Antecedent selection techniques for high-recall coreference resolution

Antecedent Selection Techniques for High-Recall Coreference Resolution
Yannick Versley
SFB 441 / Seminar f¨ur Sprachwissenschaft Abstract
larger when it is no longer possible to rely on surfacesimilarity.
We investigate methods to improve the re- To overcome the limit of recall that is encoun- call in coreference resolution by also trying tered when only relying on surface features, newer to resolve those definite descriptions where systems for coreference resolutions (Daum´e III and no earlier mention of the referent shares the Marcu, 2005; Ponzetto and Strube, 2006; Versley, same lexical head (coreferent bridging). The 2006; Ng, 2007, inter alia) use lexical semantic in- problem, which is notably harder than iden- formation as an indication for semantic compati- bility in the absence of head equality. Most cur- tions which have the same lexical head, has rent systems integrate the identification of discourse- been tackled with several rather different ap- new definites (i.e., cases like “the sun” or “the man proaches, and we attempt to provide a mean- that Ben met yesterday”, which are definite, but ingful classification along with a quantita- not anaphoric) with the antecedent selection proper, tive comparison. Based on the different mer- which implies that the gain obtained for new features its of the methods, we discuss possibilities to is dependent on the feature’s usefulness both in find- improve them and show how they can be ef- ing semantically related mentions and for the use in Introduction
One goal of this paper is to provide a better under- standing of these information sources by comparing Coreference resolution, the task of grouping men- proposed (and partly new) approaches for resolv- tions in a text that refer to the same referent in the ing coreferent bridging by separately considering real world, has been shown to be beneficial for a the task of antecedent selection (i.e., presupposing number of higher-level tasks such as information ex- that discourse-new markables have been identified traction (McCarthy and Lehnert, 1995), question an- beforehand). Although state of the art methods for swering (Morton, 2000) and summarisation (Stein- modular discourse-new detection (Uryupina, 2003; Poesio et al., 2005) do not achieve near-perfect accu- While the resolution of pronominal anaphora and racy for discourse-new detection, the results we give tracking of named entities is possible with good for antecedent selection represent an upper bound accuracy, the resolution of definite NPs (having a on recall and precision for the full coreference task, common noun as their head) is usually limited to and we think that this upper bound will be useful for the cases that Vieira and Poesio (2000) call directcoreference, where both coreferent mentions have Lascarides, 1998) is a much broader concept, the term ‘corefer- the same head. The other cases, called coreferent ent bridging’ is potentially confusing, as many cases are exam-ples of perfectly well-behaved anaphoric definite noun phrases.
bridging by Vieira and Poesio1, are notably harder Because we want to emphasise the important difference to the because the number of potential candidates is much more easily resolved cases of same-head coreference, we willstick with ‘coreferent bridging’ as the only term that has been 1Because bridging (in the sense of Clark, 1975, or Asher and established for this in the literature.
Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 496–505, Prague, June 2007. c 2007 Association for Computational Linguistics the design of features in both systems using a mod- • synonymy: The antecedent and the anaphor are ular approach, such as (Poesio et al., 2005), where the decision on discourse-newness is taken before- hand, and those that integrate discourse-new classifi-cation with the actual resolution of coreferent bridg- • hyperonymy: The anaphor is a strict generali- ing cases. In contrast to earlier investigations (Mark- ert and Nissim, 2005; Garera and Yarowsky, 2006), we provide a more extensive overview on features • near-synonymy: The anaphor and antecedent and also discuss properties that influence their com- are semantically related but not synonyms in Several approaches have been proposed for the treatment of coreferent bridging. Poesio et al. (1997)use WordNet, looking for a synonymy or hypernymy Of course, not all cases of coreferent bridging realise relation (additionally, for coordinate sisters in Word- such a lexical relation, as sometimes the anaphor Net). The system of Cardie and Wagstaff (1999) takes up information introduced elsewhere than in uses the node distance in WordNet (with an upper the lexical noun phrase head (Peter was found dead limit of 4) as one component in the distance measure in his flat . . . the deceased), or the coreference rela- that guides their clustering algorithm. Harabagiu tion is forced by the discourse structure, without the et al. (2001) use paths through Wordnet, using not only synonym and is-a relations, but also parts, mor- phological derivations, gloss texts and polysemy,which are weighted with a measure based on the re- lation types and number of path elements. Other ap- proaches use large corpora to get an indication for bridging relations: Poesio et al. (1998) use a general word association metric based on common terms oc-curing in a fixed-width window, Gasperin and Vieira Typical cases of coreference include cases like (2004) use syntactic contexts of words in a large cor- 1,2a (hypernym) or 1,2b (compatible but non- pus to induce a semantic similarity measure (similar to the one introduced by Lin, 1998), and then use example of associative bridging between the NP lists of the n nouns that are (globally) most sim- “the door” and its antecedent to “the house”; it ilar to a given noun. Markert and Nissim (2005) is inferred that the door must be part of the house mine the World Wide Web for shallow patterns like mentioned earlier (since doors are typically part of “China and other countries”, indicating an is-a rela- a house), which is not compatible with coreferent tionship. Finally, Garera and Yarowsky (2006) pro- bridging, but is also ranked highly by association pose an association-based approach using nouns that occur in a 2-sentence window before a definite de- While hypernym relations (as found by hypernym scription that has no same-head antecedent.
lookup in WordNet, or patterns indicating such rela-tions in unannotated texts) are usually a strong in- Lexical vs. Referential Relations
dicator of coreference, they can only cover some One important property of these information sources of the cases, while the near-synonymous cases are is the kind of lexical relations that they detect. The left undiscovered. Similarity and association mea- lexical relations that we expect in coreferent bridg- sures can help for the cases of near-synonymy. How- ever, while similarity measures (such as WordNetdistance or Lin’s similarity metric) only detect cases • instance: The antecedent is an instance of the of semantic similarity, association measures (such as the ones used by Poesio et al., or by Garera and Yarowsky) also find cases of associative bridg- Land (country/state/land)
Medikament (medical drug)
highest ranked words, with very rare words removed ∗: RU 486, an abortifacient drugLin98: Lin’s distributional similarity measure (Lin, 1998)RFF: Geffet and Dagan’s Relative Feature Focus measure (Geffet and Dagan, 2004)TheY: association measure introduced by Garera and Yarowsky (2006)TheY:G2: similar method using a log-likelihood-based statistic (see Dunning 1993) this statistic has a preference for higher-frequency terms PL03: semantic space association measure proposed by Pad´o and Lapata (2003) Table 1: Similarity and association measures: most similar items ing like 1a,b; the result of this can be seen in ta- hand, wordnets usually have limited coverage both ble (2): while the similarity measures (Lin98, RFF) in terms of lexical items and in terms of relations list substitutable terms (which behave like synonyms encoded (as their construction is necessarily labor- in many contexts), the association measures (Garera intensive), and – as Markert and Nissim remark and Yarowsky’s TheY measure, Pad´o and Lapata’s – they do not (and arguably should not) contain association measure) also find non-compatible asso- context-dependent relations that do not hold gener- ciations such as country–capital or drug–treatment, ally but only in some rather specific context, for ex- which is why they are commonly called relation- ample steel being anaphorically described as a com- free. For the purpose of coreference resolution, how- modity in a financial text. Context-dependent rela- ever we do not want to resolve “the door” to the an- tions, Markert and Nissim argue, can be found using tecedent “the house” as the two descriptions do not shallow patterns (for example, steel and other com- corefer, and it may be useful to filter out non-similar modities), since a use in such a context would mean that the idiosyncratic conceptual relation holds inthat context. Wordnets also have usually have poor Information Sources
(or non-existant) coverage of named entities, whichare especially relevant for instance relations; this Different resources may be differently suited for kind of instance relations can often be found in large the recognition of the various relations.
