Claim Decomposition

Each sub-claim must be a single, self-contained factual assertion that can be verified or refuted without relying on the truth value of other sub-claims. This atomicity ensures that evidence retrieval is targeted and that the failure to verify one component does not cascade. For example, the statement 'Tesla, founded by Elon Musk in 2003, launched the Model S in 2012' decomposes into three independent claims: (1) Tesla was founded by Elon Musk, (2) Tesla was founded in 2003, and (3) Tesla launched the Model S in 2012. Each can be checked against a distinct source.

The level of decomposition is a critical design choice. Over-decomposition can strip necessary context, making a claim nonsensical ('Elon Musk' is not a verifiable claim), while under-decomposition leaves compound assertions that are difficult to verify. Optimal granularity targets predicate-level units—a single subject-predicate-object triple. For instance, 'The CEO announced record profits and a new product' should be split at the conjunction, as the evidence for 'record profits' and 'a new product' will likely reside in different documents or sections.

A sub-claim must be a self-contained proposition that does not require its original textual context to be understood. Pronouns and ambiguous references must be resolved. The sentence 'He announced it on Tuesday' is not a valid sub-claim. Decontextualization transforms it into 'Elon Musk announced the acquisition of Twitter on October 27, 2022.' This process often involves entity linking and coreference resolution to replace all anaphora with their explicit referents, ensuring the claim is interpretable by a retrieval model in isolation.

After decomposition, sub-claims can be categorized by their verification profile. Verifiable claims are factual assertions with an objective truth value that can be checked against a knowledge base. Unverifiable claims include subjective opinions ('the best car ever'), future predictions, or statements requiring privileged access. A robust system must classify these types to avoid hallucinating evidence for an opinion. This categorization often uses a natural language inference (NLI) model to predict whether a claim is factual, opinion-based, or a conditional statement.

Complex claims often contain logical operators like negation, conditionals, and quantifiers that must be preserved in the decomposition. The claim 'No employees were injured in the accident' should not be decomposed into 'Employees were injured in the accident' with a separate negation flag, as this loses the semantic relationship. Instead, the negation is an integral part of the atomic claim. Similarly, a conditional like 'If the deal closes, the company will relocate' must remain as a single hypothetical unit, as its truth is contingent on an unrealized premise.

Every sub-claim must maintain a provenance pointer back to the exact segment of the original source text from which it was derived. This is critical for explainability and error analysis. If a sub-claim is found to be false, the system must trace the error back to the original statement to determine if the source was incorrect or if the decomposition process itself introduced a distortion. This is typically implemented by storing the character-level offsets of the source span alongside each generated sub-claim in a structured manifest.