AlidarAsvarov/quran-caucasus

Quran Parallel Corpus for Languages of the Caucasus

A verse-aligned parallel corpus of Quran translations into languages of the Caucasus, plus the Arabic original, three Russian translations, and an English translation as pivot/anchor texts. Built for NMT research on low-resource Caucasian languages: several of these languages (Lezgian, Kabardian, Adyghe, Dargwa, Karachay-Balkar) have very little other parallel data of this size and quality.

All texts follow the canonical Kufan/Hafs numbering (114 surahs, 6236 ayahs; the basmala is verse 1:1). Any translation-specific numbering in the printed sources (e.g. Flügel) has already been mapped to this scheme, so every file aligns to every other file by (surah, ayah).

Data format

One CSV per translation, columns:

column	meaning
surah	chapter number, 1–114
ayah_start	first ayah covered by this row
ayah_end	last ayah covered by this row (= ayah_start for normal rows)
text	the translation of that ayah range

ayah_start < ayah_end marks a merged range: the printed translation renders several ayahs as one passage that cannot be split. Only two files contain such rows (dar-gamidov: 144, krc-ebzeev: 1); all other files are strictly one row per ayah.

surah,ayah_start,ayah_end,text
1,2,2,"(Вири) гьамд (шукур) хьуй Аллагьдиз, (вири) алемрин Раббидиз,"

Translations

config	language	ISO	translator / edition	rows	bracketed rows
ar-uthmani	Arabic (original, Uthmani script)	ar	—	6236	0%
ru-kuliev	Russian	ru	Elmir Kuliev	6236	16%
ru-abu-adel	Russian	ru	Abu Adel (translation + tafsir insertions)	6236	94%
ru-krachkovsky	Russian	ru	Ignaty Krachkovsky (academic, literal; publ. 1963)	6236	1%
en-sahih	English	en	Saheeh International	6236	48%
az-musayev	Azerbaijani	az	Alikhan Musayev	6236	51%
ce-magomedov	Chechen	ce	Magomedov	6236	22%
ce-ibragimov	Chechen	ce	Ibragimov	6236	91%
lez-yamin	Lezgian	lez	Yamin-hajji Magomedov (Ямин хаджи Магомедов)	6236	95%
ka-mikeladze	Georgian	ka	Mikeladze	6236	9%
kbd-quazha	Kabardian (East Circassian)	kbd	Quazha Muhemedqeyr, Boliy Mustafar, Nalo Zaur (Хъуажь Мухьэмэдхъейр, Болий Мустафар, Нало Заур)	6236	26%
ady-autlev	Adyghe (West Circassian)	ady	Faiz Uthman Autlev	6225	1%
dar-gamidov	Dargwa	dar	Gamidov (1995, Yupiter, Makhachkala)	6086	<1%
krc-ebzeev	Karachay-Balkar	krc	Abu-Yusuf Ebzeev	6227	55%

"Bracketed rows" = share of rows containing () or [] interpolations; see Translator interpolations below.

Provenance and processing

• ru-kuliev, ru-abu-adel, az-*, ce-*, lez-*, ka-*, kbd-* were downloaded from the public reader API of quranacademy.org (2026-07). HTML/QTML markup and footnotes were stripped; the translators' inline bracketed clarifications were kept.
• ru-krachkovsky is the Tanzil text (tanzil.net, ID ru.krachkovsky) of Ignaty Krachkovsky's literal academic translation, published posthumously in 1963. Note its idiosyncrasies: the muqatta'at are transcribed as plain letter strings (2:1 is «Алм.»), and the unfinished draft's parenthetical passages were kept by the editors.
• en-sahih is the Tanzil text (tanzil.net, ID en.sahih) of the Saheeh International translation. It marks grammatically implied words with square brackets [] (48% of rows); no round brackets and no merged verses.
• ar-uthmani is the Uthmani text from github.com/nuqayah/quran-text (Tanzil-style plain text). The basmala embedded in the first ayah of surahs 2–114 was stripped, so 1:1 is the only basmala row — consistent with the translations.
• ady-autlev was parsed from an electronic text (2009, cp1251).
• dar-gamidov was OCR'd from a 673-page scan of the 1995 printed edition (LLM OCR, line-exact, with lexicon-based de-hyphenation). The book uses the Flügel (1834) verse numbering — the translation was evidently made from Krachkovsky's Russian version — and was realigned to the canonical scheme via a Flügel→Kufan concordance. Flügel merges became ayah_start–ayah_end ranges (144 rows); Flügel splits were concatenated. Alignment was validated with a proper-name anchor audit against ru-kuliev (514 anchor hits, 0 shifts). Residual single-character OCR confusions (ь/ъ, и/н) may remain — the text was never fully proofread by a native speaker.
• krc-ebzeev was parsed from a PDF of the printed edition. Per-surah forewords and standalone commentary paragraphs between ayahs were dropped; inline parenthetical glosses were kept.

