9. (8 / i1 )  89.2  6 ( 4/ 2) 56.7  3 (3 / .0 )  359.2 ahead of speech synthesis, in a text-to-speech system.  It can presume the global sentence structure through  lexical information without any analysis in the deep  structm\'e. It is also possible to consider the pause  length inserted after each clause in relation to the lex~  ical information in LDG. In other words, pauses are  more fl\'equently inserted after the clause of the higher  conjunction levels than those of the lower levels. Corn  sequently, the pause length and location can be more  efficiently controlled with the LDG conjunction levels.  To develop a text-to-speech onversion system with  LDG, it is necessary to prepare the LDG conjunction  level information of a large nmnber of conjunct equiv-  alents such as conjunctive particles. Statistical data  should also be collected fi\'om human speech and read-  ing, in regard to the correlations between pause length  and the LDG conjunction levels. This substantial data  is added to the lexical information to be used for speech  synthesis in cooperation with pronunciation and ac-  cent.  3.2 Data  Analysis  To confirm the correlation between the conjunction  level and the pause length, we have analyzed speech  data spoken by a professional news announcer (male),  reading newspapers and magazines at a regular speed.  Wc extracted conjunctive particles and verbs, auxiliary  verbs and adjectives in adverbial form from the speech  data, and classified these words by the LDG conjunc-  tion level. The average pause length for each level was  calculated for two separate cases; words preceding a  comma and words without a comma. See Figure 2 and  Table 2.  For words without a cormna (marked with white bars  in Fig. 2), the result shows that the higher the con-  junction level is, the longer the average pause length is  (except for LEV.0, which is a particle for quotation).  This tendency basically does not depend on whether  or not a comma exists after the words. However, for  words with a comma (marked with black bars in Fig. 2)  pause length of Lev. 3 is shorter than that of Lev. 4.  We suppose that the reason for this phenomenon is that a comma adds modality to the words and length-  ens the pauses, as described in the previous section.  Taking the comma effect into consideration, we can  conclude that there is a solid correlation between the  LDG conjunction level and the pause length.  LEV.0 ("to"(&) and "tte"(o\'C)functioning in a  similar way as quotation marks), however, requires a  careful observation. This conjunction level, the highest  rank, can contain every element, even an independent  sentence. In this case, the relation between the con-  junctive particle and its preceding clause is so weak  that a pause tends to be inserted BEFORE the con-  junctive particle, not after it. Therefore, in the present  data, a pause was inserted after theparticle in only one  case out of seven.  LEV.0   LEV. I   LEV.2   LEV.3   LEV.4   LEV.~  adv .   adv,  + \'re  . . . . . . . . . .  I  II I  I  i i  " - - \' - - - -1   I  I t ......  0 i00 200 300 400 500 600  \\[msec\\]  r\'n: without a comma .,,, : with a comma  Figure 2 Average Pause Length Diagram  In Table 1, no level is assigned to two of tile most  314  fl\'equent groups: verbs and auxiliary verbs in adverbial  tbrm, and verbs and auxiliary verbs in tile same form  with a conjunctive particle "te"(-C). These groups  are difficult to allocate to a single level, as dmy are  used in expressing many factors such its parataxis,  cause, means~ attendant circumstances, and because  they vary semantically and syntactically. However, in  reference to the pause length data, the adverbial verbs  ill the former group might fall into LEV.1 or LEV.2,  while those in the latter group with "te"(-\\[) might  full into LEV.4 or LEV.5. Conventionally, these two  groups; are often treated as one "adverbial form", al-  dlough many functional diiferences have been pointed  out between them. Our data supports the difference  with respect o the pause length. There are identical  tendenci~ between two adw;rbial forms of adjectives:  ("-k,F\'(~//) and "-kute"(~ a 9=)) and two adverbial  forms of pseudo adjectives: ("-ni"(~-=-) and "-de"(  r)).  4 Concluding Remarks  We have proposed a practical method for a global  structure anMyzing algorithm of Japanese long sen-  tences with lexical information (Lexical Discourse  Grammar: LDG). This model assumes that Japanese  conjunctive particles convey modality, and modality  structure can basically be detected by lexical informa-  tion. We assign a ~conjunction level\' to each conjunc-  tive particle and reduce the number of possible syntac-  tic structures of Japanese long sentences. In addition,  we assume that all conjunctive particles have a mod-  ification preference according to their modality. This  preference assigns priorities to the possible structures  of tile sentences.  We applied LDG to a prosodic information control  method in a Japanese text-to-speech onversion system  to confirm the conjunction level experimentally. This  method controls pause location and length in speech  synthesis with the conjunction level in LDG, using only  lexical information with no need fi~r syntactic analy-  sis. Even so, it can tune tile pause length more finely  than methods without sentence structure presumption.  AnMyzing speech data, we confirmed a correlation be-  tween the level of a function word and the length of  a pause inserted ai\'~er that word. We are now in the  process of developing a speech synthesis ystem with  this method,by defining the default pause length for  each conjunction level. In future research, LDG will  also be applied to other prosodic information (rhythm  and intonation).  There can be little doubt that LDG will be more  etfective when two or nmre conjunct equivalents ofdif-  ferent levels appear in one sentence, since the LDG con-  junction levels are closely related to the inter-clausal  dependency. Unfortur~ately there were few such cases  in the data used in this paper. In future work, we will  collect such data to proove this hypothesis, in so doing  will refine our method to improveits ability to analyze  long Japanese sentences. '
1.0	P10-1130	A00-1031	b'1 Introduction Unsupervised learning of hierarchical syntactic structure from free-form natural language text is a hard problem whose eventual solution promises to benefit applications ranging from question answering to speech recognition and machine translation. A restricted version of this problem that targets dependencies and assumes partial annotation ? sentence boundaries and part-of-speech (POS) tagging ? has received much attention. Klein and Manning (2004) were the first to beat a simple parsing heuristic, the right-branching baseline; today?s state-of-the-art systems (Headden et al, 2009; Cohen and Smith, 2009; Spitkovsky et al, 2010a) are rooted in their Dependency Model with Valence (DMV), still trained using variants of EM. Pereira and Schabes (1992) outlined three major problems with classic EM, applied to a related problem, constituent parsing. They extended classic inside-outside re-estimation (Baker, 1979) to respect any bracketing constraints included with a training corpus. This conditioning on partial parses addressed all three problems, leading to: (i) linguistically reasonable constituent boundaries and induced grammars more likely to agree with qualitative judgments of sentence structure, which is underdetermined by unannotated text; (ii) fewer iterations needed to reach a good grammar, countering convergence properties that sharply deteriorate with the number of non-terminal symbols, due to a proliferation of local maxima; and (iii) better (in the best case, linear) time complexity per iteration, versus running time that is ordinarily cubic in both sentence length and the total number of non-terminals, rendering sufficiently large grammars computationally impractical. Their algorithm sometimes found good solutions from bracketed corpora but not from raw text, supporting the view that purely unsupervised, selforganizing inference methods can miss the trees for the forest of distributional regularities. This was a promising break-through, but the problem of whence to get partial bracketings was left open. We suggest mining partial bracketings from a cheap and abundant natural language resource: the hyper-text mark-up that annotates web-pages. For example, consider that anchor text can match linguistic constituents, such as verb phrases, exactly: ..., whereas McCain is secure on the topic, Obama <a>[VP worries about winning the pro-Israel vote]</a>. To validate this idea, we created a new data set, novel in combining a real blog?s raw HTML with tree-bank-like constituent structure parses, gener1278 ated automatically. Our linguistic analysis of the most prevalent tags (anchors, bold, italics and underlines) over its 1M+ words reveals a strong connection between syntax and mark-up (all of our examples draw from this corpus), inspiring several simple techniques for automatically deriving parsing constraints. Experiments with both hard and more flexible constraints, as well as with different styles and quantities of annotated training data ? the blog, web news and the web itself, confirm that mark-up-induced constraints consistently improve (otherwise unsupervised) dependency parsing. 2 Intuition and Motivating Examples It is natural to expect hidden structure to seep through when a person annotates a sentence. As it happens, a non-trivial fraction of the world?s population routinely annotates text diligently, if only partially and informally.1 They inject hyper-links, vary font sizes, and toggle colors and styles, using mark-up technologies such as HTML and XML. As noted, web annotations can be indicative ofphrase boundaries, e.g., in a complicated sentence: In 1998, however, as I <a>[VP established in <i>[NP The New Republic]</i>]</a> and Bill Clinton just <a>[VP confirmed in his memoirs]</a>, Netanyahu changed his mind and ... In doing so, mark-up sometimes offers useful cues even for low-level tokenization decisions: [NP [NP Libyan ruler] <a>[NP Mu?ammar al-Qaddafi]</a>] referred to ... (NP (ADJP (NP (JJ Libyan) (NN ruler)) (JJ Mu)) (?? ?) (NN ammar) (NNS al-Qaddafi)) Above, a backward quote in an Arabic name confuses the Stanford parser.2 Yet mark-up lines up with the broken noun phrase, signals cohesion, and moreover sheds light on the internal structure of a compound. As Vadas and Curran (2007) point out, such details are frequently omitted even from manually compiled tree-banks that err on the side of flat annotations of base-NPs. Admittedly, not all boundaries between HTML tags and syntactic constituents match up nicely: ..., but [S [NP the <a><i>Toronto Star</i>][VP reports [NP this][PP in the softest possible way]</a>,[S stating only that ...]]] Combining parsing with mark-up may not be straight-forward, but there is hope: even above, 1Even when (American) grammar schools lived up to their name, they only taught dependencies. This was back in the days before constituent grammars were invented. 2http://nlp.stanford.edu:8080/parser/ one of each nested tag?s boundaries aligns; and Toronto Star?s neglected determiner could be forgiven, certainly within a dependency formulation. 3 A High-Level Outline of Our Approach Our idea is to implement the DMV (Klein and Manning, 2004) ? a standard unsupervised grammar inducer. But instead of learning the unannotated test set, we train with text that contains web mark-up, using various ways of converting HTML into parsing constraints. We still test on WSJ (Marcus et al, 1993), in the standard way, and also check generalization against a hidden data set ? the Brown corpus (Francis and Kucera, 1979). Our parsing constraints come from a blog ? a new corpus we created, the web and news (see Table 1 for corpora?s sentence and token counts). To facilitate future work, we make the final models and our manually-constructed blog data publicly available.3 Although we are unable to share larger-scale resources, our main results should be reproducible, as both linguistic analysis and our best model rely exclusively on the blog. Corpus Sentences POS Tokens WSJ? 49,208 1,028,347 Section 23 2,353 48,201 WSJ45 48,418 986,830 WSJ15 15,922 163,715 Brown100 24,208 391,796 BLOGp 57,809 1,136,659 BLOGt45 56,191 1,048,404 BLOGt15 23,214 212,872 NEWS45 2,263,563,078 32,119,123,561 NEWS15 1,433,779,438 11,786,164,503 WEB45 8,903,458,234 87,269,385,640 WEB15 7,488,669,239 55,014,582,024 Table 1: Sizes of corpora derived from WSJ and Brown, as well as those we collected from the web. 4 Data Sets for Evaluation and Training The appeal of unsupervised parsing lies in its ability to learn from surface text alone; but (intrinsic) evaluation still requires parsed sentences. Following Klein and Manning (2004), we begin with reference constituent parses and compare against deterministically derived dependencies: after pruning out all empty subtrees, punctuation and terminals (tagged # and $) not pronounced where they appear, we drop all sentences with more than a prescribed number of tokens remaining and useautomatic ?head-percolation? rules (Collins, 1999) to convert the rest, as is standard practice. 3http://cs.stanford.edu/?valentin/ 1279 Length Marked POS Bracketings Length Marked POS Bracketings Cutoff Sentences Tokens All Multi-Token Cutoff Sentences Tokens All Multi-Token 0 6,047 1,136,659 7,731 6,015 8 485 14,528 710 684 1 of 57,809 149,483 7,731 6,015 9 333 10,484 499 479 2 4,934 124,527 6,482 6,015 10 245 7,887 365 352 3 3,295 85,423 4,476 4,212 15 42 1,519 65 63 4 2,103 56,390 2,952 2,789 20 13 466 20 20 5 1,402 38,265 1,988 1,874 25 6 235 10 10 6 960 27,285 1,365 1,302 30 3 136 6 6 7 692 19,894 992 952 40 0 0 0 0 Table 2: Counts of sentences, tokens and (unique) bracketings for BLOGp, restricted to only those sentences having at least one bracketing no shorter than the length cutoff (but shorter than the sentence). Our primary reference sets are derived from the Penn English Treebank?s Wall Street Journal portion (Marcus et al, 1993): WSJ45 (sentences with fewer than 46 tokens) and Section 23 of WSJ? (all sentence lengths). We also evaluate on Brown100, similarly derived from the parsed portion of the Brown corpus (Francis and Kucera, 1979). While we use WSJ45 and WSJ15 to train baseline models, the bulk of our experiments is with web data. 4.1 A News-Style Blog: Daniel Pipes Since there was no corpus overlaying syntactic structure with mark-up, we began constructing a new one by downloading articles4 from a newsstyle blog. Although limited to a single genre ? political opinion, danielpipes.org is clean, consistently formatted, carefully edited and larger than WSJ (see Table 1). Spanning decades, Pipes? editorials are mostly in-domain for POS taggers and tree-bank-trained parsers; his recent (internetera) entries are thoroughly cross-referenced, conveniently providing just the mark-up we hoped to study via uncluttered (printer-friendly) HTML.5 After extracting moderately clean text and mark-up locations, we used MxTerminator (Reynar and Ratnaparkhi, 1997) to detect sentence boundaries. This initial automated pass begot multiple rounds of various semi-automated clean-ups that involved fixing sentence breaking, modifying parser-unfriendly tokens, converting HTML entities and non-ASCII text, correcting typos, and so on. After throwing away annotations of fractional words (e.g., <i>basmachi</i>s) and tokens (e.g., <i>Sesame Street</i>-like), we broke up all markup that crossed sentence boundaries (i.e., loosely speaking, replaced constructs like <u>...][S...</u> with <u>...</u> ][S <u>...</u>) and discarded any 4http://danielpipes.org/art/year/all 5http://danielpipes.org/article print.php? id=. . . tags left covering entire sentences. We finalized two versions of the data: BLOGt, tagged with the Stanford tagger (Toutanova and Manning, 2000; Toutanova et al, 2003),6 and BLOGp, parsed with Charniak?s parser (Charniak, 2001; Charniak and Johnson, 2005).7 The reason for this dichotomy was to use state-of-the-art parses to analyze the relationship between syntax and mark-up, yet to prevent jointly tagged (and non-standard AUX[G]) POS sequences from interfering with our (otherwise unsupervised) training.8 4.2 Scaled up Quantity: The (English) Web We built a large (see Table 1) but messy data set, WEB ? English-looking web-pages, pre-crawled by a search engine. To avoid machine-generated spam, we excluded low quality sites flagged by the indexing system. We kept onlysentence-like runs of words (satisfying punctuation and capitalization constraints), POS-tagged with TnT (Brants, 2000). 4.3 Scaled up Quality: (English) Web News In an effort to trade quantity for quality, we constructed a smaller, potentially cleaner data set, NEWS. We reckoned editorialized content would lead to fewer extracted non-sentences. Perhaps surprisingly, NEWS is less than an order of magnitude smaller than WEB (see Table 1); in part, this is due to less aggressive filtering ? we trust sites approved by the human editors at Google News.9 In all other respects, our pre-processing of NEWS pages was identical to our handling of WEB data. 6http://nlp.stanford.edu/software/ stanford-postagger-2008-09-28.tar.gz 7ftp://ftp.cs.brown.edu/pub/nlparser/ parser05Aug16.tar.gz 8However, since many taggers are themselves trained on manually parsed corpora, such as WSJ, no parser that relies on external POS tags could be considered truly unsupervised; for a fully unsupervised example, see Seginer?s (2007) CCL parser, available at http://www.seggu.net/ccl/ 9http://news.google.com/ 1280 5 Linguistic Analysis of Mark-Up Is there a connection between mark-up and syntactic structure? Previous work (Barr et al, 2008) has only examined search engine queries, showing that they consist predominantly of short noun phrases. If web mark-up shared a similar characteristic, it might not provide sufficiently disambiguating cues to syntactic structure: HTML tags could be too short (e.g., singletons like ?click <a>here</a>?) or otherwise unhelpful in resolving truly difficult ambiguities (such as PPattachment). We began simply by counting various basic events in BLOGp. Count POS Sequence Frac Sum 1 1,242 NNP NNP 16.1% 2 643 NNP 8.3 24.4 3 419 NNP NNP NNP 5.4 29.8 4 414 NN 5.4 35.2 5 201 JJ NN 2.6 37.8 6 138 DT NNP NNP 1.8 39.5 7 138 NNS 1.8 41.3 8 112 JJ 1.5 42.8 9 102 VBD 1.3 44.1 10 92 DT NNP NNP NNP 1.2 45.3 11 85 JJ NNS 1.1 46.4 12 79 NNP NN 1.0 47.4 13 76 NN NN 1.0 48.4 14 61 VBN 0.8 49.2 15 60 NNP NNP NNP NNP 0.8 50.0 BLOGp +3,869 more with Count ? 49 50.0% Table 3: Top 50% of marked POS tag sequences. Count Non-Terminal Frac Sum 1 5,759 NP 74.5% 2 997 VP 12.9 87.4 3 524 S 6.8 94.2 4 120 PP 1.6 95.7 5 72 ADJP 0.9 96.7 6 61 FRAG 0.8 97.4 7 41 ADVP 0.5 98.0 8 39 SBAR 0.5 98.5 9 19 PRN 0.2 98.7 10 18 NX 0.2 99.0 BLOGp +81 more with Count ? 16 1.0% Table 4: Top 99% of dominating non-terminals. 5.1 Surface Text Statistics Out of 57,809 sentences, 6,047 (10.5%) are annotated (see Table 2); and 4,934 (8.5%) have multitoken bracketings. We do not distinguish HTML tags and track only unique bracketing end-points within a sentence. Of these, 6,015 are multi-token ? an average per-sentence yield of 10.4%.10 10A non-trivial fraction of our corpus is older (pre-internet) unannotated articles, so this estimate may be conservative. As expected, many of the annotated words are nouns, but there are adjectives, verbs and other parts of speech too (see Table 3). Mark-up is short, typically under five words, yet (by far) the most frequentlymarked sequence of POS tags is a pair. 5.2 Common Syntactic Subtrees For three-quarters of all mark-up, the lowest dominating non-terminal is a noun phrase (see Table 4); there are also non-trace quantities of verb phrases (12.9%) and other phrases, clauses and fragments. Of the top fifteen ? 35.2% of all ? annotated productions, only one is not a noun phrase (see Table 5, left). Four of the fifteen lowest dominating non-terminals do not match the entire bracketing ? all four miss the leading determiner, as we saw earlier. In such cases, we recursively split internal nodes until the bracketing aligned, as follows: [S [NP the <a>Toronto Star][VP reports [NP this] [PP in the softest possible way]</a>,[S stating ...]]] S? NP VP? DT NNP NNP VBZ NP PP S We can summarize productions more compactly by using a dependency framework and clipping off any dependents whose subtrees do not cross a bracketing boundary, relative to the parent. Thus, DT NNP NNP VBZ DT IN DT JJS JJ NN becomes DT NNP VBZ, ?the <a>Star reports</a>.? Viewed this way, the top fifteen (now collapsed) productions cover 59.4% of all cases and include four verb heads, in addition to a preposition and an adjective (see Table 5, right). This exposes five cases of inexact matches, three of which involve neglected determiners or adjectives to the left of the head. In fact, the only case that cannot be explained by dropped dependents is #8, where the daughters are marked but the parent is left out. Most instances contributing to this pattern are flat NPs that end with a noun, incorrectly assumed to be the head of all other words in the phrase, e.g., ... [NP a 1994 <i>New Yorker</i> article] ... As this example shows, disagreements (as well as agreements) between mark-up and machinegenerated parse trees with automatically percolated heads should be taken with a grain of salt.11 11In a relatively recent study, Ravi et al (2008) report that Charniak?s re-ranking parser (Charniak and Johnson, 2005) ? reranking-parserAug06.tar.gz, also available from ftp://ftp.cs.brown.edu/pub/nlparser/ ? attains 86.3% accuracy when trained on WSJ and tested against Brown; its nearly 5% performance loss out-of-domain is consistent with the numbers originally reported by Gildea (2001). 