text corpora. The high-precision patterns that Mark- ally, it would be expected that using a wordnet ert and Nissim use only occur infrequently, but the is the best solution if we are interested in an isa- approach using shallow patterns allows to perform the search of the World Wide Web, which somewhat we would extract the pairs boy, (person) and boy, ice-cream , in the hope that the former While some near-synonyms can be found by look- pair occurs comparatively more often and gets a ing at the distance in a wordnet, they may be far apart from each other because of ontological mod-eling decisions, or lexical items not covered by the Experiments on Antecedent Selection
wordnet. Similarity and association measures canprovide greater coverage for these near-synonym re- In a setting similar to Markert and Nissim (2005), we evaluate the precision (proportion of correctcases in the resolved cases) and recall (correct cases The measures both of Lin (1998) and of Pad´o and to all cases) for the resolution of discourse-old def- Lapata (2003, 2007) are distributional methods; for inite noun phrases. Before trying to resolve coref- each word, they create a distribution of the contexts erent bridging cases, we look for compatible an- they occur in, and similarity between two words is tecedent candidates with the same lexical head and calculated as the similarity of these distributions.2 resolve to the nearest such candidate if there is one.
The difference in these two methods is the repre-sentation of the contexts. While Lin uses contexts For our experiments, we used the first 125 articles that are expected to determine semantic preferences of the coreferentially annotated T¨uBa-D/Z corpus of (like being in the direct object position of one verb), written newspaper text (Hinrichs et al., 2005), to- Pad´o and Lapata only use the co-occuring words, talling 2239 sentences with 633 discourse-old defi- weighted by syntax-based distance. For example, in nite descriptions, and the latest release of GermaNet(Kunze and Lemnitzer, 2002), which is the German- Unlike Markert and Nissim, we did not limit the Lin’s approach would yield ↑subj :like for Peter evaluation to discourse-old noun phrases where an and ↑dobj :like for ice-cream, while Pad´o and antecedent is in the 4 preceding sentences, but also Lapata’s approach would yield the contexts like included cases where the antecedent is further away.
(with a weight of 1.0) and ice-cream (with a As a real coreference resolution system would have weight of 0.5) for Peter. As a consequence, Pad´o to either resolve them correctly or leave them unre- and Lapata’s measure is more robust against data solved, we feel that this is less unrealistic and thus sparseness but also finds related non-similar terms preferable even when it gives less optimistic evalu- (which are ultimately unwanted for coreference res- ation results. Because overall precision is a mixture olution). Pad´o and Lapata show their dependency- of the precision of the same-head resolver and the based measure to perform better in a word sense precision of the resolution for coreferent bridging, disambiguation task than the measure of Lund et al.
which is lower than that for same-head cases, we (1995), on which Poesio et al. (1998) based their ex- forcibly get less precision if we resolve more coref- periments and which is based on the surface distance erent bridging cases. As it is always possible to im- prove overall precision by resolving fewer cases of We also reimplemented the approach of Gar- coreferent bridging, we separately mention the pre- era and Yarowsky (2006), who extract potential cision for coreferent bridging cases alone (i.e., num- anaphor-antecedent pairs from unlabeled texts and ber of correct coreferent bridging cases by all re- rank these potentially related pairs by the mutual in- solved coreferent bridging cases), which we deem formation statistic. As an example, in a text like In our evaluation, we included hypernymy search and a simple edge-based distance based on Ger- maNet, as well as a baseline using semantic classes(automatically determined by a combination of sim- 2Both measures use a weighted Jaccard metric on mutual ple named entity classification and GermaNet sub- information vectors to calculate the similarity. See Weeds andWeir (2005) for an overview of other measures.
