ahmetyaylalioglu/turkish-news-summarization-collection

Turkish News Summarization Dataset (RSS-Crawled)

This dataset contains Turkish news articles collected from publicly available RSS feeds and the corresponding web pages. It is designed to support both abstractive and extractive summarization experiments, as well as comparative studies between these paradigms.

Motivation

Many Turkish summarization resources are either small, domain-specific, or focused on a single summarization style. This dataset aims to provide:

• Full news article text for input,
• Multiple “summary-like” targets derived from editorial metadata,
• Sentence-level pseudo-labels to enable extractive baselines and supervised extractive training.

Data Collection

1. RSS ingestion: A large list of Turkish news RSS feeds is parsed to collect article URLs and RSS-provided descriptions (when available).
2. Web crawling: Each article URL is fetched and the main article text is extracted using a robust HTML-to-text pipeline.
3. Metadata extraction:
   • summary_title is taken from the RSS item title.
   • short_summary is taken from the page meta description (og:description, description, or twitter:description) if available; otherwise it falls back to the RSS description when present.
4. Filtering: Non-article links (e.g., video/gallery) and pages with too little extracted text are skipped.

Note: The dataset may contain different writing styles (breaking news, short reports, long-form articles), depending on sources and feed categories.

Recommended Targets

The dataset provides two summary fields:

• summary_title: Title-only summary (very short).
• short_summary: Short editorial description / meta summary (typically 1–2 sentences; may be empty for some sources).

For abstractive summarization training, a common choice is:

• target_summary = short_summary if short_summary != "" else summary_title

This helps reduce the “title-only target” effect (overly short targets that may encourage copying behavior).

Extractive Pseudo-Labels

To support extractive summarization baselines and training, the dataset includes:

• Sentence segmentation in sentences
• Binary labels ext_labels indicating which sentences are selected
• ext_selected: concatenation of selected sentences

Pseudo-labeling approach (high level):

• Each sentence is scored against the chosen reference summary (prefer short_summary, else summary_title) using a ROUGE-1 F1 overlap heuristic.
• Top-k sentences are selected as extractive targets.

These pseudo-labels are automatically generated and should be treated as weak supervision.

Dataset Format

Each example is stored as a JSON object (JSONL: one JSON per line) with the following schema:

Core fields

• id (string): deterministic ID derived from URL
• url (string): article URL
• feed (string): RSS feed URL
• source_domain (string): domain name of the source
• title (string): RSS title
• text (string): extracted main article text

Summary fields

• summary_title (string): title-based summary
• short_summary (string): meta/RSS description-based short summary (may be empty)
• rss_desc (string): raw RSS description (if available)
• page_desc (string): extracted meta description / lead text (if available)

Extractive package

• sentences (list[string]): sentence list from text
• ext_labels (list[int]): 0/1 labels for sentence selection
• ext_selected (string): concatenation of selected sentences
• ref_for_labels (string): which reference was used (short_summary or summary_title)

Meta

• n_sentences (int)
• text_chars (int)

Example

{
  "id": "354a96ae73d18e02",
  "url": "https://example.com/news/123",
  "source_domain": "example.com",
  "title": "Example headline",
  "text": "Full article text ...",
  "summary_title": "Example headline",
  "short_summary": "Short editorial description of the article.",
  "sentences": ["Sentence 1 ...", "Sentence 2 ..."],
  "ext_labels": [1, 0],
  "ext_selected": "Sentence 1 ..."
}