SARCASM AND MENTAL HEALTH DETECTION FROM SOCIAL MEDIA USING TRANSFORMER-BASED TECHNIQUES

Abstract

Social media platforms generate massive amounts of textual content every day, containing valuable information related to human emotions, opinions, sarcasm, and psychological behavior. Understanding such contextual and emotionally implicit information has become an important research prob lem in Natural Language Processing and Artificial Intelligence. This the sis presents transformer-based approaches for sarcasm detection and mental health classification from social media text using contextual representation learning and Explainable Artificial Intelligence techniques. For sarcasm de tection, a transformer-based framework named SarcasmBERTa is proposed by integrating RoBERTa with Multi-Head Attention mechanisms to cap ture contextual and semantic relationships from multilingual social media text. The framework was evaluated on English, Hindi, and Bengali datasets and demonstrated superior performance compared to several baseline and contemporary models. For mental health detection, the proposed X-MHD framework combines DistilBERT, knowledge graph embeddings, and Ex plainable Artificial Intelligence techniques. The model effectively identifies emotional well-being conditions such as depression, anxiety, bipolar disorder, and suicide-related behaviour from online textual content.