A Benchmark for Knowledge Intensive Language Tasks

http://kiltbenchmark.com/

The KILT benchmark is described in the following paper:

@inproceedings{petroni2020kilt,
 title={KILT: a Benchmark for Knowledge Intensive Language Tasks},
 author={Fabio Petroni and Aleksandra Piktus and Angela Fan and Patrick Lewis and Majid Yazdani and Nicola De Cao and James Thorne and Yacine Jernite and Vassilis Plachouras and Tim Rockt{\"{a}}schel and Sebastian Riedel},
 booktitle={arXiv:2009.02252},
 year={2020}
}

https://arxiv.org/abs/2009.02252

Setup the env

conda create -n kilt37 -y python=3.7 && conda activate kilt37
pip install -r requirements.txt

KILT knowledge source

The KILT knowledge source can be downloaded here: kilt_knowledgesource.json (34.76GiB).
It is based on the 2019/08/01 Wikipedia dump.
We use mongoDB to index the knowledge base (but you can use any json-based db).
To import the knowledge source in mongoDB run:

wget http://dl.fbaipublicfiles.com/KILT/kilt_knowledgesource.json
mongoimport --db kilt --collection knowledgesource --file kilt_knowledgesource.json

Structure of each record

{
 'wikipedia_title': 'Email marketing',
 'wikipedia_id': 1101759, 
 'text': ['p1', 'p2',...., 'pn'], # list of paragraph text
 'anchors': [{"text":,"href":,"paragraph_id":,"start":,"end":} ]  , 
 'categories': 'comma separated list of categories'
 'history': # some info from wikipedia, including original url
 'wikidata_info': # wikidata info
 }

Query the knowledge source

from kilt.knowledge_source import KnowledgeSource

# get the knowledge souce
ks = KnowledgeSource()

# count entries - 5903530
ks.get_num_pages()

# get page by id
page = ks.get_page_by_id(27097632)

# get pages by title
page = ks.get_page_by_title("Michael Jordan")

KILT data

Examples:

download the data

mkdir data
python scripts/donwload_all_kilt_data.py
python scripts/get_triviaqa_input.py

You can also download and use the KILT data through the HuggingFace's nlp library.

Note that we release only the input for the test sets, without answers. Test answers are used for the KILT challenge on EvalAI where participants can upload their models’ predictions and be listed on the public leaderboard (there are strict submission limits to discourage overfitting on test data).

KILT data format

{'id': # original data point id if available otherwise unique id
 'input': # question / claim / sentence / etc
 'output': [ # each element might contain an answer, a provenance or both
    {
    'answer': # answer in textual form
    'provenance': [
        # evidence set for the answer from the KILT ks
        {
            'wikipedia_id':  # *mandatory* 
            'title': 
            'section': 
            'start_paragraph_id': 
            'start_character': 
            'end_paragraph_id':
            'end_character': 
            'bleu_score': # wrt original evidence
            'meta': # dataset/task specific
        }
        ] 
      }
    ]
 'meta': # dataset/task specific
 }

KILT data catalogue

dataset	task	train	dev	test
FEVER	Fact Checking	fever-train-kilt.jsonl (104,966 lines, 38.9MiB)	fever-dev-kilt.jsonl (10,444 lines, 6.17MiB)	fever-test_without_answers-kilt.jsonl (10,100 lines, 839kiB)
AIDA CoNLL-YAGO	Entity Linking	aidayago2-train-kilt.jsonl (18,395 lines, 70.1MiB)	aidayago2-dev-kilt.jsonl (4,784 lines, 21.1MiB)	aidayago2-test_without_answers-kilt.jsonl (4,463 lines, 14.4MiB)
WNED-WIKI	Entity Linking	-	wned-dev-kilt.jsonl (3,396 lines, 12.9MiB)	wned-test_without_answers-kilt.jsonl (3,376 lines, 13.3MiB)
WNED-CWEB	Entity Linking	-	cweb-dev-kilt.jsonl (5,599 lines, 90.2MiB)	cweb-test_without_answers-kilt.jsonl (5,543 lines, 100MiB)
T-REx	Slot Filling	trex-train-kilt.jsonl (2,284,168 lines, 1.75GiB)	trex-dev-kilt.jsonl (5,000 lines, 3.80MiB)	trex-test_without_answers-kilt.jsonl (5,000 lines, 896kiB)
Zero-Shot RE	Slot Filling	structured_zeroshot-train-kilt.jsonl (147,909 lines, 71.4MiB)	structured_zeroshot-dev-kilt.jsonl (3,724 lines, 2.27MiB)	structured_zeroshot-test_without_answers-kilt.jsonl (4,966 lines, 1.22MiB)
Natural Questions	Open Domain QA	nq-train-kilt.jsonl (87,372 lines, 51.9MiB)	nq-dev-kilt.jsonl (2,837 lines, 7.94MiB)	nq-test_without_answers-kilt.jsonl (1,444 lines, 334kiB)
HotpotQA	Open Domain QA	hotpotqa-train-kilt.jsonl (88,869 lines, 52.8MiB)	hotpotqa-dev-kilt.jsonl (5,600 lines, 3.97MiB)	hotpotqa-test_without_answers-kilt.jsonl (5,569 lines, 778kiB)
TriviaQA	Open Domain QA	triviaqa-train_id-kilt.jsonl* (61,844 lines, 102MiB)	triviaqa-dev_id-kilt.jsonl* (5,359 lines, 9.81MiB)	triviaqa-test_id_without_answers-kilt.jsonl* (6,586 lines, 123kiB)
ELI5	Open Domain QA	eli5-train-kilt.jsonl (272,634 lines, 548MiB)	eli5-dev-kilt.jsonl (1,507 lines, 14.1MiB)	eli5-test_without_answers-kilt.jsonl (600 lines, 99kiB)
Wizard of Wikipedia	Dialogue	wow-train-kilt.jsonl (94,577 lines, 71.9MiB)	wow-dev-kilt.jsonl (3,058 lines, 2.42MiB)	wow-test_without_answers-kilt.jsonl (2,944 lines, 1.29MiB)

^* run python scripts/get_triviaqa_input.py to get the question associated with each id

Additional data

For Entity Linking, in addition to the AIDA CoNLL-YAGO train set, the whole knowledge source can be used as training data by exploiting hyperlinks. To facilitate experimentation, we release such data in KILT format following the splits of BLINK:

• blink-train-kilt.jsonl (9M lines)
• blink-dev-kilt.jsonl (10,000 lines)

We also provide a script to map the TAC-KBP 2010 dataset to the knowledge source and format of KILT.

Run the retrieval evaluation

Please follow this README.

Troubleshooting

If the module cannot be found, preface the python command with PYTHONPATH=.

If the experiments fail on GPU memory allocation, try reducing batch size.

License

KILT is MIT licensed. See the LICENSE file for details.