BERA: Feature-Enriched Multi-Task Transformer for Sentiment and Emotion Analysis of Bangladeshi E-Commerce Review

Abstract

The rapid expansion of e-commerce platforms in Bangladesh has generated a massive volume of user-generated product reviews written in a highly multilingual and code-mixed style. These reviews frequently combine Bangla, English, and Romanized Bangla (Banglish), along with emojis, informal expressions, and stylistic variations. Traditional sentiment analysis techniques, which primarily focus on polarity classification, often fail to capture the rich emotional depth and linguistic complexities present in such real-world customer feedback. This study proposes BERA (Bangla E-commerce Review Architecture), a feature-enriched, multi-task transformer-based framework built upon BanglaBERT. The model integrates lexical, semantic, stylistic, and code-mixing features with a review-level attention mechanism to jointly perform sentiment classification (Positive, Negative, Neutral/Mixed) and fine-grained emotion detection (Happiness, Sadness, Anger, Fear, Love). Through systematic hyperparameter tuning, enhanced feature engineering, and dataset expansion, the proposed system achieves approximately 93% accuracy in sentiment classification and 68% accuracy in emotion classification. Furthermore, the trained model has been successfully deployed as a production-ready Inference API and wrapped with an intuitive web application, making it practically usable for e-commerce platforms and researchers. This work contributes to the advancement of Bangla natural language processing and offers a scalable solution for multilingual sentiment and emotion analysis..