Inspiration
We wanted to make political claims easier to verify against Switzerland’s official voting booklets. The challenge is multilingual, document-grounded fact checking with reliable evidence.
What it does
Fact Attack uses Apertus to classify a claim as Entailment, Neutral or Contradiction. It supports German, French and Italian, including cross-lingual cases, and returns the relevant source pages as evidence.
How we built it
We extract the booklet structure, identify the section and pages relevant to the claim, and provide this focused context to Apertus v1.5. If section detection fails, BM25 retrieval selects relevant pages. Evidence is returned as original PDF text with page numbers.
Challenges we ran into
The booklets contain opposing viewpoints, legal details and multiple languages. We had to distinguish who is making an argument, preserve important conditions and avoid overwhelming the model with the full document.
Accomplishments that we're proud of
- Multilingual claim verification with CLI, Docker and Streamlit demo
- 0.977 Macro-F1 on the Advanced task
- 0.982 Macro-F1 on the Beginner task
- 0.980 evidence Hit@5
- Around 55% fewer input tokens through structured context selection
What we learned
The right context is more valuable than more context. Selecting the relevant booklet section improved accuracy, reduced token usage and produced more reliable evidence.
What's next for Apertus-Team-Fact-Attack-2026
We plan to improve claims without clear source references, evaluate more unseen voting booklets and investigate remaining errors in speaker attribution and contradiction detection.