RNNCore: Lexicon Aided Recurrent Neural Networkfor Sentiment Analysis

Abstract

Sentiment Analysis (SA) or Opinion Mining can help in identifying subjective information conveyed by user reviews for various automation tasks such as building better recommendation systems, identifying user trends, monitoring and customer support. This paper focus on the sentiment score detection. Traditional SA algorithms suffer from low accuracies in identifying true user intents. However, with the advent of Deep Learning many NLP tasks including Sentiment Analysis have become feasible with accuracies comparable to that of human experts. Additional advantage of Deep Learning in contrast to supervised learning is that in deep learning a manually tuned features set is not required. Deep Learning algorithm such as Convolution Neural Networks (CNN), Long Short Term Memory (LSTM), Recurrent Neural Networks (RNN) and various other have successfully been applied to SA. RNN in particular is well suited for this task, however most the works done over RNNs require large supervised training sets which are usually not available for all domains. This work proposes a new method called RNNCore which can make use of the pre-trained word embedding from Stanford Core NLP in conjunction with RNN to improve on accuracy and reduce time complexity. Comparison between the results of RNNCore, RNN and OneR method on the IMDB review dataset suggests that RNNCore yield 92.60% F1-measure which is a marked improvement of 17.74% as compared with a simple RNN approach for Sentiment Analysis.