Emotion Style Transfer with a Specified Intensity Using Deep Reinforcement Learning

Anonymous

Abstract: Text style transfer is a widely explored task in natural language generation which aims to change the stylistic properties of the text while retaining its style-independent content. In this work, we propose the task of emotion style transfer with a specified intensity in an unsupervised setting. The aim is to rewrite a given sentence, in any emotion, to a target emotion while also controlling the intensity of the target emotion. Emotions are gradient in nature, some words/phrases represent higher emotional intensity, while others represent lower intensity. In this task, we want to control this gradient nature of the emotion in the output. Additionally, we explore the issues with the existing datasets and address them. A novel BART-based model is proposed that is trained for the task by direct rewards. Unlike existing work, we bootstrap the BART model by training it to generate paraphrases so that it can explore lexical and syntactic diversity required for the output. Extensive automatic and human evaluations show the efficacy of our model in solving the problem.