Toward Entrained Response Generation for Neural Conversation Model

Synchronization of words in conversation, called entrainment, is generally observed in human-human conversation. It is reported that entrainment has a high correlation with dialogue engagement. In this work, we consider using several entrainment scores based on the word, co-occurrence to evaluate the entrainment of system generation. We try to build a neural conversation model optimized for these entrainment scores by using reinforcement learning. Experimental results showed that the proposed neural conversation model optimized by the entrainment scores achieved a better entrainment score than general neural conversation models.