A framework for utterance disambiguation in dialogue

A framework for utterance disambiguation in dialogue Pont Lurcock, Peter Vlugter, Alistair Knott Department of Computer Science University of Otago New Zealand {pont,pvlugter,alik}@cs.otago.ac.nz Abstract We discuss the data sources available for ut- terance disambiguation in a bilingual dialogue system, distinguishing global, contextual, and user-specific domains, and syntactic and seman- tic levels. We propose a framework for combin- ing the available information, and techniques for increasing a stochastic grammar’s sensitivity to local context and a speaker’s idiolect. 1 Introduction Resolving the ambiguities present in an incom- ing utterance is a key task in natural language processing. Interpreting an utterance, whether semantically, syntactically or phonologically, is typically construed as a two-stage process: the first stage involves deriving a set of all possi- ble analyses, using relatively well-defined prin- ciples, and the second stage involves selecting between these analyses, using principles that are harder to define and formalize. This paper considers principles for disam- biguating utterances in a human-machine dia- logue system. Our main goal is to present a framework for integrating the different sources of information relevant to this task, and the available techniques for making use of this in- formation. Working with a dialogue system highlights the need for such a framework; there are additional types of ambiguity that need to be considered (such as dialogue act ambiguity), and a particularly diverse set of informational resources to consult (many of which relate to the current dialogue context). At the same time, there are additional opportunities available—in particular, the ability to ask the user clarifica- tion questions if a given ambiguity cannot be otherwise resolved. We begin in Section 2 with an analysis of the kinds of disambiguation needed in our dialogue system, and of the information available to the system to perform the task. In Section 3, we outline a general framework for performing dis- ambiguation in a dialogue system. In Sections 4 and 5, we describe how our system implements this framework. We conclude in Section 6 with an account of our current work. 2 Sources of information available for utterance disambiguation Utterance interpretation is typically modelled as a pipeline process, beginning with phono- logical interpretation, proceeding with syntactic analysis and semantic analysis, and concluding with discourse or dialogue attachment. We will illustrate by showing the pipeline used in our own bilingual English/M¯aori dialogue system, Te Kaitito (see e.g. Knott and Wright (2003); Knott et al. (2004)). Our pipeline is shown in Figure 1. Since our input is a written sentence Dialogue act identification Sentence parsing Presupposition resolution Figure 1: The utterance interpretation pipeline we begin with sentence parsing, using the LKB system (Copestake and Flickinger, 2000). LKB works with HPSG-like grammars; the gram- mar we use is bilingual, simultaneously encod- ing a fragment of English and M¯aori (Knott et al., 2002). LKB’s parser delivers ‘flat’ seman- tic representations in a format called Minimal Recursion Semantics (MRS) (Copestake et al., 1999), which are turned into Discourse Repre- sentation Structures (DRSs) distinguishing be- tween an assertion and a set of presuppositions …

Content truncated to 500 words