Age and Second language acquisition and processing

Age and Second Language Acquisition and Processing: A Selective Overview David Birdsong University of Texas at Austin This article provides a selective overview of theoretical issues and empirical findings relating to the question of age and second language acquisition (L2A). Both behav- ioral and brain-based data are discussed in the contexts of neurocognitive aging and cognitive neurofunction in the mature individual. Moving beyond the classical notion of “deficient” L2 processing and acquisition, we consider the complementary question of learner potential in postado- lescent L2A. The outcome of second language acquisition (L2A) among adults is demonstrably different in many respects from the out- come of first language acquisition (L1A) among children. Depart- ing from this basic observation, researchers attempt to under- stand the various sources of age-related effects in L2A. The present article is an overview of facts and theoretical issues concerning age and L2A. This contribution considers both behavioral data and brain-based processing data. The review in- cludes findings and controversies in the areas of neurocognitive development and aging, and cognitive neurofunction in the ma- ture brain. A comprehensive treatment of the facts and issues is not possible in the space available. It is hoped, nevertheless, that this selective offering provides useful scaffolding for other articles in Correspondence concerning this article should be addressed to David Birdsong, Department of French and Italian, University of Texas, 1 University Station B7600, Austin, TX 78712-0224. Internet: birdsong@ ccwf.cc.utexas.edu 9 10 Age and L2 Acquisition and Processing this volume that examine cognitive and neural aspects of L2 use and acquisition. Background and Terminology Over the past 20 or so years, a great deal of empirical research on the age question in L2A has focused on the end state of L2A, not on rates of attainment or on stages of L2 de- velopment. The developmental literature and comparative rate (adult vs. child) literature are certainly not without interest, and overviews of this research can be found in Klein (1995), Marinova- Todd, Marshall, and Snow (2000), and Pienemann, Di Biase, Kawaguchi, and Hakansson (2005). However, it is essential that the end state receive its share of attention, because it is evidence from the end state that de- termines the upper limits of L2 attainment. Knowing the poten- tial of the learner permits inferences about the nature of puta- tive constraints on acquisition, including their relative strength and ultimate impact on learning (see Long, 1990, pp. 253–259). Accordingly, the end state is the focus of the present article. Both as a matter of logic and as a matter of theoretic adequacy, it is im- portant to recognize that when comparing L1A and L2A, a super- ficial difference in ends does not necessarily imply an underlying difference in means. Nor does similarity of ends/products nec- essarily imply similar means/processes. Thus, for example, with respect to the question of Universal Grammar’s (UG) mediating role in L2A, we understand that nativelikeness at the L2A end state does not always imply access to UG.1 By the same token, it is clear that nonnativelike linguistic behaviors are not necessar- ily evidence of lack of access to UG. Researchers must be wary of linking end-state differences in L1A and L2A exclusively to a loss of general learning ability or exclusively to some erosion of any putative mechanism(s) responsible for successful L1A. Thus, linkages between product and process are to be established only with due caution. Birdsong 11 In the literature, the terms end state, final state, steady state, ultimate attainment, and asymptote are used more or less inter- changeably to refer to the outcome of L2A. Note that “ultimate attainment” has occasionally and erroneously been used as a syn- onym for nativelike proficiency. However, the term properly refers to the final product of L2A, whether this be nativelike attainment or any other outcome. For divergent views of the construct of “end state” in L2A, see Larsen-Freeman (2005) and White (2003). For discussion of operationalizing the L2A end state, see Birdsong (2004). Researchers have explored several biographical variables that might be predictive of L2A outcomes. Age of acquisition (AoA) is understood as the age at which learners are immersed in the L2 context, typically as immigrants. This landmark is distinct from age of first exposure (AoE), which can occur in a formal schooling environment, visits to the L2 country, extended contact with relatives who are L2 speakers, and so forth. Researchers tend to equate the terms late L2A, postadolescent L2A, and postpubertal L2A; these are typically operationalized as AoA of >12 years. Length of residence (LoR) refers to the amount of time spent immersed in the L2 context. Because residence does not guarantee