Prosodic cues for backchannels and short questions: Really?

Short questions can be ambiguous even after considering their preceding contexts. Hence, prosody may be useful for disambiguating different types of questions and their uses. For example, question bias has been linked to the presence of certain pitch accents. This paper presents a corpus study of very short questions and the contribution of prosodic cues to discourse disambiguation. This study focuses on backchannel questions which are by nature highly biased and yet sit between genuine questions and genuine backchannels. The study finds LDA and SVM classifiers do not perform better than random at separating backchannel and question really based on these prosodic cues. This means that, while intonation differs between broad categories of questions, theories that try to integrate prosodic cues with semantics and discourse require more than intonation, the final rise and the other usual prosodic suspects like duration and intensity.