Domain: Civic / Democracy
Switzerland's direct democracy asks citizens to vote several times a year on complex legal, social and financial issues. For each national vote, the federal government publishes an official voting booklet (Abstimmungsbüchlein) explaining the proposal, its background, the arguments of supporters and opponents, and the position of the Federal Council and Parliament. These materials are a rich source of complex, structured and unstructured information — proposal descriptions, legal and financial detail, arguments, qualifications, dates, numbers and statements across different linguistic contexts — and an authoritative, relatively balanced basis for understanding what's being voted on. They therefore make a valuable real-world document corpus for investigating how well multilingual models like Apertus understand and reason over complex documents. During campaigns, voters meet many claims about proposals, and checking them against the official materials is difficult and time-consuming. This challenge explores how Apertus can be applied to this document-grounded claim-verification task, while remaining multilingual.
Determining the relationship between a natural-language claim and information in a document is hard: relevant information may be spread across sections, expressed with qualifications or negations, or held in tables and other structures. Given an official Swiss voting booklet and a claim, the system must decide whether the document:
Core task: Document + Claim → Entailment / Contradiction / Neutral. This is specifically about document extraction and document-grounded natural language inference (NLI) using Apertus. It is not general question-answering, and does not require deciding whether a claim is objectively true in the real world — only whether Apertus can reliably determine the relationship between a claim and the source document.
The objective is to build a practical, multilingual claim-verification application powered by Apertus that can operate over complex Swiss government documents.
The challenge should demonstrate both:
A straightforward baseline is to provide the complete voting booklet together with the claim directly to Apertus.
However, voting booklets can contain substantial amounts of text and complex document structure. Supplying the entire document may result in high context length and token consumption.
Participants are therefore encouraged to investigate whether the document can be processed, selected or compressed before being provided to Apertus, while maintaining reliable NLI performance. The challenge deliberately leaves the technical approach open.
In particular, the system should be able to: