A Multilingual and Multi-Target Dataset for Hate Speech and Target Detection on Telegram
MuLTa-Telegram is a curated dataset comprising about 4,000 manually annotated messages collected from public Telegram channels in Italian and Polish. It is designed to support research in hate speech detection, with a particular focus on:
- The presence or absence of hate speech
- Identification of specific target groups of hate
- Identification of the target groups in each message’s content (mentioned group)
The dataset is described in the following paper:
MuLTa-Telegram: A Fine-Grained Italian and Polish Dataset for Hate Speech and Target Detection
The repository includes:
- The full annotated dataset (JSON format)
- Annotation guidelines
| Language | Total Messages | Hate Speech | Non-Hate |
|---|---|---|---|
| Italian | 2,002 | 411 (20.5%) | 1,591 |
| Polish | 1,934 | 257 (12.9%) | 1,684 |
Annotations cover 9 primary identity groups, including:
- Ethnicity/Origin (People of Color, Romani, Other)
- LGBT+
- Women
- Religious groups (Jewish, Muslim, Christian)
- People with disabilities
- Other or No Target (as applicable)
Each message is also labeled for mentioned target group independently of whether it is hateful.
- Messages were retrieved using a snowball sampling approach applied to public Telegram channels known for high toxicity or disinformation.
- Pre-selection of messages used a keyword-based filtering matrix covering all target categories.
- Annotations were performed by expert native speakers following a structured guideline developed in collaboration with civil society organizations.
- All data were anonymized in accordance with applicable privacy regulations.
Additional models, data, and documentation developed during the HATEDEMICS project are available in the HATEDEMICS GitHub repository.
The dataset is publicly available under the Creative Commons Attribution Non-Commercial (CC BY NC).
You are free to share and adapt the material for research purposes, provided appropriate credit is given.
This work was supported by the European Union’s CERV fund under Grant Agreement No. 101143249 (HATEDEMICS) and the Horizon Europe research and innovation programme under Grant Agreement No. 101135437 (AI-CODE)