System paper for CoNLL-2012 shared task: Hybrid Rule-based Algorithm for Coreference Resolution.

Proceedings of the Joint Conference on EMNLP and CoNLL: Shared Task, pages 118–121, Jeju Island, Korea, July 13, 2012. c⃝2012 Association for Computational Linguistics Hybrid Rule-based Algorithm for Coreference Resolution ∗ Heming Shou1,2 Hai Zhao1,2† 1Center for Brain-Like Computing and Machine Intelligence, Department of Computer Science and Engineering, Shanghai Jiao Tong University 2MOE-Microsoft Key Laboratory for Intelligent Computing and Intelligent Systems Shanghai Jiao Tong University shouhm@gmail.com, zhaohai@cs.sjtu.edu.cn Abstract This paper describes our coreference resolu- tion system for the CoNLL-2012 shared task. Our system is based on the Stanford’s dcore- f deterministic system which applies multiple sieves with the order from high precision to low precision to generate coreference chains. We introduce the newly added constraints and sieves and discuss the improvement on the o- riginal system. We evaluate the system using OntoNotes data set and report our results of average F-score 58.25 in the closed track. 1 Introduction In this paper, our coreference resolution system for CoNLL-2012 shared task (Pradhan et al., 2012) is summarized. Our system is an extension of Stan- ford’s multi-pass sieve system, (Raghunathan et al., 2010) and (Lee et al., 2011), by adding novel con- straints and sieves. In the original model , sieves are sorted in decreasing order of precision. Initially each mention is in its own cluster. Mention clusters are combined by satisfying the condition of each sieve in the scan pass. Through empirical studies, we pro- posed some extensions and algorithms for further- more enhancing the performance. ∗This work was partially supported by the National Natural Science Foundation of China (Grant No. 60903119 and Grant No. 61170114), the National Research Foundation for the Doc- toral Program of Higher Education of China under Grant No. 20110073120022, the National Basic Research Program of Chi- na (Grant No. 2009CB320901) and the European Union Sev- enth Framework Program (Grant No. 247619). †corresponding author Many other existing systems applied supervised or unsupervised (Haghighi and Klein, 2010) learn- ing models. The classical resolution algorithm was proposed by (Soon et al., 2001). Semantic knowl- edge like word associations was involved by (Kob- dani et al., 2011). Most of the supervised learning models in CoNLL-2011 shared task (Chang et al., 2011)(Bj¨orkelund and Nugues, 2011) used classi- fiers (Maximum Entropy or SVM) to train the mod- els for obtaining the pairwise mention scores. How- ever, the training process usually takes much longer time than unsupervised or deterministic systems. In contrast, (Raghunathan et al., 2010) proposed a rule- based model which obtained competitive result with less time. Two considerable extensions to the Stanford mod- el in this paper are made to guarantee higher pre- cision and recall. First, we recorded error pattern- s from outputs of the original Stanford system and found that the usual errors are mention boundary mismatches, pronoun mismatches and so on. To avoid the irrational coreference errors, we added some constraints to the mention detection for elim- inating some unreasonable mention boundary mis- matches. Second, we added some constraints in the coreference sieves based on the errors on the …

Content truncated to 500 words