A Machine Learning Approach for Textual Sentiment Classification

Abstract

The proliferate use of microblogs and social networks have become valuable sources to determine individuals’ opinion about an entity, product, topic, events, and politics etc. Main challenges implicated by the literature review include; contextual understanding and domain adaptation. Another challenge is that models trained on one domain may not generalize well to other domains due to differences sentiment expressions, or domain-specific terms. Therefore, textual sentiment analysis has become hotspot for research purpose. The paper proposes deep learning neural network models: LSTM and Bidirectional LSTM aims to automatically predict the sentiment polarity from given user posted text review into positive or negative class. Further, feature vectors are formed for each input sentence using Global vector for word representation (Glove) algorithm. Our proposed model utilizes 300-dimensional Glove for feature embedding. This high dimensional pre train vector contains semantically closer words. The impact of varying nature of datasets on the performance of both models for sentiment analysis is also investigated. Experiment is conducted on two Amazon product reviews datasets. The proposed research concluded that BLSTM achieved higher accuracy than Single LSTM and also outperforms the state -of-the-art models on document level reviews. Both models excel in capturing the temporal dependencies and linguistic structures. Highest accuracies of 94% and 96% are achieved on Amazon food reviews and Mobile reviews respectively using the BLSTM based model. The proposed model proved robust to changing sentiment trends or evolving language use as the datasets used are real-time datasets of thousands of users from different countries having variation in sentiment expression. The proposed model grasps the contextual understanding and domain adaptation very well.