sumption), as well as an evolved version of Markert grammatical relations, was carried out on a subset of all sentences (those with length ≤ 30), with an unlexicalised PCFG parser and subsequent extrac- tion of dependency relations (Versley, 2005). For 0.58 0.68
the last approach, where dependency relations were needed but labeling accuracy was not as important, we used a deterministic shift-reduce parser that Foth and Menzel (2006) used as input source in hybrid For all three approaches, we lemmatised the words by using a combination of SMOR (Schmid 0.64 0.65
et al., 2004), a derivational finite-state morphology for German, and lexical information derived from Prec.NSH: precision for coreferent bridging cases the lexicon of a German dependency parser (Foth (1): consider candidates in the 4 preceding sentences and Menzel, 2006). We mitigated the problem of vo- (2): consider candidates in the 16 preceding sentences cabulary growth in the lexicon, due to German syn- (3): also try candidates such that the anaphor is thetic compounds, by using a frequency-sensitive unsupervised compound splitting technique, and(for semantic similarity) normalised common person and location names to ‘(person)’ and ‘(location)’, re-spectively.
and Nissim’s approach, which is presented in (Ver- Same-head resolution (including a check for sley, 2007). For the methods based on similarity modifier compatibility) allows to correctly resolve and association measures, we implemented a simple 49.8% of all cases, with a precision of 86.5%.
ranking by the respective similarity or relatedness The most simple approach for coreferent bridging, value. Additionally, we included an approach due to just resolving coreferent bridging cases to the near- Gasperin and Vieira (2004), who tackle the problem est possible antecedent (only checking for number of similarity by using lists of most similar words to a agreement), yields very poor precision (12% for the certain word, based on a similarity measure closely coreferent bridging cases), and as a result, the re- related to Lin’s. They allow resolution if either (i) call gain is very limited. If we use semantic classes the candidate is among the words most similar to the (based on both GermaNet and a simple classification anaphor, (ii) the anaphor is among the words most for named entities) to constrain the candidates and similar to the candidate, (iii) the similarity lists of then use the nearest number- and gender-compatible anaphor and candidate share a common item. We antecedent4, we get a much better precision (35% tried out several variations in the length of the simi- for coreferent bridging cases), and a much better lar words list (Gasperin and Vieira used 15, we also recall of 61.1%. Hyponymy lookup in GermaNet, tried lists with 25, 50 and 100 items). The third pos- without a limit on sentence distance, achieves a re- sibility that Gasperin and Vieira mention (a common call of 57.5% (with a precision of 67% for the re- item in the similarity lists of both anaphor and an- solved coreferent bridging cases), whereas using the tecedent) resolves some correct cases, but leads to a best single pattern (Y wie X, which corresponds to much larger number of false positives, which is whywe did not include it in our evaluation.
3Arguably, it would have been more convenient to use a sin- To induce the similarity and association measures gle parser for all three approaches, but differing tradeoffs be-tween speed on one hand and accuracy for relevant information presented earlier, we used texts from the German and/or fitness of representation on the other hand made the re- newspaper die tageszeitung, comprising about 11M spective parser or chunker a compelling choice.
sentences. For the extraction of anaphor-antecedent In German, grammatical gender is not as predictive as in English as it does not reproduce ontological distinctions. For candidates, we used a chunked version of the cor- persons, grammatical and natural gender almost always coin- pus (M¨uller and Ule, 2002). The identification of cide, and we check gender equality iff the anaphor is a person.
the English Y s such as X), with a distance limit of similar words that share this feature.
4 sentences5, on the Web only improves the recall By replacing mutual information values with RFF to 54.3% (with a lower precision of 55% for coref- values in Lin’s association measure, Geffet and Da- erent bridging cases). This is in contrast to the re- gan were able to significantly improve the propor- sults of Markert and Nissim, who found that Web tion of substitutable words in the list of the most sim- pattern search performs better than wordnet lookup; ilar words. In our experiments, however, using the see (Versley, 2007) for a discussion. Ranking all RFF-based similarity measure did not improve the candidates that are within a distance of 4 hyper- similarity-list-based resolution or the simple rank- /hyponymy edges in GermaNet by their edge dis- ing, to the contrary, both recall and precision are less tance, we get a relatively good recall of 60.5%, but than for the Weighted Jaccard measure that we used the precision (for the coreferent bridging cases) is only at 39%, which is quite poor in comparison.
