Individual Differences in Implicit and Explicit Learning.

Reber (1993) argued that implicit (unaware, unintentional) learning and memory processes are primary. This argument was based on the observation that what can be learned and also remembered implicitly often appears to exceed what can be learned and remembered explicitly (intentionally, with awareness). Implicit processes are primary in the sense that they evolve earlier in childhood and are more immune than their explicit counterparts to the effects of age-related declines in cognitive function. Reber et al. (1991) explained these findings in evolutionary terms: Implicit learning and memory are the “functional instantiations of a phylogenetically primitive system that developed before the emergence of conscious functioning” and are dissociable from conscious cognitive processes “such as problem solving, decision-making, recognition and recall from short- and long-term memory … and so on” (p. 888). Reber et al. argued that implicit processes should therefore display the following properties that reflect this dissociation: Implicit processes should (a) be more robust in the face of psychiatric and neurological insult than explicit processes, (b) display fewer individual differences and tighter population variances, and (c) operate largely independently of standard measures of cognitive capability, such as IQ.

To test the latter two claims, Reber et al. (1991) asked 20 undergraduate psychology majors to take part in an AG implicit learning task and a series-solution explicit learning task during which participants searched for rules.¹

Relevance to SLA Theory and L2 Pedagogy.

What are the patterns of cognitive abilities—or aptitude complexes (Robinson, 2001)—that facilitate incidental L2 learning? Reber's (1993) position, described previously, is that the abilities drawn on in implicit AG learning are nonconscious and evenly distributed in all populations: There is no aptitude for implicit learning. Very similar claims to those of Reber have been made by Krashen (1982) and Zobl (1995) with regard to adult L2 learning. Unconscious acquisition of natural L2s exceeds what can be learned consciously, and acquisition is insensitive to measures of the conscious cognitive processes assessed via various language learning aptitude tests, such as the Modern Language Aptitude Test (MLAT; Carroll & Sapon, 1959). By acquisition, Krashen meant incidental learning while processing comprehensible input for meaning, during which L2 learners unintentionally pick up knowledge of the constraints on grammar. However, it is likely that the learners become aware of these constraints to some extent during the process. Reber conceded that “incidental learning is not identical to implicit rule or pattern learning, but it is similar in the sense that it is unintentional and uncontrolled” and reported a number of studies of incidental learning that also show it to be “unrelated to gross measures of high-level cognitive function” (Reber & Allen, 2000, p. 238). Reviewing studies of incidental learning that require “processing of verbal or complex material” and thus greater depth of processing, Reber and Allen conceded that individual differences (IDs) in cognitive abilities might influence learning in these domains. However, they noted that “the issue of individual differences in implicit learning has gone almost completely unexplored” (p. 239), and there is currently no research that enables a direct comparison of the influence of IDs in cognitive abilities on implicit and incidental learning.

Therefore, to directly examine the generalizability of Reber et al.'s (1991) finding for implicit AG learning to incidental natural L2 learning, I added a third experimental condition to the implicit and explicit conditions of the Reber et al. study. In this condition, learners processed sentences in a novel language (Samoan) for meaning and assessed the extent to which they had learned aspects of grammar as a result. As described subsequently, learners first rote-learned the meaning of Samoan words and were then directed only to answer yes/no questions about the meaning of sentences containing the words. They were not directed to learn the rules of Samoan grammar and morphology, and any learning of these was therefore incidental to the intentionally primary experimental goal of comprehending the input to answer the questions. In addition to the measure of WAIS-R IQ that Reber et al. obtained, I added a measure of language learning aptitude (Language Aptitude Battery for the Japanese [LABJ]; Sasaki, 1996). This consisted of subtests of grammatical sensitivity, paired-associate rote memory, and phonemic sensitivity and pattern analysis. I also added a measure of working memory capacity, assessed via a reading-span test (Osaka & Osaka, 1992). These additional individual difference measures enabled the comparability of learning in the implicit and incidental conditions to be made not only with respect to the influence of IQ but also with respect to other variables that are more likely related to incidental L2 learning, thus examining Krashen's (1982) claim about the aptitude independence of acquisition.

Rationale for Hypotheses.

