Improving Personality Consistency in Conversation with Commonsense Knowledge

Maintaining persona consistency is crucial for an open-domain dialogue system. Existing methods focus on extracting user persona from historical context to guide persona-consistent response generation. Actually, the historical conversations are usually noisy and uninformative, and thus the target response generation is challenging, as well as easily falling into sub-optimization, especially for the well-known semantic gap between the potential persona information in the history and the target response, which can be bridged by dynamically injecting commonsense knowledge for inference. Hence, in this paper, we propose a persona-consistent dialogue model, consisting of a unified word-concept representing framework and a commonsense reasoning component, to reduce the semantic distance between persona and response. Extensive experiments demonstrate that our proposed model achieves significant enhancements in persona coherence.