Claim extraction refers to identifying discrete, verifiable propositions within larger text bodies—a critical first step in automated fact-checking and verification pipelines:
Definition challenges:
- Claim granularity: Should each claim be atomic (single proposition) or allow compound structures?
- Implicit claims: How to handle unstated but implied propositions?
- Subjectivity: Separating factual claims from opinion or speculation
- Context dependence: Extracting claims while preserving contextual information (temporal, spatial qualifiers)
Common extraction approaches:
- Rule-based: Pattern matching for claim structures
- Supervised learning: Training on annotated claim corpora
- LLM-based: Using language models with prompting to define and extract claims
- Hybrid: Combining multiple signals
Task-specific considerations:
- Knowledge-based QA: Extracting discrete facts from long-form answers
- Code generation: Treating each code snippet or function as a claim
- Math reasoning: Treating each arithmetic operation as an extractable claim
- Scientific writing: Extracting citation claims (author lists, paper titles, publication years)
Granularity tradeoff:
- Coarse-grained extraction (whole-sentence claims) reduces annotation burden but hides specific errors
- Fine-grained extraction (sub-sentence atomic units) enables precise error localization but increases extraction complexity
Key papers in this wiki
- 2023-chern-factool — Proposes definition-based claim extraction using LLM prompting with specific granularity guidance per task domain; demonstrates that clear claim definition enables effective downstream verification across diverse task types