Deep CNN-LSTM with combined kernels from multiple branches for IMDb review sentiment analysis

Abstract

Deep learning neural networks have made significant progress in the area of image and video analysis. This success of neural networks can be directed towards improvements in textual sentiment classification. In this paper, we describe a novel approach to sentiment analysis through the use of combined kernel from multiple branches of convolutional neural network (CNN) with Long Short-term Memory (LSTM) layers. Our combination of CNN and LSTM schemes produces a model with the highest reported accuracy on the Internet Movie Database (IMDb) review sentiment dataset. Additionally, we present multiple architecture variations of our proposed model to illustrate our attempts to increase accuracy while minimizing overfitting. We experiment with numerous regularization techniques, network structures, and kernel sizes to create five high-performing models for comparison. These models are capable of predicting the sentiment polarity of reviews from the IMDb dataset with accuracy above 89%. Firstly, the accuracy of our best performing proposed model surpasses the previously published models and secondly it vastly improves upon the baseline CNN+LSTM model. The capability of the combined kernel from multiple branches of CNN based LSTM architecture could also be lucrative towards other datasets for sentiment analysis or simply text classification. Furthermore, the proposed model has the potential in machine learning in video and audio.