README.md

metadata

language:
  - tr
license: cc-by-sa-4.0
task_categories:
  - sentence-similarity
tags:
  - turkish
  - semantic-textual-similarity
  - sts
  - dataset-construction
size_categories:
  - n<1K
pretty_name: Turkish STS Dataset
configs:
  - config_name: default
    data_files: sts_tr_dataset.parquet

Turkish STS Dataset

81 Turkish sentence pairs, scored for semantic similarity on a 0–5 scale by two annotators, together with the 527-sentence raw pool part of it was drawn from.

from datasets import load_dataset

ds = load_dataset("Erenyanic/sts-tr-dataset")

Contributors and method

Contributor	Pairs created	Sentence source
Erenyanic	43	News RSS feeds and Wikipedia, via the collector in this repo
nursimakgul	38	Web sources on film, cooking and general science

Pair creation was split, but scoring was not. Both contributors scored all 81 pairs independently, and the published score is the mean of the two ratings. Every row therefore reflects two people's judgement rather than one.

The scored dataset (sts_tr_dataset.parquet)

Column	Type	Description
sentence1	string	First Turkish sentence
sentence2	string	Second Turkish sentence
score	double	Semantic similarity, 0–5

Score distribution, rounded to the nearest whole score:

Band	0	1	2	3	4	5
Half one — news / Wikipedia (43)	7	10	5	5	6	10
Half two — film / cooking / science (38)	3	2	10	5	7	11
Total	10	12	15	10	13	21

81 pairs, mean 2.67, spanning the full range.

What the scores mean

5	Tamamen aynı anlam. Sözcükler farklı olabilir, aktarılan bilgi aynıdır.
4	Aynı olayı/olguyu anlatır, önemsiz bir ayrıntıda ayrışır.
3	Aynı konu ve aynı durum, ancak biri daha ayrıntılı veya kapsamı dar.
2	Aynı konu, farklı içerik. Ortak bir bağlam var, aktarılan bilgi farklı.
1	Yalnızca gevşek bir bağ var: ortak sözcükler veya ortak alan, farklı anlam.
0	İlgisiz.

Fractional scores were used for cases falling between two levels.

How it was built

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. The result is 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.

Both halves apply the same idea to different material: find one subject that two independent sources have each described in their own words. That yields genuine paraphrases whose lexical divergence is real rather than manufactured, and shifting how far apart the two subjects are moves a pair down the scale.

Half one — news and encyclopedia text (43 pairs). 527 sentences were collected from Turkish news RSS feeds and Wikipedia, each annotated with the context needed to reason about pairing it: which article it came from, which story it covers, which topic it belongs to. Pairs were then chosen by hand from that pool, using the metadata to hit the whole similarity range rather than sampling blindly. The high end comes from different outlets reporting one event; the low end from different topics.

Half two — film, cooking and general science (38 pairs). Sentences collected from web sources covering the same subject in different places: two descriptions of one film, two definitions of one natural phenomenon, two notes on one dish. The high end comes from two sources describing one subject — Buharlaşma, sıvı haldeki suyun ısı etkisiyle gaz haline geçmesidir against Buharlaşma, bir sıvının yüzeyinden gaz haline dönüşerek havaya karışması olayıdır. Descending the scale, a subject is paired with a neighbouring one (iklim against hava durumu, yıldırım against gök gürültüsü), then with an unrelated one (İmam bayıldı against Ekvator). This half reaches everyday domains a news and encyclopedia corpus does not.

No pre-paired or pre-labelled corpus was used at any stage. Existing Turkish paraphrase and NLI corpora would have supplied ready-made pairs, but the pairing decision is the substance of the task, so every pairing here was made by hand.

The raw pool (raw_sentences.csv)

The source for half one: 527 sentences, 352 from news and 175 from Wikipedia, across 13 topic categories.

Column	Description
id	Row identifier
id2	The sentence this one was paired with, if any
sentence	The Turkish sentence
domain	haber (news) or vikipedi (Wikipedia)
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. Pairing material available in the pool:

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

• sts_tr_dataset.parquet — the 81 scored pairs, merged
• pairs_to_score.csv — the 43 pairs of half one, with their averaged scores
• sts_data.csv — the 38 pairs of half two, with their averaged scores
• raw_sentences.csv — the 527-sentence source pool behind half one
• collect_raw.py — the collector
• make_dataset.py — extract builds a scoring sheet from the id2 column; build merges the scored CSVs into Parquet

The two halves are kept as separate CSVs and merged at build time, so each row's provenance stays traceable after the merge.

Reproducing

python3 collect_raw.py            # rebuild the raw pool (stdlib only)
python3 make_dataset.py extract   # pairs -> scoring sheet
python3 make_dataset.py build     # scored sheets -> merged parquet

RSS feeds only expose current articles, so re-collecting returns different news. The Wikipedia portion is stable. The pool snapshot was collected 2026-08-05.

Known limitations

• 81 pairs is small. This is a coursework-scale dataset. It is enough to sanity-check an embedding model, not to benchmark one — a correlation computed over 81 points has wide error bars.
• Only the averaged score is published. Each pair was rated by both annotators, but the two individual ratings were not retained, so inter-annotator agreement cannot be recomputed from this file.
• The halves differ in composition. Half one averages 101 characters per sentence and has a mean score of 2.35; half two averages 77 characters and a mean of 3.04. Since the halves also differ in domain and register, style and score are not fully independent.
• Source URLs were recorded for half one only. Every row of the raw pool carries its outlet and article URL. The web sources behind half two were not recorded per sentence, so those 38 pairs are not traceable to a page.
• Within half one, domain is unevenly spread across pairs: 36 news↔news, 7 news↔Wikipedia, 1 Wikipedia↔Wikipedia.
• Six Wikipedia articles in the pool are missing lead1 because their first sentence exceeded the 210-character collection cap. Check role numbers rather than assuming any two rows sharing a doc_id are adjacent.
• Event clustering in the pool is lexical (rare-token overlap between headlines from different outlets). High precision, but not exhaustive.
• The news portion 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 of the raw pool records its outlet and article URL.
• Web pages on film, cooking and general science, collected by nursimakgul. Individual page URLs were not recorded for these 38 pairs.

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