Tokenization is not consistent across datasets

[JohnGiorgi]

Tokenization is not the same across datasets. I don't know how big the issue is, but JNLPBA seems to have coarser tokenization than the other datasets.

For example, in JNLPBA "interleukin-n" is kept together while in the other datasets it appears as "interleukin", "-", "n".

Replace the JNLPBA corpus here with the one from: https://github.com/spyysalo/jnlpba

This will involve

• Removing duplicates
• Creating the valid split
• Outputting in both IOB and IOBES, and for each entity type