Integrated Multiclass Sentiment Analysis of Movie Reviews: Combining Fine-Tuned DistilBERT, LightGBM and XAI for Interpretation
Maharin Afroj, Md. Polash Islam, Shirin Sultana, Md. Saifur Rahman · 2024
Sentiment analysis of movie reviews plays a critical role in understanding audience perspectives and predicting trends in the entertainment industry. This work presents an integrated approach that encourages a fine-tuned DistilBERT model for feature extraction, followed by LightGBM for classification, and SHAP (Shapley Additive Explanations) for model interpretability. By combining these advanced techniques, our approach achieves a high accuracy of 97%, significantly outperforming traditional methods. The use of DistilBERT enables precise contextual understanding of textual data while offering a more efficient and lightweight alternative to the full BERT model. LightGBM provides efficient and scalable classification, and SHAP ensures transparent and interpretable model decisions, allowing us to understand key factors driving sentiment predictions. This integrated framework enhances accuracy and also provides valuable insights into the model’s behavior, making it a robust tool for sentiment analysis in movie reviews.