Unconscious learning of complex information

Unconscious Learning of Complex Information

Martina L. Van Slooten, Zoltan Dienes & Eric-Jan Wagenmakers

University of Amsterdam University of Sussex

Abstract

The purpose of this study was to evaluate the paper of Dulany, Carlson and Dewey (1984), in which they argue that arti cial grammar can only be learned consciously. In order to show that arti cial grammar could be learned unconsciously as well as consciously, we deconstructed their argumentation using a task with subjective measures and using materials for which chunks were not the relevant structure to be learned. We found that subject could have a task performance above change level, while indicating they were guessing or using intuition. The inconsistency between their performance and their knowledge, indicates they learned unconsciously. Secondly we found that, contrary to Dulany et al. believes, subjects did not learn the chucks of information but were in uenced by repetitions structures. Additionally, we showed that the use mean rule validity can give a biased e ect of the unconscious learning of grammar. Finally we found a drop in performance for unconscious learning, when subject have divide their attention between learning the grammar and a secondary task. This can imply that learning of grammar relies on working memory even when knowledge is unconscious.

Keywords: Implicit learning; Arti cial Grammar; Repetition; Bigram.

Unconscious Learning

Acquiring a new language can be quite di cult. It takes a long period of time to fully master a new grammar and it can be easily unlearned when not practised enough. On the contrary, when it comes to your own languages everything seems to come intuitively and you do not have to think about the rules to understand that an expression is correct or incorrect. You can probably even correct a non-native speaker, without being able to tell which rules they used wrong. This intuitively knowledge of grammar can already been found in children of a young age. Children know how to use their native language in everyday life, without being learned the correct grammar. The acquiring of this intuitive knowledge has been a topic of interest for a long period of time and can not only be found in languages,

also motor skills and social interactions seems to be learned intuitively (Berry & Dienes, 1993; Chartrand & Bargh, 1999; Norman & Price, 2012; Opacic, Stevens, & Tillmann, 2009; Reed, McLeod, & Dienes, 2010).

In 1967 Reber rst introduces the term 'implicit learning' to describe the acquiring of intuitive knowledge. He argued that implicit learning is a process that consist out of two critical features: (a) it is an unconscious process and (b) it yields abstract knowledge. That is, people are able to learn abstract structures, without being consciously aware that they have learned this structure (Reber, 1967, 1989). For example, a native speaker can have abstract knowledge about how to formulate correct sentences, but is unconscious of the exact rules he used.

Because the learning of natural languages can be di cult to study in a lab setting, Reber (1967) introduced the arti cial grammar paradigm. With this paradigm subjects do not learn a natural languages but instead learn a nite state grammar, which is simpli cation of a natural grammar. To make a nite state grammar, a nite number of states (e.g. circles) are drawn, these states are connected with each other by arrows that can be drawn to any state in any direction (Figure 1). Each arrow represents a letter, by following the arrows through the diagram a string of letters is made. Because di erent paths can be followed, di erent letter strings of the same nite state grammar can be made (Dienes & Seth, in press). In the arti cial grammar paradigm subject are asked to memorize and learn several of the letter strings, without being told that there is an underlying grammar. Only after learning the letter strings, the subjects are told that the strings have rules. Next, the subjects are shown new letter strings and are asked to classify the letter strings on whether they are following the rules or not. Reber found that, depending on the nite state grammar, the subjects were able to classify 60-70 percent of the letter strings correct. This while in free report the subjects state they were not consciously aware of any rules. He argued that the inconsistency between the percentage correct and not consciously knowing the rules, indicates that subjects learned unconsciously.

Although many of the experiments that followed showed similar results (Reber & Lewis, 1977; Reber & Allen, 1978; Allen & Reber, 1980; Lewicki, Czyzewska, & Ho man, 1987; Dienes, Broadbent, & Berry, 1991), free report is not the best method to indicate the subject's true knowledge. Free report causes a problem because subjects do not always tell what they actually know. For example, when a subject is not con dent or certain enough to state their true knowledge, the subject can choose to say that he did not knew the grammar at all. Because most of the early studies are based on free report, sceptics of implicit learning argue that implicit learning is not an unconscious process. They argue that it is merely the learning of simpli ed rules that tend to give the right answer more often than not. Subjects only state that they do not know the grammar, because they are not certain enough about the simpli ed rules they used (Dulany, Carlson, & Dewey, 1984; Shanks, Lamberts, & Goldstone, 2005; Perruchet & Pacteau, 1990). Dulany et al. (1984) found support this theory by using an objective measurement instead of free report. In an experiment similar to Reber, subjects were asked to memorize letter strings and classify them on whether they were following the rules or not. Instead of asking the subject what the rules were in free report, the subject had to indicate what made the letter string correct or incorrect. That is, the subject had to underline the parts that made a letter string correct and cross out the parts that made a letter sting incorrect. Dulany found that, although

