A development of a sensible dialogue agent that acquires personas from user utterances
Kazuki Kondo, Takuto Sakuma, Shōhei Kato · 2023
With the recent development of natural language processing, the demand for personalized dialogue agents has increased. One way to personalize an agent is training a language model using a persona chat dataset incorporating the user’s information as personas. However, this method is inconsistent for dialogues unrelated to the persona, and the pre-defined persona does not change. This research aims to make an agent friendlier to users by adding personas extracted from the user’s utterances as well as the agent’s utterances. We conducted two user experiments and reconfirmed the improvement in performance by adding personas. We discuss the personality and friendliness acquired by the proposed method.