Using winobias

[sashavor]

Hello!

I'm trying to use WinoBias as hosted on HuggingFace

I don't quite understand why there is (always) a repetition in the coreference_clusters for each one of the entries, e.g.

['The', 'developer', 'argued', 'with', 'the', 'designer', 'because', 'she', 'did', 'not', 'like', 'the', 'design', '.']
['0', '1', '7', '7']

['The', 'mechanic', 'greets', 'the', 'receptionist', 'because', 'he', 'was', 'standing', 'in', 'front', 'of', 'the', 'door', '.']
['3', '4', '6', '6']

I.e. the last two indices are always identical. Is there a reason for this?

Thank you for any help!

[JieyuZhao]

Hi, yes, because the last two indices refer to the token position for the pronoun "he" or "she" or "him" or "her" which is only one single token. And for the conll style, if it is one single token, the start position and end position will be the same.

[sashavor]

Thank you for your answer!

I am trying to make a fill-in-the-blank style version of this dataset, so removing the token at that index position would be the right way to do that, right?

[JieyuZhao]

Sorry for the late reply. Yes, if you want to predict the pronoun, the [7,7] or [6,6] should be the index.

[sashavor]

Great, thank you so much!