Self-supervised Sentiment Classification based on Semantic Similarity Measures and Contextual Embedding using metaheuristic optimizer

In recent years, considerable research attention has been paid to supervised machine learning methods for Sentiment Analysis (SA). The performance of these methods heavily depends on hyperparameter tuning and varies across different contexts. In addition, getting a massive amount of labeled training data is time- and labor-consuming. As a result, unsupervised machine learning methods are getting more attention. This paper proposes a Selfsupervised sentiment analysis method that semantically generates pseudo-labels (positive or negative) for each sample using the text similarity measures. Moreover, a sentiment classifier composed of the RoBERTa transformer and a Gated Recurrent Unit (GRU) is trained by these labeled data. What is more, Whale optimization Algorithm (WOA) is employed to find the optimal values of hyperparameters. The evaluation results demonstrate that the proposed method outperforms other methods in terms of accuracy, precision, and recall.