Prosody and discourse structure: Issues and experiments

In this report we summarize the approach of examining structural relations between prosody and discourse semantics by means of experimental phonetics.1 We sketch the computational basis of the project, its method-ological setting, and we present the results of three exper-iments on: • the role of local prominence in disambiguation of focal adverbs • the role of relative prominence in anaphora resolution • the role of global prominence [register features] in discourse structure. 1. METHODS Research on prosody, and particularly on discourse pros-ody, requires lots of data and explicit theories of dis-course and of prosody to generalize over this data. In our project data has been recorded mainly from Digital Radio [Deutschlandfunk, 32 Khz, 16 bit, digital broadcast] and from commercially available literature readings in CD quality [1]. In order to process hours of data computa-tional tools have been developed which allow for the automatic labelling and processing of prosody.2 On the theoretical side explicit linguistic theories of dynamic semantics have been used- in case of our project the choice fell on DRT (Discourse Representation Theory [2]). The prosody was modelled within the TSM framework (Tone Sequence Model [3]). 1.1. Computational prerequisites In order to guarantee the highest possible compatibility all the tools for description, processing, synthesis and recognition of prosody have been integrated into the ESPS/waves environment. Prosodic description is based on a set of ToBI like labels, which have been developed and tested on german intona-tional structures [4], [5]. The set of labels has been 1. Research described in this study has been supported by a grant from the Deutsche Forschungsgemeinschft (DFG). 2. For details on our speech database go to