Mapping Syntax and Prosody

The relationship between prosodic structure and syntactic structure has remained a controversial and unresolved area, partly due to the lack of rich corpora of natural speech, and partly due to the complexity involved in both syntax and prosody. Chom-sky & Halle (1968, p. 372) state that " although there is a substantial literature on intonational and prosodic features in English, it is largely restricted to citation of examples, and we cannot draw on it for any significant insight into processes of a general nature. " More recently, Ladd (1996, p. 334) adds that " in the standard theory, the correspondence between syntactic constituent types and prosodic ones is highly variable, since the make-up of the prosodic constituents is influenced by a variety of essentially linear factors. " Despite the status quo, different views on the mapping from syntax to prosody have been proposed: (1) Syntax alone determines most of prosodic structure (2) speakers use prosody that signals the syntactic information only when ambiguity is involved (Snedeker & Truswell 2003, Allbritton et al. 1996); (3) Many linguistic and para-linguistic factors along with syntax determine a prosodic struc-Besides the linguistic and psycholinguistic studies, machine learning approaches have been employed to improve the performance of TTS or ASR either using hand-built rules Much linguistic and psycholinguistic research is limited in that it has often relied on data from intuition , or small collections of recorded speech. As for machine learning approaches using syntactic parser, a concern is reflected by Taylor & Black (1998) who argue that " [a]lthough we argued. .. against using syntactic parsers for phrase break assignment, our reason stem from the basic inaccuracy of these parsers, not because syntactic parsers themselves are unhelpful. " Thus, one way to overcome the limitation of dataset and the inaccuracy of the full syntactic parser is to employ machine learning approaches on a large corpus of natural speech with features that are more accurate than those of a full syntactic parser, and at the same time that contains richer syntactic structural information than part of speech. The outcome of the parser is, then, similar to the " flattened syntactic structure " (Chomsky & Halle 1968; Langendoen 1975). This paper presents an experimental prediction of prosodic information (pitch accents and boundary tones) based on shallow syntactic structure and grammatical relations, together with part of speech, basic syllable information and constituent length. The working hypothesis is …