NER Model Features


I’m training an NER model to recognize a custom entity that is specific to my domain. I have a lot of example documents where the first word in the document is an example of the entity I’m after. However, in real data the entity is likely to appear anywhere in the document.

What features does the NER model use to detect entities? Does it use word position? Does it use word context? etc?

Basically, I want to avoid training a model that is biased towards selecting the first word(s) as being the entity that I am interested in.


I think I found my answer here - Basically I want to find a better dataset that is representative of real word data…

The model will definitely pay attention to whether the word is the first one of the document. The features include a window of up to four words either side of the target word, subword features, the previously tagged entity, and the currently open entity.