Adding the KILT knowledge source and tasks

[yjernite]

This adds Wikipedia pre-processed for KILT, as well as the task data. Only the question IDs are provided for TriviaQA, but they can easily be mapped back with:

import nlp

kilt_wikipedia = nlp.load_dataset('kilt_wikipedia')

kilt_tasks = nlp.load_dataset('kilt_tasks')
triviaqa = nlp.load_dataset('trivia_qa', 'unfiltered.nocontext')
triviaqa_map = {}
for k in ['train', 'validation', 'test']:
    triviaqa_map = dict([(q_id, i) for i, q_id in enumerate(triviaqa[k]['question_id'])])
    kilt_tasks[k + '_triviaqa'] = kilt_tasks[k + '_triviaqa'].filter(lambda x: x['id'] in triviaqa_map)
    kilt_tasks[k + '_triviaqa'].map(lambda x: {'input': triviaqa[split][triviaqa_map[x['id']]]['question']})

It would be great to have the dataset by Monday, which is when the paper should land on Arxiv and @fabiopetroni is planning on tweeting about the paper and facebookresearch repository for the datasett

[fabiopetroni]

Two minor comments:

1. It would be nice to keep the same order for the datasets as in the paper (FC,EL,SF,NQ,D)
2. wrong GitHub link https://facebookresearch.github.io/KILT/ -> https://github.com/facebookresearch/KILT

[fabiopetroni]

as before for the order

[fabiopetroni]

same

[fabiopetroni]

Thanks a lot Yacine! It looks good to me, just a couple of minor cosmetic comments. But it's good to go otherwise

[thomwolf]

Feel free to merge when you are happy with it @yjernite :-)