AlmaznayaGroza/cyrillic-language-classification

Cyrillic Language Classification Corpus

This corpus contains texts in 26 languages using the Cyrillic alphabet, designed for the task of automatic language identification.

Summary

• Languages: 26 individual Cyrillic languages, plus 17 mixed language combinations created through augmentation
• Source: texts collected from Wikipedia and augmented using various methods
• Size: 20,232 examples in total, including 18,334 original articles collected from Wikipedia and 1,898 texts created through data augmentation
• Format: CSV with 16 columns including detailed metadata for each example
• Performance: 99.46% accuracy achieved with XLM-RoBERTa after 12 training epochs

Detailed description

This corpus was compiled as part of an academic project with the aim of creating a resource for the automatic identification of languages using the Cyrillic alphabet, including minority languages rarely represented in language corpora. The resulting model demonstrates excellent performance with only 11 errors out of 2,043 test examples, achieving near-perfect calibration and robust performance across different text lengths.

Methodological choice

Belarusian variants merger

Belarusian presents a unique sociolinguistic situation with 2 systems used on Wikipedia:

• be (narkamoŭka / наркамоўка): official standardized variant
• be-tarask (taraškevica / тарашкевіца): historical pre-Soviet variant

Justification for the merger

This (mainly, but not exclusively) orthographic duality does not reflect a fundamental linguistic difference but rather distinct political and cultural choices. The merger of both variants in the final corpus is justified by several methodological considerations:

1. Linguistic unity: both variants represent the same language with different orthographic et lexical conventions

2. Practical applications: for natural language processing applications, this unification enables the creation of more robust models capable of handling natural orthographic variations in Belarusian

3. Representativeness: the unification provides a more complete and balanced representation of the Belarusian language in its contemporary usage patterns

This approach respects linguistic diversity while optimizing model efficiency for practical classification applications.

Rich linguistic mixes

A unique feature of this corpus is the inclusion of mixed texts combining various closely related Cyrillic languages, including:

• East Slavic mixes (ru_uk_mix, ru_be_mix, etc.)
• South Slavic mixes (bg_mk_mix, sr_bg_mix)
• Finno-Ugric mixes (koi_kv_mix, udm_koi_mix)
• Caucasian mixes (ab_kbd_mix, ab_ce_mix) These combinations can be used to test the robustness of classifiers in the face of multilingual content, a situation frequently encountered in the media of multicultural regions.

Data augmentation methods

Three augmentation approaches were used to enrich the corpus, particularly benefiting minority and under-resourced languages:

1. Synthetic text generation: creation of new texts using statistical models based on n-grams of characters (particularly useful for poorly endowed languages)

2. Text perturbation: introduction of controlled orthographic variations to simulate dialectal variations or common typographical errors

3. Multilingual text creation: two strategies were employed for mixing closely related languages:

• alternating sentences in different languages to simulate natural code-switching
• creating blocks of text in different languages to represent multilingual documents

Languages included

The corpus covers 26 Cyrillic languages, grouped by language family:

• East Slavic languages: Russian (ru), Ukrainian (uk), Belarusian (be), Rusyn (rue);
• South Slavic languages: Bulgarian (bg), Macedonian (mk), Serbian (sr);
• Turkic languages: Kazakh (kk), Kyrgyz (ky), Tatar (tt), Bashkir (ba), Chuvash (cv), Tuvin (tyv), Yakut (sah);
• Iranian languages: Tajik (tg), Ossetian (iron dialect) (os);
• Finno-Ugric languages: Komi (kv), Komi-Permyak (koi), Udmurt (udm), Mari (mhr), Erzya (myv);
• Mongolian languages: Mongolian (mn), Buryat (bxr);
• Caucasian languages: Abkhaz (ab), Kabardian (kbd), Chechen (ce).

Special attention has been paid to minority languages such as Abkhaz, Kabardian, Komi, and Meadow Mari, which are rarely included in standard training corpora. The final model achieves excellent performance across all language families, with 20 languages achieving perfect classification accuracy.

