How does informativeness affect prosodic prominence?

Abstract

A key function of prosodic prominence is to mark the most informative words in an utterance. However, informativeness has been conceptualised as, e.g., focus, given/new status or predictability; it is not clear how these are related. Furthermore, prominence is constrained by metrical prosodic structure. We present a new framework for prominence production: informativeness and prosodic factors are constraints on the probabilistic alignment of words with metrical structure. Informativeness operates on two levels, focus and lexical “accentability” (predictability, part-of-speech). Foci align with nuclear accents, however, this is affected by prosodic and “accentability” constraints. Accent prediction models (nuclear, non-nuclear, or unaccented) are presented for the Switchboard corpus. Consistent with our predictions, nuclear accents are more likely later in a phrase, and on focused words. The likelihood of nuclear and non-nuclear accents is affected by prosodic constraints (e.g., rhythm) and “accentability”. The implications for the role of prosody in language production are discussed.

Keywords:

• Prosody
• Information structure
• Discourse processing

Acknowledgements

The research on which this work is based was completed as part of a Scottish Enterprise Edinburgh–Stanford Link grant. The author is now supported by a British Academy postdoctoral fellowship.

Thanks to Bob Ladd, Fernanda Ferreira and Mark Steedman for useful discussions in developing the ideas presented here and feedback on earlier drafts of this article. I am grateful to Jean Carletta, Jason Brenier, Neil Mayo, Dan Jurafsky, David Beaver, Jonathan Kilgour, and Shipra Dingare for their efforts in constructing the corpus, and Leonardo Badino for his Switchboard n-gram data. Also thanks to Martin Corley and an anonymous reviewer for useful suggestions on the statistical analysis and its presentation. Lastly, thanks to the audience at the Experimental and Theoretical Advances in Prosody conference at Cornell in April 2008 for encouraging feedback on the original presentation of this paper.