The signal is coming from inside the noun phrase! Tracking semantic proto-role inferences during sentence processing

Lucas Y. Li, Zander Lynch, Marten Van Schijndel


Abstract
Semantic roles between a predicate and argument can be decomposed into proto-role properties (e.g.,Instigation). We introduce a novel LLM feature attribution method, Generalized Contextual Decomposition for Transformers (GCD-T), which we use to probe which parts of a sentence enable models to infer proto-role properties. We compare our findings with human inferences.
Anthology ID:
2026.scil-main.40
Volume:
Proceedings of the Society for Computation in Linguistics 2026
Month:
July
Year:
2026
Address:
San Diego, CA
Editors:
Rob Voigt, Alex Warstadt, Naomi Feldman, Tal Linzen
Venues:
SCiL | WS
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
427–441
Language:
URL:
https://preview.aclanthology.org/ingest-acl-workshops/2026.scil-main.40/
DOI:
Bibkey:
Cite (ACL):
Lucas Y. Li, Zander Lynch, and Marten Van Schijndel. 2026. The signal is coming from inside the noun phrase! Tracking semantic proto-role inferences during sentence processing. In Proceedings of the Society for Computation in Linguistics 2026, pages 427–441, San Diego, CA. Association for Computational Linguistics.
Cite (Informal):
The signal is coming from inside the noun phrase! Tracking semantic proto-role inferences during sentence processing (Li et al., SCiL 2026)
Copy Citation:
PDF:
https://preview.aclanthology.org/ingest-acl-workshops/2026.scil-main.40.pdf