README.md

metadata

annotations_creators:
  - found
  - machine-generated
language:
  - myv
  - ru
language_creators:
  - found
  - machine-generated
license:
  - cc-by-sa-4.0
multilinguality:
  - translation
pretty_name: Erzya-Russian parallel corpus
size_categories:
  - 10K<n<100K
source_datasets:
  - original
tags:
  - erzya
  - mordovian
task_categories:
  - translation
task_ids: []

Dataset Card for slone/myv_ru_2022

Dataset Description

• Homepage:
• Repository:
• Paper:
• Leaderboard:
• Point of Contact:

Dataset Summary

This is a corpus of parallel Erzya-Russian words, phrases and sentences, collected in the paper "The first neural machine translation system for the Erzya language".

The corpus consists of the following parts:

name	size	composition
train	74503	parallel words, phrases and sentences, mined from dictionaries, books and web texts
dev	1500	parallel sentences mined from books and web texts
test	1500	parallel sentences mined from books and web texts
mono	333651	Erzya sentences mined from books and web texts, translated to Russian by a neural model

The dev and test splits contain sentences from the following sources

name	size	description
wiki	600	Aligned sentences from linked Erzya and Russian Wikipedia articles
bible	400	Paired verses from the Bible (https://finugorbib.com)
games	250	Aligned sentences from the book "Сказовые формы мордовской литературы", И.И. Шеянова, 2017, НИИ гуманитарых наук при Правительстве Республики Мордовия, Саранск
tales	100	Aligned sentences from the book "Мордовские народные игры", В.С. Брыжинский, 2009, Мордовское книжное издательство, Саранск
fiction	100	Aligned sentences from modern Erzya prose and poetry (https://rus4all.ru/myv)
constitution	50	Aligned sentences from the Soviet 1938 constitution

To load the first three parts (train, validation and test), use the code:

from datasets import load_dataset
data = load_dataset('slone/myv_ru_2022')

To load all four parts (included the back-translated data), please specify the data files explicitly:

from datasets import load_dataset
data_extended = load_dataset(
    'slone/myv_ru_2022',
    data_files={'train':'train.jsonl', 'valiation': 'dev.jsonl', 'test': 'test.jsonl', 'mono': 'back_translated.jsonl'}
)

Supported Tasks and Leaderboards

• translation: the dataset may be used to train ru-myv translation models. There are no specific leaderboards for it yet, but if you feel like discussing it, welcome to the comments!

Languages

The main part of the dataset (train, dev and test) consists of "natural" Erzya and Russian sentences, translated to the other language by humans. There is also a larger Erzya-only part of the corpus (mono), translated to Russian automatically.

Dataset Structure

Data Instances

All data instances have three string fields: myv, ru and src (the last one is currently meaningful only for dev and test splits), for example:

{'myv': 'Сюкпря Пазонтень, кие кирвазтизе Титэнь седейс тынк кисэ секе жо бажамонть, кона палы минек седейсэяк!',
 'ru': 'Благодарение Богу, вложившему в сердце Титово такое усердие к вам.',
 'src': 'bible'}

Data Fields

• myv: the Erzya text (word, phrase, or sentence)
• ru: the corresponding Russian text
• src: the source of data (only for dev and test splits)

Data Splits

[More Information Needed]

Dataset Creation

Curation Rationale

[More Information Needed]

Source Data

Initial Data Collection and Normalization

[More Information Needed]

Who are the source language producers?

[More Information Needed]

Annotations

Annotation process

[More Information Needed]

Who are the annotators?

No human annotation was involved in data collection.

Personal and Sensitive Information

All data was collected from public sources, so no sensitive information is expected in them. However, some sentences collected, for example, from news articles or LiveJournal posts, can contain personal data.

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

Most of the dataset has been collected by automatical means, so it may contain errors and noise. Some types of these errors are systemic: for example, the words for "Erzya" and "Russian" are often aligned together, because they appear in the corresponding Wikipedias on similar positions.

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

[More Information Needed]

Citation Information

[More Information Needed]