Responsive Utterance Generation to Show Disagreement to Narrative in Attentive Listening

Abstract

Conversational agents, such as communication robots, are expected to play a role in listening to narratives. To recognize these robots as the listeners, it is essential for them to have a function that indicates their attentive listening to a narrative. The basic response strategy in attentive listening involves generating responsive utterances that show acceptance. However, the narratives occasionally contain self-deprecation or modesty. In such cases, the listener must be able to produce a response that shows disagreement with the narrator's utterance, that is, a disagreement response. This study demonstrates the feasibility of generating appropriate disagreement responses. First, we define a method to tag the timing and expression of the disagreement response to narrative data in a non-real-time environment, and verify that the method enables to construct a corpus to which the disagreement response timing and response expression are comprehensively and stably assigned, respectively. Next, we implement detection and classification methods for disagreement response timing and expressions, respectively, based on a pre-trained transformer-based model and verify the feasibility of generating disagreement responses using the response corpus through experiments.