SSentiA: A Self-supervised Sentiment Analyzer for classification from unlabeled data

In recent years, supervised machine learning (ML) methods have realized remarkable performance gains for sentiment classification utilizing labeled data. However, labeled data are usually expensive to obtain, thus, not always achievable. When annotated data are unavailable, the unsupervised tools are exercised, which still lag behind the performance of supervised ML methods by a large margin. Therefore, in this work, we focus on improving the performance of sentiment classification from unlabeled data. We present a self-supervised hybrid methodology SSentiA (Self-supervised Sentiment Analyzer) that couples an ML classifier with a lexicon-based method for sentiment classification from unlabeled data. We first introduce LRSentiA (Lexical Rule-based Sentiment Analyzer), a lexicon-based method to predict the semantic orientation of a review along with the confidence score of prediction. Utilizing the confidence scores of LRSentiA, we generate highly accurate pseudo-labels for SSentiA that incorporates a supervised ML algorithm to improve the performance of sentiment classification for less polarized and complex reviews. We compare the performances of LRSentiA and SSSentA with the existing unsupervised, lexicon-based and self-supervised methods in multiple datasets. The LRSentiA performs similarly to the existing lexicon-based methods in both binary and 3-class sentiment analysis. By combining LRSentiA with an ML classifier, the hybrid approach SSentiA attains 10%–30% improvements in macro F1 score for both binary and 3-class sentiment analysis. The results suggest that in domains where annotated data are unavailable, SSentiA can significantly improve the performance of sentiment classification. Moreover, we demonstrate that using 30%–60% annotated training data, SSentiA delivers similar performances of the fully labeled training dataset.