A little question about ELMo

UKPLab / elmo-bilstm-cnn-crf

BiLSTM-CNN-CRF architecture for sequence tagging using ELMo representations.

Apache License 2.0

389 stars 81 forks source link

A little question about ELMo #24

Open nlp-wh opened 5 years ago

nlp-wh commented 5 years ago

Hello！

According to my understanding, ELMo in this code directly calls the officially provided pre-training model without parameter updates. Because it is trained in unlabeled corpora in the general field, is this suitable for a specific field, such as biomedicine?

Best Regards

nreimers commented 5 years ago

Hi. That is correct. The ELMo embeddings are not updated. This way it is also implemented in Allen Nlp.

If you have a specific domain, you could update the language model for that domain. In the ELMo Paper they did some experiments with fine tuning the language model for a domain

nlp-wh commented 5 years ago

Thank you very much, I am reading ELMo paper and considering fine-tuning.