awesome-nlp-polish

A curated list of resources dedicated to Natural Language Processing (NLP) in polish. Models, tools, datasets.

Table of Contents:

• Polish text data
• Models and embeddings
• Libraries and tools
• Papers, articles, blogs
• Contribution

Polish text datasets

Task oriented datsets

• The KLEJ (Kompleksowa Lista Ewaluacji Językowych) benchmark is a set of nine evaluation tasks for the Polish language understanding.
• PolEval datasets -
  • Hate speech classification -distinguish between normal/non-harmful tweets (class: 0) and tweets that contain any kind of harmful information (class: 1) [PolEval 2019 Task6] [mirror GDrive]
• Polish CDSCorpus - The dataset for compositional distributional semantics. Polish CDSCorpus consists of 10K Polish sentence pairs which are human-annotated for semantic relatedness and entailment.
• Wroclaw Corpus of Consumer Reviews Sentiment (WCCRS) - corpus of Polish reviews annotated with sentiment at the level of the whole text (text) and at the level of sentences (sentence) for the following domains: hotels, medicine, products and university (reviews*)
• Ermlab Opineo dataset- opineo reviews - GDrive
• HateSpeech corpus contains over 2000 posts crawled from public Polish web.http://zil.ipipan.waw.pl/HateSpeech
• Polish analogy dataset - example: "Ateny Grecja Bagdad Irak" - useful for word embeddings evaluation
• NKJP - National Corpus of Polish. It contains classic literature, daily newspapers, specialist periodicals and journals, transcripts of conversations, and a variety of short-lived and internet texts. Only a small sub-corpus is available for download (GNU GLP v.3). Direct contact and maybe necessary to get the full corpus.
• PolEmo 2.0 Sentiment Analysis Dataset for CoNLL
• Polish Music Dataset- Polish Music Dataset is the largest dataset with information about artists, songs and lyrics in Poland (now only Hip Hop artists).

Raw texts

• Clean Polish OSCAR - preprosessed polish oscar corpus, removed: foreign sentences(non-polish), non-valid polish senteces (eg. enums), corpus preprocessed by @Ermlab

• OSCAR or Open Super-large Crawled ALMAnaCH coRpus - is a huge multilingual corpus obtained by language classification and filtering of the Common Crawl corpus. Contains 109GB or 49GB of polish text.

• Polish Wikipedia dump - regular monthly copy of Polish wikipedia. More then 4GB of text.

• Opus - the open parallel corpus - you can select languages and download only polish file

  • Polish OpenSubtitles v2018 - sentences 45.9M, polish tokens 287.1M ,collection of translated movie subtitles from opensubtitles raw txt corpus (unpacked 7.2GB) tokenized txt corpus (unpacked 7.6GB).
  • ParaCrawl v5 sentences 6.4M, polish tokens 157.1M raw txt corpus (unpacked 1.1GB) tokenized txt corpus

• Polish Parliamentary Corpus text from proceedings of Polish Parliament, Sejm and Senate

Models and Embeddings

Polish Transformer models

• Polish Roberta Model - model was trained on a corpus consisting of Polish Wikipedia dump, Polish books and articles, Polish Parliamentary Corpus
• PoLitBert - Polish RoBERTA model trained on Polish Wikipedia, Polish literature and Oscar. Major assumption is that quality text will give good model.
• PolBert - Polish BERT model. Model was trained with code provided in Google BERT's github repository. Merge with huggingface/Transformers
• Allegro HerBERT - Polish BERT model trained on Polish Corpora using only MLM objective with dynamic masking of whole words.
• SlavicBert - multilingual BERT model -BERT, Slavic Cased: 4 languages(Bulgarian,Czech, Polish, Russian), 12-layer, 768-hidden, 12-heads, 110M parameters, 600Mb. There is also another SlavicBert model http://docs.deeppavlov.ai/en/master/features/models/bert.html but I have problems to convert it to pytorch.

Other models

• ELMO embeddings - A model of ELMo embeddings for Polish language trained on large textual corpora (KGR10).

• Zalando Flair polish models - Contextual string embeddings that capture latent syntactic-semantic information that goes beyond standard word embeddings. There are two models "pl-forward and pl-backward"

• IPIPAN Word2vec polish models

• Wrocław University of Science and Technology Word2Vec - Distributional language models for Polish trained on different corpora (KGR10, NKJP, Wikipedia).

• FastText polish model FB - train on: Common Crawl, Wikipedia

• FastText KGR10 polish model binary

• Universal Sentence Encoder Multilingual - sentence embeddings, it covers 16 languages (including Polish)

• BPEmb: Subword Embeddings includes polish - easy to use with Flair

Language processing tools and libraries

• Morfologik (Java) and pyMorfologik (Python wrapper) - dictionary-based morphological analyzer

• Morfeusz - morphological analyzer. See also Elasticsearch plugin

• Stempel (Python port) - algorithmic stemmer. See also Elasticsearch plugin

• spaCy for Polish - extend spaCy, a popular production-ready NLP library, to fully support Polish language.

• spacy-pl by IPI PAN - integrating existing Polish language tools and resources into the spaCy pipeline

• KRNNT Polish morphological tagger - KRNNT is a morphological tagger for Polish based on recurrent neural networks Paper

• Stanza (Python) - NLP analysis package from Stanford University. Stanza is a Python natural language analysis package. It contains tools, which can be used for: sentence/word tokenizing, to generate base forms of words, parts of speech and morphological features, syntactic dependency parsing, recognizing named entities. Contains Polish model

• A curated list of Polish abbreviations for NLTK sentence tokenizer based on Wikipedia text

Papers, articles, blog post

• Benchmarks of some of polish NLP tools - Single-word lemmatization and morphological analysis, Multi-word lemmatization,Disambiguated POS tagging, Dependency parsing, Shallow parsing, Named entity recognition, Summarization etc.
• Github Repo with list of polish: word embeddings and language models (Word2vec, fasttext, Glove, Elmo) - https://github.com/sdadas/polish-nlp-resources
• Polish Word Embeddings Review - Evaluation of polish word embeddings: word2vec, fastext etc. prepared by various research groups. Evaluation is done by words analogy task.
• Polish Sentence Evaluation- contains evaluation of eight sentence representation methods (Word2Vec, GloVe, FastText, ELMo, Flair, BERT, LASER, USE) on five polish linguistic tasks
• TRAINING ROBERTA FROM SCRATCH - THE MISSING GUIDE - complete user guide for trainning Roberta model with use of Huggingface/Transformers for polish

Contribution

If you have or know valuable materials (datasets, models, posts, articles) that are missing here, please feel free to edit and submit a pull request. You can also send me a note on LinkedIn or via email:[email protected].