SwissGov-RSD: Detecting Cross-Lingual Semantic Differences in Swiss Government Websites

🎯 Challenge Overview

Domain: Public Sector / Multilingual Government Content / NLP

→ Context & Background

Swiss government websites are published in the official Swiss languages and English, and every language version is expected to carry equivalent information. In practice, translation errors and unsynchronised content updates introduce semantic differences between versions. SwissGov-RSD is a benchmark of 224 multi-parallel Swiss government documents in English, German, French and Italian, with human annotations of semantic differences for the EN–DE, EN–FR and EN–IT document pairs.

→ Problem Description

Given two related documents in different languages, predict for each token how strongly its meaning differs from the corresponding content in the other document. Systems should identify both symmetric differences and asymmetric ones such as additions and omissions.

→ Primary Objective

Earlier experiments show that prompted LLMs struggle substantially with cross-lingual RSD on SwissGov-RSD. The goal is to explore how effectively Apertus can address the task beyond the baseline prompting setup. Encouraged directions include:

• Alternative prompting strategies
• Decomposition or chunking approaches
• Structured generation and self-consistency
• Fine-tuning or adapters
• Post-processing — and combinations of the above

The aim is to find high-performing, efficient approaches and to understand which strategies let Apertus recognise fine-grained semantic differences across languages.

The participants should reproduce a baseline from the SwissGov paper with the Apertus model that they are using to see how much improvement their method brought.

🔧 Resources, Tools & Support

→ Datasets

• Available via Hugging Face (ZurichNLP/SwissGov-RSD) or GitHub (ZurichNLP/SwissGov-RSD); usage is explained in both.