1281 Count Constituent Production Frac Sum 1 746 NP? NNP NNP 9.6% 2 357 NP? NNP 4.6 14.3 3 266 NP? NP PP 3.4 17.7 4 183 NP? NNP NNP NNP 2.4 20.1 5 165 NP? DT NNP NNP 2.1 22.2 6 140 NP? NN 1.8 24.0 7 131 NP? DT NNP NNP NNP 1.7 25.7 8 130 NP? DT NN 1.7 27.4 9 127 NP? DT NNP NNP 1.6 29.0 10 109 S ? NP VP 1.4 30.4 11 91 NP? DT NNP NNP NNP 1.2 31.6 12 82 NP? DT JJ NN 1.1 32.7 13 79 NP? NNS 1.0 33.7 14 65 NP? JJ NN 0.8 34.5 15 60 NP? NP NP 0.8 35.3 BLOGp +5,000 more with Count ? 60 64.7% Count Head-Outward Spawn Frac Sum 1 1,889 NNP 24.4% 2 623 NN 8.1 32.5 3 470 DT NNP 6.1 38.6 4458 DT NN 5.9 44.5 5 345 NNS 4.5 49.0 6 109 NNPS 1.4 50.4 7 98 VBG 1.3 51.6 8 96 NNP NNP NN 1.2 52.9 9 80 VBD 1.0 53.9 10 77 IN 1.0 54.9 11 74 VBN 1.0 55.9 12 73 DT JJ NN 0.9 56.8 13 71 VBZ 0.9 57.7 14 69 POS NNP 0.9 58.6 15 63 JJ 0.8 59.4 BLOGp +3,136 more with Count ? 62 40.6% Table 5: Top 15 marked productions, viewed as constituents (left) and as dependencies (right), after recursively expanding any internal nodes that did not align with the bracketing (underlined). Tabulated dependencies were collapsed, dropping any dependents that fell entirely in the same region as their parent (i.e., both inside the bracketing, both to its left or both to its right), keeping only crossing attachments. 5.3 Proposed Parsing Constraints The straight-forward approach ? forcing mark-up to correspond to constituents ? agrees with Charniak?s parse trees only 48.0% of the time, e.g., ... in [NP<a>[NP an analysis]</a>[PP of perhaps the most astonishing PC item I have yet stumbled upon]]. This number should be higher, as the vast majority of disagreements are due to tree-bank idiosyncrasies (e.g., bare NPs). Earlier examples of incomplete constituents (e.g., legitimately missing determiners) would also be fine in many linguistic theories (e.g., as N-bars). A dependency formulation is less sensitive to such stylistic differences. We begin with the hardest possible constraint on dependencies, then slowly relax it. Every example used to demonstrate a softer constraint doubles as a counter-example against all previous versions. ? strict ? seals mark-up into attachments, i.e., inside a bracketing, enforces exactly one external arc ? into the overall head. This agrees with head-percolated trees just 35.6% of the time, e.g., As author of <i>The Satanic Verses</i>, I ... ? loose ? same as strict, but allows the bracketing?s head word to have external dependents. This relaxation already agrees with head-percolated dependencies 87.5% of the time, catching many (though far from all) dropped dependents, e.g., . . . the <i>Toronto Star</i> reports . . . ? sprawl ? same as loose, but now allows all words inside a bracketing to attach external dependents.12 This boosts agreement with headpercolated trees to 95.1%, handling new cases, e.g., where ?Toronto Star? is embedded in longer mark-up that includes its own parent ? a verb: . . . the <a>Toronto Star reports . . .</a> . . . ? tear ? allows mark-up to fracture after all, requiring only that the external heads attaching the pieces lie to the same side of the bracketing. This propels agreement with percolated dependencies to 98.9%, fixing previously broken PP-attachment ambiguities, e.g., a fused phrase like ?Fox News in Canada? that detached a preposition from its verb: ... concession ... has raised eyebrows among those waiting [PP for <a>Fox News][PP in Canada]</a>. Most of the remaining 1.1% of disagreements are due to parser errors. Nevertheless, it is possible for mark-up to be torn apart by external heads from both sides. We leave this section with a (very rare) true negative example. Below,?CSA? modifies ?authority? (to its left), appositively, while ?AlManar? modifies ?television? (to its right):13 The French broadcasting authority, <a>CSA, banned ... Al-Manar</a> satellite television from ... 12This view evokes the trapezoids of the O(n3) recognizer for split head automaton grammars (Eisner and Satta, 1999). 13But this is a stretch, since the comma after ?CSA? renders the marked phrase ungrammatical even out of context. 1282 6 Experimental Methods and Metrics We implemented the DMV (Klein and Manning, 2004), consulting the details of (Spitkovsky et al, 2010a). Crucially, we swapped out inside-outside re-estimation in favor of Viterbi training. Not only is it better-suited to the general problem (see ?7.1), but it also admits a trivial implementation of (most of) the dependency constraints we proposed.14 5 10 15 20 25 30 35 40 45 4.5 5.0 5.5 WSJk bpt lowest cross-entropy (4.32bpt) attained at WSJ8 x-Entropy h (in bits per token) on WSJ15 Figure 1: Sentence-level cross-entropy on WSJ15 for Ad-Hoc? initializers of WSJ{1, . . . , 45}. Six settings parameterized each run: ? INIT: 0? default, uniform initialization; or 1 ? a high quality initializer, pre-trained using Ad-Hoc? (Spitkovsky et al, 2010a): we chose the Laplace-smoothed model trained at WSJ15 (the ?sweet spot? data gradation) but initialized off WSJ8, since that ad-hoc harmonic initializer has the best cross-entropy on WSJ15 (see Figure 1). ? GENRE: 0? default, baseline training on WSJ; else, uses 1? BLOGt; 2? NEWS; or 3? WEB. ? SCOPE: 0 ? default, uses all sentences up to length 45; if 1, trains using sentences up to length 15; if 2, re-trains on sentences up to length 45, starting from the solution to sentences up to length 15, as recommended by Spitkovsky et al (2010a). ? CONSTR: if 4, strict; if 3, loose; and if 2, sprawl. We did not implement level 1, tear. Overconstrained sentences are re-attempted at successively lower levels until they become possible to parse, if necessary at the lowest (default) level 0.15 ? TRIM: if 1, discards any sentence without a single multi-token mark-up (shorter than its length). ? ADAPT: if 1, upon convergence, initializes retraining on WSJ45 using the solution to <GENRE>, attempting domain adaptation (Lee et al, 1991). These make for 294 meaningful combinations. We judged each one by its accuracy on WSJ45, using standard directed scoring ? the fraction of correct dependencies over randomized ?best? parse trees. 14We analyze the benefits of Viterbi training in a companion paper (Spitkovsky et al, 2010b), which dedicates more space to implementation and to the WSJ baselines used here. 15At level 4, <b> X<u> Y</b> Z</u> is over-constrained. 7 Discussion of Experimental Results Evaluation on Section 23 of WSJ and Brown reveals that blog-training beats all published stateof-the-art numbers in every traditionally-reported length cutoff category, with news-training not far behind. Here is a mini-preview of these results, for Section 23 of WSJ10 and WSJ? (from Table 8): WSJ10 WSJ? (Cohen and Smith, 2009) 62.0 42.2 (Spitkovsky et al, 2010a) 57.1 45.0 NEWS-best 67.3 50.1 BLOGt-best 69.3 50.4 (Headden et al, 2009) 68.8 Table 6: Directed accuracies on Section23 of WSJ{10,? } for three recent state-of-the-art systems and our best runs (as judged against WSJ45) for NEWS and BLOGt (more details in Table 8). Since our experimental setup involved testing nearly three hundred models simultaneously, we must take extreme care in analyzing and interpreting these results, to avoid falling prey to any looming ?data-snooping? biases.16 In a sufficiently large pool of models, where each is trained using a randomized and/or chaotic procedure (such as ours), the best may look good due to pure chance. We appealed to three separate diagnostics to convince ourselves that our best results are not noise. The most radical approach would be to write off WSJ as a development set and to focus only on the results from the held-out Brown corpus. It was initially intended as a test of out-of-domain generalization, but since Brown was in no way involved in selecting the best models, it also qualifies as a blind evaluation set. We observe that our best models perform even better (and gain more ? see Table 8) on Brown than on WSJ ? a strong indication that our selection process has not overfitted. Our second diagnostic is a closer look at WSJ. Since we cannot graph the full (six-dimensional) set of results, we begin with a simple linear regression, using accuracy on WSJ45 as the dependent variable. We prefer this full factorial design to the more traditional ablation studies because it allows us to account for and to incorporate every single experimental data point incurred along the 16In the standard statistical hypothesis testing setting, it is reasonable to expect that p% of randomly chosen hypotheses will appear significant at the p% level simply by chance. Consequently, multiple hypothesis testing requires re-evaluating significance levels ? adjusting raw p-values, e.g., using the Holm-Bonferroni method (Holm, 1979). 1283 Corpus Marked Sentences All Sentences POS Tokens All Bracketings Multi-Token Bracketings BLOGt45 5,641 56,191 1,048,404 7,021 5,346 BLOG?t45 4,516 4,516 104,267 5,771 5,346 BLOGt15 1,562 23,214 212,872 1,714 1,240 BLOG?t15 1,171 1,171 11,954 1,288 1,240 NEWS45 304,129,910 2,263,563,078 32,119,123,561 611,644,606 477,362,150 NEWS?45 205,671,761 205,671,761 2,740,258,972 453,781,081 392,600,070 NEWS15 211,659,549 1,433,779,438 11,786,164,503 365,145,549 274,791,675 NEWS?15 147,848,358 147,848,358 1,397,562,474 272,223,918 231,029,921 WEB45 1,577,208,680 8,903,458,234 87,269,385,640 3,309,897,461 2,459,337,571 WEB?45 933,115,032 933,115,032 11,552,983,379 2,084,359,555 1,793,238,913 WEB15 1,181,696,194 7,488,669,239 55,014,582,024 2,071,743,595 1,494,675,520 WEB?15 681,087,020 681,087,020 5,813,555,341 1,200,980,738 1,072,910,682 Table 7: Counts of sentences, tokens and (unique) bracketings for web-based data sets; trimmed versions, restricted to only those sentences having at least one multi-token bracketing, are indicated by a prime (?). way. Its output is a coarse, high-level summary of our runs, showing which factors significantly contribute to changes in error rate on WSJ45: Parameter (Indicator) Setting ?? p-value INIT 1 ad-hoc @WSJ8,15 11.8 *** GENRE 1 BLOGt -3.7 0.06 2 NEWS -5.3 ** 3 WEB -7.7 *** SCOPE 1 @15 -0.5 0.40 2 @15?45 -0.4 0.53 CONSTR 2 sprawl 0.9 0.23 3 loose 1.0 0.15 4 strict 1.8 * TRIM 1 drop unmarked -7.4 *** ADAPT 1 WSJ re-training 1.5 ** Intercept (R2Adjusted = 73.6%) 39.9 *** We use a standard convention: *** forp < 0.001; ** for p < 0.01 (very signif.); and * for p < 0.05 (signif.). The default training mode (all parameters zero) is estimated to score 39.9%. A good initializer gives the biggest (double-digit) gain; both domain adaptation and constraints also make a positive impact. Throwing away unannotated data hurts, as does training out-of-domain (the blog is least bad; the web is worst). Of course, this overview should not be taken too seriously. Overly simplistic, a first order model ignores interactions between parameters. Furthermore, a least squares fit aims to capture central tendencies, whereas we are more interested in outliers ? the best-performing runs. A major imperfection of the simple regression model is that helpful factors that require an interaction to ?kick in? may not, on their own, appear statistically significant. Our third diagnostic is to examine parameter settings that give rise to the best-performing models, looking out for combinations that consistently deliver superior results. 