RabidUmarell/hungarian-toxic-comments

Hungarian Toxic Comments

The first openly available Hungarian dataset for toxic comment classification, introduced in:

Hatvani, P., & Yang, Z. Gy. (2025). Automated detection of toxic comments in Hungarian. Annales Mathematicae et Informaticae, 61, 108-117. DOI: 10.33039/ami.2025.10.007

Dataset Description

This dataset contains 654 manually annotated Hungarian-language comments collected from social media and political news forums. Each comment is annotated across five toxicity dimensions using a 1-5 Likert scale. The scores are averaged across three human annotators.

Sources

Comments were collected from four platforms to ensure topical diversity:

Source	Type
Reddit	Social media (offensive/casual)
napiszar.com	Shock/humor site
mandiner.hu	Hungarian political news forum
kuruc.info	Hungarian political news forum

Annotation Schema

Each comment is rated on a 1-5 Likert scale (1 = not at all, 5 = extremely) across five categories, following the Jigsaw Toxic Comment Classification taxonomy:

Column	Description
text	The original Hungarian comment
toxic	General toxicity (hostility, rudeness)
obscene	Crude or sexually explicit language
threat	Explicit or implied threats of harm
insult	Personal attacks or name-calling
identity_hate	Hate speech targeting identity groups (ethnicity, gender, sexuality, religion, etc.)

Scores are averages across three annotators (Randolph's kappa_free = 0.525, "slight agreement"), reflecting the inherent subjectivity of toxicity perception.

Dataset Statistics

• Total comments: 654
• Language: Hungarian (hu)
• Annotation scale: 1-5 Likert per dimension
• Number of annotators: 3
• Inter-annotator agreement: Randolph's kappa_free = 0.525

Example Entries

text	toxic	obscene	threat	insult	identity_hate
"Bárcsak minden buzit kirúgnának!!!!"	3.0	1.0	1.0	1.0	4.33
"Írjon neked a boncmester, jegyzőkönyvet."	2.0	1.0	4.0	1.0	1.0
"Attól hogy blokkolod, még mindig igaza van, és te továbbra is egy idióta vagy."	2.0	1.0	1.0	3.0	1.0

Intended Use

This dataset is intended for:

• Training and evaluating toxic content classifiers for Hungarian
• Benchmarking multilingual NLP models on low-resource languages
• Research into culturally specific patterns of online toxicity
• Development of content moderation tools for Hungarian platforms

Do not use this dataset to build systems that target, discriminate against, or harm individuals or groups.

Models Trained on This Dataset

Three BERT-based models were fine-tuned using this dataset (with data augmentation via typographical error simulation and masked token replacement):

Model	F1 Score
RabidUmarell/hubert-embedding-setfit-toxic	0.937
RabidUmarell/toxic-hubert	0.874
RabidUmarell/toxic-mbert	0.790

The best-performing model, huBERT-SetFit, uses sentence embeddings from huBERT with a logistic regression head trained via the SetFit framework - achieving strong results without requiring data augmentation.

A held-out test set is also available at RabidUmarell/hu-toxic-test-set.

Limitations

• Small size (654 comments) may limit generalizability to all domains of Hungarian online discourse
• Moderate inter-annotator agreement (kappa_free = 0.525) reflects the subjective nature of toxicity - annotations should be treated as soft labels
• Source bias: data skews toward politically charged content and shock-site comments; other domains (e.g. sports forums, e-commerce reviews) are not represented
• Temporal snapshot: collected comments reflect a specific period and may not capture evolving slang or new forms of toxic expression

Citation

If you use this dataset, please cite:

@article{hatvani2025toxic,
  title     = {Automated detection of toxic comments in Hungarian},
  author    = {Hatvani, P{\'e}ter and Yang, Zijian Gy{\H{o}}z{\H{o}}},
  journal   = {Annales Mathematicae et Informaticae},
  volume    = {61},
  pages     = {108--117},
  year      = {2025},
  doi       = {10.33039/ami.2025.10.007},
  url       = {https://ami.uni-eszterhazy.hu}
}

Authors

• Peter Hatvani - ELTE Research Centre for Linguistics; Pázmány Péter Catholic University, Doctoral School of Linguistics
• Zijian Gyozo Yang - ELTE Research Centre for Linguistics

License

This dataset is released for research use only. It may not be used for commercial purposes, to build production moderation systems without appropriate safeguards, or in any application that targets, discriminates against, or harms individuals or groups. Redistribution must retain this notice and cite the companion paper.