Skip to content

Field level confidence scores and how to read them on an exhibitor contract

OnboardingUpdated 2026-08-188 min read

In short

Field level confidence scores are probabilities between zero and one that an extraction system attaches to each value it pulls from a document. On an exhibitor contract, a high field score with a low word score means the system found the right box and misread the characters in it, which is a different repair from finding the wrong box.

An exhibitor contract comes back from extraction with the contracted area recorded as 18 square metres. The stand is 180. The system reported a confidence of 0.96 on that field, the coordinator saw 0.96 and moved on, and the floor plan carried a hole in it for six weeks.

Field level confidence scores get used as a single gate in most onboarding pipelines, and the gate is on the wrong number. Microsoft (2026) documents four separate confidences in its extraction service, each describing a different thing that can go wrong, and reading only one of them produces exactly the failure above.

What does a confidence of 0.95 actually claim?

Microsoft (2026) defines field confidence as an estimated probability between 0 and 1 that the prediction is correct, and gives the plain reading: a value of 0.95 indicates the prediction is likely correct 19 times out of 20.

That sentence is more useful than it looks, because it converts directly into an expected error count. If your pipeline auto-accepts 1,200 field instances in a month and every one of them scored 0.95, you should expect about 1,200 times 0.05, or 60 wrong values, sitting in your accepted pile. Nobody plans for 60. Teams read 0.95 as a pass mark and behave as though the accepted pile is clean.

Raise the bar to 0.99 and the same 1,200 fields carry an expected 12 errors. Drop it to 0.90 and you are expecting 120. The arithmetic is trivial and it is the only thing that makes a confidence threshold a decision rather than a setting somebody typed once.

Whether those expected counts hold on your documents is a separate question, and a large one. A score of 0.95 that is right 87 per cent of the time on your certificates is a score that will quietly betray every threshold built on it. Measuring that gap is calibration work and it has its own post.

Four confidences, and what each one is about

Microsoft (2026) separates them clearly, and the separation is what makes diagnosis possible.

Document type confidence describes how closely the analysed document resembles the documents the model was trained on, and a low score points at template or structural variation. A contract from a new sales region with a redesigned header will score low here before any field is considered.

Field level confidence describes the model's confidence in the position of the value it extracted, which is to say whether it found the right box.

Word confidence attaches to each transcribed word and describes the character recognition. Microsoft is explicit that field results and read results should both be checked, since a low read confidence and a low field confidence call for different fixes.

Selection mark confidence covers ticks, boxes and check marks, and it carries the detection of the mark and its state together.

Four numbers describing four different failures, and a pipeline that stores only the field score has thrown away three quarters of the diagnostic information the system already computed for it.

The separation has research practice behind it. Huang and colleagues (2019), setting up the ICDAR competition on scanned receipt understanding over 1,000 annotated receipt images, split the work into three graded tasks: locating the text on the page, reading the characters, and pulling out the key values. Those are three distinct ways for the same document to fail, and any pipeline that reports one score has merged them back together.

Reading a contract line by line

Take one exhibitor contract through extraction and look at both scores per field. Figures below are invented to make the readings concrete.

The exhibiting company legal name scores 0.97 on the field and 0.62 on its worst word. The system is confident it found the name box and much less confident about the characters inside it. On inspection, "Kramer & Sons" has come through as "Kramer 8 Sons", which will fail every downstream match against your account record and will be reported as a missing company rather than as a misread one.

The booth number scores 0.71 on the field and 0.99 on its words. The characters are certain and the location is not, because this contract prints the hall number and the booth number in adjacent boxes with similar labels. The value extracted may be perfectly transcribed and belong to the wrong field entirely.

The total contract value scores 0.94 on the field and 0.93 on its words, which is the boring case and needs nothing.

The contracted area scores 0.96 on the field and 0.58 on its words, which is the 18 against 180 from the opening. Field confidence was never going to catch a dropped digit, because the box was exactly where the model expected it.

The pattern is worth stating plainly. High field score with low word score means look at the characters. Low field score with high word score means look at where the value came from. Low on both usually means the document itself is poor, and that is a scanning problem to send back to the exhibitor. High on both is the only combination you can accept without looking.

Combining field and word confidence

The simplest composite that works is the product of the field score and the lowest word score in the extracted value.

On the four fields above that gives 0.97 times 0.62, which is 0.60, for the company name. The booth number gives 0.71 times 0.99, or 0.70. The contract value gives 0.94 times 0.93, or 0.87. The contracted area gives 0.96 times 0.58, or 0.56.

