A Novel Linguistic-Aware Memory Structure for Enhancing the Response Generation

People will think of conversation experiences during a chat. Inspired by the phenomena, some previous works exploit relevant conversation utterances to enhance the response generation process. However, these methods only focus on semantic information while ignoring linguistic information. As a result, relevant sentences are not fully utilized. To address the problem, we design a novel memory structure to store linguistic information. Specifically, we conduct a memory bank on the relevant utterances, consisting a series of key-value pairs. And the memory structure is combined during the generation process. We conduct experiments on two public conversation datasets, and the results demonstrate the effectiveness of our memory structure.