eMLM: A New Pre-training Objective for Emotion Related Tasks

Abstract

Bidirectional Encoder Representations from Transformers (BERT) have been shown to be extremely effective on a wide variety of natural language processing tasks, including sentiment analysis and emotion detection. However, the proposed pre-training objectives of BERT do not induce any sentiment or emotion-specific biases into the model. In this paper, we present Emotion Masked Language Modeling, a variation of Masked Language Modeling, aimed at improving the BERT language representation model for emotion detection and sentiment analysis tasks. Using the same pre-training corpora as the original BERT model, Wikipedia and BookCorpus, our BERT variation manages to improve the downstream performance on 4 tasks for emotion detection and sentiment analysis by an average of 1:2% F1. Moreover, our approach shows an increased performance in our task-specific robustness tests. We make our code and pre-trained model available at https://github.com/tsosea2/eMLM.