Identifying dialogue acts: linguistics or statistics?

At some point in their processing, natural language dialogue systems typically perform some pragmatic analysis to compute the dialogue act (or speech act, dialogue move, illo-cution, etc.) underlying an utterance. The Verbmobil system, despite the fact that it is not engaged in the dialogue but only mediates it, is no exception. At present, both symbolic reasoning and statistical recognition are used for this purpose. This paper discusses the relative merits of both approaches and of various ways of integrating the two, which are currently subject to experimentation. 1 Dialogue acts in Verbmobil Natural language dialogue systems typically analyze utterances on various levels of abstraction , and often the \deepest" representation they construct involves dialogue acts (DAs). These are constructs that abstract from speciic lexical, syntactic, and semantic information, and represent only the pragmatic force that is relevant for the processing system. In the Verbmobil project Wahlster 1993], DAs are organized in a hierarchy reeecting their speci-city, and | to a certain extent | domain-dependency: the most speciic DAs are geared towards the task of appointment scheduling. Figure 1 shows the hierarchy of DAs used in phase I of the project Jekat et al. 1995]. At present, the list is being revised for phase II; nonetheless, we can base the following discussion on the hierarchy shown in the gure. Since Verbmobil employs several lines of analysis, varying on the scale from \deep" to \shal-low", diierent methods for recognizing the DA of an utterance are in use. Speciically, within the DIAKON component (dialogue processing and contextual evaluation) both statistical and rule-based recognition schemes are at work. The following two sections brieey describe these schemes. Then, section four examines the relative merits of the competing approaches and discusses possibilities for integrating them into a single framework. 2 Recognizing dialogue acts, linguistically In Verbmobil, the contextual evaluation (ConEval) module is given the syntactic/semantic representation of an utterance and asked to perform certain disambiguations and to compute 1