Distributional Semantics for Resolving Bridging Mentions

Proceedings of Recent Advances in Natural Language Processing, pages 192–199, Hissar, Bulgaria, Sep 7–9 2015. Distributional Semantics for Resolving Bridging Mentions Tim Feuerbach and Martin Riedl and Chris Biemann Language Technology Group, CS Department, TU Darmstadt, Germany uni@spell.work, {riedl, biem}@cs.tu-darmstadt.de Abstract We explore the impact of adding distri- butional knowledge to a state-of-the-art coreference resolution system. By inte- grating features based on word and context expansions from a distributional thesaurus (DT), automatically mined IS-A relation- ships and shallow syntactical clues into the Berkeley system (Durrett and Klein, 2013), we are able to increase its F1 score on bridging mentions, i.e. coreferent men- tions with non-identical heads, by 8.29 points. Our semantic features improve over the Web-based features of Bansal and Klein (2012). Since bridging mentions are a hard but infrequent class of coreference, this leads to merely small improvements in the overall system. 1 Introduction Automatically recognizing coreference – relating lexical items that refer to the same entity or con- text in a text – is an important semantic processing step for text understanding tasks such as fact ex- traction, information retrieval, and entity linking. A common problem of coreference systems is their inability to resolve bridging mentions, i.e. coreferent mentions with non-identical heads (Vieira and Poesio, 2000). For example, a system requires semantic knowledge to detect the hyper- nymic relationship that holds between mentions like a preliminary agreement and the pact. Simi- larly, modeling selectional preference relies on in- formation beyond the pronoun context itself. There are two different kinds of approaches em- ployed in the past to make this knowledge avail- able as features to a coreference resolution system. The first class uses manually crafted resources like WordNet or Wikipedia (Poesio et al., 2004; Ponzetto and Strube, 2006). Despite their quality, they may decrease the performance when added to the system (Lee et al., 2011; Zhou et al., 2011). Further disadvantages are their limited size, slow growth and general-purpose nature. In contrast, using unsupervised/semi-supervised methods for generating knowledge is only limited by the size of input data and adapts to the target domain. We present features exploiting automatically obtained distributional knowledge, following the distributional hypothesis formulated by Harris (1954) that words in similar contexts bear simi- lar meanings. For that we resort to a distribu- tional thesaurus (DT; Lin, 1998) listing semanti- cally similar terms, as well as hyponym-hypernym relations (IS-As) acquired with Hearst patterns (Hearst, 1992), both made available by the JoBim- Text Project (Biemann and Riedl, 2013). When added to the state-of-the-art Berkeley Corefer- ence Resolution System (Durrett and Klein, 2013), these features show a significant positive impact on bridging mentions. 2 Related Work Our work is very similar to Bansal and Klein (2012), who created, among others, features based on IS-As, distributional clusters, and pronoun con- texts. However, we chose to use a DT’s list of sim- ilar words instead of clustering, and dependency relations as context features instead of N-gram neighborhood. We will compare our approach to Bansal and Klein’s features below. Distributional methods …

Content truncated to 500 words