1 / 26

A Cross -Lingual ILP Solution to Zero Anaphora Resolution

A Cross -Lingual ILP Solution to Zero Anaphora Resolution. Ryu Iida & Massimo Poesio (ACL-HLT 2011). Zero-anaphora resolution. Anaphoric function in which phonetic realization of anaphors is not required in “pro-drop” languages Based on speaker and hearer’s shared understanding

paniz
Télécharger la présentation

A Cross -Lingual ILP Solution to Zero Anaphora Resolution

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. A Cross-Lingual ILP Solution to Zero Anaphora Resolution Ryu Iida & Massimo Poesio (ACL-HLT 2011)

  2. Zero-anaphora resolution • Anaphoric function in which phonetic realization of anaphors is not required in “pro-drop” languages • Based on speaker and hearer’s shared understanding • φ: zero-anaphor (non-realized argument) • Essential: 64.3% of anaphors in Japanese newspaper articles are zeros (Iida et al. 2007)

  3. Research background • Zero-anaphora resolution has remained an active area for Japanese (Seki et al. 2002, Isozaki&Hirao 2003, Iida et al. 2007, Imamura et al. 2009, Sasano et al. 2009, Taira et al. 2010) • The availability of the annotated corpora such that provided by SemEVAL2010 task10 “Multi-lingual coreference (Recasens et al.2010) is leading to renewed interest (e.g. Italian) • Mediocre results obtained on zero anaphors by most systems in SemEVAL • e.g. I-BART’s recall on zeros < 10%

  4. Resolving zero-anaphors requires • The simultaneous decisionof • Zero-anaphor detection: find phonetically unrealized arguments of predicates (e.g. verbs) • Antecedent identification: search for an antecedent of a zero-anaphor • Roughly correspond to anaphoricity determination and antecedent identification in coreference resolution • Denis&Baldridge(2007) proposed a solution to optimize the outputs from anaphoricity determination and antecedent identification by using Integer Linear Programming (ILP)

  5. Main idea • Apply Denis&Baldridge (2007)’s ILP framework to zero-anaphora resolution • Extend the ILP framework into a two-way to make it more suitable for zero-anaphora resolution • Focus on Italian and Japanese zero-anaphora to investigate whether or not our approach is useful across languages • Study only subject zero-anaphors (only type in Italian)

  6. Topic of contents • Research background • Denis&Baldridge (2007)’s ILP model • Proposal: extending the ILP model • Empirical evaluations • Summary & future directions

  7. Denis&Baldrige (2007)’s ILP formulation of base model • object function • If , mentions i and j are coreferentand mention j is an anaphor : 1 if mentions i and j are coreferent; otherwise 0

  8. Denis&Baldrige (2007)’s ILP formulation of joint model • object function • If , mentions i and j are coreferentand mention j is an anaphor; otherwise j is non-anaphoric : 1 if mention j is an anaphor; otherwise 0

  9. 3 constraints in ILP model characteristics of coreference relations • Resolve only anaphors:if mention pair ij is coreferent,mention j must be anaphoric • Resolve anaphors:if mention j is anaphoric, it must be coreferent with at least one antecedent • Do not resolve non-anaphors:if mention j is non-anaphoric, it should be have no antecedents transitivity of coreference chains

  10. Proposal: extending the ILP framework • Denis&Baldridge’s original ILP-based model is not suitable for zero-anaphora resolution  Two modifications • Applying best-first solution • Incorporating a subject detection model

  11. 1. Best-first solution • Select at most one antecedent for an anaphor • “Do-not-resolve-anaphors” constraint is too weak • Allow the redundant choice of more than one candidate antecedent • Lead to decreasing precision on zero-anaphora resolution • “Do-not-resolve-anaphors” constraint is replaced with “Best First constraint (BF)” that blocks selection of more than one antecedent:

  12. 2. Integrating subject detection model • Zero-anaphor detection • Difficulty in zero-anaphora resolution comparing to pronominal reference resolution • Simply relying on the parser is not enough • most dependency parsers are not very accurate at identifying grammatical roles • detecting subject is crucial for zero-anaphor detection

  13. 2. Integrating subject detection model • Resolve only non-subjects:if a predicate j syntactically depends on a subject,the predicate j should have no antecedent of its zero anaphor : 1 if predicate j syntactically depends on a subject; otherwise 0

  14. Experiment 1: zero-anaphors • Compare the baseline models with the extended ILP-based models • Use the Maximum Entropy model to create base classifiers in the ILP framework and baselines • Featuredefinitions basically follow the previous work (Iida et al. 2007) and (Poesio et al. 2010)

  15. Two baseline models • PAIRWISE classification model (PAIRWISE) • Antecedent identification and anaphoricity determination are simultaneously executed by a single classifier (as in Soon et al. 2001) • Anaphoricity Determination-then-Search antecedent CASCADEd model (DS-CASCADE) • Filter out non-anaphoric candidate anaphors using an anaphoricity determination model • Select an antecedent from a set of candidate antecedents of anaphoric anaphors using an antecedent identification model

  16. Data sets • Italian (Wikipedia articles) • LiveMemories text corpus 1.2 (Rodriguez et al. 2010) • Data set on the SemEval2010: Coreference Resolution in Multiple Languages • #zero-anaphors: train 1,160 / test 837 • Japanese (newspaper articles) • NAIST text corpus (Iida et al. 2007) ver.1.4ß • #zero-anaphors: train 29,544 / test 11,205

  17. Creating subject detection models • Data sets • Italian: 80,878 tokens in TUT corpus (Bosco et al. 2010) • Japanese: 1753 articles (i.e. training dataset) in NAIST text corpus merged with Kyoto text corpus • dependency arc is judged as positive if its relation is subject; as negative otherwise • Induce a maximum entropy classifier based on the labeled arcs • Features • Italian: lemmas, PoS tags and morphological information automatically computed by TextPro (Pianta et al. 2008) • Japanese: similar features as Italian except gender and number information

  18. Results for zeroanaphors +BF: use best first constraint, +SUBJ: use subject detection model

  19. Experiment 2: all anaphors • Investigate performance of all anaphors (i.e. NP- coreference and zero-anaphors) • Use the same data set and same data separation • Italian: LiveMemories text corpus 1.2 • Japanese: NAIST text corpus 1.4ß • Performance of each model are compared in terms of MUC score • Different types of referring expressions display very different anaphoric behavior • Induce 2 different models for NP-coreference and zero-anaphora respectively

  20. Results for all anaphors

  21. Summary • Extended Denis&Baldridge (2007)’s ILP-based coreference resolution model by incorporating modified constraints & a subject detection model • Our results show the proposed model obtained improvement on both zero-anaphora resolution and overall coreference resolution

  22. Future directions • Introduce more sophisticated antecedent identification model • Test our model for English constructions resembling zero-anaphora • Null instantiations in SEMEVAL 2010 ‘Linking Events and their Participants in Discourse’ task • Detect generic zero-anaphors • Have no antecedent in the preceding context • e.g. the Italian and Japanese translation of • I walked into the hotel and (they) said …

  23. Data sets on English coreference • Use ACE-2002 data set • Data set is classified into the two subset • Pronouns and NPs

  24. Details of experiment: English training data test data train: NPs train: zeros test: NPs test: zeros models: NP coreference models: zero anaphora outputs: NPs outputs: zeros outputs: all anaphors

  25. Results: all anaphors (English)

More Related