the subjects did not know the precise underlying grammar, the subjects came up with similar simpli ed rules that could predict the proportion correct. He concluded that the subjects were consciously making up simpli ed rules to better remember the letter string and therefore they did not learn the letter strings unconsciously.

Figure 1. Example of an Arti cial Grammar diagram.

Dulany showed with his experiment that subjects were able to learn certain parts of a letter string. In other words, he showed that subjects were able to learn that a \X" can follow an \V" or that a \M" can start a letter string. This knowledge can also be referred to as structural knowledge (Dienes & Scott, 2005). The fact that subjects acquired structural knowledge, does not have to imply that the subject learned consciously. It might be the case that unconscious knowledge guided the subjects in the underlining of the letter strings. What for the subjects could feel like randomly marking parts of a letter sting to ful l the requirements of the researcher, could be in fact be the applying of unconscious knowledge. When the subjects are not asked how they acquired the rules, it is never certain whether the subject really learned the rules in the training phase or just guessed correctly. This second form of knowledge can also be called judgement knowledge and it refers to a subject having higher order thoughts about their learning process or not (Dienes & Scott, 2005). Because the di erence between structural knowledge and judgement knowledge has to be made, we took another look at the experiment of Dulany et al. (1984) in an attempt to nd out whether subject can learn grammar unconsciously.

This study will try to capture the subjects' judgment knowledge, by asking the subjects whether a letter strings was underlined based on a guess, their intuition, memory or a rule (Dienes & Scott, 2005). When a subject indicates to be guessing, their judgement had no basis at all, they literally classi ed the letter string at random. Meaning, the subject did not have conscious structural knowledge or had conscious judgement knowledge. When a subject used intuition, the subject had some con dence about the judgement, because their attention was drawn to a part of the item that felt especially right or wrong. In this case the subject did have conscious judgment knowledge, because he knew the letter string was right or wrong, but did not have any conscious structural knowledge about what made

the letter string right or wrong. When a subject uses memory or rule, the subject either had a recollection from the training phase or used a rule learned during the training phase to classify the letter string right. Both the rule and memory attribution indicate that the subject had conscious judgement, because he knew the letter is was right or wrong, and conscious structural knowledge, because he could indicate what made them right or wrong. By asking the subject who certain they are about the rules they underlined, a di erence between unconscious and conscious structural knowledge but also between unconscious and conscious judgment knowledge can be made.

To further support the hypotheses that subjects are able to learn arti cial grammar unconsciously, the original arti cial grammar paradigm will be manipulated by adding an extra condition. In the original paradigm the subjects were not aware of the underlying rules until the test phase. In the manipulated task the subjects are immediately told that the letter strings have a set of rules and by guring out these rules they will be better in memorizing the letter strings (Dienes & Scott, 2005). By asking the subjects to actively search for the rules, the subjects should acquire more conscious structural knowledge than the subject who were not aware of the rules. Because there is more conscious structural knowledge, it requires higher load of the working memory than just memorizing the letter strings (Roberts & MacLeod, 1995; Dienes & Altmann, 1997; Frensch, Wenke, & R•unger, 1999; Waldron & Ashby, 2001). Dienes & Scott (2005) found that the classi cation performance only dropped when subject were asked to search for rules and had to divide their attention. This, while their performance stayed the same when the subjects could fully concentrate on the grammar or when their knowledge of the letter strings was unconscious. This indicates that only asking to search for rules should not a ect the subjects performances on the task, because one task this is not too demanding. When giving the subjects an additional secondary task, the conscious performance for the subjects that are actively searching for rules should become worse, while the unconscious performance should stay unchanged.

