Coreference Resolution with ILP-based Weighted Abduction

Proceedings of COLING 2012: Technical Papers, pages 1291–1308, COLING 2012, Mumbai, December 2012. Coreference Resolution with ILP-based Weighted Abduction Naoya Inoue1 Ekaterina Ovchinnikova2 Kentaro Inui1 Jerr y Hobbs2 (1) Tohoku University, 6-6-05 Aoba, Aramaki, Aoba-ku, Sendai, 980-8579, Japan (2) USC/ISI, 4676 Admiralty Way, Marina del Rey, CA 90292, USA naoya-i@ecei.tohoku.ac.jp,katya@isi.edu,inui@ecei.tohoku.ac.jp,hobbs@isi.edu ABSTRACT This paper presents the first systematic study of the coreference resolution problem in a general inference-based discourse processing framework. Employing the mode of inference called weighted abduction, we propose a novel solution to the overmerging problem inherent to inference-based frameworks. The overmerging problem consists in erroneously assuming distinct entities to be identical. In discourse processing, overmerging causes establishing wrong coreference links. In order to approach this problem, we extend Hobbs et al. (1993)’s weighted abduction by introducing weighted unification and show how to learn the unification weights by applying machine learning techniques. For making large-scale processing and parameter learning in an abductive logic framework feasible, we employ a new efficient implementation of weighted abduction based on Integer Linear Programming. We then propose several linguistically motivated features for blocking incorrect unifications and employ different large-scale world knowledge resources for establishing unification via inference. We provide a large-scale evaluation on the CoNLL-2011 shared task dataset, showing that all features and almost all knowledge components improve the performance of our system. KEYWORDS: weighted abduction, coreference resolution, Integer Linear Programming. 1291 1 Introduction In this paper, we explore coreference resolution in a discourse processing framework based on a mode of inference called weighted abduction (Hobbs et al., 1993). This framework is appealing because it is a realization of the observation that we understand new material by linking it with what we already know. It instantiates in natural language understanding the more general principle that we understand our environment by coming up with the best explanation for the observables in the environment. Hobbs et al. (1993) show that the lowest-cost abductive proof provides the solution to a whole range of natural language pragmatics problems, such as word sense disambiguation, anaphora and metonymy resolution, interpretation of noun compounds and prepositional phrases and detection of discourse relations. For examples of the application of weighted abduction to discourse processing see (Charniak and Goldman, 1991; Inoue and Inui, 2011; Ovchinnikova et al., 2011; Ovchinnikova, 2012). If weighted abduction is applied to discourse processing, coreference links naturally follow as a by-product of constructing best explanations. In weighted abduction, coreference resolution is equal to unification of predications; see Sec. 3.1. Similarly, if deductive model building is applied to discourse interpretation, coreference links result from the model minimality. Both inference approaches are based on the idea that predications having the same predicates describe the same situation and therefore their arguments can be assumed to be equal if no logical contradictions follow. If the necessary knowledge is missing from the knowledge base, both the deductive and the abductive procedures are likely to miss relevant coreference links and establish wrong links (overmerge entities). The overmerging problem is a serious obstacle in applying reasoning …

Content truncated to 500 words