Positive Feedback in Hierarchical Connectionist Models: Applications to Language Production

Recent connectionist models of the perception and production of words make use of positive feedback from later to earlier levels of processing. This paper focuses on production and identifies several specific effects of phoneme-tomorpheme feedback. In addition. I orgue that there is support for the use of this kind of feedback in production from experimental and naturalistic studies of slips of the tongue. One feature of highly parallel network models of word and letter perception is the existence of positive feedback from the word to the letter level (Adams, 1979; McClelland & Rumelhart, 1981; Rumelhart & McClelland, 1982). So, not only do activated letter units send activation to all words that contain them, but the reverse happens as well. Words send their activation back to the letters that comprise them. This mutual backscratching between words and letters is an elegant mechanism for allowing lexical knowledge to augment stimulus information at the letter level. More generally, positive feedback from “later” to “earlier” levels of processing is a simple way of letting knowledge influence the identification of perceptual features and objects in models that employ connectionist principles. This paper considers the utility of this kind of positive feedback in another information processing domain, the production of ordered behavior. In particular, I will focus on the production of spoken words and consider, first, what positive feedback can contribute to production models in general, and second, what empirical justification there is for adding it to models of human language production.