We attribute this to two factors: Firstly, Geffet The results for Garera and Yarowsky’s TheY al- and Dagan’s evaluation emphasises the precision in gorithm are quite disconcerting – recall and the pre- terms of types, whereas the use in resolving coref- cision on coreferent bridging cases are lower than erent bridging does not punish unrelated rare words the respective baseline using (wordnet-based) se- being ranked high – since these are rare, the like- mantic class information or Pad´o and Lapata’s asso- lihood that they occur together, changing a reso- ciation measure. The technique based on Lin’s simi- lution decision, is quite low, whereas rare related larity measure does outperform the baseline, but still words that are ranked high can allow a correct res- suffers from bad precision, along with Pad´o and La- olution. Secondly, Geffet and Dagan focus on high- pata’s association measure. In other words, the simi- frequency words, which makes sense in the context larity and association measures seem to be too noisy of ontology learning, but the applicability for tasks to be used directly for ranking antecedents. The ap- like coreference resolution (directly or in the ap- proach of Gasperin and Vieira performs compara- proach of Gasperin and Vieira) also depends on a bly to the approach using Web-based pattern search sensible treatment of lower-frequency words.
(although the precision is poorer than for the best- Using the framework of Weeds et al. (2004), we performing pattern for German, “X wie Y ” – X found that the bias of lower frequency words for such as Y , it is comparable to that of other patterns).
preferring high-frequency neighbours was higher forRFF (0.58 against 0.35 for Lin’s measure). Weeds Improving Distributional Similarity?
and Weir (2005) discuss the influence of bias to- While it would be na¨ıve to think that the methods wards high- or low-frequency items for different purely based on statistical similarity measures could tasks (correlation with WordNet-derived neighbour reach the accuracy that can be achieved with a hand- sets and pseudoword disambiguation), and it would constructed lexicalised ontology, it would of course not be surprising if the different high-frequency bias be nice if we could improve the quality of the se- mantic similarity measure used in ranking and themost-similar-word lists.
Combining Information Sources
Geffet and Dagan (2004) propose an approach The information sources that we presented earlier to improve the quality of the feature vectors used and the corpus-based methods based on similarity or association measures draw from different kinds of weighting features using the mutual information evidence and thus should be rather complementary.
value between the word and the feature, they pro- To put it another way, it should be possible to get pose to use a measure they call Relative Feature Fo- the best from all methods, achieving the recall of the cus: the sum of the similarities to the (globally) most high-recall methods (like using semantic class in- 5There is a degradation in precision for the pattern-based 6Simple ranking with RFF gives a precision of 33% for approach, but not for the GermaNet-based approach, which is coreferent bridging cases, against 39% for Lin’s original mea- why we do not use a distance limit for the GermaNet-based ap- sure; for an approach based on similarity lists, we get 39% and everything else). Very surprisingly, Garera and Yarowsky’s TheY approach, despite starting out at a lower precision (31%, against 39% for Lin and 42% for PL03), profits much more from the semantic fil- ter and reaches the best precision (47%), whereas Lin’s semantic similarity measure profits the least.
