Retrieval-Polished Response Generation for Chatbot

Liang Zhang, Yan Yang, Jie Zhou, Chengcai Chen, Liang He

Abstract: Chatbot communication, in which a robot communicates with a human being in natural language in an open domain, has achieved significant progress. However, it still suffers from problems such as a lack of diversity and contextual relevance. In this paper, we propose a retrieval-polished (RP) model for response generation that polishes a draft response based on a retrieved prototype. In particular, we first adopt a prototype selector to retrieve a contextually similar prototype. Then, a generation-based polisher is designed to obtain a polished response. Finally, we introduce a polished response filter to choose whether the final reply should be the retrieved response or the polished response. Extensive experiments on a dialog corpus show that our method outperforms retrieval-based and generation-based chatbots with respect to fluency, contextual relevance, and response diversity. Specifically, our model achieves substantial improvement compared with several strong baselines.