search

Alibaba-NLP / ACE

[ACL-IJCNLP 2021] Automated Concatenation of Embeddings for Structured Prediction
Other
296 stars 44 forks source link

Ask about results in Table 2. #17

Closed fym0503 closed 2 years ago

fym0503 commented 2 years ago

image Hi, I read your paper and find in Table 2, for NER tasks on Es, your methods bring very large gains compared with other languages. Do you have some intuition on why it can bring such large gains ?

wangxinyu0922 commented 2 years ago

The BERT model we used for es is a finetuned model from huggingface models. I'm not sure whether the model is trained with a wrong dataset split on CoNLL dataset and I find that the dataset split in README.md is not identical to the official dataset split. I may try to use another Spanish BERT model for the dataset later.

fym0503 commented 2 years ago

Ok, I understand.