Closed yhcc closed 2 years ago
Thanks for pointing this out.
I've just checked this issue. It turns out that the affected entities (in the testing set) are 0 for CoNLL03, and 25 for OntoNotes 5. ACE04 and ACE05 do not rely on these codes.
Hence, at maximum (assuming all the affected entities are not recalled), it will cause a 0.2% decrease in the recall rate, and an even smaller decrease in the F1 score on OntoNotes 5. Hence, this will not affect our conclusion in the paper.
Best, Enwei
I will fix this code, and report the new experimental results on OntoNotes 5.
Thanks again.
Never mind. It is not a big issue. Thank you for your reply.
In case anyone may be interested, here are the "new" experimental results:
CoNLL 2003 | OntoNotes 5 | |
---|---|---|
Reported F1 | 93.65 | 91.74 |
New F1 | 93.61 | 91.69 |
In the "new" processing code, we use the original sentences as the reference points when concatenating the sentences to document-level inputs. More specifically, we concatenate sentences such that the resulting sequences do not exceed the length restricted by PLM, i.e., 512.
The code will be updated in the next version.
Interesting framework for NLP. The processing for the over-long document may leak the label information in https://github.com/syuoni/eznlp/blob/5d7b7c2cd7131842ef99b5d0b5dacf53e530900c/eznlp/model/bert_like.py#L415 if the document is too long, this line of code will force the truncation to happen right before the entity (for NER task). Although this will not affect too many samples, I believe we should not use any information from the label.