anderson, a. 1 , murphy, b. 3 & poesio, m. 1, 2
1 CIMeC, University of Trento
2 Department of Computing and Electronic Systems, University of Essex
3 School of Computer Science, Carnegie Mellon University
Understanding the neural basis of concept representation is fundamental to cognitive science, relevant to linguistic disorders, and has implications for artificial intelligence. Multivariate pattern analyses of fMRI data have identified systematic differences between neural representations of concrete object categories (e.g. animals and tools). Neural representations have been interpreted in terms of our experience interacting with objects and the associated combinations of neural sensorimotor mechanisms activated. Although recent work has distinguished between concrete and abstract concepts, little is known of the structuring of abstract sub-categories which are less easy to fit to a genus-differentia model. Nonetheless taxonomic classification schemes spanning abstract and concrete words have been devised (e.g. WordNet, DOLCE). Alternatively, concepts could be topically bound to the situation within which they occur and represented within specialised topical processing areas.
To investigate the explanatory value of taxonomic and topical classification schemes to neural organisation we devised a test set of seventy words, 50% unambiguously relating to the topic law, the others music. The words were cross classified using seven taxonomic categories from WordNet. Abstract taxonomic classes were WordNet top level categories with high abstractness scores, e.g. attribute (illegality, melody), communication (accusation, song), event (trial, festival). Concrete classes were tool (truncheon, violin) and location (jail, auditorium). Seven participants, cued by word stimuli, imagined situations associated with the word in a 4T MRI scanner. Neural network classification analyses found:
*All taxonomic and topical categories can be classified from fMRI data.
*There is no significant difference between topical or taxonomic classification accuracy.
*Taxonomy within topic structuring is more appropriate to describe abstract taxonomic classes and topic within taxonomy for concrete classes.
*One abstract and both concrete taxonomic classes can be predicted across participants. Topics can be predicted across participants, particularly for the abstract taxonomic classes.