exposure to and use of the L2, researchers quantify the actual amount of contact L2 learners have with the L2 (in spoken and written modalities) and the relative use of the L1 ver- sus the L2 in day-to-day activities. Other experiential variables include amount of formal training in the L2 as a foreign language (e.g., grammar courses, corrective phonetics) as well as amount of exposure to the L2 in so-called content courses, where nonnatives are enrolled in high school, vocational, or university classes in the L2 country. Endogenous variables of interest to L2A researchers in- clude the following: motivation (with several subtypes relating to outcome, e.g., motivation to pass for a native, motivation to acquire lexico-grammatical accuracy), psycho-social integration with the L2 culture, aptitude (with several presumed components, 12 Age and L2 Acquisition and Processing including imitative ability, working memory capacity, metalin- guistic awareness, etc.), and learning styles and strategies. These are understood to be continuous, not all-or-nothing, variables. For overviews of these variables, see Do¨rnyei and Skehan (2003) and Doughty (2003). AoA and L2 Ultimate Attainment It is widely recognized that AoA is predictive of L2A out- comes, in the simple sense that AoA is observed to significantly correlate negatively with attained L2 proficiency at the end state. This conclusion is based on the results of more than two dozen experimental studies; see Birdsong (2005) and DeKeyser and Larson-Hall (2005) for overviews. The areas of language most commonly investigated are morphosyntax and pronunciation. Typically, morphosyntax errors in production or grammaticality judgments increase with advancing AoA, as does degree of judged nonnative accent. Across many studies that examine AoA and other factors that might be related to L2 success, it has emerged that, of all the above-mentioned experiential variables, AoA is reliably the strongest predictor of ultimate attainment. This is not to say that other variables, indeed some that are confounded with AoA, are not predictive. In many cases, variables such as LoR and AoE are controlled statistically or included as factors in the experimental design. The Age Function From the actual behavioral data, a recurrent finding is that a linear function captures the relationship between AoA and out- come over the span of AoA (i.e., when considering aggregate data from both early- and late-AoA subjects). In 10 surveyed studies, the range of correlations is .45 to .77, with a median of about .64 (all absolute values).2 The slope of the age function varies (i.e., it is steeper or shallower) as a function of such factors as L1-L2 Birdsong 13 pairing, amount of L2 use, task, education in the L2, and so on. It is also not surprising to find, given what is known about learn- ing and cognitive performance over the life span (Schaie, 1994; Weinert & Perner, 1996), that there is less intersubject variation in outcome among early arrivals than among late arrivals. When data from early- and late-AoA subjects are disaggre- gated, inconsistent results are obtained, producing a clouded pic- ture of the timing and geometry of the age function. For example, DeKeyser (2000) studied 57 Hungarian L1 English L2 subjects with AoA ranging from 1 to 40 years, all with at least 10 years of U.S. residence. On a grammaticality judgment test using some items from Johnson and Newport (1989) along with some novel items, a significant correlation of AoA with scores was obtained (r = −.63, p < .001). However, when DeKeyser broke out the data by early- and late-arriving subjects, neither set of data yielded a significant correlation with AoA (early arrivals n = 15, r = −.24, ns; late arrivals n = 42, r = −.04, ns). Another illustration of the disparate results of analyses of aggregate versus disaggregated data is seen in the comparison of the results of Johnson and Newport (1989) and Birdsong and Molis (2001). Johnson and Newport looked at accuracy on a 276- item grammaticality judgment by a group of Chinese and Korean natives (n = 46) with English as their L2. The Birdsong and Molis study was a strict replication of Johnson and Newport, but in this case, the subjects were Spanish natives (n = 61). Over all subjects and AoAs, Johnson and Newport found a strong linear relationship between AoA and accuracy (r = −.77, p < .01). This finding was reproduced by Birdsong and Molis (r = −.77, p < .0001). However, when the subjects were divided into AoA groups of ≤16 years and >16 years, the analyses produced divergent results. Figure 1 represents these differences. The pattern of results seen in Johnson and Newport (1989) is a decline in scores with increasing AoA for early arrivals (r = −.87, p < .01) and an essentially random distribution of scores for the older-arriving group (r = −.16, ns). A quite different pattern was obtained by Birdsong and Molis (2001). For early arrivals, 14 Age and L2 Acquisition and