Feature distributions and experimental evaluation in a connectionist model of semantic memory

This paper describes evaluative work carried out with a connectionist model of semantic memory investigated by Rumelhart, and later by McClelland et al. Two critical issues were investigated, the nature of the semantic associations learned by the model, and the ability of an adapted version of the model to simulate human priming data. Analysis of the correspondence between the semantic associations developed in the network and those expected from human data indicated two sources of concern. First, the semantic associations learned were susceptible to small changes in the training set. Second, surprising semantic relations were found (e.g. ‘animal’ was more similar to the plants in the model). Priming was studied by adding a cascade mechanism to the basic model. Certain aspects of the training data and model architecture made direct simulation of standard human semantic priming difficult, since primes and targets necessarily consisted of a concept and a relation term, e.g. prime ‘oak is’ — target ‘tree is’. Responses consisted of activation of appropriate features for the target term. Two interesting findings emerged. First, dissimilar primes slowed performance compared to unprimed processing. Second, the major causal factor in the priming effects obtained was due to the correspondence between relation terms rather than similarity of concepts. In light of these findings, we discuss the value, for connectionist modelling, of training data that encodes distribution information derived from the study of human semantic memory.