Detecting depression and suicidal rhetoric in social media texts using deep learning and big language models

Social media platforms are widely used communication environments in which individuals express their emotions and thoughts through textual content, and these data provide valuable insights for identifying mental health–related risks. The global increase in suicide rates, particularly among young populations, has brought research on the use of social media data in early warning systems to the forefront. The aim of this thesis is to examine the effectiveness of Deep Learning and Large Language Model (LLM)–based approaches in detecting depression and suicidal rhetoric from social media texts. In this study, traditional feature extraction methods (TF-IDF), static word representations (GloVe), and contextual embedding–based LLM approaches (BERT, RoBERTa) are comparatively analyzed. Using datasets collected from the Reddit platform, classical machine learning algorithms such as Logistic Regression, Support Vector Machines, and LightGBM, as well as deep learning architectures including Convolutional Neural Networks, Long Short-Term Memory, and Bidirectional Long Short-Term Memory models, are evaluated. To enhance the interpretability of model outputs, SHAP and LIME methods are employed. The experimental results demonstrate that the text representation method plays a decisive role in classification performance. LLM-based models with contextual learning capabilities outperform static approaches and achieve recall values of up to 95%, particularly in suicide risk detection. The findings indicate that the proposed approaches offer a practical and effective solution for early warning and decision support systems in the field of mental health.