Project Details
Projekt
The pragmatic test: how humans and LLMs decode presupposed meaning
Applicants
Dr. Nadine Bade; Professorin Dr. Miriam Butt
Subject Area
General and Comparative Linguistics, Experimental Linguistics, Typology, Non-European Languages
Applied Linguistics, Computational Linguistics
Applied Linguistics, Computational Linguistics
Term
since 2026
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 579385859
The project investigates the robustness and reliability of Large Language Models (LLMs) in interpreting presuppositions through parallel human–machine experiments. Presuppositions play a central role in persuasive communication strategies, such as framing, because they introduce information indirectly into the shared discourse. This mechanism of accommodation contributes significantly to their persuasive force and explains why expressions like again frequently appear in political slogans, policy documents, or advertising. While descriptive research has highlighted the importance of presuppositions in persuasive and manipulative communication, only few experimental studies exist on the actual effectiveness of presupposed content. Initial findings also indicate that LLMs struggle to reliably identify and resolve presuppositions—particularly in contexts where the risk of misinformation is high. Our project establishes a systematic empirical foundation by comparing the performance of LLMs and human participants on presupposition resolution tasks. Different communicative contexts are considered in order to integrate both lexical triggers and discourse factors into the modeling. Building on these results, we develop computational models that combine semantic and pragmatic analyses to enable improved detection and resolution of presuppositions. Methodologically, we combine computational work on the automatic detection and annotation of framing phenomena with experimental studies on the interpretation of presuppositions. The project centers on three aims: (i) evaluating LLMs through behavioral tests modeled on established psycholinguistic methods, (ii) examining whether semantic parsing and formal representations can improve LLM training and fine-tuning, and (iii) developing new resources—such as annotated corpora and methodological guidelines—for experimental and computational research on presuppositions. In the long run, the project seeks to clarify the extent to which LLMs can serve as tools for detecting presupposition-based biases in sensitive communicative settings. At the same time, we expect to contribute to theoretical pragmatics by refining and extending formal models of presuppositions in light of findings from parallel human–machine experiments.
DFG Programme
Priority Programmes
