{ "paper_id": "W12-0117", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T04:13:49.753045Z" }, "title": "Using Sense-labeled Discourse Connectives for Statistical Machine Translation", "authors": [ { "first": "Thomas", "middle": [], "last": "Meyer", "suffix": "", "affiliation": { "laboratory": "", "institution": "Idiap Research Institute", "location": { "addrLine": "Rue Marconi 19", "postCode": "1920", "settlement": "Martigny", "country": "Switzerland" } }, "email": "thomas.meyer@idiap.ch" }, { "first": "Andrei", "middle": [], "last": "Popescu-Belis", "suffix": "", "affiliation": { "laboratory": "", "institution": "Idiap Research Institute", "location": { "addrLine": "Rue Marconi 19", "postCode": "1920", "settlement": "Martigny", "country": "Switzerland" } }, "email": "andrei.popescu-belis@idiap.ch" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "This article shows how the automatic disambiguation of discourse connectives can improve Statistical Machine Translation (SMT) from English to French. Connectives are firstly disambiguated in terms of the discourse relation they signal between segments. Several classifiers trained using syntactic and semantic features reach stateof-the-art performance, with F1 scores of 0.6 to 0.8 over thirteen ambiguous English connectives. Labeled connectives are then used into SMT systems either by modifying their phrase table, or by training them on labeled corpora. The best modified SMT systems improve the translation of connectives without degrading BLEU scores. A threshold-based SMT system using only high-confidence labels improves BLEU scores by 0.2-0.4 points.", "pdf_parse": { "paper_id": "W12-0117", "_pdf_hash": "", "abstract": [ { "text": "This article shows how the automatic disambiguation of discourse connectives can improve Statistical Machine Translation (SMT) from English to French. Connectives are firstly disambiguated in terms of the discourse relation they signal between segments. Several classifiers trained using syntactic and semantic features reach stateof-the-art performance, with F1 scores of 0.6 to 0.8 over thirteen ambiguous English connectives. Labeled connectives are then used into SMT systems either by modifying their phrase table, or by training them on labeled corpora. The best modified SMT systems improve the translation of connectives without degrading BLEU scores. A threshold-based SMT system using only high-confidence labels improves BLEU scores by 0.2-0.4 points.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Current approaches to Statistical Machine Translation (SMT) have difficulties in modeling longrange dependencies between words, including those that are due to discourse-level phenomena. Among these, discourse connectives are words that signal rhetorical relations between clauses or sentences. Their translation often depends on the exact relation signaled in context, a feature that current SMT systems were not designed to capture, hence their frequent mistranslations of connectives (see Section 2 below).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In this paper, we present a series of experiments that aim to use, in SMT systems, data with automatically labeled discourse connectives. Section 3 first presents the data sets used in our experiments. We designed classifiers that attempt to assign sense labels to ambiguous discourse connectives, and their scores compare favorably with the state-of-the-art for this task, as shown in Section 4. In particular, we consider WordNet relations and temporal expressions as well as candidate translations of connectives as additional features (Section 4.2).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "However, our main goal is not the disambiguation of connectives per se, but the use of the labels assigned to connectives as additional input to an SMT system. To the best of our knowledge, our experiments are the first attempts to combine connective disambiguation and SMT. Three solutions to this combination are compared in Section 5: modifying phrase tables, and training on data labeled manually, or automatically, with senses of connectives. We further show that a modified SMT system is best used when the confidence for a given label is high (Section 6). The paper concludes with a comparison to related work (Section 7) and an outline of future work (Section 8).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Discourse connectives such as although, however, since or while form a functional category of lexical items that are frequently used to mark coherence or discourse relations such as explanation, synchrony or contrast between units of text or discourse. For example, in the Europarl corpus from years 199x (Koehn, 2005) , the following nine lexical items, which are often (though not always) discourse connectives, are among the 400 most frequent tokens over a total of 12,846,003 (in parentheses, rank and number of occurrences): after (244th/6485), although (375th/4062), however (110th/12,857), indeed (334th/4486), rather (316th/4688), since (190th/8263), still (168th/9195), while (390th/3938), yet (331st/4532) -see also (Cartoni et al., 2011) . Discourse connectives can be difficult to translate, because many of them can signal different relations between clauses in different contexts. Moreover, if a wrong connective is used in translation, then a text becomes incoherent, as in the two examples below, taken from Europarl and translated (EN/FR) with Moses (Koehn et al., 2007) In the first example, the connective generated by SMT (m\u00eame si, literally \"even if\") signals a concession and not a contrast, for which the connective mais should have been used (as in the reference). In the second example, the connective depuis que (literally \"from the time\") generated by SMT expresses a temporal relation and not a causal one, which should have been conveyed e.g. by the French car.", "cite_spans": [ { "start": 305, "end": 318, "text": "(Koehn, 2005)", "ref_id": "BIBREF7" }, { "start": 726, "end": 748, "text": "(Cartoni et al., 2011)", "ref_id": "BIBREF2" }, { "start": 1061, "end": 1087, "text": "Moses (Koehn et al., 2007)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Discourse Connectives in Translation", "sec_num": "2" }, { "text": "Such examples suggest that the disambiguation of connectives prior to translation could help SMT systems to generate a correct connective in the target language. Of course, depending on the language pair, some ambiguities can be carried over from the source to the target language, so they need not be solved. Still, improving the overall translation of discourse connectives should increase the overall coherence of MT output, with a potential large impact on perceived quality.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discourse Connectives in Translation", "sec_num": "2" }, { "text": "For both tasks, the disambiguation of connectives and SMT, different training and testing data sets are available. This section shows how we made use of these resources and how we augmented them by manual and automated annotation of the senses of discourse connectives.