Hypothesis 1 (Population generalizability): Reber et al.'s (1991) finding will be replicated. It is likely that the population in this replication and Reber et al.'s undergraduate psychology major participants differed in terms of the strength and patterning of some cognitive abilities drawn on in learning natural L2s. However, the two populations had almost identical mean IQ scores. I therefore expected that implicit AG learning would be nonsignificantly correlated with IQ, whereas explicit learning would be significantly positively correlated with IQ and display significantly greater variance than implicit learning.

Hypothesis 2 (Content generalizability): AG learning and incidental L2 learning will be unrelated. I did not expect implicit and incidental learning to show very similar effects for learners. I expected that scores on transfer tests of implicit and incidental learning would be nonsignificantly correlated, given that the content domain of the AG and natural language as well as the processing conditions these domains implicate are very different. Artificial grammar letter strings are free of semantic content. Processing AG stimuli can therefore be expected to result in predominantly bottom-up, shallow data-driven processing and maintenance memory rehearsal of the contingencies between letters in the strings presented during training. There can be no doubt that the associative learning and memory that such processing leads to contributes to SLA to some extent, as it does in incidental first language (L1) learning (Saffran, Newport, Aslin, Tunick, & Barrueco, 1997). However, processing sentences of a natural L2 for meaning additionally requires explicit, top-down conceptually driven processing of their semantic content and the subsequent encoding and elaborative rehearsal of this content in episodic and working memory (Craik & Tulving, 1975).²

Both of these kinds of processing, although differing in the extent of their joint contribution to AG versus incidental natural language learning, are similar in important respects. Both require attention (Hsiao & Reber, 1998; Schmidt, 2001).³

Hypothesis 3: Individual differences in cognitive abilities will not influence AG learning. In line with this account, AG learning demands attention but not the higher-order cognitive abilities that are drawn on in conceptually driven processing of semantic content. Learning the constraints on sequencing the AG stimuli should therefore take place independently of the influence not only of IQ but also working memory and the LABJ measure of aptitude.

Hypothesis 4: Individual differences in cognitive abilities will influence incidental Samoan learning. Although individual differences were not expected to influence AG learning, I expected significant positive correlations of aptitude and working memory with incidental learning of Samoan. This expectation is based on the reasons given previously, as both working memory and the grammatical sensitivity and paired-associates subtests of aptitude (although not the phonological sensitivity subtest) are measures of ability to process and elaboratively rehearse semantic information.

Chunking, Frequency, and Exemplar-Specific Implicit Learning.

Within the field of SLA, N. Ellis (2001) has argued that chunking in working memory—facilitated by frequency of exposure to instances of L2 input—is one of the basic processes that gives rise to L2 learning, as it does to the learning of many other kinds of category and concept. Others have argued that the availability of domain-specific representations in the form of abstract rules, such as those characterized in Universal Grammar (UG) approaches to SLA, are needed to explain much of how L2s are acquired. Such rules can be triggered and set by a single exposure to a domain-specified class of language input (see White, 2003). This debate concerning the nature of representations (are they rule- or instance-based, domain-specific or not?) is currently ongoing in many other areas of cognitive psychology (see Pothos, in press). The second study described here (Knowlton and Squire, 1996) addressed this issue. According to Knowlton and Squire's study, implicit AG learning gave rise to abstract rule-based as well as instance-based representations. They argued that instance-based representations were facilitated by their manipulation of the frequencies of specific chunks in the input during AG learning, whereas the abstract rules learned were independent of these manipulations. In replicating their experiment, I was interested in adapting their methodology to also examine this issue in the context of incidental natural L2 learning.

Knowlton and Squire (1996) operationalized chunk-strength following proposals by Servan-Schreiber and Anderson (1990) and the similar proposals of Perruchet and Pacteau (1990), which argued that the process of competitive-chunking was responsible for the learning that occurred during the implicit conditions of Reber's (1976) experiments. Chunks were operationalized as bigrams (pairs of letters) and trigrams (sets of three letters) occurring together in training set examples, and chunk-strength was calculated by weighting chunks for their frequency of occurrence throughout the period of exposure during training.⁴

Relevance to SLA Theory and L2 Pedagogy.