Normalization applied to all Cyrillic-script texts:

• Palochka: the letter Ӏ is always U+04C0 (capital palochka); the small palochka and look-alikes (Latin I/l, digit 1, |, Ukrainian І/і) were replaced in Chechen, Lezgian, Kabardian, Adyghe, Dargwa texts.
• Latin homoglyphs inside Cyrillic words (a, c, e, o, p, x, y, …, artifacts of italics OCR) were mapped back to Cyrillic; NFC applied. Vowel-length diacritics used by the Chechen translations (а̄, о̄, ӣ, …) are combining marks and were preserved.

Known gaps and defects

• Missing verses (lacunae in the sources, not numbering variants):
  • ady-autlev: 7:196–206 absent (11 ayahs).
  • dar-gamidov: 16:47 absent (misprint skips a verse number).
  • krc-ebzeev: 41:39–46 absent (the source PDF prints 41:30–38 twice instead; the duplicate was cut). 36:55–56 is a merged row.
• 44 rows contain unbalanced brackets. Ten of them (five spans) are legitimate commentary crossing verse boundaries (ady-autlev 23:6–7; dar-gamidov and ru-krachkovsky 4:157–159, 23:6–7 — a parenthesis opens in one verse and closes in a later one; the Dargwa spans mirror Krachkovsky's because that translation was made from his version). The other 34 are typos present in the sources themselves: ce-ibragimov 13, ce-magomedov 11, kbd-quazha 6, lez-yamin 2, ka-mikeladze 1, dar-gamidov 1. Span-removal regexes leave the stray bracket character behind on these rows — filter them or clean up leftover ()[] characters afterwards.
• ce-magomedov occasionally uses /…/ slashes instead of parentheses for interpolations, and sometimes a stray mix of the two.
• No row has empty text; every file covers all ayahs except the lacunae above.

Translator interpolations in brackets

Most translations insert words that are not in the Arabic original, marked with () or []. This matters for NMT: bracketed material is target-side content with no source counterpart, so training against the bracket-free Arabic (or a sparsely bracketed translation) teaches the model to over-generate, and commentary in references inflates length and skews BLEU/chrF differently for each language.

The brackets are not one phenomenon. Four distinct uses occur in this corpus:

1. Pure commentary (tafsir) — explanations with no counterpart in the Arabic: Господу миров [всех творений] (ru-abu-adel). Safe to remove.
2. Grammatically implied words — elided subjects/objects that the target language needs for a complete sentence: (yoluna) yox! (az-musayev). Removing these can leave an awkward or broken sentence.
3. Glosses of loanwords — гьамд (шукур) (lez-yamin), where the parenthesis explains an Arabic loan with a native synonym. Removal is safe.
4. Name equivalents — Исраила (Израиля) (ru-kuliev). Trivial either way.

Conventions differ per translator. ru-abu-adel is explicit: [] marks tafsir insertions, () marks words implied by the Arabic — so dropping [] spans while keeping () content (minus the parentheses) is a principled middle ground for that file. ce-ibragimov uses [] the same way and just as heavily. en-sahih uses [] exclusively, but for grammatically implied words (rule 2) rather than tafsir. The remaining translations use () for everything, so span type must be judged statistically or by sampling.

