README.md

metadata

license: mit
task_categories:
  - text-classification
language:
  - id
size_categories:
  - 10K<n<100K

Raw dataset for "Fact-Aware Fake-news Classification for Indonesian Language"

Disclaimer: Beta version, contains imbalanced representation of domain-specific NON-HOAX samples. We will release a new training and evaluation suite soon as a replacement of this dataset.

Data originates from https://turnbackhoax.id/ (Mafindo data 2018-2023);
The attributes of data are:

1. Label_id: Binary class labels ("HOAX"==1 ; "NON-HOAX"==0).
   
2. Label: Binary class labels ("HOAX" or "NON-HOAX").
   
3. Title: Claim or headline of news article.
   
4. Title_cleaned: Preprocessed claim, by removing tag label at the beginning of the sentence.
   
5. Content: the content of news article.
   
6. Fact: The summary of factual evidence that is either supporting or contradicting the correponding claim.
   
7. References: URL link of news article and the corresponding verdict or factual evidence as the justification of the news article.
   
8. Classification: Fine-grained classification labels for the news article:
   'CekFakta', 'Fabricated Content', 'False Connection', 'False Context', 'Impostor Content',
   'Manipulated Content', 'Misleading Content', 'Satire', 'nan'.
   
   

Example of usage:

>>> from datasets import load_dataset
>>> train_dataset = load_dataset(
...     "nlp-brin-id/fakenews-id-brin",
...     split="train",
...     keep_default_na=False,
... ).select_columns(['Label_id', 'Title_cleaned', 'Content', 'Fact'])