Word Category-Specific Semantic Deficits in Mild Cognitive Impairment: A BERT Embedding Analysis of Picture...

Mild cognitive impairment (MCI), a transitional state between normal aging and dementia, is associated with subtle but systematic disruptions in semantic memory, yet the representational mechanisms underlying these changes remain poorly understood. Grounded in neuropsychological evidence that semantic vulnerability in early neurodegeneration is selective across lexical categories, this study investigates whether MCI is associated with the category-specific contraction of semantic representational space. We predicted that concrete words would show reduced semantic dispersion in MCI, whereas abstract words and function words would remain mostly unaffected.

Narrative transcripts from 102 participants with MCI and 142 neurotypical controls were drawn from the DementiaBank database. Transcripts were preprocessed and processed independently using the Bidirectional Encoder Representations From Transformers (BERT) base-uncased model to extract contextualized embeddings from the final hidden layer. Semantic dispersion was quantified as Euclidean distances from individual word embeddings to group-specific centroids within each word category. Differences were analyzed using linear mixed-effects models with fixed effects for group, word category, and their interaction, controlling for word length and log-transformed lexical frequency, with random intercepts for participants and random slopes for word category within participants.

Linear mixed-effects modeling showed significantly reduced semantic dispersion in the MCI group overall, driven specifically by the concrete word category (Holm-adjusted p = .001). No significant group differences were observed for function or abstract words.

MCI is associated with a contracted semantic space selectively for concrete words, suggesting the early degradation of perceptually grounded semantic representations, while abstract and function word representations remain relatively preserved. These findings highlight BERT embeddings as a sensitive, theoretically interpretable tool for detecting category-specific semantic deficits in MCI, offering a mechanistic account of language changes that complements traditional neuropsychological approaches.

https://doi.org/10.23641/asha.33060893.