关键词: 大语言模型, 特征表征, 语言能力, 数据驱动范式, 语言学转向

Abstract:

This paper examines the fundamental paradigm shift in linguistic research instigated by the rise of large language models (LLMs), taking the linguistic concept of “features” as its starting point. Traditional feature unification grammar relies on manually defined, discrete symbolic systems, aiming to characterize language competence through rule-based deduction. In contrast, LLMs implicitly build high-dimensional, continuous, and context-sensitive vector representations via statistical learning from massive text corpora, thereby achieving probabilistic modeling of linguistic systems. This transition from “rule-making” to “pattern discovery” not only challenges the epistemological foundations of classical linguistics but also underscores the inherent nature of language as a dynamic probabilistic system. Confronted with the cognitive impact of artificial intelligence, the paper contends that linguistics must proactively embrace a new “data-driven” paradigm. While elucidating the statistical patterns captured by these models, the field should reclaim its disciplinary mission as a bridge connecting human and machine language understanding, thereby contributing to the independent innovation of Chinese linguistics in the digital intelligence era.

Key words: Large language model, feature representation, language competence, data-driven paradigm, linguistic transformation