Discriminative-Transformer for Emotion Classification

Abstract

The development of social media encourages people to freely express their attitudes on many political issues, movies, products, etc., on the Internet. Emotion classification helps to understand the emotions contained in people’s natural language descriptions, and mining the emotions in people’s language helps organizations understand people’s responses to their actions. In recent years, more and more researches on emotion classification have been published. Although many methods in these works have achieved good results, most of these methods ignore the discrimination in the classification process. Therefore, we propose a Transformer model that imposes constraints on the margin size between categories to improve the discriminativeness of the model in emotion classification tasks. The experimental results show that the model in this paper can learn the latent text representation and make the model discriminative. We prove that improving the model’s discriminativeness helps improve the model’s performance in emotion classification tasks.