7.1 WSJ Baselines Just two parameters apply to learning from WSJ. Five of their six combinations are state-of-the-art, demonstrating the power of Viterbi training; only the default run scores worse than 45.0%, attained by Leapfrog (Spitkovsky et al, 2010a), on WSJ45: Settings SCOPE=0 SCOPE=1 SCOPE=2 INIT=0 41.3 45.0 45.2 1 46.6 47.5 47.6 @45 @15 @15?45 7.2 Blog Simply training on BLOGt instead of WSJ hurts: GENRE=1 SCOPE=0 SCOPE=1 SCOPE=2 INIT=0 39.6 36.9 36.9 1 46.5 46.3 46.4 @45 @15 @15?45 The best runs use a good initializer, discard unannotated sentences, enforce the loose constraint on the rest, follow up with domain adaptation and benefit from re-training ? GENRE=TRIM=ADAPT=1: INIT=1 SCOPE=0 SCOPE=1 SCOPE=2 CONSTR=0 45.8 48.3 49.6 (sprawl) 2 46.3 49.2 49.2 (loose) 3 41.3 50.2 50.4 (strict) 4 40.7 49.9 48.7 @45 @15 @15?45 The contrast between unconstrained learning and annotation-guided parsing is higher for the default initializer, still using trimmed data sets (just over a thousand sentences for BLOG?t15 ? see Table 7): INIT=0 SCOPE=0 SCOPE=1 SCOPE=2 CONSTR=0 25.6 19.4 19.3 (sprawl) 2 25.2 22.7 22.5 (loose) 3 32.4 26.3 27.3 (strict) 4 36.2 38.7 40.1 @45 @15 @15?45 Above, we see a clearer benefit to our constraints. 1284 7.3 News Training on WSJ is also better than using NEWS: GENRE=2 SCOPE=0 SCOPE=1 SCOPE=2 INIT=0 40.2 38.8 38.7 1 43.4 44.0 43.8 @45 @15 @15?45 As with the blog, the best runs use the good initializer, discard unannotated sentences, enforce the loose constraint and follow up with domain adaptation ? GENRE=2; INIT=TRIM=ADAPT=1: Settings SCOPE=0 SCOPE=1 SCOPE=2 CONSTR=0 46.6 45.4 45.2 (sprawl) 2 46.1 44.9 44.9 (loose) 3 49.5 48.1 48.3 (strict) 4 37.7 36.8 37.6 @45 @15 @15?45 With all the extra training data, the best new score is just 49.5%. On the one hand, we are disappointed by the lack of dividends to orders of magnitude more data. On the other, we are comforted that the system arrives within 1% of its best result ? 50.4%, obtained with a manually cleaned up corpus ? now using an auto-generated data set. 7.4 Web The WEB-side story is more discouraging: GENRE=3 SCOPE=0 SCOPE=1 SCOPE=2INIT=0 38.3 35.1 35.2 1 42.8 43.6 43.4 @45 @15 @15?45 Our best run again uses a good initializer, keeps all sentences, still enforces the loose constraint and follows up with domain adaptation, but performs worse than all well-initialized WSJ baselines, scoring only 45.9% (trained at WEB15). We suspect that the web is just too messy for us. On top of the challenges of language identification and sentence-breaking, there is a lot of boiler-plate; furthermore, web text can be difficult for news-trained POS taggers. For example, note that the verb ?sign? is twice mistagged as a noun and that ?YouTube? is classified as a verb, in the top four POS sequences of web sentences:17 POS Sequence WEB Count Sample web sentence, chosen uniformly at random. 1 DT NNS VBN 82,858,487 All rights reserved. 2 NNP NNP NNP 65,889,181 Yuasa et al 3 NN IN TO VB RB 31,007,783 Sign in to YouTube now! 4 NN IN IN PRP$ JJ NN 31,007,471 Sign in with your Google Account! 17Further evidence: TnT tags the ubiquitous but ambiguous fragments ?click here? and ?print post? as noun phrases. 7.5 The State of the Art Our best model gains more than 5% over previous state-of-the-art accuracy across all sentences of WSJ?s Section 23, more than 8% on WSJ20 and rivals the oracle skyline (Spitkovsky et al, 2010a) on WSJ10; these gains generalize to Brown100, where it improves by nearly 10% (see Table 8). We take solace in the fact that our best models agree in using loose constraints. Of these, the models trained with less data perform better, with the best two using trimmed data sets, echoing that ?less is more? (Spitkovsky et al, 2010a), pace Halevy et al (2009). We note that orders of magnitude more data did not improve parsing performance further and suspect a different outcome from lexicalized models: The primary benefit of additional lower-quality data is in improved coverage. But with only 35 unique POS tags, data sparsity is hardly an issue. Extra examples of lexical items help little and hurt when they are mistagged. 8 Related Work The wealth of new annotations produced in many languages every day already fuels a number of NLP applications. Following their early and wide-spread use by search engines, in service of spam-fighting and retrieval, anchor text and link data enhanced a variety of traditional NLP techniques: cross-lingual information retrieval (Nie and Chen, 2002), translation (Lu et al, 2004), both named-entity recognition (Mihalcea and Csomai, 2007) and categorization (Watanabe et al, 2007), query segmentation (Tan and Peng, 2008), plus semantic relatedness and word-sense disambiguation (Gabrilovich and Markovitch, 2007; Yeh et al., 2009). Yet several, seemingly natural, candidate core NLP tasks ? tokenization, CJK segmentation, noun-phrase chunking, and (until now) parsing ? remained conspicuously uninvolved. Approaches related to ours arise in applications that combine parsing with named-entity recognition (NER). For example, constraining a parser to respect the boundaries of known entities is standard practice not only in joint modeling of (constituent) parsing and NER (Finkel and Manning, 2009), but also in higher-level NLP tasks, such as relation extraction(Mintz et al, 2009), that couple chunking with (dependency) parsing. Although restricted to proper noun phrases, dates, times and quantities, we suspect that constituents identified by trained (supervised) NER systems would also 1285 Model Incarnation WSJ10 WSJ20 WSJ? DMV Bilingual Log-Normals (tie-verb-noun) (Cohen and Smith, 2009) 62.0 48.0 42.2 Brown100 Leapfrog (Spitkovsky et al, 2010a) 57.1 48.7 45.0 43.6 default INIT=0,GENRE=0,SCOPE=0,CONSTR=0,TRIM=0,ADAPT=0 55.9 45.8 41.6 40.5 WSJ-best INIT=1,GENRE=0,SCOPE=2,CONSTR=0,TRIM=0,ADAPT=0 65.3 53.8 47.9 50.8 BLOGt-best INIT=1,GENRE=1,SCOPE=2,CONSTR=3,TRIM=1,ADAPT=1 69.3 56.8 50.4 53.3 NEWS-best INIT=1,GENRE=2,SCOPE=0,CONSTR=3,TRIM=1,ADAPT=1 67.3 56.2 50.1 51.6 WEB-best INIT=1,GENRE=3,SCOPE=1,CONSTR=3,TRIM=0,ADAPT=1 64.1 52.7 46.3 46.9 EVG Smoothed (skip-head), Lexicalized (Headden et al, 2009) 68.8 Table 8: Accuracies on Section 23 of WSJ{10, 20,? } and Brown100 for three recent state-of-the-art systems, our default run, and our best runs (judged by accuracy on WSJ45) for each of four training sets. be helpful in constraining grammar induction. Following Pereira and Schabes? (1992) success with partial annotations in training a model of (English) constituents generatively, their idea has been extended to discriminative estimation (Riezler et al, 2002) and also proved useful in modeling (Japanese) dependencies (Sassano, 2005). There was demand for partially bracketed corpora. Chen and Lee (1995) constructed one such corpus by learning to partition (English) POS sequences into chunks (Abney, 1991); Inui and Kotani (2001) used n-gram statistics to split (Japanese) clauses. We combine the two intuitions, using the web to build a partially parsed corpus. Our approach could be called lightly-supervised, since it does not require manual annotation of a single complete parse tree. In contrast, traditional semi-supervised methods rely on fully-annotated seed corpora.189 Conclusion We explored novel ways of training dependency parsing models, the best of which attains 50.4% accuracy on Section 23 (all sentences) of WSJ, beating all previous unsupervised state-of-the-art by more than 5%. Extra gains stem from guiding Viterbi training with web mark-up, the loose constraint consistently delivering best results. Our linguistic analysis of a blog reveals that web annotations can be converted into accurate parsing constraints (loose: 88%; sprawl: 95%; tear: 99%) that could be helpful to supervised methods, e.g., by boosting an initial parser via self-training (McClosky et al, 2006) on sentences with mark-up. Similar techniques may apply to standard wordprocessing annotations, such as font changes, and to certain (balanced) punctuation (Briscoe, 1994). We make our blog data set, overlaying mark-up and syntax, publicly available. Its annotations are 18A significant effort expended in building a tree-bank comes with the first batch of sentences (Druck et al, 2009). 75% noun phrases, 13% verb phrases, 7% simple declarative clauses and 2% prepositional phrases, with traces of other phrases, clauses and fragments. The type of mark-up, combined with POS tags, could make for valuable features in discriminative models of parsing (Ratnaparkhi, 1999). A logical next step would be to explore the connection between syntax and mark-up for genres other than a news-style blog and for languages other than English. We are excited by the possibilities, as unsupervised parsers are on the cusp of becoming useful in their own right ? recently, Davidov et al (2009) successfully applied Seginer?s (2007) fully unsupervised grammar inducer to the problems of pattern-acquisition and extraction of semantic data. If the strength of the connection between web mark-up and syntactic structure is universal across languages and genres, this fact could have broad implications for NLP, with applications extending well beyond parsing. Acknowledgments Partially funded by NSF award IIS-0811974 and by the Air Force Research Laboratory (AFRL), under prime contract no. FA8750-09-C-0181; first author supported by the Fannie & John Hertz Foundation Fellowship. We thank Angel X. Chang, Spence Green, Christopher D. Manning, Richard Socher, Mihai Surdeanu and the anonymous reviewers for many helpful suggestions, and we are especially grateful to Andy Golding, for pointing us to his sample Map-Reduce over the Google News crawl, and to Daniel Pipes, for allowing us to distribute the data set derived from his blog entries.'	b'1 In t roduct ion   A large number of current language processing sys-  tems use a part-of-speech tagger for pre-processing.  The tagger assigns a (unique or ambiguous) part-of-  speech tag to each token in the input and passes its  output o the next processing level, usually a parser.  Furthermore, there is a large interest in part-of-  speech tagging for corpus annotation projects, who  create valuable linguistic resources by a combination  of automatic processing and human correction.  