An Investigation of Controllable Neural Conversation Model with Dialogue Acts

Abstract

Dialogue act is known as an essential component of the dialogue system, which captures the user's intention and produces the appropriate response. In this paper, we propose a controllable response generation model given dialogue acts. Recent neural conversation models are based on the end-to-end approach that learns a mapping a mapping between dialogue histories and response utterances. However, it was difficult to control the contents of the response generated by the model. Several models tackled the problem of generating responses under the specified dialogue acts as a condition; however, these models still have problems on conditioned generations. In this paper, we introduced an extended framework of the generative adversarial network that optimizes both conditioned generator and discriminator which explicitly classifies dialogue act classes. Experimental results showed that our conditional response generation model improved both the response quality and controllability of neural conversation generation.