Health Misinformation in the Covid-19 Era - Detecting Misinformation on Bi-lingual Corpora using Lexical Features

ABSTRACT

Social media use spiked amid the COVID-19 pandemic, resulting in an increase in fake news proliferation, especially health misinformation. Many misinformation detection studies have primarily focused on English texts, and of these, very few have examined linguistic features (syntactic, lexical, and semantic). Lexical features such as number of upper-case letters have been shown to improve misinformation detection in English and non-English texts, however, use of lexical features is still in its infancy, and thus warrants further investigation. Therefore, a novel lexical-based health misinformation detection model is proposed using machine learning techniques, specifically focusing on two languages, namely, English, and standard Malay. A new dataset containing fake and real news were developed from a fact- checking portal and local media, targeting news related to COVID-19. Common natural language processing tasks including filtering, tokenization, stemming etc. and lexical feature extraction were administered prior to data modelling. Evaluation on a dataset containing 1060 fake and real news each show Random Forest to yield the best performance with 99.6% for F-measure and accuracy of 96%, followed closely by Support Vector Machine. A similar observation was noted for the Malay corpus. Improved health misinformation detection was observed when linguistic features were included as part of the model, hence implying that the features can be successfully used in detecting fake news. © 2022 IEEE.