Using Syntactic Dependencies to Solve Coreferences

Proceedings of the Joint Conference on EMNLP and CoNLL: Shared Task, pages 64–70, Jeju Island, Korea, July 13, 2012. c⃝2012 Association for Computational Linguistics Using Syntactic Dependencies to Solve Coreferences Marcus Stamborg Dennis Medved Peter Exner Pierre Nugues Lund University Lund, Sweden cid03mst@student.lu.se, dt07dm0@student.lth.se Peter.Exner@cs.lth.se, Pierre.Nugues@cs.lth.se Abstract This paper describes the structure of the LTH coreference solver used in the closed track of the CoNLL 2012 shared task (Pradhan et al., 2012). The solver core is a mention classifier that uses Soon et al. (2001)’s algorithm and features extracted from the dependency graphs of the sentences. This system builds on Bj¨orkelund and Nugues (2011)’s solver that we extended so that it can be applied to the three languages of the task: English, Chinese, and Arabic. We designed a new mention detection module that removes pleonastic pronouns, prunes constituents, and recovers mentions when they do not match ex- actly a noun phrase. We carefully redesigned the features so that they reflect more com- plex linguistic phenomena as well as discourse properties. Finally, we introduced a minimal cluster model grounded in the first mention of an entity. We optimized the feature sets for the three lan- guages: We carried out an extensive evalua- tion of pairs of features and we complemented the single features with associations that im- proved the CoNLL score. We obtained the re- spective scores of 59.57, 56.62, and 48.25 on English, Chinese, and Arabic on the develop- ment set, 59.36, 56.85, and 49.43 on the test set, and the combined official score of 55.21. 1 Introduction In this paper, we present the LTH coreference solver used in the closed track of the CoNLL 2012 shared task (Pradhan et al., 2012). We started from an earlier version of the system by Bj¨orkelund and Nugues (2011), to which we added substantial im- provements. As base learning and decoding algo- rithm, our solver extracts noun phrases and posses- sive pronouns and uses Soon et al. (2001)’s pairwise classifier to decide if a pair corefers or not. Similarly to the earlier LTH system, we constructed a primary feature set from properties extracted from the depen- dency graphs of the sentences. 2 System Architecture The training and decoding modules consist of a mention detector, a pair generator, and a feature ex- tractor. The training module extracts a set of positive and negative pairs of mentions and uses logistic re- gression and the LIBLINEAR package (Fan et al., 2008) to generate a binary classifier. The solver ex- tracts pairs of mentions and uses the classifier and its probability output, Pcoref(Antecedent, Anaphor), to determine if a pair corefers or not. The solver has also a post processing step to recover some mentions that do not match a noun phrase constituent. 3 Converting Constituents to Dependency Trees Although the input to coreference solvers are pairs or sets of constituents, many systems use concepts from dependency grammars to decide if a pair is coreferent. The most frequent one is the con- stituent’s head that solvers need then …

Content truncated to 500 words