A BERT-CNN Based Approach on Movie Review Sentiment Analysis

Abstract

Sentiment analysis plays a vital role in the decision-making of multiple fields. Specifically, in movies and television, audiences’ reviews can help with casting, the direction of the plot, etc. To further improve the performance of the original BERT model, a BERT-CNN-based approach is proposed in this paper to do sentiment analysis on IMDb dataset. Although their performances are nearly the same throughout the research, the BERT-CNN approach is better at negative sentiment detection. It got an average elevation of 3.6% in accuracy after the ensemble. Apart from that, topic modeling is also performed to show that most negative reviews are commented on from multiple aspects instead of criticizing only, making sentiment analysis of movie reviews a complex problem.