A COMPARATIVE STUDY OF CNN-BILSTM AND TF-IDF–NAIVE BAYES FOR SENTIMENT CLASSIFICATION ON MOVIE REVIEWS: PERFORMANCE AND EFFICIENCY TRADE-OFFS

Abstract

Sentiment analysis on movie reviews has become an important task for understanding public opinion, yet many high-performing deep learning models proposed in the literature rely on increasingly complex architectures such as multichannel embeddings, kernel-based projections, or attention-augmented transformers. This added complexity often comes without a clear report of computational cost, making it difficult to judge whether the performance gain justifies the resource trade-off. This study investigates whether a deliberately simplified Convolutional Neural Network combined with a Bidirectional Long Short-Term Memory (CNN-BiLSTM) architecture can achieve competitive performance compared to a classical baseline, TF-IDF combined with Naive Bayes, while explicitly reporting training efficiency. Both models were trained and evaluated on the IMDB Dataset of 50,000 movie reviews, split into 80% training, 10% validation, and 10% testing sets. Experimental results show that the proposed CNN-BiLSTM achieved an accuracy of 86.58%, precision of 90.38%, recall of 81.88%, F1-score of 85.92%, and AUC of 0.946, while the TF-IDF-Naive Bayes baseline achieved an accuracy of 85.98%, precision of 86.58%, recall of 85.16%, F1-score of 85.86%, and AUC of 0.932. The proposed method required approximately 203 seconds of training time, compared to 0.07 seconds for the baseline. These findings indicate that the performance improvement offered by the deep learning approach is modest relative to its substantially higher computational cost, providing practical guidance for method selection in resource-constrained settings.