metadata

license: cc-by-sa-4.0
task_categories:
  - text-classification
language:
  - cs
pretty_name: Czech SNLI

Dataset Card for Czech SNLI

Czech translation of the Stanford Natural Language Interface (SNLI) dataset with manual annotation of a SNLI subset. In addition to the entailment/contradiction/neutral inference, a "bad translation" class was added.

The annotation was done by students of NLP or computational linguistics. 1499 same pairs were annotated by two students to check IAA.

Dataset Details

The annotation for Czech premise-hypothesis pairs is done on 165390 pairs from train, test, and dev parts of the SNLI in the following distribution:

train: 159650
dev: 2860
test: 2880

The premise-hypothesis pairs were translated using the LINDAT Translation at https://lindat.mff.cuni.cz/services/translation/. The CUBBITT model was published as:

Popel, M., Tomkova, M., Tomek, J. et al. Transforming machine translation: a deep learning system reaches news translation quality comparable to human professionals. Nat Commun 11, 4381 (2020). https://doi.org/10.1038/s41467-020-18073-9

Annotation

From the 165390 pairs, 151470 (91.58%) were considered understandable (i.e., they were not marked as "bad translation" but the translation may not be accurate enough to preserve the entailment).

Inter-Annotator Agreement

Two random annotators obtained the same dataset. The kappa score is 0.67 (substantial agreement).

Confusion matrix

Full report on the agreement

     Simple Kappa Coefficient
              --------------------------------
              Kappa                     0.6757
              ASE                       0.0146
              95% Lower Conf Limit      0.6470
              95% Upper Conf Limit      0.7044

                 Test of H0: Simple Kappa = 0

              ASE under H0              0.0154
              Z                         43.9031
              One-sided Pr >  Z         0.0000
              Two-sided Pr > |Z|        0.0000

Dataset Formats

Dataset is available as TSV and JSONL.

The JSONL version only contains pairs that were not annotated as "bad translation". In case of multiple annotations, only the agreed pairs (where both annotators agreed) are selected. The JSONL contains 149660 sentence pairs.