Vaccine Hesitancy Detection Using BERT for Multiple Social Media Platforms

Abstract

Vaccination has been proven to be an effective measure to prevent the spread of harmful diseases. Despite its efficacy, the moves towards vaccine hesitancy have been receiving global attention. Vaccine hesitancy issues have been openly discussed across major social media platforms including Facebook, Reddit, Twitter, Instagram and YouTube. The spread of vaccine hesitancy-related posts is propagated substantially, causing greater threats to public health. Consequently, various state-of-the-art machine learning techniques have been proposed to analyse vaccine-hesitant related posts in social media. One of the most recent approaches is the transfer learning method using a pre-trained Bidirectional Encoder Representations from Transformers (BERT) model. Despite vaccine hesitancy being a prevalent issue across multiple social media platforms, only a few studies have utilised data from multiple social media platforms to detect vaccine hesitancy. To address this research gap, the use of BERT as one of the new language representation models is adopted to train from a collection of vaccine hesitancy related data from multiple social media platforms. Moreover, this study employs the Support Vector Machine (SVM) and Logistic Regression (LR) models and compare their performances against the BERT method. The objectives of this research are threefold; to establish a consolidated dataset from multiple social media sources for use in vaccine hesitancy detection, to evaluate the effectiveness of using mono-platform versus multi-platform vaccine hesitancy data on the performance of different machine learning models and to apply a transfer learning method using BERT in vaccine hesitancy detection. A collection of 193,023 labelled vaccine hesitant posts were aggregated from three (3) social media platforms which includes Facebook, Reddit, and Twitter. The results demonstrate that the BERT model performs the best and achieved an F1-score of 0.93, while both the SVM and LR achieved F1-scores of 0.90 when detecting vaccine hesitancy from multiple social media platforms. Our proposed research also revealed that models trained with multi-platform data perform at least 15% better than models trained with mono-platform data when tested with multi-platform data.