The Theory and Use of Clarification Requests in Dialogue

Clarification requests are an important, relatively common and yet under-studied dialogue device allowing a user to ask about some feature (e.g. the mea ning or form) of an utterance, or part thereof. They can take many different forms (often hi ghly elliptical) and can have many different meanings (requesting various types of infor mation). This thesis combines empirical, theoretical and implementational work to provi de a study of the various types of clarification request that exist, give a theoretical analys is thereof, and show how the results can be applied to add useful capabilities to a prototype comp utational dialogue system. A series of empirical studies (corpus-based and experiment al) are described which establish a taxonomy of the possible types of clarification reques t together with information about their meaning and usage, about the phrase types and conditio ns that trigger them and their particular forms and interpretations, and about the likely methods of responding to them. A syntactic and semantic analysis using the HPSG framework i s given which extends the work of (Ginzburg and Cooper, 2004) to cover the main classes of the above taxonomy, and to account for the clarificational potential of those word an d phrase types which commonly cause clarification requests. This is shown to have interest ing implications for the semantics of various lexical and phrasal types, in particular suggest ing hat noun phrases be given a simple witness-set based representation. Finally, the theoretical analysis and empirical findings ar e pplied within a HPSG grammar and a prototype text-based dialogue system, CLARIE. Imp le ented in Prolog using the TrindiKit, the system combines the information-state-bas ed dialogue management of GoDiS (Larsson et al., 2000) and the HPSG-based ellipsis resoluti on of SHARDS (Ginzburg et al., 2001a) and adds the capability to interpret and respond to us er clarification requests, and generate its own clarifications where necessary to deal with inc omprehensible or contradictory input, resolve unknown or ambiguous reference, and learn ou t-of-vocabulary words.