Speaker-independent context update rules for dialogue management

This paper describes a dialogue management system in which an attempt is made to factor out a declarative theory of context updates in dialogue from a procedural theory of generating and interpreting utterances in dialogue. 1 Background: declarative and procedural resources for text processing In computational linguistics, a very useful distinction can be drawn between declarative models of language, which specify what constitutes a ‘correct’ or ‘proper’ sentence or discourse, and procedural models, which specify how a proper sentence or discourse can be interpreted or generated. This distinction is virtually ubiquitous in computational models of sentence structure and sentence processing. In these models, a sentence grammar is a declarative construct, and a sentence parser or sentence generator consults the grammar in a systematic way at each iteration in order to interpret or create sentences. The idea of systematicity is very important. For instance, in a chart parser the process of sentence parsing is broken down into a sequence of procedural operations each of which has exactly the same general form: a search of the set of grammar rules, and the creation of a new chart edge if the search is successful. In fact, the benefits of thinking of a parser as a procedural module consulting a declarative grammatical resource are most clearly seen in the fact that the procedural component can be expressed systematically in this way. The declarative/procedural distinction is also increasingly common in computational treatments of extended monologues. There are several overtly declarative theories of the structure of such texts (many of them stemming from the work of Mann and Thompson (1988) and Grosz and Sidner (1986)), and several models of text generation and text interpretation which make reference to these declarative theories (see e.g. Hovy (1993) and Marcu (2000) for a summary of generation and interpretation methods respectively). Again, the most attractive feature of the declarative/procedural distinction is that the procedural algorithms envisaged are very systematic. In models of dialogue structure, a clean separation between declarative and procedural models has proved more elusive. By analogy with the cases of sentences and monologic discourse just described, what is required is a declarative model of ‘wellformed dialogue’, together with procedural mechanisms that consult this model in a systematic way to produce and interpret contributions to a dialogue. To begin with, what is a declarative theory of dialogue structure? In this paper, we will assume a theoretical perspective in which dialogue moves are represented as context-update operations (see e.g. Traum et al (1999)). An utterance in a dialogue is understood as a function which (when defined) takes the current dialogue context and outputs a new dialogue context that constitutes the input to the next utterance in the dialogue. The declarative theory of dialogue coherence will therefore be a theory about the legal ways in which the dialogue context can be utterance interpretation Dialogue context utterance processing utterance processing interlocutor (I) utterance processing rules A’s utterance I’s utterance utterance generation A B