Turning Unstructured Text Into Fields

Three components, not one

1. Extraction — pulling fields from the document or text
2. Validation — checking against master data and internal consistency
3. Routing — confident items proceed, uncertain items reach a person

Extraction alone might reach ninety per cent. Extraction with validation reaches the same and knows which ten per cent it got wrong, which is what makes automation safe.

Validation is where accuracy comes from

• Does the referenced entity exist in your records?
• Do the components sum to the stated total?
• Is the date within a plausible range?
• Is this a duplicate of something already processed?
• Are required fields present and in the expected format?

Report accuracy per field

A ninety-five per cent per-field rate across twelve fields means barely half of documents are perfect. Reporting a single accuracy figure without saying which is misleading.

Straight-through rate is the one that determines how much work actually disappears, and it is the number to agree on before building.

Design the review interface carefully

The review screen determines whether the system saves time. Source document beside the output, cursor on the first uncertain field, keyboard-only completion, and corrections captured.

Ten seconds per review makes a twenty per cent review rate workable. Two minutes per review means people approve without looking, which is worse than no review.

Preserve the source

Always store the original document linked to the extracted data. When a value is questioned months later, the answer is in the source, and the audit conversation is short.

It is also what makes reprocessing possible when extraction improves.

Frequently asked questions