Automated analysis of the semantics in narratives by persons with Primary Progressive Aphasia.

Anonymous

Abstract: Objective To detect differences in the semantics of the spontaneous language of persons with separate primary progressive aphasia syndromes (PPA) using automated Information Control Unit derivation. The resulting semantic clusters, that describe properties of the language of the participants, are evaluated for their use in a predictive model to identify speakers with PPA. Because language is relatively easy to acquire, its quantification may provide a low cost instrument that augments other spontaneous speech analyses in clinical practice.Materials and methodsSpeech was recorded from 15 control, 8 nonfluent variant (nfvPPA) and 8 semantic variant (svPPA) speakers, each describing the actions and figures on a given picture. Transcriptions of the recordings of controls were used to compute a prototypical description. Clustering was used to identify topics. The semantic distance between the prototype and the participant's language was used to quantify the degree to which the language of persons with PPA deviates from normal language. A classifier was used to classify individual fragments.ResultsThe vocabulary of speakers with PPA was found to be less diverse in speakers with PPA. Different clusters were identified automatically that correspond with categories of objects and actions in the image. In several clusters, speakers with PPA showed significant deviations in word usage from the prototype derived from control speakers.Group level differences were found. Compared to control speakers, nfvPPA speakers use fewer auxiliary verbs and verbs that describe figure actions, and fewer nouns that describe non-human figures. SvPPA speakers use more auxiliary verbs and fewer words that describe figures and actions. Both use more words that are semantically remote.A Random Forest classifier out-performed baseline with F1 scores of 0.80 (classes: control vs. PPA) and 0.71 (classes: control vs. nfvPPA vs. svPPA).DiscussionParticipants in both PPA groups show a profile that is distinct from that of control speakers. A surprise finding is that whereas nfvPPA is usually associated with speech motor problems, our study also finds their language deviating on the level of semantics.

Paper Type: long

Research Area: NLP Applications

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: Dutch