metadata

dataset_info:
  features:
    - name: src_sent_id
      dtype: float64
    - name: src_sent
      dtype: string
    - name: tgt_sent_id
      dtype: float64
    - name: tgt_sent
      dtype: string
    - name: sim
      dtype: float64
    - name: sim_pnlz
      dtype: float64
    - name: src_doc_hash
      dtype: string
    - name: tgt_doc_hash
      dtype: string
    - name: docs_sim
      dtype: float64
    - name: src_id
      dtype: int64
  splits:
    - name: train
      num_bytes: 39447584
      num_examples: 76400
  download_size: 15646643
  dataset_size: 39447584
license: cc-by-4.0
task_categories:
  - translation
language:
  - ru
  - myv
size_categories:
  - 10K<n<100K

"e-mordovia-articles-2023": a parallel Russian-Erzya news dataset

This is a semi-aligned dataset of Erzya and Russian news articles, crawled from https://www.e-mordovia.ru.

Dataset Description

Dataset Summary

This is a dataset of news arcticles collected from https://www.e-mordovia.ru, the official portal of the state authorities of the Republic of Mordovia.

The Russian and Erzya articles have been paired using heuristics, then split into sentences and automatically aligned on the sentence level.

The dataset contains unaligned sentences and sentnces that were aligned poorly. To extract a relatively high-quality parallel part, it is recommended to filter the sentence pairs by sim>=0.5.

Supported Tasks and Leaderboards

The dataset is intended to be used as training data for machine translation models.

Languages

The languages are Erzya (myv) and Russian (rus). Erzya, a language from the Mordvinic branch of the Uralic family, is one of the three official languages in Mordovia, alongside with Russian and Moksha.

Both languages are written in the Cyrillic script.

Dataset Structure

Data Instances

A typical datapoint is a pair of Erzya and Russian sentences (if they were aligned) or a single sentence in one of the languages (if the algorithm failed to align them).

Each sentence is associated with a document (a news article).

A typical aligned instance may look like

{'src_sent_id': 0.0,
 'src_sent': 'Мордовиянь Прявтось Артём Здунов таштамковонь 30-це чистэ важоди Пятигорскойсэ, косо неть читнестэ юты «Больше, чем путешествие» Весероссиянь туристической промксось.',
 'tgt_sent_id': 0.0,
 'tgt_sent': 'Глава Мордовии Артем Здунов 30 сентября работает в Пятигорске, где в эти дни проходит Всероссийский туристический слёт «Больше, чем путешествие» .',
 'sim': 0.8276803313967693,
 'sim_pnlz': 0.4310691599466438,
 'src_doc_hash': '31c1ff1900eb69d9',
 'tgt_doc_hash': 'c3270fd2c8ccdfa9',
 'docs_sim': 0.762437117256569,
 'src_id': 0}

An unaligned instance may look like

{'src_sent_id': None,
 'src_sent': None,
 'tgt_sent_id': 14.0,
 'tgt_sent': 'Только в такой связке работа будет эффективной.',
 'sim': None,
 'sim_pnlz': None,
 'src_doc_hash': '31c1ff1900eb69d9',
 'tgt_doc_hash': 'c3270fd2c8ccdfa9',
 'docs_sim': 0.762437117256569,
 'src_id': 0}

Data Fields

src_sent_id: id of the sentence in the Erzya article (or empty)
src_sent: the Erzya sentence (or empty)
`tgt_sent_id': id of the sentence in the Russian article (or empty)
tgt_sent: the Russian sentence (or empty)
sim: similarity of the Russian and Erzya sentences (or empty): a product of their LaBSE cosine similarities and their shortest-to-longest ratio of character lengths.
sim_pnlz: a penalized similarity of the Russian and Erzya sentences (or empty); based on whether this number is positive, the decision was made to align the sentences or not.
src_doc_hash: unique identifier of the Erzya article
tgt_doc_hash: unique identifier of the Russian article
docs_sim: the similiarity between the documents, computed as the aggregated similarity of the individual sentences in them.
src_id: numeric id of the document pair

Data Splits

The whole dataset is a train split. In the future, we may create an additional dev split out of the articles published after the current data has been collected.

Dataset Creation

Curation Rationale

The dataset has been curated in order to boost the quality of machine translation for Erzya.

Source Data

Initial Data Collection and Normalization

The data has been scraped from the https://www.e-mordovia.ru website.

The Erzya and Russian articles were considered as a possible pair if their included the same image. Then each candidate pair of articles was split into sentences with the razdel Python package and aligned using the slone_nmt package and the slone/LaBSE-en-ru-myv-v2 sentence encoder.

No other text preprocessing has been performed.

Who are the source language producers?

The language producers are the writers and the editors of the e-mordovia.ru portal, typically anonymous.

Annotations

The dataset does not contain any additional annotations.

Personal and Sensitive Information

The dataset may contain personal names of the people figurating in the news. However, as all these news are public anyway, we do not see any additional risks in using this data for training NLP models.

Considerations for Using the Data

Social Impact of Dataset

We hope that by enabling better machine translation from and to Erzya, this dataset would contribute to improve the digital presence and overall prestige of the Erzya language.

Discussion of Biases

As a portal of a public authority in Russian Federation, a country not very famous for its freedom of speech, e-mordovia may be presenting its own particular coverage and interpretation of some internal, federal and international events. Therefore, while we recommend using the collected dataset for teaching the Erzya language to models (and maybe even to humans), its usage as a source of non-linguistic knowledge should be approached with extreme caution.

Other Known Limitations

The original articles were not originally intended to serve as parallel texts, and the writers who translated them between Erzya and Russian might have altered the content someitmes. Our automatic metrics of sentence similarity were designed to filter out such cases, these metrics are error-prone themseles. Therefore, whereas the aligned part of the dataset seems to be highly parallel, the translations should be treated as noisy rather than 100% accurate. Filtering by the sim and doc_sim fields is recommended for obtaining a highly parallel subset.

Additional Information

Dataset Curators

The dataset was collected and aligned by Sergey Kuldin, Isai Gordeev and David Dale.

Licensing Information

According to the note on the source website (https://www.e-mordovia.ru/podderzhka-portala/ob-ispolzovanii-informatsii-sayta/):

All materials of the official website of the state authorities of the Republic of Mordovia can be reproduced in any media, on Internet servers or on any other media without any restrictions on the volume and timing of publication. This permission applies equally to newspapers, magazines, radio stations, TV channels, websites and Internet pages. The only condition for reprinting and retransmission is a link to the original source. No prior consent to reprint is required from the editorial board of the official website of the state authorities of the Republic of Mordovia.

Here is the original note in Russian:

Все материалы официального сайта органов государственной власти Республики Мордовия могут быть воспроизведены в любых средствах массовой информации, на серверах сети Интернет или на любых иных носителях без каких-либо ограничений по объему и срокам публикации. Это разрешение в равной степени распространяется на газеты, журналы, радиостанции, телеканалы, сайты и страницы сети Интернет. Единственным условием перепечатки и ретрансляции является ссылка на первоисточник. Никакого предварительного согласия на перепечатку со стороны редакции официального сайта органов государственной власти Республики Мордовия не требуется.

In short, you can use and even redistribute the data, but when redistributing, you must give a reference to the original website, https://www.e-mordovia.ru.

Based on this note, we concluded that a CC-BY license would be appropritate for this dataset.

Citation Information

[TBD]