The ranking inverts. On field confidence alone the company name at 0.97 looked like the safest value on the page and the booth number at 0.71 looked like the riskiest. On the composite the company name is second worst and the booth number is comfortably ahead of it. A review queue sorted on field confidence would have shown the coordinator the booth number and hidden the two fields that were actually wrong.

Use the minimum word score rather than the mean. A value of "1,000,000" where nine characters score above 0.98 and one scores 0.41 is a value with a wrong digit in it, and the mean hides that completely.

The composite is a heuristic with no probabilistic meaning, so treat it as a sort key for a review queue and keep it out of reports. The number that belongs in a report comes from scoring extraction against a held out gold set, where correctness is known rather than estimated.

How should a low score change what you do?

Differently per field, which is the part most pipelines skip.

A low score on the exhibiting company legal name should route to review, because everything downstream joins on that value and a wrong one creates a duplicate account. A low score on the signatory's job title should route nowhere, because nobody reads it. The score is identical and the document is identical, and the two correct responses are opposites.

The routing decision belongs to a threshold per field, chosen from your own error data, and picking those thresholds off a risk coverage curve is a measurable exercise rather than a matter of taste.

When a value does go to review, keep it on the screen. A coordinator confirming or correcting a proposed 180 works in a few seconds. A coordinator handed a blank box and a scanned page has to find the number, and that takes an order of magnitude longer. Systems that clear low confidence fields to empty in the name of safety are quietly making the review queue expensive enough that people stop working it.

Selection marks are where contracts go quietly wrong

Exhibitor contracts are full of ticks. Corner booth, yes or no. Insurance acknowledged. Rebooking option taken. Marketing package selected.

Microsoft (2026) scores these separately, with the selection mark confidence covering both the detection of the mark and its state. A confidence of 0.55 on a tick means the system is close to guessing, and the difference between selected and unselected on a rebooking option is a revenue line.

Ticks fail in ways text does not. A pen mark that overlaps two boxes. A form filled in on screen and printed, where the tick is a font glyph. A scan where the box has come through faint. None of these produce a garbled string that looks obviously wrong, and a coordinator reviewing a summary screen sees a clean yes.

Set a much higher acceptance threshold on selection marks than on text fields, and always show the cropped image of the mark next to the extracted state in review. It costs a few pixels and it removes most of the argument.

Where this stops

Confidence is the system's own opinion of its work, and a system can be confidently wrong. The classic case is a field the model has seen in one position ten thousand times: it will report high confidence on a document where that field has moved, because everything about the layout says the value should be there.

Confidence is also unavailable in places. Not every field returns a score, and older versions of extraction services return them for some result types and not others. A pipeline that assumes a score exists on every field will treat a missing score as zero and route perfectly good documents to a person, or treat it as one and route nothing.

The deepest limit is that a confidence number carries no information about consequence. The model has no idea that the general liability limit matters and the fax number does not, and it will happily return 0.99 on the fax number and 0.71 on the limit. Consequence is your input, and it comes from the field list and the onboarding rules rather than from anything the model can see on the page.

Pull last month's extraction output for one document type and store the word confidences alongside the field confidences, which most pipelines already receive and then discard. Sort by the product of the two and read the worst twenty. If any of them were auto-accepted, you have found your first real defect list, and it will not be the list you expected. Where classification put those documents in the first place is measured separately, in document type classification accuracy.

Questions people ask about field level confidence scores

What does a field confidence of 0.95 mean?
It is the system's estimate that this particular prediction is correct, expressed as a probability. Microsoft documents it as meaning the prediction is likely correct 19 times out of 20. Across a batch of a thousand fields all scored at 0.95, you would expect roughly 50 wrong values if the estimate holds on your documents.
What is the difference between field confidence and word confidence?
Field confidence reflects how sure the system is that it located the right value for a labelled field. Word confidence reflects how sure the optical character recognition is about the characters it transcribed. A field can be located perfectly and transcribed badly, or transcribed perfectly from entirely the wrong part of the page.
Should low confidence values be discarded?
No. Route them to a person and keep the value visible, because a low confidence extraction is usually close to correct and much faster to fix than to retype. Discarding the value forces a coordinator to find the number on the page from scratch, which costs more time than confirming or amending what the system proposed.

Related reading

All onboarding operations articles