From Bronze to Benchmark: Validating Parliamentary Speech Labels in Spanish for Policy Agenda Research

About the project

**Principal Investigator:** Bastián González-Bustamante **Research Team:** Christopher Klamm and Tom Bellens This project develops a validated benchmark extension for the Spanish language within the Oxford Computational Science Group (OCPSG) project on benchmarking LLMs and fine-tuned models for multilingual policy agenda annotation in parliamentary speeches. In the current silver-standard workflow, speeches in Spanish are excluded because our entropy threshold retains only a minimal and highly imbalanced set of observations. However, the broader bronze-standard pool contains a much larger and substantially more diverse set of candidate observations, which makes speeches in Spanish an ideal case for methodological innovation rather than exclusion. The project will test whether moderate-uncertainty labels can be transformed into reliable data through targeted human validation and adjudication protocols. The main outputs will be a curated Spanish-focused validated dataset release on Hugging Face and a working paper suitable for subsequent submission to a Web of Science journal. This project, as a small spin-off of the OCPSG Benchmarking LLMs project, will strengthen Universidad Diego Portales’ visibility in cutting-edge computational social science and demonstrate its active role in internationally networked AI research. The broader OCPSG project is explicitly focused on benchmark development, documentation, and reproducible evaluation workflows for LLMs and fine-tuned models applied to multilingual policy agenda annotation in parliamentary speeches.