{ "paper_id": "D09-1043", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T16:39:24.497912Z" }, "title": "Perceptron Reranking for CCG Realization", "authors": [ { "first": "Michael", "middle": [], "last": "White", "suffix": "", "affiliation": { "laboratory": "", "institution": "The Ohio State University Columbus", "location": { "region": "OH", "country": "USA" } }, "email": "mwhite@ling.osu.edu" }, { "first": "Rajakrishnan", "middle": [], "last": "Rajkumar", "suffix": "", "affiliation": { "laboratory": "", "institution": "The Ohio State University Columbus", "location": { "region": "OH", "country": "USA" } }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "This paper shows that discriminative reranking with an averaged perceptron model yields substantial improvements in realization quality with CCG. The paper confirms the utility of including language model log probabilities as features in the model, which prior work on discriminative training with log linear models for HPSG realization had called into question. The perceptron model allows the combination of multiple n-gram models to be optimized and then augmented with both syntactic features and discriminative n-gram features. The full model yields a stateof-the-art BLEU score of 0.8506 on Section 23 of the CCGbank, to our knowledge the best score reported to date using a reversible, corpus-engineered grammar.", "pdf_parse": { "paper_id": "D09-1043", "_pdf_hash": "", "abstract": [ { "text": "This paper shows that discriminative reranking with an averaged perceptron model yields substantial improvements in realization quality with CCG. The paper confirms the utility of including language model log probabilities as features in the model, which prior work on discriminative training with log linear models for HPSG realization had called into question. The perceptron model allows the combination of multiple n-gram models to be optimized and then augmented with both syntactic features and discriminative n-gram features. The full model yields a stateof-the-art BLEU score of 0.8506 on Section 23 of the CCGbank, to our knowledge the best score reported to date using a reversible, corpus-engineered grammar.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "In this paper, we show how discriminative training with averaged perceptron models (Collins, 2002) can be used to substantially improve surface realization with Combinatory Categorial Grammar (Steedman, 2000, CCG) . Velldal and Oepen (2005) and Nakanishi et al. (2005) have shown that discriminative training with log-linear (maximum entropy) models is effective in realization ranking with Head-Driven Phrase Structure Grammar (Pollard and Sag, 1994, HPSG) . Here we show that averaged perceptron models also perform well for realization ranking with CCG. Averaged perceptron models are very simple, just requiring a decoder and a simple update function, yet despite their simplicity they have been shown to achieve state-of-the-art results in Treebank and CCG parsing (Huang, 2008; Clark and Curran, 2007a) as well as on other NLP tasks.", "cite_spans": [ { "start": 83, "end": 98, "text": "(Collins, 2002)", "ref_id": "BIBREF13" }, { "start": 192, "end": 213, "text": "(Steedman, 2000, CCG)", "ref_id": null }, { "start": 216, "end": 240, "text": "Velldal and Oepen (2005)", "ref_id": "BIBREF33" }, { "start": 245, "end": 268, "text": "Nakanishi et al. (2005)", "ref_id": "BIBREF22" }, { "start": 428, "end": 457, "text": "(Pollard and Sag, 1994, HPSG)", "ref_id": null }, { "start": 770, "end": 783, "text": "(Huang, 2008;", "ref_id": "BIBREF20" }, { "start": 784, "end": 808, "text": "Clark and Curran, 2007a)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Along the way, we address the question of whether it is beneficial to incorporate n-gram log probabilities as baseline features in a discriminatively trained realization ranking model. On a limited domain corpus, Velldal & Oepen found that including the n-gram log probability of each candidate realization as a feature in their log-linear model yielded a substantial boost in ranking performance; on the Penn Treebank (PTB), however, Nakanishi et al. found that including an n-gram log prob feature in their model was of no benefit (with the use of bigrams instead of 4-grams suggested as a possible explanation). With these mixed results, the utility of n-gram baseline features for PTBscale discriminative realization ranking has been unclear. In our particular setting, the question is: Do n-gram log prob features improve performance in broad coverage realization ranking with CCG, where factored language models over words, partof-speech tags and supertags have previously been employed (White et al., 2007; Espinosa et al., 2008 )?", "cite_spans": [ { "start": 435, "end": 451, "text": "Nakanishi et al.", "ref_id": null }, { "start": 993, "end": 1013, "text": "(White et al., 2007;", "ref_id": "BIBREF37" }, { "start": 1014, "end": 1035, "text": "Espinosa et al., 2008", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "We answer this question in the affirmative, confirming the results of Velldal & Oepen, despite the differences in corpus size and kind of language model. We show that including n-gram log prob features in the perceptron model is highly beneficial, as the discriminative models we tested without these features performed worse than the generative baseline. These findings are in line with results with incremental parsing with perceptrons, where it is suggested that a generative baseline feature provides the perceptron algorithm with a much better starting point for learning. We also show that discriminative training allows the combination of multiple n-gram models to be optimized, and that the best model augments the n-gram log prob features with both syntactic features and discriminative n-gram features. The full model yields a stateof-the-art BLEU (Papineni et al., 2002) score of 0.8506 on Section 23 of the CCGbank, which is to our knowledge the best score reported to date using a reversible, corpus-engineered grammar.", "cite_spans": [ { "start": 858, "end": 881, "text": "(Papineni et al., 2002)", "ref_id": "BIBREF25" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The paper is organized as follows. Section 2 reviews previous work on broad coverage realization with OpenCCG. Section 3 describes our approach to realization reranking with averaged perceptron models. Section 4 presents our evaluation of the perceptron models, comparing the results of different feature sets. Section 5 compares our results to those obtained by related systems and discusses the difficulties of cross-system comparisons. Finally, Section 6 concludes with a summary and discussion of future directions for research.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "CCG (Steedman, 2000) is a unification-based categorial grammar formalism which is defined almost entirely in terms of lexical entries that encode sub-categorization information as well as syntactic feature information (e.g. number and agreement). Complementing function application as the standard means of combining a head with its argument, type-raising and composition support transparent analyses for a wide range of phenomena, including right-node raising and long distance dependencies. An example syntactic derivation appears in Figure 1 , with a long-distance dependency between point and make. Semantic composition happens in parallel with syntactic composition, which makes it attractive for generation.", "cite_spans": [ { "start": 4, "end": 20, "text": "(Steedman, 2000)", "ref_id": "BIBREF30" } ], "ref_spans": [ { "start": 536, "end": 544, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Background 2.1 Surface Realization with CCG", "sec_num": "2" }, { "text": "OpenCCG is a parsing/generation library which works by combining lexical categories for words using CCG rules and multi-modal extensions on rules (Baldridge, 2002) to produce derivations. Surface realization is the process by which logical forms are transduced to strings. OpenCCG uses a hybrid symbolic-statistical chart realizer (White, 2006) which takes logical forms as input and produces sentences by using CCG combinators to combine signs. Edges are grouped into equivalence classes when they have the same syntactic category and cover the same parts of the input logical form. Alternative realizations are ranked using integrated n-gram or perceptron scoring, and pruning takes place within equivalence classes of edges. To more robustly support broad coverage surface realization, OpenCCG greedily assembles fragments in the event that the realizer fails to find a complete