Semantic Label Accuracy

Explanation

Semantic label accuracy checks whether the meaning-level labels a document-extraction system assigns to content are correct — that is, whether a captured value is tagged as the right thing (this number is the 'sanctioned loan amount', this string is the 'borrower PAN', this block is the 'guarantor clause') rather than merely being read off the page. This matters because even perfectly transcribed text causes downstream errors if it is filed under the wrong meaning: a correct figure mapped to the wrong field can misstate exposure or misroute a KYC check. It is measured by the Semantic Label Accuracy metric (the calcMethod is given simply as Semantic Label Accuracy, so report the rate of correctly assigned semantic labels rather than inventing a formula). Implement it by comparing the system's assigned labels against a human-verified ground-truth set, logging per-label correctness, confusion between label types, and offending documents as audit evidence; run this validation continuously over sampled extractions in the pipeline. The threshold requires greater-than-or-equal-to 98% semantic label correctness; falling below it triggers investigation and remediation of the labelling model or schema mapping before affected extractions are trusted by core systems.