In case someone wonders whether it is possible to use external resources for solving, here's the answer:
The training/dev the data is probably not large enough to train any generalizable neural model from scratch, and it wasn't really intended to do so.
So any external knowledge (transfer learning, embeddings, knowledge bases, anything) is allowed to help as long as the test data aren't touched during learning, as in any standard NLP task.
Ivan
Posted by: ivan.habernal @ Jan. 4, 2018, 7:06 a.m.