realization.", "cite_spans": [ { "start": 146, "end": 163, "text": "(Baldridge, 2002)", "ref_id": "BIBREF1" }, { "start": 331, "end": 344, "text": "(White, 2006)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Background 2.1 Surface Realization with CCG", "sec_num": "2" }, { "text": "To illustrate the input to OpenCCG, consider the semantic dependency graph in Figure 2 . In Figure 2 : Semantic dependency graph from the CCGbank for He has a point he wants to make [. . . ] , along with gold-standard supertags (category labels) the graph, each node has a lexical predication (e.g. make.03) and a set of semantic features (e.g. NUM sg); nodes are connected via dependency relations (e.g. ARG0 ). (Gold-standard supertags, or category labels, are also shown; see Section 2.4 for their role in hypertagging.) Internally, such graphs are represented using Hybrid Logic Dependency Semantics (HLDS), a dependency-based approach to representing linguistic meaning (Baldridge and Kruijff, 2002) . In HLDS, each semantic head (corresponding to a node in the graph) is associated with a nominal that identifies its discourse referent, and relations between heads and their dependents are modeled as modal relations.", "cite_spans": [ { "start": 182, "end": 190, "text": "[. . . ]", "ref_id": null }, { "start": 675, "end": 704, "text": "(Baldridge and Kruijff, 2002)", "ref_id": "BIBREF0" } ], "ref_spans": [ { "start": 78, "end": 86, "text": "Figure 2", "ref_id": null }, { "start": 92, "end": 100, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Background 2.1 Surface Realization with CCG", "sec_num": "2" }, { "text": "Our starting point is an enhanced version of the CCGbank (Hockenmaier and Steedman, 2007 )-a corpus of CCG derivations derived from the Penn Treebank-with Propbank (Palmer et al., 2005) roles projected onto it (Boxwell and White, 2008) . To engineer a grammar from this corpus suitable for realization with OpenCCG, the derivations are first revised to reflect the lexicalized treatment of coordination and punctuation assumed by the multi-modal version of CCG that is implemented in OpenCCG (White and Rajkumar, 2008) . Further changes are necessary to support semantic dependencies rather than surface syntactic ones; in Figure 1 : Syntactic derivation from the CCGbank for He has a point he wants to make [. . . ] particular, the features and unification constraints in the categories related to semantically empty function words such complementizers, infinitivalto, expletive subjects, and case-marking prepositions are adjusted to reflect their purely syntactic status.", "cite_spans": [ { "start": 57, "end": 88, "text": "(Hockenmaier and Steedman, 2007", "ref_id": "BIBREF18" }, { "start": 164, "end": 185, "text": "(Palmer et al., 2005)", "ref_id": "BIBREF24" }, { "start": 210, "end": 235, "text": "(Boxwell and White, 2008)", "ref_id": "BIBREF7" }, { "start": 492, "end": 518, "text": "(White and Rajkumar, 2008)", "ref_id": "BIBREF36" }, { "start": 708, "end": 716, "text": "[. . . ]", "ref_id": null } ], "ref_spans": [ { "start": 623, "end": 631, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Realization from an Enhanced CCGbank", "sec_num": "2.2" }, { "text": "In the second step, a grammar is extracted from the converted CCGbank and augmented with logical forms. Categories and unary type changing rules (corresponding to zero morphemes) are sorted by frequency and extracted if they meet the specified frequency thresholds. A separate transformation then uses a few dozen generalized templates to add logical forms to the categories, in a fashion reminiscent of (Bos, 2005) . As shown in Figure 2 , numbered semantic roles are taken from PropBank when available, and more specific relations are introduced in the categories for closedclass items such as determiners.", "cite_spans": [ { "start": 404, "end": 415, "text": "(Bos, 2005)", "ref_id": "BIBREF6" } ], "ref_spans": [ { "start": 430, "end": 438, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Realization from an Enhanced CCGbank", "sec_num": "2.2" }, { "text": "After logical form insertion, the extracted and augmented grammar is loaded and used to parse the sentences in the CCGbank according to the gold-standard derivation. If the derivation can be successfully followed, the parse yields a logical form which is saved along with the corpus sentence in order to later test the realizer. Currently, the algorithm succeeds in creating logical forms for 98.85% of the sentences in the development section (Sect. 00) of the converted CCGbank, and 97.06% of the sentences in the test section (Sect. 23). Of these, 95.99% of the development LFs are semantic dependency graphs with a single root, while 95.81% of the test LFs have a single root. The remaining cases, with multiple roots, are missing one or more dependencies required to form a fully connected graph. Such missing dependencies usually reflect remaining inadequacies in the logical form templates.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Realization from an Enhanced CCGbank", "sec_num": "2.2" }, { "text": "An error analysis of OpenCCG output by Rajkumar et al. (2009) recently revealed that out of 2331 named entities (NEs) annotated by the BBN corpus (Weischedel and Brunstein, 2005) , 238 were not realized correctly. For example, multiword NPs like Texas Instruments Japan Ltd. were realized as Japan Texas Instruments Ltd. Accordingly, inspired by Hogan et al.'s (2007) 's Experiment 1, Rajkumar et al. used the BBN corpus NE annotation to collapse certain classes of NEs. But unlike Hogan et al.'s experiment where all the NEs annotated by the BBN corpus were collapsed, Rajkumar et al. chose to collapse into single tokens only NEs whose exact form can be reasonably expected to be specified in the input to the realizer. For example, while some quantificational or comparatives phrases like more than $ 10,000 are annotated as MONEY in the BBN corpus, Rajkumar et al. only collapse $ 10,000 into an atomic unit, with more than handled compositionally according to the semantics assigned to it by the grammar. Thus, after transferring the BBN annotations to the CCGbank corpus, Rajkumar et al. (partially) collapsed NEs which are CCGbank constituents according to the following rules: (1) completely collapse the PERSON, ORGANIZATION, GPE, WORK OF ART major class type entitites; (2) ignore phrases like three decades later, which are annotated as DATE entities; and (3) collapse all phrases with POS tags CD or NNP(S) or lexical items % or $, ensuring that all prototypical named entities are collapsed.", "cite_spans": [ { "start": 39, "end": 61, "text": "Rajkumar et al. (2009)", "ref_id": "BIBREF27" }, { "start": 146, "end": 178, "text": "(Weischedel and Brunstein, 2005)", "ref_id": "BIBREF35" }, { "start": 346, "end": 367, "text": "Hogan et al.'s (2007)", "ref_id": null }, { "start": 1078, "end": 1105, "text": "Rajkumar et al. (partially)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Realization from an Enhanced CCGbank", "sec_num": "2.2" }, { "text": "It is worth noting that improvements in our corpus-based grammar engineering processincluding a more precise treatment of punctuation, better named entity handling and the addition of catch-all logical form templates-have resulted in a 13.5 BLEU point improvement in our baseline realization scores on Section 00 of the CCGbank, from a score of 0.6567 in (Espinosa et al., 2008) to 0.7917 in (Rajkumar et al., 2009) , contributing greatly to the state-of-the-art results reported in Section 4. A further 4.5 point improvement is obtained from the use of named entity classes in language modeling and hypertagging (Rajkumar et al., 2009) , as described next, and from our perceptron reranking model, described in Section 3.", "cite_spans": [ { "start": 355, "end": 378, "text": "(Espinosa et al., 2008)", "ref_id": "BIBREF15" }, { "start": 392, "end": 415, "text": "(Rajkumar et al., 2009)", "ref_id": "BIBREF27" }, { "start": 613, "end": 636, "text": "(Rajkumar et al., 2009)", "ref_id": "BIBREF27" } ], "ref_spans": [], "eq_spans": [], "section": "Realization from an Enhanced CCGbank", "sec_num": "2.2" }, { "text": "As in (White et al., 2007; Rajkumar et al., 2009) , we use factored language models (Bilmes and Kirchhoff, 2003) over words, part-of-speech tags and supertags 1 to score partial and complete realizations. The trigram models were created using the SRILM toolkit (Stolcke, 2002) on the standard training sections (02-21) of the CCGbank, with sentence-initial words (other than proper names) uncapitalized. While these models are considerably smaller than the ones used in (Langkilde-Geary, 2002; Velldal and