Authors - Nethika Alagarathnam, Dhanushka Jayasinghe, WU Wickramaarachchi Abstract - Social media platforms, especially Twitter, have become trending sources for public health monitoring, as individuals often share personal experiences related to symptoms, diagnoses, and health concerns. However, detecting personal health mentions (PHMs) in such noisy, short text environments remains challenging. This study investigates about evaluating and comparing three neural architectures including Long Short-Term Memory with word embeddings, a fine tuned Bidirectional Encoder Representations from Transformer model (BERT) and a compact TinyBERT model distilled from BERT. Using a labeled corpus of health related tweets, all models were trained under identical preprocessing, optimization, and evaluation conditions with accuracy, precision, recall, and F1- score assessed on a test set. The results reveal clear performance differences across three architectural paradigms. LSTM baseline demonstrated strong learning on the training set but found significant overfitting and failed to perform on unseen data. But in contrast, the transformer models BERT and TinyBERT delivered a decent balanced performance reflecting the good ability to capture contextual semantics noise in tweets. While BERT achieved the highest overall performance. Notably, TinyBERT provided a competitive and alternate suite for deployment in constrained environment. These findings highlight the effectiveness of transformer architectures for Personal Health Mention detection and practical insights for building efficient and accurate public health monitoring system using social media data.
Open Zoom