Nice summary of Mart’s work.
I take issue with the implication there exists an agreed upon definition of true “natural language” search. Purists would say it’s a QA interface. Is an effective QA interface *still* search? It would seem to me you’re approaching digital-assistants at that point (and the use-case is no longer truly research).
Also, yes the capabilities and future of neural-nets and deep-learning shouldn’t be discounted — but online learning (non-manually updating weights) is nothing new.
Further, what about the structural concerns the legal profession (e.g. the billable hour)?… which create much stronger and deeply embedded disincentives to administrative efficiency than the choice of legal research tool ever will. It’s all first-order pressure against development and adoption of legal technology (not to mention the data/technical challenges these same structural issues introduce with creating the tech in the first place).