Abstract
Existing persona-based dialogue generation models focus on the semantic consistency between personas and responses. However, various influential factors can cause persona inconsistency, such as the speaking style in the context. Existing models perform inflexibly in speaking styles on various-persona-distribution datasets, resulting in persona style inconsistency. In this work, we propose a dialogue generation model with persona selection classifier to solve the complex inconsistency problem. The model generates responses in two steps: original response generation and rewriting responses. For training, we employ two auxiliary tasks: (1) a persona selection task to fuse the adapted persona into the original responses; (2) consistency inference to remove inconsistent persona information in the final responses. In our model, the adapted personas are predicted by an NLI-based classifier. We evaluate our model on the persona dialogue dataset with different persona distributions, i.e., the persona-dense PersonaChat dataset and the persona-spare PersonalDialog dataset. The experimental results show that our model outperforms strong models in response quality, persona consistency, and persona distribution consistency.