Advancing Bengali Sentiment Analysis: A Benchmark Study with ELMo and XLNet Architecture

Even though Bengali is the sixth most spoken language in the world, many of its natural language processing domains remain underexplored, with sentiment analysis being one of the core areas.This study applies ELMo (Embeddings from Language Models) and XLNet (Transformer-XL) for the first time in Bengali sentiment classification.A dataset of over 40,000 user comments was used, which was collected from YouTube online platforms and used to train both models on binary classification with positive and negative labels and ternary classification with positive, neutral, and negative labels.Experimental results show that the ELMo two-class model achieved the highest accuracy of 71%, while the XLNet twoclass model reached 56%.These findings highlight the potential of context-rich representations like ELMo and XLNet for Bengali sentiment analysis, while also revealing the challenges of more nuanced ternary classification.Overall, the research provides new insights into leveraging cutting-edge deep learning models for limited-resource languages such as Bengali.