A Hybrid Transformer Architecture for Multi-Class Mental Illness Prediction Using Social Media Text

Project Code :TCPGPY1821

Objective

This project develops a hybrid transformer-based system to classify multiple mental illnesses. It combines models like BERT, RoBERTa, GPT-2, and DistilBERT with deep learning layers such as CNN and BiLSTM to improve accuracy. The system processes raw text, trains on labeled data, and predicts categories like Anxiety, Depression, and Stress. A user-friendly Flask web app allows secure input and displays predictions. The approach enhances detection by capturing context and sequential patterns, supporting early mental health identification.

Abstract

This project presents a hybrid transformer-based architecture for the multiclass classification of mental illnesses using social media text. Four hybrid models have been developed and evaluated: BERT + CNN, RoBERTa + BiLSTM, GPT2 + BiLSTM, and DistilBERT + CNN. These models are trained to identify and classify mental health conditions into these categories: Anxiety', 'Bipolar', 'Depression', 'Normal', 'Personality disorder', 'Stress', 'Suicidal'. The project uses preprocessed textual data from social media platforms and leverages contextual embeddings and sequential learning to improve classification accuracy. A web-based application is developed using Flask, with modules for user registration, login, prediction, and logout. The system allows users to input text and receive a predicted mental health status. The backend is supported by transformer models and deep learning architectures, ensuring efficient prediction and scalability. Evaluation metrics such as accuracy, precision, recall, and F1-score are used to compare models and determine the best-performing hybrid.

Keywords: Mental illness, Social media, Transformer, BERT, CNN, BiLSTM, RoBERTa, GPT2, DistilBERT, Flask

NOTE: Without the concern of our team, please don't submit to the college. This Abstract varies based on student requirements.