Dataset Card for JSICK

Dataset Summary

From official GitHub:

Japanese Sentences Involving Compositional Knowledge (JSICK) Dataset.

JSICK is the Japanese NLI and STS dataset by manually translating the English dataset SICK (Marelli et al., 2014) into Japanese. We hope that our dataset will be useful in research for realizing more advanced models that are capable of appropriately performing multilingual compositional inference.

JSICK-stress Test set

The JSICK-stress test set is a dataset to investigate whether models capture word order and case particles in Japanese. The JSICK-stress test set is provided by transforming syntactic structures of sentence pairs in JSICK, where we analyze whether models are attentive to word order and case particles to predict entailment labels and similarity scores.

The JSICK test set contains 1666, 797, and 1006 sentence pairs (A, B) whose premise sentences A (the column sentence_A_Ja_origin) include the basic word order involving ga-o (nominative-accusative), ga-ni (nominative-dative), and ga-de (nominative-instrumental/locative) relations, respectively.

We provide the JSICK-stress test set by transforming syntactic structures of these pairs by the following three ways:

• scrum_ga_o: a scrambled pair, where the word order of premise sentences A is scrambled into o-ga, ni-ga, and de-ga order, respectively.
• ex_ga_o: a rephrased pair, where the only case particles (ga, o, ni, de) in the premise A are swapped
• del_ga_o: a rephrased pair, where the only case particles (ga, o, ni) in the premise A are deleted

Languages

The language data in JSICK is in Japanese and English.

Dataset Structure

Data Instances

When loading a specific configuration, users has to append a version dependent suffix:

import datasets as ds

dataset: ds.DatasetDict = ds.load_dataset("hpprc/jsick")
print(dataset)
# DatasetDict({
#     train: Dataset({
#         features: ['id', 'premise', 'hypothesis', 'label', 'score', 'premise_en', 'hypothesis_en', 'label_en', 'score_en', 'corr_entailment_labelAB_En', 'corr_entailment_labelBA_En', 'image_ID', 'original_caption', 'semtag_short', 'semtag_long'],
#         num_rows: 4500
#     })
#     test: Dataset({
#         features: ['id', 'premise', 'hypothesis', 'label', 'score', 'premise_en', 'hypothesis_en', 'label_en', 'score_en', 'corr_entailment_labelAB_En', 'corr_entailment_labelBA_En', 'image_ID', 'original_caption', 'semtag_short', 'semtag_long'],
#         num_rows: 4927
#     })
# })

dataset: ds.DatasetDict = ds.load_dataset("hpprc/jsick", name="stress")
print(dataset)
# DatasetDict({
#     test: Dataset({
#         features: ['id', 'premise', 'hypothesis', 'label', 'score', 'sentence_A_Ja_origin', 'entailment_label_origin', 'relatedness_score_Ja_origin', 'rephrase_type', 'case_particles'],
#         num_rows: 900
#     })
# })

base

An example of looks as follows:

{
    'id': 1,
    'premise': '子供たちのグループが庭で遊んでいて、後ろの方には年を取った男性が立っている',
    'hypothesis': '庭にいる男の子たちのグループが遊んでいて、男性が後ろの方に立っている',
    'label': 1, // (neutral)
    'score': 3.700000047683716,
    'premise_en': 'A group of kids is playing in a yard and an old man is standing in the background',
    'hypothesis_en': 'A group of boys in a yard is playing and a man is standing in the background',
    'label_en': 1, // (neutral)
    'score_en': 4.5,
    'corr_entailment_labelAB_En': 'nan',
    'corr_entailment_labelBA_En': 'nan',
    'image_ID': '3155657768_b83a7831e5.jpg',
    'original_caption': 'A group of children playing in a yard , a man in the background .',
    'semtag_short': 'nan',
    'semtag_long': 'nan',
}

stress

An example of looks as follows:

{
    'id': '5818_de_d',
    'premise': '女性火の近くダンスをしている',
    'hypothesis': '火の近くでダンスをしている女性は一人もいない',
    'label': 2,  // (contradiction)
    'score': 4.0,
    'sentence_A_Ja_origin': '女性が火の近くでダンスをしている',
    'entailment_label_origin': 2,
    'relatedness_score_Ja_origin': 3.700000047683716,
    'rephrase_type': 'd',
    'case_particles': 'de'
}

Data Fields

base

A version adopting the column names of a typical NLI dataset.

