Information Extraction

Information extraction (IE) converts text into structured records: entities, fields, relations, events, or table rows. Named entity recognition finds typed spans; IE adds schema constraints such as invoice id, total, due date, buyer, supplier, and source offsets. It connects naturally to document understanding when layout matters.

Predicting structured records

An extraction system predicts a record

$$\hat{r}=\{(f_j,v_j,s_j,e_j){\}}_{j=1}^m,$$

where $f_j$ is a field name, $v_j$ is a value, and $(s_j,e_j)$ identifies the supporting span. Sequence models estimate span labels; rule systems can use typed patterns; generative systems should still return evidence spans. Evaluation often checks field-level exact match, not only token F1.

Field exact match is the indicator $1\{{\hat{v}}_f=v_f\}$ for one schema field. Field exact accuracy averages that indicator across required fields and records, so a nearly correct value still counts as wrong when exact downstream values matter.

Worked extraction example

For two invoice snippets, an IE system should return typed fields with evidence, not just highlighted text:

source text	extracted id	extracted total	extracted due date
Invoice 104 total $42.10 due 2026-08-01	104	$42.10	2026-08-01
Invoice 105 total $17.00 due 2026-08-09	105	$17.00	2026-08-09

There are 2 records and 3 required fields per record, so field exact accuracy is $6/6=1.0$ here. The example is deliberately narrow, but it shows the artifact IE must produce: a schema-conformant record that downstream systems can validate.

Caveats

Exact schemas are brittle. Real documents contain missing fields, repeated totals, handwritten corrections, OCR errors, and values split across layout regions. Low-confidence extractions should carry source spans for review. For names or products, pair extraction with entity linking and matching, then report field-level and record-level metrics in evaluation of NLP systems.

References

• Jurafsky and Martin, Speech and Language Processing, 3rd ed. draft
• Finkel, Grenager, and Manning, Incorporating Non-local Information into Information Extraction Systems

Nav

Section — Natural Language Processing

← Named Entity Recognition Entity Linking and Matching →