This is a large-scale Arabic–Russian parallel corpus containing 15,467,945 sentence pairs aggregated from diverse sources: open subtitle and document corpora (OPUS), TED Talks, lexicographic dictionaries, religious texts (Quran, hadith collections, Bible), conversational phrasebooks, community-contributed examples, and news articles from political and diplomatic open sources.

Curated by: Arabov Mullosharaf Kurbonovich
Funded by: Not applicable (independent research)
Shared by: ArabicNLPWorld
Language(s): Arabic (ar), Russian (ru)
License: MIT
Total pairs: 15,467,945
Correlation (Arabic vs Russian length): 0.5457

Dataset Sources

Repository: https://huggingface.co/datasets/ArabicNLPWorld/arabic-russian-translation-corpus
Paper: Not available
Demo: Not available

Uses

Direct Use

Training and evaluating Arabic–Russian machine translation systems
Cross-lingual word embeddings and sentence representation learning
Bilingual dictionary induction and lexicographic studies
Linguistic analysis of Arabic–Russian alignment
Low-resource NLP research leveraging large parallel data

Out-of-Scope Use

Tasks requiring human gold-standard annotations without any preprocessing
Surveillance, profiling, or discriminatory applications
Sensitive domain inference without additional filtering

Dataset Structure

Each instance contains a parallel pair:

arabic (str): Arabic source text
russian (str): Russian target translation

The corpus has no predefined train/validation/test splits — users should create their own splits according to task requirements.

Dataset Creation

Curation Rationale

Arabic–Russian is an under‑resourced language pair compared to English‑centric corpora. This dataset combines large‑scale automatically collected data (OPUS, TED) with carefully selected lexicographic, religious, phrasebook, and news resources to improve domain coverage, lexical richness, and cultural relevance.

Source Data

The corpus is built from the following sources:

Source	Pairs	Percentage
OPUS	14,924,037	96.48%
TED Talks	375,463	2.43%
Religion (Quran, hadith)	82,302	0.53%
Dictionary (Baranov + Borisov)	43,662	0.28%
Bible	31,102	0.20%
Tatoeba	9,044	0.06%
News (political/diplomatic)	1,683	0.01%
Conversation (phrasebook)	652	<0.01%

Detailed source description

1. OPUS – A large collection of open parallel corpora maintained by the University of Helsinki (primarily UN/EU documents and subtitles).

2. TED Talks – Translated transcripts of TED conference speeches.

3. Dictionary sources (43,662 pairs):

Arabic–Russian Dictionary by Kh.K. Baranov (latest edition, first published 1960s) — a standard reference for Arabic‑Russian lexicography.
Russian–Arabic Dictionary by V.M. Borisov (latest edition) — complementary bidirectional coverage.

4. Religious texts (82,302 + 31,102 = 113,404 pairs):

Quran with Russian translation by Elmir Kuliev (Кулиев) — a widely used modern Russian interpretation, based on the 2022 revised edition.
Sahih Muslim — one of the six major hadith collections in Sunni Islam, with Russian translation.
41 Hadith of al‑Nawawi — a well‑known collection of prophetic traditions.
Bible – Old and New Testament excerpts with Russian Synodal translation.

5. News (1,683 pairs) – Political and diplomatic news articles collected from open sources. These entries add journalistic style and contemporary vocabulary to the corpus.

6. Russian–Arabic phrasebook (Conversation) – Everyday conversational phrases.

7. Tatoeba – Community‑contributed example sentences.

Data Collection and Processing

All pairs were collected from publicly available or digitised sources, deduplicated, and minimally filtered (non‑empty strings only). No aggressive length filtering was applied to preserve natural variance.

Processing tools: Custom Python scripts with standard NLP preprocessing utilities.

Who are the source data producers?

Lexicographers: Kh.K. Baranov (Arabic–Russian dictionary), V.M. Borisov (Russian–Arabic dictionary)
Religious scholars/translators: Imam Muslim (Sahih Muslim), Imam al‑Nawawi (40/41 Hadith), Elmir Kuliev (Quran translation), Biblical translation teams (Russian Synodal Bible)
News sources: Open diplomatic statements.
Other contributors: TED conference speakers, Tatoeba community, OPUS maintainers, anonymous phrasebook compilers

Annotations

No additional annotations were added beyond the existing translations. The parallel pairs are provided as‑is.

