K-HATERS: A Hate Speech Detection Corpus in Korean with Target-Specific Ratings

This dataset introduces K-HATERS, the largest hate speech detection corpus in Korean, shared with our EMNLP'23 findings paper. [link]

GitHub: ssu-humane/K-HATERS

---

K-HATERS has been updated to K-HATERS 2.0! (2026-10-07)

K-HATERS 2.0 adds fine-grained target labels to the corpus. While K-HATERS provides the target category of a hateful comment (e.g., Gender), it does not provide the target itself (e.g., Female, Male, LGBTQ). K-HATERS 2.0 adds a new field target that fills this gap. The comments, labels, target categories, rationales, and the data split remain the same as K-HATERS.

What's new in 2.0

• New field target: the fine-grained target for each target category of a comment.
• The target_label value region was renamed to region_race.

How fine-grained targets were annotated

• Fine-grained targets were annotated by Gemini-2.5-Flash given the comment and its target category, and then manually reviewed and normalized.

The following table shows the fine-grained targets in each target category.

Target category	Target
Political	Progressive, Conservative
Region & Race	China, Japan, Korea, North Korea, USA, Afghanistan, Korean-Chinese, Black, White
Job	Journalist, Public Official, Legal Professional, Politician, Athlete, Entertainer, Educator, Religious Worker, Healthcare Worker
Gender	Female, Male, LGBTQ
Age	Elderly, Young Adult, Minor, Middle-Aged
Disabled	Mental Disability, Developmental Disability, Physical Disability
Religion	Islam, Protestantism, Buddhism, Catholicism

---

Data collection and annotation

• We collected news comments published through the politics, society and world news sections in Naver News over two months in 2021.
• All comments were annotated through CashMission, a crowdsourcing service run by SELECTSTAR.
• Labeling guidelines are available as a part of SELECTSTAR open datasets (in Korean). [link]
• [2.0] Fine-grained targets were annotated by Gemini-2.5-Flash given the comment and its target category, and then manually reviewed and normalized.

Data distribution

We provide the split data of 172,158/10,000/10,000 for training, validation, and test purposes, ensuring the label distribution is maintained.

Data description

>>> data['train'][42]
{"id": 42, "text": "군대도 안간 놈 이 주둥아리 는 씽씽하네..보수 놈 들..군대는 안가고 애국이냐..#@이름#,#@이름#,", "label": "L1_hate", "target_label": ["political"], "offensiveness_rationale": [[7, 8], [11, 15], [27, 28]], "target_rationale": [[24, 26], [46, 51], [52, 57]], "target": [{"category": "political", "fine": "conservative"}]}

• Abusive language categories (label)
  • L2_hate: Comments with explicit forms of hate expressions toward one of the groups of protected attributes (e.g., gender, age, race, ...)
  • L1_hate: Comments with more implicit forms of hate expressions
  • Offensive: Comments that express offensiveness but not toward a protected attribute group
  • Normal: The rest comments
• Multi-label target categories (target_label): list of offensiveness targets. A comment can have zero or multiple targets.
  • List of target categories: gender, age, region & race, religion, politics, job, disability, individuals, and others.
• Annotators' rationales for the strength of ratings (offensiveness_rationale): lists providing annotators' rationales for the strength of ratings. The list includes the start and end indices of highlight spans.
• Annotators' rationales for the target of offensiveness (target_rationale)
• [2.0] Fine-grained targets (target): list of the specific targets that the hate is directed at, with one entry per target category.
  • category: corresponds to target_label.
  • fine: the fine-grained target within that category (e.g., conservative for political). It is null when the fine-grained target cannot be determined from the comment.

How to use it

from datasets import load_dataset
 
data = load_dataset('humane-lab/K-HATERS')
train = data['train']
valid = data['validation']
test =  data['test']

Where are the ratings?

The original dataset was labeled by annotators for thirteen rating variables on a 3-point Likert scale, each indicating the offensiveness strength toward a target. In this paper, we proposed a label transformation scheme leading to 4-class abusive language categories and multi-label target labels for efficient modeling. The above code is to access the transformed dataset. Please refer to the details of the transformation method in our paper.

If you want to use the dataset with rating variables, please refer to the following code.

from datasets import load_dataset
 
data = load_dataset('humane-lab/K-HATERS-Ratings')

The rating distribution is as follows:

Acknowledgements

The dataset construction was supported by DATUMO (SELECTSTAR) through the "2022 AI Dataset Supporting Business" program.

Dataset usage

This dataset is shared under CC-BY 4.0. According to this license, you are free to use the dataset as long as you provide appropriate attribution (e.g., citing our paper).

@article{park2023haters,
  title={K-HATERS: A Hate Speech Detection Corpus in Korean with Target-Specific Ratings},
  author={Park, Chaewon and Kim, Suhwan and Park, Kyubyong and Park, Kunwoo},
  journal={Findings of the EMNLP 2023},
  year={2023}
}