IMO it would be interesting to try to combine the two approaches (curation + auto tagging).
It starts out with the user scaffolding an initial hierarchy, then (after enough usage to provide meaningful data for ML predictions) the ML model predicts on subsequent entries, and asks the user for approval (which feeds a reinforcement learning model)
This is indeed the plan. We're currently working on generating embeddings for the all the bookmarks stored, and one of the usecases of this is going to be clustering. If a bookmark is similar to all other bookmarks in a list, the model can suggest adding those bookmarks to the list. Still a manual operation, but with ML assistance.
IMO it would be interesting to try to combine the two approaches (curation + auto tagging).
It starts out with the user scaffolding an initial hierarchy, then (after enough usage to provide meaningful data for ML predictions) the ML model predicts on subsequent entries, and asks the user for approval (which feeds a reinforcement learning model)