A Latent Variable Model for Discourse-aware Concept and Entity Disambiguation

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 491–500, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics A Latent Variable Model for Discourse-aware Concept and Entity Disambiguation Angela Fahrni and Michael Strube Heidelberg Institute for Theoretical Studies gGmbH Schloss-Wolfsbrunnenweg 35 69118 Heidelberg, Germany (angela.fahrni|michael.strube)@h-its.org Abstract This paper takes a discourse-oriented per- spective for disambiguating common and proper noun mentions with respect to Wikipedia. Our novel approach mod- els the relationship between disambigua- tion and aspects of cohesion using Markov Logic Networks with latent variables. Considering cohesive aspects consistently improves the disambiguation results on various commonly used data sets. 1 Introduction “I have to review a paper”, the super- visor moaned from the office. “Please don’t disturb me until I’m done with the review.” His student nodded, went to the cafeteria, sat down in the sunshine and started to read yesterday’s paper. This text snippet illustrates two aspects that have been neglected by previous disambiguation ap- proaches. (1) The interpretation of different men- tions, i.e. common and proper nouns, is deter- mined by different notions of context: some men- tions depend more on a local sentence-level con- text (paper in read yesterday’s paper; the global context is misleading), some more on a global one (review in I’m done with the review; the local con- text is not discriminative), some on both global and local context (paper in review a paper). (2) The context relevant to disambiguate a mention depends on how it is embedded into discourse and is not bound to the surface form of a mention (pa- per in the first sentence vs. paper in the last one). Starting from this observation, we argue that the context relevant to disambiguate a mention cor- relates with its cohesive scope, i.e. the text span within which a mention establishes cohesive re- lations. Therefore, we propose to disambiguate mentions differently depending on their cohesive scopes (Section 2). We distinguish between three different cohesive scopes of mentions and model them as latent variables using Markov Logic Net- works (Section 3). The use of latent variables al- lows us to learn and predict the cohesive scope and the disambiguation of a mention jointly. This comes with the advantage that the learning of the scope assignment does not need annotated data by itself but is guided by the annotations available for the target prediction task, i.e. the disambiguation. In this paper, we focus on concept and entity disambiguation1 with respect to an inventory de- rived from Wikipedia and compare (1) to a state- of-the-art approach that treats all mentions alike and uses the same features for disambiguation, (2) to a pipeline-based approach, and (3) to other state-of-the-art approaches (Section 4). While early work disambiguated concepts us- ing the local context (Csomai and Mihalcea, 2008), current research focuses on exploiting the global document context (Milne and Witten, 2008; Kulkarni et al., 2009; Ratinov et al., 2011; Fahrni and Strube, 2012; Cheng and Roth, 2013). Al- though such …

Content truncated to 500 words