KoMultiText: Korean Multi-task Dataset for Classifying Biased Speech

• paper link

Dataset Summary

KoMultiText is a comprehensive Korean multi-task text dataset designed for classifying biased and harmful speech in online platforms. The dataset focuses on tasks such as Preference Detection, Profanity Identification, and Bias Classification across multiple domains, enabling state-of-the-art language models to perform multi-task learning for socially responsible AI applications.

Key Features

• Large-Scale Dataset: Contains 150,000 comments, including labeled and unlabeled data.
• Multi-task Annotations: Covers Preference, Profanity, and nine distinct types of Bias.
• Human-Labeled: All labeled data is annotated by five human experts to ensure high-quality and unbiased annotations.
• Real-world Relevance: Collected from "Real-time Best Gallery" of DC Inside, a popular online community in South Korea.

Labels

---

Dataset Creation

Source Data

• Origin: Comments collected from "Real-time Best Gallery" on DC Inside.
• Annotation Process:
  • Human Annotation: Five human annotators independently labeled all comments in the dataset to ensure accuracy and minimize bias.
  • Labeling Process: Annotators followed strict guidelines to classify comments into Preference, Profanity, and nine types of Bias. Discrepancies were resolved through majority voting and discussion.
• Dataset Composition:
  • Labeled Data: 40,361 comments (train/test split).
  • Unlabeled Data: 110,000 comments for potential pretraining or unsupervised learning. Veiw Dataset

How to Load the Dataset

from datasets import load_dataset

# Load the dataset 
dataset = load_dataset("Dasool/KoMultiText")

# Access train and test splits
train_dataset = dataset["train"]
test_dataset = dataset["test"]

Code

• Korean BERT-based fine-tuning code. Github

Citation

@misc{choi2023largescale,
      title={Large-Scale Korean Text Dataset for Classifying Biased Speech in Real-World Online Services}, 
      author={Dasol Choi and Jooyoung Song and Eunsun Lee and Jinwoo Seo and Heejune Park and Dongbin Na},
      year={2023},
      eprint={2310.04313},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Contact

• [email protected]