UZH_Pandas at SimpleText2024: Multi-Prompt Minimum Bayes Risk with Diverse Prompts

This paper serves as a summary of further experiments of the paper "SimpleText Best of Labs in CLEF-2023: Scien-tific Text Simplification Using Multi-Prompt Minimum Bayes Risk Decoding" [1], adapted to the SimpleText2024 Shared Task 3.1 dataset. We observe how candidate simplifications generated by the off-the-shelf Llama3 perform differently depending on the prompt, and whether Minimum Bayes Risk (MBR) re-ranking is beneficial with underperforming candidates. Finally, on a small sample, we investigate the agreement of simplification candidate re-rankings between MBR and a human annotator.