Comparative Study of Machine Learning and Deep Learning Approaches for Aspect-Based Sentiment Analysis in Bengali Texts with Emojis

The increasing availability of Bengali textual data has led to a growing demand for effective sentiment analysis techniques. Aspect-Based Sentiment Analysis (ABSA) plays a critical role in identifying sentiments tied to specific aspects or features of products, services, or topics. However, the inclusion of emoji in ABSA for Bengali text remains an underexplored area, as emojis are often excluded from datasets in prior research. The aim of this comparative study is to fill this gap by comparing the performance of machine learning, ensemble learning and deep learning models for ABSA on Bengali texts containing emoji. The study investigated how emojis impact the identification of aspect-specific sentiments. Our experiments on the newly curated Bengali dataset demonstrate the effectiveness of both traditional and deep learning methods in aspect-based sentiment analysis. Without emoji, the Random Forest achieved an accuracy of 84.8% in aspect classification. Whereas with emojis, the accuracy decreased to 84.5%. However, the Sentiment-level accuracy increased from 70.5% to 70.6% with the inclusion of emojis. The BERT model achieved an aspect classification accuracy of 88.5% without emojis but the accuracy decreased to 88.4% with emojis. On the other hand, in Sentiment classification without emojis our BERT model achieved 89.4% accuracy compared to an increased accuracy of 90.1% with the inclusion of emojis indicating a trade-off between overall sentiment accuracy and aspect classification accuracy.