metadata

size_categories:
  - 1M<n<10M
language:
  - de
  - en
  - es
  - fr
multilinguality:
  - multilingual
task_categories:
  - token-classification
configs:
  - config_name: German
    data_files:
      - split: train
        path: data/DiS-ReX/german.txt
  - config_name: English
    data_files:
      - split: train
        path: data/DiS-ReX/english.txt
  - config_name: French
    data_files:
      - split: train
        path: data/DiS-ReX/french.txt
  - config_name: Spanish
    data_files:
      - split: train
        path: data/DiS-ReX/spanish.txt

Dataset origin: https://github.com/dair-iitd/DiS-ReX

DiS-ReX: A Multilingual Dataset for Distantly Supervised Relation Extraction

DiS-ReX, a multilingual dataset for distantly supervised relation extraction. The dataset has over 1.5 million instances, spanning 4 languages (English, Spanish, German and French). Our dataset has 36 positive relation types + 1 no relation (NA) class.

Format

The dataset folder has 5 text files

english.txt
german.txt
french.txt
spanish.txt
rel2id.txt

For files named <language>.txt, each line is a unique instance represented as a Python dictionary. An example is shown below:

{"token": ["at", "the", "58th", "annual", "grammy", "awards", "in", "february", "the", "eagles", "joined", "by", "leadon", "touring", "guitarist", "steuart", "smith", "and", "co-writer", "jackson", "browne", "performed", "\"take", "it", "easy\"", "in", "honor", "of", "frey"], "h": {"name": "steuart smith", "id": "Q3498822", "pos": [15, 17]}, "t": {"name": "eagles", "id": "Q2092297", "pos": [9, 10]}, "relation": "http://dbpedia.org/ontology/associatedBand"}

Here the keys and values have the following meaning:

token: A list representing the context sentence. Every element in the list represents a word.
h: A dictionary for head entity. has the following keys:
- name: name of the head
- entityid: wikidata id for the entity
- pos: a tuple of the form [start index, end index] according to head entity's positition in the token list
t: A dictionary for tail entity. has the following keys:
- name: name of the tail
- entityid: wikidata id for the entity
- pos: a tuple of the form [start index, end index] according to tail entity's positition in the token list
relation: relation for the tuple (head entity, tail entity)

The dataset format is same as presented in OpenNRE. For a bag with more than one possible relations, the instances are repeated with a different value for the relation key. An example is shown below:

{"token": ["huxley", "who", "had", "twice", "visited", "the", "soviet", "union", "was", "originally", "not", "anti-communist", "but", "the", "ruthless", "adoption", "of", "lysenkoism", "by", "joseph", "stalin", "ended", "his", "tolerant", "attitude"], "h": {"name": "joseph stalin", "id": "Q855", "pos": [19, 21]}, "t": {"name": "the soviet union", "id": "Q15180", "pos": [5, 8]}, "relation": "http://dbpedia.org/ontology/country"}
{"token": ["huxley", "who", "had", "twice", "visited", "the", "soviet", "union", "was", "originally", "not", "anti-communist", "but", "the", "ruthless", "adoption", "of", "lysenkoism", "by", "joseph", "stalin", "ended", "his", "tolerant", "attitude"], "h": {"name": "joseph stalin", "id": "Q855", "pos": [19, 21]}, "t": {"name": "the soviet union", "id": "Q15180", "pos": [5, 8]}, "relation": "http://dbpedia.org/ontology/deathPlace"}

The file named rel2id.txt contains relation types and the corresponding indices we use during training our model.

Cite

The dataset is a part of the pre-print DiS-ReX: A Multilingual Dataset for Distantly Supervised Relation Extraction. We also release our baseline results using mBERT+Bag Attention and present it in our paper. If you use or extend our work, please cite the following paper:

@misc{bhartiya2021disrex,
      title={DiS-ReX: A Multilingual Dataset for Distantly Supervised Relation Extraction}, 
      author={Abhyuday Bhartiya and Kartikeya Badola and Mausam},
      year={2021},
      eprint={2104.08655},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}