A Statistical Approach to Anaphora Resolution
Summary
The paper addresses pronominal anaphora resolution: deciding which noun phrase a pronoun refers to. Ge, Hale and Charniak avoid hand-written rules and instead build a probabilistic model trained on a small part of the Penn Wall Street Journal Treebank that was annotated with coreference links. Several cues are combined, through Bayes' rule and independence assumptions, into one score for each candidate antecedent. These cues are:
- syntactic position and distance, captured as "Hobbs distance" from a modified Hobbs algorithm;
- the gender, number and animacy suggested by the candidate's words;
- selectional fit between the candidate and the pronoun's governing head;
- how often the referent has already been mentioned.
The highest-scoring candidate is chosen. With all factors combined, the system resolves 82.9% of pronouns correctly.
A second experiment uses the resolver itself to learn word gender without supervision, by counting which pronouns refer back to each noun. Adding this learned lexicon raises accuracy to 84.2%. The work matters because nearly all of the system's knowledge sits in simple probability tables, so the approach is easy to extend.
Key points
- The training data is a small coreference-annotated subset of the Penn Wall Street Journal Treebank. The Hobbs algorithm was adapted to Treebank trees, which lack some of the nodes Hobbs assumed, and it proposes 15 candidates per pronoun.
- The final score multiplies four estimated components:
- the Hobbs-distance probability;
- the pronoun-given-antecedent-word (gender/animacy) probability;
- a head-dependent word probability taken from Charniak's parser statistics, with heads clustered to reduce sparse data;
- a mention-count probability, using bucketed counts conditioned on sentence position.
- When an antecedent contains several candidate words, Dunning's likelihood-ratio test selects the most informative one for the gender statistics.
- Accuracy is 82.9% with all factors and 84.2% after adding the unsupervised gender information.
- In a blind evaluation of the learned genders, assignments based on the Hobbs method were 70.3% accurate, better than the last-noun baseline. Accuracy grows with the number of references to a word, but most referents occur only once, following a Zipf-like distribution, which lowers the overall figure.
Cited by
- Evaluation Tool for Rule-based Anaphora Resolution Methods
- Unsupervised Learning of Contextual Role Knowledge for Coreference Resolution
- Bootstrapping Path-Based Pronoun Resolution
- Noun Phrase Coreference as Clustering
- EM Works for Pronoun Anaphora Resolution
- An Expectation Maximization Approach to Pronoun Resolution
- Coreference-inspired Coherence Modeling
- Statistical Anaphora Resolution in Biomedical Texts
- Active Learning for Anaphora Resolution
- Unsupervised Coreference Resolution in a Nonparametric Bayesian Model
- Using Semantic Relations to Refine Coreference Decisions
- An Incremental Entity-Mention Model for Coreference Resolution with Restrictive Antecedent Accessibility
- PoCoS - Potsdam Coreference Scheme
- Automatic Pronominal Anaphora Resolution in English Texts
- A Mention-Synchronous Coreference Resolution Algorithm Based On the Bell Tree
- Multi-Lingual Coreference Resolution With Syntactic Features
- Aiding Pronoun Translation with Co-Reference Resolution
- Supervised Noun Phrase Coreference Research: The First Fifteen Years
- An Algorithm for Anaphora Resolution in Spanish Texts
- A Probabilistic Method for Analyzing Japanese Anaphora Integrating Zero Pronoun Detection and Resolution
- A Machine Learning Approach to Coreference Resolution of Noun Phrases
- Analysis of Syntax-Based Pronoun Resolution Methods
- A Corpus-Based Evaluation of Centering and Pronoun Resolution
- Improving Pronoun Resolution by Incorporating Coreferential Information of Candidates
- A Twin-Candidate Model for Learning-Based Anaphora Resolution
Cite
@inproceedings{ge1998statistical,
title = {{A Statistical Approach to Anaphora Resolution}},
author = {Ge, Niyu and Hale, John and Charniak, Eugene},
booktitle = {Sixth Workshop on Very Large Corpora},
year = {1998},
url = {https://aclanthology.org/W98-1119/}
}References
- Chinatsu Aone and Scott William Bennett, 1996. Evaluating Automated and Manual Acquisition off Anaphora Resolution Strategies, pages 302-315. Springer. find
- Susan E. Brennan, Marilyn Walker Friedman, and Carl J. Pollard. 1987. A centering approach to pronouns. In Proc. 25th Annual Meeting of the A CL, pages 155-162. Association of Computational Linguistics. find
- Eugene Charniak. 1997. Statistical parsing with a context-free grammar and word statistics. In Proceedings of the 14th National Conference on Artificial Intelligence, Menlo Park, CA. AAAI Press/MIT Press. find
- Ted Dunning. 1993. Accurate methods for the statistics of surprise and coincidence. Computational Linguistics, 19(1), March. find our card
- Vasileios Hatzivassiloglou and Kathleen R. McKeown. 1997. Predicting the semantic orientation of adjectives. In Proc. 35th Annual Meeting of the ACL, pages 174-181. Association of Computational Linguistics. find
- Jerry R. Hobbs. 1976. Pronoun resolution. Technical Report 76-1, City College, New York. find
- Shalom Lappin and Herbert J. Leass. 1994. An algorithm for pronominal anaphora resolution. Computational Linguistics, pages 535-"561. find our card
- Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of english: the penn treebank. Computational Linguistics, 19:313-330. find our card
- Ruslan Mitkov. 1997. Factors in anaphora resolution: they are not the only things that matter, a case study based on two different approaches. In Proceedings of the A CL 'g7/EA CL 'g7 Workshop on Operational Factors in Practical, Robust Anaphora Resolution. find our card
- J. Ross Quinlan. 1993. C~.5 Programs for Machine Learning. Morgan Kaufmann Publishers. find
This page in Markdown → https://md.research.elamur.ai/free/papers/ge-1998-statistical-approach-anaphora-resolution/