Which Loss function is used for NER and Transformer tokenizer? #8211
-
|
I'm trying to run the 'en_core_web_trf' on my own training set. The loss after each epoch doesn't seem to improve and the model is performing very bad compared to 'en_core_web_sm' . I'm not able make sense what exactly is the 'Transformer' loss shown on console along with the 'NER' loss. I would also like to know which Loss functions are used for the 'Transformer' and the 'NER'. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
Sorry for the delayed reply on this. The NER loss calculation is here, it's pretty basic but optimized for whole-entity matching rather than best-effort token overlap. Loss for the Transformer is just the backpropagated loss from other components in the pipeline, in your case the NER. In your case you should just be able to ignore it. You can see the source here. |
Beta Was this translation helpful? Give feedback.
Sorry for the delayed reply on this. The NER loss calculation is here, it's pretty basic but optimized for whole-entity matching rather than best-effort token overlap.
Loss for the Transformer is just the backpropagated loss from other components in the pipeline, in your case the NER. In your case you should just be able to ignore it. You can see the source here.