Unsupervised Coreference Resolution in a Nonparametric Bayesian Model
Summary
Coreference resolution is usually handled by supervised classifiers that decide whether pairs of mentions co-refer. This approach needs annotated data and can produce clusterings that are not internally consistent. Haghighi and Klein propose a fully generative, unsupervised alternative in which entities are mixture components. A hierarchical Dirichlet process shares entities and parameters across a whole corpus and infers the number of entities instead of fixing it in advance.
Within each document, mentions are generated from two sources:
- global entity properties: head words, entity type, gender and number;
- a sequential salience model.
As a result, pronouns are linked to entity properties rather than to particular entities. Inference is done with Gibbs sampling.
The authors built the model incrementally on ACE 2004 development material. They then tested it, without further tuning, on MUC-6. The system received gold mention boundaries, heads and mention types, but no NER labels or POS tags. It reaches 70.3 MUC F1, which is near some supervised systems. This suggests that unlabeled text alone can support competitive coreference.
Key points
- Development data: the ACE 2004 Arabic and Chinese treebank material, rendered into English (95 documents, roughly 3,905 mentions). Results are averaged over five sampling runs.
- On development data, the basic infinite mixture (Dirichlet process, Chinese restaurant process) gets only 54.5 F1, largely because pronouns end up in their own clusters. Pronoun-property and salience components fix much of this.
- MUC-6 test results:
- 70.3 MUC F1 when unlabeled ACE ENGLISH-NWIRE data is added, versus 73.4 for the supervised McCallum and Wellner (2004) and 53.6 for the unsupervised Cardie and Wagstaff (1999).
- 89.8 F1 on proper mentions only.
- The most frequent error is failing to link proper and nominal mentions of the same entity, for example in appositives.
- Used as an unsupervised NER tagger on MUC-6 proper mentions, the model achieves 61.2% accuracy, compared with 46.4% for always predicting PERSON.
Cited by
- Unsupervised Event Coreference Resolution with Rich Linguistic Features
- Unsupervised Event Coreference Resolution
- Evaluation Metrics For End-to-End Coreference Resolution Systems
- EM Works for Pronoun Anaphora Resolution
- Improving Event Coreference Resolution by Modeling Correlations between Event Coreference Chains and Document Topic Structures
- Specialized Models and Ranking for Coreference Resolution
- The Same-Head Heuristic for Coreference
- Simple Coreference Resolution with Rich Syntactic and Semantic Features
- Coreference Resolution in a Modular, Entity-Centered Model
- Bootstrapping coreference resolution using word associations
- Deterministic Coreference Resolution Based on Entity-Centric, Precision-Ranked Rules
- A Pronoun Anaphora Resolution System based on Factorial Hidden Markov Models
- Unsupervised Ranking Model for Entity Coreference Resolution
- Transferring Coreference Resolvers with Posterior Regularization
- Multigraph Clustering for Unsupervised Coreference Resolution
- Unsupervised Models for Coreference Resolution
- Supervised Noun Phrase Coreference Research: The First Fifteen Years
- Deep Neural Networks for Coreference Resolution for Polish
- Joint Unsupervised Coreference Resolution with Markov Logic
- Translation-Based Projection for Multilingual Coreference Resolution
- A Constraint-Based Hypergraph Partitioning Approach to Coreference Resolution
- Large-Scale Cross-Document Coreference Using Distributed Inference and Hierarchical Models
- Coreference Resolution with Reconcile
- A Discriminative Hierarchical Model for Fast Coreference at Large Scale
Cite
@inproceedings{haghighi2007unsupervised,
title = {{Unsupervised Coreference Resolution in a Nonparametric Bayesian Model}},
author = {Haghighi, Aria and Klein, Dan},
booktitle = {Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics},
year = {2007},
url = {https://aclanthology.org/P07-1107/}
}References
- I. Bhattacharya and L. Getoor. 2006. A latent dirichlet model for unsupervised entity resolution. SIAM conference on data mining. find
- Claire Cardie and Kiri Wagstaff. 1999. Noun phrase coreference as clustering. EMNLP. find our card
- Hal Daume and Daniel Marcu. 2005. A Bayesian model for supervised clustering with the Dirichlet process prior. JMLR. find
- Pascal Denis and Jason Baldridge. 2007. Global, joint determination of anaphoricity and coreference resolution using integer programming. HLT-NAACL. find our card
- Thomas Ferguson. 1973. A bayesian analysis of some non-parametric problems. Annals of Statistics. find
- Niyu Ge, John Hale, and Eugene Charniak. 1998. A statistical approach to anaphora resolution. Sixth Workshop on Very Large Corpora. find our card
- Xiaoqiang Luo, Abe Ittycheriah, Hongyan Jing, Nanda Kambhatla, and Salim Roukos. 2004. A mention-synchronous coreference resolution algorithm based on the bell tree. ACL. find our card
- Andrew McCallum and Ben Wellner. 2004. Conditional models of identity uncertainty with application to noun coreference. NIPS. find
- Brian Milch, Bhaskara Marthi, Stuart Russell, David Sontag, Daniel L. Ong, and Andrey Kolobov. 2005. Blog: Probabilistic models with unknown objects. IJCAI. find
- Vincent Ng and Claire Cardie. 2002. Improving machine learning approaches to coreference resolution. ACL. find our card
- NIST. 2004. The ACE evaluation plan. find
- W. Soon, H. Ng, and D. Lim. 2001. A machine learning approach to coreference resolution of noun phrases. Computational Linguistics. find our card
- Yee Whye Teh, Michael Jordan, Matthew Beal, and David Blei. 2006. Hierarchical dirichlet processes. Journal of the American Statistical Association, 101. find
- Marc Vilain, John Burger, John Aberdeen, Dennis Connolly, and Lynette Hirschman. 1995. A model-theoretic coreference scoring scheme. MUC-6. find our card
This page in Markdown → https://md.research.elamur.ai/free/en/papers/haghighi-2007-unsupervised-coreference-resolution-nonparametric-bayesian/