NER - improve recall #10127
-
|
I know there is no confidence scores for NER component, which is very strange, but is there a way to control recall for this component? like setting threshold settings, recall is very important for me comparing with precision. |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments
-
|
One of the things that I have tried and got a bit of improvement is to train my own tok2vec model using this: https://spacy.io/usage/embeddings-transformers#pretraining Anything else where should look into, would love to get any help! Thanks |
Beta Was this translation helpful? Give feedback.
-
|
No, we don't really have any parameters you can fiddle with in the NER component to directly influence this. It's from 2019 but you can see a comment about how to approach this here. It's worth keeping in mind that we don't hide these parameters out of forgetfulness - the NER architecture we use is a bit unusual (it's the same as our parser) and it's not always clear how parameters common in other systems map to it. What you can do is use a spancat instead and use a lower than default threshold. |
Beta Was this translation helpful? Give feedback.
No, we don't really have any parameters you can fiddle with in the NER component to directly influence this. It's from 2019 but you can see a comment about how to approach this here. It's worth keeping in mind that we don't hide these parameters out of forgetfulness - the NER architecture we use is a bit unusual (it's the same as our parser) and it's not always clear how parameters common in other systems map to it.
What you can do is use a spancat instead and use a lower than default threshold.