[4Yin2] Interactive session 2

[4Yin2-36] A Comparative Analysis of Methods on Extracting Speaker Information in Dialogue Models Conditioned on Speaker Information

〇Hiroki Yasukawa¹, Seitaro Shinagawa¹, Masahiro Mizukami², Hiroaki Sugiyama², Katsuhito Sudoh¹, Satoshi Nakamura¹ (1.Nara Institute of Science and Technology, 2.NTT Communication Science Laboratories)

Keywords:dialogue model, persona information

In recent dialogue research, many models have been proposed in which sentences describing the speaker's information are input into a encoder with dialogue sentences to generate a response sentence that reflects the speaker's information.
However, such models have a problem that they generate responses that are strongly conditioned on the words appearing in the persona description.
In this study, we aimed to alleviate this conditioning by inputting distributed representations obtained by encoding Persona Descriptions with VAE.
As a result, we were able to generate a response that is not contradict the persona description while relaxing the conditioning.