Joint Unsupervised Coreference Resolution with Markov Logic

Summary

Poon and Domingos address coreference resolution without labeled data. Supervised systems need costly annotated mention pairs, and earlier unsupervised models, notably Haghighi & Klein (2007), were clearly less accurate. The authors build a cluster-based model in Markov logic that reasons jointly over all mentions and enforces transitivity implicitly. Anaphoricity is folded into the same process instead of being decided beforehand. Because the model is expressed as weighted first-order formulas, linguistic cues such as apposition and predicate nominals can be added easily. Those cues create large dependent groups of mentions, which Gibbs-sampling approaches cannot handle well. For learning, the authors extend existing Markov logic algorithms to hidden predicates, producing the first general-purpose unsupervised learner for Markov logic. Inference uses Lazy-MC-SAT. On MUC-6 and ACE data, the system beats Haghighi & Klein even with very little training text, and it often equals or surpasses strong supervised systems. This shows that joint inference can largely make up for missing labels.

Key points

Learning combines a gradient-based method (preconditioned scaled conjugate gradient, building on Lowd & Domingos) with samples from Lazy-MC-SAT, and the approach is built on the open-source Alchemy package.
On MUC-6, the full model MLN-HAN (better heads, apposition, predicate nominals) scores about 9% above Haghighi & Klein (2007) and about 6% above the supervised McCallum & Wellner (2005). Its B³ results are 77.4 precision, 67.6 recall and 72.2 F1.
Even when trained on single documents, the system's F1 is more than 7% higher than Haghighi & Klein's model trained on 60 documents. It also beats that model trained on 381 documents.
On ACE-2004, MLN-HAN reaches a B³ F1 of 70.0 (BNEWS) and 72.3 (NWIRE). On ACE-2 it clearly beats the supervised Ng (2005), ties Denis & Baldridge (2007) on NWIRE, and trails it somewhat on BNEWS and NPAPER.
To show the gains are not an artifact of merging all mentions, the authors also report pairwise precision and the error in cluster counts. Remaining errors mostly involve nominals and proper nouns with differing heads, such as "Mr. Bush" vs. "the White House".

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Topics: Coreference resolution

Cited by

68 papers of our library cite it (of the 140673 whose bibliographies we read); 25 of them have a page here, 43 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{poon2008joint,
  title = {{Joint Unsupervised Coreference Resolution with Markov Logic}},
  author = {Poon, Hoifung and Domingos, Pedro},
  booktitle = {Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing},
  year = {2008},
  doi = {10.3115/1613715.1613796},
  url = {https://doi.org/10.3115/1613715.1613796}
}

References

19 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

Amit, B. & Baldwin, B. 1998. Algorithms for scoring coreference chains. In Proc. MUC-7. find our card
Bakir, G.; Hofmann, T.; Schölkopf, B.; Smola, A.; Taskar, B. and Vishwanathan, S. (eds.) 2007. Predicting Structured Data. MIT Press. find
Culotta, A.; Wick, M.; Hall, R. and McCallum, A. 2007. First-order probabilistic models for coreference resolution. In Proc. NAACL-07. find our card
Denis, P. & Baldridge, J. 2007. Joint determination of anaphoricity and coreference resolution using integer programming. In Proc. NAACL-07. find our card
Getoor, L. & Taskar, B. (eds.) 2007. Introduction to Statistical Relational Learning. MIT Press. find
Haghighi, A. & Klein, D. 2007. Unsupervised coreference resolution in a nonparametric Bayesian model. In Proc. ACL-07. find our card
Kautz, H.; Selman, B.; and Jiang, Y. 1997. A general stochastic approach to solving problems with hard and soft constraints. In The Satisfiability Problem: Theory and Applications. AMS. find
Klein, D. & Manning, C. 2003. Accurate unlexicalized parsing. In Proc. ACL-03. find our card
Kok, S.; Singla, P.; Richardson, M.; Domingos, P.; Sumner, M.; Poon, H. & Lowd, D. 2007. The Alchemy system for statistical relational AI. http://alchemy.cs.washington.edu/. ↗
Lowd, D. & Domingos, D. 2007. Efficient weight learning for Markov logic networks. In Proc. PKDD-07. find
Luo, X.; Ittycheriah, A.; Jing, H.; Kambhatla, N. and Roukos, S. 2004. A mention-synchronous coreference resolution algorithm based on the bell tree. In Proc. ACL-04. find our card
McCallum, A. & Wellner, B. 2005. Conditional models of identity uncertainty with application to noun coreference. In Proc. NIPS-04. find
Ng, V. 2005. Machine Learning for Coreference Resolution: From Local Classification to Global Ranking. In Proc. ACL-05. find our card
Poon, H. & Domingos, P. 2006. Sound and efficient inference with probabilistic and deterministic dependencies. In Proc. AAAI-06. find
Poon, H. & Domingos, P. 2007. Joint inference in information extraction. In Proc. AAAI-07. find
Poon, H. & Domingos, P. 2008. A general method for reducing the complexity of relational inference and its application to MCMC. In Proc. AAAI-08. find
Richardson, M. & Domingos, P. 2006. Markov logic networks. Machine Learning 62:107–136. find
Vilain, M.; Burger, J.; Aberdeen, J.; Connolly, D. & Hirschman, L. 1995. A model-theoretic coreference scoring scheme. In Proc. MUC-6. find our card
Wei, W.; Erenrich, J. and Selman, B. 2004. Towards efficient sampling: Exploiting random walk strategies. In Proc. AAAI-04. find

↑ back to top

📄 Joint Unsupervised Coreference Resolution with Markov Logic