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data Used in Our Experiments", "sec_num": "3" }, { "text": "One of the most important resources for discourse connectives in English is the Penn Discourse Treebank (Prasad et al., 2008) . The PDTB provides a discourse-layer annotation over the Wall Street Journal Corpus (WSJ) and the Penn Treebank syntactic annotation. The discourse annotation consists of manually annotated senses for about 100 types of explicit connectives, for implicit ones, and their clause spans. For the entire WSJ corpus of about 1,000,000 tokens there are 18,459 instances of annotated explicit connectives. The senses that discourse connectives can signal are organized in a hierarchy with 4 toplevel senses, followed by 16 subtypes on the second level and 23 detailed subsenses on the third level. Studies making use of the PDTB to build classifiers usually split the WSJ corpus into Sections 02-21 for training and Section 23 for testing (as we did for our disambiguation experiments, see Section 4). From the PDTB, we extracted the 13 most frequent and most ambiguous connectives: after, although, however, indeed, meanwhile, nevertheless, nonetheless, rather, since, still, then, while, and yet. This set shows in particular that connectives signaling contrastive or temporal senses are the most ambiguous ones, hence they are also potentially difficult to translate, as this ambiguity is often not preserved across languages (Danlos and Roze, 2011) . We used the senses from the second PDTB hierarchy level (as the third level is too fine-grained for EN/FR translation) and generated the training and testing sets listed with statistics in Table 1 (Section 4).", "cite_spans": [ { "start": 104, "end": 125, "text": "(Prasad et al., 2008)", "ref_id": "BIBREF18" }, { "start": 1349, "end": 1372, "text": "(Danlos and Roze, 2011)", "ref_id": "BIBREF4" } ], "ref_spans": [ { "start": 1564, "end": 1571, "text": "Table 1", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Data for the Disambiguation of Discourse Connectives", "sec_num": "3.1" }, { "text": "In principle, classifiers trained on PDTB data can be applied directly to label connectives over the English side of the Europarl corpus (Koehn, 2005) used for training and testing SMT. However, to control the difference in register from newswire texts to formal political speech, and to allow for future studies of other languages, we also performed manual annotation (Cartoni et al., 2011) of five connectives over the Europarl corpus (although, even though, since, though and while).", "cite_spans": [ { "start": 137, "end": 150, "text": "(Koehn, 2005)", "ref_id": "BIBREF7" }, { "start": 369, "end": 391, "text": "(Cartoni et al., 2011)", "ref_id": "BIBREF2" } ], "ref_spans": [], "eq_spans": [], "section": "Data for the Disambiguation of Discourse Connectives", "sec_num": "3.1" }, { "text": "The manual annotation was performed on subsets of Europarl v5 (years 199x) for the first few hundred occurrences of each connective. Instead of a potentially difficult and costly annotation of senses, as in the PDTB, we performed translation spotting, asking annotators to highlight the translation of each of the five connectives in the French side of the corpus. From the list of all observed translations one can then cluster the necessary sense labels, as some target language connectives clearly signal only one sense or, in cases where ambiguity is preserved, one can group the equally ambiguous connectives under one composite label. For example, while is sometimes translated to the French discourse connectives tandis que or alors que which both preserve the ambiguity of while signaling a temporal or contrastive sense. With this method we built the data sets listed with statistics in Table 2 below (Section 4).", "cite_spans": [], "ref_spans": [ { "start": 896, "end": 903, "text": "Table 2", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "Data for the Disambiguation of Discourse Connectives", "sec_num": "3.1" }, { "text": "The translation data for our SMT experiments has been often used in other MT research work and is freely distributed for the shared tasks of the Workshop on Machine Translation (WMT) 1 .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data for Statistical Machine Translation", "sec_num": "3.2" }, { "text": "For training our SMT systems, the EN/FR Europarl corpus v5 was used in three ways to integrate data with labeled discourse connectives into SMT: no changes (for MT phrase table modifications), integration of manually annotated data and integration of automatically labeled data. These methods are described below in Section 5 -here, we gather descriptions of the corresponding data. These test sets might appear small compared to the amount of data normally used for SMT system testing. In our system evaluation however, apart from automated scoring, we also had to perform manual counts of improved translations, which is why we could not evaluate more than a hundred sentences (Section 5). When counting manually for test set (i), it was downsampled to the same amount of 35 and 62 sentences as for sets (g) and (h), by extracting the first occurrences of each connective.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Data for Statistical Machine Translation", "sec_num": "3.2" }, { "text": "In all experiments, we use the Moses Phrasebased SMT decoder (Koehn et al., 2007) specialized classifier, by using the Stanford Maximum Entropy classifier package (Manning and Klein, 2003) . Maximum Entropy is known to handle discrete features well and has been applied successfully to connective disambiguation before (see Section 7).", "cite_spans": [ { "start": 61, "end": 81, "text": "(Koehn et al., 2007)", "ref_id": "BIBREF7" }, { "start": 163, "end": 188, "text": "(Manning and Klein, 2003)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Data for Statistical Machine Translation", "sec_num": "3.2" }, { "text": "An initial set of features can directly be obtained from the PDTB (and must hence be considered as oracle features): the (capitalized) connective token, its POS tag, first word of clause 1, last word of clause 1, first word of clause 2 (the one containing the explicit connective), last word of clause 2, POS tag of the first word of clause 2, type of first word of clause 2, parent syntactical categories of the connective, punctuation pattern of the sentences. Apart from these standard features in discourse connective disambiguation we used WordNet (Miller, 1995) to compute lexical similarity scores with the lesk metric (Banerjee and Pedersen, 2002) for all the possible combinations of nouns, verbs and adjectives in the two clauses, as well as antonyms found for these word groups. In addition, we used features that are likely to help detecting temporal relations and were obtained from the Tarsqi Toolkit (Verhagen and Pustejovsky, 2008) , which annotates English sentences automatically with the TimeML annotation language for temporal expressions. For example, in the sentence The crimes may appear small, but the prices can be huge (PDTB Section 2, WSJ file 0290), for example, our features would indicate the antonyms small vs. huge that signal the contrast, along with a temporal ordering of the event appear before the event can.", "cite_spans": [ { "start": 553, "end": 567, "text": "(Miller, 1995)", "ref_id": "BIBREF12" }, { "start": 915, "end": 947, "text": "(Verhagen and Pustejovsky, 2008)", "ref_id": "BIBREF19" } ], "ref_spans": [], "eq_spans": [], "section": "Data for Statistical Machine Translation", "sec_num": "3.2" }, { "text": "We report the classifier performances as F1 scores for each connective (weighting precision and recall equally) in Table 1 , testing on Section 23 of the PDTB. This sense classifier will be referred to as Classifier PT in the rest of the paper, in particular when used for the SMT experiments.", "cite_spans": [], "ref_spans": [ { "start": 115, "end": 122, "text": "Table 1", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Data for Statistical Machine Translation", "sec_num": "3.2" }, { "text": "In an attempt to improve Classifier PT, we added a new type of feature, resulting in Classifier PT+.