Add ConceptNet

[cthoyt]

Closes #15

This PR adds the ConceptNet dataset to PyKEEN. It is distributed as a single gzipped TSV file, so it gets automatically split into training/testing/validations sets. This PR required a small extension to the SingleTabbedDataset base class to allow for the specification of usecols since there are 5 columns in the file (edge URL, relation, head, tail, metadata).

CC @isspek this could be useful for you given your interest in #2

Results

ConceptNet (create_inverse_triples=False)
Name        Entities    Relations      Triples
----------  ----------  -----------  ---------
Training    13506431    17396623      27259908
Testing     13506431    17396623       3407488
Validation  13506431    17396623       3407489
Total       -           -             34074885

Caveats

There's a small concern that the library has some relations ending in _inverse, which PyKEEN interprets in a special way and throws this warning (when running python -m pykeen.datasets.conceptnet to make a summary):

INFO:pykeen.triples.triples_factory:Some triples already have suffix _inverse. Creating TriplesFactory based on inverse triples

TODO

• Implement new class
• Improve data downloader
• Finish an merge Optimize triples splitting and cleanup algorithms #187 for improved split speeds
• Add summary statistics to the PR
• Deal with warning for _inverse suffixes

[mberr]

Not necessarily required for this PR, but it might be interesting to allow keeping the meta-data for triples, e.g. to support qualifiers, cf. e.g. https://www.aclweb.org/anthology/2020.emnlp-main.596/

[cthoyt]

@mberr this PR is technically ready, but the implementation of dataset cleanup using np.isin is restrictively slow for this large dataset and I have yet to be able to generate a split in < 3 hours (after which I gave up each time)

[mberr]

@mberr this PR is technically ready, but the implementation of dataset cleanup using np.isin is restrictively slow for this large dataset and I have yet to be able to generate a split in < 3 hours (after which I gave up each time)

Is this for randomized cleanup? This one moves one triple at a time, right?

[mberr]

@cthoyt I think I have an idea for another cleanup implemention 🙂

[cthoyt]

I did some benchmarking of the cleanup algorithm last week - deterministic is way faster assuming you're okay with its caveats. It's currently the default.

[mberr]

@cthoyt #187

[mberr]

I removed the blocked label since #187 has been merged to master.

[cthoyt]

@PyKEEN-bot test

[mberr]

This is an unrelated change, right?

[cthoyt]

correct, but it was missing and this is a dataset related PR so I threw it in there (the docs weren't showing it becuase I forgot to include it in all)