Problem about the label of NER dataset

[WangYuxuan93]

Hi,
I wanted to reproduce the results in the paper and downloaded the dataset from https://xglue.blob.core.windows.net/xglue/xglue_full_dataset.tar.gz
However, I found that the label of the NER dataset is quite different from CoNLL 03.
For example, for instance "Peter Blackburn", in the "en.train" file, the labels are:

Peter I-PER
Blackburn I-PER

While in the original CoNLL 03 file, the labels are:

Peter B-PER
Blackburn I-PER

Briefly, I found that almost all the labels that start with 'B' are replaced with 'I', and only a few labels starting with 'B' still remained in the NER dataset of xglue.
Therefore, I wonder if there is any problem with the NER dataset in xglue.

[Eureka6174]

Hi,

We converted it from IOB2 tag to IOB tag. You could find the detailed description at here.

And our evaluation scripts support both of these two tags. Please feel free to use any of them.

[WangYuxuan93]

Thanks a lot for the information!
BTW, do the different tagging rules have any influence on the final score?
It seems that the IOB tagging rule will lead to very few tags starting with B. Won't this cause any label bias problem?

[Eureka6174]

The performances are almost same, but I can't find the the detailed numbers.