README.md

metadata

annotations_creators:
  - machine-generated
  - manual-partial-validation
language_creators:
  - expert-generated
language:
  - id
license: unknown
multilinguality:
  - monolingual
size_categories:
  - 100K<n<1M
source_datasets:
  - SQuAD-ID
task_categories:
  - text-classification
task_ids:
  - natural-language-inference
pretty_name: SQuAD-ID-NLI
dataset_info:
  features:
    - name: premise
      dtype: string
    - name: hypothesis
      dtype: string
    - name: label
      dtype:
        class_label:
          names:
            '0': entailment
            '1': neutral
            '2': contradiction
  config_name: squadid-nli
  splits:
    - name: train
      num_bytes: 103934750
      num_examples: 236890
    - name: validation
      num_bytes: 10831375
      num_examples: 23748
    - name: test
      num_bytes: 10969750
      num_examples: 23746
  download_size: 125735875
  dataset_size: 284384

Dataset Card for SQuAD-ID-NLI

Table of Contents

• Dataset Description
  • Dataset Summary
  • Supported Tasks and Leaderboards
  • Languages
• Dataset Structure
  • Data Instances
  • Data Fields
  • Data Splits
• Dataset Creation
  • Curation Rationale
  • Source Data
  • Annotations
  • Personal and Sensitive Information
• Considerations for Using the Data
  • Social Impact of Dataset
  • Discussion of Biases
  • Other Known Limitations
• Additional Information
  • Dataset Curators
  • Licensing Information
  • Citation Information
  • Contributions

Dataset Description

• Repository: Hugging Face
• Point of Contact: Hugging Face
• Experiment: Github

Dataset Summary

The SQuAD-ID-NLI dataset is derived from the SQuAD-ID question answering dataset, utilizing named entity recognition (NER), chunking tags, Regex, and embedding similarity techniques to determine its contradiction sets. Collected through this process, the dataset comprises various columns beyond premise, hypothesis, and label, including properties aligned with NER and chunking tags. This dataset is designed to facilitate Natural Language Inference (NLI) tasks and contains information extracted from diverse sources to provide comprehensive coverage. Each data instance encapsulates premise, hypothesis, label, and additional properties pertinent to NLI evaluation.

Supported Tasks and Leaderboards

• Natural Language Inference for Indonesian

Languages

Indonesian

Dataset Structure

Data Instances

An example of test looks as follows.

{
  "premise": "Beberapa keluarga Yunani Bizantium berasal dari tentara bayaran Norman selama periode Restorasi Comnenian, ketika kaisar Bizantium mencari prajurit Eropa Barat. Raoulii adalah keturunan dari orang Italia-Norman bernama Raoul, Petraliphae adalah keturunan dari Pierre d'Aulps, dan kelompok klan Albania yang dikenal sebagai Maniakate diturunkan dari Normandia yang bertugas di bawah George Maniaces dalam ekspedisi Sisilia tahun 1038.", 
  "hypothesis": "Dari mana beberapa famili tentara bayaran Norman berasal? Yunani Bizantium", 
  "label": 0
}

Data Fields

The data fields are:

• premise: a string feature
• hypothesis: a string feature
• label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).

Data Splits #TODO

The data is split across train, valid, and test.

split	# examples
train	236890
valid	23748
test	23746

Dataset Creation

Curation Rationale

Indonesian NLP is considered under-resourced. We need NLI dataset to fine-tuning the NLI model to utilizing them for QA models in order to improving the performance of the QA's.

Source Data

Initial Data Collection and Normalization

We collect the data from the prominent QA dataset in Indonesian. The annotation fully by the original dataset's researcher.

Who are the source language producers?

This synthetic data was produced by machine, but the original data was produced by human.

Personal and Sensitive Information

There might be some personal information coming from Wikipedia and news, especially the information of famous/important people.

Considerations for Using the Data

Discussion of Biases

The QA dataset (so the NLI-derived from them) is created using premise sentences taken from Wikipedia and news. These data sources may contain some bias.

Other Known Limitations

No other known limitations

Additional Information

Dataset Curators

This dataset is the result of the collaborative work of Indonesian researchers from the University of Indonesia, Mohamed bin Zayed University of Artificial Intelligence, and the Korea Advanced Institute of Science & Technology.

Licensing Information

The license is Unknown. Please contact authors for any information on the dataset.