Latent Structure Perceptron with Feature Induction for Unrestricted Coreference Resolution

Proceedings of the Joint Conference on EMNLP and CoNLL: Shared Task, pages 41–48, Jeju Island, Korea, July 13, 2012. c⃝2012 Association for Computational Linguistics Latent Structure Perceptron with Feature Induction for Unrestricted Coreference Resolution Eraldo Rezende Fernandes Departamento de Inform´atica PUC-Rio Rio de Janeiro, Brazil efernandes@inf.puc-rio.br C´ıcero Nogueira dos Santos Brazilian Research Lab IBM Research Rio de Janeiro, Brazil cicerons@br.ibm.com Ruy Luiz Milidi´u Departamento de Inform´atica PUC-Rio Rio de Janeiro, Brazil milidiu@inf.puc-rio.br Abstract We describe a machine learning system based on large margin structure perceptron for unre- stricted coreference resolution that introduces two key modeling techniques: latent corefer- ence trees and entropy guided feature induc- tion. The proposed latent tree modeling turns the learning problem computationally feasi- ble. Additionally, using an automatic feature induction method, we are able to efficiently build nonlinear models and, hence, achieve high performances with a linear learning algo- rithm. Our system is evaluated on the CoNLL- 2012 Shared Task closed track, which com- prises three languages: Arabic, Chinese and English. We apply the same system to all lan- guages, except for minor adaptations on some language dependent features, like static lists of pronouns. Our system achieves an offi- cial score of 58.69, the best one among all the competitors. 1 Introduction The CoNLL-2012 Shared Task (Pradhan et al., 2012) is dedicated to the modeling of coreference resolution for multiple languages. The participants are provided with corpora for three languages: Ara- bic, Chinese and English. These corpora are pro- vided by the OntoNotes project and, besides accu- rate anaphoric coreference information, contain var- ious annotation layers such as part-of-speech (POS) tagging, syntax parsing, named entities (NE) and se- mantic role labeling (SRL). The shared task consists in the automatic identification of coreferring men- tions of entities and events, given predicted infor- mation on other OntoNotes layers. We propose a machine learning system for coref- erence resolution that is based on the large margin structure perceptron algorithm (Collins, 2002; Fer- nandes and Milidi´u, 2012). Our system learns a pre- dictor that takes as input a set of candidate men- tions in a document and directly outputs the clus- ters of coreferring mentions. This predictor com- prises an optimization problem whose objective is a function of the clustering features. To embed clas- sic cluster metrics in this objective function is prac- tically infeasible since most of such metrics lead to NP-hard optimization problems. Thus, we introduce coreference trees in order to represent a cluster by a directed tree over its mentions. In that way, the prediction problem optimizes over trees instead of clusters, which makes our approach computationally feasible. Since coreference trees are not given in the training data, we assume that these structures are la- tent and use the latent structure perceptron (Fernan- des and Brefeld, 2011; Yu and Joachims, 2009) as the learning algorithm. To provide high predicting power features to our model, we use entropy guided feature induc- tion (Fernandes and Milidi´u, 2012). By using this technique, we automatically generate several fea- ture templates that …

Content truncated to 500 words