C3EL: A Joint Model for Cross-Document Co-Reference Resolution and Entity Linking

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 846–856, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. C3EL: A Joint Model for Cross-Document Co-Reference Resolution and Entity Linking Sourav Dutta Max-Planck Institute for Informatics Saarbr¨ucken, Germany sdutta@mpi-inf.mpg.de Gerhard Weikum Max-Planck Institute for Informatics Saarbr¨ucken, Germany weikum@mpi-inf.mpg.de Abstract Cross-document co-reference resolution (CCR) computes equivalence classes over textual mentions denoting the same entity in a document corpus. Named-entity link- ing (NEL) disambiguates mentions onto entities present in a knowledge base (KB) or maps them to null if not present in the KB. Traditionally, CCR and NEL have been addressed separately. However, such approaches miss out on the mutual syn- ergies if CCR and NEL were performed jointly. This paper proposes C3EL, an unsuper- vised framework combining CCR and NEL for jointly tackling both problems. C3EL incorporates results from the CCR stage into NEL, and vice versa: additional global context obtained from CCR improves the feature space and performance of NEL, while NEL in turn provides distant KB fea- tures for already disambiguated mentions to improve CCR. The CCR and NEL steps are interleaved in an iterative algorithm that focuses on the highest-confidence still un- resolved mentions in each iteration. Ex- perimental results on two different corpora, news-centric and web-centric, demonstrate significant gains over state-of-the-art base- lines for both CCR and NEL. 1 Introduction With the advent of large knowledge bases (KB) like DBpedia, YAGO, Freebase, and others, enti- ties (people, places, organizations, etc.) along with their attributes and relationships form the basis of smart applications like search, analytics, recom- mendations, question answering, and more. The major task that arises in both the KB construc- tion process and the entity-centric applications in- volves precise recognition, resolution, and link- ing of named entities distributed across web pages, news articles, and social media. Named Entity Recognition (NER) deals with the identification of entity mentions in a text and their classification into coarse-grained semantic types (person, location, etc.) (Finkel et al., 2005; Nadeau & Sekine, 2007; Ratinov & Roth, 2009). This in- volves segmentation of token sequences to obtain mention boundaries, and mapping relevant token spans to pre-defined entity categories. For exam- ple, NER on the text Einstein won the Nobel Prize identifies the mentions “Einstein” and “No- bel Prize” and marks them as person and misc type, respectively. Named Entity Linking (NEL)1 involves the dis- ambiguation of textual mentions, based on context and semantic information, and their mapping to proper entities in a KB (Bunescu & Pas¸ca, 2006; Cucerzan, 2007; Milne & Witten, 2008; Hoffart et al., 2011; Ratinov et al., 2011; Cornolti et al., 2013). For example, in the above text, the mention “Einstein” is linked to the physicist Albert Einstein. Entity Co-reference Resolution (CR) (Haghighi & Klein, 2010; Ng, 2010; Lee et al., 2013) is essentially a clustering task to identify mentions (and anaphoras) within a document referring to the same entity, thus computing equivalence classes or mention groups. For example, mentions Albert Einstein and Nobel …

Content truncated to 500 words