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Classifier PT+: With Candidate Translations as Features", "sec_num": "4.2" }, { "text": "Namely, we used candidate translations of discourse connectives from a baseline SMT system (not adapted to connectives). To find these values, a Moses baseline decoder was used to translate the PDTB data, which was then word-aligned (En- glish source with target French) by using GIZA++ (Och and Ney, 2003) . In this alignment, we searched for the translation equivalents of the 13 PDTB connectives by using a hand-crafted dictionary of possible French translations. When the translation candidate is not ambiguous -e.g. bien que as a translation for while clearly signals a concession -its specific sense label was added as the value of an additional feature. In some cases, however, the values of the features are not determined (and are set to NONE): either when the SMT system or GIZA++ failed in translating or aligning a connective, or when the target connective was just as ambiguous as the source one (e.g. while translated as tandis que, which can be labeled both temporal or contrast). Overall, this procedure led to an accuracy gain of Classifier PT+ with respect to Classifier PT of about 0.1 to 0.6 F1 score for some of the connectives, as can be seen in the last column of Table 1 .", "cite_spans": [ { "start": 287, "end": 306, "text": "(Och and Ney, 2003)", "ref_id": "BIBREF14" } ], "ref_spans": [ { "start": 1187, "end": 1194, "text": "Table 1", "ref_id": "TABREF3" } ], "eq_spans": [], "section": "Classifier PT+: With Candidate Translations as Features", "sec_num": "4.2" }, { "text": "As explained in Section 3.1, we performed manual annotation of connective senses in Europarl as well, to provide labeled instances directly in the data used for SMT training and to account for the register change. For the Europarl data sets, we built a new MaxEnt classifier (called Classifier EU) using the same feature set as Classifier PT. However, all features were this time extracted automatically (no oracle). In particular, we used Charniak and Johnson's (2005) parser to then extract the syntactic features. In Table 2 , we report the results of Classifier EU, again in terms of F1 scores. For all three classifiers, PT, PT+ and EU, the F1 scores are in a range of 0.6 and 0.8, thus comparing favorably to the state-of-the-art for discourse connective disambiguation with detailed senses (Section 7). Classifier EU also compares favorably to PT and PT+, as seen for instance for since (0.80 vs. 0.78) or although (0.67 vs. 0.60-0.66).", "cite_spans": [ { "start": 440, "end": 469, "text": "Charniak and Johnson's (2005)", "ref_id": "BIBREF3" } ], "ref_spans": [ { "start": 520, "end": 527, "text": "Table 2", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "Classifier EU: Trained on Europarl Data", "sec_num": "4.3" }, { "text": "In this section, we report on experiments that study the effect of discourse connective labeling on SMT. The experiments differ with respect to the method used for taking advantage of the labels, but also with respect to the data sets and the sense classifiers that are used.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Use of Labeled Connectives for SMT", "sec_num": "5" }, { "text": "The variation in MT quality can be estimated in several ways. On the one hand, we use the BLEU metric (Papineni et al., 2002) with one reference translation as is most often done in current SMT research 2 . To improve confidence in the BLEU scores, especially when test sets are small, we also compute BLEU scores using bootstrapping of data sets (Zhang and Vogel, 2010) ; the test sets are re-sampled a thousand times and the average BLEU score is computed from individual sample scores. The BLEU approach is not likely, however, to be sensitive enough to the small differences due to the correction of discourse connectives (less than one word per sentence). We therefore additionally resort to a manual evaluation metric, referred to as \u2206Connectives, which counts the occurrences of connectives that are better translated by our modified systems compared to the baseline ones. Table 3 : MT systems dealing with manually and automatically (PT, PT+, EU) sense-labeled connectives: BLEU scores (including bootstrapped ones) and variation in the translation of individual connectives (\u2206Connectives, as a percentage). The description of each condition and the baseline BLEU scores are in the text of the article.", "cite_spans": [ { "start": 102, "end": 125, "text": "(Papineni et al., 2002)", "ref_id": "BIBREF16" }, { "start": 347, "end": 370, "text": "(Zhang and Vogel, 2010)", "ref_id": "BIBREF22" } ], "ref_spans": [ { "start": 880, "end": 887, "text": "Table 3", "ref_id": null } ], "eq_spans": [], "section": "Evaluation Metrics for MT", "sec_num": "5.1" }, { "text": "A first way of using labeled connectives is to modify the phrase table of an SMT system previously trained/tuned on data sets (a)/(d) from Section 3.2, in order to force it to translate each specific sense of a discourse connective (as indicated by its label) with an acceptable equivalent selected among those learned from the training data. Of course, this only handles cases when connectives are translated by explicit lexical items (typically, target connectives) and not by more complex grammatical constructs. The phrase table modification is done as follows. Based on a small dictionary of the five connective types of Table 2 , their acceptable French equivalents and the possible senses, the initial phrase table is searched for phrases containing a connective and each occurrence is inspected to find out which sense is reflected in the translation. If the sense is non-ambiguous, then the table entry is modified to include the label, and the probability score is set to 1 in order to maximize the chance that the respective translation is found during decoding. For instance, for every phrase table entry where while is translated as alors que, this corresponds to a contrastive use and while is changed into while CONTRAST. Or, for the entries where while is translated as bien que, the lexical entry is changed into while CONCESSION. However, when the source entry is as ambiguous as the target one, no modification is made. This means that during decoding (testing) with labeled sentences, these entries will never be used.", "cite_spans": [], "ref_spans": [ { "start": 626, "end": 633, "text": "Table 2", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "Phrase Table Modification", "sec_num": "5.2" }, { "text": "The results of the SMT system are shown in experiments 1 and 