For both applications, a tagger with the highest  possible accuracy is required. The debate about  which paradigm solves the part-of-speech tagging  problem best is not finished. Recent comparisons  of approaches that can be trained on corpora (van  Halteren et al, 1998; Volk and Schneider, 1998) have  shown that in most cases statistical aproaches (Cut-  ting et al, 1992; Schmid, 1995; Ratnaparkhi, 1996)  yield better results than finite-state, rule-based, or  memory-based taggers (Brill, 1993; Daelemans etal.,  1996). They are only surpassed by combinations of different systems, forming a "voting tagger".  Among the statistical approaches, the Maximum  Entropy framework has a very strong position. Nev-  ertheless, a recent independent comparison of 7 tag-  gets (Zavrel and Daelemans, 1999) has shown that  another approach even works better: Markov mod-  els combined with a good smoothing technique and  with handling of unknown words. This tagger, TnT,  not only yielded the highest accuracy, it also was the  fastest both in training and tagging.  The tagger comparison was organized as a "black-  box test": set the same task to every tagger and  compare the outcomes. This paper describes the  models and techniques used by TnT together with  the implementation.  The reader will be surprised how simple the under-  lying model is. The result of the tagger comparison  seems to support he maxime "the simplest is the  best". However, in this paper we clarify a number  of details that are omitted in major previous pub-  lications concerning tagging with Markov models.  As two examples, (Rabiner, 1989) and (Charniak  et al, 1993) give good overviews of the techniques  and equations used for Markov models and part-of-  speech tagging, but they are not very explicit in the  details that are needed for their application. We ar-  gue that it is not only the choice of the general model  that determines the result of the tagger but also the  various "small" decisions on alternatives.  The aim of this paper is to give a detailed ac-  count of the techniques used in TnT. Additionally,  we present results of the tagger on the NEGRA cor-  pus (Brants et al, 1999) and the Penn Treebank  (Marcus et al, 1993). The Penn Treebank results  reported here for the Markov model approach are  at least equivalent to those reported for the Maxi-  mum Entropyapproach in (Ratnaparkhi, 1996). For  a comparison to other taggers, the reader is referred  to (Zavrel and Daelemans, 1999).  2 Arch i tec ture   2.1 The Under ly ing Model   TnT uses second order Markov models for part-of-  speech tagging. The states of the model represent  tags, outputs represent the words. Transition prob-  abilities depend on the states, thus pairs of tags.  Output probabilities only depend on the most re-  cent category. To be explicit, we calculate  argmax P(tilti-1, ti-2)P(wilti P(tr+l ItT)  ti...iT  (i)  for a given sequence of words w I . . .  W T of length T.  tl... tT are elements of the tagset, the additional  224 tags t - l ,  to, and tT+l are beginning-of-sequence  and end-of-sequence markers. Using these additional  tags, even if they stem from rudimentary process-  ing of punctuation marks, slightly improves tagging  results. This is different from formulas presented  in other publications, which just stop with a "loose  end" at the last word. If sentence boundaries are  not marked in the input, TnT adds these tags if it  encounters one of \\[.!?;\\] as a token.  Transition and output probabilities are estimated  from a tagged corpus. As a first step, we use the  maximum likelihood probabilities /5 which are de-  rived from the relative frequencies:  Unigrams: /5(t3) = f(t3) (2)  N  f(t2, t3) (3)  Bigrams: P(t31t~)= f(t2)  f(ta, t2, t3)  Trigrams: /5(t3ltx,t~) - f ( t l , t2)  (4)  Lexical: /5(w3 It3) - /(w3, t3) (5)  f(t3)  for all tl, t2, t3 in the tagset and w3 in the lexi-  con. N is the total number of tokens in the training  corpus. We define a maximum likelihood probabil-  ity to be zero if the corresponding nominators and  denominators are zero. As a second step, contex-  tual frequencies are smoothed and lexical frequences  are completed by handling words that are not in the  lexicon (see below).  2.2 Smooth ing   Trigram probabilities generated from a corpus usu-  ally cannot directly be used because of the sparse-  data problem. This means that there are not enough  instances for each trigram to reliably estimate the  probability. Furthermore, setting a probability to  zero because the corresponding trigram never oc-  cured in the corpus has an undesired effect. It causes  the probability of a complete sequence to be set to  zero if its use is necessary for a new text sequence,  thus makes it impossible to rank different sequences  containing a zero probability.  The smoothing paradigm that delivers the best  results in TnT is linear interpolation of unigrams,  bigrams, and trigrams. Therefore, we estimate a  trigram probability as follows:  P(t3ltl ,  t2) = AlP(t3) + Ag_/5(t31t2) + A3/5(t3\\[t1, t2)  (6) /5 are maximum likelihood estimates of the proba-  bilities, and A1 + A2 + A3 = 1, so P again represent  probability distributions.  We use the context-independent variant of linear  interpolation, i.e., the values of the As do not depend  on the particular trigram. Contrary to intuition,  this yields better esults than the context-dependent  variant. Due to sparse-data problems, one cannot es-  timate a different set of As for each trigram. There-  fore, it is common practice to group trigrams by fre-  quency and estimate tied sets of As. However, we  are not aware of any publication that has investi-  gated frequency groupings for linear interpolation i part-of-speech tagging. All groupings that we have  tested yielded at most equivalent results to context-  independent linear interpolation. Some groupings  even yielded worse results. The tested groupings  included a) one set of As for each frequency value  and b) two classes (low and high frequency) on the  two ends of the scale, as well as several groupings  in between and several settings for partitioning the  classes.  The values of Ax, A2, and A3 are estimated by  deleted interpolation. This technique successively  removes each trigram from the training corpus and  estimates best values for the As from all other n-  grams in the corpus. Given the frequency counts  for uni-, bi-, and trigrams, the weights can be very  efficiently determined with a processing time linear  in the number of different rigrams. The algorithm  is given in figure 1. Note that subtracting 1 means  taking unseen data into account. Without this sub-  traction the model would overfit the training data  and would generally ield worse results.  2.3 Handling of Unknown Words  Currently, the method of handling unknown words  that seems to work best for inflected languages is  a suffix analysis as proposed in (Samuelsson, 1993).  Tag probabilities are set according to the word\'s end-  ing. The suffix is a strong predictor for word classes,  e.g., words in the Wall Street Journal part of the  Penn Treebank ending in able are adjectives (JJ) in  98% of the cases (e.g. fashionable, variable), the rest  of 2% are nouns (e.g. cable, variable).  The probability distribution for a particular suf-  fix is generated from all words in the training set  that share the same suffix of some predefined max-  imum length. The term suffix as used here means  "final sequence of characters of a word" which is not  necessarily a linguistically meaningful suffix.  Probabilities are smoothed by successive abstrac-  tion. This calculates the probability of a tag t  given the last m letters li of an n letter word:  P(t l ln - r ,+l , . . . ln) .  The sequence of increasingly  more general contexts omits more and more char-  acters ofthe suffix, such that P(t l ln_m+2,. . . , ln) ,   P(t \\ [ ln-m+3,. . . , l~) ,  . . . ,  P(t) are used for smooth-  ing. The recursiou formula is  P(t l l ,_ i+l,  . . . ln)  = P( t l ln - i+ l , . . .  In) + O iP ( t l l~- , . . . ,  ln) (7)  1 +0~  225 set  )%1 ---- )%2 = )%3 = 0  fo reach  t r ig ram tl,t2,t3 with  f(t l ,t2,t3) > 0  depending on the maximum of the fo l low ing  three  va lues :   f(tl ,t2,ts)-- 1 case f (h,t2)- I  " increment )%3 by f ( t l , t2 , t3)   f(t2,t3)-I case f(t2)- I  " increment )%2 by f ( t l , t2 , t s )   f ( t3 ) - - i   case N-1  " increment )%1 by f(t l ,t2,t3)  end  end  normalize )%1, )%2, )%3  Figure 1: Algorithm for calculting the weights for context-independent li ear interpolation )%1, )%2, )%3 when  the n-gram frequencies are known. N is the size of the corpus? If the denominator in one of the expressions  is 0, we define the result of that expression to be 0.  for i = m. . .  0, using the maximum likelihood esti-  mates/5  from frequencies in the lexicon, weights Oi  and the initialization  P(t) =/5(t) .  (8)  The maximum likelihood estimate for a suffix of  length i is derived from corpus frequencies by  P(t i /~- i+l ,  .. l~) = f (t ,  1~-~+1,... l~) ? (9 )   For the Markov model, we need the inverse condi-  tional probabilities P ( / , - i+ l , . . .  lnlt) which are ob-  tained by Bayesian inversion.  A theoretical motivated argumentation uses the  standard eviation of the maximum likelihood prob-  abilities for the weights 0i (Samuelsson, 1993)?  This leaves room for interpretation.  1) One has to identify a good value for m, the  longest suffix used? The approach taken for TnT is  the following: m depends on the word in question.  We use the longest suffix that we can find in the  training set (i.e., for which the frequency is greater  than or equal to 1), but at most 10 characters. This  is an empirically determined choice?  2) We use a context-independent approach for 0i,  as we did for the contextual weights )%i. It turned  out to be a good choice to set al 0i to the standard  deviation of the unconditioned maximum likelihood  probabilities of the tags in the training corpus, i.e.,  weset  1  $  Oi = ~--~(/5(tj) - 15)2 (10)  s 1 j= l   for all i = 0 . . .  m - 1, using a tagset of s tags and  the average  $  /5 = 1 ~/5(t j )  (11)  8  j----I  This usually yields values in the range 0.03 .. .  0.10.  3) We use different estimates for uppercase and  lowercase words, i.e., we maintain two different suffix  tries depending on the capitalization of the word.  This information improves the tagging results?  4) Another freedom concerns the choice of the  words in the lexicon that should be used for suf-  fix handling. Should we use all words, or are some  of them better suited than others? Accepting that  unknown words are most probably infrequent, one  can argue that using suffixes of infrequent words in  the lexicon is a better approximation for unknown  words than using suffixes of frequent words. There-  fore, we restrict the procedure of suffix handling to  words with a frequency smaller than or equal to some  threshold value. Empirically, 10 turned out to be a  good choice for this threshold.  