摘要
本文探讨“大型语言模型是什么”的问题。为此对大模型的评判标准展开实验观察,对大模型的基础设施关联度预测进行直观分析,构建关联度预测的一种形式化LC,进而研究关联度预测的语义解释。在此基础上讨论大模型的真实性挑战、共识挑战、内容属性挑战和非封闭性挑战。主要发现包括:语元关联度是体现人类语言习惯的可自动提取的语言痕迹;关联度预测具有语境相关的统计性质;LC具有弱共识性实质语义;LC是一个非概念化公理系统。这些特点颠覆了科学理论、形式化方法和软件的传统理念在人工智能领域的主导地位,是大模型输出既出人预料、又符合语言习惯的深层原因。
To explore the problem of what a large language model is,we conduct experimental observation on the evaluation criteria for large language models,intuitively analyze the infrastructure of large language models—correlation degree prediction,of which a formalization LC is constructed and semantic interpretations are explored.On top of these,four challenges of truthfulness,consensus,content attribute,and non-closeness for large language models are discussed.The main findings include:the correlation degrees between tokens are automatically extractable language traces that reflect human language habits;correlation degree prediction has the context-sensitive statistical property;LC has a substantive semantics of weak consensus;LC is a non-conceptualized axiomatic system.These radically differ from the traditional notions of scientific theory,formal methods,artificial intelligence(AI)and software,and are the deep reasons why large language models can behave unexpectedly yet consistent with human language habits.
作者
陈小平
CHEN Xiaoping(School of Computer Science and Technology,University of Science and Technology of China,Hefei 230026,China)
出处
《智能系统学报》
CSCD
北大核心
2023年第4期894-900,共7页
CAAI Transactions on Intelligent Systems
基金
国家自然科学基金项目(92048301,U1613216)
国家重点研发计划项目(2020YFB1313602)。
关键词
大模型
形式化
语义
概念化
弱共识
large language models
formalization
semantics
conceptualization
weak consensus
