Simple Maximum Entropy Models for Multilingual Coreference Resolution

Proceedings of the Joint Conference on EMNLP and CoNLL: Shared Task, pages 83–87, Jeju Island, Korea, July 13, 2012. c⃝2012 Association for Computational Linguistics Simple Maximum Entropy Models for Multilingual Coreference Resolution Xinxin Li, Xuan Wang, Xingwei Liao Computer Application Research Center Harbin Institute of Technology Shenzhen Graduate School Shenzhen, China lixxin2@gmail.com Abstract This paper describes our system participat- ing in the CoNLL-2012 shared task: Mod- eling Multilingual Unrestricted Coreference in Ontonotes. Maximum entropy models are used for our system as classifiers to deter- mine the coreference relationship between ev- ery two mentions (usually noun phrases and pronouns) in each document. We exploit rich lexical, syntactic and semantic features for the system, and the final features are selected us- ing a greedy forward and backward strategy from an initial feature set. Our system partici- pated in the closed track for both English and Chinese languages. 1 Introduction In this paper, we present our system for the CoNLL- 2012 shared task which aims to model coreference resolution for multiple languages. The task of coref- erence resolution is to group different mentions in a document into coreference equivalent classes (Prad- han et al., 2012). Plenty of machine learning al- gorithms such as Decision tree (Ng and Cardie, 2002), maximum entropy model, logistic regres- sion (Bj¨orkelund and Nugues, 2011), Support Vec- tor Machines, have been used to solve this problem. Meanwhile, the CoNLL-2011 shared task on En- glish language show that a well-designed rule-based approach can achieve a comparable performance as a statistical one (Pradhan et al., 2011). Our system treats coreference resolution problem as classification problem by determining whether every two mentions in a document has a corefer- ence relationship or not. We use maximum entropy (ME) models to train the classifiers. Previous work reveal that features play an important role on coref- erence resolution problem, and many different kinds of features has been exploited. In this paper, we use many different lexical, syntactic and semantic fea- tures as candidate features, and use a greedy forward and backward approach for feature selection for ME models. 2 System Description The framework of our system is shown in figure 1. It includes four components: candidate mention selec- tion, training example generation, model generation, and decoding algorithm for test data. The details of each component as described below. 2.1 Candidate Mention Selection In both training and test sets, our system only con- sider all noun phrases (NP) and pronouns (PRP, PRP$) as candidate mentions for both English and Chinese. The mentions in each sentence are ob- tained from given syntactic tree by their syntactic label. Other phrases in the syntactic tree are omit- ted due to their small proportion. For example, in the English training dataset, our candidate mentions includes about 91% of golden mentions. 2.2 Training Example Generation There are many different training example gen- eration algorithms, e.g., McCarthy and Lehnert’s method, Soon et al.s method, Ng and Cardies method (Ng, 2005). For our baseline system, we choose Soon et al.’s method because it …

Content truncated to 500 words