Implicit and Explicit Knowledge Decay at Different Rates: A Dissociation Between Priming and Recognition in Artificial Grammar Learning Richard J. Tunney University College London, UK Abstract. An experiment tested the common assumption that implicit and explicit knowledge are forgotten at different rates. In a training phase participants responded to sequences of letters generated by a finite-state grammar by pressing corresponding letters on a keyboard. A control group responded to randomized sequences. Participants were tested immediately following training and after intervals of seven and fourteen days. During each test participants responded to the letters of old and new sequences, and performed a concurrent recognition test. Priming was indexed by the time taken to respond to the sequences. In the immediate test both priming and recognition were substantially greater than the control group. In the delayed tests the level of priming remained unchanged but recognition had declined. The data indicate that priming and recognition decay at different rates, and are discussed with reference to both single and dual process models of memory. Key words: implicit learning, forgetting, retention, priming, recognition A key claim of many theories of cognition is that priming and recognition are dissociable, reflecting distinct implicit and explicit memory processes. In particular, explicit memory, such as recognition, is thought to be relatively fragile in comparison to implicit forms of memory, such as priming. For example, there is evidence that implicit processes are preserved in the face of amnesia or divided attention while explicit processes are certainly impaired (e.g., Gabrieli, 1998). One commonly held assumption about memory is that implicit and explicit knowledge are forgotten or decay at different rates (e.g., Berry & Dienes, 1993; Reber, 1989). Despite this, little work has directly questioned whether this is indeed the case. I thank Martina Niemela for helping collect some of the data, and David Shanks for much help and support during my time at UCL. This work was supported by the ESRC Centre for Economic Learning and Social Evolution. DOI: 10.1027//1618-3169.50.2.124 Early work using the traditional materials of implicit memory, such as word-identification or wordstem completion, produced equivocal results. Some workers have observed different patterns of forgetting on direct and indirect tests, while others have found similar patterns. For instance, Tulving, Schacter, and Stark (1982) found priming on wordfragment completion persisted over days and weeks, while recognition declined. More recently McBride and Dosher (1997) compared implicit and explicit versions of stem-completion that differed solely in whether participants were instructed to use the first word that came to mind or a word that they had seen earlier. In contrast to Tulving et al. (1982), McBride and Dosher (1997) observed similar patterns of decay on both measures. A good deal of research has examined the relationship between priming and recognition in implicit learning paradigms. A common feature of many implicit learning tasks is the acquisition of sequential
Decay of Implicit and Explicit Knowledge 125 information in which participants learn by responding to events constrained by a simple system of rules. Participants knowledge may be expressed either indirectly as priming, via reduction in response latency to predictable targets, or directly via recognition, prediction, or generation tests thought to require explicit knowledge. It seems that many of our preferences, skills, and habits are learned in this way. One such paradigm, in which knowledge is thought to be retained for long periods, is artificial grammar learning. Artificial Grammar Learning In a typical artificial grammar learning experiment participants study sequences of letters that, unknown to them, are generated by an artificial grammar (see Figure 1). After this study period, participants are informed that the training sequences were constructed using some simple rules. Participants are then presented with previously unseen sequences, some of which can be generated by the grammar and some of which cannot, and are asked to discriminate between the two on the basis of their adherence to the grammar. Reber and Allen (1978) found the percentage of correct classifications immediately following study to be 81%. Two years later, and with no intervening exposure to the study materials, the same participants classified 68% of sequences correctly (Allen & Reber, 1980). Participants were not informed during the original study that they would be recalled two years later so it seems unlikely that they would have spent any of the intervening period Figure 1. The artificial grammar used to generate stimuli. Grammatical sequences are generated by traversing the nodes of the grammar from left to right and collecting letters from each transition. Ungrammatical sequences were generated by rearranging one or two letters of grammatical sequences so that they no longer conform to the grammar. rehearsing the materials. These grammaticality decisions are frequently considered to be based on implicit knowledge of the grammar. Although there has been substantial debate about what kind of information participants actually learn by studying artificial languages, there is some consensus that much of the knowledge expressed at test is composed of fragments of the sequences studied earlier (e.g., Dulany, Carlson, & Dewey, 1984; Perruchet & Pacteau, 1990). For instance, if a participant studies a sequence such as MTVRVX, at test they might remember having seen the fragment MTV. The presence of this fragment in a previously unseen sequence shown at test makes the new sequence similar to the old one enabling it to be processed more fluently than dissimilar sequences. This fluency primes the classification of the new sequences. Dual process models of memory argue that in addition to fluency, discriminations may also be predicated on conscious recollection. One piece of evidence suggests that familiarity and recollective components of recognition might decay at different rates. Higham, Vokey, and Pritchard (2000; Experiment 2) asked participants to study sequences generated by two grammars. Immediately after the study period participants were asked to classify old and new sequences, some of which belonged to the earlier studied grammars and some of which belonged to neither. According to the logic adopted by Higham et al. (2000), the attribution of sequences to the correct grammar was due to recollective and familiarity based influences on classification, while the attribution of sequences to the wrong grammar reflected familiarity based influences alone. When Higham et al. repeated the test seven days later they found that classifications based on recollection had decreased but those based on familiarity had not, suggesting that implicit and explicit memory of the grammars had decayed at different rates. However, whether or not recollective and familiarity influences on recognition truly reflect dissociable memory processes, is a source of considerable debate (see Foster & Jelicic, 1999). Priming in artificial grammar learning supports reaction time advantages as well as discrimination. Knowledge of fragments reflects the transitional dependencies of the grammar, and as in serial reaction time learning, this priming supports a reaction time advantage for probable over improbable transitions (e.g., Cleeremans, 1993). Dissociations have been observed between discrimination-based measures that make reference to the study episode and reaction time savings that make no such reference. For instance, Seger (1998) observed that although participants were able to discriminate between grammatical and ungrammatical sequences when composed of a different set of letters, they were unable to transfer
126 Richard J. Tunney the reaction time advantage for old over new sequences that they had exhibited in the original vocabulary. This finding converges with serial reaction time learning in the suggestion that recognition and motor-based learning can be dissociated. In particular, motor-learning appears to be relatively spared by amnesia (e.g., Willingham, Nissen, & Bullemer, 1989). More importantly, at least some forms of motor-learning appear to be retained for long periods of time (Adams, 1987).The nature of the motor-responses seems to make a difference to the rate of decay. Continuous responses, such as tracking, may be retained for longer periods than discrete responses such as button presses (Adams, 1987). Deficits in explicit memory, such as amnesia or Alzheimer s disease, do not appear to impair the acquisition of sequences of motor responses. Nissen, Willingham, and Hartman (1989) examined the retention of sequence knowledge in both amnesic and age-matched controls. Both the amnesics and the controls retained a reaction time advantage for old over new sequences over a one-week period. Similar results have been observed in patients with Alzheimer s disease (Knopman, 1991). The retention of sequence knowledge is not, however, unbounded. Willingham and Dumas (1997) observed that that normal participants did not retain sequence knowledge over a one-year interval. However, because recognition performance in these studies was not measured, it is not clear whether priming and recognition decay at the same or different rates. Moreover, because recognition of this form of sequence learning is generally poor, even in normal control groups (e.g., Destrebecqz & Cleeremans, 2001), any estimate of the decay of recognition over time would, if observed, be confounded by floor effects. Stimuli that consist of sequences of letters generated by an artificial grammar have the advantage that recognition is relatively higher than for stimuli composed of response locations, thus permitting forgetting to be observed. To date no study has examined the rates at which recognition is retained over time relative to priming (in the form of reaction-time savings) in artificial grammar learning. In the following experiment participants were trained to respond to sequences of letters generated by an artificial grammar by pressing the corresponding letters on a keyboard. Immediately after, they were asked to respond to more sequences, some old and some new. After each sequence participants were asked whether they recognized the sequence. Rates of decay were measured by repeating the test on a further two occasions with an interval of seven days between sessions. Method Participants Twenty-four young adults from the University College London community volunteered for this study, twelve male and twelve female. All were naïve to the purpose of the study. Participants were assigned to either a trained group or a control group. Materials There were two kinds of training sequence and two kinds of test sequence. Nineteen grammatical sequences were generated using the finite-state grammar shown in Figure 1. Sequences were composed of the letters M, T, V, R, and X and varied in length from between 3 and 7 letters (mean length = 6.31 letters). The trained group were presented with the grammatical sequences five times