The issues raised by Knowlton and Squire's (1996) study are also important for SLA theory and pedagogy. N. Ellis (2002, this issue) has argued that frequency effects and chunking in working memory are both powerful influences on the nature and extent of both implicit and explicit SLA. If so, then chunk-strength during naturalistic exposure to L2 input and manipulations of it during controlled instructional exposure should both facilitate further L2 learning. Providing L2 learners with an input flood that contains frequent instances of a particular form has in fact been proposed as an implicit technique for inducing attention to form and subsequent internalization and use of the targeted form during classroom instruction (see Doughty & Williams, 1998).

However, if L2 learners in the present replication display the same influence of chunk-strength on grammaticality judgments in both AG and incidental L2 conditions, as Knowlton and Squire's (1996) participants did, there are interesting implications for these theoretical and pedagogic proposals. This would imply that manipulations of chunking and frequency in the input contribute to the development of L2 knowledge but that other processes resulting in rule-based abstraction also take place independently of these. It would also suggest that although pedagogically manipulating the frequency of a chunk of input to learners can lead to knowledge of collocational restrictions and lexical learning (N. Ellis, 2002; Mintz, 2003; Robinson, 1989, 1990) and promote automatic access to and recognition of these (Logan, 1988; Robinson, 1997a; Robinson & Ha, 1993), it might be relatively ineffective in developing knowledge of which sentences are grammatically acceptable (as revealed by the poor performance of Knowlton and Squire's participants on the ungrammatical items in the grammaticality judgment transfer test). Therefore, to directly examine the generalizability of Knowlton and Squire's finding to natural L2 learning, I also assessed the effects of chunk-strength on transfer test judgments in the third Samoan learning condition.

Rationale for Hypotheses.

Hypothesis 5 (Population generalizability): Knowlton and Squire's (1996) finding will be replicated. I expected chunk-strength to negatively influence accurate judgments of ungrammatical but not grammatical AG transfer test strings.

Hypothesis 6. Effects of chunk-strength on AG learning will be unrelated to individual differences in explicit memory and grammatical sensitivity. Knowlton and Squire (1996) found a similar influence of chunk-strength on AG learning by amnesics and a control group with intact explicit memory abilities. Therefore, I did not expect measures of working memory or rote memory to correlate significantly and positively with transfer test judgments of AG items with high chunk-strength. Effects of chunk-strength on semantics-free implicit AG learning do not strongly implicate higher levels of explicit memory encoding necessary for deep semantic, conceptually driven processing. Rather, these effects are attributable to shallow data-driven processing and lower-level associative memory processes (Craik & Tulving, 1975; Graf & Ryan, 1990; Robinson, 1995). Similarly, I expected grammatical sensitivity to the rules of natural languages to be nonsignificantly correlated with learning rules of semantics-free AGs. The MLAT and LABJ tests of grammatical sensitivity do not measure simple pattern recognition but, rather, the ability to identify and discriminate between the grammatical roles (e.g., subject) and their relationship to semantic roles (e.g., agent), which jointly determine whether word orders in presented sentences are similar or dissimilar.

Hypothesis 7 (Content generalizability): The same effect of chunk-strength will be found in AG and incidental Samoan learning. In line with claims by N. Ellis (2001) about the psycholinguistic plausibility of chunks as the units of language on which processes in working memory operate and the subsequent contribution of their frequency in the input to SLA, I expected chunk-strength to influence incidental natural L2 learning.⁵

Hypothesis 8: Effects of chunk-strength on incidental natural L2 learning will be related to individual differences in explicit memory and grammatical sensitivity. I further expected that the effect of chunk-strength would be strongest for those with high explicit rote- and working-memory abilities. Additionally, I expected that measures of these explicit memory abilities would correlate significantly and positively with correct acceptance of high-chunk grammatical incidental learning transfer set sentences. Finally, I expected individual differences in grammatical sensitivity to correlate significantly and positively with correct judgments of all sentences (grammatical or ungrammatical; high in chunk-strength or not) during the incidental learning transfer test.

Learning Task Sequences.

Participants in the experimental group completed three experimental tasks. The first two tasks, replicating the tasks presented to participants in Reber et al. (1991), included an explicit and an implicit learning task, described in the following subsections. Order of presentation of these tasks, which took around 40 minutes to complete, was counterbalanced as in Reber et al. Following this, the third incidental learning task began on a separate day. This task involved an initial vocabulary learning phase, followed by two separate training sessions on separate days, followed, finally, by a series of three posttests (immediate, 1-week delayed, and 6-month delayed). The control group performed only the vocabulary learning phase and the immediate posttest of incidental Samoan learning. All of the training tasks and transfer tasks as well as the explanation of task procedures were administered via Macintosh computers. In each case, the researcher was on hand to clarify any questions participants had about these procedures.

