Integrating Background and General Knowledge for Dialogue Generation

The traditional sequence-to-sequence model generates responses that are smooth but empty in their content. Background-based dialogue is one solution that uses the context’s unstructured knowledge to generate informative responses. The key point of background-based dialogue is knowledge extraction, but some conversations have poor performance in knowledge selection due to insufficient information. At the same time, to improve the satisfaction of the responses, this paper can enhance the amount of conversational knowledge while allowing the model to carry some emotional awareness by selecting external knowledge sources with emotional information. In this paper, we introduce the CEC model, which utilizes graph attention and a double-matching matrix for the selection of external and background knowledge. The generation process is conducted within each decoding step, considering the selected knowledge’s content. We conduct experiments on the Holl-E dataset. According to the experimental results, our model CEC outperforms the previous model in terms of performance.