Detection of feeling through back-channels in spoken dialogue

We investigate the usage of back-channel information in the information navigation dialogue between an expert guide and a user. By back-channel feedback, we mean the user’s verbal short response, which expresses his state of the mind during the dialogue. Its prototypical lexical entries include “hai” in Japanese and “yes” or “right” in English, however, we do not count explicit affirmative responses as back-channels. Previously, there were several works[1, 2] which attempted to automatically generate back-channel responses for smooth communication between the user and the system. Recently, the back-channel information is included in the framework of dialogue act tagging in the game-playing dialogue[3] and meetings[4]. In the information navigation dialogue, in which an expert guide presents a list of recommendation spots, it is expected that the prosodic pattern of the back-channel conveys the para-linguistic information, that is, it suggests the positive/negative feeling on the recommended candidate. We also presume that the human expert guide detects such feelings expressed via back-channels, and chooses to continue the explanation of the current topic if the user seems interested, or change the topic otherwise. Thus, we investigate the back-channel patterns observed in the Kyoto Tour Guide Dialog Corpus.