ChatGPT等大型语言模型对语言学理论的挑战与警示  

Challenges and Warnings from Large Language Models like ChatGPT on Linguistic Theories

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 [2]北京大学中文系/中国语言学研究中心,北京100871

出　　处：《当代修辞学》2024年第1期2-17,共16页Contemporary Rhetoric

基　　金：澳门大学讲座教授研究与发展基金(CPG2023-00004-FAH);启动研究基金(SRG2022-00011-FAH);国家社会科学基金专项项目“新时代中国特色语言学基本理论问题研究”(项目编号:19VXK06)的资助。

摘　　要：本文首先简介ChatGPT怎样颠覆自然语言处理的工作范式和语言学的有关信条,然后介绍Piantadosi (2023)怎样通过对比现代大型语言模型的工作机理与生成语言学的研究路径,得出以下这个极具批判性和震撼力的结论:现代机器学习已经颠覆并绕过了乔姆斯基方法的整个理论框架。为了便于大家的阅读与理解,本文围绕着下面三个问题来介绍和评述Piantadosi (2023):1)现代语言模型能不能看作是一种语言理论 2)现代语言模型及其工作原理能不能禁得住语言学家的质疑 3)现代语言模型及其表现能不能反驳乔姆斯基的语言学主张?最后,文章指出,面对现代大型语言模型在语言运用上的出色表现,我们应该摒弃鸵鸟法、权威法和先验法,采用正视问题、勇于怀疑、不断试错与验证的科学法。This article first introduces how ChatGPT discards the working paradigm of natural language processing(NLP)and challenges linguistic beliefs.It then discusses the critical and groundbreaking conclusion which was made by comparing the mechanisms of modern large language models with the research path of generative linguistics in Piantadosi(2023):modern machine learning has subverted and bypassed the entire theoretical framework of Chomsky's approach.To facilitate reading and understanding,this article focuses on the following three questions to introduce and evaluate Piantadosi's(2023)work:1)Can modern language models be considered genuine theories of language?2)Can modern language models and their principles be attested under linguistic scrutiny?3)Can modern language models and their performance refute Chomsky's linguistic claims?Finally,the article suggests that in the face of the outstanding performance of modern large language models in language application,we should abandon the methods of tenacity,authority,and a priori,and instead adopt a method of science that involves confronting problems,embracing skepticism,and continuously experimenting and validating.

关 键 词：ChatGPT 现代/大型语言模型 乔姆斯基/生成语言学的主张 科学法 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] H08[自动化与计算机技术—控制科学与工程]