Oepen, 2005) , the training data does have the advantage of being in the same domain and genre. The models employ interpolated Kneser-Ney smoothing with the default frequency cutoffs. The best performing model interpolates three component models using rankorder centroid weights: (1) a word trigram model;", "cite_spans": [ { "start": 6, "end": 26, "text": "(White et al., 2007;", "ref_id": "BIBREF37" }, { "start": 27, "end": 49, "text": "Rajkumar et al., 2009)", "ref_id": "BIBREF27" }, { "start": 84, "end": 112, "text": "(Bilmes and Kirchhoff, 2003)", "ref_id": "BIBREF4" }, { "start": 261, "end": 276, "text": "(Stolcke, 2002)", "ref_id": "BIBREF31" }, { "start": 470, "end": 493, "text": "(Langkilde-Geary, 2002;", "ref_id": "BIBREF21" }, { "start": 494, "end": 518, "text": "Velldal and Oepen, 2005)", "ref_id": "BIBREF33" } ], "ref_spans": [], "eq_spans": [], "section": "Factored Language Models", "sec_num": "2.3" }, { "text": "(2) a word model with semantic classes replacing named entities; and (3) a trigram model that chains a POS model with a supertag model, where the POS model (P ) conditions on the previous two POS tags, and the supertag model (S) conditions on the previous two POS tags as well as the current one, as shown below:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Factored Language Models", "sec_num": "2.3" }, { "text": "p P S ( F i | F i\u22121 i\u22122 ) = p(P i | P i\u22121 i\u22122 )p(S i | P i i\u22122 )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Factored Language Models", "sec_num": "2.3" }, { "text": "(1) Training data for the semantic class-replaced model was created by replacing (collapsed) words with their NE classes, in order to address data sparsity issues caused by rare words in the same semantic class. For example, the Section 00 sentence Pierre Vinken , 61 years old , will join the board as a nonexecutive director Nov. 29 . becomes PERSON , DATE:AGE DATE:AGE old , will join the ORG DESC:OTHER as a nonexecutive PER DESC DATE:DATE DATE:DATE . During realization, word forms are generated, but are then replaced by their semantic classes for scoring using the semantic class-replaced model, similar to Oh and Rudnicky (2002) .", "cite_spans": [ { "start": 614, "end": 636, "text": "Oh and Rudnicky (2002)", "ref_id": "BIBREF23" } ], "ref_spans": [], "eq_spans": [], "section": "Factored Language Models", "sec_num": "2.3" }, { "text": "Note that the use of supertags in the factored language model to score possible realizations is distinct from the prediction of supertags for lexical category assignment: the former takes the words in the local context into account (as in supertagging for parsing), while the latter takes features of the logical form into account. This latter process we call hypertagging, to which we now turn.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Factored Language Models", "sec_num": "2.3" }, { "text": "A crucial component of the OpenCCG realizer is the hypertagger (Espinosa et al., 2008) , or supertagger for surface realization, which uses a maximum entropy model to assign the most likely lexical categories to the predicates in the input logical form, thereby greatly constraining the realizer's search space. 2 Figure 2 shows gold-standard supertags for the lexical predicates in the graph; such category labels are predicted by the hypertagger at run-time. As in recent work on using supertagging in parsing, the hypertagger operates in a multitagging paradigm (Curran et al., 2006) , where a variable number of predictions are made per input predicate. Instead of basing category assignment on linear word and POS context, however, the hypertagger predicts lexical categories based on contexts within a directed graph structure representing the logical form (LF) of the sentence to be realized. The hypertagger generalizes Bangalore and Rambow's (2000) method of using supertags in generation by using maximum entropy models with a larger local context.", "cite_spans": [ { "start": 63, "end": 86, "text": "(Espinosa et al., 2008)", "ref_id": "BIBREF15" }, { "start": 565, "end": 586, "text": "(Curran et al., 2006)", "ref_id": "BIBREF14" } ], "ref_spans": [ { "start": 314, "end": 322, "text": "Figure 2", "ref_id": null } ], "eq_spans": [], "section": "Hypertagging", "sec_num": "2.4" }, { "text": "During realization, the hypertagger returns a \u03b2best list of supertags in order of decreasing probability. Increasing the number of categories returned clearly increases the likelihood that the most-correct supertag is among them, but at a corresponding cost in chart size. Accordingly, the hypertagger begins with a highly restrictive value for \u03b2, and backs off to progressively less-restrictive values if no complete realization can be found using the set of supertags returned. Clark and Curran (2007b) have shown this iterative relaxation strategy to be highly effective in CCG parsing.", "cite_spans": [ { "start": 480, "end": 504, "text": "Clark and Curran (2007b)", "ref_id": "BIBREF11" } ], "ref_spans": [], "eq_spans": [], "section": "Hypertagging", "sec_num": "2.4" }, { "text": "As Collins (2002) observes, perceptron training involves a simple, on-line algorithm, with few iterations typically required to achieve good performance. Moreover, averaged perceptrons-which Input: training examples (x i , y i ) Initialization: set \u03b1 = 0, or use optional input model Algorithm:", "cite_spans": [ { "start": 3, "end": 17, "text": "Collins (2002)", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Perceptron Reranking", "sec_num": "3" }, { "text": "for Figure 3 : Averaged perceptron training algorithm approximate voted perceptrons, a maximummargin method with attractive theoretical properties-seem to work remarkably well in practice, while adding little further complexity. Additionally, since features only take on nonzero values when they appear in training items requiring updates, perceptrons integrate feature selection with, and often produce quite small models, especially when starting with a good baseline.", "cite_spans": [], "ref_spans": [ { "start": 4, "end": 12, "text": "Figure 3", "ref_id": null } ], "eq_spans": [], "section": "Perceptron Reranking", "sec_num": "3" }, { "text": "t = 1 . . . T , i = 1 . . . N z i = argmax y\u2208GEN(x i ) \u03a6(x i , y) \u2022 \u03b1 if z i = y i \u03b1 = \u03b1 + \u03a6(x i , y i ) \u2212 \u03a6(x i , z i ) Output: \u03b1 = T t=1 N i=1 \u03b1 ti /T N", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Perceptron Reranking", "sec_num": "3" }, { "text": "The generic averaged perceptron training algorithm appears in Figure 3 . In our case, the algorithm trains a model for reranking the n-best realizations generated using our existing factored language model for scoring, with the oracle-best realization considered the correct answer. Accordingly, the input to the algorithm is a list of pairs (x i , y i ), where x i is a logical form, GEN(x i ) are the n-best realizations for x i , and y i is the oraclebest member of GEN(x i ). The oracle-best realization is determined using a 4-gram precision metric (approximating BLEU) against the reference sentence.", "cite_spans": [], "ref_spans": [ { "start": 62, "end": 70, "text": "Figure 3", "ref_id": null } ], "eq_spans": [], "section": "Perceptron Reranking", "sec_num": "3" }, { "text": "We have followed Huang (2008) in using oracle-best targets for training, rather than gold standard ones, in order to better approximate test conditions during training. However, following Clark & Curran (2007a) , during training we seed the realizer with the gold-standard supertags, augmenting the hypertagger's \u03b2-best list, in order to ensure that the n-best realizations are generally of high quality; consequently, the gold standard realization (i.e., the corpus sentence) usually appears in the n-best list. 3 In addition, we use a hypertagger trained on all the training data, to improve hypertagger performance, while excluding the cur-rent training section (in jack-knifed fashion) from the word-based parts of the language model, in order to make the language model scores more realistic. It remains for future work to determine whether using a different compromise between ensuring high-quality training data and remaining faithful to the test conditions would yield better results.", "cite_spans": [ { "start": 17, "end": 29, "text": "Huang (2008)", "ref_id": "BIBREF20" }, { "start": 188, "end": 210, "text": "Clark & Curran (2007a)", "ref_id": "BIBREF10" }, { "start": 513, "end": 514, "text": "3", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Perceptron Reranking", "sec_num": "3" }, { "text": "Since realization of the