Integrate label-studio for manual curation and annotation

[brownag]

I have worked a bit with label-studio and I would like to integrate it further as a trial way of graphical annotation of various digital assets. Here is a Soil Taxonomy themed Gist for text-based named entity recognition: https://gist.github.com/brownag/c520c0c52fe272341ed7c6ae3c404f05

For basic stuff, we can test it right on the labelstud.io website. Importing arbitrary "content" and specifying a custom template on their website is pretty straightforward.

Playground

With a few standard custom annotation workflows (specified like the Gist above) I think we could have a relatively slick way of identifying key markers of "concepts" in SoilKnowledgeBase. These would then be further processed, defined, categorized, linked to external ontologies and internal resources, etc.

Here are places where this may apply:

• identifying logical elements in Soil Taxonomy clauses
• annotating figures in NSSH (Tracking of media associated with pdftohtml #1)
• block diagrams from SoilWeb API (@dylanbeaudette)

The idea is that the annotation of text and images produces JSON markup that refers to unique line/character position, XY position/regions on images, etc. and assigns a particular label. This could be used for guiding e.g. optical character recognition, extraction of elements for processing or referencing in new contexts etc.

These types of annotations would need to be tied to a specific instance of an asset (e.g. via SHA hash) to ensure that they at least invalidate (need to be re-verified) if the resource changes.

[brownag]

Gists for sample Soil + Bird annotation with label-studio:

birds.R -- sample Munsell "chip" point annotation workflow for soil/bird photos by 4theloveofsoil (on twitter)

2020-12-30-21-48-09-result.json -- Contents of a sample label-studio keypointlabels output of image X,Y positions and labels SoilColor, BirdColor