Data-driven Multilingual Coreference Resolution using Resolver Stacking

Proceedings of the Joint Conference on EMNLP and CoNLL: Shared Task, pages 49–55, Jeju Island, Korea, July 13, 2012. c⃝2012 Association for Computational Linguistics Data-driven Multilingual Coreference Resolution using Resolver Stacking Anders Bj¨orkelund and Rich´ard Farkas Institute for Natural Language Processing University of Stuttgart {anders,farkas}@ims.uni-stuttgart.de Abstract This paper describes our contribution to the CoNLL 2012 Shared Task.1 We present a novel decoding algorithm for coreference res- olution which is combined with a standard pair-wise coreference resolver in a stacking approach. The stacked decoders are evaluated on the three languages of the Shared Task. We obtain an official overall score of 58.25 which is the second highest in the Shared Task. 1 Introduction In this paper we present our contribution to the CoNLL 2012 Shared Task (Pradhan et al., 2012). We follow the standard architecture where mentions are extracted in the first step, then they are clustered using a pair-wise classifier (see e.g., (Ng, 2010)). For English, the set of extracted mentions is filtered by removing non-referential occurrences of certain pronouns. Our coreference resolver at its core re- lies on a pair-wise classifier. To overcome the prob- lems associated with the isolated pair-wise deci- sions, we devised a novel decoding algorithm which compares a mention to partially built clusters. For our Shared Task contribution we combined this al- gorithm with conventional pair-wise decoding algo- rithms in a stacking approach. In the Shared Task evaluation, our system re- ceived an overall official score of 58.25, which is the second highest among the sixteen participants.2 1The system is available for download on http://www. ims.uni-stuttgart.de/˜anders/ 2The overall score is the average of MUC, B3, and CEAFE, averaged over all three languages 2 Mention Extraction Since all mentions are not annotated in Shared Task data, but only mentions that take part in coreference chains, training a general-purpose anaphoricity clas- sifier is non-trivial. We thus implemented a high- recall, low-precision mention extraction module that allows the coreference resolver to see most of the possible mentions, but has to learn to sort out the non-referential mentions. The mention extraction module relies mainly on the syntactic parse tree, but also on named entities (which were only provided for English in the pre- dicted versions of the Shared Task data). Since the part-of-speech tags vary a bit across the languages, so do our extraction rules: For Arabic, we extract all NP’s, and all terminals with part-of- speech tags PRP and PRP$; for Chinese, we extract all NP’s, and all terminals with part-of-speech tags PN and NR; for English, we extract all NP’s, all ter- minals with part-of-speech tags PRP and PRP$, and all named entities. Early experiments indicated that the English coreference resolver frequently makes mistakes re- lated to non-referential instances of the pronouns it (often referred to as expletive or pleonastic in the lit- erature), we, and you (generic mentions, which are not annotated according to the OntoNotes annota- tion guidelines). To address this issue, we developed a referential/non-referential mention classifier in order to identify these mentions. The …

Content truncated to 500 words