Mental Health Sentiment Analysis: Exploring an Optimized BERT with Deep Encodings

Abstract

Deep learning technologies have significantly advanced the field of Natural Language Processing (NLP) in the recent past. A promising line of research in the existing study of text sentiment is the analysis of medical texts, which can potentially find numerous uses in medical diagnosis. However, sentiment analysis in the medical field remains challenging due to the complexity of medical language and context-aware interpretation of domain-specific ontologies. This paper focuses on the medical field and uses deep encoding techniques such as Fast-Text, Word2Vec, along with BERT and RoBERTa at the output layer for sentiment detection. The analysis specifically targets seven key mental health classes: Anxiety, Bipolar, Depression, Normal, Personality Disorder, Stress, and Suicidal. Text data were preprocessed using text cleaning, stop word elimination, and lemmatization to enhance the quality of the input data and improve the effectiveness of models. Experiments were carried out on a mental health dataset to analyze performance after the integration of deep encoding with diverse deep learning models. Based on the results, transformer-based models outperformed various other networks, achieving more than 95% accuracy. This study provides a basis for the selection of appropriate models in achieving accurate sentiment analysis within the medical field and is useful for research on designing efficient model frameworks.