This study successfully developed deep learning models, including GPT-2, BERT, LSTM, and CNN, to detect stress states in social media texts, particularly focusing on Twitter. The models demonstrated effectiveness in identifying stress signals, with SHAP analysis revealing both similarities and differences in word importance among the models. BERT and GPT-2 showed consistency in recognizing stress-related and non-stress-related language patterns, while CNN presented distinct terms. Further analysis using BERT revealed that non-stress users had higher social media engagement and more positive sentiment in their posts compared to stress users. These findings highlight the potential of utilizing social media metrics and sentiment analysis to understand and identify user stress levels.
🗝 Keywords - Text Classification