Repository navigation
Replies: 4 comments
|
The competency question that existing terms already answer (the history of a value: who or what asserted it, when, and how it was reviewed) is recorded as decided work in #236. That issue also adds the other candidate questions above to the UC-01 ORSD as pending, linked here. |
Proposed direction for points 1 to 5A change of premise first: an extraction run should record everything it produces as ADIRO data, with no side file in another format. So nothing can be parked "as evidence" (#96's JSON/JSONL files): what ADIRO cannot say yet is added to ADIRO. That makes points 1 to 4 needed; point 5 is not. Checked against what a typical title-block extraction actually keeps for each page:
A common need shows what the terms must answer: before processing a page again, did an earlier run already read text from its title block and from the whole page, and was that text empty? 1. Text: a reading, not a property of the region
2. Outcomes: as few new terms as possible
3. Derived images: keep the crop, and say the reading used it
4. Runs: an activity made of steps
5. Permitted uses: not in ADIROKept by whoever holds the drawings, outside the TBox. For a published set it is the same question as #231 (Croissant Two more things the same work needs
Issues to open once agreed
Comments welcome on any point, in particular the three choices above: a new outcome value vs a new class, a derived-image subclass vs a plain |
DecisionAgreed with the maintainer. Each point below becomes an issue, listed at the end. Premise. An extraction run records everything it produces as ADIRO data. What ADIRO cannot say yet is added to ADIRO; nothing is parked in a side file. 1. Text read from a region or a page: a reading
2. Outcomes
3. The images a model saw: frames, not filesA typical extraction renders each page at a fixed resolution, feeds the detector a uniformly scaled-down copy, maps the detected box back to the rendering, crops that box for the reading model (scaled down only above a size cap), and also reads the whole rendered page. Every one of these images is the page, or a region of it, at some resolution, and every change between them is a uniform scale. So:
4. Runs: an activity made of steps
5. Permitted uses: not in ADIROKept by whoever holds the drawings. For a published set, it is #231's question (a dataset manifest's usage terms). Also decided
Issues follow in the next comment. |
|
The issues for the decision above:
#236 (the history of a value) stays as it is; #237 extends its question with the run, the step and the configuration. |
Uh oh!
There was an error while loading. Please reload this page.
Motivation
A title-block extraction run produces more than the field values ADIRO already models. For each page of each file it also yields the text it read, whether it found a title block at all, why it failed when it did, and images cropped from the page for a person to check. Extraction tools usually keep this in formats of their own. The aim is to be able to keep everything a run produces as ADIRO data instead, so one vocabulary serves the extraction, the people who review it, and the queries that use it. That works only if ADIRO can say what that data is.
This Discussion is a brainstorm: what is missing, which competency questions need it, and what should stay outside the ontology. It is not a proposal to add terms yet. Once a direction is agreed for a point, a GitHub issue records the decision and a pull request implements it.
What already exists
Most of what a run produces is covered already:
adiro_provenance:DrawingFile(contentHash,storageLocation,fileModifiedAt,sourceLocation) andFilePage(pageNumber,belongsToFile), from #192 and #229.FieldAssertionwithassertsFieldKind(theTitleblockFieldSchemeconcepts,UnidentifiedFieldfor a caption that maps to none),hasLiteralValue,capturedCaption,hasConfidence, andhasOutcome(NotFound,Illegible,NotPrinted,Empty).InferenceMetawithinferredBy(a person, or a software agent),inferredWith,inferredAt. A person's correctionsupersedesa model's value, andhasAcceptance(Pending,Accepted,Rejected) withacceptedBy,acceptedAtandacceptedUnderrecords the review.adiro_drawing_metadata:Titleblockwith anadiro_geometry:BoundingBox(bboxXYXY).InferenceMetais reused, and that bulky per-occurrence evidence lives in indexed JSON/JSONL files referenced from the graph rather than as individuals. It deliberately assumed no storage mechanism.What a run produces that ADIRO cannot say yet
Titleblock), on the page (FilePage), or on a reading of the region (anInferredEntitywith its ownInferenceMeta, since two runs can read different text)?dcterms:language), and on what?hasOutcomesays why one field has no value. Nothing says that a run found no title block on a page, or that it failed on a page, so no field assertions exist at all. Today "not processed", "processed, nothing found" and "processed, failed" all look the same in the graph.prov:wasDerivedFrom). Is a crop such a file, or is it fully reproducible from the page and theBoundingBox, so that recording it adds nothing?Competency questions this would serve
Candidates, to be checked against the UC-01 ORSD before any term is added. None of them is answerable today.
Out of scope
All reactions