How to add "incorrect_spans" feature into pretrained spacy model

[Ashufet]

Hi Team,

Currently, I am using the pre-trained spacy model "en_core_web_lg" for NER prediction. There are some words for which model is giving the wrong prediction like -

"TestData" being predicted as "Person".

I wanted to use the latest feature("incorrect_spans") of spacy but I am confused about how to do that. As per documentation,
I need to update config file of ner as mentioned below -

from spacy.pipeline.ner import DEFAULT_NER_MODEL
config = {
   "moves": None,
   "update_with_oracle_cut_size": 100,
   "model": DEFAULT_NER_MODEL,
   "incorrect_spans_key": "incorrect_spans",
}
nlp.add_pipe("ner", config=config)

But when I am doing it, I am getting an error as "ner is already added in pipe", which seems reasonable but then how to update ner config file.

Can anyone help here?

[svlandeg]

Hi,

I'm afraid this process is not as smooth yet as we would like (and we definitely want to address this in the future).

If I understand your use-case correctly, you want to fine-tune an existing NER model. That means you don't want to use add_pipe, because that will create a blank model from scratch. Instead, you want to create a new pipeline with your required components by creating a config file that sources the relevant components:

[components.ner]
source = "en_core_web_lg"
component = "ner"

This will pull in the trained NER component from the other pipeline into your new pipeline.

Now, unfortunately the currently pretrained pipelines (e.g. `en_core_web_lg 3.1.0) have this setting by default:

incorrect_spans_key = null

and there is currently no way yet to change it. We will implement this in the future. For now, a kind-of-silly workaround is to either edit your en_core_web_lg package directly or make a copy and edit that. When you do, make the edit in two places:

• change the incorrect_spans_key in the config.cfg in the root directory
• make the same change in the cfg file in the ner subdirectory.

Once that's done, you can source that (edited) ner component as described, prepare your training data with incorrect examples in doc.spans["incorrect_spans"], mix in normal correct examples to avoid catastrophic forgetting, and run spacy train to hopefully get a better, fine-tuned model!

[Ashufet]

Thanks for the update, surely will try the workaround way and update here soon.

[ywsada]

Having the exact same problem, is this functionality implemented now, 2 years later? Or do I still need to perform the workaround? Thanks in advance.

[svlandeg]

As far as I know we haven't actually gotten around to implementing a better workflow for this yet. Sorry - but it's a good reminder for our (ever growing ;-)) TODO list!

[ywsada]

I see, the workaround is fine for the time being, thanks for your response!