metadata

dataset_info:
  features:
    - name: src_lang
      dtype: string
    - name: src_sent
      dtype: string
    - name: tgt_lang
      dtype: string
    - name: tgt_sent
      dtype: string
  splits:
    - name: kaa_eng
      num_bytes: 19047157
      num_examples: 100000
    - name: kaa_rus
      num_bytes: 27731049
      num_examples: 100000
    - name: kaa_uzb
      num_bytes: 30608474
      num_examples: 100000
  download_size: 46148914
  dataset_size: 77386680
configs:
  - config_name: default
    data_files:
      - split: kaa_eng
        path: data/kaa_eng-*
      - split: kaa_rus
        path: data/kaa_rus-*
      - split: kaa_uzb
        path: data/kaa_uzb-*
language:
  - en
  - ru
  - uz
  - kaa
pretty_name: dilmash
size_categories:
  - 100K<n<1M
license: mit
task_categories:
  - translation
tags:
  - dilmash
  - karakalpak

Dilmash: Karakalpak Parallel Corpus

This repository contains a parallel corpus for the Karakalpak language, developed as part of the research paper "Open Language Data Initiative: Advancing Low-Resource Machine Translation for Karakalpak".

Dataset Description

The Karakalpak Parallel Corpus is a collection of 300,000 sentence pairs, designed to support machine translation tasks involving the Karakalpak language. It includes:

Uzbek-Karakalpak (100,000 pairs)
Russian-Karakalpak (100,000 pairs)
English-Karakalpak (100,000 pairs)

Usage

This dataset is intended for training and evaluating machine translation models involving the Karakalpak language.

To load and use dataset, run this script:

from datasets import load_dataset

dilmash_corpus = load_dataset("tahrirchi/dilmash")

Dataset Structure

Data Instances

Size of downloaded dataset files: 77.4 MB
Size of the generated dataset: 46.1 MB
Total amount of disk used: 123.5 MB

An example of 'kaa_eng' looks as follows.

{'src_lang': 'kaa_Latn',
 'src_sent': 'Pedagogikalıq ideal balaǵa ıktıyatlılıq penen katnasta bolıw principine bárqulla, úlken hám kishi jumıslarda súyeniwdi talan etedi.',
 'tgt_lang': 'eng_Latn',
 'tgt_sent': 'The ideal of education demands that the principle of treating children with care be observed at all times, in both big and small matters.'
}

Data Fields

The data fields are the same among all splits.

src_lang: a string feature that contains source language.
src_sent: a string feature that contains sentence in source language.
tgt_lang: a string feature that contains target language.
tgt_sent: a string feature that contains sentence in target language.

Data Splits

split_name	num_examples
kaa_eng	100000
kaa_rus	100000
kaa_uzb	100000

Data Sources

The corpus comprises diverse parallel texts sourced from multiple domains:

23% sentences from news sources
34% sentences from books (novels, non-fiction)
24% sentences from bilingual dictionaries
19% sentences from school textbooks

Additionally, 4,000 English-Karakalpak pairs were sourced from the Gatitos Project (Jones et al., 2023)[https://aclanthology.org/2023.emnlp-main.26].

Data Preparation

The data mining process involved local mining techniques, ensuring that parallel sentences were extracted from translations of the same book, document, or article. Sentence alignment was performed using LaBSE (Language-agnostic BERT Sentence Embedding) embeddings. For more information, plase refet to our paper.

Citation

If you use this dataset in your research, please cite our paper:

@misc{mamasaidov2024openlanguagedatainitiative,
      title={Open Language Data Initiative: Advancing Low-Resource Machine Translation for Karakalpak}, 
      author={Mukhammadsaid Mamasaidov and Abror Shopulatov},
      year={2024},
      eprint={2409.04269},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2409.04269}, 
}

Gratitude

We are thankful to these awesome organizations and people for helping to make it happen:

David Dalé: for advise throughout the process
Perizad Najimova: for expertise and assistance with the Karakalpak language
Nurlan Pirjanov: for expertise and assistance with the Karakalpak language
Atabek Murtazaev: for advise throughout the process
Ajiniyaz Nurniyazov: for advise throughout the process

We would also like to express our sincere appreciation to Google for Startups for generously sponsoring the compute resources necessary for our experiments. Their support has been instrumental in advancing our research in low-resource language machine translation.

Contacts

We believe that this work will enable and inspire all enthusiasts around the world to open the hidden beauty of low-resource languages, in particular Karakalpak.

For further development and issues about the dataset, please use m.mamasaidov@tahrirchi.uz or a.shopolatov@tahrirchi.uz to contact.