Official ROUGE added

[ghost]

To fix #68 and to have official scores (compared to #69), official ROUGE script is added (through pyrouge)

Dependencies

• pyrouge package is used.
• Perl is needed to run the official ROUGE script
• libxml-parser-perl might be needed : see this issue

_I'm not sure how to add these dependencies in Travis, if someone could provide help here 👍 _

Notes

Implementation works great. However, a few notes :

• If the ROUGE metric is called several time in a row, the official ROUGE script crash. To fix this, I reloaded scorers after calling compute_metrics. I think the performance are not impacted.
• ROUGE official script does not work for the japanese test (but it works for the french one...). It's a problem in the official rouge script. In the test, I omit the ROUGE metric for this japanese example. We might need to add a mention of this in the README.
• Finally, in the case where several candidates are given, I didn't compute the ROUGE metric for each of them : this would mean to write files and run the official script for each of them. This would be way too long. So I run the official script only once and retrieve only the average scores, not individual ones.

Licence

There is issue with licence : pyrouge didn't include the official script in their repo because the licence of the official script is unknown.

To make nlg-eval easy to use I already included the official script. My opinion is :

• Anyway the official script is not available anymore, and people who want to use ROUGE need to download it from unofficial source.
• Even if the licence is unknown, people use it everywhere without mentionning any licence. Keeping this separate just make it difficult for user.
• As long as we are transparent and very clear on the README, this should not be an issue.

[msftclas]

All CLA requirements met.

[temporaer]

Hey @astariul , thanks for the PR, this is great! Currently the tests fail because pyrouge isn't found. Could you add it to the dependencies in setup.py please?

[ghost]

I need help on this one ^^

[temporaer]

setting aside that travis complains about a memory error, if I run the tests locally I get a permission denied for the ROUGE-1.5.5.pl script. Apparently, executable permissions aren't being set. This seems to be an issue in the pyrouge package. A workaround would be to subclass Rouge155 from pyrouge/Rouge155.py, and overload its evaluate() method by copying the function body and changing what was in line 333 to command = ["perl", self._bin_path] + options.

The next error that pops up is that the wordnet exceptions db file cannot be opened. It appears that this page suggests to [re]build this database to get things to work on windows. Our build server is on linux though, so that maybe the question is where the file is from?