Topic-aware response selection for dialog systems

It is challenging for a persona-based chitchat system to return responses consistent with the dialog context and the persona of the agent. This particularly holds for a retrieval-based chitchat system that selects the most appropriate response from a set of candidates according to the dialog context and the persona of the agent. A persona usually has some dominant topics (e.g., sports, music). Adhering to these topics can enhance the consistency of responses. However, previous studies rarely explore the topical semantics of the agent’s persona in the chitchat system, which often fails to return responses coherent with the persona. In this paper, we propose a Topic-Aware Response Selection (TARS) model, capturing multi-grained matching between the dialog context and a response and also between the persona and a response at both the word and the topic levels, to select the appropriate topic-aware response from the pool of response candidates. Empirical results on the public persona-based empathetic conversation (PEC) data demonstrate the promising performance of the TARS model for response selection.