vonqo/mongolian-nlp

Useful resources for Mongolian NLP

This repo will contain a list of useful resources for Mongolian NLP. Feel free to contribute.

Datasets

• DATASET LJSpeech like male voice TTS dataset created from the Mongolian Bible
  • used in tugstugi/pytorch-dc-tts
  • use dl_and_preprop_dataset.py to download the audio files
• DATASET LJSpeech like Kalmyk (West Mongolian) female voice TTS dataset created from the Kalmyk Bible (2 hours)
• DATASET 300 hours Kalmyk synthetic STT dataset created by a voice conversion model
  • each WAV has a different text created from Kalmyk books
  • source voice is the Kalmyk Bible female TTS
  • target voices are from the VCTK dataset
  • an example WAV: https://twitter.com/tugstugi/status/1409111296897912835
• DATASET Eduge news classification dataset provided by Bolorsoft LLC
  • used to train the Eduge.mn production news classifier
  • 75K news with 9 categories: урлаг соёл, эдийн засаг, эрүүл мэнд, хууль, улс төр, спорт, технологи, боловсрол and байгал орчин
• DATASET 11-11.mn government agency complaint dataset
  • 80K with 5 categories: санал хүсэлт, гомдол, шүүмжлэл, талархал and өргөдөл
• DATASET online news corpus
  • 700 million words
• DATASET Digital Archive of Mongolian Newspapers 1990-1995 of the British Library
• Common Crawl Mongolian dataset
• opendata.burtgel.gov.mn
  • DATASET 220K Mongolian personal names
  • DATASET 90K Mongolian clan/family names
  • DATASET 192K Mongolian company names
• DATASET Mongolian provinces (aimags and sums) names
• DATASET 195 country (with capital cities) names in Mongolian
• DATASET 250 Mongolian most frequent words from Mongolian news, books and Wikipedia articles. (total 670M words / 2M unique words).
  • These words could be used also as the stop words.
• DATASET 500 Mongolian abbreviations
• DATASET Mongolian NER dataset created from Mongolian politics and sport news
  • 10K sentences annotated by tugstugi and enod using doccano
  • 4 categories LOCATION (6453/1753), PERSON (2839/1698), ORGANIZATION (4453/1970) and MISC (3716/2617)
• DATASET Mongolian POS dataset of the National University of Mongolia
  • 100k words
  • used POS tagsets
• DATASET Traditional Mongolian synthetic OCR dataset created from Mongolian song lyrics and dictionary
  • 80K images
  • without any data augmentation, for augmenting data use external libraries like albumentations.
• DATASET Handwritten Mongolian Cyrillic Characters Database of the Mongolian University of Science and Technology
  • 28x28 gray scale, 350k images
  • dataset description
• DATASET Mongolian Wordnet of the National University of Mongolia
  • 26875 words, 2979 glosses, 23665 synsets, 213 examples
• DATASET Mongolian Inflectional Morphology from UniMorph 4.0
  • 2085 lemmas and 14592 inflections (+ morpheme segmentations)
• DATASET Mongolian Derivational Morphology from MorphyNet
  • 1410 lemmas, 1629 derivations, and 229 derivational suffixes.
• DATASET Multilingual Spoken Words multilingual keyword spotting dataset
  • 2200 Mongolian keywords, 44000 audio files
  • example keywords: аав, байна, бэлдэж, дүрслэх, ламын, олов, сонирхож, түүний, хаанаас, хуулиар, чиглэсэн
• DATASET Small Kalmyk text corpus
  • newspaper, poetry etc.

Mongolian Text-to-Speech

• PYTORCH tugstugi/pytorch-dc-tts
  • DEMO Colab online demo
  • DATASET LJSpeech like male voice dataset created from the Mongolian Bible
• TF tugstugi/Tacotron-2 fork of Rayhane-mamah/Tacotron-2 adapted for the Mongolian Bible dataset
  • DEMO Colab online demo
  • DEMO speaker adaptation Colab online demo for the former Mongolian president Elbegdorj. The Tacotron model trained with the 5 hours Mongolian Bible dataset was fine tuned with a 10 minutes dataset created from a Elbegdorj's speech.
• PYTORCH Chimege TTS demo
  • 1x female
  • NVIDIA/tacotron2 + NVIDIA/waveglow
• DEMO HMM TTS online demo of the National University of Mongolia
  • 1x male and 2x female voices
• DEMO Yet another HMM? TTS online demo from “Мон Спийч Ай Ти” ХХК
  • demo server is currently down
  • 1x male and 1x female
  • female voice samples
• SAMPLES Tacotron2 TTS demo samples of Ikon.MN
  • 1x female (35h)
  • NVIDIA/tacotron2 + NVIDIA/waveglow
• DEMO HMM based TTS online demo of the Inner Mongolian university
  • 1x female
• DEMO MTL-Tacotron TTS demo samples of the Inner Mongolian university & National University of Singapore
  • 1x female
• TF ttslr/MonTTS Inner Mongolian TTS training code
  • SAMPLES Speech samples
  • DATASET SAMPLES MonSpeech of the Inner Mongolia University
  • dataset and pretrained models are not available
• PRODUCT NVDA/HTS screen reader developed by Innovation Development Center for the blind
  • 1x female (National University of Mongolia voice)
• PYTORCH/DEMO Kalmyk TTS demo Kalmyk is a Mongolic language spoken in Russia
  • dataset created from the Kalmyk Bible (2 hours)
  • NVIDIA/tacotron2 + NVIDIA/waveglow
• PYTORCH/DEMO Kalmyk TTS demo from Silero Kalmyk is a Mongolic language spoken in Russia
  • snakers4/silero-models

