Discovering the Sounds of Discourse Structure Extended Abstract

It is widely accepted that discourses are composed of segments and that the recognition of segment boundaries is essential to a determination of discourse meaning (Grosz and Sidner, 1986). Written language has orthographic cues such as section headings, paragraph boundaries, and punctuation which can assist in identifying discourse structure. In spoken language, into-national variation provides essential information about disconrse structure. For instance, it may be used to mark structural features of discourse at the global level, such as segment boundaries. Intonation also provides more local information about relations among utterances within a segment , for example indicating whether phrases are parenthetical. It can also help distinguish between different interpretations of phrases that can function either as cue phrases that indicate discourse segment bomldaries or sententially to convey domain information. Finally, variations in intona-tional prominence may be used to convey information about the discourse status of entities referred to by definite noun phrases and pronouns. An understanding of intonational variation and the ways in which it carries information about discourse characteristics of spoken language is important for computer-based interpretation and gener-atien of speech. From the interpretation perspective , this understanding may provide new techniques for identifying discourse structure. From the generation perspective, it would lead to more natural synthetic speech, making it possible to produce comtmter speech that is easier for people to understand and less susceptible to misinterpretation. Three major challenges have faced researchers attempting to discover the relationship between intonational features and the structure of spoken discourse. First, the collection of corpora of spontaneous speech has required the development of David Ahn provided invaluable technical musistance. new experimental methodologies. Whereas it is straightforward to have the same text read by many speakers, it is much more difficult to obtain similar samples of spontaneous speech from nmltiple speakers. Second, techniques must be developed to obtain reliable segmentations and la-belings of the corpora. Because discourse structure is rooted in semantics rather than syntax, this has proved more difficult than tagging corpora for sentence structure. Third, measures of agreement among segmentations must be designed. In this area too, the semantic nature of discourse structure leads to a more complex problem than comparing sentence parse structures. This talk will begin with a summary of pilot studies that demonstrated reliable correlations of discourse structure and intonational fea-It will then tbcus on a new corpus of direction-giving monologues, the Boston Directions Corpus (Nakatani et al., …