2 in Table 3 , respectively test-ing over data set (g) (7 manually annotated sentences for each of the 5 connectives) and over set (i), in which the 5 connectives were automatically labeled with Classifier EU. In the first test, the translations of 29% of the connectives are improved by the modified system, while 20% are degraded and 51% remain unchanged -thus reflecting an overall 10% improvement in the translations of connectives (\u2206Connectives). However, for this test set, the BLEU score is about 3 points below the baseline SMT system that used the same phrase table without modification of labels and scores (not shown in Table 3 ). In experiment 2, however, the BLEU score of the modified system is in the same range as the baseline one (22.13 vs. 22.76) . As for \u2206Connectives, as it was not possible to score manually all the 10,311 connectives, we sampled 35 sentences and found that 34% of the connectives are improved, 20% are degraded and 46% remain unchanged, again reflecting an improvement in the translation of connectives. This shows that piping automatic labeling and SMT with a modified phrase table does not degrade the overall BLEU score, while increasing \u2206Connectives.", "cite_spans": [ { "start": 808, "end": 825, "text": "(22.13 vs. 22.76)", "ref_id": null } ], "ref_spans": [ { "start": 66, "end": 73, "text": "Table 3", "ref_id": null }, { "start": 692, "end": 699, "text": "Table 3", "ref_id": null } ], "eq_spans": [], "section": "Phrase Table Modification", "sec_num": "5.2" }, { "text": "We explored a more principled way to integrate external labels into SMT, by using labeled data (manually or automatically) for training, so that the system directly learns a modified phrase table which allows the translation of labeled data (automatically) when testing.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Training on Tagged Corpora", "sec_num": "5.3" }, { "text": "We report first two experiments using the manual gold annotation for the five connective types over Europarl excerpts, used for training. When used also for testing (experiment 3 in Table 3) , this can be seen as an oracle experiment, measuring the translation improvement when connective sense labeling is perfect. However, in experiment 4, the SMT system uses the output of an automatic labeler. For training/tuning we used data sets (b)/(e), Section 3.2.", "cite_spans": [], "ref_spans": [ { "start": 182, "end": 190, "text": "Table 3)", "ref_id": null } ], "eq_spans": [], "section": "Manual Gold Annotation", "sec_num": "5.3.1" }, { "text": "In experiment 3, for test set (g), 32% of the connectives were translated better by the modified system, 57% remained the same, and 11% were degraded. In experiment 4, over a 35 sentence sample of the bigger test set (i), 26% were improved, 66% remained the same, and only 8% were degraded. The baseline SMT system (not shown in Table 3 ) was built with the same amounts of unlabeled training and tuning data. Overall, the BLEU scores of our modified systems are similar to the baseline ones, though still lower -41.58 vs. 42.77 for experiment 3, and 22.43 vs. 22.76 for experiment 4, also confirmed by the bootstrapped scores.", "cite_spans": [], "ref_spans": [ { "start": 329, "end": 336, "text": "Table 3", "ref_id": null } ], "eq_spans": [], "section": "Manual Gold Annotation", "sec_num": "5.3.1" }, { "text": "Another comparison shows that the system trained on manual annotations (exp. 4) outperforms the system using a modified phrase table (exp. 2) in terms of BLEU scores (22.43 vs. 22.13) and bootstrapped ones (24.00 vs. 23.63).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Manual Gold Annotation", "sec_num": "5.3.1" }, { "text": "We evaluated an SMT system trained on data that was automatically labeled using the classifiers in Section 4. This method provides a large amount of imperfect training data, and uses no manual annotations at all, except for the initial training of the classifiers. For these experiments (5 and 6 in Table 3 ), the BLEU scores as well as the manual counts of improved connectives are lower than in the preceding experiments because, overall, less training/tuning data was used -about 15% of Europarl, data sets (c) and (f) in Section 3.2. The baseline system was built over the same amount of data, with no labels.", "cite_spans": [], "ref_spans": [ { "start": 299, "end": 306, "text": "Table 3", "ref_id": null } ], "eq_spans": [], "section": "Automated Annotation", "sec_num": "5.3.2" }, { "text": "Testing here was performed over the slightly bigger test set (h) with 62 sentences (13 connective types). The occurrences were tagged with Classifier PT prior to translation (exp. 5). Compared to the baseline system, the translations of 16% of the connectives were improved, while 60% remained the same and 24% were degraded. In experiment 6, the 10,311 UN occurrences for 5 connective types were first tagged with Classifier EU. Evaluated on a sample of 62 sentences, 16% of the connectives were improved, while 66% remained the same and 18% were degraded. Despite less training data, in terms of BLEU, the difference to the respective baseline system (scores not shown in Table 3 ) is similar in both experimental settings: 19.78 vs. 20.11 for experiment 6 (automated annotation), compared to 22.43 vs. 22.76 for experiment 4 (manual annotation).", "cite_spans": [], "ref_spans": [ { "start": 674, "end": 681, "text": "Table 3", "ref_id": null } ], "eq_spans": [], "section": "Automated Annotation", "sec_num": "5.3.2" }, { "text": "Finally, we carried out two experiments (7 and 8) with Classifier PT+, which uses as additional features the translation candidates and has a higher accuracy than PT (Section 4.2). As a result, the translation of connectives (\u2206Connectives) is indeed improved compared (respectively) to experiments 5 and 6, as it appears from lines 7-8 of Table 3 . Also, the BLEU scores of the corresponding SMT systems are increased in experiments 7 vs. 5 and in 8 vs. 6, and are now equal to the baseline ones (for experiment 8: 20.14 vs. 20.11, or, bootstrapped, 21.55 vs. 21.55) .", "cite_spans": [ { "start": 496, "end": 566, "text": "(for experiment 8: 20.14 vs. 20.11, or, bootstrapped, 21.55 vs. 21.55)", "ref_id": null } ], "ref_spans": [ { "start": 339, "end": 346, "text": "Table 3", "ref_id": null } ], "eq_spans": [], "section": "Automated Annotation", "sec_num": "5.3.2" }, { "text": "The results of experiments 7/8 vs. 5/6 indicate that improved classifiers for connectives also improve SMT output as measured by \u2206Connectives, with BLEU remaining fairly constant, and therefore are worth investigating in more depth in the future. When comparing manual (experiments 3/4) vs. automated annotation (experiments 5/6/7/8) and their use in SMT, the differences in the scores (BLEU and \u2206Connectives) highlight a trade-off: manually annotated data used for training leads to better scores, but noisier and larger training data that is annotated automatically is an acceptable solution when manual annotations are not available.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Automated Annotation", "sec_num": "5.3.2" }, { "text": "As shown with the above experiments, the accuracy of the connective classifiers influences SMT quality. We therefore hypothesize that an SMT system dealing with labeled connectives would best be used when the confidence of the classifier is high, while a generic SMT system could be used for lower confidence values.