OneAdder/mari-pdf-corpus

Mari Press and School Books Corpus

Description

This is a text corpus of Meadow Mari and Hill Mari languages. The dataset is plain text without any additional annotations except for basic meta data. Please note, that the texts have been extracted from PDFs, so they are quite dirty.

Source

The corpus is based on texts from Mari Lab PDFs of magazines and school books.

Composition

The first part of the corpus includes the following magazines:

name	names_in_corpus	years	languages
У вий	U_Viy_*	1926–1936	Meadow Mari with rare presence of Hill Mari
Пиалан илыш	пиалан_илыш_*	1926–1936	Meadow Mari
Родина верч	родина_верч_*	1942–1945	Meadow Mari
Марий альманах	марий_альманах_*	1946–1949	Meadow Mari
Ончыко	ончыко_*	1954-2022	Meadow Mari with occasional presence of Hill Mari
Марий сандылак	марий_сандалык_*	2008-2018	Meadow Mari with occasional presence of Hill Mari
У сэм	у_сэм_*	1930-1934	Hill Mari
Йакшар знамӹ	йакшар_знамӹ_*	1935-1936	Hill Mari
У сем	у_сем_*	1990-2018	Hill Mari

Note 1: У вий, Пиалан илыш, Родина верч, Марий альманах and Ончыко are names of the same magazine in different years.

Note 2: writing system had minor changes in 1938, if only modern writing is needed, filter by year field.

The second part includes school books on geography, mathematics and other subjects from between 1933-1951.

Preprocessing

Due to the fact that those are PDFs, a huge amount of preprocessing has been done to convert them to PDFs. At the beggining the texts have been split into four categories:

• clean: clean PDFs that have a text layer in UTF-8
• ocr: OCR PDFs have an automatically recognized text layer in UTF-8
• fix: broken encoding PDFs that have inconsistent ASCII encoding
• remap: slightly broken PDFs that have a text layer in UTF-8 but some charachters are incorrect

PDFs from first and second categories have been used as-is. PDFs from the third categories have been assumed to be in cp1251 and converted to UTF-8. For the third and fourth categories, the incorrect symbols have been remapped to the correct ones.

Filtering

Sometimes Mari magazines contain sections in Russian language. I have trained a logistic regression classifier on Meadow-Mari-Russian parallel corpus. It yields zero if a sentence is in Mari and 1 if in Russian. Then I classified each line in sections of PDF texts (separated be two line breaks) and calculated mean score for each section. Sections that are definitely Russian (>0.8) have been removed.

Metrics

	Total	Hill Mari only	Meadow Mari only	No OCR	Modern Writing
texts	878	176	563	166	701
tokens*	45 576 320	5 081 705	32 426 529	8 192 294	41 520 480

*nltk.tokenize.casual.TweetTokenizer not very reliable as there is a lot of rubbish from PDF

Cite

@online{mari_press_and_school_books_corpus,
  author = {Andrei Chemyshev and Michael Voronov},
  title = {Mari Press and School Books Corpus},
  year = 2025,
  url = {https://huggingface.co/datasets/OneAdder/mari-pdf-corpus},
  urldate = {2025-07-17}
}