Social Media Analysis for Social Good

Data on social media is abundant and offers valuable information that can be utilised for a range of purposes. Users share their experiences and opinions on various topics, ranging from their personal life to the community and the world, in real-time. In comparison to conventional data sources, social media is cost-effective to obtain, is up-to-date and reaches a larger audience. By analysing this rich data source, it can contribute to solving societal issues and promote social impact in an equitable manner. In this thesis, I present my research in exploring innovative applications using and machine learning to identify patterns and extract actionable insights from social media data to ultimately make a positive impact on society. First, I evaluate the impact of an intervention program aimed at promoting inclusive and equitable learning opportunities for underrepresented communities using social media data. Second, I develop EmoBERT, an emotion-based variant of the BERT model, for detecting fine-grained emotions to gauge the well-being of a population during significant disease outbreaks. Third, to improve public health surveillance on social media, I demonstrate how emotions expressed in social media posts can be incorporated into health mention classification using an intermediate task fine-tuning and multi-feature fusion approach. I also propose a multi-task learning framework to model the literal meanings of disease and symptom words to enhance the classification of health mentions. Fourth, I create a new health mention dataset to address the imbalance in health data availability between developing and developed countries, providing a benchmark alternative to the traditional standards used in digital health research. Finally, I leverage the power of pretrained language models to analyse religious activities, recognised as social determinants of health, during disease outbreaks.