Evidence reasoning refers to the process of combining and reasoning over multiple pieces of evidence to reach a conclusion about a claim's veracity or the correctness of a statement. In the context of fact-checking and misinformation detection, evidence reasoning addresses the challenge that a single piece of evidence may be insufficient; instead, multiple documents or sentences often need to be considered jointly to correctly verify a claim.
Key challenges in evidence reasoning include:
- Noisy retrieval: Retrieved evidence may contain irrelevant or misleading information.
- Incomplete evidence: The full set of supporting or contradicting evidence may not be easily accessible.
- Multi-hop reasoning: Some claims require connecting information across multiple sentences or documents.
- Evidence aggregation: Combining signals from multiple evidence sources into a single veracity judgment.
Graph-based and neural network approaches have been particularly effective at capturing dependencies among evidence pieces, using mechanisms like graph attention to focus on the most relevant evidence during reasoning.
Key papers
- Zhou et al. (2019) — GEAR: Graph-based Evidence Aggregating and Reasoning for Fact Verification — Proposes fully-connected evidence graph with Evidence Reasoning Network (ERNet) for multi-hop information propagation; uses attention-based message passing to enable evidence nodes to communicate; proposes three aggregation strategies (attention, max, mean) for combining reasoned evidence
- 2020-liu-kgat-fact-verification — Uses edge kernels for fine-grained propagation of evidence clues across a graph
Related topics
- fact-checking (application)
- natural-language-inference (related reasoning task)
- graph-neural-networks (common methodology)