Mari NLP

Datasets

Corpora

• Meadow Mari monoligual corpus 1.4M texts, over 20M word occurrences, 19 genres.

• Meadow Mari corpora (web-corpora.net) web corpora with 5.53M word occurrences in the main monolingual corpus and 3.59M Mari and 15.11 Russian word occurrences in social media corpus (commentaries in VK).

• Mari-language Korp (Vienna) web corpora with 57.38M tokens in Meadow Mari corpus and 6.25M tokens in Hill Mari corpus.

• Hill Mari Corpus (tilda) Hill Mari corpus with 63522 word occurences in latin transcription.

• Tatoeba corpus of parallel sentences, 3869 pairs for Meadow Mari and Russian, 72 sentences for Hill Mari and Russian.

• Wiki-dumps hours of Meadow and Hill Mari audio with transcriptions.

Pretrained models

• UralicNLP pretrained morphological analysators/generators and lemmatisation for uralic languages. Includes Meadow Mari and Hill Mari.

• TartuNLP “Smugri” Multilingual Neural Machine Translation model for low-resource Finno-Ugric languages. Includes Meadow Mari (average to-lang scores: 8.51 bleau, 43.42 chrf, 38.76 chrf++), Hill Mari (average to-lang scores: 7.30 bleau, 40.81 chrf, 36.40 chrf++).

• Trained adapters on wikipedia corpus for Meadow Mari. BERT. XLM-R.

Word Similarity

• Fasttext Mari Word Embeddings. Meadow Mari. Hill Mari.

Methods/Software

• GiellaLT tools for building morphological analysers, proofing tools and dictionaries. Meadow Mari repository. Hill Mari repository.

• Apertium morphological analysis and generation, PoS-tagging.

Miscellaneous

• Mari-lab community