Exploring NLP for Mental Health Insights: Multi-Class Classification of Online Forum Texts

With the increasing incidence of mental health issues, there is a real need for early detection, which is currently limited by stigma and ignorance. This study attempts to explore multi-class classification models to analyze mental health problems through social media text. The goal of the classification model is to categorize text to one of six categories of the mental health problems and thus to provide patterns of the language which might serve as an early indication of these problems. After data collection and labeling, the dataset was resampled to balance the dataset for model training. Some of the important steps for data preprocessing included tokenization, the removal of unnecessary characters and labels, and one-hot encoding. To further understand the language used in expressing the different conditions, word clouds and bigram analyses were conducted. The training utilized transformer models and machine learning approaches, including RoBERTabase, T5, and MentalBERT. The results revealed that T5-small and MentalBERT achieved the highest accuracy of $83 \%$, notably outperforming RoBERTa-base, which reached an accuracy of $\mathbf{8 1 \%}$.