Feature request: add neural models for dep parsing?

[ym-han]

Hi David,

This looks like a great package. Just wanted to say that it might be nice to have neural models for dependency parsing as well. Perhaps one fairly easy option might be to wrap Stanza's models? (Though of course a native Julia option would be nice.)

[dellison]

Thank you @ym-han, and I'm sorry I didn't see this until now!

I agree that that would be really good to have in Julia. I think my own feeling is that I would rather see something like this in a separate package that imports DependencyTrees and whatever libraries it would need for the model- maybe Flux for the neural model or PyCall for wrapping Stanza. My idea for DependencyTrees.jl was just to provide data structures and parsing algorithms for working with dependency parses with as few dependencies as possible. In fact, at first I was using LightGraphs.jl for the tree representation, but spent a lot of effort rewriting things to avoid the need for the external dependency, since I needed only a tiny bit of what that package provides. Then again, this was just my own design decision motivated by my own opinion, and could be worth revisiting.

At some point last year, I had an implementation of the Kiperwasser & Goldberg BiLSTM transition parser from their 2016 paper using DependencyTrees.jl and Flux.jl, but I remember that the performance I was seeing was quite poor compared to the results they reported getting in the paper, and I never managed to figure out where my error was. That was before my big refactor, and I'm not doing much NLP these days so it hasn't been at the front of my mind quite as much), but seeing this issue is a good reminder to revisit this! :) And now that I mention it, I'm remembering that I also implemented a port of this parser and got slightly better speed and performance that I did with their implementation. Maybe I'll see if I can dig that back up and get it released as a package- it's not neural (it uses an averaged perceptron) but that might be a good start, if nothing else.

In any case, I guess I'd say that I'm probably not very inclined to work on this feature myself from within the package, but I would be really happy to review any pull requests I receive and reconsider my view, or help with the development of any packages that use DependencyTrees.jl to provide this! I'll leave the issue open since it seems like there could be some good discussion on this point in the future. :D