the corrected sentences of ABCN.dev.gold.bea19.m2 only get a f0.5 score 86.45 on Codalab platform.

[14484...@qq.com]

 The process we used:

1.Took the original ABCN.dev.gold.bea19.m2 file 

2.Created the corrected sentences by applying the changes described in the  m2 file

3.Then submit 

is it normal ？ or I forget to do something？

[BEA 2019 Shared Task Organisers]

Yes, that is expected.

This was also raised by someone else in the group: link, and I added a FAQ about it. 

This reason is because hypothesis edits must be extracted automatically, while reference edits were defined by humans. Consequently, auto edits and human edits don't always match. It is also worth mentioning that human edits are sometimes inconsistent and that this is a known problem in GEC evaluation; the Conll shared tasks and M2 scorer suffered from a similar limitation. 

[Abhijeet Awasthi]

Hi,

  Does it mean that the sentences obtained by applying edits in the m2 file to incorrect sentences are not the "original" correct sentences (because auto-edits and human edits don't always match)? 

[BEA 2019 Shared Task Organisers]

No, the corrected sentences are always the same, it's just the edit spans that are different; e.g.

Gold edit: [was eating -> has eaten]

Auto edit: [was -> has], [eating -> eaten]

If you apply the edits in auto or gold, you will end up with the same corrections.