Another problem with the objective measurement of Dulany's is that subjects are told to underline a part of a letter string. Underlining parts of the letter string indicates that letter strings can only be remembered when learning speci c features of that letter string, for example that \VX" is allowed in a letter sting but \XV" is not. But contrary, various articles show that subjects are also sensitive for learning patterns, repetition or even symmetry (Lai & Poletiek, 2011; Jiang et al., 2012; Rohrmeier & Rebuschat, 2012). By asking the subject to underline parts of a letter string, it rules out the possibility that a subject learned a rule through pattern, repetition or symmetry, because underlining these features is not possible. For example, when the rule of a letter string is 'use all the ve letters once' (e.g., the pattern 12345) the letter strings \XTMVR" and \XTRMV" are both correct. By forcing the subject to underline a part in the letter sting, it is uncertain whether the subject thought that part made the string grammatically correct or that he was just unable to indicate the correct pattern rule because he had to underline a feature. Because there is the possibility that letter strings can be learned by patterns or repetition this experiment did not only use the traditional nite state grammar, but half of the letter strings did have a repetition structure as well.

used the same objective measure that Dulany et al. (1984) used, but added the judgement attributions guess, intuition, memory and rule of Dienes and Scott (2005) paper. By adding the judgement attributions, it becomes certain whether a subject had higher order thoughts when classifying the items. Additionally, the subjects will either get an arti cial grammar task that is highly demanding or not, depending on whether they got a secondary task to perform. The additional task will show that performance for conscious learning will drop in the high demanding task but will remain the same unconsciously learning. Lastly, the arti cial grammar paradigm will have bigram and repetition grammar, to show that grammar is not only learned in chuncks.

Methods

This study used the two-grammar design of Dienes & Altmann (1997). Subjects were either trained on A grammar or B grammar, but all subjects were tested on the same items. The test items consisted out of an equal mixture of A and B grammar items. For subjects trained on the A grammar, the A grammar test items had to be classi ed correctly and the B grammar test items incorrectly. This holds the same for subjects trained on the B grammar, only the other way around.

Next, the subjects were either placed in the rule search condition where they were instructed to actively search for rules during the training phase, or in the memory condition where the rules were not mentioned and the subjects were just asked to memorize the letter strings during the training phase. Lastly, half the subjects in both conditions were asked to perform a secondary task making the task more demanding.

Participants

For this experiment we intended to stop recruiting participants when nding a Bayes factor of 4 or 1/4 for subjects scoring above chance level for the guess and intuition attributions in both the bigram and repetition grammar. Due to a lack of time this stopping rule was not met. Instead 49 volunteers of the University of Sussex and the University of Amsterdam between the age of 18 and 40 were randomly assigned over the di erent conditions.

Materials

Both in the A and the B grammar the letters M, T, V, R and X were used to make ve letter long strings. In the training phase 20 letter strings were formulated for each grammar. The grammar followed either a repetition structure, for example the rst letter is repeated at end, or a bigram structure, for example \RX" is allowed. In the training phase there was a mixture of items that was allowed in both the A and B grammar and items speci c to the A or B grammar (Table 1).

For the test phase 40 new strings were created. 20 of the letter strings were grammatical for A and therefore ungrammatical for group B. The other 20 letter strings

were grammatical for B and therefore ungrammatical for group A. Ungrammatical letter strings could either violate the rules by using the wrong repetitions structure, or by using the wrong bigram structure (Appendix A). To nd the best grammar for this experiment ve pilot studies were run in order to get the right balance between a di cult and easy task.

E-prime 2 was used to present the letter strings. Each string was presented for 10 sec on a separate frame, displayed centrally in black on a white background. When a subject was in the condition with a secondary task, the TempoPerfect metronome played the subjects 45 beats per minute through a headphone.

Table 1: Rules for the di erent grammars.

Grammar String A repetition 12134 B repetition 12341 Shared repetition 12324 A bigrams RX, RT, VR, XM B bigrams VX, VM, MT, XT Shared bigrams XR, TR, RV, XV, MV, TM, MX, TX, MR, RM, TV, VT Procedure

All subject started the experiment with the training phase. Subject in the memory condition were told they participated in a simple memory experiment and that they had to remember as much as possible of all of the items. Subjects in the rule search condition were told they participated in a problem solving experiment and that the order of the letters was determined by a set of rules. Figuring out the rules would help them when tested on their knowledge later. After the instructions, the training phase began and the subjects were presented each letter strings separately, for 10 seconds. This was repeated three times. When a subject was in the condition with a secondary task, the subject was asked to verbally generate random number between 0 and 9 on the beat of a metronome, while either memorizing or guring out the rules of the letter strings. After the training phase all subjects were told that the order of the letter was determined by complex a set of rules. The new letter strings (e.g. the test items) either follow those complex rules or the violate the rules. For every test item, they had to determine whether the letter string followed or violated the rule. Following, what part of the item made the letter string right or wrong and last they had to indicate the basis of their judgment with the guess, intuition, rule or memory attributions.

