An Alternative Deep Model for Turn-Based Conversations

Abstract

In this paper, we propose a new model for building a conversational dialogue system which provides natural, realistic and flexible interaction between human and machine based on large movie subtitles dataset. Our models are a generative model that is autonomously generated word-by-word, opening up the possibility of working on many different languages. To address this goal, we extend the hierarchical recurrent encoder-decoder neural network (HRED) for dealing with extracting features from long input turns and generating long output turns. Furthermore, we also apply an attention mechanism in order to attend to particular sentences on source side when predicting a turn response. The models are trained end-to-end without labeling data. The experiments show that our proposed model has improved 36% on BLUE score compared to the HRED model. It also shows many potential improvements in chatbot models.