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Classifier Confidence Scores", "sec_num": "6" }, { "text": "We experimented with the confidence scores of Classifier EU, which assigns a score between 0 and 1 to each of its decisions on the connectives' labels. (All processing is automatic in these ex- Figure 1 : Use of a combined system (COMB) that directs the input sentences either to a system trained on a senselabeled corpus (TTC) or to a baseline one (BASE), depending on the confidence of the connective classifier. The x-axis shows the threshold above which TTC is used -BASE being used below it -and the y-axis shows the BLEU scores of COMB with respect to TTC and BASE. periments, and the evaluation is done solely in terms of BLEU). We defined a threshold-based procedure to combine SMT systems: if the confidence for a sense label is above a certain threshold, then the sentence is translated by an SMT system trained on labeled data from experiment 4 (or \"tagged corpus\", hence noted TTC), and if it is below the threshold, it is sent to a baseline system (noted BASE). The resulting BLEU scores of the combined system (COMB) obtained for various threshold values are shown in Figure 1 for two connectives.", "cite_spans": [], "ref_spans": [ { "start": 194, "end": 202, "text": "Figure 1", "ref_id": null }, { "start": 1082, "end": 1090, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Classifier Confidence Scores", "sec_num": "6" }, { "text": "Firstly, we considered all the 1,572 sentences from the UN corpus which contained the connective although, labeled either as contrast or concession. We show BLEU scores of the COMB system for several thresholds in the interval of observed confidence scores, along with the scores of BASE and TTC, in Figure 1(a) . The results show that the scores of COMB increase with the value of the threshold, and that for at least one value of the threshold (0.95) COMB outperforms both TTC and BASE by 0.20 BLEU points.", "cite_spans": [], "ref_spans": [ { "start": 300, "end": 311, "text": "Figure 1(a)", "ref_id": null } ], "eq_spans": [], "section": "Classifier Confidence Scores", "sec_num": "6" }, { "text": "To confirm this finding with another connective, we took the first 1,572 sentences containing the connective since from the UN corpus. The BLEU scores for COMB are shown for the range of observed confidence values (0.4-1.0) in Figure 1(b) . For several values of the threshold, COMB outperforms both BASE and TTC, in par-ticular for 0.85, with a difference of 0.39 BLEU points.", "cite_spans": [], "ref_spans": [ { "start": 227, "end": 238, "text": "Figure 1(b)", "ref_id": null } ], "eq_spans": [], "section": "Classifier Confidence Scores", "sec_num": "6" }, { "text": "The significance of the observed improvement was tested as follows. For each of the two connectives, we split the test sets of 1,572 sentences each in five folds, and compared for each fold the scores of COMB for the best performing threshold (0.95 or 0.85) with the highest of BASE or TTC (i.e. BASE for although and TTC for since). We performed a paired t-test to compute the significance of the difference, and found p = 0.12 for although. This value, although slightly above the conventional boundary of 0.1, shows that the five pairs of scores reflect a significant difference in quality. Similarly, when performing a t-test for since, the difference in scores is found significant at the 0.01 level (p = 0.005). Of course, COMB is always significantly better than the lower of BASE or TTC (p < 0.05). In the future, the system combination will be tested for all connectives, and the respective values of the thresholds will be set on tuning, not on test data.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Classifier Confidence Scores", "sec_num": "6" }, { "text": "Discourse parsing (Marcu, 2000) has proven to be a difficult task, even when complex models (CRFs, SVMs) are used (Wellner, 2009; Hernault et al., 2010) . The performance of discourse parsers is in a range of 0.4 to 0.6 F1 score.", "cite_spans": [ { "start": 18, "end": 31, "text": "(Marcu, 2000)", "ref_id": "BIBREF11" }, { "start": 114, "end": 129, "text": "(Wellner, 2009;", "ref_id": "BIBREF21" }, { "start": 130, "end": 152, "text": "Hernault et al., 2010)", "ref_id": "BIBREF6" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "7" }, { "text": "With the release of the PDTB, recent research focused on the disambiguation of discourse connectives as a task in its own right. For the disambiguation of explicit connectives, the state-of-theart performance for labeling all types of connectives in English is quite high. In the PDTB data, the disambiguation of discourse vs. non-discourse uses of connectives reaches 97% accuracy (Lin et al., 2010) . The labeling of the four main senses from the PDTB sense hierarchy (temporal, contingency, comparison, expansion) reaches 94% accuracy (Pitler and Nenkova, 2009) -however, the baseline accuracy is already around 85% when using only the connective token as a feature. Various methods for classification and feature analysis have been proposed (Wellner et al., 2006; Elwell and Baldridge, 2008) . Other studies have focused on the analysis of highly ambiguous discourse connectives only. Miltsakaki et al. (2005) report classification results for the connectives since, while and when. Using a Maximum Entropy classifier, they reach 75.5% accuracy for since, 71.8% for while and 61.6% for when. As the PDTB was not completed at that time, the data sets and labels are not exactly identical to the ones that we used above (see Section 4).", "cite_spans": [ { "start": 382, "end": 400, "text": "(Lin et al., 2010)", "ref_id": "BIBREF8" }, { "start": 538, "end": 564, "text": "(Pitler and Nenkova, 2009)", "ref_id": "BIBREF17" }, { "start": 745, "end": 767, "text": "(Wellner et al., 2006;", "ref_id": "BIBREF20" }, { "start": 768, "end": 795, "text": "Elwell and Baldridge, 2008)", "ref_id": "BIBREF5" }, { "start": 889, "end": 913, "text": "Miltsakaki et al. (2005)", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "7" }, { "text": "The disambiguation of senses signaled by discourse connectives can be seen as a word sense disambiguation (WSD) problem for functional words (as opposed to WSD for content words, which is more frequently studied). The integration of WSD into SMT has especially been studied by Carpuat and Wu (2007) , who used the translation candidates output by a baseline SMT system as word sense labels. This is similar to our use of translation candidates as an additional feature for classification in Section 4.2. Then, the output of several classifiers based on linguistic features was weighed against the translation candidates output by the baseline SMT system. With this procedure, their WSD+SMT system improved the BLEU scores by 0.4-0.5 for the English/Chinese pair. Chang et al. (2009) use a LogLinear classifier with linguistic features in order to disambiguate the Chinese particle 'DE' that has five different context-dependent uses (modifier, preposition, relative clause etc.). When the classifier is used to annotate the particle prior to SMT, the output of the translation