Field Extraction Accuracy

Explanation

Field Extraction Accuracy checks whether the values the AI system pulls out of a document are correct, which matters in document-extraction because even when a field is detected, a wrong value, such as a misread loan amount or transposed account number, can flow straight into a banking decision and cause material harm. It is measured by the Field Extraction F1 Score, computed as 2*(Precision*Recall)/(Precision+Recall), balancing how many extracted values were correct (precision) against how many of the true values were correctly captured (recall). To implement it, maintain a labelled ground-truth dataset of documents with the correct value for each field, compare the model's extracted values against those labels to count true positives, false positives, and false negatives, and log precision, recall, and F1 as evidence; this runs in pre-deployment validation and as ongoing monitoring on sampled production extractions. The thresholds are a target of at least 0.90, with investigation below 0.85 and remediation below 0.80 (proposed): a score in the investigate band prompts a reviewer to examine why values are wrong, while a score in the remediate band requires corrective action such as retraining or improving pre-processing. Its Immediate priority reflects that inaccurate extracted values directly undermine the reliability of automated processing.