Using entity specific custom features in NER

[mapadofu]

Generically, is there a way to use additional/custom features as inputs to NER model? I vaguely recall that earlier versions of the software did support this and how to do it was described in the documentation. However I haven't been able to locate this in the current version.

Specifically, I've got some custom entity types similar to serial numbers. I can flag most instances of them using a rule-based matcher, and could use that matcher to set custom attributes on the token that would indicate "this token 'looks like' a serial number". The idea would be to incorporate these highly informative features into the NER model that I plan on training on the relevant dataset.

[polm]

What you may be remembering is that you can customize the attributes used in the tok2vec layer. This can capture information like prefixes/suffixes, whether a token is numeric, etc.

The default tok2vec only supports Token attributes (the ones here). You can implement a custom tok2vec layer to use arbitrary information, though it's a little more complicated, and will be slower. You can see an example of someone doing that here).