split-words

CharSplit repacked with poetry, published as pypi split-words-- all credit goes to the original author.

pip install split-words

# replace charsplit with split_words in the following, e.g.
from split_words import Splitter
...

CharSplit - An ngram-based compound splitter for German

url

Splits a German compound into its body and head, e.g.

Autobahnraststätte -> Autobahn - Raststätte

Implementation of the method decribed in the appendix of the thesis:

Tuggener, Don (2016). Incremental Coreference Resolution for German. University of Zurich, Faculty of Arts.

TL;DR: The method calculates probabilities of ngrams occurring at the beginning, end and in the middle of words and identifies the most likely position for a split.

The method achieves ~95% accuracy for head detection on the Germanet compound test set.

A model is provided, trained on 1 Mio. German nouns from Wikipedia.

Usage

Train a new model:

training.py --input_file --output_file

from command line, where input_file contains one word (noun) per line and output_file is a json file with computed n-gram probabilities.

Compound splitting

In python

>> from charsplit import Splitter
>> splitter = Splitter()
>> splitter.split_compound("Autobahnraststätte")

returns a list of all possible splits, ranked by their score, e.g.

[(0.7945872450631273, 'Autobahn', 'Raststätte'),
(-0.7143290887876655, 'Auto', 'Bahnraststätte'),
(-1.1132332878581173, 'Autobahnrast', 'Stätte'), ...]

By default, Splitter uses the data from the file charsplit/ngram_probs.json. If you retrained the model, you may specify a custom file with

>> splitter = Splitter(ngram_path=<json_data_file_with_ngram_probs>)