Prior

The e ects of our study will be reported using Bayes factors (B). It is broadly accepted that a B of 4 or higher indicates substantial evidence for the alternative hypothesis, while a B of 1/4 or smaller is substantial evidence for the null hypothesis rather than the alternative

hypothesis. Everything in between 4 and 1/4 will be seen as inconclusive evidence for either hypothesis, meaning that the data is insensitive for distinguishing the alternative and null hypothesis (Dienes, 2014; Lee & Wagenmakers, 2014). We expected an e ect of 5% for unconscious learning for both the bigram and repetition structure, because this is slightly less than the e ect Dienes & Scott (2005) found in their paper. Additionally, it is expected that the subjects will learn from the training phase. In other words, is expected the performance will be larger than chance level. Therefore a half-normal prior with a SD of 5% above baseline will be used.

Results

Table 2 shows the proportion of di erent attribution subjects used for the bigram and the repetition structure. For both the bigram and repetition structure holds that, most of the items were marked as guess or intuition (61%). In other words the subjects were most of the time unconscious of the grammar. A 2X2X2 (attention [full vs divided] by training [memory vs rule search] by structure [bigram vs repetition) mixed model ANOVA on the proportion unconscious attributions shows that there is no interaction e ect B = .17, indicating that the di erent conditions did not a ect the number of times a subject marked their knowledge as unconscious.

Table 2: Proportion of di erent attribution for the bigram and the repetition structure.

Structure Bigram Repetition Guess 0.12 (0.125) 0.12 (0.115) Intuition 0.19 (0.127) 0.19 (0.117) Memory 0.08 (0.097) 0.08 (0.089) Rule 0.12 (0.147) 0.11 (0.144)

Grammar and classi cation accuracy

Looking at the classi cation performance for the two structures we found that the subjects were only able to correctly classify the items with a repetition structure (M = .62, SD =.48) and not the items with the bigram structure (M = .50, SD = .50). Apparently the manipulation did not work for the items with a bigram structure and will therefore not be analysed further. A 2X2X2 (training [memory vs rule search] by attention [full vs divided] by attribution [conscious vs unconscious]) mixed model ANOVA on the proportion of correct classi cations gave evidence for training interaction on the unconscious and conscious attributions B >100. When analysing the three way interaction into two separate partial interaction for the unconscious and conscious attributions, there was evidence for a partial interaction for the unconscious attributions, B = 55.31, but there was no evidence for the conscious attributions B = 1.04. When we look Figure 2, the e ect is di erent to the e ect we expected. For subjects that had to divide their attention the performance

was worse for the subjects in the memory condition than in the rule search condition. This while the performance of the subjects who had the possibility to fully concentrate on the grammar, stayed the same in the memory and rule search condition.

Figure 2. Proportion correct classi cation for training on attention for the conscious and unconscious attributions.

Grammar and marked rules

In order to see whether the marked rules could predict the proportion correct, the mean validity of rules was calculated (Dulany et al., 1984). The mean validity of rules is the probability that the underlined rule correctly categorizes a string, given the presence of the speci c rule it should represents. In Figure 3 we plotted the mean validity of rules against the proportion correct of the unconscious and conscious attributions for the bigram and repetition structure. Similar to the results of Dulany a steep regression is found for each case, indicating that the underlined rules predicted the proportion correct.

To see whether the marked rules actually represented the grammar, the proportion of times that an ungrammatical string contained the position of violation was compared to the proportion of time this same position was used by the subject trained on the other grammar. We found that subjects included the position of violation less often (M = 5.75, SD = 2.8), compared to the subjects trained on the other grammar (M = 7.85, SD = 2.4, B = 52 ). Therefore it seems that the subjects were repelled away from the position of violation. Although this shows that the subjects were sensitive to the position of violation, the applied rules were not completely rational, because otherwise the position of violation would marked as incorrect.

