Decomposition boosts claim verification only with granular, sub-claim aligned evidence; repeated claim-level evidence degrades performance. Noisy sub-claim labels propagate errors unless using conservative abstention. New dataset features annotated evidence spans.
Key Points
- 1.Requires strict evidence alignment for gains
- 2.Abstention curbs error propagation
- 3.Inconsistent results from overlooked bottlenecks
Impact Analysis
Guides better frameworks prioritizing evidence synthesis and label calibration.
Technical Details
SAE vs SRE setups on PHEMEPlus, MMM-Fact, COVID-Fact.