newsph_nli / README.md
albertvillanova's picture
Replace YAML keys from int to str (#2)
70f6170
|
raw
history blame
5.4 kB
metadata
annotations_creators:
  - machine-generated
language_creators:
  - found
language:
  - tl
license:
  - unknown
multilinguality:
  - monolingual
size_categories:
  - 100K<n<1M
source_datasets:
  - original
task_categories:
  - text-classification
task_ids:
  - natural-language-inference
paperswithcode_id: newsph-nli
pretty_name: NewsPH NLI
dataset_info:
  features:
    - name: premise
      dtype: string
    - name: hypothesis
      dtype: string
    - name: label
      dtype:
        class_label:
          names:
            '0': '0'
            '1': '1'
  splits:
    - name: train
      num_bytes: 154510599
      num_examples: 420000
    - name: test
      num_bytes: 3283665
      num_examples: 9000
    - name: validation
      num_bytes: 33015530
      num_examples: 90000
  download_size: 76565287
  dataset_size: 190809794

Dataset Card for NewsPH NLI

Table of Contents

Dataset Description

Dataset Summary

First benchmark dataset for sentence entailment in the low-resource Filipino language. Constructed through exploting the structure of news articles. Contains 600,000 premise-hypothesis pairs, in 70-15-15 split for training, validation, and testing.

Supported Tasks and Leaderboards

[More Information Needed]

Languages

The dataset contains news articles in Filipino (Tagalog) scraped rom all major Philippine news sites online.

Dataset Structure

Data Instances

Sample data: { "premise": "Alam ba ninyo ang ginawa ni Erap na noon ay lasing na lasing na rin?", "hypothesis": "Ininom niya ang alak na pinagpulbusan!", "label": "0" }

Data Fields

[More Information Needed]

Data Splits

Contains 600,000 premise-hypothesis pairs, in 70-15-15 split for training, validation, and testing.

Dataset Creation

Curation Rationale

We propose the use of news articles for automatically creating benchmark datasets for NLI because of two reasons. First, news articles commonly use single-sentence paragraphing, meaning every paragraph in a news article is limited to a single sentence. Second, straight news articles follow the “inverted pyramid” structure, where every succeeding paragraph builds upon the premise of those that came before it, with the most important information on top and the least important towards the end.

Source Data

Initial Data Collection and Normalization

To create the dataset, we scrape news articles from all major Philippine news sites online. We collect a total of 229,571 straight news articles, which we then lightly preprocess to remove extraneous unicode characters and correct minimal misspellings. No further preprocessing is done to preserve information in the data.

Who are the source language producers?

The dataset was created by Jan Christian, Blaise Cruz, Jose Kristian Resabal, James Lin, Dan John Velasco, and Charibeth Cheng from De La Salle University and the University of the Philippines

Annotations

Annotation process

[More Information Needed]

Who are the annotators?

Jan Christian Blaise Cruz, Jose Kristian Resabal, James Lin, Dan John Velasco and Charibeth Cheng

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[Jan Christian Blaise Cruz] (mailto:jan_christian_cruz@dlsu.edu.ph)

Licensing Information

[More Information Needed]

Citation Information

@article{cruz2020investigating, title={Investigating the True Performance of Transformers in Low-Resource Languages: A Case Study in Automatic Corpus Creation}, author={Jan Christian Blaise Cruz and Jose Kristian Resabal and James Lin and Dan John Velasco and Charibeth Cheng}, journal={arXiv preprint arXiv:2010.11574}, year={2020} }

Contributions

Thanks to @anaerobeth for adding this dataset.