Name	Description
id	The ids (the same with original SICK).
premise	The first sentence in Japanese.
hypothesis	The second sentence in Japanese.
label	The entailment label in Japanese.
score	The relatedness score in the range [1-5] in Japanese.
premise_en	The first sentence in English.
hypothesis_en	The second sentence in English.
label_en	The original entailment label in English.
score_en	The original relatedness score in the range [1-5] in English.
semtag_short	The linguistic phenomena tags in Japanese.
semtag_long	The details of linguistic phenomena tags in Japanese.
image_ID	The original image in 8K ImageFlickr dataset.
original_caption	The original caption in 8K ImageFlickr dataset.
corr_entailment_labelAB_En	The corrected entailment label from A to B in English by (Karouli et al., 2017).
corr_entailment_labelBA_En	The corrected entailment label from B to A in English by (Karouli et al., 2017).

stress

Name	Description
id	Ids (the same with original SICK).
premise	The first sentence in Japanese.
hypothesis	The second sentence in Japanese.
label	The entailment label in Japanese
score	The relatedness score in the range [1-5] in Japanese.
sentence_A_Ja_origin	The original premise sentences A from the JSICK test set.
entailment_label_origin	The original entailment labels.
relatedness_score_Ja_origin	The original relatedness scores.
rephrase_type	The type of transformation applied to the syntactic structures of the sentence pairs.
case_particles	The grammatical particles in Japanese that indicate the function or role of a noun in a sentence.

Data Splits

name	train	validation	test
base	4,500		4,927
original	4,500		4,927
stress			900
stress-original			900

Annotations

To annotate the JSICK dataset, they used the crowdsourcing platform "Lancers" to re-annotate entailment labels and similarity scores for JSICK. They had six native Japanese speakers as annotators, who were randomly selected from the platform. The annotators were asked to fully understand the guidelines and provide the same labels as gold labels for ten test questions.

For entailment labels, they adopted annotations that were agreed upon by a majority vote as gold labels and checked whether the majority judgment vote was semantically valid for each example. For similarity scores, they used the average of the annotation results as gold scores. The raw annotations with the JSICK dataset are publicly available. The average annotation time was 1 minute per pair, and Krippendorff's alpha for the entailment labels was 0.65.

Additional Information

• verypluming/JSICK
• Compositional Evaluation on Japanese Textual Entailment and Similarity
• JSICK: 日本語構成的推論・類似度データセットの構築

Licensing Information

CC BY-SA 4.0

Citation Information

@article{yanaka-mineshima-2022-compositional,
    title = "Compositional Evaluation on {J}apanese Textual Entailment and Similarity",
    author = "Yanaka, Hitomi  and
      Mineshima, Koji",
    journal = "Transactions of the Association for Computational Linguistics",
    volume = "10",
    year = "2022",
    address = "Cambridge, MA",
    publisher = "MIT Press",
    url = "https://aclanthology.org/2022.tacl-1.73",
    doi = "10.1162/tacl_a_00518",
    pages = "1266--1284",
}

@article{谷中 瞳2021,
  title={JSICK: 日本語構成的推論・類似度データセットの構築},
  author={谷中 瞳 and 峯島 宏次},
  journal={人工知能学会全国大会論文集},
  volume={JSAI2021},
  number={ },
  pages={4J3GS6f02-4J3GS6f02},
  year={2021},
  doi={10.11517/pjsai.JSAI2021.0_4J3GS6f02}
}

Contributions

Thanks to Hitomi Yanaka and Koji Mineshima for creating this dataset.