【作者】 杨阳；

【导师】 蔡清；

【作者基本信息】 华东师范大学 ， 认知神经科学， 2021， 硕士

【摘要】 随着心理学领域对“心理字典”的研究和人工智能领域的蓬勃发展,人类对于语义概念组织形式的研究取得了长足进展,但仍未知其全貌。计算机领域基于大量文本的词嵌入语义模型在语言加工场景中得到了广泛应用,但其模型背后的机制未能得到有效的解释,也限制了它在深度语义加工中的应用。在心理学领域,先前工作已提出自由联想网络可通过覆盖足够多的词汇(节点)和联想(边),构建蕴含多维度语义信息及真实心理特征的理想模型,但至今尚未建立在数据规模上可与计算机领域语义模型比肩的中文联想语义网络数据库。建立大规模心理语义网络对于比较以上模型的优缺点,以及进一步推动对语义概念的研究及应用必不可少。认知神经科学研究已发现,语义加工涉及到广泛分布的脑区,并由此认为语义概念可能遵循分布式表征。近年方法学的发展,特别是多体素模式分析方法,使得我们可以对语义的多维特征进行更好的考察。另一方面,由于缺乏有效的理论模型,先前研究中对语义条件的探查多数限于抽象/具体、有生命/无生命等个别人为指定的范畴,可能忽略重要的心理语义特征。为了全面和准确的考察语义的神经表征,合理选取语义网络模型和减少实验设计的人为偏差至关重要。本研究首先运用自由联想范式建立大规模中文词联想网络SWOW-ZH,在此基础上,我们将基于大规模词汇联想的语义网络和基于大规模文本的词嵌入模型(Word2Vec和Concept Net)作为理论模型,分别与基于功能磁共振的神经表征上进行了比较。为了避免人为选取语义范畴/社区带来的偏差,我们完全基于数据驱动选取刺激材料,即对以上三个理论模型分别使用社区探测算法,得到9个共有的稳定社区中的72个双字词作为刺激材料,并分别建立了从粗粒度到细粒度的三个层级(社区层级、集群层级、节点层级)对应的表征相似性矩阵。通过功能磁共振下的语义判别任务和多体素表征相似性分析的方法,对语义加工时不同类型和层级的理论模型和脑表征的相似性进行了探查。结果发现,在全脑层面,各概念层级的语义神经表征都显著与SWOW-ZH更相似,而与词嵌入模型相似性较低。在社区层级上,与SWOW-ZH显著相关的区域包括广泛的颞顶叶与外侧枕叶区域;在节点层级上,更多的显著相关集中在颞前叶区域。此结果1、揭示了自由联想网络比词嵌入模型更接近大脑中的语义概念表征,这可能表示了基于自由联想的语义网络与基于文本的词嵌入模型相比,可能编码了更多广义的语义特征;2、发现了概念的多重层级结构体现在从颞枕联合区到颞极的功能梯度上;3、前颞叶脑区参与了不同语义层级信息的编码,这一结果直接印证了轮辐模型(Hub and Spoke Model)提出的关于前颞叶作为语义多模态信息整合枢纽的假说。以上发现进一步深化了我们对语义神经表征机制的了解,并指出了基于大规模实验数据进行语义网络构建对于各领域语义研究的重要性。更多还原

【Abstract】 Research on the mental organization of semantic concepts has made remarkable advances by virtue of psychological studies on mental lexicon and the flourishing of artificial intelligence,but the full picture is still unknown.In the field of computer science,a large number of word embedding models derived from text corpus have been created and are widely used in real semantic scenarios,but the underlying mechanisms have not been clearly addressed,which has constrained the applications associated with deep semantic processing.In the field of psychology,free association networks have shown the capability in constructing models that contain multidimensional semantic information and real psychological features by covering large numbers of words(nodes)and associations(edges).However,no semantic association network database in Chinese of comparable scale to the mainstream corpus-derived models has been established.Thus,establishing psychologically valid large-scale semantic network is necessary for the comparison of the above-mentioned models and for advancing the research of semantic concepts.Research in cognitive neuroscience has discovered that semantic processing involves widely distributed brain regions,suggesting distributed encoding of semantics.Recent methodological developments,especially multi-voxel pattern analysis methods,have allowed investigations of semantic features in a multidimensional approach.On the other hand,due to the lack of effective theoretical models,previously investigated semantic conditions have been mostly limited to manually specified categories such as abstract vs.concrete and animate vs.inanimate,where potentially significant neurosemantic features might be ignored.In order to comprehensively and accurately investigate the neural representations of semantics,it is crucial to properly select semantic network models and reduce subjective bias in experimental design.The present study uses word association paradigm to build a large-scale association network of Chinese words,the SWOW-ZH.Neural representations of word concepts are compared with the representations in two types of theoretical models,one being the SWOW-ZH,and the other being the word embedding models,Word2 Vec and Concept Net.To avoid bias in acquiring the target semantic categories,stimulus words are determined in a data-driven approach: community detection algorithm is applied on each of the three theoretical models,resulting in 72two-character words from 9 reliable communities that are common across the models.Representational dissimilarity matrices of the stimuli are constructed at three levels of granularity,namely the community level(the coarsest),the cluster level,and the node level(the most fine-grained).Regions of interests across the whole brain are searched for the representational similarities to different theoretical models at different granularities,based on data measured by functional magnetic resonance imaging during a semantic judgment task.Neural representations of concepts are found to be most consistent with the SWOW-ZH at all levels of concept granularity at the whole-brain level.Specifically,at community level,regions that are most similar to SWOW-ZH include a wide range of temporoparietal and lateral occipital areas.While at node level,significant similarities are mostly localized in the anterior temporal lobe.First,these findings reveal that free association network resembles the neural representation of semantic concepts to a larger than the word embedding models,suggesting a better characterization of mental features of concepts.Second,neural semantic representations are found hierarchically structured along the gradient from temporooccipital cortices to the anterior temporal lobe.Third,the anterior temporal lobe encodes semantic information of different granularities,supporting the Hub-and-spoke Model that proposes the anterior temporal area as the hub of integrating semantic features from multiple sources.Overall,the findings have advanced our understanding of the neural representations of semantic knowledge,and have highlighted the significance of semantic network constructed from large-scale behavioral data to the studies of semantics in various fields.更多还原