during training. The order in which the grammatical sequences were presented was randomized within each block. Control participants were presented with scrambled versions of the training sequences Ð the letters in each grammatical sequence was randomized in each of the five blocks. Thus control participants were not exposed to any of the contingencies in the grammatical sequences but to sequences composed of the same letters and length as the grammatical sequences. The same nineteen grammatical sequences were presented at test mixed with nineteen ungrammatical sequences. These were the grammatical and ungrammatical sequences used by Brooks and Vokey (1991). The ungrammatical sequences were differentiated from the grammatical sequences by rearranging two or more letters into positions not permitted by the grammar, but were of identical length and composed of the same letters as the grammatical sequences. The whole test set was presented twice on each session in a random order to both groups of participants. The materials used for each test were identical on all three sessions. Training Procedure At the start of the experiment participants were given a brief verbal description of the experiment. The following instructions were presented on screen. Thank you for agreeing to take part in this experiment. Your task is simple Ð you will see sequences of letters appear one letter at a time in the center of the screen. The words START and END signal the start and end of each sequence. At these prompts simply press the space bar. During each sequence simply press the corresponding
Decay of Implicit and Explicit Knowledge 127 letter on the keyboard as quickly as you can. Once you have pressed the key a new letter will appear and again press the corresponding key. Try to work as quickly as you can while making as few errors as possible. The training phase began once the participant had read the instructions. Participants were not informed at this stage that they would later receive a recognition test or that they would be asked to return for subsequent tests. Each sequence began after the word START appeared on screen and participants pressed the space bar. Each letter of the sequence appeared on screen for an unlimited time and disappeared when the corresponding key was pressed on the keyboard. Two hundred milliseconds later the next letter appeared, and so on. The end of each sequence was indicated by the appearance of the word END, which prompted participants to press the space bar to begin the next sequence. The training sequences were presented five times in different randomized orders. Testing procedure At the end of the training phase participants were presented with the following set of instructions. You may now take a short break. This part of the experiment is identical to the first Ð each sequence of letters begins with the word START and ends with the word END. At these prompts press the space bar. In the center of the screen you will then see a letter, simply press the corresponding key on the keyboard as quickly as you can. Once you have pressed the key a new letter will appear and again press the corresponding key. Try to work as quickly as you can while making as few errors as possible. After the end of each sequence you will be asked if you recognize the sequence from the previous training phase. After the end prompt of each sequence participants were asked whether they had seen the sequence before by pressing one of two buttons marked yes, I have seen the sequence before and no, I have not seen the sequence before. There were three test phases: the first immediately followed the training phase, the second 7 days later, and the third 14 days later. The second and third test phases were identical to the first with the exception that participants were instructed to try to recognize sequences from the original training phase and not the previous test phase. After the initial test phase participants were asked to return for a follow up session 7 days later, they were not informed of the third session until after the second session. Results Reaction-time savings were calculated by subtracting the mean-response time for grammatical sequences from the mean-response time for ungrammatical sequences. The time taken to respond to the START and END prompts was not entered into any analysis. In addition, extreme reaction times to individual letters were removed from the analyses. These were defined as reaction times to individual letters that exceeded ð 2 standard deviations from each participant s mean reaction time. Note that reaction times to whole sequences were not omitted since the priming scores were calculated by averaging response times to sequences over individual letters. Table 1 shows the mean priming score for each group on each session. The reaction-time savings were entered into a repeated-measures ANOVA with group as the between-subjects factor and session as the within-subjects factor. There was a large effect of group (F(1, 22) = 11.14, MSE = 16952.513, p.003; all tests were conducted to a criterion of.05.), but no effect of Session (F 1), nor an interaction between the two (F 1). Clearly, participants trained on grammatical sequences retained their reaction-time advantage for old over new sequences over the intervals between test sessions. The mean endorsement rates for grammatical and ungrammatical sequences, and resulting d values are shown in Table 2. The d values were entered into a Table 1. Mean Reactions times for Grammatical and Ungrammatical Test Sequences (Figures in Parentheses Are Standard Errors) Control Trained Interval (days) 0 7 14 0 7 14 Grammatical 4423.88 4257.96 4281.74 4355.32 4065.74 3908.51 (252.30) (260.96) (276.17) (252.27) (231.40) (215.58) Ungrammatical 4409.33 4267.60 4280.12 4449.71 4169.41 4011.25 (260.95) (259.90) (280.62) (237.51) (219.60) (196.83) Priming Ð14.55 9.64 Ð1.61 94.39 103.68 102.75 (20.10) (12.29) (16.84) (29.84) (40.37) (40.81)
