Using Neural Word Embeddings in the Analysis of the Clinical Semantic Verbal Fluency Task

The Semantic Verbal Fluency Task is a common neuropsychological assessment for cognitive disorders: patients are prompted to name as many words from a semantic category as possible in a time interval; the count of correctly named concepts is assessed. Patients often organise their retrieval around semantically related clusters. The definition of clusters is usually based on handmade taxonomies and the patient's performance is manually evaluated. In order to overcome limitations of such an approach, we propose a statistical method using distributional semantics. Based on transcribed speech samples from 100 French elderly, 53 diagnosed with Mild Cognitive Impairment and 47 healthy, we used distributional semantic models to cluster words in each sample and compare performance with a taxonomic baseline approach in a realistic classification task. The distributional models outperform the baseline. Comparing different linguistic corpora as basis for the models, our results indicate that models trained on larger corpora perform better.