A Mixed Deterministic Model for Coreference Resolution

Proceedings of the Joint Conference on EMNLP and CoNLL: Shared Task, pages 76–82, Jeju Island, Korea, July 13, 2012. c⃝2012 Association for Computational Linguistics A Mixed Deterministic Model for Coreference Resolution Bo Yuan1, Qingcai Chen, Yang Xiang, Xiaolong Wang2 Liping Ge, Zengjian Liu, Meng Liao, Xianbo Si Intelligent Computing Research Center, Key Laboratory of Network Oriented Intelligent Computation, Computer Science and technology Department, Harbin Institute of Technology Shenzhen graduate School, Shenzhen, Guangdong, 518055, China {yuanbo.hitsz1, windseedxy, qingcai.chen, geliping123, autobotsonearth, dream2009gd, sixianbo}@gmail.com wangxl@insun.hit.edu.cn2 Abstract This paper presents a mixed deterministic model for coreference resolution in the CoNLL-2012 shared task. We separate the two main stages of our model, mention detection and coreference resolution, into several sub-tasks which are solved by machine learning method and deterministic rules based on multi-filters, such as lexical, syntactic, semantic, gender and number information. We participate in the closed track for English and Chinese, and also submit an open result for Chinese using tools to generate the required features. Finally, we reach the average F1 scores 58.68, 60.69 and 61.02 on the English closed task, Chinese closed and open tasks. 1 Introduction The coreference resolution task is a complicated and challenging issue of natural language processing. Although many sub-problems, such as noun phrase to noun phrase and pronouns to noun phrase, are contained in this issue, it is interesting that humans do not get too confused when they determine whether two mentions refer to the same entity. We also believe that automatic systems should copy the human behavior (Kai-Wei et al., 2011). In our understanding, the basis for human making judgment on different sub-problems is different and limited. Although there are some complicated and ambiguous cases in this task, and we are not able to cover all the prior knowledge of human mind, which plays a vital role in his solution, the mixed deterministic model we constructed can solve a big part of this task. We present a mixed deterministic model for coreference resolution in the CoNLL-2012 shared task (Sameer et al., 2011). Different methods such as Relaxation labeling (Emili et al., 2011), Best-Link (Kai-Wei et al., 2011), Entropy Guided Transformation Learning (Cicero et al., 2011) and deterministic models (Heeyoung et al., 2011), were attempted in the CoNLL-2011 shared task (Sameer et al., 2011). The system performance reported by the task shows that a big part of this task has been solved but some sub-problems need more exploration. We also participate in the Chinese closed and open tracks. However, the lack of linguistic annotations makes it more difficult to build a deterministic model. Basic solutions such as Hobbs Algorithm and Center Theory have been listed in (Wang et al., 2002; Jun et al., 2007). The recent research on Chinese contains non-anaphors detection using a composite kernel (Kong Fang, et al., 2012(a)) and a tree kernel method to anaphora resolution of pronouns (Kong Fang et al., 2012(b)). We accept the thought of Stanford (Karthik et al., 2010; Heeyoung et al., 2011). In Stanford system the coreference resolution task …

Content truncated to 500 words