system improves by up to 1.49 BLEU score for phrase-based Chinese to English translation. Ma et al. (2011) use a Maximum Entropy model to POS tag English collocational particles (e.g. come down/by, turn against, inform of ) more specifically than a usual POS tagger does (where only one label is given to all particles). The authors claim the usefulness of such a particle tagger for English/Chinese translation, but do not show its actual integration into an MT system.", "cite_spans": [ { "start": 277, "end": 298, "text": "Carpuat and Wu (2007)", "ref_id": "BIBREF1" }, { "start": 763, "end": 782, "text": "Chang et al. (2009)", "ref_id": "BIBREF2" }, { "start": 1167, "end": 1183, "text": "Ma et al. (2011)", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "7" }, { "text": "These approaches, as well as ours, show that integrating discourse information into SMT is promising and deserves future examination. The disambiguation of word senses, including function words, can improve SMT output when the senses are annotated in a pre-processing step that uses classifiers based on linguistic features at the semantic and discourse levels, which are not available to a state-of-the-art SMT systems.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Related Work", "sec_num": "7" }, { "text": "This paper has presented methods and results for the disambiguation of temporal and contrastive discourse connectives using MaxEnt classifiers with syntactic and semantic features, in English texts, in terms of senses intended to help SMT. These classifiers have been used to perform experiments with connective-annotated data applied to EN/FR SMT systems. The results have shown an improvement in the translation of connectives for fully automatic systems trained on either handlabeled or automatically-labeled data. Moreover, BLEU scores were significantly improved by 0.2-0.4 when such systems were only used for connectives that had been disambiguated with high confidence.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion and Future Work", "sec_num": "8" }, { "text": "In future work we plan to improve the sense classifiers using additional features, to improve their integration with SMT, and to unify our data sets through additional manual annotations over Europarl. The applicability of the method to other languages will also be demonstrated experimentally.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion and Future Work", "sec_num": "8" }, { "text": "The scores are generated by the NIST MTeval script version 11b, available from www.itl.nist.gov/iad/ mig/tools/.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "We are grateful for the funding of this work to the Swiss National Science Foundation (SNSF), under the COMTIS Sinergia Project n. CRSI22 127510 (see www.idiap.ch/comtis/).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgments", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "An Adapted Lesk Algorithm for Word Sense Disambiguation Using WordNet", "authors": [ { "first": "Satanjeev", "middle": [], "last": "Banerjee", "suffix": "" }, { "first": "Ted", "middle": [], "last": "Pedersen", "suffix": "" } ], "year": 2002, "venue": "Computational Linguistics and Intelligent Text Processing", "volume": "2276", "issue": "", "pages": "117--171", "other_ids": {}, "num": null, "urls": [], "raw_text": "Satanjeev Banerjee and Ted Pedersen. 2002. An Adapted Lesk Algorithm for Word Sense Disam- biguation Using WordNet. In Alexander Gel- bukh, editor, Computational Linguistics and Intelli- gent Text Processing, LNCS 2276, pages 117-171. Springer, Berlin/Heidelberg.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Improving Statistical Machine Translation using Word Sense Disambiguation", "authors": [ { "first": "Marine", "middle": [], "last": "Carpuat", "suffix": "" }, { "first": "Dekai", "middle": [], "last": "Wu", "suffix": "" } ], "year": 2007, "venue": "Proc. of EMNLP-CoNLL", "volume": "", "issue": "", "pages": "61--72", "other_ids": {}, "num": null, "urls": [], "raw_text": "Marine Carpuat and Dekai Wu. 2007. Improving Sta- tistical Machine Translation using Word Sense Dis- ambiguation. Proc. of EMNLP-CoNLL, pages 61- 72, Prague.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "How Comparable are Parallel Corpora? Measuring the Distribution of General Vocabulary and Connectives", "authors": [ { "first": "Bruno", "middle": [], "last": "Cartoni", "suffix": "" }, { "first": "Sandrine", "middle": [], "last": "Zufferey", "suffix": "" }, { "first": "Thomas", "middle": [], "last": "Meyer", "suffix": "" }, { "first": "Andrei", "middle": [], "last": "Popescu-Belis", "suffix": "" }, { "first": ".", "middle": [ ";" ], "last": "Pi-Chuan Chang", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Jurafsky", "suffix": "" }, { "first": "Christopher", "middle": [ "D" ], "last": "Manning", "suffix": "" } ], "year": 2009, "venue": "Proc. of the Fourth Workshop on Statistical Machine Translation at EACL-2009", "volume": "", "issue": "", "pages": "78--86", "other_ids": {}, "num": null, "urls": [], "raw_text": "Bruno Cartoni, Sandrine Zufferey, Thomas Meyer, and Andrei Popescu-Belis. 2011. How Comparable are Parallel Corpora? Measuring the Distribution of General Vocabulary and Connectives. Proc. of the 4th Workshop on Building and Using Comparable Corpora (BUCC), pages 78-86, Portland, OR. Pi-Chuan Chang, Dan Jurafsky, and Christopher D. Manning. 2009. Disambiguating 'DE' for Chinese- English Machine Translation. Proc. of the Fourth Workshop on Statistical Machine Translation at EACL-2009, Athens.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Coarseto-fine n-best Parsing and MaxEnt Discriminative Reranking", "authors": [ { "first": "Eugene", "middle": [], "last": "Charniak", "suffix": "" }, { "first": "Mark", "middle": [], "last": "Johnson", "suffix": "" } ], "year": 2005, "venue": "Proc. of the 43rd Annual Meeting of the ACL", "volume": "", "issue": "", "pages": "173--180", "other_ids": {}, "num": null, "urls": [], "raw_text": "Eugene Charniak and Mark Johnson. 2005. Coarse- to-fine n-best Parsing and MaxEnt Discriminative Reranking. Proc. of the 43rd Annual Meeting of the ACL, pages 173-180, Ann Arbor, MI.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Traduction (Automatique) des Connecteurs de Discours", "authors": [ { "first": "Laurence", "middle": [], "last": "Danlos", "suffix": "" }, { "first": "Charlotte", "middle": [], "last": "Roze", "suffix": "" } ], "year": 2011, "venue": "Actes 18e Conf\u00e9rence sur le Traitement Automatique des Langues Naturelles (TALN)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Laurence Danlos and Charlotte Roze. 2011. Traduc- tion (Automatique) des Connecteurs de Discours. Actes 18e Conf\u00e9rence sur le Traitement Automa- tique des Langues Naturelles (TALN), Montpellier.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Discourse Connective Argument Identification with Connective Specific Rankers", "authors": [ { "first": "Robert", "middle": [], "last": "Elwell", "suffix": "" }, { "first": "Jason", "middle": [], "last": "Baldridge", "suffix": "" } ], "year": 2008, "venue": "Proc. of the 2nd IEEE International Conference on Semantic Computing (ICSC)", "volume": "", "issue": "", "pages": "198--205", "other_ids": {}, "num": null, "urls": [], "raw_text": "Robert Elwell and Jason Baldridge. 