n-best lists for training is the most time-consuming part of the process, in our current implementation we perform this step once, generating event files along the way containing feature vectors for each candidate realization. The event files are used to calculate the frequency distribution for the features, and minimum cutoffs are chosen to trim the feature alphabet to a reasonable size. Training then takes place by iterating over the event files, ignoring features that do not appear in the alphabet. As Figure 3 indicates, training consists of calculating the topranked realization according to the current model \u03b1, and performing an update when the top-ranked realization does not match the oracle-best realization. Updates to the model add the feature vector \u03a6(x i , y i ) for the missed oracle-best realization, and subtract the feature vector \u03a6(x i , z i ) for the mistakenly top-ranked realization. The final model averages the models across the T iterations over the training data, and N test cases within each iteration.", "cite_spans": [], "ref_spans": [ { "start": 535, "end": 543, "text": "Figure 3", "ref_id": null } ], "eq_spans": [], "section": "Perceptron Reranking", "sec_num": "3" }, { "text": "Note that while training the perceptron model involves n-best reranking, realization with the resulting model can be viewed as forest rescoring, since scoring of all partial realizations is integrated into the realizer's beam search. In future work, we intend to investigate saving the realizer's packed charts, rather than event files, and integrating the unpacking of the charts with the perceptron training algorithm.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Perceptron Reranking", "sec_num": "3" }, { "text": "The features we employ in our perceptron models are of three kinds. First, as in the log-linear models of Velldal & Oepen and Nakanishi et al., we incorporate the log probability of the candidate realization's word sequence according to our factored language model as a single feature in the perceptron model. Since our language model linearly interpolates three component models, we also include the log prob from each component language model as a feature, so that the combination of these components can be optimized. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Perceptron Reranking", "sec_num": "3" }, { "text": "For the experiments reported below, we used a lexico-grammar extracted from Sections 02-21 of our enhanced CCGbank, a hypertagging model incorporating named entity class features, and a trigram factored language model over words, named entity classes, part-of-speech tags and supertags, as described in the preceding section. BLEU scores were calculated after removing the underscores between collapsed NEs. Events were generated for each training section separately. As already noted, the hypertagger and POS/supertag language model was trained on all the training sections, while separate word-based models were trained excluding each of the training sections in turn. Event files for 26530 training sentences with complete realizations were generated in 7 hours and 16 minutes on a cluster using one commodity server per section, with an average n-best list size of 18.2. Perceptron models were trained on single machines; details for three of the models appear in Table 2 . The complete set of models is listed in Table 3 .", "cite_spans": [], "ref_spans": [ { "start": 968, "end": 975, "text": "Table 2", "ref_id": "TABREF3" }, { "start": 1018, "end": 1025, "text": "Table 3", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "Experimental Conditions", "sec_num": "4.1" }, { "text": "Realization results on the development section are given in Table 4 . As the first block of rows after the baseline shows, of the models incorporating a single kind of feature, only the one with the ngram log prob features beats the baseline BLEU The second block of rows shows that both the discriminative n-gram features and the syntactic features provide a substantial boost when used with the ngram log prob features, with the syntactic features yielding a more than 3 BLEU point gain. The final row shows that the full model works best, though the boosts provided by the syntactic and discriminative n-gram features are clearly not independent. The BLEU point trends are mirrored in the percentage of exact match realizations, which goes up by more than 10% from the baseline. The percentage of complete (i.e., non-fragmentary) realizations, however, goes down; we expect that this is due to the time taken up by our current naive method of feature extraction, which does not cache the features calculated for partial realizations. Realization results on the standard test section appear in Table 5 , confirming the gains made by the full model over the baseline. 5 We calculated statistical significance for the main results on the development section using bootstrap random sampling. 6 After re-sampling 1000 times, significance was calculated using a paired t-test (999 d.f.). The results indicated that lp-only exceeded the baseline, lp-ngram and lp- syn exceeded lp-only, and the full model exceeded lp-syn, with p < 0.0001 in each case. Table 6 presents four examples where the full model improves upon the baseline. Example sentence wsj 0020.10 in Table 6 is a case where the perceptron successfully weights the component ngram models, as the lp-ngram model and those that build on it get it right. Note that here, the modifier ordering in small video-viewing is not specified in the logical form and either ordering is possible syntactically. In wsj 0024.2, number agreement between the conjoined subject noun phrase and verb is obtained only with the full model. This suggests that the full model is more robust to cases where the grammar is insufficiently precise (number agreement is enforced by the grammar in only the simplest cases). Example wsj 0034.9 corrects a VP ordering mismatch, where the corpus sentence is clearly preferred to the one where into oblivion is shifted to the end. Finally, wsj 0047.13 corrects an animacy mismatch on the wh-pronoun, in large part due to the high negative weight assigned to the discriminative n-gram feature PER-SON , which. Note that the full model still differs from the original sentence in its placement of the adverb reportedly, choosing the arguably more natural position following the auxiliary. most similar systems to ours are those of Nakanishi et al. (2005) and Hogan et al. (2007) , as they both involve chart realizers for reversible grammars engineered from the Penn Treebank. While direct comparisons across systems cannot really be made when inputs vary in their semantic depth and specificity, we observe that our all-sentences BLEU score of 0.8506 exceeds that of Hogan et al., who report a top score of 0.6882 (though with coverage near 100%), and also surpasses Nakanishi et al.'s score of 0.7733, despite their results being limited to sentences of length 20 or less (with 91% coverage). Velldal & Oepen's (2005) system is also closely related, as noted in the introduction, but as their experiments are on a limited domain corpus, their results cannot be compared at all meaningfully.", "cite_spans": [ { "start": 1169, "end": 1170, "text": "5", "ref_id": null }, { "start": 2804, "end": 2827, "text": "Nakanishi et al. (2005)", "ref_id": "BIBREF22" }, { "start": 2832, "end": 2851, "text": "Hogan et al. (2007)", "ref_id": "BIBREF19" }, { "start": 3368, "end": 3392, "text": "Velldal & Oepen's (2005)", "ref_id": "BIBREF33" } ], "ref_spans": [ { "start": 60, "end": 67, "text": "Table 4", "ref_id": "TABREF7" }, { "start": 1096, "end": 1103, "text": "Table 5", "ref_id": "TABREF8" }, { "start": 1548, "end": 1555, "text": "Table 6", "ref_id": "TABREF10" }, { "start": 1660, "end": 1667, "text": "Table 6", "ref_id": "TABREF10" } ], "eq_spans": [], "section": "Results", "sec_num": "4.2" }, { "text": "As alluded to above, realization systems cannot be easily compared, even on the same corpus, when their inputs are not the same. This point is dramatically illustrated in Langkilde-Geary's (2002) system, where a BLEU score of 0.514 is reported for minimally specified inputs on PTB Section 23, while a score of 0.757 is reported for the 'Per-mute, no dir' case (which perhaps most closely resembles our inputs), and a score of 0.924 is reported for the most fully specified inputs; note, however, that in the latter case word order is determined by sibling order in the inputs, an assumption not commonly made. As another example, Guo et al. (2008) report a competitive result of 0.7440 (with 100% coverage) using a dependencybased