Cascading Collective Classification for Bridging Anaphora Recognition using a Rich Linguistic Feature Set

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 814–820, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Cascading Collective Classification for Bridging Anaphora Recognition Using a Rich Linguistic Feature Set Yufang Hou1, Katja Markert2, Michael Strube1 1 Heidelberg Institute for Theoretical Studies gGmbH, Heidelberg, Germany (yufang.hou|michael.strube)@h-its.org 2School of Computing, University of Leeds, UK scskm@leeds.ac.uk Abstract Recognizing bridging anaphora is difficult due to the wide variation within the phenomenon, the resulting lack of easily identifiable surface markers and their relative rarity. We develop linguistically motivated discourse structure, lexico-semantic and genericity detection fea- tures and integrate these into a cascaded mi- nority preference algorithm that models bridg- ing recognition as a subtask of learning fine- grained information status (IS). We substan- tially improve bridging recognition without impairing performance on other IS classes. 1 Introduction In bridging or associative anaphora (Clark, 1975; Prince, 1981; Gundel et al., 1993), the antecedent and anaphor are not coreferent but are linked via a variety of contiguity relations.1 In Example 1, the phrases a resident, the stairs and the lobby are bridg- ing anaphors with the antecedent One building.2 (1) One building was upgraded to red status while peo- ple were taking things out, and a resident called up the stairs to his girlfriend, telling her to keep sending things down to the lobby. Bridging is an important problem as it affects lin- guistic theory and applications alike. For exam- ple, without bridging resolution, entity coherence between the first and second coordinated clause in 1We exclude comparative anaphora where anaphor and an- tecedent are in a similarity/exclusion relation, indicated by ana- phor modifiers such as other or similar (Modjeska et al., 2003). 2Examples are from OntoNotes (Weischedel et al., 2011). Bridging anaphora are set in boldface; antecedents in italics. Example 1 cannot be established. This is a prob- lem both for coherence theories such as Centering (Grosz et al., 1995) (where bridging is therefore in- corporated as an indirect realization of previous en- tities) as well as applications relying on entity co- herence modelling, such as readability assessment or sentence ordering (Barzilay and Lapata, 2008). Full bridging resolution needs (i) recognition that a bridging anaphor is present and (ii) identification of the antecedent and contiguity relation. In re- cent work, these two tasks have been tackled sep- arately, with bridging recognition handled as part of information status (IS) classification (Markert et al., 2012; Cahill and Riester, 2012; Rahman and Ng, 2012). Each mention in a text gets assigned one IS class that describes its accessibility to the reader at a given point in a text, bridging being one possible class. We stay within this framework. Bridging recognition is a difficult task, so that we had to report very low results on this IS class in pre- vious work (Markert et al., 2012). This is due to the phenomenon’s variety, leading to a lack of clear sur- face features for recognition. Instead, we formulate in this paper novel discourse structure …

Content truncated to 500 words