Pick the Better and Leave the Rest: Leveraging Multiple Retrieved Results to Guide Response Generation

Neural Response Generation (NRG) has achieved remarkable progress recently, though they suffer from the safe response problem. Some researchers proposed leveraging the retrieval-based chatbots' results to enhance the NRG models to generate diverse and informative responses. However, picking helpful information inside the multiple retrieved references and avoiding errors and noise brought from retrieval systems is still challenging. This paper proposes a variational neural response generating framework in which the validity of retrieved results is measured and those useful ones are taken as guidance explicitly. Moreover, if all the retrieved results fail to provide sufficient information, the framework also can let the model regress to a regular query-based NRG automatically. According to the thorough experimental comparisons with other retrieval-guided models, our proposed model can better utilize the useful information of retrieved results to generate appropriate and diverse responses.