Single and Multibranch CNN-Bidirectional LSTM for IMDb Sentiment Analysis

Abstract

Online users now frequently use the internet to voice opinions, ask for advice, or choose products and services based on the feedback of others. This provides a window into the way users feel about specific topics. The study of natural language processing, a sub-category of sentiment analysis, takes on this task by extracting meaning out of user text through observing the way in which words are grouped and used. Machine learning techniques have made significant advances which allow us to further explore mechanisms for interpreting such data. This research aims to use the internet movie database (IMDb) dataset and the Keras API to compare single and multibranch CNN-Bidirectional LSTMs of various kernel sizes (Maas et al. Learning word vectors for sentiment analysis. In: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, vol. 1, pp. 142–150, 2011; Chollet et al. Keras. 2015. https://keras.io). The results show that while only time to train varies between single and multibranch models, their maximum accuracies are close in range. The highest accuracy model was the single branch with kernel size 9 with an accuracy of 89.54%. While slightly more accurate than the multibranch model with 88.94%, the time savings for the single branch is approximately of 64% (2 h and 20 min).