[3Xin4] Poster session 1

[3Xin4-44] Training dialogue response generation model using dialogue sentences with detailed empathy adjustment

〇Takeshi Homma¹, Tomoki Fujihara², Kanta Watanabe³, Yasuhiro Sogawa¹ (1.Hitachi, Ltd., 2.Tohoku Univ., 3.NAIST)

Keywords:dialogue system, empathy, end-to-end learning

We investigate a training method for dialogue response generation models that generate responses that make users perceive the empathy of the dialogue systems. Specifically, we train four response generation models by changing the empathy types in training sentences. As the empathy types, we use empathy dimensions proposed in the psychological research. We conduct a subjective evaluation of the generated sentences to compare evaluators' perceived empathy. The evaluation result shows that the models trained with sentences with empathy types of "personal distress" and "perspective taking" elicit the highest empathetic feeling to the evaluators.