Intonational Features of Local and Global Discourse Structure

1. ABSTRACT We present results of a study of the relationship between intonational features including pitch range, timing, and amplitude and aspects of discourse structure defined in terms of Grosz and Sidner's (1986) model of discourse. We compare structural labelings of AP news text with prosodic/acoustic features examined from recordings of the same text read by a professional newscaster. We find significant correlations between prosodic/acoustic characteristics and both local and global aspects of discourse structure identified by our labelers. Our results have applications for speech synthesis and, potentially, for speech recognition. The hypothesis that discourse structure is signalled by variation in intonational features such as pitch range, timing, and amplitude has been examined in studies such as [1, 2, 3, 4, 5, 6, 7]. However, as Brown and her colleagues note [2, p. 27]: "... until an independent theory of topic-structure is formulated, much of our argument in this area is in danger of circularity." In this paper we examine the relationship between discourse structure and variation in intonational features using just such an independent model of discourse structure, that proposed by Grosz and Sidner [8] (G&S). We present results of an empirical study comparing intonational features of read text with elements of both the local and global structure of discourse. Our study has immediate application to the generation of appropriate intonational features for synthetic speech, and future applicability to the recognition of discourse structure in speech recognition tasks. Our corpus consisted of AP news stories recorded by a professional speaker. The intonational features we considered included pitch range, contour, timing, and amplitude. The discourse structural elements we examined at the local level included parentheticals, quotations, tags, and indirect reported speech; at the global level, we studied discourse segmentation-the division of a discourse into constituents that provide the basis for determining discourse meaning. The discourses were labeled by two groups: one group labeled from text; the other group labeled from text while listening to the recorded speech. In this paper, we describe similarities and differences in the segmentations elicited in these two conditions. Our experiments provide support for three hypotheses. First, instructions can be devised, based on the G&S model, that enable subjects to analyze discourses with considerable similarity. Second, discourse structure is marked intonationally, although the relationship between structure and intonational features is a complex one; a given discourse structural feature may be sig-naled by several intonational features, either …