Day 1

O4

Assessing the Political Fairness of Multilingual LLMs: A Case Study Based on a 21-Way Multiparallel EuroParl Dataset

• Links: * Arxiv - LREC'26 | Schedule - Github
• Takeaways:
  • BLEU variants not comparable across language pairs
  • Borda count - Wikipedia is a way to fairly rank, rating = number of candidates below them in ranking
    • Originally for voting
    • Used in IR to aggregate ratings and benchmarking models on tasks w/ different metrics

Appeal, Align, Divide? Stance Detection for Group-Directed Messages in German Parliamentary Debates

• Links: LREC
• Group appeals:
  • Negative: liberals want to X -> bad
  • +: we stand w/ millions of commuters
• Benchmark + LLMs evaluation
• LLL “our work is similar in spirit to”
• QQQ/PPP “Our work comes with limitations, the most obvious one being the high computational and, for closed- source models, financial cost of LLMs,”

Day 2

Once upon a Kernel: Extracting Important Events from Narratives

• Links
  • LREC
• Kernel events in stories are key ones that survive across retellings
• Idea history: Narrative functions (1928), cardinal functions, kernel events (chatman 78)
• Fairy tales b/c simple
• Operational definition (=operationalization): three criteria
• Gold standard dataset of 50 fairytales 3.5k-ish kernels

Temporal Expression Recognition in Legal Transcripts

• Taxonomy
  • calendar date, duration, vague period (‘sometime in july’) etc.

Multilingual, Multimodal Pipeline for Creating Authentic and Structured Fact-Checked Claim Dataset

TODO really really cool, need to look deeper

• [2601.07985] Multilingual, Multimodal Pipeline for Creating Authentic and Structured Fact-Checked Claim Dataset