Dataset Card for JaNLI

Dataset Summary

The JaNLI (Japanese Adversarial NLI) dataset, inspired by the English HANS dataset, is designed to necessitate an understanding of Japanese linguistic phenomena and to illuminate the vulnerabilities of models.

Languages

The language data in JaNLI is in Japanese (BCP-47 ja-JP).

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/janli")
print(dataset)
# DatasetDict({
#     train: Dataset({
#         features: ['id', 'premise', 'hypothesis', 'label', 'heuristics', 'number_of_NPs', 'semtag'],
#         num_rows: 13680
#     })
#     test: Dataset({
#         features: ['id', 'premise', 'hypothesis', 'label', 'heuristics', 'number_of_NPs', 'semtag'],
#         num_rows: 720
#     })
# })

dataset: ds.DatasetDict = ds.load_dataset("hpprc/janli", name="original")
print(dataset)
# DatasetDict({
#     train: Dataset({
#         features: ['id', 'sentence_A_Ja', 'sentence_B_Ja', 'entailment_label_Ja', 'heuristics', 'number_of_NPs', 'semtag'],
#         num_rows: 13680
#     })
#     test: Dataset({
#         features: ['id', 'sentence_A_Ja', 'sentence_B_Ja', 'entailment_label_Ja', 'heuristics', 'number_of_NPs', 'semtag'],
#         num_rows: 720
#     })
# })

base

An example of looks as follows:

{
  'id': 12,
  'premise': '若者がフットボール選手を見ている',
  'hypothesis': 'フットボール選手を若者が見ている',
  'label': 0,
  'heuristics': 'overlap-full',
  'number_of_NPs': 2,
  'semtag': 'scrambling'
}

original

An example of looks as follows:

{
  'id': 12,
  'sentence_A_Ja': '若者がフットボール選手を見ている',
  'sentence_B_Ja': 'フットボール選手を若者が見ている',
  'entailment_label_Ja': 0,
  'heuristics': 'overlap-full',
  'number_of_NPs': 2,
  'semtag': 'scrambling'
}

Data Fields

base

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

Name	Description
id	The number of the sentence pair.
premise	The premise (sentence_A_Ja).
hypothesis	The hypothesis (sentence_B_Ja).
label	The correct label for the sentence pair (either entailment or non-entailment); in the setting described in the paper, non-entailment = neutral + contradiction (entailment_label_Ja).
heuristics	The heuristics (structural pattern) tag. The tags are: subsequence, constituent, full-overlap, order-subset, and mixed-subset.
number_of_NPs	The number of noun phrase in a sentence.
semtag	The linguistic phenomena tag.

original

The original version retaining the unaltered column names.

Name	Description
id	The number of the sentence pair.
sentence_A_Ja	The premise.
sentence_B_Ja	The hypothesis.
entailment_label_Ja	The correct label for this sentence pair (either entailment or non-entailment); in the setting described in the paper, non-entailment = neutral + contradiction
heuristics	The heuristics (structural pattern) tag. The tags are: subsequence, constituent, full-overlap, order-subset, and mixed-subset.
number_of_NPs	The number of noun phrase in a sentence.
semtag	The linguistic phenomena tag.

Data Splits

name	train	validation	test
base	13,680		720
original	13,680		720

Annotations

The annotation process for this Japanese NLI dataset involves tagging each pair (P, H) of a premise and hypothesis with a label for structural pattern and linguistic phenomenon. The structural relationship between premise and hypothesis sentences is classified into five patterns, with each pattern associated with a type of heuristic that can lead to incorrect predictions of the entailment relation. Additionally, 11 categories of Japanese linguistic phenomena and constructions are focused on for generating the five patterns of adversarial inferences.

For each linguistic phenomenon, a template for the premise sentence P is fixed, and multiple templates for hypothesis sentences H are created. In total, 144 templates for (P, H) pairs are produced. Each pair of premise and hypothesis sentences is tagged with an entailment label (entailment or non-entailment), a structural pattern, and a linguistic phenomenon label.

The JaNLI dataset is generated by instantiating each template 100 times, resulting in a total of 14,400 examples. The same number of entailment and non-entailment examples are generated for each phenomenon. The structural patterns are annotated with the templates for each linguistic phenomenon, and the ratio of entailment and non-entailment examples is not necessarily 1:1 for each pattern. The dataset uses a total of 158 words (nouns and verbs), which occur more than 20 times in the JSICK and JSNLI datasets.

Additional Information

• verypluming/JaNLI
• Assessing the Generalization Capacity of Pre-trained Language Models through Japanese Adversarial Natural Language Inference

Licensing Information

CC BY-SA 4.0

Citation Information

@InProceedings{yanaka-EtAl:2021:blackbox,
  author    = {Yanaka, Hitomi and Mineshima, Koji},
  title     = {Assessing the Generalization Capacity of Pre-trained Language Models through Japanese Adversarial Natural Language Inference},
  booktitle = {Proceedings of the 2021 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP (BlackboxNLP2021)},
  url       = {https://aclanthology.org/2021.blackboxnlp-1.26/},
  year      = {2021},
}

Contributions

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