Datasets:

wikis2t
/

COVERED-data

license: cc-by-sa-4.0
language:
  - en
task_categories:
  - text-generation
tags:
  - relation-extraction
  - wikidata
  - wikipedia
  - knowledge-graph
  - information-extraction
size_categories:
  - 1M<n<10M
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl
      - split: validation
        path: valid.jsonl
      - split: test
        path: test.jsonl

COVERED-data

COVERED aligns Wikipedia sentences with Wikidata triples for benchmarking generative relation extraction. It is built with the WikiS2T framework.

Statistics

Split	Sentences	Triples	Entities	Relations
all	2,382,570	4,863,548	307,146	1,139
train	1,909,789	3,898,094	245,716	1,106
valid	238,719	485,865	30,713	746
test	234,062	479,589	30,717	752

The split is done at the entity level and stratified by category (80/10/10, seed 42). All sentences of an entity are in the same split.

Categories

17 Wikidata categories (selected by P31 instance of): 12 general and 5 domain-specific, sorted by number of sentences.

Category	Group	Entities	Sentences	Triples	Sentences per entity
human	general	27,919	1,068,288	2,099,478	38.26
film	general	27,119	281,062	679,600	10.36
album	general	25,163	176,853	378,261	7.03
river	general	28,082	151,683	314,948	5.40
literary_work	general	26,037	126,154	286,555	4.85
business	general	24,951	120,977	200,681	4.85
building	general	17,921	74,550	124,268	4.16
taxon	domain-specific	24,434	69,757	137,885	2.85
organization	general	18,053	63,992	100,237	3.54
mountain	general	24,953	63,734	141,636	2.55
human_settlement	general	25,247	46,661	107,502	1.85
painting	general	9,057	42,498	112,950	4.69
street	general	4,114	29,171	52,359	7.09
chemical	domain-specific	8,523	24,386	39,200	2.86
gene	domain-specific	9,443	22,341	52,554	2.37
asteroid	domain-specific	4,183	13,391	23,310	3.20
star	domain-specific	1,947	7,072	12,124	3.63
All		307,146	2,382,570	4,863,548	7.76

Format

One JSON object per line, one sentence per object.

{
  "category": "album",
  "entity": "Q100234970_Utopia",
  "sentence_id": "1_0",
  "sentence": "Utopia is the fourth studio album by American rapper Travis Scott.",
  "triples": [
    {"hop": 1, "triple_index": 11, "subject_id": "Q100234970", "subject_label": "Utopia",
     "predicate_id": "P175", "predicate_label": "performer",
     "object_id": "Q13605596", "object_label": "Travis Scott"},
    {"hop": 2, "subject_id": "Q13605596", "subject_label": "Travis Scott",
     "predicate_id": "P106", "predicate_label": "occupation",
     "object_id": "Q2252262", "object_label": "rapper", "via_object_id": "Q13605596"}
  ]
}

Field	Meaning
`category`	One of the 17 categories above
`entity`	Source entity, `<QID>_<title>`
`sentence_id`	Sentence index in the entity's article
`sentence`	Wikipedia sentence
`triples[].hop`	1: subject is the source entity. 2: subject is an entity object of a 1-hop triple aligned to this sentence
`triples[].subject_id`, `predicate_id`, `object_id`	Wikidata ids
`triples[].*_label`	English labels
`triples[].object_id`	Empty string for literal objects (dates, quantities, text); the literal value is in `object_label`
`triples[].via_object_id`	2-hop only: the entity through which the triple is reached

License

The sentences come from Wikipedia and are released under CC BY-SA 4.0. Wikidata triples are CC0.