Information extraction

Information extraction identifies information expressed in text or other unstructured material and casts it into a defined structure. Outputs may include entity mentions, dates, events or relations. Extraction is interpretation: a source sentence can contain negation, uncertainty, conditional language or a date that changes the meaning of a candidate claim. A structured result should therefore retain a route back to the source span.

In practice, extraction is constrained by a target schema and reviewed against examples of correct and incorrect outputs. A system may identify a trial name accurately while misreading a planned start as an actual start. The extractor's confidence, where available, does not establish factual truth. Validation has to address both whether the source was read correctly and whether the source itself is authoritative.

For a defensible record, extraction creates a proposal rather than an automatic fact. The exact evidence span, source date and any qualifiers need to remain visible when the proposal is evaluated. Slot filling is a narrower formulation focused on requested fields; knowledge-base population decides how accepted extractions enter a larger structured collection.

The common errors are systematic and can be catalogued. Hedges and modality are dropped, so that expects to or may become a plain assertion. Negation and scope are lost. Coreference is resolved wrongly, so that a statement about a subsidiary is attached to its parent. Numbers are read without their units or periods. Table cells are read without their headers. A statement quoted from a third party is taken as the author's own. Each of these produces a claim that is fluent, plausible and wrong, and none is caught by checking that the extracted value appears somewhere in the page.

Quality control therefore operates at the level of the claim and its evidence. Keep the evidence span, sample extractions for human review stratified by source and by relation type, and measure precision and recall separately for each relation, since an overall figure hides the relations on which the system is poor. The article on separating source errors from extractor errors describes a way to decide, for each error found, which side caused it.

Extraction produces the candidates that knowledge-base population decides whether to admit, and slot filling is the variant in which the relation and the subject are given in advance. Its outputs are inputs to truth discovery only after they have been aligned to common subjects and fields. Calibrated confidence from the extractor, as discussed under calibration, is more useful to downstream steps than an uncalibrated score.