摘要
ChatGPT等大语言模型在语义理解和常识推理方面表现优秀,其技术奥秘在于模型开发者在对词语进行向量表示时,遵循分布式语义学原理,采用了“嵌入”这种代数方法。但是,经典的测试语言运用等智能水平的“图灵测试”,难以识别欺骗和回避等作弊手段,因此,计算机科学家设计了“新图灵测试”,其中的威诺格拉德模式挑战与语言学关系密切。这一模式以代词消歧为测试点设计句子对和问题,但是经过大规模语料训练的语言模型可以凭借词汇上的统计相关性,而不是靠真正理解句子的意义来给出正确答案。为克服这一缺陷,学者们又发展出WinoGrande数据集,提高了数据的规模和难度,确保它们无法通过网络搜索等手段来得到正确答案。我们用威诺格拉德模式设计了无偏向双重句子对测试ChatGPT,展示了大语言模型在语义理解和常识推理方面已达到接近人类的水平。当然,从具身模拟假说来看,大语言模型不可能像人一样富有体验性地理解人类自然语言。最后,我们呼吁:语言学家应该积极参与构建WinoGrade测试集之类的工作,在人工智能时代扩展自身的学术研究领域。
Abstract
This paper examines the approaches to test the performance of ChatGPT in semantic understanding and common-sense reasoning. It first reviews the remarkable performance of ChatGPT and other large language models (LLM) in semantic understanding and common-sense reasoning, and explains how they use distributional semantics and token vector representations to process language. Then, it critiques the classical “Turing Test” and introduces some alternative new tests such as Winograd Schema Challenge (WSC) and its upgraded version WinoGrande datasets, which are more relevant to linguistic studies. Next, it presents the results of applying ChatGPT to Winograd Schema sentences, showing that LLM have reached human-level performance in semantic understanding and common-sense reasoning. Finally, it argues that these language models can “understand” human natural languages, and the development of tests like the Winograd Schema is an opportunity for linguists to expand their academic field in the era of artificial intelligence.
关键词
ChatGPT /
大语言模型 /
新图灵测试 /
威诺格拉德模式挑战 /
WinoGrande数据集
Key words
ChatGPT /
large language models /
semantic understanding and common-sense reasoning /
(New) Turing Test /
Winograd Schema challenge /
WinoGrande dataset
袁毓林.
如何测试ChatGPT的语义理解与常识推理水平?——兼谈大语言模型时代语言学的挑战与机会[J]. 语言战略研究. 2024, 9(1): 49-63 https://doi.org/10.19689/j.cnki.cn10-1361/h.20240105
Yuan Yulin.
How to Test ChatGPT’s Performance in Semantic Understanding and Common-Sense Reasoning: Challenges and Opportunities of Linguistics in the Era of Large Language Models[J]. Chinese Journal of Language Policy and Planning. 2024, 9(1): 49-63 https://doi.org/10.19689/j.cnki.cn10-1361/h.20240105
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
基金
本课题的研究得到澳门大学讲座教授研究与发展基金(CPG2023–00004–FAH)和启动研究基金(SRG2022–00011–FAH)及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”(19VXK06)资助,谨此谢忱。