Structure and format

The corpus is divided into 3 stratified sets maintaining balanced representation across languages and source types:

• train_corpus.csv: training set (80% of data)
• validation_corpus.csv: validation set (10% of data)
• `test_corpus.csv: test set (10% of data, 2,043 examples)

Each CSV file contains 16 columns with rich metadata:

• language: language code or language combination (43 unique values total)
• title: title or identifier of the example
• text: textual content in Cyrillic script
• pageid: Wikipedia page identifier for original articles
• category: thematic category the text belongs to
• type: collection method (ordered, random, etc.)
• url: source URL for Wikipedia articles
• token_count: number of tokens in the text
• page_id: alternative page identifier
• detected_languages: additional language detection metadata
• is_multilingual: boolean flag for mixed language texts
• main_category: primary category classification
• source_type: origin of the text (original, augmented, etc.)
• source: detailed source information
• mixing_strategy: method used for creating mixed language examples
• original_id: reference to original text for augmented examples

Benchmark performance

The corpus was used to fine-tune an XLM-RoBERTa model (`papluca/xlm-roberta-base-language-detection) with the following configuration:

• training epochs: 12
• batch size: 32 (training), 64 (evaluation)
• learning rate: 8e-5
• maximum sequence length: 64 tokens
• warmup steps: 200
• weight decay: 0.01

Performance metrics

The fine-tuned model achieved exceptional results:

• overall accuracy: 99.46% (2,032/2,043 correct predictions)
• top-3 accuracy: 99.76%
• error rate: 0.54% (only 11 errors total)
• calibration quality: ECE = 0.0029 (near-perfect)
• average model confidence: 99.53%

Error analysis

The model demonstrates sophisticated linguistic understanding:

• 45.5% of errors are linguistically justified (all occurred with artificial mixtures containing the predicted language)
• most common confusion: Buryat (bxr) → Yakut (sah) and Udmurt (udm) → Kabardinian-Chechen mix (kbd_ce_mix), with 2 errors each
• only 9.1% of errors occur within the same linguistic family, showing very good discrimination between related languages
• strong metacognitive capabilities: errors occur primarily in moderate confidence ranges (50-80%), with near-perfect calibration (ECE = 0.0029) for high-confidence predictions

Use cases and applications

This corpus and the resulting model could be particularly valuable for:

• content moderation on multilingual platforms with Cyrillic scripts
• automatic translation systems requiring accurate source language identification
• digital humanities research involving Cyrillic language documents
• educational applications for learning about Cyrillic language diversity
• linguistic research on automatic classification of related languages

Limitations

• The texts are sourced from Wikipedia, which may create a bias toward a formal encyclopedic style
• Language mixtures are artificially created and may not perfectly represent natural code-switching situations
• Language representation reflects natural web availability, with some minority languages having fewer examples despite augmentation efforts
• Model performance may vary on texts significantly different from Wikipedia articles in style or domain

Ethical considerations

This corpus was created with particular attention to linguistic diversity and the preservation of minority languages. The inclusion of under-resourced Cyrillic languages reflects a commitment to linguistic equity in natural language processing tools. ll data was collected from publicly available Wikipedia articles, respecting copyright and attribution requirements.

Citations

If you use this corpus in your research, please cite it as follows: @misc{RassamCyrillic2025, author = {Rassam, Géraldine}, title = {Multi-Cyrillic Language Identification Dataset}, year = {2025}, publisher = {Hugging Face}, howpublished = {\url{https://huggingface.co/datasets/AlmaznayaGroza/cyrillic-language-classification}} }

License

This corpus is distributed under a CC BY-SA 4.0 license.

Note on metadata

Due to limitations of the Hugging Face interface, some information could not be correctly represented in the dataset metadata:

License: Although this corpus is registered under the MIT License in the metadata, the intention is that it be used and shared under the terms of the CC BY-SA 4.0 (Creative Commons Attribution-ShareAlike 4.0) license.

Languages: The metadata does not represent all the languages included in this corpus. In addition to the languages listed, this corpus also contains:

• Rusyn (rue)
• Komi (kv)
• Komi-Permyak (koi)
• Mari (mhr)
• Erzya (myv)
• Kabardinian (kbd)
• Abkhazian (ab) (Full list in the #Included Languages section)

Language mixtures: This corpus also contains examples of texts mixing several Cyrillic languages (identified by codes such as ru_uk_mix, ab_ce_mix, etc.), a specific characteristic that is not representable in standard metadata.