A lemmatizer for German language text
Germalemma lemmatizes Part-of-Speech-tagged German language words. To do so, it combines a large lemma dictionary (an excerpt of the TIGER corpus from the University of Stuttgart), functions from the CLiPS "Pattern" package, and an algorithm to split composita.
Easy option: Installing from PyPI via
You can install the package from PyPI via
pip install -U germalemma
Alternative option: Downloading and installing from source
Only do this if you don't install germalemma via pip:
In order to use GermaLemma, you will need to install some additional packages (see Requirements section below) and then download the TIGER corpus from the University of Stuttgart. You will need to use the CONLL09 format, not the XML format. The corpus is free to use for non-commercial purposes (see License Agreement).
Then, you should convert the corpus into pickle format for faster loading by executing
germalemma/__init__.py and passing the path to the corpus file in CONLL09 format:
python germalemma/__init__.py tiger_release_[...].conll09
This will place a
lemmata.pickle file in the
data directory which is then automatically loaded.
Part-of-Speech (POS) Tagging
You have set up GermaLemma to use the TIGER corpus (as explained above). You have tokenized your text (e.g. with NLTK). You have POS-tagged your tokens. Now you can use GermaLemma:
from germalemma import GermaLemma lemmatizer = GermaLemma() # passing the word and the POS tag ("N" for noun) lemma = lemmatizer.find_lemma('Feinstaubbelastungen', 'N') print(lemma) # -> lemma is "Feinstaubbelastung"
Valid POS tags
You can pass POS tags from the STTS tagset, however, only four POS tags can be processed:
- 'N...' (nouns)
- 'V...' (verbs)
- 'ADJ...' (adjectives)
- 'ADV...' (adverbs)
All other POS tags will result in a
ValueError so you should wrap the call to
find_lemma in a try-except block.
GermaLemma's accuracy was evaluated using a sample of 696 POS tagged and manually lemmatized words from a sample of paragraphs from proceedings of the European Parliament, Goethe's "Werther", Kafka's "Verwandlung" and a news article from the website of the WZB (see samples in folder "eval_texts").
Under the assumption that the POS tag is correct (only those words were selected), GermaLemma finds the correct lemma in 99.43% of the cases. For comparison, Pattern achieved 95.11% for the same sample.
- Python 3.6 or newer
- required package Pyphen
- optional package PatternLite (This package is optional but highly recommended as it boosts the lemmatizer's accuracy.)
Apache License 2.0. See LICENSE file.
The TIGER corpus is not part of this repository and has to be downloaded separately under separate license conditions.