Mongolian Language Model

• MODEL 5-gram binary LM generated by KenLM on a 670M word dirty corpus.
  • it can be used either with mozilla/DeepSpeech: ./generate_trie alphabet.txt mn_5gram.binary trie
  • or in tugstugi/mongolian-speech-recognition
• TF / PYTORCH tugstugi/mongolian-bert pretrained Mongolian BERT models
  • trained by tugstugi, enod and sharavsambuu
  • nabar sponsored 5x TPUs.
• PYTORCH bayartsogt-ya/albert-mongolian pretrained Mongolian ALBERT
• PYTORCH robertritz/NLP ULMFiT experiments
• PYTORCH huggingface.co/bayartsogt/mongolian-gpt2 Mongolian GPT-2 model
• PYTORCH huggingface.co/bayartsogt/mongolian-roberta-base Mongolian Roberta base model

Mongolian Speech Recognition

• PYTORCH tugstugi/mongolian-speech-recognition
  • DEMO Chimege Speech Recognition
  • a proprietary dataset is used
• PRODUCT Chinese and traditional Mongolian voice input from aicloud.com
  • direct link to the APK file
  • seems to be working only for simple cases (or it works only for Southern Mongolian dialects...)
  • same system but for windows (according to someone, you have to register with a Chinese identity card to use it)
• DEMO Speech recognition of the Inner Mongolian university
  • seems to be non functional
• PRODUCT Huawei cloud ASR supports minority languages such as Mongolian, Tibetan, and Uyghur.
• PRODUCT Google Cloud Speech-to-text
  • 20% WER on a 3000 audio private test dataset
• PYTORCH Wav2Vec2 XLSR finetuned on Mongolian Common Voice
  • DEMO Colab online demo
  • 50% WER
• PYTORCH Wav2Vec2 XLSR trained on Kalmyk dataset
  • pretrained on 500 hours Kalmyk TV recordings and 1000 hours Mongolian speech recognition dataset
  • finetuned on 300 hours synthetic Kalmyk STT dataset created by voice conversion
  • 50% WER on a private test set created from Kalmyk TV recordnings, on clean voice recordings, it should have much lower WER
  • DEMO https://huggingface.co/tugstugi/wav2vec2-large-xlsr-53-kalmyk
• TF coqui.ai mongolian speech recognition trained on Mongolian CommonVoice
  • 90.08% WER

Mongolian Script

• DEMO Cyrillic to Mongolian script converter demo of the Inner Mongolian university
• DEMO Mongolian script OCR demo of the Inner Mongolian university
• PYTORCH tugstugi/bichig2cyrillic Mongolian script to (and back) cyrillic converter
  • DEMO Cyrillic to Mongolian Colab online demo
• PYTORCH tugstugi/image2bichig Traditional Mongolian OCR using CRNN
  • DEMO OCR Colab online demo
  • DATASET Traditional Mongolian synthetic OCR dataset

Mongolian Text Classification

• TF2 sharavsambuu/mongolian-text-classification
• SKLEARN / DEMO simple SVM Colab notebook classifying the Eduge dataset with around 91% accuracy.
  • SentencePiece model from tugstugi/mongolian-bert is used as the text tokenizer.

Mongolian Named Entity Recognition

• DATASET Mongolian NER dataset created from Mongolian politics and sport news
  • for more info see datasets
• PYTORCH enod/mongolian-bert-ner BERT based Mongolian NER
  • uses tugstugi/mongolian-bert Mongolian pre-trained BERT models
• DEMO NER demo of the National University of Mongolia

Misc

• PYTORCH tugstugi/forced_aligner Mongolian forced alignment tool using Rayhane-mamah/Tacotron-2 and readbeyond/aeneas
  • DEMO Colab online demo
• TF2 cyrillic transliteration Colab notebook sharavsambuu/cyrillic-mongolian-transliteration
• DATASET 1M back-translated MN->EN sentence dataset download link
  • sharavsambuu/english-mongolian-nmt-dataset-augmentation
• DICTIONARY Mongolian digitalized dictionaries from Center for Northeast Asian of the Tohoku University in Japan
  • for usage see Digitizing the Mongolian Language: An Introduction to the Polyglot “Online Dictionaries and Full-text Search of Mongolian Languages and Written Manchu”
  • it includes also IPA pronuncations for Mongolian words