Since limiting the distance to the 4 previous sen- tences had quite a devastating effect for the approach based on Lin’s similarity measure (which achieves 39% precision when all the candidates are avail- able and 30% precision if it choses the most se-mantically similar out of the candidates that are in the last 4 sentences), we also wanted to try and ap- ply the distance-based filtering after finding seman- The approach we tried was as follows: we rank all candidates using the similarity function, and keep only the 3 top-rated candidates. From these 3 top- rated candidates, we keep only those within the last 4 sentences. Without filtering by semantic class, this improves the precision to 41% (from 30% for lim- iting the distance beforehand, or 39% without lim- 0.68 0.73
0.70 0.72
iting the distance). Adding filtering based on se-mantic classes to this (only keeping those from the (2): consider candidates in the 16 preceding sentences(3) 3 top-rated candidates which have a compatible se- : also try candidates such that the anaphor is mantic class and are within the last 4 sentences), weget a much better precision of 53%, with a recall that can still be seen as good (57.8%). In compari-son with the similarity-list-based approach, we get amuch better precision than we would get for meth- formation, or similarity and association measures), ods with comparable recall (the version with the 100 with a precision closer to the most precise method most similar items has 44% precision, the version using GermaNet. In the case of web-based patterns, with 50 most similar items and matching both ways Versley (2007) combines several pattern searches on the web and uses the combined positive and nega- Applying this distance-bounding method to Gar- tive evidence to compute a composite score – with a era and Yarowsky’s association measure still leads suitably chosen cutoff, it outperforms all single pat- to an improvement over the case with only seman- terns both in terms of precision and recall. First re- tic and gender checking, but the improvement (from solving via hyponymy in GermaNet and then using 47% to 50%) is not as large as with the semantic the pattern-combination approach outperforms the similarity measure or Pad´o and Lapata’s association semantic class-based baseline in terms of recall and is reasonably close to the GermaNet-based approach For the final system, we back off from the most in terms of precision (i.e., much better than the ap- precise information sources to the less precise. Start- proach based only on the semantic class).
ing with the combination of GermaNet and pattern- As a first step to improve the precision of the based search on the World Wide Web, we begin corpus-based approaches, we added filtering based by adding the distance-bounded semantic similarity- on automatically assigned semantic classes (per- based resolver (LinBnd) and resolution based on sons, organisations, events, other countable objects, the list of 25 most similar words (following the approach of Gasperin and Vieira 2004). This re- et al. (1995) with wordnet relations and pattern sults in visibly improved recall (from 62% to 68%), search on a fixed-size corpus.7 However, they eval- while the precision for coreferent bridging cases uate only on a small subset of discourse-old definite does not suffer much. Adding resolution based on descriptions (those where a wordnet-compatible se- Lin’s semantic similarity measure and Garera and mantic relation was identified and which were rea- Yarowsky’s TheY value leads to a further improve- sonably close to their antecedent), and they did not ment in recall to 69.7%, but also leads to a larger distinguish coreferent from associative bridging an- tecedents. Although the different evaluation methoddisallows a meaningful comparison, we think that Conclusion
the more evolved information sources we use (Pad´oand Lapata’s association measure instead of Lund In this paper, we compared several approaches to re- et al’s, combined pattern search on the World Wide solve cases of coreferent bridging in open-domain Web instead of search for patterns in a fixed-size corpus), as well as the additional information based sources can match the precision of the hypernymy on semantic similarity, lead to superior results when information encoded in GermaNet, or that of using a combination of high-precision patterns with theWorld Wide Web as a very large corpus, it is possi- Ongoing and Future Work
ble to achieve a considerable improvement in terms Both the distributional similarity statistics and the of recall without sacrificing too much precision by association measure can profit from more training data, something which is bound by availability of Very interestingly, the distributional methods similar text (Gasperin et al., 2004 point out that us- based on intra-sentence relations (Lin, 1998; ing texts from a different genre strongly limits the Pad´o and Lapata, 2003) outperformed Garera and usefulness of the learned semantic similarity mea- Yarowsky’s (2006) association measure when used sure), and by processing costs (which are more se- for ranking, which may due to sparse data problems rious for distributional similarity measures than for or simply too much noise for the latter. For the asso- non-grammar-related association measures, as the ciation measures, the fact that they are relation-free also means that they can profit from added semantic Based on existing results for named entity coref- erence, a hypothetical coreference resolver combin- The novel distance-bounded semantic similarity ing our information sources with a perfect detec- method (where we use the most similar words in the tor for discourse-new mentions would be able to previous discourse together with a semantic class- achieve a precision of 88% and a recall of 83% con- based filter and a distance limit) comes near the pre- sidering all full noun phrases (i.e., including names, cision of using surface patterns, and offers better ac- but not pronouns). This is both much higher than curacy than Gasperin and Vieira’s method of using state-of-the art results for the same data set (Versley, 2006, gets 62% precision and 70% recall), but such By combining existing higher-precision informa- accuracy may be very difficult to achieve in prac- tion sources such as hypernym search in GermaNet tice, as perfect (or even near-perfect) discourse-new and the Web-based approach presented in (Vers- detection does not seem to achievable in the near fu- ley, 2007) together with similarity- and association- ture. Preliminary experiments show that the inte- based resolution, it is possible to get a large im- gration of pattern-based information leads to an in- provement in recall even compared to the combined crease in recall of 0.6% for the whole system (or GermaNet+Web approach or an approach combin- 46% more coreferent bridging cases), but the inte- ing GermaNet with a semantically filtered version gration of distributional similarity (loosely based on of Garera and Yarowsky’s TheY approach.
