--- dataset_info: features: - name: translation dtype: string splits: - name: ru_kbd num_bytes: 8395185 num_examples: 99958 - name: kbd_ru num_bytes: 21911512 num_examples: 122080 - name: ru_ady num_bytes: 4068940 num_examples: 53987 - name: ady_ru num_bytes: 15731589 num_examples: 57305 download_size: 26373815 dataset_size: 50107226 configs: - config_name: default data_files: - split: ru_kbd path: data/ru_kbd-* - split: kbd_ru path: data/kbd_ru-* - split: ru_ady path: data/ru_ady-* - split: ady_ru path: data/ady_ru-* language: - ady - kbd - ru task_categories: - translation pretty_name: Circassian-Russian Parallel Corpus size_categories: - 100K adiga ai logo

Circassian-Russian Parallel Corpus v1.0

This is a high-quality dataset containing over 330,000 parallel text pairs for machine translation between Russian and the Circassian language in its two literary dialects: East Circassian (Kabardian, `kbd`) and West Circassian (Adyghe, `ady`). ## About Circassian Circassian is an indigenous language of the Northwest Caucasus region. The language is notable for its complex phonological system (featuring 50+ consonants), ergative-absolutive alignment, and polysynthetic structure. Today, Circassian is spoken by approximately 1.5 million people worldwide, with significant diaspora communities across Turkey, Jordan, Syria, and other countries. Despite its cultural significance and unique linguistic features, Circassian remains underrepresented in digital spaces and natural language processing research, making it a prototypical low-resource language. This dataset represents the largest publicly available parallel corpus for the Circassian language to date. ## Dataset Significance and Applications This dataset was created to address the scarcity of digital resources for Circassian and to enable the development of: - Neural machine translation systems between Russian and both Circassian literary dialects - Language preservation tools and educational applications for Circassian communities - Computational linguistic research on polysynthetic and ergative-absolutive languages - etc. The corpus contains a diverse range of texts, including literary works, folk stories, textbooks, dictionaries, technical content, and everyday expressions, providing broad coverage of various domains and linguistic phenomena. ## Dataset Structure ### Data Instances Each example is a dictionary containing a translation pair: ```json { "translation": { "kbd": "Уи пщэдджыжь фIыуэ!", "ru": "Доброе утро!" } } ``` ### Data Splits | Split | Source | Target | Examples | |-------|--------|--------|----------| | kbd_ru | East Circassian | Russian | 120,218 | | ru_kbd | Russian | East Circassian | 99,956 | | ady_ru | West Circassian | Russian | 57,305 | | ru_ady | Russian | West Circassian | 53,989 | | Total | | | 331,468 | ## Dataset Creation ### Source Data The data was collected from publicly available sources, including dictionaries, books, articles, literary works, etc. For literary works, special processing was applied to create high-quality translation pairs: - Works without existing Russian translations were translated from scratch into Russian and then split into sentence pairs - Works with existing translations were refined by converting "artistic" translations into more precise, literal translations better suited for the machine translation training task. This approach ensured that the parallel text pairs maintain semantic equivalence while being more suitable for training translation models than purely literary translations that might prioritize style over precision. ### Data Quality and Processing All data was verified for quality either manually or through automated processes. The dataset underwent rigorous cleaning and preprocessing: - **Quality Verification**: The majority of entries were checked for accuracy and completeness by a native speaker. - **Standardization**: All [palochka](https://en.wikipedia.org/wiki/Palochka)-like characters (1, l, I, І, ӏ) were standardized to uppercase palochka (Ӏ). - **Script Consistency**: Look-alike Latin letters were converted to their Cyrillic equivalents to maintain script consistency. - **Deduplication**: Duplicate entries were identified and removed to ensure dataset quality. - **Filtering**: Entries with imbalanced text lengths or other quality issues were filtered out. - **Normalization**: Special characters and spacing were normalized across all texts. These processing steps helped ensure the dataset's consistency and reliability for machine translation tasks. ## Considerations for Using the Data This dataset was created to support the development of machine translation systems for the Circassian language, contributing to language preservation and digital accessibility for Circassian speakers. Researchers working with this data should be aware that: - The two Circassian dialects (Kabardian and Adyghe) are structurally related but differ in phonology, vocabulary, and some grammatical features - The corpus contains texts from various time periods and domains, which may exhibit stylistic and lexical differences - While extensive cleaning has been performed, some translation pairs may still contain minor inaccuracies - For optimal results in machine translation tasks, models should be trained on each dialect separately ## License This dataset is licensed under the [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/) license. This means you are free to share and adapt the material for any purpose, even commercially, under the following terms: - Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made. ## Citation Information If you use this dataset in your research or applications, please cite it as: ```bibtex @dataset{adiga-ai_circassian-russian_v1, author = {Anzor Qunash}, title = {Circassian-Russian Parallel Text Corpus v1.0}, year = {2025}, publisher = {adiga.ai}, url = {https://huggingface.co/datasets/adiga-ai/circassian-parallel-corpus} } ```