Overcoming Language Barriers: Multilingual Analysis of the 2023 Swiss Privacy Law's Impact
Contributors
Contact person:
Data manager (2):
-
1.
ETH Zurich
- 2. Information Processing and Telecommunications Center, Universidad Politécnica de Madrid
-
3.
Carnegie Mellon University
Description
This repository contains the complete research pipeline for automatically analyzing privacy policies in a multilingual setting (validated for English, German, French, and Italian). The project operationalizes this pipeline in the context of (1) a revision in Swiss privacy law and (2) the use of automated policy generators.
Repository Structure
This repository is organized into three main subdirectories, each serving a specific purpose in the research pipeline:
1. Analysis
The Analysis directory contains most statistical analyses and data processing scripts.
Key Components:
- Data preprocessing (including text cleaning and removal of personal identifiers with Presidio)
- Creation of the three relevant groups for the analysis (CH, CH & EU, and EU)
- Main statistical and semantic analysis scripts presented in the paper
- Analysis of policy clusters based on generator use
2. Data
The Data directory contains the relevant datasets and corpora used throughout the paper.
Key Components:
- Annotated datasets:
1. LLM-annotated full original dataset after removal of personal identifiers with Presidio ("swiss-gdpr_annotated.parquet")
2. Final annotated and grouped dataset used for all statistical analyses ("swiss-gdpr_annotated_groups.parquet")
3. Log of the LLM annotations ("run.log")
- CrUX dataset used for website popularity rankings as well as the website budgeting list used to scrape the initial dataset
- Embeddings of the policies used for the cluster analysis
3. LLM
The LLM directory contains all files related to the LLM-based data analysis.
Key Components:
- The codebooks, human annotations (reference benchmark), and evaluations for all three initial annotation phases ("Annotations")
- The validation of the models against the final set of human annotations ("Validation")
- The scripts for the large-scale policy evaluation using OpenAI's GPT-5 ("Evaluation")
Citation
If you use this work in your research, please cite it as:
```
Accepted at PETS '26 [Citation details to be added upon publication]
```
We kindly ask you to cite the paper and not the dataset itself. Please find a more detailed list of funding sources in the paper's Acknowledgments section.
License
This work and its artifacts are licensed under a CC-BY 4.0 license.
Contact
For questions about this research, please contact Luka Nenadic at lnenadic@ethz.ch.
Notes
Artifact Appendix
Paper title: Overcoming Language Barriers: Multilingual Analysis of the 2023 Swiss Privacy Law's Impact
Requested Badge(s):
- [X] Available
- [ ] Functional
- [ ] Reproduced
Description
This is the artifact repository for the paper "Overcoming Language Barriers: Multilingual Analysis of the 2023 Swiss Privacy Law's Impact" published at PETS 2026.4. The authors are Luka Nenadic (ETH Zurich), David Rodriguez (Information Processing and Telecommunications Center, Universidad Politécnica de Madrid), and Joseph Calandrino (Carnegie Mellon University).
The repository contains all datasets (after the removal of personal identifiers with Presidio) as well as the statistical and LLM-based analyses used for the paper.
Security/Privacy Issues and Ethical Concerns
Given that we (1) have removed personal identifiers with Presidio and (2) do not use any artifacts desabling security mechanisms, run vulnerable code, or collect user data, our artifacts pose only minimal security and privacy risks. It remains, nonetheless, likely that Presidio could not successfully remove all personal identifiers from the publicly accessible privacy policies we use in our paper.
Environment
The entire artifact repository is hosted on Zenodo (DOI: 10.5281/zenodo.20512192) and can be downloaded as a .ZIP file. We explain the repository's structure above.
All analyses can be performed on a personal computer, except for the LLM-based analyses, for which we comprehensively describe how we used OpenAI's API. The statistical analyses were performed with Python and R (version 4.5.0 in R Studio version 2026.01.0).
Accessibility
The entire artifact repository is hosted on Zenodo (DOI: 10.5281/zenodo.20512192) and can be downloaded as a .ZIP file. We explain the repository's structure above.
Notes on Reusability
While we have validated our multilingual disclosure-assessment method for English, German, French, and Italian, it could be useful in different languages as well (upon careful validation in the target languages).
Finally, and as we note in the paper, the LLM-based policy analysis cannot guarantee perfect reproducibility, since model outputs may vary slightly across runs.
Files
Artifacts.zip
Additional details
Funding
- Swiss National Science Foundation
- Digital Market Regulation: Monitoring Enforcement and Evaluating Outcomes 10002634
- Ministerio de Ciencia, Innovación y Universidades
- Re-InitS PID2024-155230OB-C43
Dates
- Accepted
-
2026-06-01PETS '26