Explicit Learning Task.

This was the forced choice, series-solution problem task used by Reber et al. (1991, p. 891). As in Reber et al., 12 series-solution problems were randomly selected from a set of 21 and presented in a fixed random order (see Appendix A). Participants were told they would see a series of letters and that they should choose the letter that best completed the series. Two problem types were randomly interspersed to make the task more demanding (although participants were not made aware of this during the orientation phase). Alphabetic problems (e.g., ABCBCDCDE_; D* or C) depend on repeating alphabetic sequences, ABC, BCD, CDE, in which the last two letters of one triplet are repeated as the first two letters of the following triplet. Nonalphabetic problems (e.g., CDEADCA_; E* or D) depend on mirroring or reversing pairs of letters in the sequence, CD, EA, DC, AE. Six questions of each type were included and were presented on a computer screen. Participants responded by keyboard letter-stroke to complete the series with one of the two choices presented. As in Reber et al., no feedback was given, and following answers, the next problem appeared. As in Reber et al., percent-correct scores for this task were calculated.⁷

Implicit Learning Task.

Training. As in Reber et al. (1991), participants were told that they were taking part in a memory task. During training, participants viewed strings of letters generated by a miniature artificial “Reber” grammar. These were presented for 3 seconds each (in Times font, size 14). After viewing each string, participants were instructed to write it down on a piece of paper in response to the prompt “Now write down the letter sequence you saw.” After writing down each sequence, they then pressed any key in response to the “Press any key to continue” prompt, and a new sequence appeared for 3 seconds. The paper sheets were removed immediately following each of the two training trials, giving no opportunity for analytic study. In each training trial, participants saw 26 strings (e.g., XXVT, VXJJJJ, XXVJ, XXXVTV) generated by the rules of the AG. Reber et al. did not report their transfer and training set stimuli strings.⁸

Transfer and chunk-strength. As in Reber et al. (1991) and Knowlton and Squire (1996), participants were told after training that the strings they had viewed followed a set of rules and that they would then see some more strings. They were asked to decide if the strings followed the same rules. If so, participants were instructed to press C for “correct,” and if not, to press M for “mistake.” This was not a speeded task. There were 32 transfer set items (see Appendix B). As in Reber et al. (1991) and Knowlton and Squire (1996), percent-correct scores for this task were calculated.

Transfer set items were of four types: grammatical high and low chunk-strength and ungrammatical high and low chunk-strength. Chunk-strength of these items is reported in Appendix B. Chunk-strength was calculated by Knowlton and Squire (1994, 1996) on the basis of whether a transfer set string contained bigrams (e.g., XV) or trigrams (e.g., XVT) that had appeared in the input during training and the frequency with which those bigrams and trigrams had appeared in the two training trials. “The associative strength of a chunk was defined as the frequency with which the chunk occurred in the 16 items of the training set” (Knowlton & Squire, 1994, p. 85). So, if a string XVTX occurred in the transfer set, and only the above bigram XV and trigram XVT chunks had occurred in training, with frequencies of five and two times respectively, then the chunk-strength for that transfer set item was 7. “The mean associative strength of the chunks in each test item was then calculated by averaging the obtained frequency scores across all of the chunks in the item” (Knowlton & Squire, 1994, p. 85). Therefore, the mean associative chunk-strength of the item XVTX, with two chunks, occurring a combined total of seven times during training, is 3.5. The same procedure was followed in determining the chunk-strength of the incidental transfer set items.

Incidental Learning Task.

Training. The 10 training trials were spread over a number of days. The first phase involved learning new vocabulary and was followed by two training sessions (each session contained five trials) spread from 2 to 5 days apart, depending on participant availability. An immediate posttest followed the second training session.

Vocabulary learning. Participants were asked to memorize the meanings of 27 new Samoan words: 1 article, 15 nouns, and 11 verbs. Samoan was a language unknown to the participants. English and Japanese translation equivalents, written in English and Japanese, were provided on a study sheet—for example, taalo “play, asobu”; faitau “read, yomu”; tusi “book, hon” (see Appendix C). Participants were asked to study them until they could recall all translation equivalents on a test.

