[4Yin2] Interactive session 2

[4Yin2-33] Style-controllable response generation model in chat dialogue

〇Itsuki Oishi¹, Masayasu Atsumi¹ (1.soka university)

Keywords:chat dialogue, style control, generative model, deep learning

In recent years, there has been active research on chat dialogue using generative models based on deep learning. Training a response generation model requires large datasets, which often consist of dialogue data from multiple speakers. This raises the issue of uncertain response styles. This is a problem that cannot be overlooked in chat dialogues, as it leads to a decrease in the "Human-like consistency" of the response. Therefore, in this research, we build a style-controllable response generation model based on GPT-2, which generates responses by inputting utterances and user IDs. This model is designed to determine the response topic from the utterance and the response style from the user ID. In addition, since subjective evaluation based on a small number of evaluators lacks reliability, in this study, objective evaluation in performed using a text classifier which estimates respondent.