muzaffercky/kurdish-theses-yok

Kurdish Academic Theses (Kurmanji, Zazaki, Sorani)

Text and abstracts of academic theses that YÖK Tez Merkezi, the Turkish national thesis repository, lists under Kurdish. Each row is one thesis.

Subsets

A thesis is filed under the language of its text. There is one subset per language:

Subset	Language	Theses	Words	Abstracts in that language
kmr_Latn	Kurmanji (Northern Kurdish), Latin script	408	18,574,237	361
zza_Latn	Zazaki, Latin script	29	1,162,986	21
ckb_Arab	Sorani (Central Kurdish), Arabic script	4	105,254	4
hac_Arab	Arabic-script theses whose text is damaged (see below)	6	250,276	0

hac_Arab is not usable as clean text. The code is the one the language identifier gave these six theses, and it stands for Gurani, but they are Sorani and Kurmanji theses printed in old fonts. Their letters come out mapped wrong — ثەسچبٚ دەکەٚێذ where the page reads differently. The subset is there so the theses are not lost, and so they stay apart from the clean Sorani in ckb_Arab. Their Turkish and English abstracts are intact.

from datasets import load_dataset

kurmanji = load_dataset("muzaffercky/kurdish-theses-yok", "kmr_Latn", split="train")
zazaki = load_dataset("muzaffercky/kurdish-theses-yok", "zza_Latn", split="train")
sorani = load_dataset("muzaffercky/kurdish-theses-yok", "ckb_Arab", split="train")
damaged = load_dataset("muzaffercky/kurdish-theses-yok", "hac_Arab", split="train")

Fields

Every subset has the same fields.

Field	Type	Description
tez_id	string	Thesis ID at YÖK Tez Merkezi
title	string	Thesis title
url	string	Link to the thesis at YÖK Tez Merkezi
word_count	int	Number of words in text
text	string	The body of the thesis: headings and paragraphs, separated by blank lines. The abstracts and the bibliography are not in it.
summaries	dict	The abstracts of the thesis, one key per language. A key is null when the thesis has no abstract in that language.

The keys of summaries are the same in every row:

Key	Abstract in	Heading it is printed under	Found
kmr_Latn	Kurmanji	KURTE	363
zza_Latn	Zazaki	KILMNUS, KILMVATE, KILMÎYE	21
ckb_Arab	Sorani	پوختە	4
tur_Latn	Turkish	ÖZET	365
eng_Latn	English	ABSTRACT, SUMMARY	394

An abstract is typically 200 to 280 words.

What it can be used for

• Pre-training: train a language model on text, or keep training an existing one, so it learns the language. No labels are needed.
• Summarization: text → an entry of summaries.
• Parallel data: the abstracts of one thesis say the same thing in two or three languages, so two entries of summaries are a parallel pair. The pairs are aligned by abstract, not by sentence; split and align the sentences to get sentence pairs.
• Translation: train or test a translation model on those pairs.

The abstract a thesis prints in its front matter is taken out of text, so the summary is not inside the text it summarizes. A thesis that prints its abstract a second time elsewhere keeps that second copy in text.

How it was made

1. Download — the PDFs are fetched from YÖK Tez Merkezi.
2. Extract — MinerU 4.0 reads each PDF's layout and labels its blocks as headings, text or contents-page entries. Scanned pages are read with OCR. Tables, images, page numbers, headers and footers are dropped. Text is normalized to Unicode NFC.
3. Sections — the blocks are cut into sections at the headings.
   • A section under a summary heading becomes that language's abstract. It is cut at its keywords line. If GlotLID v3 says it is not in the language its heading names, it is discarded.
   • A section under a ÇAVKANÎ (bibliography) heading is discarded.
   • A section of 50 words or more that GlotLID labels neither Kurmanji, Zazaki nor Sorani is discarded. This removes Turkish and English pages and chapters.
   • Everything else is text.
4. Filter — GlotLID labels the whole text once, and that decides the subset. A thesis with fewer than 1,000 words, or with any other label, is dropped.

Of 457 downloaded theses, 447 are in the dataset.

Known limits

• The text comes from automatic layout analysis and OCR. There are misread characters, and a heading can be attached to the wrong paragraph.
• Some abstracts are missing or under the wrong heading. On some pages MinerU puts the body of an abstract before its heading. The abstract is then lost, or filed under the previous heading and discarded by the language check. Where the contents page names an abstract that was not recovered, the entry is null.
• An abstract can start with title-page lines, such as Master Thesis or the author's name.
• Language is judged per section, not per sentence. A Turkish or English quotation inside a Kurmanji section stays in text, and so does any section under 50 words.
• Zazaki is one subset. GlotLID has two labels for it (diq_Latn, kiu_Latn) and does not separate these theses by them reliably.
• Clean Sorani is only four theses. Thirteen more theses in Arabic script use old fonts whose text cannot be extracted correctly. Six of them are in hac_Arab, damaged; the other seven are left out.
• One thesis is in the wrong subset. 458722 is in kmr_Latn but is Sorani written in Latin letters.
• Reference lists under a heading other than ÇAVKANÎ are still in text.

Source and license

The theses are public documents published by YÖK Tez Merkezi; url links each row to its source. The copyright of each thesis belongs to its author.

The dataset was produced by the kurdish-kurmanji-thesis pipeline.