Erenyanic/sts-tr-dataset

Turkish STS — Raw Sentence Pool

527 Turkish sentences collected as raw material for building a Semantic Textual Similarity (STS) dataset.

The sentences are unpaired and unscored. Pairing and scoring are done in a later stage; this repository is the pool those pairs will be drawn from, annotated with the metadata that makes principled pairing possible.

Why this exists

The hard part of building an STS dataset is not finding sentences — it is deciding which two sentences to put together.

Pair two Turkish sentences at random and roughly 99% of them score 0. You end up with a dataset that is almost entirely zeros and measures nothing. A usable STS dataset needs pairs deliberately spread across the whole similarity range, and that spread has to come from a pairing strategy, not from luck.

This repository holds a sentence pool where that strategy is possible because each sentence carries the context needed to reason about it: which article it came from, which story it covers, and what topic it belongs to.

Columns

Column	Description
id	Row identifier
sentence	The Turkish sentence
domain	haber (news, 352) or vikipedi (Wikipedia, 175)
category	One of 13 topics (gundem, ekonomi, spor, bilim, tarih, …)
source	Outlet or wikipedia
doc_id	Sentences from the same article share this.
role	baslik / spot for news; lead1–lead3 for Wikipedia
event_id	Headlines from different outlets covering the same story share this.
url	Source article URL

doc_id and event_id are what turn a flat list of sentences into a pairing resource.

What you can build from it

Without writing a single sentence yourself:

Pairing	Count	Typical similarity
Cross-outlet, same event (event_id)	20 pairs / 16 events	High — independently written paraphrases
Headline + its own lede (doc_id, baslik+spot)	165 pairs	Medium-high
Adjacent Wikipedia lead sentences (consecutive lead numbers)	98 pairs	Medium
Non-adjacent sentences from one article	55 pairs	Low-medium
Same category, different doc_id	many	Low
Different category and domain	many	Zero

An event_id example — the same incident, reported independently by two newspapers (E02):

hurriyet — New York'a giden uçak rahatsızlanan yolcu için İstanbul'a indi

cumhuriyet — Dubai-New York seferi yapan uçak İstanbul Havalimanı'na iniş yaptı

The two headlines share barely any wording, yet report the same event. That is what makes naturally occurring cross-outlet coverage more useful than a generated paraphrase: the lexical divergence is real, not manufactured.

Note on two "wrong" clusters

Clusters E14 and E16 group stories that are not the same event:

• E14 — two unrelated yapay zekâ stories
• E16 — two different explosions (a volcano in Guatemala, a chemical plant in Brazil), linked by patlama and tahliye edildi

These were left in deliberately. High lexical overlap with low semantic similarity is the most valuable kind of row in an STS dataset — it is what separates a real model from a bag-of-words baseline. The extremes of the scale are nearly free signal; every model gets paraphrases and unrelated pairs right.

Files

• raw_sentences.csv — the sentence pool
• PAIRING_PROTOCOL.md — pairing recipes, a suggested 80-pair quota, a Turkish-language 0–5 scoring rubric, Turkish morphological perturbation patterns (-mA negation, argument swap via case marking, -DI/-mIş evidentiality), and pre-submission balance checks
• collect_raw.py — the collector, re-runnable

Reproducing

python3 collect_raw.py   # stdlib only, no dependencies

RSS feeds only expose current articles, so a later run returns different news. The Wikipedia portion is stable. This snapshot was collected 2026-08-05.

Known limitations

• Six Wikipedia articles are missing lead1 because their first sentence exceeded the 210-character cap. Check role numbers rather than assuming any two rows sharing a doc_id are adjacent.
• Event clustering is lexical (rare-token overlap between headlines from different outlets). It is high-precision after a document-frequency filter, but it is not exhaustive — related stories worded very differently are missed.
• gundem is the largest category (126 sentences) because it is the category with the most outlets, which is also what makes cross-outlet event clustering work there.
• News is a single-day snapshot, so topics cluster around whatever was in the news that day.

Sources and licence

• Wikipedia lead paragraphs via the Wikimedia REST API — text is CC BY-SA 4.0.
• News headlines and ledes from the public RSS feeds of Hürriyet, Cumhuriyet, Anadolu Ajansı, BBC Türkçe, Euronews Türkçe, Habertürk, Milliyet and Independent Türkçe. Each row records its outlet and article URL.

The repository is licensed CC BY-SA 4.0, the licence of the Wikipedia portion. News headlines are short excerpts included for research and educational use with per-row attribution and source URLs, in the manner of sentence-pair corpora such as MRPC. Rights in those headlines remain with their publishers; if you redistribute this data, keep the source and url columns intact.

Intended use

Coursework: constructing a Turkish STS dataset of up to 80 scored sentence pairs, in the standard STS format — sentence1, sentence2, and a score in the range 0–5.