A New Aligned Simple German Corpus

This repository contains the data and code for a Simple German Corpus.

If you use the corpus in your research, please cite our paper

@article{toborek2022simplegerman,
  title={A New Aligned Simple German Corpus},
  author={Toborek, Vanessa and Busch, Moritz and Bo{\ss}ert, Malte and Bauckhage, Christian and Welke, Pascal}
}

About

The German language knows two versions of plain language: Einfache Sprache (Simple German) and Leichte Sprache, the latter being controlled by the Netzwerk Leichte Sprache.

Currently, there are only few works that build a parallel corpus between Simple German and German and the corresponding data is often not available. Such a (potentially expanded) corpus may be used to implement automatic machine learning translation from German to Simple German. While the data might currently not be sufficient, the goal of this work is to lay the foundations for such a corpus by:

1. Scraping websites with parallel versions for German and Simple German
2. Implementing various algorithms presented in the literature to form a corpus that contains aligned, "translated" sentences.

Installation

We recommend creating a virtual python environment, e.g. using anaconda

conda create -n simple-german python=3.10

and installing the required packages

conda activate simple-german
pip install -r requirements
python -m spacy download de_core_news_lg

Before using the repo, you must edit the file defaultvalues.py. Within it, you need to define the variable repository_location, the absolute path to the folder of this repository. E.g. dataset=/home/bob/Simple-German-Corpus. You can also change any of the other variables.

Usage

Please note, that downloading is throttled by a 5 second delay to reduce network traffic. You can change this in crawler/utilities.py

To run the whole code, simply setup the environment and run python main.py. This calls both main_crawler.py as well as main_matching.py. The crawler downloads the archived websites and parses all contents. The matcher calculates all corresponding match distances. Beware that the latter might take a lot of time, even though it is parallelized. The end result can be found in the results/ folder.

To run other code, i.e. evaluate.py in the evaluation folder, use python -m evaluation.evaluate. For manual alignment python -m evaluation.align_by_hand might come in handy. (This tool is by no means fully fleshed out)

Tools/Libraries

The code was written using python version 3.10.4.
All python packages needed to run the code can be found in the "requirements" file and installed using pip install -r requirements. We recommend using a virtualenv.

• spacy.io python library for natural language processing supporting a plethora of languages, including German
• BeautifulSoup4 python library for html parsing / crawling.