2008. Discourse Connective Argument Identification with Connec- tive Specific Rankers. Proc. of the 2nd IEEE International Conference on Semantic Computing (ICSC), pages 198-205, Santa Clara, CA.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "HILDA: A Discourse Parser using Support Vector Machine classification", "authors": [ { "first": "Hugo", "middle": [], "last": "Hernault", "suffix": "" }, { "first": "Helmut", "middle": [], "last": "Prendinger", "suffix": "" }, { "first": "David", "middle": [ "A" ], "last": "", "suffix": "" }, { "first": "Mitsuru", "middle": [], "last": "Ishizuka", "suffix": "" } ], "year": 2010, "venue": "Dialogue and Discourse", "volume": "3", "issue": "1", "pages": "1--33", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hugo Hernault, Helmut Prendinger, David A. duVerle, and Mitsuru Ishizuka. 2010. HILDA: A Discourse Parser using Support Vector Machine classification. Dialogue and Discourse, 3(1):1-33.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Europarl: A Parallel Corpus for Statistical Machine Translation", "authors": [ { "first": "", "middle": [], "last": "Philipp Koehn", "suffix": "" } ], "year": 2005, "venue": "Proc. of 45th Annual Meeting of the ACL, Demonstration Session", "volume": "", "issue": "", "pages": "79--86", "other_ids": {}, "num": null, "urls": [], "raw_text": "Philipp Koehn, et al. 2007. Moses: Open Source Toolkit for Statistical Machine Translation. Proc. of 45th Annual Meeting of the ACL, Demonstration Session, pages 177-180, Prague. Philipp Koehn. 2005. Europarl: A Parallel Corpus for Statistical Machine Translation. Proc. of MT Sum- mit X, pages 79-86, Phuket.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "A PDTB-styled End-to-end Discourse Parser", "authors": [ { "first": "Ziheng", "middle": [], "last": "Lin", "suffix": "" }, { "first": "Min-Yen", "middle": [], "last": "Hwee Tou Ng", "suffix": "" }, { "first": "", "middle": [], "last": "Kan", "suffix": "" } ], "year": 2010, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ziheng Lin, Hwee Tou Ng, and Min-Yen Kan. 2010. A PDTB-styled End-to-end Discourse Parser. Tech- nical Report TRB8/10, School of Computing, Na- tional University of Singapore.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "POS Tagging of English Particles for Machine Translation", "authors": [ { "first": "Jianjun", "middle": [], "last": "Ma", "suffix": "" }, { "first": "Degen", "middle": [], "last": "Huang", "suffix": "" }, { "first": "Haixia", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Wenfeng", "middle": [], "last": "Sheng", "suffix": "" } ], "year": 2011, "venue": "Proc. of MT Summit XIII", "volume": "", "issue": "", "pages": "57--63", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jianjun Ma, Degen Huang, Haixia Liu, and Wenfeng Sheng. 2011. POS Tagging of English Particles for Machine Translation. Proc. of MT Summit XIII, pages 57-63, Xiamen.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Optimization, MaxEnt Models, and Conditional Estimation without Magic", "authors": [ { "first": "Christopher", "middle": [], "last": "Manning", "suffix": "" }, { "first": "Dan", "middle": [], "last": "Klein", "suffix": "" } ], "year": 2003, "venue": "Tutorial at HLT-NAACL and 41st ACL conferences", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Christopher Manning and Dan Klein. 2003. Opti- mization, MaxEnt Models, and Conditional Estima- tion without Magic. Tutorial at HLT-NAACL and 41st ACL conferences, Edmonton, AB and Sapporo.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "The Theory and Practice of Discourse Parsing and Summarization. A Bradford Book", "authors": [ { "first": "Daniel", "middle": [], "last": "Marcu", "suffix": "" } ], "year": 2000, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Daniel Marcu. 2000. The Theory and Practice of Discourse Parsing and Summarization. A Bradford Book. The MIT Press, Cambridge, MA.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "WordNet: A Lexical Database for English", "authors": [ { "first": "A", "middle": [], "last": "George", "suffix": "" }, { "first": "", "middle": [], "last": "Miller", "suffix": "" } ], "year": 1995, "venue": "Communications of the ACM", "volume": "38", "issue": "11", "pages": "39--41", "other_ids": {}, "num": null, "urls": [], "raw_text": "George A. Miller. 1995. WordNet: A Lexical Database for English. Communications of the ACM, 38(11):39-41.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Experiments on Sense Annotations and Sense Disambiguation of Discourse Connectives. Proc. of the 4th Workshop on Treebanks and Linguistic Theories (TLT)", "authors": [ { "first": "Eleni", "middle": [], "last": "Miltsakaki", "suffix": "" }, { "first": "Nikhil", "middle": [], "last": "Dinesh", "suffix": "" }, { "first": "Rashmi", "middle": [], "last": "Prasad", "suffix": "" }, { "first": "Aravind", "middle": [], "last": "Joshi", "suffix": "" }, { "first": "Bonnie", "middle": [], "last": "Webber", "suffix": "" } ], "year": 2005, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Eleni Miltsakaki, Nikhil Dinesh, Rashmi Prasad, Ar- avind Joshi, and Bonnie Webber. 2005. Exper- iments on Sense Annotations and Sense Disam- biguation of Discourse Connectives. Proc. of the 4th Workshop on Treebanks and Linguistic Theories (TLT), Barcelona.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "A Systematic Comparison of Various Statistical Alignment Models", "authors": [ { "first": "Josef", "middle": [], "last": "Franz", "suffix": "" }, { "first": "Hermann", "middle": [], "last": "Och", "suffix": "" }, { "first": "", "middle": [], "last": "Ney", "suffix": "" } ], "year": 2003, "venue": "Computational Linguistics", "volume": "29", "issue": "1", "pages": "19--51", "other_ids": {}, "num": null, "urls": [], "raw_text": "Franz Josef Och and Hermann Ney. 2003. A System- atic Comparison of Various Statistical Alignment Models. Computational Linguistics, 29(1):19-51.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Minimum Error Rate Training in Statistical Machine Translation", "authors": [ { "first": "Franz Josef", "middle": [], "last": "Och", "suffix": "" } ], "year": 2003, "venue": "Proc. of the 41st Annual Meeting of the ACL", "volume": "", "issue": "", "pages": "160--167", "other_ids": {}, "num": null, "urls": [], "raw_text": "Franz Josef Och. 2003. Minimum Error Rate Train- ing in Statistical Machine Translation. Proc. of the 41st Annual Meeting of the ACL, pages 160-167, Sapporo.", "links": null }, "BIBREF16": { "ref_id": "b16", "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": "Proc. of 40th Annual Meeting of the ACL", "volume": "", "issue": "", "pages": "311--318", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: A method for Au- tomatic Evaluation of Machine Translation. Proc. of 40th Annual Meeting of the ACL, pages 311-318, Philadelphia, PA.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Using Syntax to Disambiguate Explicit Discourse Connectives in Text", "authors": [ { "first": "Emily", "middle": [], "last": "Pitler", "suffix": "" }, { "first": "Ani", "middle": [], "last": "Nenkova", "suffix": "" } ], "year": 2009, "venue": "Proc. of the 47th Annual Meeting of the ACL and the 4th International Joint Conference of the AFNLP (ACL-IJCNLP), Short Papers", "volume": "", "issue": "", "pages": "13--16", "other_ids": {}, "num": null, "urls": [], "raw_text": "Emily Pitler and Ani Nenkova. 