128 Richard J. Tunney Table 2. Mean Endorsement Rates and d Values for Grammatical and Ungrammatical Test Sequences (Figures in Parentheses Are Standard Errors) Control Trained Interval (days) 0 7 14 0 7 14 Grammatical.64.63.62.66.62.63 (.04) (.06) (.05) (.03) (.04) (.04) Ungrammatical.60.63.58.37.46.48 (.05) (.06) (.06) (.04) (.05) (.04) d 0.10 Ð0.04 0.12 0.80 0.43 0.41 (0.08) (0.13) (0.10) (0.10) (.08) (0.11) repeated-measures ANOVA with group as the between-subjects factor and session as the within-subjects factor. There was an effect of group (F(1, 22) = 18.878, MSE =.224, p.001), an effect of session (F(2, 44) = 5.398, MSE = 0.08, p =.008), and an interaction between the two (F(2, 44) = 3.299, MSE =.254, p =.046). It seems that the trained group s d values decline across sessions. The interaction suggests that initially this decline occurs between sessions 1 and 2, but not between sessions 2 and 3. This is confirmed by a paired-sample t test on the trained group s d values. The d values decrease between sessions 1 and 2 (t(11) = 3.008, SD = 0.425, p.012), but not between sessions 2 and3(t 1). Discussion The present experiment suggests that priming (in the form of motor responses) and recognition decay, or are disrupted, at different rates. Moreover, the decline of d values do not appear to be linear: initially recognition is relatively high, but on the subsequent two sessions the d values, like reaction times, seem to remain constant. Thus the data represent a functional dissociation between priming and recognition in memory for grammatical knowledge. The decline of recognition is consistent with a traditional forgetting function of initial rapid forgetting toward asymptote first described by Ebbinghaus (1885). More recently, Wixted (1990; see also Wixted & Ebbesen, 1991), suggested that this function is a simple power function. Priming may well decay according to the same function but it is clear that the decay constant to fit each curve would be different. How should these data be interpreted? Two possibilities spring to mind. One possibility, is that the data are consistent with a dual-process model of recognition memory of the sort proposed by Gabrieli (1998; see also Knowlton & Squire, 1994). According to this model there are in fact two dissociations present in these data: the first between procedural and declarative memory, and the second between recollective and familiarity components of recognition. The high recognition scores immediately following training are predicated on both familiarity and conscious recollection. The lower recognition scores on subsequent test sessions reflect the decay of conscious recollection, but remain constant and above chance as recognition continues to be supported by priming. Reaction times remain constant because they are supported by familiarity but not conscious recollection. This interpretation is consistent with the experiment reported by Higham et al. (2000). However, dissociations between different measures of knowledge should not be taken to conclusively demonstrate separable processes until alternative explanations have been explored. A second possibility is that a unitary source of knowledge primes both motor and recognition responses. The dissociation between the two reflects not the source of knowledge, but the different ways of responding. The distinction made by Seger (1998) between judgment and motor responses seems appropriate: Recognition involves a decision whether to say either old or new, and necessarily varies according to the various sensitivities and biases of participants. It seems unlikely that motor responses require any such decisions to be made (although relative speeds may vary and are likely to be subject to noise). By this analysis, the dissociation is due not to different representational schemes but to different ways of accessing that knowledge. Ironically, priming may simply be a more direct and recognition a more indirect test of knowledge. Recently, Shanks and Perruchet (2002) have described a simple mathematical model in which a single source of memory can give rise to dissociations between priming and recognition in serial reaction time learning, although it is not clear whether this model could account for the data reported here. Across sessions the endorsement rates for grammatical sequences are similar for both the trained and control groups. However, the trained group s endorsement rates for ungrammatical sequences
