Mit tudnak valójában a nagy nyelvmodellek? Nyelvészeti kérdések az LLM-ek értelmezéséhez = What do Large Language Models actually know? Linguistic questions for the interpretation of LLMs

Váradi, Tamás (2026) Mit tudnak valójában a nagy nyelvmodellek? Nyelvészeti kérdések az LLM-ek értelmezéséhez = What do Large Language Models actually know? Linguistic questions for the interpretation of LLMs. ALKALMAZOTT NYELVTUDOMÁNY, 2026 (KSZ). pp. 144-162. ISSN 1587-1061

Preview

Text 8_Varadi.pdf - Published Version Download (327kB) | Preview

Abstract

The release of ChatGPT in late 2022 brought large language models (LLMs) to public prominence with remarkable speed. Although this phenomenal overnight success was due in large part to the system’s impressive linguistic capabilities, these were all too quickly taken for granted as attention shifted toward using LLMs mostly as seemingly omniscient information systems. This paper does not address the general intelligence of such systems or their practical applications; instead, it focuses on a narrower question: how should their performance be interpreted from a linguistic perspective? The paper argues that LLMs do not operate on traditional linguistic units such as morphemes, words, or explicit grammatical categories. Their input consists of statistically derived tokens, and their core learning task is next-token prediction. Their internal operation relies on distributed vector representations and attention-based relational processing rather than explicitly encoded symbolic rules. From this starting point, the paper asks what kind of linguistic knowledge may legitimately be attributed to such systems. It is argued that traditional linguistic categories can often be recovered from model representations, but this does not in itself prove that such categories are explicitly present in the system. Rather, distributional learning may create stable patterns that can be interpreted in linguistic terms. This claim is illustrated with examples from syntax, semantics, and discourse. LLMs can handle many phenomena traditionally described in structural terms, including agreement, long-distance dependencies, context-sensitive meaning, anaphora, and discourse coherence. At the same time, the paper emphasizes the limits of a purely distributional explanation. Meaning, reference, grounding, compositionality, systematic generalization, and interpretability remain open problems. The final part addresses the methodological difficulty of interpreting LLMs linguistically: unlike human speakers, these systems do not provide anything comparable to speaker intuitions. For this reason, the paper suggests that current empirical methods should be complemented by elicitation-based approaches analogous to those used in field linguistics. The conclusion is that if LLMs continue to display forms of linguistic behavior that, in human speakers, would count as evidence of substantial grammatical knowledge, this may require a reconsideration of some basic assumptions of theoretical linguistics.

Actions (login required)

Edit Item