Multilingual Mention Detection for Coreference Resolution

International Joint Conference on Natural Language Processing, pages 100–108, Nagoya, Japan, 14-18 October 2013. Multilingual Mention Detection for Coreference Resolution Olga Uryupina DISI, University of Trento, Italy uryupina@gmail.com Alessandro Moschitti QCRI, Qatar Foundation, Doha, Qatar amoschitti@qf.org.qa Abstract This paper proposes a novel algorithm for multilingual mention detection: we ex- tract mentions from parse trees via kernel- based SVM learning. Our approach al- lows for straightforward mention detection for any language where (not necessary per- fect) parsing resources are available, with- out any complex language-specific rule engineering. We also investigate possi- bilities for incorporating automatically ac- quired mentions into an end-to-end coref- erence resolution system. We evaluate our approach on the Arabic and Chinese por- tions of the CoNLL-2012 dataset, showing a significant improvement over the system with the baseline mention detection. 1 Introduction Accurate mention detection (MD) is a vital prereq- uisite for a variety of Natural Language Processing tasks, in particular, for Relation Extraction (RE) and Coreference Resolution (CR). If a toolkit can- not extract mentions reliably, it will obviously be unable to assign them to relations or entities. Many studies on RE and CR report evaluation figures on gold mentions: in such a setting, a sys- tem is supplied with correct mention boundaries and/or semantic classes or other relevant prop- erties. It can, in theory, be argued that such a methodology provides better insights on perfor- mance of RE and CR algorithms per se. It has been demonstrated, however, that evaluation results on gold mentions are misleading: for example, Ng (2008) shows that unsupervised CR algorithms ex- hibit promising results on gold mentions, that are not mirrored in a more realistic evaluation on au- tomatically detected mentions. The exact scope of the mention detection task varies considerably depending on the annotation guidelines. Thus, some corpora consider all the (non-embedding) NPs to be mentions, some cor- pora do not allow for non-referential mentions and some do not mark singleton referential mentions, that do not participate in coreference relations. In addition, some guidelines may restrict the annota- tion to specific semantic types. A number of linguistic studies focus on various syntactic, semantic and discourse clues that might help identify nominal constructions that cannot participate in coreference relations. Possible fea- tures include, among others, specific syntactic constructions for expletive pronouns, negation, modality and quantification (Karttunen, 1976). Several algorithms have been proposed recently, trying to tackle some of the addressed phenom- ena within a computational approach. Thus, a number of algorithms have been developed re- cently to identify expletive usages of “it” (Evans, 2001; Boyd et al., 2005; Bergsma and Yarowsky, 2011). While these approaches are potentially beneficial for mention detection in English, for other languages, neither theoretical nor compu- tational studies are available at the moment. In this paper, we use tree kernels to extract relevant syntactic patterns automatically, without assuming any prior knowledge of the input language. In this paper, we propose a learning-based so- lution to the mention detection task. We use SVMs (Joachims, 1999) with syntactic tree ker- …

Content truncated to 500 words