The rising interest in emotion detection in language is driven by the abundance of emotional expressions on Web 2.0 platforms. This paper explores the challenges in developing an automatic emotion detection system for Bengali, given the limited resources and lack of standard corpora. We describe the development of a comprehensive emotional dataset containing 20,247 texts categorized into 27 different emotional categories. Our process involved data collection, preprocessing, human and automatic labeling, and label verification. The dataset’s evaluation, reflected by a Cohen’s score of 0.89, indicates a high level of annotator agreement. Experiments conducted with machine learning, deep learning, and BERT-based models identified XLM-R as the best-performing model, achieving an F1 score of 0.79 and an accuracy of 0.82 in emotion classification tasks.