Processing Figure 1. Plot of accuracy over AoA from Birdsong and Molis (2001, p. 240). Solid regression lines are fit to the Birdsong and Molis data; dashed lines are fit to the Johnson and Newport (1989) data. Early and late AoA groups are divided at 16 years. the correlation of scores with age is not significant (r = −.24, p = .22), as this subgroup performed at ceiling. For late arrivals, the correlation is strongly negative (r = .69, p < .0001). In a reexamination of the Johnson and Newport (1989) data, Bialystok and Hakuta (1994) moved the cutoff point separating early- and late-arriving groups to 20 years. For late learners, the subsequent correlation reached significance (r = −.50, p < .05). Birdsong and Molis (2001) conducted a similar reanalysis of their data, placing the cutoff at various ages between 15 years and 27.5 years; all correlations reached significance. The meta-analysis by Birdsong (2005) of L2 end-state mor- phosyntactic and pronunciation behavioral research arrives at three main conclusions: (a) In all analyses of pooled data from early and late arrivals, age effects persist indefinitely across the span of surveyed AoA (i.e., they are not confined to a Birdsong 15 circumscribed period); (b) In analyses of disaggregated samples (and in studies that look only at late AoA), most studies find sig- nificant AoA effects for the late learners, indicating postmatu- rational declines in attainment; (c) in analyses of early-arrival data alone, AoA effects are inconsistent: Some are flat, some are random, and some are monotonically declining. Do Observed AoA Effects Suggest a Maturationally Based Critical Period? We can now take a step back and consider whether observed AoA effects can be interpreted as critical period effects.3 If what we are dealing with is in fact a period, the age effects observed in the data must be confined to a finite time span; see Bornstein (1989) for a further discussion of characteristics of a critical pe- riod. Moreover, if the effects are maturational in nature, then the age function prior to the end of maturation should look different from the age function after the end of maturation. Taken together, the requirement of finite age effects and a discontinuity in the age function synchronized with the end of maturation permute into three basic patterns (see Figure 2). One is a stretched “L” or hockey stick shape, with age-related de- clines ceasing at a point of articulation that coincides with the end of maturation. The second is an upside-down mirror image of the stretched “L,” resembling a stretched “7.” The flat portion at the top left of the image is the period where success is guaran- teed. A third possibility, laid out by Johnson and Newport (1989) and expanded by Pinker (1994), specifies a causal role of brain Figure 2. Three patterns of bounded age effects: (A) stretched “L” shape; (B) stretched “7” shape; (C) stretched “Z” shape. 16 Age and L2 Acquisition and Processing maturation in L2A age effects, with the end of age effects syn- chronized with the completion of brain maturation. This version combines features of the first two possibilities to produce the im- age of a stretched “Z.” The function begins with a period of ceiling effects, followed by a decline that ceases at the end of maturation, after which the age function flattens and no further age effects are seen. Let us consider the third possibility first. The stretched “Z” shape (Figure 2C) includes two finite periods. At the upper left portion of the image, where performance is at ceiling, we indeed observe a bounded period, which is actually a period during which age effects are absent, as there is no downward slope in the age function. The next segment is a bounded downward slope; the age effect begins prepubertally and ends at the completion of maturation. The third segment, which is unbounded, captures the hypothesized bottoming out or flattening of the age function. Johnson and Newport (1989) purport to have produced findings consistent with the timing and geometric features just described. However, instead of an orderly array of scores parallel to the x- axis—that is, the hypothesized floor effect—one finds a random dispersion of points. In other words, the crucial flattening feature of the function, whose beginning should coincide with the end of maturation, is in fact not present in the data. Moreover, as mentioned earlier, if following Bialystok and Hakuta (1994), one moves the cutoff point to 20 years, the late- arrivals data in Johnson and Newport (1989) start to look a bit more orderly. The result of the ensuing analysis is neither a ran- dom distribution nor a floor effect, but a significant negative cor- relation of AoA and performance for the late-arriving group. The stretched “L” or hockey-stick representation (Figure 2A) incorporates a sloping segment on the left that would satisfy the requirement of a bounded period during which AoA is neg- atively correlated with outcomes. It