Emotion-aware and Intent-controlled Empathetic Response Generation using Hierarchical Transformer Network

Tulika Saha, Sophia Ananiadou

Abstract: Enriching any dialogue systems to exhibit empathy is fundamental for delivering human-like conversations. Empathetic interactions in the form of empathetic dialogue generation has been studied widely in recent times. Existing models either incorporate emotion as a feature at the encoding side or as a latent variable explicitly at the decoder to condition their response on. While understanding speaker emotion is integral to expressing empathy, another aspect of being empathetic necessitates responding with an appropriate emotion (also known as emotional regulating intents) to speaker's mental state. To integrate these multiple aspects, in this paper, we propose a Hierarchical Transformer Network (HTN), an amalgamation of the recently introduced Transformer model and Hierarchical Encoder Decoder (HRED) architecture to capture the speaker emotion and dialogic context. For generating intent controlled empathetic responses, we draw insights from Reinforcement Learning (RL) to optimize rewards implicitly. The proposed approach is demonstrated on two benchmark open-domain empathetic datasets. The empirical evaluation (both automated and manual) demonstrates the system capability by way of outperforming several baselines and the state of the art models.