Towards Speaker Adaptation for Dialogue Act Recognition

Dialogue act labels are being used to represent a higher level intention of utterances during human conversation (Stolcke et al., 2000). Automatic dialogue act recognition is still an active research topic. The conventional approach is to train one generic classifier using a large corpus of annotated utterances (Stolcke et al., 2000). One aspect that makes it so challenging is that people can express the same intentions using a very different set of spoken words. Imagine how different the vocabulary used by a native English speaker or a foreigner can be. Even more, people can have different intentions when using the exact same spoken words. These idiosyncratic differences in dialogue acts make the learning of generic classifiers extremely challenging. Luckily, in many applications such as face-to-face meetings or tele-immersion, we have access to archives of previous interactions with the same participants. From these archives, a small subset of spoken utterances can be efficiently annotated. As we will later show in our experiments, even a small number of annotated utterances can make a significant differences in the dialogue act recognition performance.