End-to-End Personalized Humorous Response Generation in Untrimmed Multi-Role Dialogue System

Multi-role dialogue is challenging in natural language processing (NLP), which needs not only to understand sentences but also to simulate interaction among roles. However, the existing methods assume that only two speakers are present in a conversation. In real life, this assumption is not always valid. More often, there are multiple speakers involved. To address this issue, we propose a multi-role interposition dialogue system (MIDS) that generates reasonable responses based on the dialogue context and next speaker prediction. The MIDS employs multiply role-defined encoders to understand each speaker and an independent sequence model to predict the next speaker. The independent sequence model also works as a controller to integrate encoders with weights. Then, an attention-enhanced decoder generates responses based on the dialogue context, speaker prediction, and integrated encoders. Moreover, with the help of unique speaker prediction, the MIDS is able to generate diverse responses and allow itself to join (interpose) the conversation when appropriate. Furthermore, a novel reward function and an updating policy of reinforcement learning (RL) are applied to the MIDS, which further enable MIDS the ability to write drama scripts. The experimental results demonstrate that the MIDS offers a significant improvement to the accuracy of speaker prediction and the reduction of response generation perplexity. It is also able to interact with users without cues during real-life online conversations and avoid meaningless conversation loops while generating scripts. This paper marks the first step toward multi-role humorous dialogue generation.