Intonation: Theory, Models, and Applications

Athens, Greece
September 18-20, 1997


Bootstrapping on Citation Form Accent Assignment

Christine H. Nakatani

AT&T Labs - Research, Florham Park, NJ, USA

A new approach to accent assignment is explored in which patterns of deviation from citation form accentuation, defined at the constituent level, are automatically learned from an annotated corpus. This constituent-based approach to accent prediction bootstraps off of established knowledge about citation-form accenting, and allows for the direct integration of higher-level linguistic features, such as form of referring expression and grammatical function, with lower-level linguistic features, such as lexical class and word lemma information. Machine learning experiments show that for two speakers from a corpus of spontaneous direction-giving monologues, accent assignment can be improved by up to 4%-6% relative to a hypothetical baseline system that would produce only citation-form accentuation, giving an error rate reduction of 11 %-24%.

Full Paper

Bibliographic reference.  Nakatani, Christine H. (1997): "Bootstrapping on citation form accent assignment", In INT-1997, 255-258.