Beyond Individual Features: Computational Analysis of Linguistic Coordination in Dementia Speech

In dementia, progressive changes in spoken language, including reduced lexical diversity, increased disfluency, and syntactic simplification, are well documented. Although computational approaches have demonstrated reliable group differences in individual linguistic features, less attention has been given to the potential coordination between linguistic features and discourse-level moderation of the features with group differences.

Over five domains-syntactic, lexical, disfluency, output, and affective-nine linguistic features were extracted from 552 transcripts (309 dementia, 243 healthy controls) of individuals performing the Cookie Theft picture description task in the DementiaBank Pitt Corpus. The Mann-Whitney U test was used to examine group differences with rank-biserial correlation effect sizes. Fisher z-transformation tests were applied to compare within-group correlation to examine the moderation effect of discourse-pragmatic factors (self-references, hedging, and sentiment polarity) on different groups. Fisher z-transformation tests were used to compare inter-feature correlations between the two groups. All multiple comparisons were controlled using the Benjamini-Hochberg false discovery rate (FDR) procedure.

Out of the nine features, word complexity, repetition rate, MLU, and filler rate were significantly different between the group with dementia and healthy controls. Among them, word complexity and repetition rate showed the largest and most consistent effects. Self-referential language showed a stronger association with verbal output in the group with dementia, although the association was not significant when analyses were restricted to one visit per participant. Moreover, one inter-feature relationship (i.e., the positive correlation between short utterance ratio and word complexity) remained significant after FDR correction. This correlation was present only in healthy controls, suggesting that this preserved coordination may be disrupted in dementia.

Dementia disrupts not only individual linguistic features but also the relationships between them. The preserved coordination between syntactic simplification and lexical elaboration was found to be absent in speakers with dementia. Inter-feature coordination may, therefore, be able to reveal group differences that individual feature analyses alone cannot detect.