Datasets:
dataset_info:
features:
- name: text
dtype: string
- name: author
dtype: string
- name: title
dtype: string
- name: genre
dtype: string
- name: publisher
dtype: string
- name: year
dtype: string
splits:
- name: train
num_bytes: 120978838
num_examples: 55810
download_size: 53523399
dataset_size: 120978838
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
language:
- chm
- mhr
license: cc-by-4.0
Meadow Mari Corpus
Description
This is a text corpus of Meadow Mari language. The dataset is plain text without any additional annotations except for basic meta data which is somtetimes shifted a bit (possibly will be fixed in newer version). The genre field is in Meadow Mari, the table of contents could be found here.
Source
This corpus is a HF version of corpus by Andrei Chemyshev et al. So, if you plan on using this corpus, please cite their work:
@online{mari_monolingual_corpus,
title={Mari monolingual corpus},
author={Andrei Chemyshev and Gennadii Sabantsev and Nadezhda Timofeeva and Vasilii Semenov},
year={2023},
url={https://mari-lab.ru},
}
Composition
The corpus includes texts in various genres: prose, poetry, press, etc.
The most significant bias for this data is the dictionary data, specifically short one-sentence examples from «Марий йылме мутер» / «Словарь марийского языка» в 10 томах which is a 10 volume dictionary of Meadow Mari. That makes such examples 97.7% of the dataset entries and 6.95% of tokens.
To use this dataset as a balanced corpus, it is recommended to disable the dictionary examples. To achieve that,
disable entries with мутер in genre field (it is Meadow Mari for a dictionary).
Metrics
Length
| Total | Without Dictionary | |
|---|---|---|
| Texts | 55 810 | 1288 |
| Tokens* | 11 493 681 | 10 694 443 |
*nltk.tokenize.TweetTokenizer
Years
- First example: 1885
- Other examples: 1943-2021
Related Work
- This version of the corpus uses split by sentence (=each entry is a sentence).
- A parallel corpus of random Mari and Russian sentences can be found here.