Best way to add in 'gazzette-like' word-level features

[Zylatis]

Hey,

Really enjoying using this package but i have a quick question (which may be a bit more theoretical actually) - my understanding is that the lstm part is responsible for generating the feature inputs to the CRF. I would like to add external word-level feature inputs as well to 'lighten the burden' of the model overall (i.e. why make the network learn stuff you can just tell it?). I think i more or less understand how to just add more features into the CRF part, but would it make sense to add this at the LSTM level as well or that wouldn't make sense?

Thanks,
Z