Jiang Hong, Chen Zhengwei, Liu Yu, Yang Chun, Yuan Xiaofeng, He Rui
Zhuhai People's Hospital (The Affiliated Hospital of Beijing Institute of Technology, Zhuhai Clinical Medical College of Jinan University), Zhuhai, China.
Faculty of Medicine, Macau University of Science and Technology, Macau, China.
PLoS One. 2025 May 22;20(5):e0324270. doi: 10.1371/journal.pone.0324270. eCollection 2025.
Impairment in the semantic domain is prominent in Alzheimer's Disease (AD). We analyzed spontaneous speech in English from 148 people with probable AD (pAD) and 143 controls, and aimed to replicate these findings in a smaller Greek dataset of 28 controls and 26 pAD patients, using different language models comparatively. Static models (fastText) represented non-contextual meaning via encoding words as static vectors, while contextual models (BERT) represented the contextual meanings sensitive to syntactic structure. These models calculated semantic similarity at two levels: local similarity (between adjacent words/tokens) and global similarity (across all word/token pairs). Generative contextual models (Mistral) additionally quantified token probability within context, thereby indicating the unexpectedness in speech progression. Given that contextual meaning is syntactically sensitive, we introduced averaged dependency distance as an indicator for formal syntactic complexity. Moreover, bimodal models were introduced to evaluate how speech reflected picture-based stimuli. Results showed significant increases in global semantic similarity in the pAD group, as measured by both fastText and BERT models, which co-occurred with enlarged picture-speech semantic distance and increased in speech perplexity. Only the fastText-based global semantic similarity, which captured the contraction in conceptual semantic space, correlated with the overall cognitive decline in the AD populations. These findings together indicates that semantic space changes in AD differed across different forms of meanings and thus points to the necessity of distinguishing these forms to raveling the underlying mechanism.
语义领域的损害在阿尔茨海默病（AD）中很突出。我们分析了148名可能患有AD（pAD）的人和143名对照者的英语自发语言，并旨在使用不同的语言模型，在一个较小的希腊语数据集中（28名对照者和26名pAD患者）重复这些发现，并进行比较。静态模型（fastText）通过将单词编码为静态向量来表示非上下文意义，而上下文模型（BERT）则表示对句法结构敏感的上下文意义。这些模型在两个层面上计算语义相似度：局部相似度（相邻单词/词元之间）和全局相似度（所有单词/词元对之间）。生成式上下文模型（Mistral）还对上下文中的词元概率进行了量化，从而表明了言语进展中的意外性。鉴于上下文意义在句法上是敏感的，我们引入了平均依存距离作为形式句法复杂性的指标。此外，还引入了双峰模型来评估言语如何反映基于图片的刺激。结果显示，通过fastText和BERT模型测量，pAD组的全局语义相似度显著增加，同时图片-言语语义距离扩大，言语困惑度增加。只有基于fastText的全局语义相似度捕捉到了概念语义空间的收缩，它与AD人群的整体认知衰退相关。这些发现共同表明，AD中的语义空间变化在不同形式的意义上有所不同，因此指出了区分这些形式以揭示潜在机制的必要性。
