A Coherence-Based Approach to the Interpretation of Non-Sentential Utterances in Dialogue

Abstract

This thesis is concerned with the syntax, compositional semantics and contextually-situated interpretation of a certain kind of non-sentential utterance occuring in dialogue, namely one where the utterance, despite its ‘incomplete’ syntactic form, is intended to convey a proposition, a question or a request. Perhaps the most prominent type of such utterances is the short answer, as in “A: Who came to the party? — B: Peter.”, but there are many other types as well. Following (Morgan 1973) and (Ginzburg 1999b) and others, we will call such utterances fragments. Clearly, the interpretation of fragments is highly context dependent. We will provide evidence that there are complex syntactic, semantic and pragmatic constraints governing the use of fragments. In particular, following (Ginzburg 1999b), we will present evidence that while the main resolution must be semantic, some limited syntactic information nevertheless has to persist beyond the boundaries of sentences to allow for the formulation of certain constraints on fragments. We will argue that consequently only a theory that has at its disposal a wide array of information sources —from syntax through compositional and lexical semantics to domain and world knowledge, and reasoning about cognitive states— can do justice to the complexity of their interpretation. As we will show, however, it is desirable to encapsulate these knowledge sources as much as possible, in order to maintain computability. Our main thesis then is that the resolution of the intended content of fragments can be modelled as a by-product of the establishment of coherence in dialogue, which (following much of the work on discourse) we define as the establishment of certain connections of the content of the current utterance to the content of its discourse context. We will show that all constraints on the form and content of fragments follow from how they are connected to the context. The central role of discourse coherence in our account of fragments, together with having access to different kinds of information, distinguishes our theory from prior attempts. The work of Jonathan Ginzburg and colleagues ((Ginzburg 1999b, Ginzburg & Sag 2001) inter alia), for example, provides an approach to some types of fragments which is based on unification-operations on HPSG-signs. This approach, as we will show, fails to offer a convincing model of the interpretation of fragments where missing content is linguistically implicit and has to be inferred. Carberry (1990), on the other hand, employs computationally expensive plan-recognition techniques for the interpretion of fragments. This fails to predict certain empirical facts and we will furthermore show that the complex reasoning with cognitive states that she employs can often be replaced with much simpler inferences based on linguistic information. In this thesis, we offer an analysis of the syntax and compositional semantics of fragments, and we provide a computational and formally precise theory of how the compositional semantics is supplemented with further content via reasoning about the context—both linguistic and non-linguistic. We also describe an implementation of our approach, based on an extension of a wide-coverage grammar and an accompanying discourse reasoning component for a simple domain.