Hugging Face's logo

Datasets:
demelin
/
wino_x

Tasks:
Translation
Sub-tasks:
multiple-choice-qa
language-modeling
Languages:
English
German
French
Multilinguality:
multilingual
translation
Size Categories:
1K<n<10K
Language Creators:
machine-generated
expert-generated
Annotations Creators:
no-annotation
Source Datasets:
original
Tags:
License:
Dataset card Viewer Files Community
1
wino_x / README.md
demelin
Updated the dataset card.
79a0451
preview code
|
raw history blame contribute delete
No virus
8.96 kB
metadata
annotations_creators:
  - no-annotation
language:
  - en
  - de
  - fr
  - ru
language_creators:
  - machine-generated
  - expert-generated
license:
  - mit
multilinguality:
  - multilingual
  - translation
pretty_name: Wino-X
size_categories:
  - 1K<n<10K
source_datasets:
  - original
task_categories:
  - translation
  - coreference resolution
  - commonsense reasoning
task_ids:
  - multiple-choice-qa
  - language-modeling

Dataset Card for Wino-X

Table of Contents

Dataset Description

Dataset Summary

Wino-X is a parallel dataset of German, French, and Russian Winograd schemas, aligned with their English counterparts, used to examine whether neural machine translation models can perform coreference resolution that requires commonsense knowledge, and whether multilingual language models are capable of commonsense reasoning across multiple languages.

Supported Tasks and Leaderboards

  • translation: The dataset can be used to evaluate translations of ambiguous source sentences, as produced by translation models . A pretrained transformer-based NMT model can be used for this purpose.
  • coreference-resolution: The dataset can be used to rank alternative translations of an ambiguous source sentence that differ in the chosen referent of an ambiguous source pronoun. A pretrained transformer-based NMT model can be used for this purpose.
  • commonsense-reasoning: The dataset can also be used evaluate whether pretrained multilingual language models can perform commonsense reasoning in (or across) multiple languages by identifying the correct filler in a cloze completion task. An XLM-based model can be used for this purpose.

Languages

The dataset (both its MT and LM portions) is available in the following translation pairs: English-German, English-French, English-Russian. All English sentences included in Wino-X were extracted from publicly available parallel corpora, as detailed in the accompanying paper, and represent the dataset-specific language varieties. All non-English sentences were obtained through machine translation and may, as such, exhibit features of translationese.

Dataset Structure

Data Instances

The following represents a typical MT-Wino-X instance (for the English-German translation pair):

{"qID": "3UDTAB6HH8D37OQL3O6F3GXEEOF09Z-1",
"sentence": "The woman looked for a different vase for the bouquet because it was too small.",
"translation1": "Die Frau suchte nach einer anderen Vase für den Blumenstrauß, weil sie zu klein war.",
"translation2": "Die Frau suchte nach einer anderen Vase für den Blumenstrauß, weil er zu klein war.",
"answer": 1,
"pronoun1": "sie",
"pronoun2": "er",
"referent1_en": "vase",
"referent2_en": "bouquet",
"true_translation_referent_of_pronoun1_de": "Vase",
"true_translation_referent_of_pronoun2_de": "Blumenstrauß",
"false_translation_referent_of_pronoun1_de": "Vase",
"false_translation_referent_of_pronoun2_de": "Blumenstrauß"}

The following represents a typical LM-Wino-X instance (for the English-French translation pair):

{"qID": "3UDTAB6HH8D37OQL3O6F3GXEEOF09Z-1",
"sentence": "The woman looked for a different vase for the bouquet because it was too small.",
"context_en": "The woman looked for a different vase for the bouquet because _ was too small.",
"context_fr": "La femme a cherché un vase différent pour le bouquet car _ était trop petit.",
"option1_en": "the bouquet",
"option2_en": "the vase",
"option1_fr": "le bouquet",
"option2_fr": "le vase",
"answer": 2,
"context_referent_of_option1_fr": "bouquet",
"context_referent_of_option2_fr": "vase"}

Data Fields

For MT-Wino-X:

  • "qID": Unique identifier ID for this dataset instance.
  • "sentence": English sentence containing the ambiguous pronoun 'it'.
  • "translation1": First translation candidate.
  • "translation2": Second translation candidate.
  • "answer": ID of the correct translation.
  • "pronoun1": Translation of the ambiguous source pronoun in translation1.
  • "pronoun2": Translation of the ambiguous source pronoun in translation2.
  • "referent1_en": English referent of the translation of the ambiguous source pronoun in translation1.
  • "referent2_en": English referent of the translation of the ambiguous source pronoun in translation2.
  • "true_translation_referent_of_pronoun1_[TGT-LANG]": Target language referent of pronoun1 in the correct translation.
  • "true_translation_referent_of_pronoun2_[TGT-LANG]": Target language referent of pronoun2 in the correct translation.
  • "false_translation_referent_of_pronoun1_[TGT-LANG]": Target language referent of pronoun1 in the incorrect translation.
  • "false_translation_referent_of_pronoun2_[TGT-LANG]": Target language referent of pronoun2 in the incorrect translation.

For LM-Wino-X:

  • "qID": Unique identifier ID for this dataset instance.
  • "sentence": English sentence containing the ambiguous pronoun 'it'.
  • "context_en": Same English sentence, where 'it' is replaced by a gap.
  • "context_fr": Target language translation of the English sentence, where the translation of 'it' is replaced by a gap.
  • "option1_en": First filler option for the English sentence.
  • "option2_en": Second filler option for the English sentence.
  • "option1_[TGT-LANG]": First filler option for the target language sentence.
  • "option2_[TGT-LANG]": Second filler option for the target language sentence.
  • "answer": ID of the correct gap filler.
  • "context_referent_of_option1_[TGT-LANG]": English translation of option1_[TGT-LANG].
  • "context_referent_of_option2_[TGT-LANG]": English translation of option2_[TGT-LANG]

Data Splits

Wno-X was designed as an evaluation-only benchmark and therefore is intended to be used for zero-shot testing only. However, users are very welcome to split the data as they wish :) .

Dataset Creation

Curation Rationale

Please refer to Section 2 in the dataset paper.

Source Data

Initial Data Collection and Normalization

Please refer to Section 2 in the dataset paper.

Who are the source language producers?

Please refer to Section 2 in the dataset paper.

Annotations

Annotation process

Please refer to Section 2 in the dataset paper.

Who are the annotators?

Annotations were generated automatically and verified by the dataset author / curator for correctness.

Personal and Sensitive Information

[N/A]

Considerations for Using the Data

Social Impact of Dataset

Please refer to 'Ethical Considerations' in the dataset paper.

Discussion of Biases

Please refer to 'Ethical Considerations' in the dataset paper.

Other Known Limitations

Please refer to 'Ethical Considerations' in the dataset paper.

Additional Information

Dataset Curators

Denis Emelin

Licensing Information

MIT

Citation Information

@inproceedings{Emelin2021WinoXMW, title={Wino-X: Multilingual Winograd Schemas for Commonsense Reasoning and Coreference Resolution}, author={Denis Emelin and Rico Sennrich}, booktitle={EMNLP}, year={2021} }