An Evaluation of XAI Methods applied to Information Disorder Detection Models

• The use of RoBERTa information disorder detection models for seen and unseen data. • An exploration of XAI methods’ suitability to explain black boxes. • The application and evaluation of XAI using SHAP, LIME, and Vanilla Gradients. • An analysis of Correctness, Completeness, Consistency, Continuity, and Contrastivity. • Insights regarding model classification and explainability of longer news articles. Information disorders like mis-, dis-, and malinformation cause severe consequences from a societal perspective. Due to their complex characteristics or interconnected features, it is particularly challenging and difficult to recognise them. BERT based deep learning architectures are promising as they are capable of detecting information disorders but they potentially suffer from bias arising from the fine-tuning data or internal mechanisms. Explainable artificial intelligence methods can help to understand such classifiers better. Towards this end, we make the following contributions: (i) we provide insights regarding information disorder classification and explainability for longer news articles as they have received limited attention to date; (ii) we apply SHAP, LIME, and Vanilla Gradients to RoBERTa based detection models and domain-specific datasets containing seen and unseen data; and (iii) we exercise an evaluation of those three XAI methods by using multiple quantitative evaluation categories and metrics. Categories that are analysed include Correctness, Completeness, Consistency, Continuity, and Contrastivity. The evaluation sheds light into potential bias that originates from the datasets and assesses the suitability and effectiveness of SHAP, LIME, and Vanilla Gradients to explain the black box.