摘要
白硕近文提出,要深挖语言的第一性原理,打造“小而美”的低资源语言模型;语言是不是“学会的”,并非乔姆斯基语言理论的本质。我们认为,“语言天赋”就是乔姆斯基语言理论的第一性原理,其本质是人类与生俱来的“语言官能”赋予人类递归能力,达到“有限手段的无限使用”。但大语言模型的语言能力是学会的,其成功基于“无限语料的有限使用”。就语言技术上的可行性与语言工程上的可实施性而言,基于语言的第一性原理的“小而美”的语言模型目前尚有难以跨越或抵达的距离。人类天赋的语言官能学说目前尚无足够的科学证据来支撑,而统计学习理论却努力证明,人类婴儿的大脑可以通过概念系统感知外物、获得知识、学习语言。基于“语言是天赋能力而非学会”语言观对大语言模型的批评和对“小而美”语言模型的追求,有可能反映了学者思维的一个弱点,即陷于“理论诱导的盲区”。基于情绪建构理论的调查和实验发现,经典情绪理论所谓的情绪普遍性,实际上是实验设计的结果,却被研究者看作事实。这对普遍语法的研究具有一定的警醒作用
Abstract
Bai Shuo (2025) recently proposed that we should delve deeper into the first principle of language to create “small yet efficient” low-resource language models and argued that whether language is “learned” is not the essence of Chomskyan linguistic theory. We hold that, in contrast, “linguistic innateness” is the first principle of Chomskyan linguistic theory, since it is essential for humans’ innate “linguistic faculty” to endow humans with recursive abilities, allowing “infinite use of finite means”. The language ability of large language models, however, is learned, and their success is based on “finite use of infinite data”. As far as language technology and language engineering are concerned, it is infeasible or unworkable at this stage to construct “small yet efficient” language models on the first principle of language. While the theory of humans’ innate linguistic faculties still lacks sufficient scientific evidence, the statistical learning theory currently strives to prove that the brains of human infants can perceive external objects, acquire knowledge, and learn language through a conceptual system. The criticism of large language models and the pursuit of “small yet efficient” language models from the linguistic perspective that “language is not learned but an innate ability” may reflect a weakness in scholarly thinking, namely, the “theory-induced blindness”. As surveys and experiments guided by the theory of constructed emotion reveal, what classical emotion theories refer to as emotional universality, often regarded as fact by researchers, is actually the effect of experimental designs. This serves as a wake-up call for the study of universal grammar.
关键词
语言的第一性原理 /
“小而美”的语言模型 /
普遍语法 /
理论诱导的盲区
Key words
first principle of language /
“small yet efficient” language model /
Universal Grammar /
theory-induced blindness
袁毓林.
打造“小而美”的语言模型有没有可能性?——跟白硕博士讨论“语言的第一性原理”和“普遍语法”#br#[J]. 语言战略研究. 2026, 11(3): 28-40 https://doi.org/10.19689/j.cnki.cn10-1361/h.20260302
Yuan Yulin.
Constructing “Small Yet Efficient” Language Models: A Response to Dr. Bai on First Principle of Language and Universal Grammar[J]. Chinese Journal of Language Policy and Planning. 2026, 11(3): 28-40 https://doi.org/10.19689/j.cnki.cn10-1361/h.20260302
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
基金
澳门大学讲座教授研究与发展基金(CPG2026‒00012‒FAH)和启动研究基金(SRG2022‒00011‒FAH)。承施春宏教授、周韧教授、孙晓雪博士、编辑部和审稿专家的指正,谨此一并致以谢忱。