EmoFusion: An integrated machine learning model leveraging embeddings and lexicons to improve textual emotion classification

Abstract

Human emotions are complicated and intertwined with cognitive processes, influencing mental health, learning, and decision-making. The Web 2.0 era has seen a remarkable spike in the number of people sharing their experiences and emotions on online social media, mostly through posts or text messages. Due to inherent challenges associated with textual data, the issue of discovering the intricate relationships between texts and its inherent emotions is still an increasingly prevalent topic in AI and NLP. This paper presents EmoFusion, an integrated machine learning model that improves emotion classification in textual data by integrating pre-trained word embeddings and emotion lexicons. Instead of relying on a single emotion lexicon, EmoFusion integrates multiple emotion lexicons since a single lexicon might not fully cover all possible words or phrases linked with emotions. The proposed approach uses semantically related features to bridge the semantic gap between words and emotions, capturing a wide range of emotional nuances and resulting in better classification performance. The efficacy is further improved by employing emotion-specific pre-processing techniques. EmoFusion is evaluated using three benchmark datasets, namely Google AI GoEmotions, CBET, and TEC. The evaluation results demonstrate a significant improvement compared to six baselines and a state-of-the-art technique using different classifiers.