Reference Resolution in Context

This paper sketches an approach to pronoun reference resolution in context based on a dynamic incremental semantics for NL in polymorphic type theory. Our set-up provides full incrementality of processing, and can handle salience and pronoun resolution in context. An implementation of the system in Haskell, in ‘literate programming’ style, exists. The full literate source code can be found at http://www.cwi.nl/ jve/papers/02/rric. 1 Salience An important aspect of natural language text understanding is the resolution of anaphoric links: finding out what the pronouns that occur in the text refer to. Anaphoric reference resolution is incremental in the sense that, in general, the information that is needed to determine the reference of a pronoun in a text can be determined ‘on the go’, on the basis of: • Syntactic properties of the sentence that contains the pronoun. • Information conveyed in the previous discourse (including the introduction of salient referents, with salience determined in part by syntax). • Background information shared by speaker and hearer in an act of communication (the common ground of speaker and hearer). Surface syntactic form is an important determinant for salience. It is usually assumed, e.g., that subject is more salient than object. In John hit Bill. He was upset., the first choice for resolving the pronoun He is John, because John is the subject of the preceding sentence. In Bill got kicked by John. He was upset, the first choice for resolving He is Bill, because Bill is the subject of the preceding sentence. In both cases, John and Bill are the two obvious candidates for resolving the reference of the pronoun He, because both John and Bill have been made salient by the preceding text. There are many ways to increase one’s understanding about reference resolution in context. One method, the method to be used in this paper, is to focus on the simplest possible set-up, and to construct a formal picture of what is needed for the process of reference resolution. Just as in logic one can study reasoning by constructing and analyzing reasoning systems, so we can study reference resolution by constructing and analyzing systems for reference resolution. The set-up we focus on is a simplification in many ways, but this should be considered as a virtue rather than a defect. A respectable way to proceed in the formal sciences is to first form a clear picture of an idealized reality, and then gradually proceed to take more factors into account.