Bert getting wrong spans (GoldP task)

[rahular]

Similar to #9, I get wrong spans if I follow the instructions and run the original BERT code on the GoldP task. I solved it by normalizing the data while reading and replacing the non-breaking space (rampant especially in Russian) to the regular space u.normalize('NFKC', <text>).replace(u"\xa0", u" "). It would probably save some time for others if this is mentioned in the readme.
I would have sent a pull request, but the code changes have to be made in the BERT repository and not here. A potential solution would be to release pre-normalized UTF-8 encoded data.

[PluviophileYU]

Hi, do you mean that sometimes the label for GoldP task is not correct?