Machine Learning for Mention Head Detection in Multilingual Coreference Resolution

Proceedings of Recent Advances in Natural Language Processing, pages 747–754, Hissar, Bulgaria, 7-13 September 2013. Machine Learning for Mention Head Detection in Multilingual Coreference Resolution Desislava Zhekova CIS, University of Munich zhekova@cis.uni-muenchen.de Sandra K¨ubler Indiana University skuebler@indiana.edu Abstract This work introduces a machine learn- ing approach to the identification of men- tion heads needed for multilingual coref- erence resolution (MCR). We evaluate the method and compare it to a heuristic base- line and a rule-based approach, which are widely used in coreference resolution sys- tems. We use the CoNLL-2012 shared task data sets, which include data for Ara- bic, Chinese, and English. We show that for MCR, machine learning offers a com- petitive, flexible, and robust solution for mention head detection. 1 Introduction Coreference Resolution (CR) aims to detect all lin- guistic expressions in a given discourse that refer to real world entities. Such expressions are gen- erally called mentions. They need to be grouped into equivalence classes so that each class contains only mentions that refer to the same entity. The classes are called coreference chains. The task of CR includes not only the identification of coref- erence links between mentions, but also the de- tection of the mentions themselves. This subtask of CR has not been a main topic of interest, since most of the standard data sets for CR contained gold mention information. This situation changed in the most recent shared tasks on the topic of CR: SemEval-2010 Task 1 (Recasens et al., 2010), CoNLL-2011 (Pradhan et al., 2011) and CoNLL- 2012 (Pradhan et al., 2012). The data distributed by these tasks included syntactic annotations, and it was considered an integral part of the task for the participating systems to develop their own meth- ods to detect mention boundaries. Statistical approaches to the CR problem often recast the task to a binary classification exercise. For the latter, coreference is represented by a deci- sion model, such as the mention-pair model (Soon et al., 2001). The mention-pair model, which is the most widely used model for CR, pairs the anaphor with a potential antecedent, and deter- mines whether they are coreferent or not. Since the decisions are taken independently for each possible antecedent, a global heuristic can be used to decide between multiple positive decisions or in cases where no antecedent was found. The use of the mention-pair model implies that an instance consists of a pair of mentions, and, since vectors for machine learning (ML) need to be of a fixed length, each mention is generally represented by its syntactic head, plus informative features that describe the phrases and their con- text. As a consequence, there is an additional sub- task of CR that needs to be performed before the actual resolution process: mention head detection (MHD). This is usually done by the use of simple heuristics or manually defined sets of rules (see section 2). In this work, we will investigate a novel ML method for multilingual MHD. Multilinguality has presented additional issues to the coreference …

Content truncated to 500 words