SumO-Ro

A large-scale Romanian news corpus for three information-condensation tasks: summarization, headline generation, and keyword extraction.

Contents

• 171,475 unique news articles from 20 Romanian publications
• 514,425 task-formatted records (each article appears once per task)
• Splits: 168,475 train / 1,500 validation / 1,500 test articles

Structure

Each record has three fields:

• input_text — the article body, prefixed with a task marker (<TITLE>, <SUMMARY>, <KEYWORDS>)
• target_text — the expected output (headline, summary, or keyword list)
• task — one of title, summary, keywords

Tasks

Summarization and headline generation are abstractive. Keyword extraction is extractive: every reference keyword occurs in the corresponding article body, verified across the full corpus.

Construction

Raw collection (250,763 pages) → filtering, deduplication, quality control → 182,839 articles → keyword-extractivity filter (articles whose keywords do not occur in the body were removed, −11,364) → 171,475.

An earlier release under this name contained the 182,839-article version, before the extractivity filter. The current version supersedes it.

Intended use

Academic, non-commercial research. The data was collected from publications that consented to academic use of their articles.

Limitations

• Domain: news only — does not represent conversational, legal, medical, or literary Romanian
• Category and source distributions reflect real-world availability and are not balanced
• Summaries, headlines, and keywords are editorial fields from the source publications; style and quality vary across outlets

Citation

A paper describing this resource is in preparation. Please cite this repository until it appears.

License

CC BY-NC-SA 4.0 — attribution, non-commercial, share-alike.

Contact

corcodelgabrielaaa@gmail.com