What is document classification?
Document classification is the identification of what kind of document an upload actually is. It assigns the file to a class the organisation recognises, such as a registration certificate or a laboratory accreditation. Every subsequent decision depends on this, since the rules for extraction and expiry differ by class.
Why everything downstream depends on getting the class right
The first week of any document-heavy intake produces a queue of files named scan0001 and document final v2. An applicant uploading six certificates under one requirement behaves reasonably from their side of the form while leaving the organisation to establish which file is which. Classification is where that is settled, so an error here propagates silently into the extraction rules applied next. The resulting values then look entirely plausible to everyone who reads them.
- Extraction rules differ by class, so a misclassification corrupts every value taken.
- Applicants attach documents to whichever requirement seems closest to them.
- A misclassified document produces plausible values that nobody questions.
- Confident classification is what allows the remainder of intake to be automated.
How document classification works
- The classes are defined from the documents the scheme actually receives.
- Each uploaded file is examined as rendered pages rather than by filename.
- The classifier returns a class with a confidence value for that assignment.
- Low confidence assignments are referred to a person before extraction proceeds.
- The assigned class is checked against the requirement the applicant attached it to.
- Confirmed corrections are retained as training material for the next revision.
How document classification is evaluated
Document classification is assessed per class rather than on an overall figure. Measure the following against assessor-confirmed classes:
- accuracy per class, including the classes that arrive rarely
- confusion between classes that look similar on the page
- proportion of assignments falling below the referral threshold
- calibration of confidence against observed accuracy per band
- rate of documents arriving in no defined class at all
Limitations, risks and common failure modes
- An overall accuracy figure conceals a rare class that is never identified correctly.
- Classes defined from a pilot miss the variants a real scheme receives.
- A document bearing two certificates is assigned to whichever appears first.
- Low confidence referrals are approved in bulk when the queue lengthens.