Persona Consistent Dialogue Generation via Contrastive Learning
Zhenfeng Han, Sai Zhang, Xiaowang Zhang · 2023
The inclusion of explicit personas in generation models has gained significant attention as a means of developing intelligent dialogue agents. However, large pretrained generation models often produce inconsistent responses with persona. We investigate the model generation behavior to identify signs of inconsistency and observe inconsistent behavior patterns. In this work, we introduce contrastive learning into persona consistent dialogue generation, building on the idea that humans learn not just from positive feedback, but also from identifying and correcting undesirable behaviors. According to the inconsistent patterns, we design two strategies to construct high-quality negative samples, which are critical for contrastive learning efficacy. Experimental results demonstrate that our method can effectively improve the consistency of the responses while improving its dialogue quality on both automatic and human evaluation.