approach; however, their inputs, like those of Hogan et al., include more surface syntactic information than ours, as they specify case-marking prepositions, wh-pronouns and complementizers. In a recent experiment to assess the impact of input specificity, we found that including predicates for all prepositions in our logical forms boosted our baseline results by more than 3 BLEU points, with complete realizations found in more than 90% of the test cases, indicating that generating from a more surfacy input is indeed an easier task than generating from a deeper representation. Given the current lack of consensus on realizer input specificity, we believe it is important to keep in mind that within-system comparisons (such as those in the preceding section) are the ones that should be given the most credence.", "cite_spans": [ { "start": 171, "end": 195, "text": "Langkilde-Geary's (2002)", "ref_id": "BIBREF21" }, { "start": 631, "end": 648, "text": "Guo et al. (2008)", "ref_id": "BIBREF17" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work and Discussion", "sec_num": "5" }, { "text": "Returning to our cross-system comparison, it is perhaps surprising that Callaway (2005) reports the best PTB BLEU score to date, 0.9321, with 98.5% coverage, using a purely symbolic, handcrafted grammar augmented to handle the most frequent coverage issues for the PTB. While Callaway's inputs are unordered, word order is often determined by positional features (e.g. front) or by the type of modification (e.g. describer vs. qualifier), and parts-of-speech are included for lexical items. Additionally, in contrast to our approach, Callaway makes use of a generationonly grammar, rather than a reversible one, and his approach is less well-suited to producing n-best outputs. Nevertheless, his high scores do suggest the potential for precise grammar engineering to improve realization quality.", "cite_spans": [ { "start": 72, "end": 87, "text": "Callaway (2005)", "ref_id": "BIBREF8" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work and Discussion", "sec_num": "5" }, { "text": "While we have yet to perform a thorough error analysis, our impression is that although the current set of syntactic features substantially improves clausal constituent ordering, a variety of disfluent cases remain. More thorough investigations of features for constituent ordering in English have been performed by Ringger et al. (2004) , Filippova and Strube (2009) and Zhong and Stent (2009) , all of whom develop classifiers for determining linear order. In future work, we plan to investigate whether features inspired by these approaches can be usefully integrated into our perceptron reranker.", "cite_spans": [ { "start": 316, "end": 337, "text": "Ringger et al. (2004)", "ref_id": "BIBREF28" }, { "start": 340, "end": 367, "text": "Filippova and Strube (2009)", "ref_id": "BIBREF16" }, { "start": 372, "end": 394, "text": "Zhong and Stent (2009)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Related Work and Discussion", "sec_num": "5" }, { "text": "Also related to the present work is discriminative training in syntax-based MT (Turian et al., 2007; Watanabe et al., 2007; Blunsom et al., 2008; Chiang et al., 2009) . Not surprisingly, since MT is a harder problem than surface realization, syntaxbased MT systems have made use of less precise grammars and more impoverished (target-side) feature sets than those tackling realization ranking. With progress on discriminative training with large numbers of features in syntax-based MT, the features found to be useful for high-quality surface realization may become increasingly relevant for MT as well.", "cite_spans": [ { "start": 79, "end": 100, "text": "(Turian et al., 2007;", "ref_id": "BIBREF32" }, { "start": 101, "end": 123, "text": "Watanabe et al., 2007;", "ref_id": "BIBREF34" }, { "start": 124, "end": 145, "text": "Blunsom et al., 2008;", "ref_id": "BIBREF5" }, { "start": 146, "end": 166, "text": "Chiang et al., 2009)", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work and Discussion", "sec_num": "5" }, { "text": "In this paper, we have shown how discriminative reranking with an averaged perceptron model can be used to achieve substantial improvements in realization quality with CCG. Using a comprehensive feature set, we have also confirmed the utility of including language model log probabilities as features in the model, which prior work on discriminative training with log linear models for HPSG realization had called into question. The perceptron model allows the combination of multiple n-gram models to be optimized and then augmented with both syntactic features and discriminative n-gram features, inspired by related work in discriminative parsing and language modeling for speech recognition. The full model yields a state-of-the-art BLEU score of 0.8506 on Section 23 of the CCGbank, to our knowledge the best score reported to date using a reversible, corpusengineered grammar, despite our use of deeper, less specific inputs. Finally, the perceptron model paves the way for exploring the utility of richer feature spaces in statistical realization, including the use of linguistically-motivated and non-local features, a topic which we plan to investigate in future work.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions", "sec_num": "6" }, { "text": "With CCG, supertags(Bangalore and Joshi, 1999) are lexical categories considered as fine-grained syntactic labels.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "The approach has been dubbed hypertagging since it operates at a level \"above\" the syntax, moving from semantic representations to syntactic categories.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "As in Clark & Curran's approach, we use a single \u03b2 value during training, rather than iteratively loosening the \u03b2 value; the chosen \u03b2 value determines the size of the discrimation space.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "We have omitted Clark & Curran's root features, since the category we use for the full stop ensures that it must appear at the root of any complete derivation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Note that the baseline for Section 23 uses 4-grams and a filter for balanced punctuation(White and Rajkumar, 2008), unlike the other reported configurations, which would explain the somewhat smaller increase seen with this section.6 Scripts for running these tests are available at http://projectile.sv.cmu.edu/research/ public/tools/bootStrap/tutorial.htm", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "MichaelWhite. 2006. Efficient Realization of Coordinate Structures in Combinatory Categorial Grammar. Research on Language and Computation, 4(1):39-75. Huayan Zhong and Amanda Stent. 2009. Determining the position of adverbial phrases in English. In Proc. NAACL HLT 2009 Short Papers.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "This work was supported in part by NSF grant IIS-0812297 and by an allocation of computing time from the Ohio Supercomputer Center. Our thanks also to the OSU Clippers group and the anonymous reviewers for helpful comments and discussion.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgements", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Coupling CCG and Hybrid Logic Dependency Semantics", "authors": [ { "first": "Jason", "middle": [], "last": "Baldridge", "suffix": "" }, { "first": "-", "middle": [], "last": "Geert", "suffix": "" } ], "year": 2002, "venue": "Proc. ACL-02", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jason Baldridge and Geert-Jan Kruijff. 2002. Cou- pling CCG and Hybrid Logic Dependency Seman- tics. In Proc. ACL-02.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Lexically Specified Derivational Control in Combinatory Categorial Grammar", "authors": [ { "first": "Jason", "middle": [], "last": "Baldridge", "suffix": "" } ], "year": 2002, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jason Baldridge. 2002. Lexically Specified Deriva- tional Control in Combinatory Categorial Gram- mar. Ph.D. thesis, University of Edinburgh.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Supertagging: An Approach to Almost Parsing", "authors": [ { "first": "Srinivas", "middle": [], "last": "Bangalore", "suffix": "" }, { "first": "Aravind", "middle": [ "K" ], "last": "Joshi", "suffix": "" } ], "year": 1999, "venue": "Computational Linguistics", "volume": "25", "issue": "2", "pages": "237--265", "other_ids": {}, "num": null, "urls": [], "raw_text": "Srinivas Bangalore and Aravind K. Joshi. 1999. Su- pertagging: An Approach to Almost Parsing. Com- putational Linguistics, 25(2):237-265.