Boris Grozdanoff (Bulgarian Academy of Science)
Fregean Meaning and Probabilistic Truth in LLMs
|
Abstract: Frege's theory of sense and reference revolutionized semantics by distinguishing between a term's sense and its reference. Despite its foundational significance, Frege's theory faces competition from several influential rivals, including Russell’s descriptive theory, Wittgensteinian use theories, and Kripke's causal-historical account of reference. I observe that current Large Language Models (LLMs) offer new insights into Fregean semantics through computationally realized semantic patterns. In LLMs, sense corresponds implicitly to high-dimensional embeddings capturing statistical relationships among words in varying linguistic contexts. Reference in LLMs, arguably, arises as implicit stable clusters in these embeddings, signifying consistent associations with the same conceptual object. The notion of truth in LLMs diverges from traditional Fregean truth; instead, truth emerges probabilistically as statistical confidence at the output and unresolved philosophical challenges continue to persist. While LLMs can effectively model sense-meaning, they lack clear reference-fixing mechanisms and direct causal-historical grounding thus undermining the hypothesis that probabilistic associations do truly capture Fregean or other references. In this talk I explore the interplay between Fregean semantics and probabilistic truth generation in LLMs and I will stress on both potentials and limitations of LLMs to arrive at a successful artificial Fregean style semantics.
|
|
|
|