Although it seems that the subjects were repelled away from position of violation, it could also be the case that they were more drawn to an other parts of the string. In order to see whether subjects marked certain parts of the letter string more frequently than others, the associative chunk strength was measured. The associative chunk strength is the frequency in which bigrams that appeared in the training phase, were marked as grammatical in the test phase. We control for the associative chunk strength by applying the same rule on each of the other grammatical test items. We found that the associative chunk strength is twice as high as the control (M = .66, SD = .57, M = .33, SD = .24,

B > 100), indicating that subjects were attracted to high frequency bigrams. Although, subject were attracted to the high frequency bigrams in grammatical correct they were not repelled away from those bigrams when the letter string was grammatically incorrect (M =

.56, SD= .50, M = .55, SD = .43, B = .16).

Finally we looked at how many times a rule applied on di erent items. We found that on average every marked rule applies to 1.9 items. In other words most rules only apply to one other string.

Figure 3. Percentage correct classi cation and mean rule validity for (A) unconscious attributions with a repetition structure, (B) conscious attributions with a repetition structure, (C) unconscious attributions with bigram structure (D) conscious attributions with bigram structure.

Discussion

Previous research on the learning of arti cial grammar argued that subject were able to learn parts of the letter string consciously (Dulany et al., 1984) . The study showed that when subjects were asked to mark the rules they used to classify the grammar items, these marked rules could predict the their classi cation accuracy. Although the Dulanys results showed that the marked rules could predict the percentage correct, it was not clear whether the marking of the rules was consciously correct or guided by unconscious knowledge. This

study added a subjective measurement on which the subjects could indicate whether the marking of the rules was based on a guess, intuition, memory or a rule. When subjects would indicate they were guessing or using intuition , but classi ed the items correctly, it would indicate that the subjects learned unconsciously. The results showed that in 61% percent of the times, subjects indicated that they guessed or used intuition, indicating they were mostly unaware of the rules. This said, the subjects were able to correctly classify the items above chance level. The inconsistency between the subjects their performance and the subjective measure indicate that the subjects learned the grammar unconsciously.

To support this nding we looked at whether a secondary task would in uence the classi cation performance. Dienes and Scott (2005) found that when a task became very demanding due to two intensive cognitive tasks, the subjects conscious learning became worse, but their unconscious learning was not a ected. Surprisingly, in this study we found a drop in performance for the unconscious learning. Subjects who indicated that they were not aware of the rules, performed worse when they had to divide their attention between two task and were not informed about the underlying grammar. This, while the performance stayed the same when subjects indicated they were aware of the grammar. This nding is a reversed result of that of Dienes and Scott (2005). Dienes and Scott argued that the intensive tasks only in uence the conscious learning and not with unconscious learning because the unconscious learning is processed outside the working memory. This study found that the secondary task could also in uenced the unconscious learning. In this experiment subjects were only able to learn grammar with a repetition structure. Possibly, the learning of a repetition structure is always processed in the working memory even the when the learning is unconscious. When the knowledge of the repetition structure is learned consciously, one can prioritize what gets kept in the working memory even when space is limited. When the knowledge is unconscious there is no such control so relevant information is lost, resulting no learning of the repetition structure, unless the subject speci ed that he used a memory or a rule.

We also found that the bigram structure, broadly used in the arti cial grammar paradigm, had no e ect compared to the repetition structure. For the repetition structure the classi cation performance was above chance level, while the classi cation performance for the bigram structure was around chance level. The di erence between the repetition and bigram structure was probably due the grammar used in this experiment. In our grammar almost all bigrams that could be generated from the ve letter, but four, were correct. Therefore, the repetition structure was a better help when classifying the items. Interestingly the di erence in classi cation performance between the repetition and bigram structure was found for both the conscious and the unconscious attributions. Indicating that subject who were unaware of the grammar unconsciously choose to adapt the more simpler rules. Similar results were found in a experiment of Wan, Dienes, & Fu (2008). In this experiment subjects were shown two separate grammars, but were asked to focus on just one. They found that even when the subjects were not aware of the grammar they were able to pick the most strategically one. Indicating that, whether judgement knowledge or structural knowledge is conscious or not, subjects can strategically choose the most e ective grammar.

