Learning Noun Phrase Anaphoricity to Improve Conference Resolution: Issues in Representation and Optimization

Learning Noun Phrase Anaphoricity to Improve Coreference Resolution: Issues in Representation and Optimization Vincent Ng Department of Computer Science Cornell University Ithaca, NY 14853-7501 yung@cs.cornell.edu Abstract Knowledge of the anaphoricity of a noun phrase might be profitably exploited by a coreference sys- tem to bypass the resolution of non-anaphoric noun phrases. Perhaps surprisingly, recent attempts to incorporate automatically acquired anaphoricity in- formation into coreference systems, however, have led to the degradation in resolution performance. This paper examines several key issues in com- puting and using anaphoricity information to im- prove learning-based coreference systems. In par- ticular, we present a new corpus-based approach to anaphoricity determination. Experiments on three standard coreference data sets demonstrate the ef- fectiveness of our approach. 1 Introduction Noun phrase coreference resolution, the task of de- termining which noun phrases (NPs) in a text refer to the same real-world entity, has long been con- sidered an important and difficult problem in nat- ural language processing. Identifying the linguis- tic constraints on when two NPs can co-refer re- mains an active area of research in the commu- nity. One significant constraint on coreference, the non-anaphoricity constraint, specifies that a non- anaphoric NP cannot be coreferent with any of its preceding NPs in a given text. Given the potential usefulness of knowledge of (non-)anaphoricity for coreference resolution, anaphoricity determination has been studied fairly extensively. One common approach involves the design of heuristic rules to identify specific types of (non-)anaphoric NPs such as pleonastic pro- nouns (e.g., Paice and Husk (1987), Lappin and Le- ass (1994), Kennedy and Boguraev (1996), Den- ber (1998)) and definite descriptions (e.g., Vieira and Poesio (2000)). More recently, the problem has been tackled using unsupervised (e.g., Bean and Riloff (1999)) and supervised (e.g., Evans (2001), Ng and Cardie (2002a)) approaches. Interestingly, existing machine learning ap- proaches to coreference resolution have performed reasonably well without anaphoricity determination (e.g., Soon et al. (2001), Ng and Cardie (2002b), Strube and M¨uller (2003), Yang et al. (2003)). Nev- ertheless, there is empirical evidence that resolution systems might further be improved with anaphoric- ity information. For instance, our coreference sys- tem mistakenly identifies an antecedent for many non-anaphoric common nouns in the absence of anaphoricity information (Ng and Cardie, 2002a). Our goal in this paper is to improve learning- based coreference systems using automatically computed anaphoricity information. In particular, we examine two important, yet largely unexplored, issues in anaphoricity determination for coreference resolution: representation and optimization. Constraint-based vs. feature-based representa- tion. How should the computed anaphoricity information be used by a coreference system? From a linguistic perspective, knowledge of non- anaphoricity is most naturally represented as “by- passing” constraints, with which the coreference system bypasses the resolution of NPs that are deter- mined to be non-anaphoric. But for learning-based coreference systems, anaphoricity information can be simply and naturally accommodated into the ma- chine learning framework by including it as a fea- ture in the instance representation. Local vs. global optimization. Should the anaphoricity determination procedure be developed independently of the …

Content truncated to 500 words