Open Dtphelan1 opened 3 months ago
Great point. This is a useful feature that we don't have right now. Andy had written something like this for a covid paper, but I had questions about how to accurately generalize what he did, without introducing new math bugs of my own. :smile:
This is the old code he wrote, FYI.
My gut prefers the CLI approach - like just specify multiple annotators on the CLI and this mode would kick in.
Motivation: Often multiple annotators divide and conquer for a single annotation task. The combination of their annotations are then used as a ground truth against which we compare some NLP/LLM system's output.
To accommodate this workflow – compare one set of output against multiple annotators ground truth – there should be some way of specifying a set of multiple annotators as a single "ground truth" annotator.
One approach might be to support this at the config level; similar to
grouped-labels
, we could introduce something likegrouped-annotators
, allowing for the aliasing of multiple annotators under a single, new annotator id.Another approach might be to take this from the command-level; create some flexibility on the CLI side to support specifying multiple annotators for the ground-truth argument.