Universal Dependencies treebank is based on data samples extracted from Taiga Corpus and MorphoRuEval-2017 and GramEval-2020 shared tasks collections.
UD Russian Taiga has been developed at the School of Linguistics, National Research University Higher School of Economics in Moscow (HSE/Vyshka). The selection of texts is meant to represent those registers that have not been covered by UD Russian SynTagRus and UD Russian Google Stanford Dependencies, mainly e-communication (blogs and social media). The sentences are extracted from two open data collections. Taiga Corpus (https://tatianashavrina.github.io/taiga_site/) is an open-source corpus for machine learning collected by students as part of the curriculum of the MA Program in Computational Linguistics at HSE. MorphoRuEval 2017 text collections (https://github.com/dialogue-evaluation/morphoRuEval-2017) is an output of the RuEval shared task 'Evaluation of Russian NLP: Morphological analysis, http://www.dialog-21.ru/en/evaluation/2017/morphology/). GramEval 2020 collection (https://competitions.codalab.org/competitions/22902)[https://competitions.codalab.org/competitions/22902] is an output of the GramEval 2020 Shared Task on Russian Full Morphology and Dependency Parsing which consists of test data for five genres (social, wiki, news, fiction, poetry).
The plain text data were tokenized, lemmatized and parsed using UDpipe (http://ufal.mff.cuni.cz/udpipe) and checked manually. Corrections were made at all levels: tokenization, lemmata, pos, features, dependency relations.
blogs and social media: 50%
- source: vk.com, instagram, facebook, twitter, youtube comments
- source: stihi.ru (naïve poetry), Corpus of Russian poetry (RNC)
- source: Zhurnalnyj zal (magazines.gorky.media), RNC main corpus
- main source: lenta.ru
- source: Russian wikipedia
- train: 68.1% (43.6K tokens, 3138 sentences)
- dev: 15.8% (10.1K tokens, 945 sentences)
- test: 16.1% (10.3K tokens, 881 sentences)
- train: 43.0% (16.6K tokens, 1302 sentences)
- dev: 26.1% (10.1K tokens, 945 sentences)
- test: 30.8% (11.9K tokens, 1017 sentences)
- train: 50% (10K tokens, 880 sentences)
- test: 50% (10K tokens, 884 sentences)
We are grateful to all the contributors to the original open Russian data collections and especially to Tatiana Shavrina (Taiga, GramEval-2020) and Alena Fenogenova (MorphoRuEval-2017).
Lyashevskaya, Olga, Kira Droganova, Daniel Zeman, Maria Alexeeva, Tatiana Gavrilova, Nina Mustafina, and Elena Shakurova. (2016). Universal Dependencies for Russian: a New Syntactic Dependencies Tagset. In: Series: Linguistics, WP BRP 44/LNG/2016. Moscow.
Sorokin, Andrey, Tatiana Shavrina, Olga Lyashevskaya, Victor Bocharov, Svetlana Alexeeva, Kira Droganova, Alena Fenogenova, and Dmitry Granovsky. (2017). MorphoRuEval-2017: an Evaluation Track for the Automatic Morphological Analysis Methods for Russian. In Computational Linguistics and Intellectual Technologies, Proceedings of Dialog 2017, Moscow. No 16 (23). Vol. 1, 297-313.
Lyashevskaya, Olga, Victor Bocharov, Alexey Sorokin, Tatiana Shavrina, Dmitry Granovsky, and Svetlana Alexeeva. (2017). Text collections for evaluation of Russian morphological taggers. Jazykovedny Casopis, 68 (2), 2017: 258-267.
Shavrina, Tatiana, Olga Shapovalova. (2017) To the methodology of corpus construction for machine learning: «Taiga» syntax tree corpus and parser. In Proceedings of the International Conference "CORPORA 2017", Saint-Petersbourg, Russia.
- GramEval2020 shared task data added (genre: news wiki social fiction poetry)
- UPOS, FEAT, HEAD, DEPREL validated, minor fixes in lemmas
- data validated, minor fixes
- Major changes
- UPOS, FEAT, DEPREL manually fixed
- New texts (genre: poetry social) added.
- Dev added, train expanded, UPOS, FEAT, HEAD, DEPREL updated in accordance with the UD2 guidelines, minor tokenization updates
- First official release.
=== Machine-readable metadata (DO NOT REMOVE!) ================================ Data available since: UD v2.2 License: CC BY-SA 4.0 Includes text: yes Genre: blog fiction news poetry social wiki Lemmas: manual native UPOS: manual native XPOS: manual native Features: manual native Relations: manual native Contributors: Lyashevskaya, Olga; Rudina, Olga; Zhuravleva, Anna Contributing: elsewhere Contact: firstname.lastname@example.org ===============================================================================