the approach by Gasperin and Vieira) does not lead In independent research, Goecke et al. (2006) 7Thanks to Tonio Wandmacher for pointing this out to me at combined the original LSA-based method of Lund to a noticeable improvement over GermaNet alone; a joint entity detection and tracking model. In in isolation, the distributional similarity information HLT/EMNLP’05, pages 97–104.
did improve the recall, albeit less than information Dunning, T. (1993). Accurate methods for the statis- tics of surprise and coincidence. Computational The fact that only a small fraction of the achiev- able recall gain is currently attained seems to sug- gest that better identification of discourse-old men- ing: Using probabilistic models as predictors for tions could potentially lead to larger improvements.
a symbolic parser. In ACL 2006.
It also seems that firstly, it makes more sense to com- Garera, N. and Yarowsky, D. (2006). Resolving and bine information sources that cover different rela- generating definite anaphora by modeling hyper- tions (e.g. GermaNet for hypernymy and synonymy nymy using unlabeled corpora. In CoNLL 2006.
and the pattern-based approach for instance rela- Gasperin, C., Salmon-Alt, S., and Vieira, R. (2004).
tions) than those that yield independent evidence for How useful are similarity word lists for indirect the same relation(s), as GermaNet and the Gasperin anaphora resolution? In Proc. DAARC 2004.
and Vieira approach do for (near-)synonymy; andsecondly, that good precision is especially important in the context of integrating antecedent selection and similarity lists for resolving indirect anaphora. In discourse-new identification, which means that the ACL’04 workshop on reference resolution and its finer view that we get using antecedent selection ex- periments (compared to direct use in a coreference quality and distributional similarity. In CoLing2004.
Acknowledgements
Goecke, D., St¨uhrenberg, M., and Wandmacher, T.
Schulte im Walde, Piklu Gupta and Sandra K¨ubler (2006). Extraction and representation of seman- for useful criticism of an earlier version, and to tic relations for resolving definite descriptions. In Simone Ponzetto and Michael Strube for feedback Workshop on Ontologies in Text Technology (OTT ported in this paper was supported by the Deutsche Harabagiu, S., Bunescu, R., and Maiorano, S.
Forschungsgemeinschaft (DFG) as part of Collab- (2001). Text and knowledge mining for corefer- orative Research Centre (Sonderforschungsbereich) ence resolution. In Proceedings of the 2nd Meet- 441 “Linguistic Data Structures”.
ing of the North American Chapter of the Associa-tion of Computational Linguistics (NAACL-2001).