Personal and Sensitive Information

The dataset includes publicly available translations that may contain names of individuals (e.g., political figures, UN officials, religious personalities). No deliberate collection of private or sensitive data was performed. Users should exercise caution when applying the data to tasks involving named entity recognition or personal information extraction.

Bias, Risks, and Limitations

Known limitations:

Domain imbalance: 96.48% of the data comes from OPUS (primarily official/UN‑style documents), which may bias models toward formal, diplomatic language.
Correlation: The length correlation of 0.5457 indicates a moderate relationship, suggesting that source and target sentence lengths do not always align closely.
Alignment noise: Automatically aligned sources (OPUS, TED) may contain occasional misalignments.
Religious domain: A significant portion of the non-OPUS data comes from religious texts, which may bias models toward religious terminology.
No predefined splits: Users must create their own training/validation/test partitions.
News domain: The 1,683 news pairs are limited in size.

Risks:

Models trained on this corpus may inherit source biases (Western‑leaning, diplomatic, UN/TED styles).
Dialectal Arabic and informal spoken Russian are under‑represented.

Recommendations

Users are advised to:

Filter out extreme length ratios (e.g., >1:5) if training MT models.
Create stratified splits across sources to avoid domain leakage.
Apply task‑specific cleaning (e.g., remove very short pairs for translation tasks).
Manually validate a small random sample before production use.
For religious text applications, consult the original Arabic sources alongside Kuliev's Quran translation and the Synodal Bible.

Citation

BibTeX:

@dataset{arabicnlpworld_2026,
  title = {Arabic-Russian Translation Corpus},
  author = {Arabov, Mullosharaf Kurbonovich},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/ArabicNLPWorld/arabic-russian-translation-corpus/}
}

APA:

Arabov, M. K. (2026). Arabic‑Russian Translation Corpus [Dataset]. Hugging Face. https://huggingface.co/datasets/ArabicNLPWorld/arabic-russian-translation-corpus/

Glossary

Length correlation (0.5457): Pearson correlation between character counts of Arabic and Russian sentences.
Tokens: Word‑level segmentation by whitespace and punctuation.
OPUS: A large parallel corpus collection from the University of Helsinki.
Quran (Kuliev translation): The Russian translation of the Quran by Elmir Kuliev (Эльмир Кулиев), first published in 2002, revised edition 2022.
Sahih Muslim: One of the Kutub al‑Sittah (six major hadith books) in Sunni Islam.
40/41 Hadith of al‑Nawawi: A famous collection of prophetic traditions compiled by Imam Yahya ibn Sharaf al‑Nawawi.
Synodal Bible: The Russian Synodal translation of the Bible (canonical for the Russian Orthodox Church).

More Information

Full corpus statistics (JSON): Available in ./corpus_stats_light/stats_light.json
Readable report: Available in ./corpus_stats_light/

Updated sample statistics

Top Arabic tokens (based on 200,000 sample pairs)

Arabic token	Frequency
عَنْ	87,038
قَالَ	84,918
بْنُ	61,496
في	48,867
بْنِ	43,899
حَدَّثَنَا	40,861
ﷺ	35,511
فِي	34,612
من	34,165
أَبِي	30,767

Top Russian tokens (based on 200,000 sample pairs)

Russian token	Frequency
и	161,301
в	137,464
что	82,087
на	64,025
не	61,831
с	51,462
сказал	42,551
к	32,447
Глава	31,911
по	31,858

Most frequent letters

Arabic:

Character	Count
space	4,012,558
َ (fatha)	3,851,606
ا	2,290,816
ل	1,897,610
ِ (kasra)	1,285,437

Russian:

Character	Count
space	5,120,757
о	3,078,201
а	2,558,365
и	2,331,520
е	2,279,883

Full Statistics Summary

Total pairs: 15,467,945
Arabic characters total: 2,293,642,421
Russian characters total: 3,017,354,772
Arabic tokens total: 372,114,969
Russian tokens total: 388,707,431
Avg Arabic chars: 148.28
Avg Russian chars: 195.07
Avg Arabic tokens: 24.06
Avg Russian tokens: 25.13
Length correlation: 0.5457

Dataset Card Authors

Arabov Mullosharaf Kurbonovich (ArabicNLPWorld)

Dataset Card Contact

https://huggingface.co/ArabicNLPWorld

Downloads last month: 9

Total file size:

4.18 GB