2009. Using Syntax to Disambiguate Explicit Discourse Connectives in Text. Proc. of the 47th Annual Meeting of the ACL and the 4th International Joint Conference of the AFNLP (ACL-IJCNLP), Short Papers, pages 13-16, Singapore.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "The Penn Discourse Treebank 2.0", "authors": [ { "first": "Rashmi", "middle": [], "last": "Prasad", "suffix": "" }, { "first": "Nikhil", "middle": [], "last": "Dinesh", "suffix": "" }, { "first": "Alan", "middle": [], "last": "Lee", "suffix": "" }, { "first": "Eleni", "middle": [], "last": "Miltsakaki", "suffix": "" }, { "first": "Livio", "middle": [], "last": "Robaldo", "suffix": "" }, { "first": "Aravind", "middle": [], "last": "Joshi", "suffix": "" }, { "first": "Bonnie", "middle": [], "last": "Webber", "suffix": "" } ], "year": 2008, "venue": "Proc. of the 6th International Conference on Language Resources and Evaluation (LREC)", "volume": "", "issue": "", "pages": "2961--2968", "other_ids": {}, "num": null, "urls": [], "raw_text": "Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Milt- sakaki, Livio Robaldo, Aravind Joshi, and Bon- nie Webber. 2008. The Penn Discourse Treebank 2.0. Proc. of the 6th International Conference on Language Resources and Evaluation (LREC), pages 2961-2968, Marrakech.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Temporal Processing with the TARSQI Toolkit", "authors": [ { "first": "Marc", "middle": [], "last": "Verhagen", "suffix": "" }, { "first": "James", "middle": [], "last": "Pustejovsky", "suffix": "" } ], "year": 2008, "venue": "Proc. of the 22nd International Conference on Computational Linguistics (COLING), Demonstrations", "volume": "", "issue": "", "pages": "189--192", "other_ids": {}, "num": null, "urls": [], "raw_text": "Marc Verhagen and James Pustejovsky. 2008. Tem- poral Processing with the TARSQI Toolkit. Proc. of the 22nd International Conference on Com- putational Linguistics (COLING), Demonstrations, pages 189-192, Manchester, UK.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Classification of Discourse Coherence Relations: An Exploratory Study using Multiple Knowledge Sources", "authors": [ { "first": "Ben", "middle": [], "last": "Wellner", "suffix": "" }, { "first": "James", "middle": [], "last": "Pustejovsky", "suffix": "" }, { "first": "Catherine", "middle": [], "last": "Havasi", "suffix": "" } ], "year": 2006, "venue": "Proc. of the 7th SIGdial Meeting on Discourse and Dialog", "volume": "", "issue": "", "pages": "117--125", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ben Wellner, James Pustejovsky, Catherine Havasi, Roser Sauri, and Anna Rumshisky. 2006. Classi- fication of Discourse Coherence Relations: An Ex- ploratory Study using Multiple Knowledge Sources. Proc. of the 7th SIGdial Meeting on Discourse and Dialog, pages 117-125, Sydney.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Sequence Models and Ranking Methods for Discourse Parsing", "authors": [ { "first": "Ben", "middle": [], "last": "Wellner", "suffix": "" } ], "year": 2009, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ben Wellner. 2009. Sequence Models and Ranking Methods for Discourse Parsing. PhD thesis, Bran- deis University, Waltham, MA.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Significance Tests of Automatic Machine Translation Evaluation Metrics. Machine Translation", "authors": [ { "first": "Ying", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Stefan", "middle": [], "last": "Vogel", "suffix": "" } ], "year": 2010, "venue": "", "volume": "24", "issue": "", "pages": "51--65", "other_ids": {}, "num": null, "urls": [], "raw_text": "Ying Zhang and Stefan Vogel. 2010. Significance Tests of Automatic Machine Translation Evaluation Metrics. Machine Translation, 24(1):51-65.", "links": null } }, "ref_entries": { "FIGREF1": { "uris": null, "num": null, "type_str": "figure", "text": "Figure (a) is for although and (b) for since." }, "TABREF1": { "text": "News Commentary (NC) EN/FR development set with the following modifications: d: Phrase table: NC 2009 (2,051 sentences), no modifications.", "num": null, "content": "
e: Manual annotation: NC 2009 (2,051 sen-
tences), minus all 123 sentences containing
one of the above 5 connective types, plus 102
sentences with manually sense-labeled con-
nectives.
f: Automated annotation: NC 2009 (2,051 sen-
tences), all occurrences of the 13 PDTB sub-
set connective types have been labeled by
classifiers (in 340 sentences).
For testing our modified SMT systems, three
test sets were extracted in the following way:
g: 35 sentences from NC 2007, with 7 occur-
rences for each of the 5 connective types
above, manually labeled.
a: Modification of the phrase table: Europarl
(346,803 sentences), labeling the translation
model after training.
b: Integration of manual annotation: Europarl
(346,803 sentences), minus all 8,901 sen-
tences containing one of the above 5 connec-
tive types, plus 1,147 sentences with manu-
ally sense-labeled connectives.
c: Integration of automated annotation: Europarl
-years 199x (58,673 sentences), all occur-
rences of the 13 PDTB subset connective
types have been labeled by classifiers (in
6,961 sentences).
", "html": null, "type_str": "table" }, "TABREF2": { "text": "and a 5gram language model built over the entire French part of the Europarl corpus v5. Connective Number of occurrences and senses F1 Scores Training set: total and per sense Test set: total and per sense PT PT+ after", "num": null, "content": "
507 456 As, 51 As/Ca25 22 As, 3 As/Ca0.661.00
although267 135 Cs, 118 Ct, 14 Cp16 9 Ct, 7 Cs0.600.66
however176 121 Ct, 32 Cs, 23 Cp14 13 Ct, 1 Cs0.331.00
indeed69 37 Cd, 24 R, 3 Ca, 3 E, 2 I*2 2 R*0.50 *0.50
meanwhile117 66 Cj/S, 16 Cd, 16 S, 1410 5 S, 5 Ct/S0.320.53
Ct/S, 5 Ct
nevertheless26 15 Ct, 11 Cs6 4 Cs, 2 Ct0.440.66
nonetheless12 7 Cs, 3 Ct, 2 Cp*1 1 Cs*1.00 *1.00
rather10 6 R, 2 Al, 1 Ca, 1 Ct*1 1 Al*0.00 *0.00
since166 75 As, 83 Ca, 8 As/Ca9 4 As, 3 Ca, 20.780.78
As/Ca
still114 56 Cs, 51 Ct, 7 Cp13 9 Ct, 4 Cs0.600.66
then145 136 As, 6 Cd, 3 As/Ca6 5 As, 1 Cd0.831.00
while631 317 Ct, 140 S, 79 Cs, 4137 19 Ct, 10 S, 4 Cs,0.930.96
Ct/S, 36 Cd, 18 Cp4 Ct/S
yet80 46 Ct, 25 Cs, 9 Cp*2 2 Ct*0.5 *1.00
Total2,320 -142 -0.570.75
4 Automatically Disambiguating
Discourse Connectives
", "html": null, "type_str": "table" }, "TABREF3": { "text": "Performance of MaxEnt connective sense classifiers: Classifier PT (initial feature set) and Classifier PT+ (with candidate translation features) for 13 temporal and contrastive connectives in the PDTB. The sense labels are coded as follows.", "num": null, "content": "", "html": null, "type_str": "table" }, "TABREF4": { "text": "ConnectiveNumber of occurrences and senses F1 Size of training set: total and per sense Test set: total and per sense Score", "num": null, "content": "
although173 155 Cs, 18 Ct10 5 Cs, 5 Ct0.67
even though179 165 Cs, 14 Ct10 5 Cs, 5 Ct1.00
since413 274 S, 131 Ca, 8 S/Ca10 5 Ca, 3 S, 2 S/Ca0.80
though150 80 Cs, 70 Ct10 5 Cs, 5 Ct1.00
while280 130 Cs, 41 Ct, 89 S/Ct, 13 S/Ca, 7 S 14 4 Cs, 2 Ct, 2 S/Ct, 20.64
S/Ca, 4 S
Total1,195 -54 -0.82
", "html": null, "type_str": "table" }, "TABREF5": { "text": "", "num": null, "content": "", "html": null, "type_str": "table" } } } }