Exqrch/IndoDiscourse

Notice: We added new data and restructured the dataset on 31st October 2024 (GMT+7)
Changes:
- Group unique texts together
- The annotators of a text are now set as a list of annotator_id. Each respective column is a list of the same size of annotators_id.
- Added Polarized column

Notice 2: We rename the dataset from IndoToxic2024 to IndoDiscourse

A Multi-Labeled Dataset for Indonesian Discourse: Examining Toxicity, Polarization, and Demographics Information

Dataset Overview

IndoToxic2024 is a multi-labeled dataset designed to analyze online discourse in Indonesia, focusing on toxicity, polarization, and annotator demographic information. This dataset provides insights into the growing political and social divisions in Indonesia, particularly in the context of the 2024 presidential election. Unlike previous datasets, IndoToxic2024 offers a multi-label annotation framework, enabling nuanced research on the interplay between toxicity and polarization.

Dataset Statistics

• Total annotated texts: 28,477
• Platforms: X (formerly Twitter), Facebook, Instagram, and news articles
• Timeframe: September 2023 – January 2024
• Annotators: 29 individuals from diverse demographic backgrounds

Label Distribution

Label	Count
Toxic	2,156 (balanced)
Non-Toxic	6,468 (balanced)
Polarized	3,811 (balanced)
Non-Polarized	11,433 (balanced)

Dataset Structure

The dataset consists of texts labeled for toxicity and polarization, along with annotator demographics. Each text is annotated by at least one coder, with 44.6% of texts receiving multiple annotations. Annotations were aggregated using majority voting, excluding texts with perfect disagreement.

Features:

• text: The Indonesian social media or news text
• toxicity: List of toxicity annotations (1 = Toxic, 0 = Non-Toxic)
• polarization: List of polarization annotations (1 = Polarized, 0 = Non-Polarized)
• annotators_id: List of annotator_id that annotate the text (anonymized) -- Refer to annotator subset for each annotator_id's demographic informatino

Baseline Model Performance

Key Results:

We benchmarked IndoToxic2024 using BERT-based models and large language models (LLMs). The results indicate that:

• BERT-based models outperform LLMs, with IndoBERTweet achieving the highest accuracy.
• Polarization detection is harder than toxicity detection, as evidenced by lower recall scores.
• Demographic information improves classification, especially for polarization detection.

Additional Findings:

• Polarization and toxicity are correlated: Using polarization as a feature improves toxicity detection, and vice versa.
• Demographic-aware models perform better for polarization detection: Including coder demographics boosts classification performance.
• Wisdom of the crowd: Texts labeled by multiple annotators lead to higher recall in toxicity detection.

Ethical Considerations

• Data Privacy: All annotator demographic data is anonymized.
• Use Case: This dataset is released for research purposes only and should not be used for surveillance or profiling.

Citation

If you use IndoToxic2024, please cite:

@misc{susanto2025multilabeleddatasetindonesiandiscourse,
      title={A Multi-Labeled Dataset for Indonesian Discourse: Examining Toxicity, Polarization, and Demographics Information}, 
      author={Lucky Susanto and Musa Wijanarko and Prasetia Pratama and Zilu Tang and Fariz Akyas and Traci Hong and Ika Idris and Alham Aji and Derry Wijaya},
      year={2025},
      eprint={2503.00417},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.00417}, 
}```