Efficient Ensemble-Based Approaches to Personal Health Mention Detection

Abstract

An important component of public health surveillance is the analysis of personal health-related posts on social media platforms, known as Personal Health Mention (PHM) Detection. PHM detection is essential to quickly detecting epidemics, allowing health organisations to prepare themselves and warn the general public to take precautionary steps. One of the key complexities of this task is the informal nature of the language used in social media, which makes it difficult to understand their context. The architectures that are capable of discerning context are also computationally expensive to use and often mistrusted in the medical community due to their decision-making strategies being hidden behind a black box. In this thesis, we address each of these issues separately. We introduce four transformer-based ensemble architectures that have not been previously explored in the PHM domain and show that these architectures can achieve state-of-the art results across the domain. Combined with computationally efficient training mechanisms, our architectures also use fewer resources than existing ones. Addi tionally, we provide methods to explain the outputs produced by the architectures in order to address the concerns related to explainability.