Publication

Abstract :

Social media is rapidly reflecting mental health issues, hence offering useful yet unstructured data for early detection and intervention. Based on social media material, this paper offers an AI-powered framework for multi-class sentiment analysis to address mental health issues. From several mental healthrelated sources on Kaggle, a thorough dataset of 53,043 tagged entries spanning seven emotion categories-Normal, Depression, Suicidal, Anxiety, Stress, Bi-Polar, and Personality Disorder-was assembled. Extensive preprocessing designed for loud, informal social media language was applied to the data. Using different embedding approaches including TF-IDF, GloVe, FastText, and BERT, we assessed a broad spectrum of Transformer-based models, Deep Learning (DL), and Machine Learning (ML). Multi-Class Multi-Level approach is also implemented using ML and DL models. While DL models employing GRU with FastText reached 73.63% accuracy and an F1-score of 0.66, classic ML models with TF-IDF showed up to 75% accuracy. Of all methods, the BERT-based Transformer model beat others with a test accuracy of 79.37% and an F1-score of 0.76. These results highlight the possibility of using Transformer structures for complex sentiment categorization in the mental health field, hence providing smart and scalable early-warning systems for emotional well-being.