Open Domain Response Generation Guided by Retrieved Conversations

Open domain response generation is the task of creating a response givena user query in any topics/domain. Limited by context and referenceinformation, responses generated by current systems are often "bland"or generic. In this paper, we combine a response generation model witha retrieval system that searches for relevant utterances and responses,and extracts keywords from the retrieved results to guide the responsegeneration. Our model uses a keyword extraction module to extract twotypes of keywords in an unsupervised fashion: (1) keywords in thequery not found in the retrieved utterances (DIFFKEY),and (2) overlapping keywords among the retrieved responses(SIMKEY). Given these keywords, we use a two-stage transformer thatfirst decides where to insert the keywords in the response, and thengenerates the full response given the location of the keywords. Thekeyword extraction module and the two-stage transformer are connected ina single network, and so our system is trained end-to-end.Experimental results on Cornell Movie-Dialog corpus, Douban and Weibodemonstrate that our model outperforms state-of-the-art systems in termsof ROUGE, relevance scores and human evaluation. Source code of ourmodel is available at: ANONYMISED.