Multi-label emotion and sentiment detection from Bengali text using hybrid deep learning model

Emotion and sentiment play a significant role in many applications, such as e-commerce, public mental health, and online media security. Several works have been conducted on emotion and sentiment detection, both individually and in combination, using English texts. However, very few works are found in low-resource languages such as Bengali (endonym Bangla). Furthermore, most of these works still need improvements in accuracy, robustness, and data availability. In this research, we worked on multi-label emotion and sentiment detection using Bengali text. First, we gathered a large annotated dataset comprising five sentiment and seven emotion classes, which helped alleviate data scarcity in this area. Then, we employed a hybrid deep learning (DL) model that combines long short-term memory (LSTM) and convolutional neural network (CNN). Several experiments were conducted through two feature engineering methods (traditional and customized BERT) and seven optimizers (SGD, Adam, RMSprop, Adagrad, Adadelta, Adamax, and Nadam). The best combination turned out to be the customized BERT with the Adam optimizer. The proposed model achieved 99.35% and 99.85% accuracy for sentiment and emotion classifications, respectively. Furthermore, the proposed model achieved very high precision, recall, F-score, kappa, Matthews correlation coefficient (MCC), and area under the curve (AUC) values for both the sentiment and emotion detections. • We conducted multi-label classification of emotions and sentiments in Bengali texts. • We recruited volunteers to annotate 40,911 Bengali posts from social media platforms. • The annotated dataset is made publicly available. • We utilized a hybrid LSTM+CNN classifier with customized BERT feature engineering. • The proposed model achieved ∼ 99% accuracy, precision, recall, and F1-score.