References
Hinrichs, E., K¨ubler, S., and Naumann, K. (2005). A Asher, N. and Lascarides, A. (1998). Bridging. Jour- unified representation for morphological, syntac- nal of Semantics, 15(1):83–113.
tic, semantic and referential annotations. In ACL Cardie, C. and Wagstaff, K. (1999). Noun phrase Workshop on Frontiers in Corpus Annotation II: coreference as clustering. In Proceedings of the Joint Conference on Empirical Methods in Natu- Kunze, C. and Lemnitzer, L. (2002). Germanet – ral Language Processing and Very Large Corpora representation, visualization, application. In Pro- (EMNLP/VLC 1999), pages 82–89.
Clark, H. H. (1975). Bridging. In Schank, R. C. and Lin, D. (1998). Automatic retrieval and clustering Nash-Webber, B. L., editors, Proceedings of the of similar words. In Proc. CoLing/ACL 1998.
1975 workshop on Theoretical issues in natural Lund, K., Atchley, R. A., and Burgess, C.
language processing, pages 169–174, Cambridge, (1995). Semantic and associative priming in high- MA. Association for Computing Machinery.
dimensional semantic space. In Proc. of the 17th Daum´e III, H. and Marcu, D. (2005).
Annual Conference of the Cognitive Science Soci- scale exploration of effective global features for Conference and Conference on Empirical Meth- knowledge sources for nominal anaphora resolu- ods in Natural Language Processing, pages 1–8.
tion. Computational Linguistics, 31(3):367–402.
Uryupina, O. (2003). High-precision identification McCarthy, J. F. and Lehnert, W. G. (1995). Using of discourse new and unique noun phrases. In decision trees for coreference resolution. In IJCAI Proceedings of the ACL Student Workshop.
Morton, T. S. (2000). Coreference for NLP applica- types. In Proceedings of the Fourth Workshop on Treebanks and Linguistic Theories (TLT 2005).
M¨uller, F. H. and Ule, T. (2002). Annotating topo- logical fields and chunks – and revising POS tags to noun phrase coreference resolution in German at the same time. In Proceedings of the Nineteenth newspaper text. In Konferenz zur Verarbeitung International Conference on Computational Lin- Nat¨urlicher Sprache (KONVENS 2006).
Ng, V. (2007). Shallow semantics for coreference coreferent bridging in German newspaper text.
resolution. In IJCAI 2007, pages 1689–1694.
In Proceedings of GLDV-Fr¨uhjahrstagung 2007, Pad´o, S. and Lapata, M. (2003). Constructing se- mantic space models from parsed corpora. In Pro- Vieira, R. and Poesio, M. (2000). An empirically based system for processing definite descriptions.
Pad´o, S. and Lapata, M. (2007). Dependency-based Computational Linguistics, 26(4):539–593.
construction of semantic space models. Compu- Weeds, J. and Weir, D. (2005). Co-occurrence re- tational Linguistics, to appear.
trieval: A flexible framework for lexical distri- Poesio, M., Alexandrov-Kabadjov, M., Vieira, R., Goulart, R., and Uryupina, O. (2005).
discourse-new detection help definite description Weeds, J., Weir, D., and McCarthy, D. (2004). Char- acterizing measures of lexical distributional simi- national Workshop on Computational Semantics Poesio, M., Schulte im Walde, S., and Brew, C.
(1998). Lexical clustering and definite descrip-tion interpretation. In AAAI Spring Symposiumon Learning for Discourse.
Poesio, M., Vieira, R., and Teufel, S. (1997). Re- solving bridging descriptions in unrestricted text.
In ACL-97 Workshop on Operational Factors inPractical, Robust, Anaphora Resolution For Un-restricted Texts.
Ponzetto, S. P. and Strube, M. (2006). Exploiting semantic role labeling, wordnet and wikipedia forcoreference resolution. In HLT-NAACL 2006.
Schmid, H., Fitschen, A., and Heid, U. (2004).
SMOR: A german computational morphologycovering derivation, composition and inflection.
In Proceedings of LREC 2004.
Steinberger, J., Kabadjov, M., Poesio, M., and based summarization with anaphora resolution.
In Proceedings of Human Language Technology

Source: http://www.versley.de/D07-1052.pdf