Critical Analysis of BERT and LSTM Model for Bengali Sentiment Analysis Across Varied Datasets
Md Mehrab Hossain, Iffat Ara Helal Akhi, Syeda Rafiatus Sama Raisa, Taskin Sultana Onni, Ariful Islam Rifat, Ashraful Islam · 2023
Bengali sentiment analysis is a fast-developing field with a wide range of uses, including social media monitoring, market research, and user feedback analysis. In order to develop reliable models that can accurately represent the complex emotional expressions found in Bengali text, this study explores Bengali sentiment analysis. This research stands out for its thorough examination of multiple sentiment analysis models, including both conventional BiLSTM models and cutting-edge BERT-based models. Our analysis provides insightful information about the performance of the model on three different sentiment datasets: SentiNob, Social Media Comments, and Bengali Sentiment. The best-performing models highlight differences between datasets and highlight the impact of dataset-specific features. Kowsher/model-bangla-bert is a highly adaptive performer in the SENTINOB dataset, demonstrating adaptability with a strong F1 score of 73% and Accuracy of 72%. In close pursuit, Csebuetnlp/banglabert exhibits dependability and equilibrium, achieving the highest F1 score and accuracy of 75% and 74%, respectively. With an amazing F1 score of 90%, Kowsher/model-bangla-bert emerges as the top performer in the SOCIAL MEDIA COMMENTS dataset, demonstrating remarkable precision and recall. Kowsher/model-bangla-bert is a standout performer in the BENGALI SENTIMENT dataset, exhibiting adaptability while sustaining a high F1 score of 67% and the highest Precision of 70%. We find that FastText + BiLSTM consistently achieves a competitive F1 score of 65%. By providing a thorough comparative analysis of these models across several datasets, this study not only closes a significant gap in the literature but also provides a roadmap for future research in Bengali sentiment analysis. We suggest future directions for domain adaptation, multi-class learning, and improvement. This research lays the groundwork for the creation of more responsible and effective AI models in the field of sentiment analysis in Bengali by carefully examining various models, datasets, and performance metrics.