Datasets:
The viewer is disabled because this dataset repo requires arbitrary Python code execution. Please consider
removing the
loading script
and relying on
automated data support
(you can use
convert_to_parquet
from the datasets
library). If this is not possible, please
open a discussion
for direct help.
Dataset Card for K-MHaS
Sample Code
Dataset Summary
The Korean Multi-label Hate Speech Dataset, K-MHaS, consists of 109,692 utterances from Korean online news comments, labelled with 8 fine-grained hate speech classes (labels: Politics
, Origin
, Physical
, Age
, Gender
, Religion
, Race
, Profanity
) or Not Hate Speech
class. Each utterance provides from a single to four labels that can handles Korean language patterns effectively. For more details, please refer to our paper about K-MHaS, published at COLING 2022.
Supported Tasks and Leaderboards
Hate Speech Detection
binary classification
(labels:Hate Speech
,Not Hate Speech
)multi-label classification
: (labels:Politics
,Origin
,Physical
,Age
,Gender
,Religion
,Race
,Profanity
,Not Hate Speech
)
For the multi-label classification, a Hate Speech
class from the binary classification, is broken down into eight classes, associated with the hate speech category. In order to reflect the social and historical context, we select the eight hate speech classes. For example, the Politics
class is chosen, due to a significant influence on the style of Korean hate speech.
Languages
Korean
Dataset Structure
Data Instances
The dataset is provided with train/validation/test set in the txt format. Each instance is a news comment with a corresponding one or more hate speech classes (labels: Politics
, Origin
, Physical
, Age
, Gender
, Religion
, Race
, Profanity
) or Not Hate Speech
class. The label numbers matching in both English and Korean is in the data fields section.
{'text':'์๊ผดํ๋ฑ์ํค๋ค์ด ๋ค ๋์ ธ์ผ ๋๋ผ๊ฐ ๋๋ฐ๋ก ๋ ๊ฒ๊ฐ๋ค..๋ต์ด ์๋ ์ข
์๋คใ
'
'label': [2, 3, 4]
}
Data Fields
text
: utterance from Korean online news comment.label
: the label numbers matching with 8 fine-grained hate speech classes andnot hate speech
class are follows.0
:Origin
(์ถ์ ์ฐจ๋ณ
) hate speech based on place of origin or identity;1
:Physical
(์ธ๋ชจ์ฐจ๋ณ
) hate speech based on physical appearance (e.g. body, face) or disability;2
:Politics
(์ ์น์ฑํฅ์ฐจ๋ณ
) hate speech based on political stance;3
:Profanity
(ํ์ค์์ค
) hate speech in the form of swearing, cursing, cussing, obscene words, or expletives; or an unspecified hate speech category;4
:Age
(์ฐ๋ น์ฐจ๋ณ
) hate speech based on age;5
:Gender
(์ฑ์ฐจ๋ณ
) hate speech based on gender or sexual orientation (e.g. woman, homosexual);6
:Race
(์ธ์ข ์ฐจ๋ณ
) hate speech based on ethnicity;7
:Religion
(์ข ๊ต์ฐจ๋ณ
) hate speech based on religion;8
:Not Hate Speech
(ํด๋น์ฌํญ์์
).
Data Splits
In our repository, we provide splitted datasets that have 78,977(train) / 8,776 (validation) / 21,939 (test) samples, preserving the class proportion.
Dataset Creation
Curation Rationale
We propose K-MHaS, a large size Korean multi-label hate speech detection dataset that represents Korean language patterns effectively. Most datasets in hate speech research are annotated using a single label classification of particular aspects, even though the subjectivity of hate speech cannot be explained with a mutually exclusive annotation scheme. We propose a multi-label hate speech annotation scheme that allows overlapping labels associated with the subjectivity and the intersectionality of hate speech.
Source Data
Initial Data Collection and Normalization
Our dataset is based on the Korean online news comments available on Kaggle and Github. The unlabeled raw data was collected between January 2018 and June 2020. Please see the details in our paper K-MHaS published at COLING2020.
Who are the source language producers?
The language producers are users who left the comments on the Korean online news platform between 2018 and 2020.
Annotations
Annotation process
We begin with the common categories of hate speech found in literature and match the keywords for each category. After the preliminary round, we investigate the results to merge or remove labels in order to provide the most representative subtype labels of hate speech contextual to the cultural background. Our annotation instructions explain a twolayered annotation to (a) distinguish hate and not hate speech, and (b) the categories of hate speech. Annotators are requested to consider given keywords or alternatives of each category within social, cultural, and historical circumstances. For more details, please refer to the paper K-MHaS.
Who are the annotators?
Five native speakers were recruited for manual annotation in both the preliminary and main rounds.
Personal and Sensitive Information
This datasets contains examples of hateful language, however, has no personal information.
Considerations for Using the Data
Social Impact of Dataset
We propose K-MHaS, a new large-sized dataset for Korean hate speech detection with a multi-label annotation scheme. We provided extensive baseline experiment results, presenting the usability of a dataset to detect Korean language patterns in hate speech.
Discussion of Biases
All annotators were recruited from a crowdsourcing platform. They were informed about hate speech before handling the data. Our instructions allowed them to feel free to leave if they were uncomfortable with the content. With respect to the potential risks, we note that the subjectivity of human annotation would impact on the quality of the dataset.
Other Known Limitations
[More Information Needed]
Additional Information
Dataset Curators
This dataset is curated by Taejun Lim, Heejun Lee and Bogeun Jo.
Licensing Information
Creative Commons Attribution-ShareAlike 4.0 International (cc-by-sa-4.0).
Citation Information
@inproceedings{lee-etal-2022-k,
title = "K-{MH}a{S}: A Multi-label Hate Speech Detection Dataset in {K}orean Online News Comment",
author = "Lee, Jean and
Lim, Taejun and
Lee, Heejun and
Jo, Bogeun and
Kim, Yangsok and
Yoon, Heegeun and
Han, Soyeon Caren",
booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
month = oct,
year = "2022",
address = "Gyeongju, Republic of Korea",
publisher = "International Committee on Computational Linguistics",
url = "https://aclanthology.org/2022.coling-1.311",
pages = "3530--3538",
abstract = "Online hate speech detection has become an important issue due to the growth of online content, but resources in languages other than English are extremely limited. We introduce K-MHaS, a new multi-label dataset for hate speech detection that effectively handles Korean language patterns. The dataset consists of 109k utterances from news comments and provides a multi-label classification using 1 to 4 labels, and handles subjectivity and intersectionality. We evaluate strong baselines on K-MHaS. KR-BERT with a sub-character tokenizer outperforms others, recognizing decomposed characters in each hate speech class.",
}
Contributions
The contributors of the work are:
- Jean Lee (The University of Sydney)
- Taejun Lim (The University of Sydney)
- Heejun Lee (BigWave AI)
- Bogeun Jo (BigWave AI)
- Yangsok Kim (Keimyung University)
- Heegeun Yoon (National Information Society Agency)
- Soyeon Caren Han (The University of Western Australia and The University of Sydney)
- Downloads last month
- 33