Task-Oriented Automatic Fact-Checking with Frame-Semantics

Automatic fact-checking is a critical tool for addressing the growing challenge of misinformation, particularly when verifying novel claims that lack previously curated evidence. Prior work has largely focused on unstructured or highly-curated data sources, limiting scalability and generalization. In this work, we introduce a novel paradigm for task-oriented automatic fact-checking using frame semantics to improve interpretability and evidence retrieval from high-volume structured datasets. We present a new pilot dataset of real-world factual claims grounded in two large-scale databases: U.S. congressional voting records and OECD country statistics. Across two case studies using these datasets, we demonstrate how frame elements can guide fine-grained retrieval and automatic fact-checking. Our experiments show that frame element-driven evidence retrieval improves recall by 14% and 11% over full-claim baselines in the voting and OECD case studies, respectively. We further analyze frame distributions in PolitiFact fact-checked claims and find a strong alignment with the frames targeted in our study. Overall, our results highlight frame semantics as a promising foundation for scalable, interpretable, and domain-adaptable automatic fact-checking.