语义在人脑中的分布式表征:来自自然语言处理技术的证据  

Distributed representation of semantics in the human brain:Evidence from studies using natural language processing techniques

作　　者：蒋嘉浩 赵国钰 马英博 丁国盛[3] 刘兰芳[2,4] JIANG Jiahao;ZHAO Guoyu;MA Yingbo;DING Guosheng;LIU Lanfang(Faculty of Psychology,Beijing Normal University,Beijing 100875,China;Faculty of Psychology,School of Arts and Sciences,Beijing Normal University,Zhuhai 519087,China;State Key Laboratory of Cognitive Neuroscience and Learning,Beijing Normal University&IDG/McGovern Institute for Brain Research,Beijing 100875,China;Center for Cognition and Neuroergonomics at the State Key Laboratory of Cognitive Neuroscience and Learning,Beijing Normal University,Zhuhai 519087,China)

机构地区：[1]北京师范大学心理学部,北京100875 [2]北京师范大学文理学院心理学系,珠海519087 [3]北京师范大学认知神经科学与学习国家重点实验室和IDG/麦戈文脑科学研究院,北京100875 [4]北京师范大学认知神经科学与学习国家重点实验室认知神经工效研究中心,珠海519087

出　　处：《心理科学进展》2023年第6期1002-1019,共18页Advances in Psychological Science

基　　金：国家自然科学基金青年科学基金项目(31900802)资助。

摘　　要：人脑如何表征语义信息一直以来是认知神经科学的核心问题。传统研究主要通过人为操纵刺激属性或任务要求等实验方法来定位语义表征脑区,这类方法虽然取得了诸多成果,但是依然存在难以详细量化语义信息和语境效应等问题。基于语义的分布式假设,自然语言处理(NLP)技术将离散的、难以客观量化的语义信息转变为统一的、可计算的向量形式,极大提高了语义信息的刻画精度,提供了有效量化语境和句法等信息的工具。运用NLP技术提取刺激语义信息,并通过表征相似性分析或线性回归建立语义向量与脑活动模式的映射关系,研究者发现表征语义信息的神经结构广泛分布在颞叶、额叶和枕叶等多个脑区。未来研究可引入知识图谱和多模态融合模型等更复杂的语义表示方法,将语言模型用于评估特殊人群语言能力,或利用认知神经科学实验来提高深度语言模型的可解释性。How semantics are represented in human brain is a central issue in cognitive neuroscience.Previous studies typically address this issue by artificially manipulating the properties of stimuli or task demands.Having brought valuable insights into the neurobiology of language,this psychological experimental approach may still fail to characterize semantic information with high resolution,and have difficulty quantifying context information and high-level concepts.The recently-developed natural language processing(NLP)techniques provide tools to represent the discrete semantics in the form of vectors,enabling automatic extraction of word semantics and even the information of context and syntax.Recent studies have applied NLP techniques to model the semantic of stimuli,and mapped the semantic vectors onto brain activities through representational similarity analyses or linear regression.A consistent finding is that the semantic information is represented by a vastly distributed network across the frontal,temporal and occipital cortices.Future studies may adopt multi-modal neural networks and knowledge graphs to extract richer information of semantics,apply NLP models to automatically assess the language ability of special groups,and improve the interpretability of deep neural network models with neurocognitive findings.

关 键 词：语义表征 大脑 自然语言处理 语言模型 

分 类 号：B842[哲学宗教—基础心理学]