Using Multi-Encoder Fusion Strategies to Improve Personalized Response Selection

Anonymous

Abstract: Personalized response selection systems are generally grounded on persona. However, there exists a co-relation between persona and empathy which is not explored well in these systems. Also, faithfulness to the conversation context 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 captures the interaction between persona, emotion, and entailment information of the utterances. 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 2.3% on original personas and 1.9% on revised personas in terms of hits@1 (top-1 accuracy), achieving a new state-of-the-art performance on the Persona-Chat dataset.