A Multigraph Model for Coreference Resolution

Proceedings of the Joint Conference on EMNLP and CoNLL: Shared Task, pages 100–106, Jeju Island, Korea, July 13, 2012. c⃝2012 Association for Computational Linguistics A Multigraph Model for Coreference Resolution Sebastian Martschat, Jie Cai, Samuel Broscheit, ´Eva M´ujdricza-Maydt, Michael Strube Natural Language Processing Group Heidelberg Institute for Theoretical Studies gGmbH Heidelberg, Germany (sebastian.martschat|jie.cai|michael.strube)@h-its.org Abstract This paper presents HITS’ coreference reso- lution system that participated in the CoNLL- 2012 shared task on multilingual unrestricted coreference resolution. Our system employs a simple multigraph representation of the rela- tion between mentions in a document, where the nodes correspond to mentions and the edges correspond to relations between the mentions. Entities are obtained via greedy clustering. We participated in the closed tasks for English and Chinese. Our system ranked second in the English closed task. 1 Introduction Coreference resolution is the task of determining which mentions in a text refer to the same entity. This paper describes HITS’ system for the CoNLL- 2012 Shared Task on multilingual unrestricted coref- erence resolution, where the goal is to build a system for coreference resolution in an end-to-end multilin- gual setting (Pradhan et al., 2012). We participated in the closed tasks for English and Chinese and fo- cused on English. Our system ranked second in the English closed task. Being conceptually similar to and building upon Cai et al. (2011b), our system is based on a directed multigraph representation of a document. A multi- graph is a graph where two nodes can be connected by more than one edge. In our model, nodes rep- resent mentions and edges are built from relations between the mentions. The entities to be inferred correspond to clusters in the multigraph. Our model allows for directly representing any kind of relations between pairs of mentions in a graph structure. Inference over this graph can har- ness structural properties and the rich set of encoded relations. In order to serve as a basis for further work, the components of our system were designed to work as simple as possible. Therefore our system relies mostly on local information between pairs of mentions. 2 Architecture Our system is implemented on top of the BART toolkit (Versley et al., 2008). To compute the coref- erence clusters in a document, we first extract a set of mentions M = {m1, . . . , mn} ordered according to their position in the text (Section 2.1). We then build a directed multigraph where the set of nodes is M and edges are induced by relations between mentions (Section 2.4). The relations we use in our system are coreference indicators like string match- ing or alias (Section 3). For every relation R, we compute a weight wR using the training data (Sec- tion 2.3). We then assign the weight wR to any edge that is induced by the relation R. Depending on dis- tance and connectivity properties of the graph the weights may change (Section 2.4.1). Given the con- structed graph with edge weights, we go through the mentions …

Content truncated to 500 words