2.4 Cap i ta l i za t ion   Additional information that turned out to be use-  ful for the disambiguation process for several cor-  pora and tagsets is capitalization information. Tags  are usually not informative about capitalization, but  probability distributions of tags around capitalized  words are different from those not capitalized. The  effect is larger for English, which only capitalizes  proper names, and smaller for German, which capi-  talizes all nouns.  We use flags ci that are true if wi is a capitalized  word and false otherwise? These flags are added to  the contextual probability distributions. Instead of  P(tsItl,t2) (12)  we use  P(t3, c3\\[tl , cl, t2, c2) (13)  and equations (3) to (5) are updated accordingly.  This is equivalent o doubling the size of the tagset  and using different ags depending on capitalization.  226 2.5 Beam Search  The processing time of the Viterbi algorithm (Ra-  biner, 1989) can be reduced by introducing a beam  search. Each state that receives a 5 value smaller  than the largest 5 divided by some threshold value  is excluded from further processing. While the  Viterbi algorithm is guaranteed to find the sequence  of states with the highest probability, this is no  longer true when beam search is added. Neverthe-  less, for practical purposes and the right choice of  0, there is virtually no difference between the algo-  rithm with and without a beam. Empirically, a value  of 0 = 1000 turned out to approximately double the  speed of the tagger without affecting the accuracy.  The tagger currently tags between 30~000 and  60,000 tokensper second (including file I/O) on a  Pentium 500 running Linux. The speed mainly de-  pends on the percentage of unknown words and on  the average ambiguity rate.  3 Eva luat ion   We evaluate the tagger\'s performance under several  aspects. First of all, we determine the tagging ac-  curacy averaged over ten iterations. The overall ac-  curacy, as well as separate accuracies for known and  unknown words are measured.  Second, learning curves are presented, that indi-  cate the performance when using training corpora of  different sizes, starting with as few as 1,000 tokens  and ranging to the size of the entire corpus (minus  the test set).  An important characteristic of statistical taggers  is that they not only assign tags to words but also  probabilities in order to rank different assignments.  We distinguish reliable from unreliable assignments  by the quotient of the best and second best assign-  ments 1. All assignments for which this quotient is  larger than some threshold are regarded as reliable,  the others as unreliable. As we will see below, accu-  racies for reliable assignments are much higher.  The tests are performed on partitions of the cor-  pora that use 90% as training set and 10% as test  set, so that the test data is guaranteed to be unseen  during training. Each result is obtained by repeat-  ing the experiment 10 times with different partitions  and averaging the single outcomes.  In all experiments, contiguous test sets are used.  The alternative is a round-robin procedure that puts  every 10th sentence into the test set. We argue that  contiguous test sets yield more realistic results be-  cause completely unseen articles are tagged. Using  the round-robin procedure, parts of an article are al-  ready seen, which significantly reduces the percent-  age of unknown words. Therefore, we expect even  1 By definition, this quotient is co if there is only one pos-  sible tag for a given word.  higher results when testing on every 10th sentence  instead of a contiguous et of 10%.  In the following, accuracy denotes the number of  correctly assigned tags divided by the number of to-  kens in the corpus processed. The tagger is allowed  to assign exactly one tag to each token.  We distinguish the overall accuracy, taking into  account all tokens in the test corpus, and separate  accuracies for known and unknown tokens. The lat-  ter are interesting, since usually unknown tokens are  much more difficult to process than known tokens,  for which a list of valid tags can be found in the  lexicon.  3.1 Tagging the  NEGRA corpus   The German NEGRA corpus consists of 20,000 sen-  tences (355,000 tokens) of newspaper texts (Frank-  furter Rundschau) that are annotated with parts-of-  speech and predicate-argument structures (Skut et al., 1997). It was developed at the Saarland Univer-  sity in Saarbrficken 2. Part of it was tagged at the  IMS Stuttgart. This evaluation only uses the part-  of-speech annotation and ignores structural annota-  tions.  Tagging accuracies for the NEGRA corpus are  shown in table 2.  Figure 3 shows the learning curve of the tagger,  i.e., the accuracy depending on the amount of train-  ing data. Training length is the nmnber of tokens  used for training. Each training length was tested  ten times, training and test sets were randomly cho-  sen and disjoint, results were averaged. The training  length is given on a logarithmic scale.  It is remarkable that tagging accuracy for known  words is very high even for very small training cot-  pora. This means that we have a good chance of  getting the right tag if a word is seen at least once  during training. Average percentages of unknown  tokens are shown in the bottom line of each diagram.  We exploit the fact that the tagger not only de-  termines tags, but also assigns probabilities. If there  is an alternative that has a probability "close to"  that of the best assignment, his alternative can be  viewed as almost equally well suited. The notion of  "close to" is expressed by the distance of probabil-  ities, and this in turn is expressed by the quotient  of probabilities. So, the distance of the probabili-  ties of a best tag tbest and an alternative tag tart  is expressed by P(tbest)/p(tau), which is some value  greater or equal to 1 since the best tag assignment  has the highest probability.  Figure 4 shows the accuracy when separating as-  signments with quotients larger and smaller than  the threshold (hence reliable and unreliable assign-  ments). As expected, we find that accuracies for  2For availability, please check  h~tp ://w~. col i. uni-sb, de/s fb378/negra-corpus  777 227 Table 2: Part-of-speech tagging accuracy for the NEGRA corpus, averaged over 10 test runs, training and  test set are disjoint. The table shows the percentage of unknown tokens, separate accuracies and standard  deviations for known and unknown tokens, as well as the overall accuracy.  percentage  unknowns  NEGRA corpus 11.9%  known  ace .   97.7% 0.23  unknown  acc. (x  89.0% 0.72  overall  aCE. o"  96.7% 0.29  NEGRA Corpus :  POS Learn ing  Curve   100  9O  80  70 /S   6O  50 , i  1 2 5  50.8 46.4 41.4  I i i I I I i  10 20 50 100 200 320 500  36.0 30.7 23.0 18.3 14.3 11.9 11).3  Overall  min =78.1%  max=96.7%  e Known  rain =95.7%  max=97.7%  - - -a- - -  Unknown  rain =61.2% max=89.0%  I  1000 x 1000 Training Length  St  avg. percentage unknown  Figure 3: Learning curve for tagging the NEGRA corpus. The training sets of variable sizes as well as test  sets of 30,000 tokens were randomly chosen. Training and test sets were disjoint, the procedure was repeated  10 times and results were averaged. Percentages of unknowns for 500k and 1000k training are determined  from an untagged extension.  NEGRA Corpus: Accuracy of  re l iab le  assignments  100  99  98  97  A  f " / :. Reliable rain =96.7% max=99.4%  96 I i i i i i i i i i i  2 5 10 20 50 100 500 2000 10000 threshold 0  100 97.9 95.1 92.7 90.3 86.8 84.1 81.0 76.1 71.9 68.3 64.1 62.0 % cases reliable  - 53.5 62.9 69.6 74.5 79.8 82.7 85.2 88.0 89.6 90.8 91.8 92.2 acc. of complement  Figure 4: Tagging accuracy for the NEGRA corpus when separating reliable and unreliable assignments. The  curve shows accuracies for reliable assignments. The numbers at the bottom line indicate the percentage of  reliable assignments and the accuracy of the complement set (i.e., unreliable assignments).  228 Table 5: Part-of-speech tagging accuracy for the Penn Treebank. The table shows the percentage of unknown  tokens, separate accuracies and standard deviations for known and unknown tokens, as well as the overall  accuracy.  I percentage known  unknowns acc. a  Penn Treebank 2.9% 97.0% 0.15  unknown  aCC. O"  85.5% 0.69  overall  aCE. O"  96.7% 0.15  100  9O  80  70 <  60  Penn Treebank: POS Learn ing  Curve   /  50 I I ~ i I I I i I  1 2 5 10 20 50 100 200 500  50.3 42.8 33.4 26.8 20.2 13.2 9.8 7.0 4.4  Overall  rain =78.6%  max=96.7%  Known  rain =95.2%  max=97.0%  Unknown  min =62.2%  max=85.5%  I  1000 ? 1000 Training Length  2.9 avg. percentage unknown  Figure 6: Learning curve for tagging the Penn Treebank. The training sets of variable sizes as well as test sets  of 100,000 tokens were randomly chosen. Training and test sets were disjoint, the procedure was repeated  10 times and results were averaged.  Penn Treebank: Accuracy of reliable assignments  100  99  98  97  Overall  rain =96.6%  max=99.4%  96 i i i i t i i i i i i i  2 5 10 20 50 100 500 2000 10000 threshold 0  100 97.7 94.6 92.2 89.8 86.3 83.5 80.4 76.6 73.8 71.0 67.2 64.5 % cases reliable  - 53.5 62.8 68.9 73.9 79.3 82.6 85.2 87.5 88.8 89.8 91.0 91.6 acc. of complement  Figure 7: Tagging accuracy for the Penn Treebank when separating reliable and unreliable assignments. The  curve shows accuraciesfor reliable assignments. The numbers at the bottom line indicate the percentage of  reliable assignments and the accuracy of the complement set.  229  reliable assignments are much higher than for unre-  liable assignments. This distinction is, e.g., useful  for annotation projects during the cleaning process,  or during pre-processing, sothe tagger can emit mul-  tiple tags if the best tag is classified as unreliable.  3.2 Tagging the Penn Treebank   We use the Wall Street Journal as contained in the  Penn Treebank for our experiments. The annotation  consists of four parts: 1) a context-free structure  augmented with traces to mark movement and dis-  continuous constituents, 2) phrasal categories that  are annotated as node labels, 3) a small set of gram-  matical functions that are annotated as extensions to  the node labels, and 4) part-of-speech tags (Marcus  et al, 1993). This evaluation only uses the part-of-  speech annotation.  The Wall Street Journal part of the Penn Tree-  bank consists of approx. 50,000 sentences (1.2 mil-  lion tokens).  