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Exploiting a probabilistic hierarchical model for generation", "authors": [ { "first": "Srinivas", "middle": [], "last": "Bangalore", "suffix": "" }, { "first": "Owen", "middle": [], "last": "Rambow", "suffix": "" } ], "year": 2000, "venue": "Proc. COLING-00", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Srinivas Bangalore and Owen Rambow. 2000. Ex- ploiting a probabilistic hierarchical model for gener- ation. In Proc. COLING-00.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Factored language models and general parallelized backoff", "authors": [ { "first": "Jeff", "middle": [], "last": "Bilmes", "suffix": "" }, { "first": "Katrin", "middle": [], "last": "Kirchhoff", "suffix": "" } ], "year": 2003, "venue": "Proc. HLT-03", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jeff Bilmes and Katrin Kirchhoff. 2003. Factored lan- guage models and general parallelized backoff. In Proc. HLT-03.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "A discriminative latent variable model for statistical machine translation", "authors": [ { "first": "Phil", "middle": [], "last": "Blunsom", "suffix": "" }, { "first": "Trevor", "middle": [], "last": "Cohn", "suffix": "" }, { "first": "Miles", "middle": [], "last": "Osborne", "suffix": "" } ], "year": 2008, "venue": "Proc. ACL-08: HLT", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Phil Blunsom, Trevor Cohn, and Miles Osborne. 2008. A discriminative latent variable model for statistical machine translation. In Proc. ACL-08: HLT.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Towards wide-coverage semantic interpretation", "authors": [ { "first": "Johan", "middle": [], "last": "Bos", "suffix": "" } ], "year": 2005, "venue": "Proc. IWCS-6", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Johan Bos. 2005. Towards wide-coverage semantic interpretation. In Proc. IWCS-6.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Projecting Propbank roles onto the CCGbank", "authors": [ { "first": "Stephen", "middle": [], "last": "Boxwell", "suffix": "" }, { "first": "Michael", "middle": [], "last": "White", "suffix": "" } ], "year": 2008, "venue": "Proc. LREC-08", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Stephen Boxwell and Michael White. 2008. Projecting Propbank roles onto the CCGbank. In Proc. LREC- 08.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "The types and distributions of errors in a wide coverage surface realizer evaluation", "authors": [ { "first": "Charles", "middle": [], "last": "Callaway", "suffix": "" } ], "year": 2005, "venue": "Proceedings of the 10th European Workshop on Natural Language Generation", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Charles Callaway. 2005. The types and distributions of errors in a wide coverage surface realizer evaluation. In Proceedings of the 10th European Workshop on Natural Language Generation.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "11,001 new features for statistical machine translation", "authors": [ { "first": "David", "middle": [], "last": "Chiang", "suffix": "" }, { "first": "Kevin", "middle": [], "last": "Knight", "suffix": "" }, { "first": "Wei", "middle": [], "last": "Wang", "suffix": "" } ], "year": 2009, "venue": "Proc. NAACL HLT", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "David Chiang, Kevin Knight, and Wei Wang. 2009. 11,001 new features for statistical machine transla- tion. In Proc. NAACL HLT 2009.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Perceptron training for a wide-coverage lexicalized-grammar parser", "authors": [ { "first": "Stephen", "middle": [], "last": "Clark", "suffix": "" }, { "first": "James", "middle": [], "last": "Curran", "suffix": "" } ], "year": 2007, "venue": "ACL 2007 Workshop on Deep Linguistic Processing", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Stephen Clark and James Curran. 2007a. Perceptron training for a wide-coverage lexicalized-grammar parser. In ACL 2007 Workshop on Deep Linguistic Processing.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Wide-Coverage Efficient Statistical Parsing with CCG and Log-Linear Models", "authors": [ { "first": "Stephen", "middle": [], "last": "Clark", "suffix": "" }, { "first": "James", "middle": [ "R" ], "last": "Curran", "suffix": "" } ], "year": 2007, "venue": "Computational Linguistics", "volume": "33", "issue": "4", "pages": "493--552", "other_ids": {}, "num": null, "urls": [], "raw_text": "Stephen Clark and James R. Curran. 2007b. Wide- Coverage Efficient Statistical Parsing with CCG and Log-Linear Models. Computational Linguistics, 33(4):493-552.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Incremental parsing with the perceptron algorithm", "authors": [ { "first": "Michael", "middle": [], "last": "Collins", "suffix": "" }, { "first": "Brian", "middle": [], "last": "Roark", "suffix": "" } ], "year": 2004, "venue": "Proc. ACL-04", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Michael Collins and Brian Roark. 2004. Incremen- tal parsing with the perceptron algorithm. In Proc. ACL-04.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Discriminative training methods for hidden Markov models: theory and experiments with perceptron algorithms", "authors": [ { "first": "Michael", "middle": [], "last": "Collins", "suffix": "" } ], "year": 2002, "venue": "Proc. EMNLP-02", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Michael Collins. 2002. Discriminative training meth- ods for hidden Markov models: theory and ex- periments with perceptron algorithms. In Proc. EMNLP-02.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Multi-tagging for lexicalized-grammar parsing", "authors": [ { "first": "James", "middle": [ "R" ], "last": "Curran", "suffix": "" }, { "first": "Stephen", "middle": [], "last": "Clark", "suffix": "" }, { "first": "David", "middle": [], "last": "Vadas", "suffix": "" } ], "year": 2006, "venue": "Proc. COLING/ACL-06", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "James R. Curran, Stephen Clark, and David Vadas. 2006. Multi-tagging for lexicalized-grammar pars- ing. In Proc. COLING/ACL-06.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Hypertagging: Supertagging for surface realization with CCG", "authors": [ { "first": "Dominic", "middle": [], "last": "Espinosa", "suffix": "" }, { "first": "Michael", "middle": [], "last": "White", "suffix": "" }, { "first": "Dennis", "middle": [], "last": "Mehay", "suffix": "" } ], "year": 2008, "venue": "Proc. ACL-08: HLT", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Dominic Espinosa, Michael White, and Dennis Mehay. 2008. Hypertagging: Supertagging for surface real- ization with CCG. In Proc. ACL-08: HLT.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Tree linearization in English: Improving language model based approaches", "authors": [ { "first": "Katja", "middle": [], "last": "Filippova", "suffix": "" }, { "first": "Michael", "middle": [], "last": "Strube", "suffix": "" } ], "year": 2009, "venue": "Proc. NAACL HLT 2009 Short Papers", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Katja Filippova and Michael Strube. 2009. Tree lin- earization in English: Improving language model based approaches. In Proc. NAACL HLT 2009 Short Papers.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Dependency-based n-gram models for general purpose sentence realisation", "authors": [ { "first": "Yuqing", "middle": [], "last": "Guo", "suffix": "" }, { "first": "Josef", "middle": [], "last": "Van Genabith", "suffix": "" }, { "first": "Haifeng", "middle": [], "last": "Wang", "suffix": "" } ], "year": 2008, "venue": "Proc. COLING-08", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yuqing Guo, Josef van Genabith, and Haifeng Wang. 2008. Dependency-based n-gram models for general purpose sentence realisation. In Proc. COLING-08.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "CCGbank: A Corpus of CCG Derivations and Dependency Structures Extracted from the Penn Treebank", "authors": [ { "first": "Julia", "middle": [], "last": "Hockenmaier", "suffix": "" }, { "first": "Mark", "middle": [], "last": "Steedman", "suffix": "" } ], "year": 2007, "venue": "Computational Linguistics", "volume": "33", "issue": "3", "pages": "355--396", "other_ids": {}, "num": null, "urls": [], "raw_text": "Julia Hockenmaier and Mark Steedman. 2007. CCG- bank: A Corpus of CCG Derivations and Depen- dency Structures Extracted from the Penn Treebank. Computational Linguistics, 33(3):355-396.