Finally we looked at rules the subjects underlined or crossed out in the test phase. In line with the results of Dulany et al. (1984) we found a strong relation between the

mean rule validity and the percentage correct. This might give the impression that the marked rules could predict the subject classi cation performance and were therefore learned consciously. Hence, this result was also found when subjects indicated they were guessing or using intuitions. In other words the subjects indicated they were not aware of the rules, making conscious learning in the traditional way less probable. Additionally, we found that although subject were attracted to the high frequency bigrams from the training phase during the classi cation of grammatically correct letter strings, they were not repelled away from those same bigrams when the letter string was grammatically incorrect. The learning op the high frequency bigrams is therefore less conscious than what we would expect when a subject knew what rule made a letter string correct or incorrect. Altogether, the high relationship between the mean rule validity and the classi cation performance does not re ect our other ndings, making it a susceptible method in describing the data. And perhaps is its. The mean rule validity is the probability that a underlined rule correctly categorizes a string. We found that on average every marked rule applies to 1.9 items. In other words most rules only apply to one other string. Therefore a perfect correlation between the classi cation performance and the mean rule validity occurs, adding just a little noise around the correlation.

Conclusion

This paper deconstructed the argument of Dulany et al. (1984) that arti cial grammar can only be learned consciously. We showed that subjects could have the feeling they were guessing or using intuition, while having a classi cation performance above change level. The inconsistency between their subjective measure and their performance, showed that subjects were able to learn the grammar unconsciously. Additionally we found that subjects who were unaware of the underlying repetition structure of the grammar, performed worse when they had to devide their attention between two task and were not made aware of the grammar by forehand. This result could indicate that the learning of a repetition structure is always processed in the working memory even the when the learning is unconscious. Finally, we showed that the use mean rule validity can give a biased e ect in the learning of arti cial grammar.

References

Allen, R., & Reber, A. S. (1980). Very long term memory for tacit knowledge. Cognition, 8 (2), 175{185.

Berry, D., & Dienes, Z. P. (1993). Implicit learning: Theoretical and empirical issues. Psychology Press.

Chartrand, T. L., & Bargh, J. A. (1999). The chameleon e ect: The perception{behavior link and social interaction. Journal of personality and social psychology, 76 (6), 893.

Dienes, Z. (2014). Using bayes to get the most out of non-signi cant results. Frontiers in psychology, 5 , 781.

Dienes, Z., & Altmann, G. (1997). Transfer of implicit knowledge across domains: How implicit and how abstract. How implicit is implicit learning, 5 , 107{123.

Dienes, Z., Broadbent, D., & Berry, D. C. (1991). Implicit and explicit knowledge bases in arti cial grammar learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 17 (5), 875.

Dienes, Z., & Scott, R. (2005). Measuring unconscious knowledge: Distinguishing structural knowledge and judgment knowledge. Psychological research, 69 (5-6), 338{351.

Dienes, Z., & Seth, A. K. (in press). Conscious versus unconscious processes. G. C. L. Davey (Ed.), Psychology. Wiley: Chichester.

Dulany, D. E., Carlson, R. A., & Dewey, G. I. (1984). A case of syntactical learning and judgment: How conscious and how abstract? Journal of Experimental Psychology: General, 113 (4), 541.

Frensch, P. A., Wenke, D., & R•unger, D. (1999). A secondary tone-counting task suppresses expression of knowledge in the serial reaction task. Journal of Experimental Psychology: Learning, Memory, and Cognition, 25 (1), 260.

Jiang, S., Zhu, L., Guo, X., Ma, W., Yang, Z., & Dienes, Z. (2012). Unconscious structural knowledge of tonal symmetry: Tang poetry rede nes limits of implicit learning. Consciousness and cognition, 21 (1), 476{486.

Lai, J., & Poletiek, F. H. (2011). The impact of adjacent-dependencies and staged-input on the learnability of center-embedded hierarchical structures. Cognition, 118 (2), 265{273.

Lee, M. D., & Wagenmakers, E.-J. (2014). Bayesian cognitive modeling: A practical course. Cambridge University Press.

Lewicki, P., Czyzewska, M., & Ho man, H. (1987). Unconscious acquisition of complex procedural knowledge. Journal of Experimental Psychology: Learning, Memory, and Cognition, 13 (4), 523.

Norman, E., & Price, M. C. (2012). Social intuition as a form of implicit learning: Sequences of body movements are learned less explicitly than letter sequences. Advances in Cognitive Psychology, 8 (2), 121{131.

Opacic, T., Stevens, C., & Tillmann, B. (2009). Unspoken knowledge: implicit learning of structured human dance movement. Journal of Experimental Psychology: Learning, Memory, and Cognition, 35 (6), 1570.

