Define a way to improve NER installation through adding new cases and retraining it

[danielkornev]

trafficstars

Want to contribute to DeepPavlov? Please read the contributing guideline first.

What problem are we trying to solve?:

In production systems, it is important to have a simple and non-obtrusive way to extend capabilities on NER to recognize new objects or fix recognition of the existing ones

How can we solve it?:

Web interface for adding new cases

Are there other issues that block this solution?:

[acriptis]

I suppose a good NER Service should cover the following use cases:

1. Annotate requests with NER entites (as any NER tool must do) via REST API
2. Allow Admins to view log of requests for annotation and produced results
3. Allow Admins to add new training cases into the dataset (which will fine-tune NER model instantly or via button).

It seems to be a quite simple web app.

[Rajmehta123]

I agree. The 3rd point seems to be interesting. But, we cant expect the model to fine-tune in real-time. Do you mean to transfer learning? Transfer the weights of the new data to the existing trained model?