Semi-supervised approach for tweet-level stress detection

Psychological stress detection through social media has caught the interest of researchers due to the emergence of social media platforms like Twitter as major data sources to solicit human emotions and interactions. The problem of tweet-level stress detection is solved using various supervised learning mechanisms described in the existing literature. Given the real-world scenario in which there is a scarcity of labeled data and an abundance of unlabeled data, the semi-supervised solutions form an ideal choice. In this work, a semi-supervised approach called “Self-training Method for Tweet-level Stress Detection (SMTSD)” is proposed to extend the self-training approach for utilizing the information from sarcasm in predicting the pseudo-labels. In SMTSD, logistic regression is used as the base classifier. The predicted pseudo-labels are then used to train and build supervised baseline models on all the datasets collected using Twitter’s Tweepy API. The proposed solution of SMTSD, with the usage of sarcasm in predicting the pseudo-label, surpasses the performance of the basic self-training approach. Moreover, the proposed model also outperforms the supervised models considered and also gives better performance than the state-of-the-art techniques like Bi-LSTM. It is demonstrated that as the sample size for combined data of tweets with predicted pseudo-labels increases, so does the performance of the supervised models trained on them.