license: mit
task_categories:
- text-classification
- token-classification
language:
- zh
- ja
- pt
- fr
- de
- ru
tags:
- toxicity
- hatespeech
pretty_name: Multi-Lingual Social Network Toxicity
size_categories:
- 100K<n<1M
MLSNT: Multi-Lingual Social Network Toxicity Dataset
MLSNT is a multi-lingual dataset for toxicity detection created through a large language model-assisted label transfer pipeline. It enables efficient and scalable moderation across languages and platforms, and is built to support span-level and category-specific classification for toxic content.
This dataset is introduced in the following paper:
Unified Game Moderation: Soft-Prompting and LLM-Assisted Label Transfer for Resource-Efficient Toxicity Detection
🏆 Accepted at KDD 2025, Applied Data Science Track
🧩 Overview
MLSNT harmonizes 15 publicly available toxicity datasets across 7 languages using GPT-4o-mini to create consistent binary and fine-grained labels. It is suitable for both training and evaluating toxicity classifiers in multi-lingual, real-world moderation systems.
🌍 Supported Languages
- 🇫🇷 French (
fr) - 🇩🇪 German (
de) - 🇵🇹 Portuguese (
pt) - 🇷🇺 Russian (
ru) - 🇨🇳 Simplified Chinese (
zh-cn) - 🇹🇼 Traditional Chinese (
zh-tw) - 🇯🇵 Japanese (
ja)
🏗️ Construction Method
Source Datasets
15 human-annotated datasets were gathered fromhatespeechdata.comand peer-reviewed publications.LLM-Assisted Label Transfer
GPT-4o-mini was prompted to re-annotate each instance into a unified label schema. Only examples where human and LLM annotations agreed were retained.Toxicity Categories
Labels are fine-grained categories (e.g.,threat,hate,harassment).
📊 Dataset Statistics
| Language | Total Samples | % Discarded | Toxic % (Processed) |
|---|---|---|---|
| German (HASOC, etc.) | ~13,800 | 28–69% | 32–56% |
| French (MLMA) | ~3,200 | 20% | 94% |
| Russian | ~14,300 | ~40% | 33–54% |
| Portuguese | ~21,000 | 20–44% | 26–50% |
| Japanese | ~2,000 | 10–25% | 17–45% |
| Chinese (Simplified) | ~34,000 | 29–46% | 48–61% |
| Chinese (Traditional) | ~65,000 | 37% | ~9% |
💾 Format
Each row in the dataset includes:
full_text: The original utterance or message.start_string_index: A list of start string indices (start positions of toxic spans).end_string_index: A list of end string indices (end positions of toxic spans).category_id: A list of toxic category IDs (integer values).final_label: A list of toxic category names (string values).min_category_id: The minimum toxic category ID in the row (used as the primary label).match_id: A unique identifier composed of the original dataset name and a row-level ID.
🗂️ Category ID Mapping
| ID | Friendly Name |
|---|---|
| 0 | Non Toxic |
| 1 | Threats (Life Threatening) |
| 2 | Minor Endangerment |
| 3 | Threats (Non-Life Threatening) |
| 4 | Hate |
| 5 | Sexual Content / Harassment |
| 6 | Extremism |
| 7 | Insults |
| 8 | Controversial / Potentially Toxic Topic |
🔬 Applications
- Fine-tuning multi-lingual moderation systems
- Cross-lingual toxicity benchmarking
- Training span-level and category-specific toxicity detectors
- Studying LLM label transfer reliability and agreement filtering
📜 Citation
If you use MLSNT in academic work, please cite:
@inproceedings{yang2025mlsnt,
title={Unified Game Moderation: Soft-Prompting and LLM-Assisted Label Transfer for Resource-Efficient Toxicity Detection},
author={Zachary Yang and Domenico Tullo and Reihaneh Rabbany},
booktitle={Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)},
year={2025}
}