What is key-value extraction?
Key-value extraction is the recovery of named fields and their values from a document. The key is the label the form uses, such as a registration number, and the value is whatever was entered against it. It covers the majority of routine intake, where a document carries a known set of independent fields.
Also called: Field extraction
Why most intake is answered by the simplest extraction there is
Teams specifying a document pipeline tend to design for the hardest document in the set, which is usually a dense table on a scanned annexure. The ordinary case is a page of labelled fields, and it accounts for most of the volume an intake process actually sees. Treating the common case properly, with its own confidence thresholds and its own validation, recovers more throughput than solving the hard case ever does. The failure to watch for is a flat key-value model applied to a document whose fields repeat.
- Most intake documents carry labelled fields that do not depend on each other.
- A field with a known label can be validated against reference data immediately.
- Confidence is reported per field, so correction effort lands where it is needed.
- Repeating structures defeat a flat field model and require table extraction instead.
How key-value extraction works
- The document type is classified so the expected field set is known.
- Candidate labels are located and matched against the field names of the schema.
- The value region associated with each label is identified from the page layout.
- Text within that region is recognised and coerced into the field's data type.
- Each value is validated against format rules and against records already held.
- Values below the confidence threshold are routed for human confirmation.
How key-value extraction is evaluated
Key-value extraction is judged per field, because an aggregate figure hides the one field that stops a submission. Measure the following on documents from live intake:
- precision and recall per field name
- accuracy on the fields that determine eligibility
- proportion of fields extracted above the confidence threshold
- behaviour on documents where an expected label is absent
- rate of silent substitution where a neighbouring value is captured
Limitations, risks and common failure modes
- Labels vary between regional versions of the same form, so matching degrades.
- A value adjacent to the correct one is captured with high confidence.
- Fields that repeat are collapsed into a single value with no warning.
- Date formats differ across submitting organisations and are coerced wrongly.
- Absent fields are recorded as empty rather than as unfound.