Evaluation Model for Persona Characteristics via Contrastive Learning

Abstract

Persona-aware dialogue systems can improve the consistency of system responses, user trust, and enjoyment. Filtering nonpersona-like utterances is important for constructing these systems. We proposed a novel model to capture the intensity of persona characteristics in a given utterance. We trained the model using contrastive learning, based on the similarity of the utterances' speaker. Experimental results showed that our model outperforms existing methods and baselines using pre-trained language models.