ivykopal/anli_sk

Dataset Card for ANLI-SK

Dataset Summary

ANLI-SK is a Slovak-language version of the Adversarial NLI (ANLI) dataset, created by machine-translating the original English ANLI dataset into Slovak using Google Translate. The dataset follows the same three-way natural language inference (NLI) format as the original: each example consists of a premise, a hypothesis, and a label indicating whether the hypothesis is an entailment, contradiction, or neutral with respect to the premise.

ANLI-SK was created for and used in evaluating cross-lingual transfer methods for low-resource languages, as described in Vykopal et al. (2025), "Soft Language Prompts for Language Transfer" (NAACL 2025).

• Language: Slovak (sk)
• Task: Natural Language Inference (3-way classification: entailment / neutral / contradiction)
• Source: Machine-translated from ANLI (Nie et al., 2020) via Google Translate
• Total examples: 103,000 (train: 100,000 / validation: 1,200 / test: 1,200)

Dataset Structure

Data Instances

Each instance contains:

{
  "id": "3b1e7f2a-...",
  "premise": "Slovak-language premise text.",
  "hypothesis": "Slovak-language hypothesis text.",
  "label": 0
}

Data Fields

• id (string): Unique identifier for the example (inherited from the original ANLI example).
• premise (string): The premise text, translated into Slovak.
• hypothesis (string): The hypothesis text, translated into Slovak.
• label (int, 0-2): NLI label — 0 = entailment, 1 = neutral, 2 = contradiction.

Data Splits

Split	Examples
train	100,000
validation	1,200
test	1,200

Dataset Creation

Source Data

ANLI-SK is derived from the original Adversarial NLI (ANLI) dataset (Nie et al., 2020), which was collected via a human-and-model-in-the-loop adversarial process on top of Wikipedia-derived premises.

Translation Process

All premise and hypothesis fields were machine-translated from English to Slovak using the Google Translate API. No manual post-editing or human verification of translation quality was performed; as such, the dataset may contain translation artifacts, and downstream users should be aware of this limitation when interpreting model performance.

Labels were kept unchanged from the original English dataset, under the assumption that translation preserves the entailment relationship between premise and hypothesis.

Considerations for Using the Data

Limitations

• Translations were produced automatically (Google Translate) without human review, and may contain grammatical errors, mistranslations, or altered nuances that could affect the validity of the NLI label for a given pair.
• As a machine-translated resource, ANLI-SK should be treated as a proxy for a natively Slovak NLI dataset rather than a gold-standard resource for evaluating Slovak language understanding.

Intended Use

This dataset was created to support research on cross-lingual transfer and multilingual NLP for low-resource languages, particularly Slovak. It is suitable for training and evaluating NLI models in Slovak and for studying the effects of translation-based data augmentation for low-resource languages.

Citation

If you use this dataset, please cite both the paper that introduces this Slovak translation and the original ANLI paper.

This work:

@inproceedings{vykopal-etal-2025-soft,
    title = "Soft Language Prompts for Language Transfer",
    author = "Vykopal, Ivan  and
      Ostermann, Simon  and
      Simko, Marian",
    editor = "Chiruzzo, Luis  and
      Ritter, Alan  and
      Wang, Lu",
    booktitle = "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
    month = apr,
    year = "2025",
    address = "Albuquerque, New Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.naacl-long.517/",
    doi = "10.18653/v1/2025.naacl-long.517",
    pages = "10294--10313",
    ISBN = "979-8-89176-189-6",
}

Original ANLI dataset:

@inproceedings{nie-etal-2020-adversarial,
    title = "Adversarial {NLI}: A New Benchmark for Natural Language Understanding",
    author = "Nie, Yixin  and
      Williams, Adina  and
      Dinan, Emily  and
      Bansal, Mohit  and
      Weston, Jason  and
      Kiela, Douwe",
    editor = "Jurafsky, Dan  and
      Chai, Joyce  and
      Schluter, Natalie  and
      Tetreault, Joel",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2020.acl-main.441/",
    doi = "10.18653/v1/2020.acl-main.441",
    pages = "4885--4901",
}

Dataset Curator

Ivan Vykopal (ivykopal)

Licensing Information

This dataset is derived from ANLI, which is released under a CC BY-NC 4.0 license. This translated version, ANLI-SK, inherits the same license.