Recent Trends and Advances in Deep Learning-Based Sentiment Analysis

Abstract

Sentiment analysis is a fundamental branch of natural language processing. It is an essential task of identifying and extracting sentiment in opinionated data from sources such as social media, product feedback or blogs. Deep learning-based approaches have exceeded human-level performance in areas such as computer vision and speech recognition. Deep learning is widely accepted as the most promising in machine learning. In this chapter, we survey and analyse the current trends and advances in deep learning-based sentiment analysis approaches for document-level, sentence-level and aspect-based sentiment analysis for short and long text. A detailed discussion of deep learning architectures for sentiment analysis is provided. The studied approaches are classified into coarse-grain (including document and sentence level), fine-grain (includes target and aspect level) and cross-domain. Lastly, we provide a summary and in-depth analysis of the surveyed studies, for each of the aforementioned categories. The overwhelming number of studies explored convolutional neural networks (CNNs), long short-term memory (LSTM), gated recurrent unit (GRU) and attention mechanism. For coarse-grain sentiment analysis, LSTM and CNN-based models compete on performance, but it is CNNs that offer reduced model complexity and training overhead. Fine-grain sentiment analysis requires a model to learn complex interactions between target/aspect words and opinion words. Bi-directional LSTM and attention mechanisms offer the most promise, although CNN-based models have been adept at aspect extraction. The efforts in cross-domain sentiment analysis are dominated by LSTM and attention models. Our survey of cross-domain approaches revealed the use of multitask learning, adversarial training and joint training for domain adaptation.