Suggested defaults:

• Test/evaluation sets: strip bracketed spans (and tidy whitespace), then spot-check a sample per language for sentences broken by rule 2 above. Commentary-heavy references make cross-language metric comparison unfair.
• Training data: either variant is defensible; stripping reduces hallucination pressure against Arabic sources, while keeping brackets (or keeping content and dropping only the bracket characters) works well when both sides of a pair use matched interpolation styles — e.g. lez-yamin closely mirrors ru-abu-adel in interpolation density (95% vs 94% of rows), so a Russian↔Lezgian pair built from those two has near-parallel commentary. Using both stripped and unstripped variants as augmentation is also an option.

Bracket-handling script

import re

_SPAN = {"(": re.compile(r"\([^()]*\)"), "[": re.compile(r"\[[^\[\]]*\]")}

def clean_verse(text: str, drop: str = "([", unwrap: str = "") -> str:
    """Remove bracketed spans and/or bracket characters.

    drop:   bracket kinds whose spans are removed entirely, e.g. "([" or "[".
    unwrap: bracket kinds whose characters are removed but content kept.
    Innermost spans are removed first, so nesting works. Unbalanced rows
    (40 in the corpus, see the card) keep the stray bracket character.
    """
    for kind in drop:
        pattern = _SPAN[kind]
        while pattern.search(text):
            text = pattern.sub(" ", text)
    for kind in unwrap:
        closer = {"(": ")", "[": "]"}[kind]
        text = text.replace(kind, " ").replace(closer, " ")
    text = re.sub(r"\s+([,.;:!?])", r"\1", re.sub(r"\s+", " ", text))
    return text.strip()

# examples
clean_verse("Господу миров [всех творений],", drop="[")   # -> "Господу миров,"
clean_verse("(Вири) гьамд (шукур) хьуй", drop="(")        # -> "гьамд хьуй"
clean_verse("Господу миров [всех творений],", drop="[", unwrap="(")
#   ru-abu-adel convention: drop tafsir, keep implied words

Usage

from datasets import load_dataset

lez = load_dataset("AlidarAsvarov/quran-caucasus", "lez-yamin", split="train")

Building sentence pairs for NMT — the helper below aligns any two files, handles lacunae (ayahs missing on either side are skipped), and joins texts over the union of merged ranges. For the eleven one-row-per-ayah files it degenerates to a plain (surah, ayah) join. Dependency-free:

import csv

def load(path):
    with open(path, newline="", encoding="utf-8") as f:
        return [(int(r["surah"]), int(r["ayah_start"]), int(r["ayah_end"]),
                 r["text"]) for r in csv.DictReader(f)]

def align(a, b):
    """Pair two translations -> (surah, ayah_start, ayah_end, text_a, text_b)."""
    index = lambda rows: {(r[0], y): r for r in rows
                          for y in range(r[1], r[2] + 1)}
    ia, ib = index(a), index(b)
    blocks = []
    for surah, ayah in sorted(set(ia) & set(ib)):
        ra, rb = ia[(surah, ayah)], ib[(surah, ayah)]
        if ayah == ra[1] == rb[1]:      # both sides start a fresh row here
            blocks.append([surah, ayah, ayah, [], []])
        block = blocks[-1]
        block[2] = ayah
        for rows, r in ((block[3], ra), (block[4], rb)):
            if r not in rows:
                rows.append(r)
    return [(s, y0, y1, " ".join(r[3] for r in ta), " ".join(r[3] for r in tb))
            for s, y0, y1, ta, tb in blocks]

pairs = align(load("data/ru-kuliev.csv"), load("data/dar-gamidov.csv"))

Copyright notice

The Arabic text is not subject to copyright. The translations are the intellectual property of their respective translators and publishers. They are compiled here from publicly accessible sources for research on low-resource machine translation. If you are a rights holder and object to this redistribution, please open a discussion on the dataset page and the material will be removed.

Acknowledgements

Translation texts were obtained via quranacademy.org's public reader; the Krachkovsky and Saheeh International translations come from the Tanzil project, as does the Arabic Uthmani text (via nuqayah/quran-text); the Flügel→Kufan verse concordance used for dar-gamidov was derived from the correlation chart at muhammadanism.org.