Improving NER for label Coordinate

Dear all.,

thank you again for updating prodigy and such new property that you add, I am really enjoying using prodigy for NER.

I have trained a custom named entity recognition label by the label. for the most label, it works very well

but some times flr label "COOR" which basically "Coordinate " in a different format ( all old version of data format) I have problem

here you can see the result:

as you see in the second label related to"COOR"

but when I change the text ( here I changed the second label to 12 23 25) it works

why it is so error-prone, I have calculated the metrics:


for label "COOR"

I have very good metric, I know 1 percentage mistake in a large corpus could cause some mistake, but how can I improve it.

my model is made by this way, using regex --> checking by annotation and adding some human level entity---training by prodigy for each label(too see each model) --merging labels ---training whole custom NER

do you have any hint for me to improve my model?
can I use matcher from spacy only for "COOR" to improve the NER Model

Many thanks in advacne


I would be very thankful if someone can give me a response

Hi Robertto,

Please don't tag the team to answer your question, as it wouldn't be fair of us to prioritise questions that tagged the team, and it wouldn't be helpful if everyone did this.

I think the accuracy on your data is high, and so I don't have any advice about what you should do. In general you should not expect 100% accuracy and I do not have recommendations to help you achieve 100% accuracy.

ok, I edited my question and removed the tags.

but my main question was about the reason of such behavior of the model and how can I possibly improve it.

secondly, I asked about how can I use matcher to improve it.

anyway, thank you for your nice answer!