What If Bots Feel Moods? Towards Controllable Retrieval-based Dialogue Systems with Emotion-Aware Transition Networks

Lisong Qiu,Ying Wai Shiu,Pingping Lin,Ruihua Song,Yue Liu,Dongyan Zhao,Rui Yan

DOI: https://doi.org/10.1145/3397271.3401108

2020-01-01

Abstract:For social bots, smooth emotional transitions are essential for delivering a genuine conversation experience to users. Yet, the task is challenging because emotion is too implicit and complicated to understand. Among previous studies in the domain of retrieval-based conversational model, they only consider the factors of semantic and functional dependencies of utterances. In this paper, to implement a more empathetic retrieval-based conversation system, we incorporate emotional factors into context-response matching from two aspects: 1) On top of semantic matching, we propose an emotion-aware transition network to model the dynamic emotional flow and enhance context-response matching in retrieval-based dialogue systems with learnt intrinsic emotion features through a multi-task learning framework; 2) We design several flexible controlling mechanisms to customize social bots in terms of emotion. Extensive experiments on two benchmark datasets indicate that the proposed model can effectively track the flow of emotions throughout a human-machine conversation and significantly improve response selection in dialogues over the state-of-the-art baselines. We also empirically validate the emotion-control effects of our proposed model on three different emotional aspects. Finally, we apply such functionalities to a real IoT application.