Auxiliary supertagging

[namednil]

We could add another supertagger that tries to predict (for each word) the type that comes out when we evaluate the corresponding (gold) subtree in the type algebra. At test time, we wouldn't do anything with that output but it could potentially increase parsing accuracy.

This might push the NN in the right direction to learn the type system. That hypothesis could be tested by comparing the type-unaware baseline (from ACL 2018) with and without this additional supertagging task.