Three Topics Talk Generation: A Method for Generating Responses in Dialogue Systems with Specified Knowledge

Abstract

In the recent years, generation-based dialogue systems using state-of-the-art (SoTA) transformer-based models have demonstrated impressive performance in simulating human-like conversations. Many generation-based dialogue systems use the sequential generation method, which generates response words sequentially from left to right according to the output distribution of model, based on decoding strategies such as Greedy. However, it is difficult to control the content of the responses generated by the sequential generation method, although the parameters such as minimum and maximum length can be controlled. To address this, inspired by the Three Topics Talk, which is an impromptu storytelling using three given topics, we propose a new responses generation method which generates responses preceding and following the specified knowledge (topic). The dialogue system using our proposed method has been validated to generate significantly more diverse and correct responses than baseline approaches.