Moving from human ratings to word vectors to classify people with focal dementias: Are we there yet?

Fine-grained variables based on semantic proximity of words can provide helpful diagnostic information when applied to the analysis of Verbal Fluency tasks. However, before leaving human-based ratings in favour of measures derived from distributional approaches, it is essential to assess the performance of the latter against that of the former. In this work, we analysed a Verbal Fluency task using measures of semantic proximity derived from Distributional Semantic Models of language, and we show how Machine Learning models based on them are less accurate in classifying patients with focal dementias than the same models built on human-based ratings. We discuss the possible interpretation of these results and the implications for the application of distributional semantics in clinical settings.