What I think when I think about treebanks

In this opinion piece, I present four somewhat controversial suggestions for the design of future treebanks: a) Treebanks should be based on adversarial samples, rather than pseudorepresentative samples. b) Treebanks should include multiple splits of the data, rather than just a single split, as in most treebanks today. c) They should include multiple annotations of each sentence, whenever possible, instead of adjudicated annotations. d) There is no real motivation for adhering to a notion of well-formedness, since we now have parsers based on deep learning that generalize easily and perform well on any type of graphs, and treebanks therefore do not have to limit themselves to trees or directed acyclic graphs.