Authors: H Ullrich, T Mlynář, J Drchal
Venue: arXiv preprint arXiv:2502.04955
Year: 2025
Citations: 2
Links
- arXiv: 2502.04955
Abstract
The work examines claim extraction as a dedicated subtask for fact-checking, introducing data, modeling strategies, and automatic metrics tailored to claim quality. It analyzes the trade-off between extracting many candidate claims and extracting claims that are specific, verifiable, and useful downstream.
Resources
- Video: TODO
- Slides: TODO
- Code: TODO
- Dataset: TODO
Notes
In practice, this paper helped us treat claim extraction as a first-class modeling problem rather than a preprocessing detail. Better claim quality noticeably improved retrieval focus and final verification reliability in later pipelines.