A Deep Learning–Based Framework for Dataset Creation and Sentiment Classification of English–Bengali Code-Mixed Texts

Datasets form the foundation for most Natural Language Processing (NLP) operations, such as sentiment analysis, summarization, and translation. Lack of big and diverse datasets poses a critical problem for low-resource languages like Bengali. The problem is aggravated in code-mixed environments where users tend to intermix Bengali and English on web platforms like social media, video comments, and product reviews on e-commerce websites. It is very difficult to conduct sentiment analysis of such Bengali–English code-mixed reviews, as there are no high-quality, large-scale datasets available. To address this challenge, the present research introduces a novel synthetic English–Bengali code-mixed product review dataset, specifically curated for sentiment analysis in the Bangladeshi e-commerce domain. The dataset was generated through a new deep learning–driven pipeline integrating part-of-speech tagging, translation, transliteration, and a two-tier ensemble-based sentiment annotation framework to ensure linguistic diversity and contextual realism in code-mixed expressions. The dataset, En–Bn–Code–Mixed–Two–Class–Sentiment–Dataset–, is available on the Hugging Face repository, with a total of 100,000 reviews, balanced for both positive and negative sentiments. The quality and reliability of the generated dataset were validated through quantitative, linguistic, and statistical analyses, including translation, transliteration, and sentiment analysis evaluation metrics. The proposed two-tier ensemble sentiment annotation approach, implemented through deep learning models, meta-learning models, and majority voting, achieved 93.00% accuracy, 92.16% precision, 94.00% recall, and 93.07% F1-score. This work not only provides a useful resource for code-mixed Bengali–English NLP but also establishes a scalable methodology for low-resource and multilingual text processing, opening up further possibilities of more comprehensive and inclusive studies on sentiment analysis.