Back to seenX
In development
Learning

Ontology and model-assisted learning

Connect images to concepts while keeping every claim traceable.

Bounded experiment · not implemented

What an ontology adds

An ontology defines concepts, properties, and relationships. Image, Receipt, Trip, and Place form a small starting vocabulary. Particular facts populate that model: a receipt image is associated with a specific trip. A graph database is optional; ordinary tables can represent these relationships.

“Find receipts associated with my Alpine trip.”

OCR can find merchant names and dates. Explicit relationships can connect a receipt to a trip even when the trip name is absent from the image. A model can suggest the association, but the ontology alone does not establish that it is true.

What learning means here

A pretrained model proposes structured facts from images, OCR, and metadata. Validation checks their shape, while source evidence and user confirmation establish their status. Corrections update stored knowledge. This does not change model weights or train a new model.

Keep evidence with each claim

  • The source image and supporting text or metadata.
  • The extraction method and model revision, when applicable.
  • Whether the association is proposed, confirmed, or rejected.
  • Dependencies that must be revisited after correction or source deletion.

Recorded GPS, a user label, and a visual location guess must remain distinct. An OCR place name is not proof of capture location. Changes to concept definitions deserve review separately from new facts.

A manageable experiment

Use roughly 30–50 appropriate images and label receipt-by-trip queries before testing. Compare OCR and metadata alone against the same pipeline with relationships. Record retrieval gains, false associations, extraction time, and correction effort. Keep the layer only if its benefits justify its costs.

What comes later

Entity resolution, temporal facts, and formal reasoning are useful topics after the small experiment. Automated definition changes and model fine-tuning are outside the initial scope.