Do Language Models Know Theo Has a Wife? Investigating the Proviso Problem

INRIA Sémagramme 2025 SIR01 Submission4 Authors

Keywords: Presupposition, Proviso Problem, Natural Language Inference, Explainable AI, Language Models

TL;DR: This paper investigates how language models handle presupposition projection in conditional sentences, focusing on the proviso problem through NLI classification using explainability techniques.

Abstract: We investigate the proviso problem, an unresolved issue in formal semantics concerning presupposition projection in conditionals, using four language models (RoBERTa, DeBERTa, LLaMA, and Gemma) and AI explainability methods. Our findings show that these models process presuppositions differently from humans. The results offer insights for evaluating and refining existing semantic theories.

Submission Number: 4