Fine-Grained Emotional Paraphrasing along Emotion Gradients

Abstract: Paraphrase generation, a.k.a. paraphrasing, is a common and important task in natural language processing. Emotional paraphrasing, which changes the emotion embodied in a piece of text while preserving its meaning, has many potential applications, e.g., moderating online dialogues and preventing cyberbullying. We introduce a new task of fine-grained emotional paraphrasing along emotion gradients, that is, altering the emotional intensities of the paraphrases in fine grain following smooth variations in affective dimensions while preserving the meanings of the originals. We propose a framework for addressing this task by fine-tuning text-to-text Transformers through multi-task training. We enhance several widely used paraphrasing corpus by annotating the input and target texts with their fine-grained emotion labels. With these labels, fine-tuning text-to-text Transformers on these corpus entails multi-task training. Evaluations of the fine-tuned Transformers on separate test sets show that including fine-grained emotion labels in the paraphrase task significantly improve the chance of obtaining high-quality paraphrases of the desired emotions, i.e., more than doubling the number of exact matches of desired emotions while achieving consistently better scores in paraphrase metrics such as BLEU, ROGUE, and METEOR.