To Memorize or to Predict: Prominence labeling in Conversational Speech

Proceedings of NAACL HLT 2007, pages 9–16, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics To Memorize or to Predict: Prominence Labeling in Conversational Speech A. Nenkova, J. Brenier, A. Kothari, S. Calhoun†, L. Whitton, D. Beaver, D. Jurafsky Stanford University {anenkova,jbrenier,anubha,lwhitton,dib,jurafsky}@stanford.edu †University of Edinburgh Sasha.Calhoun@ed.ac.uk Abstract The immense prosodic variation of natural con- versational speech makes it challenging to pre- dict which words are prosodically prominent in this genre. In this paper, we examine a new fea- ture, accent ratio, which captures how likely it is that a word will be realized as prominent or not. We compare this feature with traditional accent- prediction features (based on part of speech and N-grams) as well as with several linguistically mo- tivated and manually labeled information structure features, such as whether a word is given, new, or contrastive. Our results show that the linguistic fea- tures do not lead to significant improvements, while accent ratio alone can yield prediction performance almost as good as the combination of any other sub- set of features. Moreover, this feature is useful even across genres; an accent-ratio classifier trained only on conversational speech predicts prominence with high accuracy in broadcast news. Our results sug- gest that carefully chosen lexicalized features can outperform less fine-grained features. 1 Introduction Being able to predict the prominence or pitch accent status of a word in conversational speech is impor- tant for implementing text-to-speech in dialog sys- tems, as well as in detection of prosody in conversa- tional speech recognition. Previous investigations of prominence prediction from text have primarily relied on robust surface fea- tures with some deeper information structure fea- tures. Surface features like a word’s part-of-speech (POS) (Hirschberg, 1993) and its unigram and bi- gram probability (Pan and McKeown, 1999; Pan and 0Thanks to the Edinburgh-Stanford Link and ONR (MURI award N000140510388) for generous support. Hirschberg, 2000) are quite useful; content words are much more likely to be accented than function words, and words with higher probability are less likely to be prominent. More sophisticated linguis- tic features have also been used, generally based on information-structural notions of contrast, focus, or given-new. (Hirschberg, 1993). For example, in the Switchboard utterance be- low, there is an intrinsic contrast between the words “women” and “men”, making both terms more salient (words in all capital letters represent promi- nent tokens): you SEE WOMENc GOING off to WARS as WELL as MENc. Similarly the givenness of a word may help deter- mine its prominence. The speaker needs to focus the hearer’s attention on new entities in the discourse, so these are likely to be realized as prominent. Old en- tities, on the other had, need not be prominent; these tendencies can be seen in the following example. theyold have all the WATERnew theyold WANT. theyold can ACTUALLY PUMP waterold. While previous models have attempted to cap- ture global properties of words (via POS or unigram probability), they have not in general used word identity as a predictive feature, assuming either that …

Content truncated to 500 words