Integrate the key persona information and context to improve the performance of multi round dialogue generation

Abstract

In multi-round dialogue systems, we can easily find that the final reply is closely related to two points, one is the context of the dialogue, the other is the persona characteristics. But not all characters and contextual information will affect the final reply, because the final reply may only be related to some crucial characters and contextual information,the indiscriminate use of all information may even have a negative impact on the generated dialogue. So it is necessary to extract and utilize key characters and contextual information to improve the quality of the final generated response. In this paper, we show how to solve this problem through our new model and methods. Specifically, our new model consists of two parts: encoder and decoder. The encoder is mainly used to encode personas, contexts and historical responses, and the decoder generates corresponding words from the vocabulary. Then, the weight of the character and context is updated through the multi-head self-attention mechanism to affect the response generated by the decoder. The experimental results show that compared with the baseline models, our model and methods have improved in terms of metric-based evaluation.