Emotion Classification in Bengali-English Code-Mixed Data using Transformers

Emotion classification from social media text has attracted significant interest in recent years, largely because of the vast amount of unstructured and often disorganized text produced by users. While there has been significant progress in emotion classification for high-resource languages (HRLs), the challenge remains in resource-constrained languages (RCLs) like Bengali. The lack of standardized datasets and natural language processing tools further complicates this task. This challenge is intensified in the context of Bengali-English code-mixed text, which is prevalent on social media, making emotion classification even more complex. Addressing this issue, we propose a transformer-based architecture for emotion recognition in Bengali-English code-mixed text. We categorize emotions into six classes: joy, sadness, anger, disgust, fear, and surprise. To facilitate this task, a novel dataset called BECoD, comprising 5,753 code-mixed texts, was developed. We investigated the performance of several machine learning (ML), deep learning (DL), and transformer-based approaches for this classification task. The experimental results show that the Bangla-BERT-1 model outperformed all other models, achieving the highest weighted f<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf>-score (WF) of 81.75%.