Joint Learning for Coreference Resolution with Markov Logic

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1245–1254, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Joint Learning for Coreference Resolution with Markov Logic Yang Song1, Jing Jiang2, Wayne Xin Zhao3, Sujian Li1, Houfeng Wang1 1Key Laboratory of Computational Linguistics (Peking University) Ministry of Education,China 2School of Information Systems, Singapore Management University, Singapore 3School of Electronics Engineering and Computer Science, Peking University, China {ysong, lisujian, wanghf}@pku.edu.cn, jingjiang@smu.edu.sg, batmanfly@gmail.com Abstract Pairwise coreference resolution models must merge pairwise coreference decisions to gen- erate final outputs. Traditional merging meth- ods adopt different strategies such as the best- first method and enforcing the transitivity con- straint, but most of these methods are used independently of the pairwise learning meth- ods as an isolated inference procedure at the end. We propose a joint learning model which combines pairwise classification and mention clustering with Markov logic. Experimen- tal results show that our joint learning sys- tem outperforms independent learning sys- tems. Our system gives a better performance than all the learning-based systems from the CoNLL-2011 shared task on the same dataset. Compared with the best system from CoNLL- 2011, which employs a rule-based method, our system shows competitive performance. 1 Introduction The task of noun phrase coreference resolution is to determine which mentions in a text refer to the same real-world entity. Many methods have been pro- posed for this problem. Among them the mention- pair model (McCarthy and Lehnert, 1995) is one of the most influential ones and can achieve the state- of-the-art performance (Bengtson and Roth, 2008). The mention-pair model splits the task into three parts: mention detection, pairwise classification and mention clustering. Mention detection aims to iden- tify anaphoric noun phrases, including proper nouns, common noun phrases and pronouns. Pairwise clas- sification takes a pair of detected anaphoric noun phrase candidates and determines whether they re- fer to the same entity. Because these classification decisions are local, they do not guarantee that can- didate mentions are partitioned into clusters. There- fore a mention clustering step is needed to resolve conflicts and generate the final mention clusters. Much work has been done following the mention- pair model (Soon et al., 2001; Ng and Cardie, 2002). In most work, pairwise classification and mention clustering are done sequentially. A major weak- ness of this approach is that pairwise classification considers only local information, which may not be sufficient to make correct decisions. One way to address this weakness is to jointly learn the pair- wise classification model and the mention cluster- ing model. This idea has been explored to some extent by McCallum and Wellner (2005) using con- ditional undirected graphical models and by Finley and Joachims (2005) using an SVM-based super- vised clustering method. In this paper, we study how to use a different learning framework, Markov logic (Richardson and Domingos, 2006), to learn a joint model for both pairwise classification and mention clustering un- der the mention-pair model. We …

Content truncated to 500 words