Perruchet, P., & Pacteau, C. (1990). Synthetic grammar learning: Implicit rule abstraction or explicit fragmentary knowledge? Journal of experimental psychology: General, 119 (3), 264. Reber, A. S. (1967). Implicit learning of arti cial grammars. Journal of verbal learning and verbal

behavior, 6 (6), 855{863.

Reber, A. S. (1989). Implicit learning and tacit knowledge. Journal of experimental psychology: General, 118 (3), 219.

Reber, A. S., & Allen, R. (1978). Analogic and abstraction strategies in synthetic grammar learning: A functionalist interpretation. Cognition, 6 (3), 189{221.

Reber, A. S., & Lewis, S. (1977). Implicit learning: An analysis of the form and structure of a body of tacit knowledge. Cognition, 5 (4), 333{361.

Reed, N., McLeod, P., & Dienes, Z. (2010). Implicit knowledge and motor skill: What people who know how to catch dont know. Consciousness and cognition, 19 (1), 63{76.

Roberts, P. L., & MacLeod, C. (1995). Representational consequences of two modes of learning. The Quarterly Journal of Experimental Psychology, 48 (2), 296{319.

Rohrmeier, M., & Rebuschat, P. (2012). Implicit learning and acquisition of music. Topics in cognitive science, 4 (4), 525{553.

Shanks, D. R., Lamberts, K., & Goldstone, R. (2005). Implicit learning. Handbook of cognition, 202{220.

Waldron, E. M., & Ashby, F. G. (2001). The e ects of concurrent task interference on category learning: Evidence for multiple category learning systems. Psychonomic Bulletin & Review, 8 (1), 168{176.

Wan, L., Dienes, Z., & Fu, X. (2008). Intentional control based on familiarity in arti cial grammar learning. Consciousness and Cognition, 17 (4), 1209{1218.

Appendix A

Position of Violation Repetition Structure String

NA 12134 RXRTV NA 12134 RXRTM NA 12134 RTRXM NA 12134 RTRMV NA 12134 VRVTM NA 12134 VRVTX NA 12134 XMXRV NA 12134 XMXVT NA 12134 TVTRX NA 12134 TVTXM NA 12324 MRXRT NA 12324 VRXRT NA 12324 MRTRX NA 12324 VRTRX NA 12324 VRTRM NA 12324 RXMXV NA 12324 RTRRX NA 12324 VRMRT NA 12324 VRMRX NA 12324 VRXRT

Grammar B

Position of Violation Repetition Structure String

NA 12341 VXRMV NA 12341 MTXRM NA 12341 XTRMX NA 12341 XRMVX NA 12341 MXRVM NA 12341 RMVXR NA 12341 TXRMT NA 12341 XRVMX NA 12341 XRMTX NA 12341 RMXTR NA 12324 VXTXR NA 12324 VMTMR NA 12324 VMTMX NA 12324 VMRMT NA 12324 MTXTV NA 12324 MTXTR NA 12324 XTMTV NA 12324 XTMTR NA 12324 RVXVM NA 12324 RVMVX

Test grammar A violates B repetition

Position of Violation Repetition Structure String

3 12134 TVTMR 3 12134 TVTRM 3 12134 TVTMX 3 12134 TVTXR 3 12134 MRMVT 3 12134 MRMXV 3 12134 RMRVT 3 12134 RXRVT 3 12134 RXRMV 3 12134 RTRMX

Test grammar A violates B bigrams

Position of Violation Repetition Structure String

2 12324 RTVTX 2 12324 RTVTM 3 12324 TRXRV 5 12324 TRMRX 5 12324 XRMRT 2 12324 XMRMV 3 12324 MRTRV 3 12324 MVRVT 3 12324 TXMXV 5 12324 MVTVR

Test grammar B violates A repetition

Position of Violation Repetition Structure String

5 12341 XRVTX 5 12341 TRMVT 5 12341 RVTXR 5 12341 RVTMR 5 12341 XVTMX 5 12341 MVTRM 5 12341 TMRVT 5 12341 TMXVT 5 12341 MXVTM 5 12341 TXRVT

Test grammar B violates A bigrams

Position of Violation Repetition Structure String

2 12324 VMRMX 2 12324 MTVTX 2 12324 MTVTR 2 12324 XTVTR 2 12324 XTVTM 5 12324 RVTVM 5 12324 XVTVM 5 12324 RVTVX 5 12324 MVTVX 4 12324 TXVXR