[Paper Review] KOLD: Korean Offensive Language Dataset

Younghoon Jeong, Juhyun Oh+5 more|arXiv (Cornell University)|May 23, 2022

Hate Speech and Cyberbullying Detection4 citations

TL;DR

This paper introduces KOLD, the first large-scale, hierarchically annotated Korean offensive language dataset with 40,429 comments from news and YouTube, featuring offensive type, target, and target group classification, along with span-level annotations for offensiveness and targets. The study demonstrates that providing context (article/video titles) significantly improves model performance—especially in target group classification (+13.1 F1)—and highlights cultural differences in offensive language distribution compared to English datasets.

ABSTRACT

Recent directions for offensive language detection are hierarchical modeling, identifying the type and the target of offensive language, and interpretability with offensive span annotation and prediction. These improvements are focused on English and do not transfer well to other languages because of cultural and linguistic differences. In this paper, we present the Korean Offensive Language Dataset (KOLD) comprising 40,429 comments, which are annotated hierarchically with the type and the target of offensive language, accompanied by annotations of the corresponding text spans. We collect the comments from NAVER news and YouTube platform and provide the titles of the articles and videos as the context information for the annotation process. We use these annotated comments as training data for Korean BERT and RoBERTa models and find that they are effective at offensiveness detection, target classification, and target span detection while having room for improvement for target group classification and offensive span detection. We discover that the target group distribution differs drastically from the existing English datasets, and observe that providing the context information improves the model performance in offensiveness detection (+0.3), target classification (+1.5), and target group classification (+13.1). We publicly release the dataset and baseline models.

Motivation & Objective

• Address the lack of comprehensive, culturally grounded offensive language datasets in Korean, a language with distinct linguistic and cultural characteristics.
• Develop a hierarchical annotation framework that captures offensive language type, target (individual/group), and target group identity, enabling fine-grained analysis.
• Improve model interpretability by annotating offensive spans and target spans in comments for explainability.
• Investigate the impact of contextual information (e.g., article or video titles) on offensive language detection performance in low-resource, non-English settings.
• Provide a publicly released dataset and baseline models to support future research in multilingual offensive language detection and mitigate bias in NLP systems.

Proposed method

• Collected 40,429 offensive comments from NAVER news and YouTube, with associated article or video titles as contextual metadata.
• Designed a hierarchical annotation scheme labeling: (1) offensiveness (OFF/NOT), (2) target type (IND/GRP/OTH/UNT), (3) target group (e.g., race, gender, religion), and (4) span-level annotations for offensive and target terms.
• Employed a multi-annotator labeling process with 3,124 annotators, using strict guidelines and iterative updates to ensure consistency and handle ambiguous cases.
• Fine-tuned Korean BERT and RoBERTa models on the KOLD dataset for joint offensiveness detection, target classification, and span detection tasks.
• Evaluated model performance with and without context (title) to assess its impact on classification and span detection tasks.
• Conducted ablation studies to compare target group distribution in KOLD with existing English datasets, revealing significant cultural and linguistic differences.

Experimental results

Research questions

• RQ1 How does the distribution of offensive language targets and groups in Korean differ from that in existing English offensive language datasets?
• RQ2 To what extent does providing contextual information (e.g., article or video titles) improve the performance of offensive language detection models in Korean?
• RQ3 Can hierarchical annotation with span-level labeling enhance model interpretability and performance in offensive language detection tasks?
• RQ4 How effective are fine-tuned Korean BERT and RoBERTa models on multi-task learning for offensiveness detection, target classification, and span detection using KOLD?
• RQ5 What are the limitations of current multilingual models when applied to Korean offensive language, and how can they be addressed through language-specific datasets?

Key findings

• Providing context (title) improved offensiveness detection by +0.3 F1, target classification by +1.5 F1, and target group classification by +13.1 F1, demonstrating the critical role of context in low-resource, non-English settings.
• The target group distribution in KOLD differs significantly from existing English datasets, with higher proportions of offensive content targeting racial/ethnic and religious groups, highlighting cultural specificity in offensive language.
• The KOLD dataset enables effective fine-tuning of Korean BERT and RoBERTa models, achieving strong performance on offensiveness detection and target classification, though room for improvement remains in target group and span detection.
• Annotators frequently omitted target terms in offensive comments, indicating that models must rely on context to infer targets, which underscores the importance of contextual information.
• The study reveals that multilingual models trained on English data may generalize poorly to Korean due to linguistic and cultural differences, emphasizing the need for language-specific datasets.
• Despite high annotation cost and complexity, the hierarchical, context-aware annotation framework in KOLD sets a new standard for explainable and culturally grounded offensive language detection in non-English languages.