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Exploiting multi-word units in history-based probabilistic generation", "authors": [ { "first": "Deirdre", "middle": [], "last": "Hogan", "suffix": "" }, { "first": "Conor", "middle": [], "last": "Cafferkey", "suffix": "" }, { "first": "Aoife", "middle": [], "last": "Cahill", "suffix": "" }, { "first": "Josef", "middle": [], "last": "Van Genabith", "suffix": "" } ], "year": 2007, "venue": "Proc. EMNLP-CoNLL", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Deirdre Hogan, Conor Cafferkey, Aoife Cahill, and Josef van Genabith. 2007. Exploiting multi-word units in history-based probabilistic generation. In Proc. EMNLP-CoNLL.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Forest reranking: Discriminative parsing with non-local features", "authors": [ { "first": "Liang", "middle": [], "last": "Huang", "suffix": "" } ], "year": 2008, "venue": "Proc. ACL-08: HLT", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Liang Huang. 2008. Forest reranking: Discriminative parsing with non-local features. In Proc. ACL-08: HLT.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "An empirical verification of coverage and correctness for a generalpurpose sentence generator", "authors": [ { "first": "Irene", "middle": [], "last": "Langkilde-Geary", "suffix": "" } ], "year": 2002, "venue": "Proc. INLG-02", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Irene Langkilde-Geary. 2002. An empirical verifi- cation of coverage and correctness for a general- purpose sentence generator. In Proc. INLG-02.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Probabilistic methods for disambiguation of an HPSG-based chart generator", "authors": [ { "first": "Hiroko", "middle": [], "last": "Nakanishi", "suffix": "" }, { "first": "Yusuke", "middle": [], "last": "Miyao", "suffix": "" }, { "first": "Jun'ichi", "middle": [], "last": "Tsujii", "suffix": "" } ], "year": 2005, "venue": "Proc. IWPT-05", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hiroko Nakanishi, Yusuke Miyao, and Jun'ichi Tsujii. 2005. Probabilistic methods for disambiguation of an HPSG-based chart generator. In Proc. IWPT-05.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "Stochastic natural language generation for spoken dialog systems", "authors": [ { "first": "Alice", "middle": [ "H" ], "last": "Oh", "suffix": "" }, { "first": "Alexander", "middle": [ "I" ], "last": "Rudnicky", "suffix": "" } ], "year": 2002, "venue": "Computer, Speech & Language", "volume": "16", "issue": "3/4", "pages": "387--407", "other_ids": {}, "num": null, "urls": [], "raw_text": "Alice H. Oh and Alexander I. Rudnicky. 2002. Stochastic natural language generation for spoken dialog systems. Computer, Speech & Language, 16(3/4):387-407.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "The proposition bank: A corpus annotated with semantic roles", "authors": [ { "first": "Martha", "middle": [], "last": "Palmer", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Gildea", "suffix": "" }, { "first": "Paul", "middle": [], "last": "Kingsbury", "suffix": "" } ], "year": 2005, "venue": "Computational Linguistics", "volume": "", "issue": "1", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Martha Palmer, Dan Gildea, and Paul Kingsbury. 2005. The proposition bank: A corpus annotated with se- mantic roles. Computational Linguistics, 31(1).", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "BLEU: a method for automatic evaluation of machine translation", "authors": [ { "first": "Kishore", "middle": [], "last": "Papineni", "suffix": "" }, { "first": "Salim", "middle": [], "last": "Roukos", "suffix": "" }, { "first": "Todd", "middle": [], "last": "Ward", "suffix": "" }, { "first": "Wei-Jing", "middle": [], "last": "Zhu", "suffix": "" } ], "year": 2002, "venue": "Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL), Philadelphia, PA.", "links": null }, "BIBREF26": { "ref_id": "b26", "title": "Head-Driven Phrase Structure Grammar", "authors": [ { "first": "Carl", "middle": [], "last": "Pollard", "suffix": "" }, { "first": "Ivan", "middle": [], "last": "Sag", "suffix": "" } ], "year": 1994, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Carl Pollard and Ivan Sag. 1994. Head-Driven Phrase Structure Grammar. University Of Chicago Press.", "links": null }, "BIBREF27": { "ref_id": "b27", "title": "Exploiting named entity classes in CCG surface realization", "authors": [ { "first": "Rajakrishnan", "middle": [], "last": "Rajkumar", "suffix": "" }, { "first": "Michael", "middle": [], "last": "White", "suffix": "" }, { "first": "Dominic", "middle": [], "last": "Espinosa", "suffix": "" } ], "year": 2009, "venue": "Proc. NAACL HLT 2009 Short Papers", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Rajakrishnan Rajkumar, Michael White, and Dominic Espinosa. 2009. Exploiting named entity classes in CCG surface realization. In Proc. NAACL HLT 2009 Short Papers.", "links": null }, "BIBREF28": { "ref_id": "b28", "title": "Linguistically informed statistical models of constituent structure for ordering in sentence realization", "authors": [ { "first": "Eric", "middle": [], "last": "Ringger", "suffix": "" }, { "first": "Michael", "middle": [], "last": "Gamon", "suffix": "" }, { "first": "Robert", "middle": [ "C" ], "last": "Moore", "suffix": "" }, { "first": "David", "middle": [], "last": "Rojas", "suffix": "" }, { "first": "Martine", "middle": [], "last": "Smets", "suffix": "" }, { "first": "Simon", "middle": [], "last": "Corston-Oliver", "suffix": "" } ], "year": 2004, "venue": "Proc. COLING-04", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Eric Ringger, Michael Gamon, Robert C. Moore, David Rojas, Martine Smets, and Simon Corston- Oliver. 2004. Linguistically informed statistical models of constituent structure for ordering in sen- tence realization. In Proc. COLING-04.", "links": null }, "BIBREF29": { "ref_id": "b29", "title": "Discriminative language modeling with conditional random fields and the perceptron algorithm", "authors": [ { "first": "Brian", "middle": [], "last": "Roark", "suffix": "" }, { "first": "Murat", "middle": [], "last": "Saraclar", "suffix": "" }, { "first": "Michael", "middle": [], "last": "Collins", "suffix": "" }, { "first": "Mark", "middle": [], "last": "Johnson", "suffix": "" } ], "year": 2004, "venue": "Proc. ACL-04", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Brian Roark, Murat Saraclar, Michael Collins, and Mark Johnson. 2004. Discriminative language modeling with conditional random fields and the perceptron algorithm. In Proc. ACL-04.", "links": null }, "BIBREF30": { "ref_id": "b30", "title": "The syntactic process", "authors": [ { "first": "Mark", "middle": [], "last": "Steedman", "suffix": "" } ], "year": 2000, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Mark Steedman. 2000. The syntactic process. MIT Press, Cambridge, MA, USA.", "links": null }, "BIBREF31": { "ref_id": "b31", "title": "SRILM -An extensible language modeling toolkit", "authors": [ { "first": "Andreas", "middle": [], "last": "Stolcke", "suffix": "" } ], "year": 2002, "venue": "Proc. ICSLP-02", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Andreas Stolcke. 2002. SRILM -An extensible lan- guage modeling toolkit. In Proc. ICSLP-02.", "links": null }, "BIBREF32": { "ref_id": "b32", "title": "Scalable discriminative learning for natural language parsing and translation", "authors": [ { "first": "Joseph", "middle": [], "last": "Turian", "suffix": "" }, { "first": "Benjamin", "middle": [], "last": "Wellington", "suffix": "" }, { "first": "I", "middle": [ "Dan" ], "last": "Melamed", "suffix": "" } ], "year": 2007, "venue": "Proc. NIPS 19", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Joseph Turian, Benjamin Wellington, and I. Dan Melamed. 2007. Scalable discriminative learn- ing for natural language parsing and translation. In Proc. NIPS 19.", "links": null }, "BIBREF33": { "ref_id": "b33", "title": "Maximum entropy models for realization ranking", "authors": [ { "first": "Erik", "middle": [], "last": "Velldal", "suffix": "" }, { "first": "Stephan", "middle": [], "last": "Oepen", "suffix": "" } ], "year": 2005, "venue": "Proc. MT Summit X", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Erik Velldal and Stephan Oepen. 2005. Maximum en- tropy models for realization ranking. In Proc. MT Summit X.", "links": null }, "BIBREF34": { "ref_id": "b34", "title": "Online large-margin training for statistical machine translation", "authors": [ { "first": "Taro", "middle": [], "last": "Watanabe", "suffix": "" }, { "first": "Jun", "middle": [], "last": "Suzuki", "suffix": "" }, { "first": "Hajime", "middle": [], "last": "Tsukada", "suffix": "" }, { "first": "Hideki", "middle": [], "last": "Isozaki", "suffix": "" } ], "year": 2007, "venue": "Proc. EMNLP-CoNLL-07", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Taro Watanabe, Jun Suzuki, Hajime Tsukada, and Hideki Isozaki. 