Processing for meaning. After successfully completing the word list memorization test, participants completed 10 training trials. During each trial, participants viewed 45 Samoan sentences of three types (450 presentations total, with 150 tokens of each type). These types corresponded to three different rules of Samoan (see Appendix D). The first rule, which marks the ergative, inserts a particle e before the subject of a transitive verb, as in (1).

The second rule marks adverbial phrases with a locative marker, i, as in (2).

The third rule allows direct object nouns to incorporate directly into verbs, as in (3).⁹

Sentences were presented without the English glosses given here. Training set sentences appeared for 10 seconds on the computer screen. During the orientation phase, participants were asked to try as hard as possible to understand the meaning of each sentence. No grammar explanation was provided. During training, to test comprehension following the 10-second presentation of each stimulus, participants were asked to respond to yes/no comprehension questions. For example, on screen 1, they saw taalo le tama i le paka for 10 seconds. On screen 2, they saw Does the boy swim in the sea? for which there was no time limit for responding. Finally, on screen 3, they saw Correct/incorrect. Press any key to continue. Again, there was no time limit for responding. Participants responded by pressing C for “yes” or M for “no,” whereupon feedback regarding the correctness or incorrectness of the response appeared, followed by the next stimulus sentence. The mean error rate on the first trial was around 15%; on the last trial round, the mean error rate was around 1%. Each trial took around 15 minutes. Total time spent in training, including vocabulary learning, was around 3 hours 15 minutes.

Transfer: Immediate Posttest.

Computerized grammaticality judgment test and chunk-strength. After completing the last training trial, participants were informed that they would then see more sentences on a computer screen and that they must decide if the new sentences followed the same rules as those seen during training. The transfer set consisted of 27 items with the following distribution: 9 old items viewed in training (3 for each rule described previously), 9 new grammatical items (3 for each rule), and 9 new ungrammatical items (3 for each rule). Participants responded by pressing C for “correct” (grammatical, same rules) or M for “mistake” (ungrammatical, different rules). Each item remained on screen until a response was made, whereupon the next item appeared. Response time and accuracy of response were calculated. This transfer test was unspeeded, as in the implicit conditions of Reber et al. (1991) and Knowlton and Squire (1996).

The chunk-strength was calculated for the 18 new items in the transfer grammaticality judgment test following the same procedure described by Knowlton and Squire (1994, 1996) for calculating mean associative chunk-strength of AG transfer set items. For example, the new grammatical transfer set item taalo le tama i le fale contained 4 bigrams (taalo le, le tama, tama i, and i le) that had occurred with the frequencies of 7, 36, 8, and 11 times in the input over the 10 training trials, respectively. Additionally, the item contained three trigrams (taalo le tama, le tama i, and tama i le) that had occurred with the frequencies of 7, 8, and 8 times in training, respectively, giving this specific item 7 chunks that had occurred with a combined frequency of 85 times in the input. Averaging the frequency scores (85) across the number of chunks in the item (7), as Knowlton and Squire (1994, 1996) did for their AG transfer items, gives it a mean associative chunk-strength of 12.1.

Low-chunk-strength items contained a total of 22 bigrams and 10 trigrams that had appeared in the training set (32 chunks). High-chunk-strength items contained 23 bigrams and 11 trigrams (34 chunks; see Table 1). Although the low- and high-chunk-strength transfer items were therefore almost exactly matched in terms of the number of previously occurring chunks they contained (i.e., their similarity to the training set), they differed with respect to the frequencies with which the chunks they contained had occurred in training (determining the associative strength of the chunks). The mean frequency of occurrence of these bigrams and trigrams in training was 62.5 times for high-chunk-strength items and 22.6 times for low-chunk-strength items. The mean associative chunk-strength for the high chunk-strength ungrammatical transfer set items, calculated following the procedure described previously, was 24.8; for the low-chunk-strength ungrammatical items, it was 5.6; for the high-chunk-strength grammatical items, it was 24; and for the low-chunk-strength grammatical items, it was 6.8 (see Table 1 and Appendix D).¹⁰

Number of chunks contained in new items, both high and low in chunk-strength, in the incidental learning transfer grammaticality judgment test

Listening grammaticality judgment test. After the grammaticality judgment test, participants completed a listening grammaticality judgment test. For this test, the same 27 sentences that had been presented during the first grammaticality judgment test were presented aurally in a different fixed random order. Participants wrote C or M on a piece of paper after hearing each item. Sentences were taped and spoken at normal speed. A 5-second pause occurred between each sentence. As such, this was a speeded test because participants had less than 5 seconds to respond in each case.