Tagging accuracies for the Penn Treebank are  shown in table 5. Figure 6 shows the learning curve  of the tagger, i.e., the accuracy depending on the  amount of training data. Training length is the num-  ber of tokens used for training. Each training length  was tested ten times. Training and test sets were  disjoint, results are averaged. The training length is  given on a logarithmic scale. As for the NEGRA cor-  pus, tagging accuracy is very high for known tokens  even with small amounts of training data.  We exploit the fact that the tagger not only de-  termines tags, but also assigns probabilities. Figure  7 shows the accuracy when separating assignments  with quotients larger and smaller than the threshold  (hence reliable and unreliable assignments). Again,  we find that accuracies for reliable assignments are  much higher than for unreliable assignments.  3.3 Summary of Part-of-Speech Tagging  Results  Average part-of-speech tagging accuracy is between  96% and 97%, depending on language and tagset,  which is at least on a par with state-of-the-art re-  sults found in the literature, possibly better. For  the Penn Treebank, (Ratnaparkhi, 1996) reports an  accuracy of 96.6% using the Maximum Entropy ap-  proach, our much simpler and therefore faster HMM  approach delivers 96.7%. This comparison eeds to  be re-examined, since we use a ten-fold crossvalida-  tion and averaging of results while Ratnaparkhi only  makes one test run.  The accuracy for known tokens is significantly  higher than for unknown tokens. For the German  newspaper data, results are 8.7% better when the  word was seen before and therefore is in the lexicon,  than when it was not seen before (97.7% vs. 89.0%).  Accuracy for known tokens is high even with very  small amounts of training data. As few as 1000 to- kens are sufficient o achieve 95%-96% accuracy for  them. It is important for the tagger to have seen a  word at least once during training.  Stochastic taggers assign probabilities to tags. We  exploit the probabilities to determine reliability of  assignments. For a subset hat is determined uring  processing by the tagger we achieve accuracy rates  of over 99%. The accuracy of the complement set is  much lower. This information can, e.g., be exploited  in an annotation project to give an additional treat-  ment to the unreliable assignments, or to pass se-  lected ambiguities to a subsequent processing step. 4 Conc lus ion   We have shown that a tagger based on Markov mod-  els yields state-of-the-art results, despite contrary  claims found in the literature. For example, the  Markov model tagger used in the comparison of (van  Halteren et al, 1998) yielded worse results than all  other taggers. In our opinion, a reason for the wrong  claim is that the basic algorithms leave several deci-  sions to the implementor. The rather large amount  of freedom was not handled in detail in previous pub-  lications: handling of start- and end-of-sequence, the  exact smoothing technique, how to determine the  weights for context probabilities, details on handling  unknown words, and how to determine the weights  for unknown words. Note that the decisions we made  yield good results for both the German and the En-  glish Corpus. They do so for several other corpora  as well. The architecture remains applicable to a  large variety of languages.  According to current tagger comparisons (van  Halteren et al, 1998; Zavrel and Daelemans, 1999),  and according to a comparsion of the results pre-  sented here with those in (Ratnaparkhi, 1996), the  Maximum Entropy framework seems to be the only  other approach yielding comparable results to the  one presented here. It is a very interesting future  research topic to determine the advantages of either  of these approaches, to find the reason for their high  accuracies, and to find a good combination of both.  TnT is freely available to universities and re-  lated organizations for research purposes (see  http ://www. coli. uni-sb, de/-thorsten/tnt). Acknowledgements   Many thanks go to Hans Uszkoreit for his sup-  port during the development of TnT. Most of the  work on TnT was carried out while the author  received a grant of the Deutsche Forschungsge-  meinschaft in the Graduiertenkolleg Kognitionswis-  senschaft Saarbriicken. Large annotated corpora re  the pre-requisite for developing and testing part-of-  speech taggers, and they enable the generation of  high-quality language models. Therefore, I would  230  like to thank all the people who took the effort  to annotate the Penn Treebank, the Susanne Cor-  pus, the Stuttgarter Referenzkorpus, the NEGRA  Corpus, the Verbmobil Corpora, and several others.  And, last but not least, I would like to thank the  users of TnT who provided me with bug reports and  valuable suggestions for improvements. '
0.0	P10-1130	P10-2025	b'1 Introduction Unsupervised learning of hierarchical syntactic structure from free-form natural language text is a hard problem whose eventual solution promises to benefit applications ranging from question answering to speech recognition and machine translation. A restricted version of this problem that targets dependencies and assumes partial annotation ? sentence boundaries and part-of-speech (POS) tagging ? has received much attention. Klein and Manning (2004) were the first to beat a simple parsing heuristic, the right-branching baseline; today?s state-of-the-art systems (Headden et al, 2009; Cohen and Smith, 2009; Spitkovsky et al, 2010a) are rooted in their Dependency Model with Valence (DMV), still trained using variants of EM. Pereira and Schabes (1992) outlined three major problems with classic EM, applied to a related problem, constituent parsing. They extended classic inside-outside re-estimation (Baker, 1979) to respect any bracketing constraints included with a training corpus. This conditioning on partial parses addressed all three problems, leading to: (i) linguistically reasonable constituent boundaries and induced grammars more likely to agree with qualitative judgments of sentence structure, which is underdetermined by unannotated text; (ii) fewer iterations needed to reach a good grammar, countering convergence properties that sharply deteriorate with the number of non-terminal symbols, due to a proliferation of local maxima; and (iii) better (in the best case, linear) time complexity per iteration, versus running time that is ordinarily cubic in both sentence length and the total number of non-terminals, rendering sufficiently large grammars computationally impractical. Their algorithm sometimes found good solutions from bracketed corpora but not from raw text, supporting the view that purely unsupervised, selforganizing inference methods can miss the trees for the forest of distributional regularities. This was a promising break-through, but the problem of whence to get partial bracketings was left open. We suggest mining partial bracketings from a cheap and abundant natural language resource: the hyper-text mark-up that annotates web-pages. For example, consider that anchor text can match linguistic constituents, such as verb phrases, exactly: ..., whereas McCain is secure on the topic, Obama <a>[VP worries about winning the pro-Israel vote]</a>. To validate this idea, we created a new data set, novel in combining a real blog?s raw HTML with tree-bank-like constituent structure parses, gener1278 ated automatically. Our linguistic analysis of the most prevalent tags (anchors, bold, italics and underlines) over its 1M+ words reveals a strong connection between syntax and mark-up (all of our examples draw from this corpus), inspiring several simple techniques for automatically deriving parsing constraints. Experiments with both hard and more flexible constraints, as well as with different styles and quantities of annotated training data ? the blog, web news and the web itself, confirm that mark-up-induced constraints consistently improve (otherwise unsupervised) dependency parsing. 2 Intuition and Motivating Examples It is natural to expect hidden structure to seep through when a person annotates a sentence. As it happens, a non-trivial fraction of the world?s population routinely annotates text diligently, if only partially and informally.1 They inject hyper-links, vary font sizes, and toggle colors and styles, using mark-up technologies such as HTML and XML. As noted, web annotations can be indicative ofphrase boundaries, e.g., in a complicated sentence: In 1998, however, as I <a>[VP established in <i>[NP The New Republic]</i>]</a> and Bill Clinton just <a>[VP confirmed in his memoirs]</a>, Netanyahu changed his mind and ... In doing so, mark-up sometimes offers useful cues even for low-level tokenization decisions: [NP [NP Libyan ruler] <a>[NP Mu?ammar al-Qaddafi]</a>] referred to ... (NP (ADJP (NP (JJ Libyan) (NN ruler)) (JJ Mu)) (?? ?) (NN ammar) (NNS al-Qaddafi)) Above, a backward quote in an Arabic name confuses the Stanford parser.2 Yet mark-up lines up with the broken noun phrase, signals cohesion, and moreover sheds light on the internal structure of a compound. As Vadas and Curran (2007) point out, such details are frequently omitted even from manually compiled tree-banks that err on the side of flat annotations of base-NPs. Admittedly, not all boundaries between HTML tags and syntactic constituents match up nicely: ..., but [S [NP the <a><i>Toronto Star</i>][VP reports [NP this][PP in the softest possible way]</a>,[S stating only that ...]]] Combining parsing with mark-up may not be straight-forward, but there is hope: even above, 1Even when (American) grammar schools lived up to their name, they only taught dependencies. This was back in the days before constituent grammars were invented. 2http://nlp.stanford.edu:8080/parser/ one of each nested tag?s boundaries aligns; and Toronto Star?s neglected determiner could be forgiven, certainly within a dependency formulation. 3 A High-Level Outline of Our Approach Our idea is to implement the DMV (Klein and Manning, 2004) ? a standard unsupervised grammar inducer. But instead of learning the unannotated test set, we train with text that contains web mark-up, using various ways of converting HTML into parsing constraints. We still test on WSJ (Marcus et al, 1993), in the standard way, and also check generalization against a hidden data set ? the Brown corpus (Francis and Kucera, 1979). Our parsing constraints come from a blog ? a new corpus we created, the web and news (see Table 1 for corpora?s sentence and token counts). To facilitate future work, we make the final models and our manually-constructed blog data publicly available.3 Although we are unable to share larger-scale resources, our main results should be reproducible, as both linguistic analysis and our best model rely exclusively on the blog. Corpus Sentences POS Tokens WSJ? 49,208 1,028,347 Section 23 2,353 48,201 WSJ45 48,418 986,830 WSJ15 15,922 163,715 Brown100 24,208 391,796 BLOGp 57,809 1,136,659 BLOGt45 56,191 1,048,404 BLOGt15 23,214 212,872 NEWS45 2,263,563,078 32,119,123,561 NEWS15 1,433,779,438 11,786,164,503 WEB45 8,903,458,234 87,269,385,640 WEB15 7,488,669,239 55,014,582,024 Table 1: Sizes of corpora derived from WSJ and Brown, as well as those we collected from the web. 4 Data Sets for Evaluation and Training The appeal of unsupervised parsing lies in its ability to learn from surface text alone; but (intrinsic) evaluation still requires parsed sentences. Following Klein and Manning (2004), we begin with reference constituent parses and compare against deterministically derived dependencies: after pruning out all empty subtrees, punctuation and terminals (tagged # and $) not pronounced where they appear, we drop all sentences with more than a prescribed number of tokens remaining and useautomatic ?head-percolation? rules (Collins, 1999) to convert the rest, as is standard practice. 