关于大语言模型与人类语言能力差异的思考

On the Differences between Large Language Models and Human Language Abilities

在大语言模型备受关注且饱受争议的背景下,为回应部分人对于大语言模型的质疑,马霍瓦尔德(Mahowald)等人对“语言能力”的内涵进行了重新思考,并将其区分为“形式语言能力”和“功能语言能力”。文章基于马霍瓦尔德等人对于ChatGPT大语言模型两种语言能力的区分,进一步详细探讨了两种语言能力的内涵,并且对“大语言模型在何种程度上实现了或媲美了哪些语言能力”这一问题进行了深入探究。文章发现:(1)大语言模型对于两种语言能力的掌握上都仍有欠缺,尤其尚未掌握人类“句法”的核心部分;(2)大语言模型和人类的语言能力在“工作机制”上存在着本质的区别。这表明大语言模型虽然在语言处理方面取得了很大的进步,但与人类的语言能力相比,仍存在一定的差距。

In response to the growing attention and controversy surrounding Large Language Models(LLMs),Mahowald et al.(2023)have reconsidered the concept of“linguistic competence”,proposing a distinction between“formal”and“functional”linguistic competence.The article is based on the distinction between the two language abilities in the ChatGPT Large Language Models by Mahowald et al.,and further explores the connotations of the two language abilities in detail.It also delves into the question that is“to what extent the Large Language Models have achieved or have been comparable to which language abilities”.The outcome reveals:(1)LLMs still exhibit shortcomings in both areas,particularly in their failure to fully grasp the core aspects of human syntax.(2)There are fundamental differences in the mechanisms of Large Language Models and human linguistic competence.This indicates that,despite significant advancements in language processing,Large Language Models remain limited,compared to human linguistic competence.