also contains a flattened seg- ment, the beginning of which coincides with the end of matura- tion. A review of the literature (see Birdsong, 2005; DeKeyser & Larson-Hall, 2005) reveals that several analyses of disaggregated Birdsong 17 data show prematurational declines—the left portion of the stretched L. However, for later learners (i.e., those whose perfor- mance would be represented by the right segment of the stretched “L”), there is no evidence of a flat function or floor effect. Instead, for late-learner groups, there is either a random array of scores (e.g., DeKeyser, 2000; Patkowski, 1990) or a persistent decline in performance with increasing AoA (e.g., Bialystok & Miller, 1999; Birdsong, 1992). Returning to Figure 2A, we note that the appear- ance of a stretched “L” shape (i.e., the two rightward segments of the “Z”) is obtained for the Johnson and Newport (1989) data when linear functions are applied separately to early- and late-arrival data. A systematic performance decline over AoA is indeed ob- served for early arrivals (r = −.87). However, as we saw earlier, for late arrivals, a flat segment is a misleading representation of the correlation coefficient in this instance (r = −.16), as the best-fitting near-horizontal regression line actually goes through a random array of scores, not through an orderly set of points that are parallel to the x-axis. The final scenario by which age effects would be considered critical period effects is the mirror image of the one just discussed, a stretched “7” or upside-down hockey stick shape with the “blade” at the top left (Figure 2B). This is an unconventional, although often implicitly invoked, notion of a critical period function (see Birdsong, 2005, for discussion of conventional and unconventional conceptions; based on Bornstein, 1989). The leftmost part of the function is flat, with performance at ceiling. On the right portion of the image, the age gradient (i.e., the decline in ultimate attain- ment with advancing AoA) is not bounded. What is bounded is the left segment of the image, the period of peak attainment, which is often referred to as a “window of opportunity”—the temporal span during which sensitivity or learning potential is at its highest and full attainment is guaranteed. Such a period has been observed in at least one study: Birdsong and Molis (2001).4 As seen in Figure 1, a roughly flat function at ceiling is generated by the per- formance of the early-arriving AoA group of Spanish L1 speakers. This “age noneffect” is confined to a limited span, thus satisfying 18 Age and L2 Acquisition and Processing the geometric criterion and corresponding to the unconventional “window of opportunity” version of the critical period. However, because of the apparent duration of the window of opportunity, the temporal features do not conform to a maturational account of AoA effects. For their L2 learners’ results, Birdsong and Molis (pp. 241–242) conducted a series of post hoc piecewise regression analyses that included the inflection point (i.e., the terminus of the period) as a free parameter. Under these conditions, the best- fitting function placed the end of the ceiling period, and thus the beginning of the decline, at 27.5 years. In other words, the pe- riod of peak performance extends 10 or more years beyond the end of maturation. Thus, although the Birdsong and Molis re- sults reveal a stretched “7” shape and its circumscribed period of full attainment, the temporal parameters do not mesh with a maturational-effects account of L2 ultimate attainment. Divergent Conceptualizations of “Critical Period” Singleton (2005) examined several proposals for the timing of the “end of the critical period.” In most cases, these propos- als made reference to the end of the period of peak sensitiv- ity; that is, they invoked the “window of opportunity” notion of critical period. In the studies that Singleton surveyed, hypothe- sized beginnings of declines ranged from near birth to late ado- lescence. Some proposals made distinct timing claims for pho- netics/phonology versus other areas of linguistic knowledge and performance. Such so-called “multiple critical period” accounts of attainment in various language domains were advanced by Long (1990) and Seliger (1978) for the L2 context and are consistent with current neurobiological thinking about critical periods in other contexts (Knudsen, 2004). The proposals of Johnson and Newport (1989), Lenneberg (1967), Long (1990), Pinker (1994), Scovel (1988), and Seliger (1978) signaled changes that occur around puberty. Signifi- cantly, in some cases, this maturational milestone is thought to be the point at which declines in performance begin (i.e., the Birdsong 19 unconventional notion of critical period), and in other cases, this maturational milestone is thought to be the point at which per- formance declines cease (the conventional notion). Thus, a serious conceptual issue confronts proponents of a m