Guided sentence production. Following the grammaticality judgment tests, participants were given sets of words, written on separate pieces of paper and stored together in four separate packets. Participants were asked to empty the words from each packet and arrange them to make a sentence. The packets of words to complete the first three sentences (one of each ergative, locative, and incorporated rule type) were presented separately. For example, for the first ergative sentence, the words fau “build”, fale “house,” e (ergative marker), le “the,” tama “boy,” and le “the” were emptied in a jumbled order from a packet. If participants correctly ordered them to make the sequence fau e le tama le fale, they were given one point. Following this, they replaced the words in the first packet and then emptied and arranged the words in the second and third packets. Following the first three sentences, words for the second group of three sentences (locative, ergative, and incorporated) were emptied simultaneously in jumbled order from the fourth packet, making the second part of the production task much more demanding. On the basis of these responses, participants were coded as 0, 1, or 2 for production of each rule and from 0 to 6 overall for production.

Transfer: 1-Week and 6-Month Delayed Posttests.

Between 5 and 7 days after the immediate posttest, participants completed an unexpected delayed posttest. They were given time to look again at the word lists. When they demonstrated that they still remembered the meaning of these words by making 10 successive correct responses to requests for Samoan translation equivalents, they again completed the grammaticality judgment, listening grammaticality judgment, and guided sentence production tests. After 5 to 6 months, participants completed another unexpected posttest that consisted of only the grammaticality judgment and sentence production ability tests. Of the 54 original participants, only 26 returned to complete the 6-month delayed posttest.

Individual Difference Measures.

Intelligence test. As in Reber et al. (1991), a short form of the WAIS-R was used (Japanese version; Shinagawa, Kobayashi, Fujita, & Mayekawa, 1990). The full form has an extremely high reported reliability, r = .97 (Kline, 2000, p. 451). Reber et al. used a short form that consisted of picture arrangement, block design, vocabulary, and arithmetic. This short form correlates highly with the full test, with a validity and reliability coefficient of r = .95 to .96 (see Cyr & Brooker, 1984, p. 905; Reber et al., p. 892). Because some participants had already been exposed to the picture arrangement subtest during participation in an unrelated research project, it was decided to drop scores for this subtest and use a triadic WAIS-R short form (vocabulary, arithmetic, and block design), which Cyr and Brooker (p. 905) also reported as having a high validity and reliability coefficient of r = .94. As in Reber et al., I used the conversion tables in Brooker and Cyr (1986; see also Cyr & Brooker) to convert short-form scores to scaled IQ scores. Using this triadic short form, participants had a mean IQ of 113, which is almost identical to Reber et al.'s participants' mean of 110 measured using the tetradic short form (vocabulary, arithmetic, block design, and picture arrangement). However, the standard deviation was much narrower (SD = 8.2, kurtosis = 1.7, skewness = −.8) than Reber et al.'s (SD = 21.2), as was the range of scores, from 88 to 132 compared to 73 to 151 in Reber et al.'s study.

Aptitude test. The LABJ consists of three subtests: rote memory for 24 paired-associates, grammatical sensitivity to grammatical patterns in a new language, and phonemic sensitivity (the ability to match sounds and symbols). This took about 30 minutes to administer. The mean score for the 60-item test was 46.6 (SD = 8.2, kurtosis = .48, skewness = −.69). Reliability for this test was .85 (KR21).

Working-memory test. The reading-span test of working memory took about 40 minutes to administer. Participants were presented with sets of Japanese sentences written on cards and were asked to read the sentences aloud. After reading them aloud, participants were asked to recall individual words underlined in each sentence. The sets increased from two to five sentences throughout the test. I scored this using total number of correctly recalled words. The mean score on this 70-item test was 44.8 (SD = 9.3, kurtosis = −.74, skewness = −.1). Reliability for this test was .83 (KR21). All individual difference measures were administered following the 1-week delayed incidental learning posttest, at times convenient to participants.