[ENH] Document embedding

[djukicn]

Issue

Bag of words is currently the only method for vector representation of documents in text add-on. This pull request adds another way for doing it using fastText pretrained models described in:
E. Grave, P. Bojanowski, P. Gupta, A. Joulin, T. Mikolov,
Learning Word Vectors for 157 Languages.
Proceedings of the International Conference on Language Resources and Evaluation, 2018.

Description of changes

• A new widget and script for obtaining document embeddings from a given corpus are added.
• Embeddings are calculated on a server thus eliminating the need for loading large models locally.
• Currently supported languages are English, Slovenian, and German.

How to test?

Until the cluster does not get a permanent address is not hardcoded in the script. To test run Orange with setting the environment variable:

ORANGE_EMBEDDING_API_URL=https://apiv2.garaza.io orange-canvas

The environment variable can be also added in the configuration in PyCharm.

Includes

• Code changes
• Tests
• Documentation

[codecov-io]

Codecov Report

Merging #504 into master will increase coverage by 1.15%.
The diff coverage is 95.95%.

@@            Coverage Diff             @@
##           master     #504      +/-   ##
==========================================
+ Coverage   63.76%   64.91%   +1.15%     
==========================================
  Files          59       62       +3     
  Lines        6306     6539     +233     
  Branches      828      851      +23     
==========================================
+ Hits         4021     4245     +224     
- Misses       2151     2157       +6     
- Partials      134      137       +3     

[ajdapretnar]

I like this widget, it works well. I think it we should merge this as soon as it is remotely operational and then we address some minor details in future PRs.

Something that could probably be improved is what happens when the internet is off. It works quite well in Image Embedding, but here it still takes a while for the error to appear. Perhaps it can't be fixed, perhaps it can.

Also, why is auto_commit box so big? There's a weird extra space above the button, but not because of the PR - it seems to be the problem of the inherited box itself. Any ideas how to fix this?

[djukicn]

@ajdapretnar
I haven't been able to reproduce the case where it takes a while for the error to appear. It appears almost instantly for me.
As for the size of auto commit box, I'll explore it further and hopefully find a way to fix it.

[PrimozGodec]

When the internet is off there are two possible cases:

• when internet is already off when a request is sent the error is raised immediately.
• when the connection is lost during the embedding, it happens that embedded is waiting for responses that were already sent and then it times out after a minute. In this case, it takes longer to realize that the connection is off.

The last case is less probable. The same error is in the Image embedding. I think this behavior is ok from the user's point of view. Also when the internet is off and you request a specific website in the browser it will try to get it for you for some time and then write it is not reachable.

[PrimozGodec]

@ajda for me widget do not have any extra space around the auto-apply box

[PrimozGodec]

Now I see what @ajdapretnar taught with the extra space. It seems that it appears only at Ubuntu (maybe Windows). @djukicn, do you also have this extra space above the auto-commit button on the other widgets?

Now URL is set and Matjaz said that the server can be in production for the document embedder. So from my side, it can be merged. @ajda I suggest that you merge it if everything is ok from your side.

[ajdapretnar]

Perfect, I will have a final look and merge. I can't wait for this one! 🎉

[djukicn]

@PrimozGodec @ajdapretnar I've managed to get rid of the empty space by adding box=False property to auto_commit.

[ajdapretnar]

@djukicn could you please rebase the branch so that I can merge asap? :) Thanks!

[djukicn]

@ajdapretnar Done.