Argument-structured Justification Generation for Explainable Fact-checking
Résumé
Justification production is a central task in automated fact-checking, and most studies cast this task as summarization. However, the majority of previous studies presume the availability of human-written fact-checking articles, which is unrealistic in practice. In this work, we address this issue by proposing a novel approach to generate argument-based justifications to improve fact-checking. Our contribution is threefold. First, our extensive experimental setting shows that, despite lower ROUGE scores, our argument-structured summarizer produces summaries leading to better claim verification performance than the state-of-the-art summarizer in fact-checking on three different benchmarks for this task. Second, our jointly-trained summarization and evidence retrieval system outperforms the state-of-the-art method on ExClaim, the only dataset where no human-written fact-checking articles are provided during verification of news claims. Third, we show that integrating attackability evaluation into the training process of the summarizer significantly reduces hallucinated argument relations, leading to more reliable and trustworthy justification generation.
Origine | Fichiers produits par l'(les) auteur(s) |
---|