Interpretable Surprisal: Language Models Between Linguistic Structure and Neural Evidence

• Applicants: Dr. Alessandro Lopopolo; Professorin Dr. Milena Rabovsky
• Subject Area: General, Cognitive and Mathematical Psychology
• Term: since 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 578800581

This project investigates how surprisal — a probabilistic measure of word predictability estimated by language models (LMs)— relates to human language comprehension. The central question is whether surprisal reflects only surface-level statistics or also captures deeper linguistic structures such as syntax and semantics. Building on evidence that surprisal correlates with neural and behavioral markers of comprehension (e.g., the N400 ERP component, reading times), the project tests whether these effects can be traced to lexical prediction alone or are partly driven by structural and semantic processes. In parallel, it examines how LMs acquire sensitivity to linguistic structure during training, probing the role of data scale, type, and complexity. To address these issues, the project combines computational modeling, linguistic analysis, and neurocognitive experiments, aligning LM predictions with EEG and behavioral data. This integrated approach will clarify which aspects of LM behavior genuinely resemble human processing, how training shapes their sensitivity to syntax and semantics, and, more broadly, how LMs can serve as interpretable and epistemically robust tools in the cognitive and brain sciences.

• DFG Programme: Priority Programmes
• Subproject of: SPP 2556: Robust Assessment & Safe Applicability of Language Modeling: Foundations for a New Field of Language Science & Technology (LaSTing)