More the Merrier: Towards Multi-Emotion and Intensity Controllable Response Generation
Abstract
The focus on conversational systems has recently shifted to-wards creating engaging agents by embedding emotions intothem. Human emotions are highly complex as humans canexpress multiple emotions with varying intensity in a singleutterance, whereas the conversational agents convey only oneemotion in their responses. To infuse human-like behaviourin the agents, we introduce the task of multi-emotion control-lable response generation with the ability to express differentemotions with varying levels of intensity in an open-domaindialogue system. We introduce a Multiple Emotion Intensityaware Multi-party Dialogue (MEIMD) dataset having 34kconversations taken from 8 different TV Series. We proposea Multiple Emotion with Intensity-based Dialogue Genera-tion (MEI-DG) framework. The system employs two novelmechanisms: (i).Explicit Memory:to determine whether togenerate an emotion or generic word, while focusing on theintensity of the desired emotions; and (ii).Implicit Memory:to compute the number of wordsremainingto express theemotion completely, thereby regulating the generation ac-cordingly. The detailed evaluation shows that our proposedapproach attains superior performance compared to the base-line models