Would like to contribute a CCG Supertagging Demo

[jakpra]

Hi, I made an AllenNLP demo for our new tree-structured constructive CCG supertagger and would like to contribute it to the official demo collection, if this sounds interesting to you. Here are some screenshots that show different tagging decisions and confidence of the model for different prepositions in a similar context:

The demo code is in my public fork of allennlp-demo and the model code is in https://github.com/jakpra/treeconstructive-supertagging. As of now, the model code itself is written in pure PyTorch (without AllenNLP) and I added some wrappers to make the demo work.

I just talked about this with @schmmd, who pointed out that it would help if both AllenNLP was a more integral part of my code and my models were ported into allennlp-models. I am indeed considering doing that, but am not sure when I'll get to it. How strict a requirement is that? Is there any chance of getting the demo hosted as is, e.g., by sending over the readily-built docker container?

Would this type of demo be interesting to you in the first place - provided the model code logistics get resolved one way or another?

Thanks, Jakob

[schmmd]

@jakpra as we discussed, I don't think it makes much sense to include this in the AllenNLP demo unless it's demonstrating a model that was built using AllenNLP.

@dirkgr would this be a good addition to our demo suite if it were built using AllenNLP components?

[dirkgr]

It would be neat to have this in the demo, but we're holding first-party models (i.e., models that live in the allennlp-models repository) to a high standard, since we think of them as templates for how models should be written. For example, they have unit tests. If you're willing to do this work, I'd be happy to review it.

We're also talking about some more informal ways of showcasing contributed models, but those plans are not finalized yet. If you decide not to port your model to AllenNLP, I'll keep you in mind for this alternative!