3http://cs.stanford.edu/?valentin/ 1279 Length Marked POS Bracketings Length Marked POS Bracketings Cutoff Sentences Tokens All Multi-Token Cutoff Sentences Tokens All Multi-Token 0 6,047 1,136,659 7,731 6,015 8 485 14,528 710 684 1 of 57,809 149,483 7,731 6,015 9 333 10,484 499 479 2 4,934 124,527 6,482 6,015 10 245 7,887 365 352 3 3,295 85,423 4,476 4,212 15 42 1,519 65 63 4 2,103 56,390 2,952 2,789 20 13 466 20 20 5 1,402 38,265 1,988 1,874 25 6 235 10 10 6 960 27,285 1,365 1,302 30 3 136 6 6 7 692 19,894 992 952 40 0 0 0 0 Table 2: Counts of sentences, tokens and (unique) bracketings for BLOGp, restricted to only those sentences having at least one bracketing no shorter than the length cutoff (but shorter than the sentence). Our primary reference sets are derived from the Penn English Treebank?s Wall Street Journal portion (Marcus et al, 1993): WSJ45 (sentences with fewer than 46 tokens) and Section 23 of WSJ? (all sentence lengths). We also evaluate on Brown100, similarly derived from the parsed portion of the Brown corpus (Francis and Kucera, 1979). While we use WSJ45 and WSJ15 to train baseline models, the bulk of our experiments is with web data. 4.1 A News-Style Blog: Daniel Pipes Since there was no corpus overlaying syntactic structure with mark-up, we began constructing a new one by downloading articles4 from a newsstyle blog. Although limited to a single genre ? political opinion, danielpipes.org is clean, consistently formatted, carefully edited and larger than WSJ (see Table 1). Spanning decades, Pipes? editorials are mostly in-domain for POS taggers and tree-bank-trained parsers; his recent (internetera) entries are thoroughly cross-referenced, conveniently providing just the mark-up we hoped to study via uncluttered (printer-friendly) HTML.5 After extracting moderately clean text and mark-up locations, we used MxTerminator (Reynar and Ratnaparkhi, 1997) to detect sentence boundaries. This initial automated pass begot multiple rounds of various semi-automated clean-ups that involved fixing sentence breaking, modifying parser-unfriendly tokens, converting HTML entities and non-ASCII text, correcting typos, and so on. After throwing away annotations of fractional words (e.g., <i>basmachi</i>s) and tokens (e.g., <i>Sesame Street</i>-like), we broke up all markup that crossed sentence boundaries (i.e., loosely speaking, replaced constructs like <u>...][S...</u> with <u>...</u> ][S <u>...</u>) and discarded any 4http://danielpipes.org/art/year/all 5http://danielpipes.org/article print.php? id=. . . tags left covering entire sentences. We finalized two versions of the data: BLOGt, tagged with the Stanford tagger (Toutanova and Manning, 2000; Toutanova et al, 2003),6 and BLOGp, parsed with Charniak?s parser (Charniak, 2001; Charniak and Johnson, 2005).7 The reason for this dichotomy was to use state-of-the-art parses to analyze the relationship between syntax and mark-up, yet to prevent jointly tagged (and non-standard AUX[G]) POS sequences from interfering with our (otherwise unsupervised) training.8 4.2 Scaled up Quantity: The (English) Web We built a large (see Table 1) but messy data set, WEB ? English-looking web-pages, pre-crawled by a search engine. To avoid machine-generated spam, we excluded low quality sites flagged by the indexing system. We kept onlysentence-like runs of words (satisfying punctuation and capitalization constraints), POS-tagged with TnT (Brants, 2000). 4.3 Scaled up Quality: (English) Web News In an effort to trade quantity for quality, we constructed a smaller, potentially cleaner data set, NEWS. We reckoned editorialized content would lead to fewer extracted non-sentences. Perhaps surprisingly, NEWS is less than an order of magnitude smaller than WEB (see Table 1); in part, this is due to less aggressive filtering ? we trust sites approved by the human editors at Google News.9 In all other respects, our pre-processing of NEWS pages was identical to our handling of WEB data. 6http://nlp.stanford.edu/software/ stanford-postagger-2008-09-28.tar.gz 7ftp://ftp.cs.brown.edu/pub/nlparser/ parser05Aug16.tar.gz 8However, since many taggers are themselves trained on manually parsed corpora, such as WSJ, no parser that relies on external POS tags could be considered truly unsupervised; for a fully unsupervised example, see Seginer?s (2007) CCL parser, available at http://www.seggu.net/ccl/ 9http://news.google.com/ 1280 5 Linguistic Analysis of Mark-Up Is there a connection between mark-up and syntactic structure? Previous work (Barr et al, 2008) has only examined search engine queries, showing that they consist predominantly of short noun phrases. If web mark-up shared a similar characteristic, it might not provide sufficiently disambiguating cues to syntactic structure: HTML tags could be too short (e.g., singletons like ?click <a>here</a>?) or otherwise unhelpful in resolving truly difficult ambiguities (such as PPattachment). We began simply by counting various basic events in BLOGp. Count POS Sequence Frac Sum 1 1,242 NNP NNP 16.1% 2 643 NNP 8.3 24.4 3 419 NNP NNP NNP 5.4 29.8 4 414 NN 5.4 35.2 5 201 JJ NN 2.6 37.8 6 138 DT NNP NNP 1.8 39.5 7 138 NNS 1.8 41.3 8 112 JJ 1.5 42.8 9 102 VBD 1.3 44.1 10 92 DT NNP NNP NNP 1.2 45.3 11 85 JJ NNS 1.1 46.4 12 79 NNP NN 1.0 47.4 13 76 NN NN 1.0 48.4 14 61 VBN 0.8 49.2 15 60 NNP NNP NNP NNP 0.8 50.0 BLOGp +3,869 more with Count ? 49 50.0% Table 3: Top 50% of marked POS tag sequences. Count Non-Terminal Frac Sum 1 5,759 NP 74.5% 2 997 VP 12.9 87.4 3 524 S 6.8 94.2 4 120 PP 1.6 95.7 5 72 ADJP 0.9 96.7 6 61 FRAG 0.8 97.4 7 41 ADVP 0.5 98.0 8 39 SBAR 0.5 98.5 9 19 PRN 0.2 98.7 10 18 NX 0.2 99.0 BLOGp +81 more with Count ? 16 1.0% Table 4: Top 99% of dominating non-terminals. 5.1 Surface Text Statistics Out of 57,809 sentences, 6,047 (10.5%) are annotated (see Table 2); and 4,934 (8.5%) have multitoken bracketings. We do not distinguish HTML tags and track only unique bracketing end-points within a sentence. Of these, 6,015 are multi-token ? an average per-sentence yield of 10.4%.10 10A non-trivial fraction of our corpus is older (pre-internet) unannotated articles, so this estimate may be conservative. As expected, many of the annotated words are nouns, but there are adjectives, verbs and other parts of speech too (see Table 3). Mark-up is short, typically under five words, yet (by far) the most frequentlymarked sequence of POS tags is a pair. 5.2 Common Syntactic Subtrees For three-quarters of all mark-up, the lowest dominating non-terminal is a noun phrase (see Table 4); there are also non-trace quantities of verb phrases (12.9%) and other phrases, clauses and fragments. Of the top fifteen ? 35.2% of all ? annotated productions, only one is not a noun phrase (see Table 5, left). Four of the fifteen lowest dominating non-terminals do not match the entire bracketing ? all four miss the leading determiner, as we saw earlier. In such cases, we recursively split internal nodes until the bracketing aligned, as follows: [S [NP the <a>Toronto Star][VP reports [NP this] [PP in the softest possible way]</a>,[S stating ...]]] S? NP VP? DT NNP NNP VBZ NP PP S We can summarize productions more compactly by using a dependency framework and clipping off any dependents whose subtrees do not cross a bracketing boundary, relative to the parent. Thus, DT NNP NNP VBZ DT IN DT JJS JJ NN becomes DT NNP VBZ, ?the <a>Star reports</a>.? Viewed this way, the top fifteen (now collapsed) productions cover 59.4% of all cases and include four verb heads, in addition to a preposition and an adjective (see Table 5, right). This exposes five cases of inexact matches, three of which involve neglected determiners or adjectives to the left of the head. In fact, the only case that cannot be explained by dropped dependents is #8, where the daughters are marked but the parent is left out. Most instances contributing to this pattern are flat NPs that end with a noun, incorrectly assumed to be the head of all other words in the phrase, e.g., ... [NP a 1994 <i>New Yorker</i> article] ... As this example shows, disagreements (as well as agreements) between mark-up and machinegenerated parse trees with automatically percolated heads should be taken with a grain of salt.11 11In a relatively recent study, Ravi et al (2008) report that Charniak?s re-ranking parser (Charniak and Johnson, 2005) ? reranking-parserAug06.tar.gz, also available from ftp://ftp.cs.brown.edu/pub/nlparser/ ? attains 86.3% accuracy when trained on WSJ and tested against Brown; its nearly 5% performance loss out-of-domain is consistent with the numbers originally reported by Gildea (2001). 1281 Count Constituent Production Frac Sum 1 746 NP? NNP NNP 9.6% 2 357 NP? NNP 4.6 14.3 3 266 NP? NP PP 3.4 17.7 4 183 NP? NNP NNP NNP 2.4 20.1 5 165 NP? DT NNP NNP 2.1 22.2 6 140 NP? NN 1.8 24.0 7 131 NP? DT NNP NNP NNP 1.7 25.7 8 130 NP? DT NN 1.7 27.4 9 127 NP? DT NNP NNP 1.6 29.0 10 109 S ? NP VP 1.4 30.4 11 91 NP? DT NNP NNP NNP 1.2 31.6 12 82 NP? DT JJ NN 1.1 32.7 13 79 NP? NNS 1.0 33.7 14 65 NP? JJ NN 0.8 34.5 15 60 NP? NP NP 0.8 35.3 BLOGp +5,000 more with Count ? 60 64.7% Count Head-Outward Spawn Frac Sum 1 1,889 NNP 24.4% 2 623 NN 8.1 32.5 3 470 DT NNP 6.1 38.6 4458 DT NN 5.9 44.5 5 345 NNS 4.5 49.0 6 109 NNPS 1.4 50.4 7 98 VBG 1.3 51.6 8 96 NNP NNP NN 1.2 52.9 9 80 VBD 1.0 53.9 10 77 IN 1.0 54.9 11 74 VBN 1.0 55.9 12 73 DT JJ NN 0.9 56.8 13 71 VBZ 0.9 57.7 14 69 POS NNP 0.9 58.6 15 63 JJ 0.8 59.4 BLOGp +3,136 more with Count ? 62 40.6% Table 5: Top 15 marked productions, viewed as constituents (left) and as dependencies (right), after recursively expanding any internal nodes that did not align with the bracketing (underlined). Tabulated dependencies were collapsed, dropping any dependents that fell entirely in the same region as their parent (i.e., both inside the bracketing, both to its left or both to its right), keeping only crossing attachments. 5.3 Proposed Parsing Constraints The straight-forward approach ? forcing mark-up to correspond to constituents ? agrees with Charniak?s parse trees only 48.0% of the time, e.g., ... in [NP<a>[NP an a