Named Entity Recognition, Can You Selectively Train?

[hellyars]

I am playing around with the Named Entity Recognition Tool.

The vanilla model is (for the most part) correctly identifying locations and company names.  

I do not want to mess with something that is working, but the vanilla model (understandably) does not recognize contract numbers.

Is there a way to train a model to incorporate one or two new entity types/variations without negatively impacting the base model?

[gautiergodard]

Hey @hellyars 

You could probably try one of the MD5 formulas to mask your contract numbers. More info found below:

https://help.alteryx.com/20221/designer/string-functions

[AndrewSu]

@hellyars , that is possible with the "Train with New Entities" option.   see screenshot below. 

See documentation: https://help.alteryx.com/20221/designer/named-entity-recognition 

For the contract numbers in question, is there some logic/pattern to how long they can be? Perhaps a named range?  I'm thinking that you can do a a Regex Replace if that is the case.  That can be one method bypassing NER entirely. 

[hellyars]

@AndrewSu 

I tried this approach previously.  I don't recall the details at the moment, but it failed every time -- even a simple list of 20 or 200 or 2,000 company names.

This question is a little different.  Assuming the above issue is not an issue, can I train a model to identify company names and just company names without impacting its ability to identify locations (e.g., GPE).

Of course my task would be a lot easier if the Text Mining Tools could just parse sentences in a paragraph.

[AndrewSu]

@hellyars ,  I've confirmed that the ability to manually configure has an ADDITIVE effect and merely expands the list of entities it can recognize.

Please mark this post as the solution if it answers your main question so that other's in the community can benefit from our collaboration. 

Thanks!