KazLit-Parallel

KazLit-Parallel is a multilingual literary parallel corpus containing aligned text pairs in Kazakh, Russian, and English. The corpus is designed primarily for machine translation and multilingual NLP research involving Kazakh.

The current release contains 71,096 parallel text pairs across six translation directions.

Dataset Details

Dataset Description

KazLit-Parallel provides automatically aligned text segments derived from literary works available in Kazakh, Russian, and English.

The corpus was created to expand the availability of parallel data involving Kazakh, which remains comparatively underrepresented in multilingual NLP resources. In addition to Kazakh–Russian and Kazakh–English pairs, the dataset includes Russian–English parallel data derived from the same broader literary collection.

Potential applications include:

• machine translation involving Kazakh;
• multilingual representation learning;
• cross-lingual NLP;
• sentence-alignment research;
• computational literary studies;
• comparative analysis of translations across Kazakh, Russian, and English.

The corpus was constructed through automatic alignment and has not undergone exhaustive manual validation.

• Dataset name: KazLit-Parallel
• Full name: KazLit-Parallel: A Kazakh–Russian–English Literary Parallel Corpus
• Languages: Kazakh, Russian, English
• Scripts: Cyrillic for Kazakh and Russian; Latin for English
• Primary task: Machine translation
• Number of parallel pairs: 71,096
• License: CC BY 4.0, subject to the licensing note below

Dataset Structure

Each example contains a source text segment and its corresponding target-language segment.

Data Fields

• Language_pair: Translation direction in the format source_to_target, for example kaz_to_eng.
• X: Source-language text.
• y: Target-language text.
• X_lang: Language of the source text.
• y_lang: Language of the target text.
• Similarity: Similarity score associated with the aligned source–target pair.

Language Pairs

Translation direction	Number of examples
rus_to_eng	23,856
rus_to_kaz	19,832
eng_to_rus	15,690
eng_to_kaz	5,534
kaz_to_eng	3,884
kaz_to_rus	2,300
Total	71,096

The number of examples varies across translation directions, and this imbalance should be considered when constructing training and evaluation subsets.

Dataset Creation

Motivation

Parallel corpora involving Kazakh are substantially less abundant than comparable resources for higher-resource languages such as English and Russian.

KazLit-Parallel was created to provide a larger literary parallel resource involving Kazakh and to support research in machine translation, multilingual NLP, and cross-lingual text analysis.

Literary translations provide naturally occurring multilingual correspondences between texts while also exhibiting linguistic phenomena such as paraphrasing, restructuring, and non-literal translation. This makes the corpus relevant both for machine translation and for research on multilingual literary text.

Source Data

The source material consists of literary works and their available translations in Kazakh, Russian, and English.

Where multiple language versions of the same work were available, the texts were processed and aligned to identify corresponding multilingual segments.

Data Collection and Processing

The general construction pipeline consisted of:

1. collecting corresponding literary texts in Kazakh, Russian, and/or English;
2. extracting textual content;
3. preprocessing and normalizing the texts;
4. identifying corresponding multilingual works;
5. automatically aligning source and target segments;
6. computing similarity scores for candidate alignments; and
7. organizing the resulting pairs by translation direction.

The alignment process was performed automatically.

Uses

Intended Uses

KazLit-Parallel is intended as a research resource for:

• training and fine-tuning machine translation systems;
• multilingual language-model training and adaptation;
• Kazakh–Russian and Kazakh–English translation research;
• cross-lingual representation learning;
• sentence-alignment experiments;
• multilingual and comparative literary analysis;
• NLP research involving low-resource languages.

For evaluation settings requiring manually verified reference translations, researchers may wish to construct a validated subset of the corpus.

Out-of-Scope Uses

KazLit-Parallel is derived from literary texts and is therefore not intended to represent all domains or styles of contemporary language use.

Results obtained using this corpus should not automatically be assumed to generalize to domains such as legal, medical, technical, conversational, or news translation without additional evaluation.

Dataset Quality and Limitations

KazLit-Parallel was constructed using automatic text processing and alignment procedures. As a result, some alignment noise may be present.

Literary translation is also frequently non-literal. Translators may restructure sentences, paraphrase passages, combine or divide segments, or adapt expressions for the target language. Consequently, lexical similarity alone should not be interpreted as a definitive measure of translation quality.

The six translation directions are also unevenly represented, and the corpus is specific to the literary domain. These characteristics should be considered when designing experiments and interpreting results.

Licensing

KazLit-Parallel is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license for components of the dataset for which the dataset creator holds the applicable rights.

Under CC BY 4.0, these materials may be shared and adapted, including for commercial purposes, provided appropriate attribution is given.

Underlying Literary Works

The copyright status of the underlying literary works and translations is separate from the rights associated with the compilation, processing, alignment, and metadata of KazLit-Parallel.

Individual source works or translations may remain subject to copyright held by their respective authors, translators, publishers, or other rights holders. Users are responsible for ensuring that their use of the underlying material complies with applicable copyright law and licensing conditions.

Ethical Considerations

The dataset consists of previously published literary material rather than text collected directly from human participants.

Literary works may contain historically situated language, cultural assumptions, stereotypes, or other sensitive content originating from the source material. Models trained on the corpus may reproduce linguistic or cultural biases present in those texts.

Citation

If you use KazLit-Parallel in academic work, please cite the dataset:

@misc{kazlitparallel,
  title        = {KazLit-Parallel: A Kazakh--Russian--English Literary Parallel Corpus},
  author       = {Abdashim, Sagi},
  howpublished = {Hugging Face Datasets},
  year         = {2024},
  note         = {Dataset}
}