srbNLI / README.md
MilosKosRad's picture
Update README.md
4026d58 verified
metadata
dataset_info:
  features:
    - name: claim
      dtype: string
    - name: evidence
      dtype: string
    - name: label
      dtype: int64
  splits:
    - name: train
      num_bytes: 1703425
      num_examples: 971
    - name: valid
      num_bytes: 199433
      num_examples: 120
    - name: test
      num_bytes: 218204
      num_examples: 122
  download_size: 905297
  dataset_size: 2121062
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: valid
        path: data/valid-*
      - split: test
        path: data/test-*
language:
  - sr
tags:
  - Textual Entailment
  - NLI

Dataset Card for srbNLI

Dataset Details

Dataset Description

English:
The srbNLI dataset is an automatically translated version of the SciFact dataset (VerifAI version) into Serbian using the Google Translate API. It is designed for Natural Language Inference (NLI) tasks, where the goal is to predict the relationship between a premise and a hypothesis. The relationship could be classified as one of 3 labels - support, contradict, or no evidence. This dataset enables research in cross-lingual NLI tasks for Serbian.

Serbian:
srbNLI skup podataka predstavlja automatski prevedenu verziju SciFact skupa podataka (VerifAI verzija) na srpski jezik koristeći Google Translate API. Namenjen je zadacima iz oblasti Inferencije u prirodnom jeziku (NLI), gde je cilj predvideti odnos između premise i hipoteze. Ovaj odnos se može klasifikovati u jedan od tri označena: podrška, kontradikcija ili nema dokaza. Ovaj skup podataka omogućava istraživanja u oblasti međujezičke NLI zadatke na srpskom jeziku.

Uses

Direct Use

The dataset is intended for use in Natural Language Inference (NLI) tasks, specifically for Serbian language models. It can be used to train and evaluate NLI models on Serbian texts and in cross-lingual NLI tasks.

Out-of-Scope Use

This dataset is not suitable for tasks outside of NLI, such as machine translation or sentiment analysis, due to its specific design for inference tasks.

Dataset Structure

  • claim: The premise text in Serbian.
  • evidence: The hypothesis text in Serbian.
  • label: The relationship between the premise and the hypothesis (0 = support, 1 = contradict, 2 = no evidence).

Dataset Creation

Curation Rationale

The srbNLI dataset was created to enable research in NLI tasks for the Serbian language, leveraging the SciFact dataset for its content and Google Translate API for translation.

Source Data

The source data is the SciFact dataset (VerifAI version), which contains English claims and corresponding evidence sentences. These were translated into Serbian using the Google Translate API.

Bias, Risks, and Limitations

As the dataset was created through automatic translation, it may contain errors or inconsistencies arising from the translation process. The quality of the translations may vary, and the dataset may contain biases inherent in the Google Translate API. Researchers should be cautious when using it for tasks that require high-quality translations.