From coarse to fine-grained decomposition: Hierarchical question generation and learned importance for automated fact-checking

Highlights • A hierarchical, importance-aware framework for automated fact-checking • Deconstructs claims hierarchically via a multi-stage question pipeline • An importance-aware integration guided by Natural Language Inference labels • Validates the framework via sensitivity analysis and in-depth case studies Automated fact-checking is vital for combating misinformation by evaluating claim credibility against evidence. Existing methods focus on decomposing the claim into components at the same level and examining them separately with equal importance, overlooking the reality that some components are more critical to a claim’s overall validity than others. Inspired by the human progressive thinking process, in this paper, we propose HILI , a fact-checking framework via HI erarchical Question Generation and L earned I mportance Integration. HILI moves beyond the limitations of single-perspective claim decomposition and equally-weighted verdict aggregation. Instead, it hierarchically decomposes claims into increasingly fine-grained components via multi-stage question generation. Then HILI integrates importance scores to explicitly guide the model in considering the differential impact of each component on the claim’s overall veracity. Specifically, to progressively decompose claim, HILI first generates question-answer pairs organized into three distinct stages targeting the claim’s core thesis, its peripheral details, and its implicit information. Subsequently, to quantify the differential importance of each component, a Natural Language Inference model is employed, producing importance scores for determining the claim’s final veracity. Finally, a classifier reaches a holistic verdict by jointly considering the textual content of each question-answer pair and its corresponding importance scores. Our method achieves new state-of-the-art results on public datasets, with in-depth case studies demonstrating the effectiveness of its hierarchical approach. The source code is available at https://github.com/xuanling-zhang/HILI.git .