Collective Entity Resolution with Multi-Focal Attention

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 621–631, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Collective Entity Resolution with Multi-Focal Attention Amir Globerson∗and Nevena Lazic and Soumen Chakrabarti† and Amarnag Subramanya and Michael Ringgaard and Fernando Pereira Google, Mountain View CA, USA gamir@post.tau.ac.il, nevena@google.com, soumen@cse.iitb.ac.in, {asubram, ringgaard, pereira}@google.com Abstract Entity resolution is the task of linking each mention of an entity in text to the cor- responding record in a knowledge base (KB). Coherence models for entity resolu- tion encourage all referring expressions in a document to resolve to entities that are related in the KB. We explore attention- like mechanisms for coherence, where the evidence for each candidate is based on a small set of strong relations, rather than relations to all other entities in the doc- ument. The rationale is that document- wide support may simply not exist for non-salient entities, or entities not densely connected in the KB. Our proposed sys- tem outperforms state-of-the-art systems on the CoNLL 2003, TAC KBP 2010, 2011 and 2012 tasks. 1 Introduction Entity resolution (ER) is the task of mapping men- tions of entities in text to corresponding records in a knowledge base (KB) (Bunescu and Pasca, 2006; Cucerzan, 2007; Kulkarni et al., 2009; Dredze et al., 2010; Hoffart et al., 2011; Hachey et al., 2013). ER is a challenging problem because mentions are often ambiguous on their own, and can only be resolved given appropriate context. For example, the mention Beirut may refer to the capital of Lebanon, the band from New Mexico, or a drink- ing game (Figure 1). Names may also refer to en- tities that are not in the KB, a problem known as NIL detection. Most ER systems consist of a mention model, a context model, and a coherence model (Milne and Witten, 2008; Cucerzan, 2007; Ratinov et al., ∗Currently at Tel Aviv University † Currently at IIT Bombay 2011; Hoffart et al., 2011; Hachey et al., 2013). The mention model associates each entity with its possible textual representations (also known as aliases or surface forms). The context model helps resolve an ambiguous mention using textual fea- tures extracted from the surrounding context. The coherence model, the focus of this work, encour- ages all mentions to resolve to entities that are re- lated to each other. Relations may be established via the KB, Web links, embeddings, or other re- sources. Coherence models often define an objective function that includes local and pairwise candi- date scores, where the pairwise scores correspond to some notion of coherence or relation strength.1 Support for a candidate is typically aggregated over relations to all other entities in the document. One problem with this approach is that it may di- lute evidence for entities that are not salient in the document, or not well-connected in the KB. Our work aims to address this issue. We introduce a novel coherence model with an attention mechanism, where the score for each candidate only depends on …

Content truncated to 500 words