描写还是解释:由ChatGPT反思语言学的两种目标  

How versus Why:Reflections on the Two Objectives of Linguistics by Means of ChatGPT

作　　者：袁毓林 Yuan Yulin

机构地区：[1]澳门大学人文学院中国语言文学系,中国澳门999078 [2]北京大学中文系/中国语言学研究中心,北京100871

出　　处：《语言战略研究》2025年第1期62-74,共13页Chinese Journal of Language Policy and Planning

基　　金：澳门大学讲座教授研究与发展基金(CPG2024-00005-FAH)和启动研究基金(SRG2022-00011-FAH)支持

摘　　要：本文在现代大语言模型语境下反思语言学研究的两种目标之争:精确描写(语言事实,how)还是科学解释(语言能力,why)?以此为中心,讨论了一系列相关的问题,并考察了ChatGPT能否捕获长距离依存、能否理解句法与语义分离的句子、对语言的科学解释与精确描写是否对立。得出的结论是:(1)ChatGPT等大模型能够超越马尔可夫过程模型,来捕获语句中不同词语之间的长距离依存关系;能够隐式地学习基本的句法和语义知识,从而理解、识别和生成语义异常的句子。(2)对语言的精确描写和科学解释并不对立,并且前者比后者更加重要。(3)生成语法学的“原则与参数”范式下的范畴语法,对于描写人类自然语言有不可克服的困难。(4)语法学的研究取向应该是语义优先,而不是句法优先。(5)大模型的成功说明:对语言事实的准确描写远比对语言能力的抽象解释更为基本。Since its public release at the end of 2022,ChatGPT has attracted worldwide attention.Much research has been conducted on the opportunities and challenges ChatGPT has brought to linguistic studies.At the same time,scholars hold different views on the roles of ChatGPT in linguistic studies.This paper begins with Norvig’s(2011)argument on two competing goals in linguistic studies:descriptive accuracy(of linguistic performance,that is,how)and scientifi c explanation(of linguistic competence,that is,why).Centered on this issue,a series of related questions are discussed,leading to the following conclusions:(1)ChatGPT and Large Language Models(LLMs)can surpass Markov Process Model to capture long-distant dependency holding between diff erent words in a sentence.They can implicitly learn basic syntactic and semantic knowledge,enabling them to understand,recognize,and generate semantically anomalous sentences.(2)Descriptive accuracy and scientifi c explanation do not contradict each other,and the former is more important than the latter in linguistic studies.(3)Categorical grammar within the“principles and parameters”paradigm of generative grammar faces insurmountable diffi culties in describing human natural language.(4)The approach of grammar study should prioritize semantics over syntax.(5)The success of LLMs shows that the descriptive accuracy of linguistic performance is far more basic than abstract explanation of linguistic competence.

关 键 词：ChatGPT 语言模型 描写/解释 语言事实/语言能力 语义优先/句法优先 

分 类 号：H002[语言文字—语言学]