Environment

Clone this repository and make sure the python path is set appropriately.

git clone https://github.com/igorsterner/csntax
cd csntax
export PYTHONPATH=$(pwd):$PYTHONPATH

Make sure you have python installed (we used version 3.11.8) and the required dependencies.

conda create -n myenv python=3.11.8
conda activate myenv
pip install -r requirements.txt

In order to process the sentences, we use the tools from the benchmark paper. Clone that repository too

git clone https://github.com/igorsterner/acs

CSntax-GNN

Reproduce our results with the following steps, or try them out online

First, generate the required translations, syntax graphs and alignments for each of the languages and splits desired. Note that you will need to have requested access to the benchmark data (here, approval is automatic) before you proceed.

python csntax/data.py --lang1 de --lang2 en --split train

Next, train the model (by default for three different seeds):

python csntax/train.py

(add binary flags --randomize_edge_features, --randomize_pos_tags, --randomize_lang_features or set --conv_type to gat for ablations)

Finally, you can evaluate on the trained language pair and other language pairs:

python csntax/inference.py

Baseline

Code for the finetuned XLM-R model is available in csntax/baseline/. Run it as follows:

python csntax/baseline/run.py --seed 77 --model_name xlm-roberta-base

Citation

The model is described in the following publication:

@inproceedings{sterner-2025-gnn,
    author = {Igor Sterner and Simone Teufel},
    title = {Code-Switching and Syntax: A Large–Scale Experiment},
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2025",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
}