Predicting mental health using social media sentiment analysis

Abstract

This project aims to develop sentiment analysis models, capable of identifying Gen Z individuals experiencing mental health conditions. In the context of mental health, this project adopts a more comprehensive approach by performing fine-grained classification of the type of mental health condition, if any, expressed in social media texts. The project considers Reddit and Twitter as key platforms for performing sentiment analysis. The linguistic features of these platforms are analyzed to evaluate their potential for being leveraged for the task. The models used for this task include Support Vector Machine, CNN-BiLSTM and BERT variants. From the results, a BERT variant, MentalRoBERTa achieved the best accuracy performance in both the Reddit and Twitter test datasets. Additionally, incorporating lexicon-based approaches with machine learning can enhance the effectiveness of sentiment analysis. The paper highlights two key lexicon approaches - VADER and NRC Emotion. The VADER lexicon is useful for understanding the severities of mental health conditions, while the NRC Emotion lexicon extracts emotion features from text. When combined with model ensembles, the NRC Emotion lexicon further enhances the performance of sentiment analysis. To facilitate understanding of the predictions made by the models, the project also explores Explainable Artificial Intelligence techniques and deploys a web framework that allows user interactions with the models. The application aims to provide medical professionals with a seamless and intuitive interface, empowering them to identify individuals at risk of mental health conditions and promote early intervention.