jamil226/turkish_news_dataset

Turkish News Summarization Master Corpus (TNSMC)

📋 Overview

The Turkish News Summarization Master Corpus (TNSMC) is a professionally curated dataset comprising over 15,000 news articles, specifically designed for Natural Language Processing (NLP) research. This dataset serves as a specialized benchmark for training and evaluating Abstractive Text Summarization models in the Turkish language.

By addressing the scarcity of large-scale Turkish summarization datasets, TNSMC provides a robust foundation for fine-tuning state-of-the-art Transformer-based architectures such as mBART, T5, and various BERT-based Seq2Seq models.

📊 Dataset Composition & Provenance

This master corpus was synthesized from high-quality data collected from three major Turkish news outlets. To ensure research integrity, a source column has been included to allow for provenance tracking and source-specific linguistic analysis.

Primary Source	Internal Identifier	Article Count	Feature Set
Sözcü	msozcu	10,000	title, body, source
TRT Haber	trthaber	5,000	title, body, source
Milliyet	milliyet	186	title, body, source
Aggregate Total		15,186

🚀 Potential Use Cases

This corpus is engineered for high versatility across several NLP disciplines:

1. Abstractive Summarization: Developing models that generate human-like, semantic summaries (title) from complex news bodies (body).
2. Editorial Style Analysis: Comparative studies on how different journalistic entities frame identical events.
3. Cross-Domain Generalization: Evaluating how a model trained on one news source performs on another (e.g., public vs. private media).
4. Morphological Analysis: Given Turkish is an agglutinative language, this corpus provides ample data for sub-word tokenization and morphological research.
5. Zero-Shot Evaluation: Benchmarking the linguistic capabilities of large multilingual models on authentic Turkish prose.

🛠 Technical Specifications

• Language: Turkish (tr-TR)
• Data Format: Tabular CSV
• Data Curation: The raw content has been processed to minimize noise, removing HTML artifacts and irrelevant metadata to ensure high-fidelity training signals.
• Academic Intent: This data was collected and curated exclusively for educational and research purposes.

⚖️ Disclaimer & Ethics

The intellectual property rights of the original articles belong to their respective publishers: TRT Haber, Sözcü, and Milliyet. This dataset is shared under a non-commercial license for academic exploration. Users are expected to respect the fair use principles of the original content creators.

🧬 How to Initialize

You can integrate this corpus into your Python workflow using the datasets library:

from datasets import load_dataset

# Load the master corpus from the Hugging Face Hub
corpus = load_dataset("jamil226/turkish_news_dataset")

# Sample data inspection
sample = corpus['train'][0]
print(f"Source: {sample['source']}")
print(f"Summary: {sample['title']}")