国际学界关于ChatGPT语言能力的争论与思考  

ChatGPT’s Linguistic Competence:Debates and Reflections in the International Academic Community

作　　者：时仲 田英慧 司富珍[1] Shi Zhong;Tian Yinghui;Si Fuzhen

机构地区：[1]北京语言大学语言学系/乔姆斯基研究所/生物语言学与脑科学实验室,北京100083 [2]北京华文学院招生办公室,北京102206

出　　处：《语言战略研究》2025年第1期75-86,共12页Chinese Journal of Language Policy and Planning

基　　金：2024年度教育部哲学社会科学研究后期资助项目“人脑与大语言模型句法加工的比较研究”(24JHQ046);北京语言大学后期资助项目孵化课题“人脑与大语言模型句法加工的比较研究”(24HQ04);北京语言大学研究生创新基金(中央高校基本科研业务费专项资金)项目“ChatGPT对生成语法理论的影响”(23YCX057)。

摘　　要：随着ChatGPT为代表的大语言模型在应用方面取得极大成功,语言官能是否为人类独有的问题引起热议,国际学界形成两个对立的阵营。一方认为,大语言模型语言理解和产出方面达到了媲美人类的水准,对乔姆斯基的语言学理论提出了挑战,甚至足以取代生成语法的语言学理论地位。另一方则认为,人类语言习得“刺激贫乏”但生成能力惊人,而大语言模型“学习”语言依靠输入海量数据,因此,它无法对人类语言的本质问题给出合理性解释,在语言的核心属性方面与人类语言官能存在本质区别。不少实证测试也对夸大大语言模型在语言学理论中的地位的观点进行了批判。本文认为,对这一问题的讨论,首先应思考如下问题:(1)区分科学理论建构与工程应用;(2)对“可能的语言”与“不可能的语言”的区分做出原则性的预测与解释;(3)探讨自然语言习得“刺激贫乏”与大语言模型依靠“豪华型”海量数据之对立背后的深层原因;(4)对句法在人类语言和大语言模型中的地位进行更多维度和更系统的对比评测。With the great success of ChatGPT and other large language models(LLMs)in practical applica-tions,a heated debate has arisen regarding whether language faculty is unique to human beings.Two contrasting perspectives have emerged within the international academic community.One perspective argues that LLMs have achieved human-level profi ciency in language understanding and production,thereby challenging Chomsky’s linguistic theories and even potentially replacing the theoretical framework of Generative Grammar.The opposing perspective argues that while humans acquire language despite“poverty of stimulus”,demonstrating a remarkable generative capacity,LLMs“learn”language by leveraging massive data input.Therefore,LLMs fundamen-tally differ from human language faculty in their core attributes and cannot adequately explain the essential nature of human language.Empirical studies have also criticized the tendency to overstate the role of LLMs for linguistic theory.This paper argues that discussions on this issue should begin by addressing the following key issues:(1)the diff erentiation between scientifi c theory formulation and engineering applications;(2)the principled predictions and explanations regarding the distinction between“possible languages”and“impossible languages”;(3)the underlying factors accounting for the contrast between natural language acquisition under“poverty of stimulus”and LLMs’reliance on massive data input;and(4)multi-dimensional and systematic comparative evaluations of the role of syntax in human language versus LLMs.

关 键 词：ChatGPT 大语言模型 刺激贫乏 可能的语言 不可能的语言 

分 类 号：H002[语言文字—语言学]