Improving machine learning approaches to (sub)classification of PrimaryProgressive Aphasia using fine-grained linguistic features

The proposal, in short. Machine Learning approaches can perform a successful classification of Primary Progressive Aphasia (PPA) variants (Garrard et al. 2013). The accuracy of these methods for PPA sub-classifications is promising, also in very sparse contexts of connected speech productions (picture description elicitation task, generating speech samples smaller than 100 tokens). This result has been obtained by including highly informative phonetic, morpho-syntactic and semantic feature information, mainly consisting of phoneme frequency, (bi)-syllabic repetition patterns, out-of-vocabulary term frequency, cues for syntactic truncated structures and characterizing low-frequency content word distribution.