大语言模型生成文本的语义从何而来?
成素梅,彼利格亚·拜黑提亚
摘要(Abstract):
大语言模型(LLM)进入语言实践,催生在传统语义框架内难以被充分理解的语言现象。从“生成者—语言文本—接受者”处于分置状态这一事实来看,LLM文本的语义始于文本内部的形式关联,成于交互过程中的理解活动,显于接受者的意义显现。LLM文本的指称效力与训练语料中有关对象的文本化程度成正比。语义与指称的分离揭示出LLM文本语义的三重来源:语料库中分布式语义、对人类语言规范的模式学习、人机交互中文本的可理解性与接受关系。对LLM生成文本的语义揭示,深化了我们对语言本身的理解。
关键词(KeyWords): 语义来源;结构性分置;指称;可理解性;人机交互
基金项目(Foundation): 国家社会科学基金重大项目“数字智能技术与哲学发展及知识生产范式变革研究”(24&ZD320)
作者(Author): 成素梅,彼利格亚·拜黑提亚
参考文献(References):
- [1]Piantadosi S T, Hill F. Meaning without Reference in Large Language Models[PP/OL]. V2. arXiv(2022-08-12)[2026-09-05]. https://doi.org/10.48550/arXiv.2208.02957.
- [2]Brandom R B. Making It Explicit:Reasoning, Representing, and Discursive Commitment[M].Cambridge, MA:Harvard University Press, 1994.
- [3]Pavlick E. Symbols and Grounding in Large Language Models[J]. Philosophical Transactions of the Royal Society A, 2023, 381(2251):20220041.
- [4]Pedrotti A, Rambelli G, Villani C, et al. How Humans and LLMs Organize Conceptual Knowledge:Exploring Subordinate Categories in Italian[C]//Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics(Volume 1:Long Papers). Vienna:Association for Computational Linguistics, 2025:4464-4482.
- [5]Yu M, Liu L, Wu J, et al. The Stochastic Parrot on LLM’s Shoulder:A Summative Assessment of Physical Concept Understanding[PP/OL]. V1. arXiv(2025-02-13)[2026-09-05]. https://doi.org/10.48550/arXiv.2502.08946.
- [6]Frege G. On Sense and Reference[C]//Geach P, Black M. Translations from the Philosophical Writings of Gottlob Frege. Oxford:Basil Blackwell, 1952:56-78.
- [7]Russell B. On Denoting[J]. Mind, 1905, 14(56):479-493.
- [8]Kripke S A. Naming and Necessity[M]. Cambridge, MA:Harvard University Press, 1980.
- [9]Putnam H. The Meaning of“Meaning”[C]//Gunderson K. Language, Mind, and Knowledge:Minnesota Studies in the Philosophy of Science, Vol. 7. Minneapolis:University of Minnesota Press,1975:131-193.
- [10]Mandelkern M, Linzen T. Do Language Models’ Words Refer?[J]. Computational Linguistics, 2024, 50(3):1191-1200.
- [11]Xu Q, Peng Y, Nastase S A, et al. Large Language Models without Grounding Recover Nonsensorimotor but Not Sensorimotor Features of Human Concepts[J]. Nature Human Behaviour, 2025, 9(9):1871-1886.
- [12]Sοgaard A. Do Language Models Have Semantics? On the Five Standard Positions[C]//Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics(Volume 1:Long Papers). Vienna:Association for Computational Linguistics, 2025:25910-25922.