Identifying relevant cues for uncertainty in dialogue
- Master Thesis
- Announcement date
- 07 Oct 2013
- Tobias Schrank
- Research Areas
Uncertainty is ubiquitous in natural human communication. While human listeners are used to assessing the speaker’s degree of uncertainty at
any time, they are fairly good at it, too. In contrast, currently available computer systems dealing with spoken language are usually
not built to perform this task. As humans make use of such paralinguistic information quite heavily to shape communication, computer systems such as dialogue systems are likely to benefit from its use as well.
To detect uncertainty, I combine features that have been mentioned in the literature and features derived from my analysis of a corpus of
naturalistic task-oriented spoken German, the Kiel Corpus of Spontaneous Speech. Particular focus is put on the comparison between
linguistic features (e.g., lexical items, dialogue structure) and acoustic features (e.g., intensity contour, pause durations).
Finally, a classifier based on Support Vector Machines is built using these features. Experimental results conducted with this classifier
show that a relatively high proportion of uncertainty can be correctly detected (F1 .70, accuracy .74) if uninformative features are
eliminated. A classifier that employs all features performs considerably worse.