Abstract

Aims and Background: As the COVID-19 pandemic develops, there is a lot of false information floating around on social media, and the risks are huge. This is why identifying and countering disinformation campaigns is so important. Modern deep learning models that employ Natural- Language-Processing (NLP) approaches, such as Bidirectional-Encoder-Representations-from- Transformers (BERT), have been quite effective at identifying disinformation.

Objectives and Methods: To tackle the spread of false information on COVID-19, we present an explainable NLP approach that utilizes DistilBERT and Shapley-Additive-exPlanations (SHAP), two powerful and efficient frameworks. A dataset consisting of 984 assertions regarding COVID-19 that were factually verified was initially compiled. The DistilBERT model achieved superior performance in spotting COVID-19 disinformation after we doubled the dataset's sample size using backtranslation.

Results: Compared to more conventional machine learning models, it performed better on both datasets. Additionally, we used SHAP to enhance the explainability of the models, which were then tested in amid-subjects experimentation with three constraints: text(T), text+SHAP explanation( TSE), and text+SHAP explanation+source and evidence(TSESE). The goal was to increase public trust in the models' predictions.

Conclusion: Compared to the T condition, the TSE &TSESE constraints showed a substantial increase in participants' confidence and sharing of COVID-19-related information. Improving public trust and identifying COVID-19 disinformation were two important outcomes of our study.

Keywords: Misinformation detection, COVID-19, Twitter, xAI, NLP, Machine Learning