Multiplex lexical networks and artificial intelligence unravel cognitive patterns of picture naming in people with anomic aphasia

Aphasia is a language disorder which impairs people’s ability to comprehend or produce words. The mechanisms behind this disorder are not yet fully understood mainly because of the challenge of interpreting them through large-scale quantitative models. To this aim, we use artificial intelligence and knowledge graphs to investigate picture naming in people affected by anomic aphasia. Our knowledge graphs encode four aspects of associative knowledge: free word associations, synonyms, generalisations and phonological similarities. We then use these networks to compute features of target words and mistakes in producing them as recorded in a psychological mega-study with 31700 utterances. Adopting a human-centric AI approach, we train an artificial general intelligence (AGI) to predict the type of mistake (formal, semantic or mixed) according to network distances and individual level psychological norms. Our results reveal some key relationships between the multiplex structure and the errors made: (i) Network distance is found to be predictive of the error type but only when considered independently across each layer (accuracy: 73.3%, precision: 74.1%, recall 70.9%); (ii) The most predictive model is achieved when closeness centrality, rather than other psychological norms is added to the four network distances (accuracy: 80.9%, precision: 80.1%, recall 79.7%). We find that the ability to predict different types of mistakes crucially depends on the presence or absence of different aspects of knowledge in the network. In particular, removing free associations damages predictions of all mistakes the most. This indicates the importance of free associations in driving picture naming.