2007. Online large-margin training for statistical machine translation. In Proc. EMNLP- CoNLL-07.", "links": null }, "BIBREF35": { "ref_id": "b35", "title": "BBN pronoun coreference and entity type corpus", "authors": [ { "first": "Ralph", "middle": [], "last": "Weischedel", "suffix": "" }, { "first": "Ada", "middle": [], "last": "Brunstein", "suffix": "" } ], "year": 2005, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ralph Weischedel and Ada Brunstein. 2005. BBN pro- noun coreference and entity type corpus. Technical report, BBN.", "links": null }, "BIBREF36": { "ref_id": "b36", "title": "A more precise analysis of punctuation for broadcoverage surface realization with CCG", "authors": [ { "first": "Michael", "middle": [], "last": "White", "suffix": "" }, { "first": "Rajakrishnan", "middle": [], "last": "Rajkumar", "suffix": "" } ], "year": 2008, "venue": "Proc. of the Workshop on Grammar Engineering Across Frameworks (GEAF08)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Michael White and Rajakrishnan Rajkumar. 2008. A more precise analysis of punctuation for broad- coverage surface realization with CCG. In Proc. of the Workshop on Grammar Engineering Across Frameworks (GEAF08).", "links": null }, "BIBREF37": { "ref_id": "b37", "title": "Towards broad coverage surface realization with CCG", "authors": [ { "first": "Michael", "middle": [], "last": "White", "suffix": "" }, { "first": "Rajakrishnan", "middle": [], "last": "Rajkumar", "suffix": "" }, { "first": "Scott", "middle": [], "last": "Martin", "suffix": "" } ], "year": 2007, "venue": "Proc. of the Workshop on Using Corpora for NLG: Language Generation and Machine Translation (UCNLG+MT)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Michael White, Rajakrishnan Rajkumar, and Scott Martin. 2007. Towards broad coverage surface re- alization with CCG. In Proc. of the Workshop on Using Corpora for NLG: Language Generation and Machine Translation (UCNLG+MT).", "links": null } }, "ref_entries": { "TABREF1": { "num": null, "text": "dcl \u2192 np s dcl \\np Rule + Word s dcl \u2192 np s dcl \\np + bought Rule + POS s dcl \u2192 np s dcl \\np + VBD Word-Word company, s dcl \u2192 np s dcl \\np, bought Word-POS company, s dcl \u2192 np s dcl \\np, VBD POS-Word NN, s dcl \u2192 np s dcl \\np, bought Word + \u2206w bought, s dcl \u2192 np s dcl \\np + dw POS + \u2206w VBD, s dcl \u2192 np s dcl \\np + dw Word + \u2206p bought, s dcl \u2192 np s dcl \\np + dp POS + \u2206p VBD, s dcl \u2192 np s dcl \\np + dp Word + \u2206v bought, s dcl \u2192 np s dcl \\np + dv POS + \u2206v VBD, s dcl \u2192 np s dcl \\np + dv", "content": "
Feature TypeExample
LexCat + Word s/s/np + before
LexCat + POSs/s/np + IN
Rules
", "html": null, "type_str": "table" }, "TABREF2": { "num": null, "text": "", "content": "
Model#Alph-feats #FeatsAccTime
full-model 2402173576176 96.40% 08:53
lp-ngram1127437342025 94.52% 05:19
lp-syn1274740291728 85.03% 05:57
: Basic and dependency features from
Clark & Curran's (2007b) normal form model;
distances are in intervening words, punctuation
marks and verbs, and are capped at 3, 3 and 2,
respectively
model by implementing Clark & Curran's (2007b)
normal form model in OpenCCG. 4 The features of
this model are listed in Table 1; they are integer-
valued, representing counts of occurrences in a
derivation. These syntactic features are quite com-
parable to the dominance-oriented features in the
union of the Velldal & Oepen and Nakanishi et
al. models, except that our feature set does not
include grandparenting, which has been found to
have limited utility in CCG parsing. Our syntac-
tic features also include ones that measure the dis-
tance between headwords in terms of intervening
words, punctuation marks or verbs; these features
generalize the ones in Nakanishi et al.'s model.
Note that in contrast to parsing, in realization dis-
tance features are non-local, since different partial
realizations in the same equivalence class typically
differ in word order; as we are working in a rerank-
ing paradigm though, the non-local nature of these
features is unproblematic.
Third, we include discriminative n-gram fea-
tures in our model, following Roark et al.'s (2004)
approach to discriminative n-gram modeling for
speech recognition. By discriminative n-gram fea-
tures, we mean features counting the occurrences
of each n-gram that is scored by our factored lan-
guage model, rather than a feature whose value is
the log prob determined by the language model.
As Roark et al. note, discriminative training with
n-gram features has the potential to learn to nega-
", "html": null, "type_str": "table" }, "TABREF3": { "num": null, "text": "Perceptron Training Details-number of features in the alphabet, number of features in the", "content": "
model, training accuracy and training time (hours)
for 10 iterations on a single commodity server
tively weight n-grams that appear in some of the
GEN(x i ) candidates, but which never appear in
the naturally occurring corpus used to train a stan-
dard, generative language model. Since our fac-
tored language model considers words, semantic
classes, part-of-speech tags and supertags, our n-
gram features represent a considerable generaliza-
tion of the sequence-oriented features in Velldal
& Oepen's model, which never contain more than
one word and do not include semantic classes.
", "html": null, "type_str": "table" }, "TABREF5": { "num": null, "text": "", "content": "
: Legend for Experimental Conditions
score, with the other models falling well below
the baseline (though faring better than the trigram-
word LM baseline). This result confirms the im-
portance of including n-gram log prob features in
discriminative realization ranking models, in line
with Velldal & Oepen's findings, and contra those
of Nakanishi et al., even though it was Nakanishi
et al. who experimented with the Penn Treebank
corpus, while Velldal & Oepen's experiments were
on a much smaller, limited domain corpus.
", "html": null, "type_str": "table" }, "TABREF7": { "num": null, "text": "", "content": "
: Section 00 Results (98.9% coverage)-
percentage of exact match and grammatically
complete realizations, BLEU scores and average
times, in seconds
Model%Exact %Complete BLEU
baseline33.7485.040.8173
full-model40.4583.880.8506
", "html": null, "type_str": "table" }, "TABREF8": { "num": null, "text": "Section 23 Results (97.06% coverage)", "content": "", "html": null, "type_str": "table" }, "TABREF9": { "num": null, "text": "Taipei 's growing number of small video-viewing parlors to pay ... baseline,syn-only,ngram-only that measure could compel Taipei 's growing number of video-viewing small parlors to ... lp-only, lp-ngram, full-model that measure could compel Taipei 's growing number of small video-viewing parlors to ... Ref-wsj 0024.2 Esso Australia Ltd. , a unit of new york-based Exxon Corp. , and Broken Hill Pty. operate the fields ... all except full-model Esso Australia Ltd. , a unit of new york-based Exxon Corp. , and Broken Hill Pty. operates the fields ... full-model Esso Australia Ltd. , a unit of new york-based Exxon Corp. , and Broken Hill Pty. operate the fields ... Ref-wsj 0034.9 they fell into oblivion after the 1929 crash . baseline, lp-ngram they fell after the 1929 crash into oblivion . lp-only, ngram-only, syn-only, full-model they fell into oblivion after the 1929 crash . Ref-wsj 0047.13 Antonio Novello , whom Mr. Bush nominated to serve as surgeon general , reportedly has assured . . . baseline,baseline-w3, lp-syn, lp-only Antonio Novello , which Mr. Bush nominated to serve as surgeon general , has reportedly assured . . . full-model, lp-ngram, syn-only, ngram-syn Antonio Novello , whom Mr. Bush nominated to serve as surgeon general , has reportedly assured . . .", "content": "
lists our results in the context of those re-
ported for other systems on PTB Section 23. The
", "html": null, "type_str": "table" }, "TABREF10": { "num": null, "text": "Examples of realized output", "content": "
SystemCoverage BLEU %Exact
Callaway (05)98.5%0.932157.5
OpenCCG (09)97.1%0.850640.5
Ringger et al. (04)100.0%0.83635.7
Langkilde-Geary (02) 83%0.75728.2
Guo et al. (08)100.0%0.744019.8
Hogan et al. (07)\u2248100.0% 0.6882
OpenCCG (08)96.0%0.670116.0
Nakanishi et al. (05)90.8%0.7733
", "html": null, "type_str": "table" }, "TABREF11": { "num": null, "text": "", "content": "
: PTB Section 23 BLEU scores and exact
match percentages in the NLG literature (Nakan-
ishi et al.'s results are for sentences of length 20 or
less)
", "html": null, "type_str": "table" } } } }