m_lama / README.md
lhoestq's picture
lhoestq HF staff
add dataset_info in dataset metadata
a55b230
|
raw
history blame
7.11 kB
metadata
annotations_creators:
  - crowdsourced
  - expert-generated
  - machine-generated
language_creators:
  - crowdsourced
  - expert-generated
  - machine-generated
language:
  - af
  - ar
  - az
  - be
  - bg
  - bn
  - ca
  - ceb
  - cs
  - cy
  - da
  - de
  - el
  - en
  - es
  - et
  - eu
  - fa
  - fi
  - fr
  - ga
  - gl
  - he
  - hi
  - hr
  - hu
  - hy
  - id
  - it
  - ja
  - ka
  - ko
  - la
  - lt
  - lv
  - ms
  - nl
  - pl
  - pt
  - ro
  - ru
  - sk
  - sl
  - sq
  - sr
  - sv
  - ta
  - th
  - tr
  - uk
  - ur
  - vi
  - zh
license:
  - cc-by-nc-sa-4.0
multilinguality:
  - translation
size_categories:
  - 100K<n<1M
source_datasets:
  - extended|lama
task_categories:
  - question-answering
  - text-classification
task_ids:
  - open-domain-qa
  - text-scoring
paperswithcode_id: null
pretty_name: MLama
tags:
  - probing
dataset_info:
  features:
    - name: uuid
      dtype: string
    - name: lineid
      dtype: uint32
    - name: obj_uri
      dtype: string
    - name: obj_label
      dtype: string
    - name: sub_uri
      dtype: string
    - name: sub_label
      dtype: string
    - name: template
      dtype: string
    - name: language
      dtype: string
    - name: predicate_id
      dtype: string
  config_name: all
  splits:
    - name: test
      num_bytes: 125919995
      num_examples: 843143
  download_size: 40772287
  dataset_size: 125919995

Dataset Card for [Dataset Name]

Table of Contents

Dataset Description

Dataset Summary

This dataset provides the data for mLAMA, a multilingual version of LAMA. Regarding LAMA see https://github.com/facebookresearch/LAMA. For mLAMA the TREx and GoogleRE part of LAMA was considered and machine translated using Google Translate, and the Wikidata and Google Knowledge Graph API. The machine translated templates were checked for validity, i.e., whether they contain exactly one '[X]' and one '[Y]'.

This data can be used for creating fill-in-the-blank queries like "Paris is the capital of [MASK]" across 53 languages. For more details see the website http://cistern.cis.lmu.de/mlama/ or the github repo https://github.com/norakassner/mlama.

Supported Tasks and Leaderboards

Language model knowledge probing.

Languages

This dataset contains data in 53 languages: af,ar,az,be,bg,bn,ca,ceb,cs,cy,da,de,el,en,es,et,eu,fa,fi,fr,ga,gl,he,hi,hr,hu,hy,id,it,ja,ka,ko,la,lt,lv,ms,nl,pl,pt,ro,ru,sk,sl,sq,sr,sv,ta,th,tr,uk,ur,vi,zh

Dataset Structure

For each of the 53 languages and each of the 43 relations/predicates there is a set of triples.

Data Instances

For each language and relation there are triples, that consists of an object, a predicate and a subject. For each predicate there is a template available. An example for dataset["test"][0] is given here:

{
'language': 'af',
'lineid': 0, 
'obj_label': 'Frankryk', 
'obj_uri': 'Q142', 
'predicate_id': 'P1001', 
'sub_label': 'President van Frankryk', 
'sub_uri': 'Q191954', 
'template': "[X] is 'n wettige term in [Y].", 
'uuid': '3fe3d4da-9df9-45ba-8109-784ce5fba38a'
}

Data Fields

Each instance has the following fields

  • "uuid": a unique identifier
  • "lineid": a identifier unique to mlama
  • "obj_id": knowledge graph id of the object
  • "obj_label": surface form of the object
  • "sub_id": knowledge graph id of the subject
  • "sub_label": surface form of the subject
  • "template": template
  • "language": language code
  • "predicate_id": relation id

Data Splits

There is only one partition that is labelled as 'test data'.

Dataset Creation

Curation Rationale

The dataset was translated into 53 languages to investigate knowledge in pretrained language models multilingually.

Source Data

Initial Data Collection and Normalization

The data has several sources:

LAMA (https://github.com/facebookresearch/LAMA) licensed under Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) T-REx (https://hadyelsahar.github.io/t-rex/) licensed under Creative Commons Attribution-ShareAlike 4.0 International License Google-RE (https://github.com/google-research-datasets/relation-extraction-corpus) Wikidata (https://www.wikidata.org/) licensed under Creative Commons CC0 License and Creative Commons Attribution-ShareAlike License

Who are the source language producers?

See links above.

Annotations

Annotation process

Crowdsourced (wikidata) and machine translated.

Who are the annotators?

Unknown.

Personal and Sensitive Information

Names of (most likely) famous people who have entries in Google Knowledge Graph or Wikidata.

Considerations for Using the Data

Data was created through machine translation and automatic processes.

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

Not all triples are available in all languages.

Additional Information

Dataset Curators

The authors of the mLAMA paper and the authors of the original datasets.

Licensing Information

The Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). https://creativecommons.org/licenses/by-nc-sa/4.0/

Citation Information

@article{kassner2021multilingual,
  author    = {Nora Kassner and
               Philipp Dufter and
               Hinrich Sch{\"{u}}tze},
  title     = {Multilingual {LAMA:} Investigating Knowledge in Multilingual Pretrained
               Language Models},
  journal   = {CoRR},
  volume    = {abs/2102.00894},
  year      = {2021},
  url       = {https://arxiv.org/abs/2102.00894},
  archivePrefix = {arXiv},
  eprint    = {2102.00894},
  timestamp = {Tue, 09 Feb 2021 13:35:56 +0100},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2102-00894.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org},
  note      = {to appear in EACL2021}
}

Contributions

Thanks to @pdufter for adding this dataset.