The Emergence of Words:Attentional Learning in Form and MeaningTerry RegierDepartment of Psychology, University of ChicagoReceived 3 December 2003; received in revised form 16 October 2004; accepted 16 February 2005AbstractChildren improve at word learning during the 2nd year of life—sometimes dramatically. This fact hassuggested a change in mechanism, from associative learning to a more referential form of learning. Thisarticle presents an associative exemplar-based model that accounts for the improvement without achange in mechanism. It provides a unified account of children’s growing abilities to (a) learn a newword given only 1 or a few training trials (“fast mapping”); (b) acquire words that differ only slightly inphonological form; (c) generalize word meanings preferentially along particular dimensions, such asobject shape (the “shape bias”); and (d) learn 2nd labels for already-named objects, despite a persistingresistance to doing so (“mutual exclusivity”). The model explains these improvements in terms of in-creased attention to relevant aspects of form and meaning, which reduces memory interference. The in-teraction of associations and reference in word learning is discussed.Keywords: Word learning; Fast mapping; Phonological detail; Shape bias; Mutual exclusivity;Associative learning; Selective attention; Exemplar-based; Connectionist; Memory interference;Reference; Vocabulary spurt; Lexical development; Language acquisition1. IntroductionHow do children learn words? One possibility is that they rely in part on simple associativelearning (e.g., Pavlov, 1927), of the sort found in nonhuman animals. Children might link thesound ball to the idea of a ball using the same general principles that allow a dog to associate abell with an upcoming meal.This cannot be all there is to it. In learning words, children rely on many skills beyond asso-ciation, including social awareness of reference (Baldwin, Markman, Bill, Desjardins, &Irwin, 1996), inference about events that are not present (Gleitman, 1990), knowledge of syn-tax (Fisher, Gleitman, & Gleitman, 1991), and pragmatic inference (Clark, 2004), among oth-Cognitive Science 29 (2005) 819–865Copyright © 2005 Cognitive Science Society, Inc. All rights reserved.Requests for reprints should be sent to Terry Regier, Department of Psychology, University of Chicago, 5848South University Avenue, Chicago, IL 60637. E-mail: [email protected] (Bloom, 2000). Yet associative learning can account for some aspects of word learning(e.g., Merriman, 1999; Plunkett, Sinha, Møller, & Strandsby, 1992; Smith, 2000). What is notyet clear is just how much of word learning this simple idea can explain.One influential proposal is that associative learning accounts for only the initial stages ofword learning, in which words are acquired fairly slowly. Then, on this view, sometime duringthe 2nd year of life, the child has a conceptual insight into the symbolic, referential nature ofwords (Lock, 1980; McShane, 1979). This insight qualitatively changes the nature of the wordlearning mechanism, allowing the child to learn words more quickly and effectively.Several improvements in word learning occur at around this age, and when viewed together,they do suggest a mechanistic change of some sort. I argue, however, that these changes can beaccounted for without recourse to a referential insight during the 2nd year, or any other qualita-tive change in mechanism. Instead, they can be explained in a unified fashion by an associativemodel that gradually learns to attend to communicatively relevant aspects of the world(Kruschke, 1992; Smith, 1989). These learned attentional shifts reduce memory interferenceand improve word learning. This work expands on existing attentional proposals (e.g.,Merriman, 1999; Smith, 2000; Smith, Jones, Landau, Gershkoff-Stowe, & Samuelson, 2002)in bringing together these parallel changes in the child’s treatment of both form and meaning(Hespos & Spelke, 2004).This proposal is compatible with the recent suggestion that children have an earlier in-sight—at around 12 months—into the social bases of reference and communication. That in-sight may initiate word learning (Tomasello, 1999) and set in motion gradual attentional shifts.After several months, attention may have shifted enough to bring about the improved wordlearning we see in children partway through the 2nd year of life.2. Changes in word learningDuring the 2nd year, the child’s word-learning behavior changes in at least four respects:ease of learning, honing of linguistic form, honing of linguistic meaning, and the learning ofsecond labels.2.1. Ease of learningAs children first begin to produce words, their acquisition of new words is slow and errorful.Between 12 and 16 months of age, children tend to learn new words at the rate of two or threeper week (Fenson et al., 1994; Gershkoff-Stowe & Smith, 1997). Later, around 18 to 22months of age—often when the child has about 50 words in productive vocabulary—the acqui-sition of new words accelerates, sometimes sharply (a “vocabulary spurt,” but see also Ganger& Brent, 2004, who show that the increase is usually more gradual than has been assumed).Many children at this age acquire eight or nine words per week (Fenson et al., 1994), and re-ports show some children learning as many as 44 new words in a week (Dromi, 1987).Experimental studies in the laboratory provide further evidence for a shift from slow toquick word learning. Thirteen- to 15-month-olds can acquire a word–object linkage in com-prehension based on 9 to 12 training trials (Schafer & Plunkett, 1998; Woodward, Markman, &820 T. Regier/Cognitive Science 29 (2005)Fitzsimmons, 1994). By the time children are 2 to 3 years of age, one to three training trials aresufficient (Behrend, Scofield, & Kleinknecht, 2001; Carey, 1978). This rapid, efficient wordlearning is known as fast mapping. Children may retain their knowledge of such a newlylearned word for up to a month (Markson & Bloom, 1997).The topic of fast mapping has attracted considerable attention (Behrend et al., 2001; Bloom,2000; Bloom & Markson, 2001; Markson & Bloom, 1997; Waxman & Booth, 2000; see alsoDickinson, 1984; Dollaghan, 1987; Heibeck & Markman, 1987). This interest stems in partfrom the idea that fast mapping may reflect a specialized mechanism for word learning—partof an overall human predisposition for language. Another possibility, however, is that fast