numind

/

NuNER-v0.1

like 57

I haven't used the model yet. I read your blog post, I benefited very well, thank you.

https://www.numind.ai/blog/a-foundation-model-for-entity-recognition

Thank you for the great work. Do you have an example of code where this model was use for tasks like NER or QA ?

Thank you for the great work. Do you have an example of code where this model was use for tasks like NER or QA ?

Yes +1 for NER code example

Hello @etiennebcp , thank you for open sourcing this model.
I have read your blog post carefully, and wanted to thank you for the explainations and the great work.
Just to be sure I understood correctly the approach, in order to use it in a NER settings (inference) we also need the concept embedding model, right?

Concerning use cases, we have a use case for custom and specific information extraction from news and companies websites content (e.g extracting whether a company has a facility, extracting different types of niche technologies). I think this approach could be useful because each of our client project requires a specific ontology and efficient fine tuning and F1 are very important.

Thanks everyone for your feedback! To give a overall answer:

1. This model is only for NER, not QA
2. This model cannot be used in a zero-shot way. It has to be trained on some annotated NER examples to be customized to a particular NER problem.
3. The concept encoder is only there to train the original model (that we now call NuNER). You do not need the concept encoder to use NuNER.
4. The simplest way to use this model is to add a linear layer after the text embeddings to be compatible for a specific NER problem. Then you need to either train this linear layer by itself (equivalent to a logistic regression on top of the embeddings), or fine-tune the entire network.

We are going to release a collab notebook that you will be able to use and train NuNER

Also, you can directly use our tool NuMind to annotate and train automatically :)