spaCy: Polish language pipeline and models

Where to get it

The latest versions of the models are available here: http://zil.ipipan.waw.pl/SpacyPL

Requirements

Spacy in a version >= 2.2.

Installation

First, you need to install spaCy (we need versions >= 2.2). Please refer to the official documentation to do so.

For example, using Anaconda:

conda install -c conda-forge spacy

Then, after downloading the package, install it as any other python module:

python -m pip install PATH/TO/pl_spacy_model-x.x.x.tar.gz

If you can install the Morfeusz 2 bindings for Python, we reccomend using the pl_spacy_model_morfeusz version of our model, which is superior in performance. For details, please see the pl_spacy_model_morfeusz section below.

python -m pip install PATH/TO/pl_spacy_model_morfeusz-x.x.x.tar.gz

Quick start

import spacy
nlp = spacy.load('pl_spacy_model') # or spacy.load('pl_spacy_model_morfeusz')

# List the tokens including their lemmas and POS tags
doc = nlp("Granice mojego języka oznaczają granice mojego świata") # ~Wittgenstein
for token in doc:
    print(token.text, token.lemma_, token.tag_)

A more complete example

Please see this Jupyter notebook

or this poster and presentation.

What is included?

Lemmatizer

The lemmatizer is implemented as a look-up table, using a lemma dictionary imported from the Morfeusz morphological analyzer.

Tagger

The tagger has been trained on a corpus consisting of the 1 million word subcurpous of the [National Corpus of Polish](http://clip.ipipan.waw.pl/NationalCorpusOfPolish} and the 500k Frequency Corpus of the 1960s Polish language. For tasks involving Polish language only, we reccomend using the internal tagset (token.tag_ as opposed to token.pos_), because the latter is a lossy mapping of the former.

Depenendency Parser

For training a dependency parser, we've used the PDB UD treebank

Named Entity Recognizer

NER model has been trained on the 1 million word subcurpous of the National Corpus of Polish.

Flexer

The morfeusz-based model additionally incorporates a custom inflection component. It is deterministic and works by consulting both the tagger and the dictionaries of morfeusz. For details please see the notebook.

Word embeddings

Word embeddings trained on KGR10 corpus (over 4 billion of words) using Fasttext by Jan Kocoń and Michał Gawor (https://clarin-pl.eu/dspace/handle/11321/606). Our model uses only the vector representation for 800.000 most frequent words.

Please see this Jupyter notebook for a demo.

pl_spacy_model_morfeusz

This version of our model utilizes an external dependency: Morfeusz 2. We reccomend using this version, as it is significantly better for all the tasks involved. To do so you need to first install the Morfeusz 2 library, and its embeddings for Python. The detailed instructions for your architecture are available here: Morfeusz 2 bindings are installed via easy_install, and not via pip, although we are discussing the option to switch onto pip into the future. If Morfeusz 2 is not installed correctly, you will see a warning message and the model will not work as expected.

Morfeusz 2 is used within our custom pipeline component called Preprocessor. The Preprocessor first tokenizes the text, and then performs morphosyntactic analysis, and lemmatization. Since Morfeusz 2 offers multiple analyses for each token, we disambiguate these using our tagger. From 0.1.0 onwards the tagger is a integrated version of Toygger (for more info see the bibliography below). Because of this there is no separate tagger component in our pipeline (although there is a tagger in the model), and you cannot skip tagging during the processing stage.

In this version of the model, the token.tag_ attribute returns the POS tag only. All the morphological features are stored in token._.feats custom attribute.

This notebook shows some features of the model with Morfeusz 2.

Evaluation metrics

Both the poster and the article list the evaluation scores for 0.0.3 version, for up to date results please see the evaluation folder.

Change history

0.1.0
reduced the size of the model
added the flexer component
added the capabilities for full morphosyntactic analysis using Toygger
0.0.5 -- several changes
- corrected errors in NER model,
- parser now performs sentencization correctly (you can parse whole documents now, instead of individual sentences)
- parser trained on PDB treebank (instead of LFG),
- added version using Morfeusz library as an external dependency; more detailed morphosyntactic tags are available in token.tag_ attribute.
0.0.3 -- added support for spaCy 2.2

Authors

Ryszard Tuora

supervision: Łukasz Kobyliński

Citing

Ryszard Tuora and Łukasz Kobyliński, "Integrating Polish Language Tools and Resources in spaCy". In: Proceedings of PP-RAI'2019 Conference, 16-18.10.2019, Wrocław, Poland.

Bibliography

For more info about Toygger, see here:

Katarzyna Krasnowska-Kieraś. Morphosyntactic disambiguation for Polish with bi-LSTM neural networks. In Zygmunt Vetulani and Patrick Paroubek, editors, Proceedings of the 8th Language & Technology Conference: Human Language Technologies as a Challenge for Computer Science and Linguistics, pages 367–371, Poznań, Poland, 2017. Fundacja Uniwersytetu im. Adama Mickiewicza w Poznaniu.

Cogitarian/spacy-pl