BERT+BiLSTM: Boosting Performance in Movie Review Sentiment Classification

The demand for movie reviews sentiment analysis is growing rapidly nowadays. This study focuses on the development of advanced text classifiers to address complex classification tasks and proposes three models. The first model utilizes 6 encoder layers to capture information from texts and the second analyzes data using pretrained parameters of bidirectional encoder representations from transformers (BERT) with 12 encoder layers. The third one combines contextual embeddings of BERT with bi-directional long short-term memorys (BiLSTM) sequential modeling capabilities. Specifically, it leverages BERT to extract deep contextual features, which are then fed into a BiLSTM layer followed by a classification head for prediction. Experiments are conducted on a Kaggle Internet Movie Database (IMDB) and the analysis displays a trade-off between classification accuracy and total training time as well as memory consumption. The BERT+BiLSTM model ends up with an accuracy of around 65% but the longest training time per epoch and high memory usage. These results show the models ability to effectively capture contextual and sequential patterns in text, making it highly suitable for real-world textual sentiment analysis applications.