aizana1/Tuvan-VK-Corpus

Tuvan VK Corpus

Public posts and comments from 27 VKontakte communities related to the Republic of Tyva (Tuva, Russia), 2013–2025. About 350K non-empty texts in Tuvan, Russian, and a mix of both.

Посты и комментарии из 27 сообществ ВКонтакте, связанных с Республикой Тыва (2013–2025): около 350 тыс. текстов на тувинском, русском и смешанных.

The corpus was collected to study how Tuvan is actually written online: how often people write in Tuvan at all, and how often they replace the Tuvan letters ң, ө, ү with н, о, у because of the Russian keyboard layout.

Subsets

Each of the three source categories is split into a posts table and a comments table.

Config	Rows	Period
official_media_posts	12,166	2024-05 – 2025-12
official_media_comments	8,397	2024-05 – 2025-12
gov_institutions_posts	48,478	2013-02 – 2025-12
gov_institutions_comments	49,256	2013-03 – 2025-12
community_media_posts	42,644	2018-01 – 2025-12
community_media_comments	189,636	2018-02 – 2025-12
Total	350,577

from datasets import load_dataset

comments = load_dataset("aizana1/tuvan-vk-corpus", "community_media_comments", split="train")

Source communities

Official media (3): shyn31081925 (newspaper Шын, published in Tuvan), gtrktuva (GTRK Tyva), tvtyva (TV channel Tuva 24).

Government institutions (14): gov.tuva (Government of the Republic of Tyva), mintrud_tuva (labour and social policy), minobrtuva (education), tuvaminzdrav (health), tuvacult (culture), minfinrt (finance), minjust.tuva (justice), molodej_tuva (youth), mincifrart (digital development), minselkhoz_tuva (agriculture), minles17 (forestry), minsportrt (sport), sud17rf (courts), police17 (police).

Community media (10): tuvasansara, tuva_komu_za_30, tovsp1, read_kzl, kraida_tyvalar, tuva_popsa, discussion_tuva, setkil_001, chs17rus, podslushano17rus.

Fields

Posts

Field	Type	Description
post_id	int64	VK post ID, unique within a group
group	string	VK short name of the community
category	string	official_media, gov_institutions or community_media
text	string	Post text, personal data masked (see below)
lang_heuristic	string	tyv, tyv_ru_layout or ru (see below)
likes	int64	Number of likes at collection time
comments_count	int64	Number of comments at collection time
date	timestamp (UTC)	Publication time
year	int16	Publication year

Comments

Field	Type	Description
comment_id	int64	VK comment ID, unique within a group
post_id	int64	Parent post; join with the posts table on (group, post_id)
group, category, text, lang_heuristic, likes, date, year		As above

Language label

lang_heuristic is a rule-based, noisy label, not a gold annotation:

• tyv — the text contains at least one of ң, ө, ү. Reliable for "contains Tuvan", but a mostly Russian text with one Tuvan name also gets this label.
• tyv_ru_layout — no Tuvan letters, but at least one word or ending from a list of typical Tuvan patterns (e.g. мен, чок, -нын, оо-). This is meant to catch Tuvan typed on a Russian keyboard. It works reasonably on comments, but has many false positives on Russian texts: Tuvan personal names (Ооржак, -оолович) and Russian words ending in -зы, -иг, -иске (базы, подвиг, списке). In gov_institutions_posts most texts with this label are Russian.
• ru — everything else (Russian or other).

Config	tyv	tyv_ru_layout	ru
official_media_posts	6,723	564	4,879
official_media_comments	750	2,357	5,290
gov_institutions_posts	1,769	10,331	36,378
gov_institutions_comments	1,567	8,112	39,577
community_media_posts	3,326	13,329	25,989
community_media_comments	6,026	41,234	142,376

Collection and processing

• Collected in December 2025 with the VK API (wall.get, wall.getComments) from public communities. Collector code: tuvan-lang-in-vk-posts.
• Comments are a sample per post (the collector requested about 10 top-level comments per post), not all comments. Replies in threads are not included.
• No author identifiers were collected.
• Rows with empty text (media-only posts, sticker comments) were removed: 22,572 posts and 33,886 comments.
• 29 duplicate posts and 22 duplicate comments (same ID collected twice) were removed.
• Masking in text: user mentions [id…|Name] → @user; links to user profiles → <vk_user>; phone numbers → <phone>; e-mail addresses → <email>. Mentions of communities ([club…|Name]) are replaced by the community name.
• date is computed from the Unix timestamp in UTC.

The processing script is in scripts/prepare.py.

Limitations

• Names of people can still appear in plain text (e.g. a comment addressing someone by name).
• Comments are not a complete thread and are sampled per post.
• Spam and copy-pasted comments are not removed.
• The language label is heuristic (see above).

Terms of use

The texts were published by their authors on VKontakte and remain their property. The dataset is shared for non-commercial research on the Tuvan language. Do not use it to identify, profile or contact individual users. If you are an author of a text in this dataset and want it removed, please open a discussion on this page.

Citation

To be added.