Improved Response Generation Consistency in Multiturn Dialog

Abstract

The consistency of response generation in multi-turn dialog must be improved. The existing VHDER model, that extends the RNN model to dialog generation, does not distinguish turns. Thus, in this study, we improve it by adding implementing a User-RNN that corresponds to a speaker who conveys information every odd and even turn. In addition, the global vibrational transformer model, which applies the CVAE, an advanced autoencoder, enables diverse response generation. However, latent variables sampled from the entire context are diluted by speaker characteristics, which degrades the consistency of generated responses. In this study, we employed speaker-specific latent variables to generate speaker-consistent responses.