A Connectionist Model of Semantic Memory: Superordinate structure without hierarchies

A C onnectionist M odel of Sem antic M em ory: Superordinate structure w ithout hierarchies G eorge S. C ree (gcree@ uw o.ca) D epartm ent of Psychology, 1151 Richm ond Street London, O ntario N 6A 5B8 Canada K en M cR ae (kenm @ uw o.ca) D epartm ent of Psychology, 1151 Richm ond Street London, O ntario, N 6A 5B8 Canada Sym bolic, spreading-activation m odels of sem antic m em ory represent subset-superset relationships am ong concepts as distinct, hierarchical levels of nodes connected by “isa” links (e.g., Q uillian, 1968). N um erous theoretical and em pirical argum ents have been leveled against this approach (e.g., D ean & Slom an, 1995; Rum elhart & Todd, 1993), including (1) the difficulty such m odels have in accounting for fam iliarity and typicality effects, (2) that category m em bership is often unclear, (3) that item s can belong to m ultiple categories, (4) that som e categories are m ore internally coherent than others, (5) that general properties do not necessarily take longer to verify than specific properties, and (6) that som e general category m em bership relations can be verified faster than specific category m em bership relations. W e present a novel connectionist m odel of sem antic m em ory that offers potential solutions to these problem s. The m odel, an extension of M cRae, de Sa & Seidenberg's (1997) and Cree, M cRae & M cN organ's (1999) m odels of sem antic m em ory, w as trained to com pute distributed patterns of sem antic features from w ord form s. Sem antic feature production norm s w ere used to derive basic-level representations and category m em bership for 181 concepts taken from M cRae et al’s (1997) property norm s. Basic-level (e.g., dog) and superordinate-level (e.g., anim al) concepts w ere represented over the sam e set of sem antic features. The training schem e w as designed to m im ic the fact that people som etim es refer to an exem plar w ith its basic-level label, and som etim es w ith its superordinate- level label. Tw o types of training trials w ere used. In 90% of the training trials, basic-level w ord form s m apped to their sem antic representation, instantiating a one-to-one m apping. The occurrence of each of the 181 basic-level exem plars during training w as scaled by fam iliarity ratings that w ere collected from hum an participants. In the rem aining 10% of the trials, a superordinate w ord form w as trained by pairing it w ith one of its exem plars’ sem antic representations. Im portantly, each sem antic representation included in a category w as paired w ith that superordinate w ord form w ith equal frequency (i.e., typicality w as not built in). The m odel w as used to sim ulate data from typicality, superordinate-exem plar prim ing, and category- verification experim ents. In explaining the hum an data, em phasis w as placed on the role of correlations am ong features, the fam iliarity of concepts, category size, and on the distinction betw een off-line and on-line processing dynam ics. Specifically, settled attractor states for superordinate-level concepts are com posed of a greater num ber of units w ith states on the linear com ponent of the sigm oidal activation function, m aking it easier, for exam ple, for the netw ork to m ove from a superordinate representation to any other during tem poral, on-line processing. A cknow ledgm ents This w ork w as supported by an N SERC Postgraduate Fellow ship to the first author and N SERC grant RG PIN 155704 to the second author. R eferences Cree, G .S., M cRae, K . & M cN organ, C. (1999). A n attractor m odel of lexical conceptual processing: Sim ulating sem antic prim ing. Cognitive Science, D ean, W . & Slom an, S.A . (1995). A connectionist m odel of sem antic m em ory. U npublished M anuscript. M cRae, K ., de Sa, V .R. & Seidenberg, M .S. (1997). O n the nature and scope of featural representations of w ord m eaning. Journal of Experim ental Psychology: G eneral, 126, 99-130. Q uillian, M .R. (1968). Sem antic M em ory. In M . M insky [Ed.], Sem antic Inform ation Processing (pp. 216-270). Cam bridge, M A : M IT Press. Rum elhart, D .E. & Todd, P.M . (1993). Learning and connectionist representations. In D .E. M eyer and S. K ornblum [Eds.], Attention and Perform ance XIV: Synergies in experim ental psychology, artificial intelligence, and cognitive neuroscience (pp. 3-30). Cam bridge, M A : M IT Press.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.