Towards Faithful Personalized Response Selection in Retrieval Based Dialog Systems

Anonymous

Abstract: Personalized response selection systems are generally grounded on persona. However, the angle of emotion influencing response selection is not explored. Also, faithfulness to the conversation context of these systems plunges when a contradictory or an off-topic response is selected. This paper makes an attempt to address these issues by proposing a suite of fusion strategies that capture the interaction between persona, emotion, and entailment information of the utterances. A concept-flow encoder is designed which capture the relevant concept knowledge both in context and responses. Ablation studies were done on Persona-Chat dataset show that incorporating emotion, entailment improves the accuracy of response selection. We combine our fusion strategies and concept-flow encoding to train a BERT based model which outperforms the previous methods by margins larger than 1.9% on original personas and 1.7% on revised personas in terms of hits@1 (top-1 accuracy), achieving a new state-of-the-art performance on the Persona-Chat dataset.

Paper Type: long

Consent To Share Data: no