Tagging Prosody and Discourse Structure in Elicited Spontaneous Speech

Abstract

This paper motivates and describes the annotation and analysis of prosody and discourse structure for several large spoken language corpora. The annotation schema are of two types: tags for prosody and intonation, and tags for several aspects of discourse structure. The choice of the particular tagging schema in each domain is based in large part on the insights they provide in corpus-based studies of the relationship between discourse structure and the accenting of referring expressions in American English. We first describe these results and show that the same models account for the accenting of pronouns in an extended passage from one of the Speech Warehouse hotel-booking dialogues. We then turn to corpora described in Venditti [Ven00], which adapts the same models to Tokyo Japanese. Japanese is interesting to compare to English, because accent is lexically specified and so cannot mark discourse focus in the same way. Analyses of these corpora show that local pitch range expansion serves the analogous focusing function in Japanese. The paper concludes with a section describing several outstanding questions in the annotation of Japanese intonation which corpus studies can help to resolve.