Can Distributed Orthographic Knowledge Support Word Specific Long-Term Priming? Apparently So

A property of distributed representations is that related information is coded as overlapping patterns of activation over the same set of units and learning associated with one item extends to related items. Accordingly, the null (or near null) long-term priming observed between form-related words seems to pose a challenge to connectionist theories of reading that include distributed codes. In the present report, priming was assessed in a behavioral study and a computer simulation using Seidenberg and McClelland's (1989) distributed model of word identification. Contrary to our expectation, both the behavioral and simulation studies obtained robust repetition and little form priming. Furthermore, analysis of the model's performance revealed that the lack of form priming was the product of collapsing facilitatory effects between rhymes (mint-hint) and inhibitory effects between nonrhymes (pint-hint). A second behavioral experiment confirmed this prediction. A number of additional long-term priming results were also successfully modeled.