A Neural Model to Jointly Predict and Explain Truthfulness of Statements

Automated fact-checking (AFC) systems exist to combat disinformation, however their complexity usually makes them opaque to the end user, making it difficult to foster trust in the system. In this paper, we introduce the E-BART model with the hope of making progress on this front. E-BART is able to provide a veracity prediction for a claim, and jointly generate a human-readable explanation for this decision. We show that E-BART is competitive with the state-of-the-art on the e-FEVER and e-SNLI tasks. In addition, we validate the joint-prediction architecture by showing 1) that generating explanations does not significantly impede the model from performing well in its main task of veracity prediction, and 2) that predicted veracity and explanations are more internally coherent when generated jointly than separately. We also calibrate the E-BART model, allowing the output of the final model be correctly interpreted as the confidence of correctness. Finally, we also conduct and extensive human evaluation on the impact of generated explanations and observe that: explanations increase human ability to spot misinformation and make people more skeptical about claims, and explanations generated by E-BART are competitive with ground truth explanations.