Explainable Fact-Checking

The natural language generation step that summarizes the reasoning behind a veracity decision. Instead of a black-box score, the system produces a coherent paragraph explaining why a claim is supported or refuted.

• Synthesizes retrieved evidence into a logical narrative
• Highlights the specific text spans that proved decisive
• Often implemented via fine-tuned T5 or LLaMA models on datasets like LIAR-PLUS
• Enables human auditors to quickly validate machine decisions without re-reading the entire corpus

The mechanism that links every atomic claim in a justification back to its exact source document, paragraph, and line. This creates an unbroken chain of custody from raw data to final output.

• Uses document fingerprinting and content hashing to prevent tampering
• Stores retrieval metadata: source URL, retrieval timestamp, and relevance score
• Implements the W3C PROV data model for standardized lineage representation
• Critical for high-stakes domains like journalism and legal tech where citation integrity is non-negotiable

The ability to combine multiple pieces of evidence across different documents to verify a complex claim. A single sentence like 'The CEO of the company that acquired Startup X was born in Paris' requires linking a merger record with a biographical entry.

• Decomposes the claim into atomic sub-questions using Claim Decomposition
• Iteratively retrieves evidence for each sub-question
• Constructs a directed acyclic graph (DAG) of logical entailments
• Benchmarked on datasets like HotpotQA and HoVer

The process of aligning a system's predicted probability of correctness with its actual empirical accuracy. A well-calibrated XFC system says '90% confident' only when it is truly correct 90% of the time.

• Uses temperature scaling or Platt scaling on output logits
• Prevents overconfident misclassifications that erode user trust
• Reports Expected Calibration Error (ECE) as a key evaluation metric
• Allows downstream systems to set risk-appropriate thresholds for auto-publishing vs. human review

A technique that explains a decision by showing what minimal change to the input would flip the outcome. For fact-checking, this means identifying the specific entity, date, or relation that, if altered, would change a FALSE verdict to TRUE.

• Generates statements like: 'The claim would be true if the date were 2023 instead of 2022'
• Helps users understand the precise boundary between fact and error
• Implemented using contrastive explanation methods adapted from computer vision
• Valuable for debugging model bias and identifying brittle reasoning patterns

The systematic inclusion of disagreeing evidence in the final explanation to demonstrate thoroughness. An XFC system must not cherry-pick supporting documents but must explicitly address and refute contradictory sources.

• Uses stance detection to classify retrieved passages as supporting or refuting
• Generates 'However' clauses that acknowledge counter-evidence
• Weighs sources by Source Reliability Score to resolve conflicts
• Builds user trust by showing the system considered alternative viewpoints before reaching a conclusion