The Expanded Amharic News Dataset is a large-scale, ethically collected corpus of Amharic-language news articles written in Geʽez (Fidel) script, designed to support research in Natural Language Processing (NLP).

This dataset builds upon the “An Amharic News Text Classification Dataset” developed by Israel Abebe Azime and Nebil Mohammed (arXiv link), which categorized Amharic news articles into multiple topical classes.
We extend their pioneering work by significantly expanding the data sources, time span, and overall size of the corpus.

Source and Collection

Original Dataset Foundation

The initial dataset contained Amharic news articles categorized into the following topics:

Local News
International News
Sports
Entertainment
Business
Politics

Expanded Sources

The authors curated and organized news articles from 11 well-established Amharic news outlets, covering a continuous publication period from 2011 to January 2021.

Newly Collected Data

To further enrich the dataset, we collected an additional set of Amharic news articles from diverse media outlets, spanning February 2021 to December 2024.
All newly added data was ethically collected, publicly available, and carefully processed to ensure quality and consistency.

Dataset Statistics

Language: Amharic
Script: Geʽez (Fidel)
Time span: 2011 – December 2024
Number of articles: Over 144,000 news articles
Domains:
- Local News
- International News
- Sports
- Entertainment
- Business
- Politics

Intended Uses

The dataset provides a longitudinal view of Amharic news reporting, making it suitable for tasks such as:

Topic classification
Text classification
Language modeling
Temporal and diachronic analysis
Transformer-based NLP benchmarking

Ethical Considerations

The dataset was collected ethically and responsibly, using only publicly accessible news content.
No intentional harm, misuse, or malicious purpose was involved in the data collection or preparation process.

Citation

Additionally, if you plan to use the dataset in your research, we kindly ask that you cite our work as it forms a key contribution to the dataset.

The citation for our paper is:

🔗 ACL Anthology link

@inproceedings{marilign-alemu-2025-amharic,
    title = "{A}mharic News Topic Classification: Dataset and Transformer-Based Model Benchmarks",
    author = "Marilign, Dagnachew Mekonnen and Alemu, Eyob Nigussie",
    editor = "Zhang, Chen and Allaway, Emily and Shen, Hua and Miculicich, Lesly and Li, Yinqiao and M'hamdi, Meryem and Limkonchotiwat, Peerat and Bai, Richard He and T.y.s.s., Santosh and Han, Sophia Simeng and Thapa, Surendrabikram and Rim, Wiem Ben",
    booktitle = "Proceedings of the 9th Widening NLP Workshop",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.winlp-main.23/",
    doi = "10.18653/v1/2025.winlp-main.23",
    pages = "130--135",
    ISBN = "979-8-89176-351-7",
    abstract = "News classification is a downstream task ..."
}

Acknowledgments

We sincerely acknowledge the foundational contributions of Israel Abebe Azime and Nebil Mohammed, whose original Amharic news dataset made this expanded work possible.

Collaboration

As part of our ongoing commitment to research in this field, we welcome academic and research collaborations that make use of this dataset. We believe that collaborative efforts can further enhance the impact, insights, and real-world applications of this resource.

Downloads last month: 62