Update TextModels to work with Flux @0.12 and Zygote

[AdarshKumar712]

This PR updates the current version of the models available with TextModels.jl to work with [email protected] and Zygote. Will fix #1 and #19

Models updated so far:

• CRF
• NER
• POS
• Sentiment.jl (no changes)
• averagePerceptronTagger.jl (no changes)

Also updated Project.toml

Work to be done:

• Update ULMFit model
• Update Tests
• Update Docs (Need Suggestions on this!)

PS. Currently I have commented out the section of code belonging to ULMFit Model. Will uncomment it once its updated.

Please let me know if there are some mistakes.

cc @aviks

Update: Updated the tests and checked as well. For ULMFiT model, I am unable to verify the trainable part as the model is running very slow on my system. Rest of the models are working fine.
Update -2 : Documentation for some models, not available earlier is also moved from TextAnalysis.jl

[sambitdash]

I was working on fixing this. But didn't realise @AdarshKumar712 was working on it. So I have submitted all that I have done as part of PR #21.

In #21, I have made some changes to ULMFit, CRF, NER and POS. Some test cases in NER, POS are broken and have been marked as well. I feel there may be a retraining needed for NER and POS as the model is changed in Flux for the LSTM. The precomputed weights are not good enough. The retrainable parameter logic in ULMFit is probably broken. That may not be needed as well. One can provide the @functor macro the relevant params to set the right retrainable parameters.

@AdarshKumar712 feel free to use any code from #21 if you find it of use.

[AdarshKumar712]

I will surely look into the changes you made in your PR. Thanks a lot @sambitdash!

[AdarshKumar712]

Currently, CI build is giving a dependency error while precompiling because of CorpusLoaders.jl (used for ULMFit module). I have created an issue JuliaText/CorpusLoaders.jl#43 regarding the same. Once the issue resolves, hopefully CI build will work properly. And accordingly, I will also update this PR.

[aviks]

So @AdarshKumar712 is this ready to merge? There is a test failure in nightly, but I think that is an upstream issue.

[AdarshKumar712]

Yes, this is ready for now. If there will be any issues, that can be accommodated in future. On a side note, the ULMFiT model is a bit way too slow. So wasn't able to completely check for the Pretraining and fine-tuning. But rest works fine.

[tejasvaidhyadev]

looks good to me

[AdarshKumar712]

Updated as per the suggestions. Hopefully, now it's ready to merge.
cc @aviks @tejasvaidhyadev

[tejasvaidhyadev]

Looks good to me

[tejasvaidhyadev]

@aviks
should we merge it?

[aviks]

Ok, here goes...