Though I'm thinking porting only the grammar.xml and not the entire program.
For example the NLP package spaCy has a matcher API. I could write some rules for matching on grammar, but I can avoid the cold start problem by porting over the LT rules.
So it's less an LT port but a grammar matching package for spaCy augmented greatly with LT rules (grammar.xml). There will be differences too, e.g., spaCy and LT return different sets of properties to match on.
I agree if it's just using LT in Python, using the server would be best.
If there's interest I could contribute a yml translation script.
Also didn't know about the neural network feature. Looks good.