Decay of Implicit and Explicit Knowledge 129 increase across sessions. The pattern of endorsement rates might indicate that interference is the cause of the apparent decay of recognition. No such pattern is apparent in the priming data. This may indicate that the form of forgetting differs between the two systems. If so this would exclude a single process explanation, but further research would be required to examine this issue. The data do not resolve these possible interpretations but nonetheless appear to support the claim made by Reber (e.g., 1989; Allen & Reber, 1980), and others (e.g., Berry & Dienes, 1993) that implicit knowledge is more robust in the face of time than explicit knowledge. Learning and memory of this form has traditionally been referred to as implicit, in part because the indirect performance aspect of learning (i.e., priming) is preserved while direct explicit aspect of memory (i.e., recognition) is not. However, these data do not indicate that priming can occur without recognition, thus learning cannot be said to be implicit in the sense that it occurred without awareness: because priming and recognition cooccur across sessions we cannot say that either measure of learning is process pure. Indeed evidence that learning can occur without awareness, as defined by chance recognition performance has been elusive (see Shanks, in press). Thus it has been difficult to demonstrate dissociable processes of priming and recognition. The advantage of investigating the rate of decay of these two forms of memory is that chance performance on one or the other measures is not required to demonstrate that they are dissociable systems. Previous research (e.g., Nissen et al. 1989; Willingham & Dumas, 1997) has shown that priming can be retained over relatively long periods of time. The data reported here, are the first to show that recognition of the same information, while persisting for a similar length of time, shows a very different pattern of forgetting. In this sense, the data reported here support the long held view that at least some form of implicit knowledge can be retained for longer periods than explicit knowledge. References Adams, J. A. (1987). Historical review and appraisal of research on the learning, retention, and transfer of human motor skills. Psychological Bulletin, 101, 41Ð74. Allen, R., & Reber, A. S. (1980). Very long-term memory for tacit knowledge. Cognition, 8, 175Ð185. Berry, D. C., & Dienes, Z. (1993). Implicit learning: Theoretical and empirical issues. Hove, UK: LEA. Brooks, L. R., & Vokey, J. R. (1991). Abstract analogies and abstracted grammars: Comments on Reber (1989) and Mathews et al (1989). Journal of Experimental Psychology: General, 120, 373Ð383. Cleeremans, A. (1993). Mechanisms of implicit learning: Connectionist models of sequence processing. Cambridge: MIT Press. Destrebecqz, A., & Cleeremans, A. (2001). Can sequence learning be implicit? New evidence with the process dissociation procedure. Psychonomic Bulletin & Review, 8, 343Ð350. Dulany, D. E., Carlson, R. C., & Dewey, G. I. (1984). A case of syntactical learning and judgement: How conscious and how abstract? Journal of Experimental Psychology: General, 113, 541Ð555. Ebbinghaus, H. (1964). Memory: A contribution to experimental psychology. (H. A Reuger & C. E. Bussenius, Trans.). New York: Dover. (Original work published 1885). Foster, J. K., & Jelicic, M. (Eds.) (1999). Memory: Systems, Process, or Function? Oxford: Oxford University Press. Gabrieli, J. D. E (1998). Cognitive neuroscience of human memory. Annual Review of Psychology, 49, 87Ð115. Higham, P. A., Vokey, J. R., & Pritchard, J. L. (2000). Beyond dissociation logic: Evidence for controlled and automatic influences in artificial grammar learning. Journal of Experimental Psychology: General, 129, 457Ð470. Knopman, D. (1991). Long-term retention of implicitly acquired learning in patients with Alzheimer s disease. Journal of Clinical and Experimental Neuropsychology, 13, 880Ð894. Knowlton, B. J., & Squire, L. R. (1994). The information acquired during artificial grammar learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 79Ð91. McBride, D. M., & Dosher, B. A. (1997). A comparison of forgetting in an implicit and explicit memory task. Journal of Experimental Psychology: General, 126, 371Ð392. Nissen, M. J., Willingham, D. B., & Hartman, M. (1989). Explicit and implicit remembering: When is learning preserved in amnesia? Neuropsychologia, 27, 341Ð 352. Perruchet, P., & Pacteau, C. (1990). Synthetic grammar learning: Implicit rule abstraction or explicit fragmentary knowledge? Journal of Experimental Psychology: General, 119, 264Ð275. Reber, A. S. (1989). Implicit learning and tacit knowledge. Journal of Experimental Psychology: General, 118, 219Ð235. Reber, A. S., & Allen, R. (1978). Analogic and abstraction strategies in synthetic grammar learning: A functionalist interpretation. Cognition, 6, 189Ð221. Seger, C. A. (1998). Independent judgment-linked and motor-linked forms of artificial grammar learning. Consciousness and Cognition, 7, 259Ð284. Shanks, D. R. (in press). Implicit learning. In K. Lamberts & R. Goldstone (eds.), Handbook of Cognition. London: Sage. Shanks, D. R., & Perruchet, P. (2002). Dissociation between priming and recognition in the expression of sequential knowledge. Psychonomic Bulletin and Review, 9, 362Ð368. Tulving, E., Schacter, D., & Stark, H. (1982). Priming effects in word-fragment completion are independent of recognition memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 8, 336Ð342. Willingham, D. B., & Dumas, J. A. (1997). Long-term retention of a motor skill: Implicit sequence knowledge
130 Richard J. Tunney is not retained after a one-year delay. Psychological Research, 60, 113Ð119. Willingham, D. B., Nissen, M. J., & Bullemer, P. (1989). On the development of procedural knowledge. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 1047Ð1060. Wixted, J. T. (1990). Analyzing the empirical course of forgetting. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 927Ð935. Wixted, J. T., & Ebbesen, E. B. (1991). On the form of forgetting. Psychological Science, 2, 409Ð415. Richard J. Tunney Department of Psychology Keele University Keele Staffordshire, ST5 5BG United Kingdom Tel.: +44 1782 583669 Fax: +44 1782 583387 E-mail: r.tunney@keele.ac.uk Received 14 June, 2002 Revision received 9 September, 2002 Accepted 4 October, 2002