从三种复杂句看ChatGPT是不是随机鹦鹉?——语言大模型能不能理解语言意义的测试与讨论  

Tests and discussions on whether large language models can understand linguistic meanings from the perspective of three complex sentences:Is it the case that ChatGPT is a stochastic parrot

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519000 [2]北京大学中文系/中国语言学研究中心,100871

出　　处：《语言教学与研究》2025年第1期35-49,共15页Language Teaching and Linguistic Studies

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

摘　　要：本文通过考察ChatGPT在指称歧义句、花园幽径句和递归嵌套句等三种复杂句的语义理解上的表现,来讨论下列问题:ChatGPT等现代大型语言模型到底能不能理解语言的意义?它们真的只是一种不顾意义的随机鹦鹉吗?文章发现,ChatGPT在指称歧义句和一般递归结构的理解方面表现出色,在花园幽径句的理解方面表现平凡;在中心内嵌的递归结构的理解方面,跟人类一样表现不佳。由此,文章指出,ChatGPT等语言大模型能够理解语句的语言性意义,但是未必能够理解语句背后隐微的交际意义和具身意义。最后,文章展望,语言学可为建设关于多种意义的理解的新型智能科学做出独特的贡献。This paper tests ChatGPT's performance in understanding the semantic meanings of three types of complex sentences:referential ambiguity sentences,garden-path sentences,and recursive sentences,especially center-embedding recursive sentences,for answering the following questions:Can modern large language models like ChatGPT truly understand the meaning of natural languages?Are they merely stochastic parrots that disregard meaning?The paper finds that ChatGPT excels in understanding referential ambiguity sentences and normal recursive structures but performs averagely in garden-path sentences.And similar to humans,it struggles in confusion to understand center-embedding recursive structures.Based on these tests,the paper argues that while large language models like ChatGPT can comprehend the linguistic meaning of sentences,they may not fully grasp the subtle communicative and embodied meanings behind them.Finally,the paper looks ahead,suggesting that linguistics study can make a unique contribution to the development of a new intelligent science that understands the multifaceted dimensions of meaning.

关 键 词：指称歧义句 花园幽径句 递归嵌套句 ChatGPT 语言大模型 语义理解 语言性意义 交际意义 具身意义 

分 类 号：G63[文化科学—教育学]