Regularization in Short-Term Memory for Serial Order - CiteSeerX

Journal of Experimental Psychology: Learning, Memory, and Cognition 2005, Vol. 31, No. 2, 351–358

Copyright 2005 by the American Psychological Association 0278-7393/05/$12.00 DOI: 10.1037/0278-7393.31.2.351

OBSERVATIONS

Regularization in Short-Term Memory for Serial Order Matthew Botvinick and Lauren M. Bylsma University of Pennsylvania Previous research has shown that short-term memory for serial order can be influenced by background knowledge concerning regularities of sequential structure. Specifically, it has been shown that recall is superior for sequences that fit well with familiar sequencing constraints. The authors report a corresponding effect pertaining to serial recall errors. Undergraduate participants performed immediate serial recall on sequences of pseudowords generated on the basis of an artificial grammar. After extensive experience with this material, recall errors displayed a bias toward regularizing responses, response sequences more probable, with respect to the artificial grammar, than the originally presented stimulus sequence. This regularization effect squares well with recent trace redintegration and Bayesian models of serial recall, and appears to represent an analog of the schema-based error patterns observed in other domains of memory.

of serial recall in general (Botvinick & Plaut, 2003; Hartley & Houghton, 1996; Henson, 1998). It is thus important to understand its characteristics and underlying causes. To date, work examining the impact of background knowledge on short-term sequence memory has focused almost exclusively on recall accuracy. In contrast, very little attention has been paid to patterns of error. This is striking, considering the central importance of error analysis in the larger serial recall literature. In this article, we present an experiment that focused on errors of serial recall in a domain containing sequential structure. The basic finding to be reported is that such errors display a tendency toward regularization: a bias toward high-probability sequences, and, in particular, toward sequences that are higher in probability than the to-be-recalled stimulus. Preliminary evidence for such a regularizing tendency was presented by Mayzner and Schoenberg (1964) on the basis of observations from the bigram frequency paradigm. They focused on cases in which two adjacent letters in a low bigram-frequency stimulus were reversed in the recall response, noting that such exchanges tended to yield a higher frequency bigram than the one appearing in the stimulus. Mayzner and Schoenberg interpreted this as a regularization effect, suggesting that “low-frequency digrams in the string might frequently be inverted in recall to their high-frequency complements” (p. 399). However, although this finding is suggestive, it does not conclusively establish the presence of regularization. The problem is that the observed increment in bigram frequency could well have occurred by chance. Given that the stimulus strings were composed of low-frequency bigrams, it seems plausible that even random exchanges of adjacent letters would yield higher frequency bigrams. The case of Mayzner and Schoenberg (1964) makes it evident that, in order to demonstrate regularization, it is not sufficient to show that errors tend to be more well-formed or regular than stimulus sequences. Rather, it must be shown that regularizing

In a classic study of immediate serial recall, Baddeley, Conrad, and Hull (1965) demonstrated that consonant strings were better recalled when they contained high-frequency letter transitions than when they contained low-frequency transitions. This bigram frequency effect (later replicated by Kantowitz, Ornstein, & Schwartz, 1972; Mayzner & Schoenberg, 1964) provides one of numerous demonstrations that short-term memory for serial order can be influenced by background knowledge concerning domain-specific regularities in sequential structure. Even before Baddeley et al.’s (1965) study, Miller and Selfridge (1951) had observed better recall for sequences of words that contained high-frequency word transitions than for sequences containing low-frequency transitions. More recently, a number of studies have shown better recall for nonwords containing high-frequency, versus low-frequency, phoneme sequences (a phonotactic frequency effect; Gathercole, 1995; Gathercole, Frankish, Pickering, & Peaker, 1999; Gathercole, Willis, Emslie, & Baddeley, 1991; Grant et al., 1997; Roodenrys & Hinton, 2002; van Bon & van der Pijl, 1997). In each of these instances, knowledge concerning regularities in sequential structure exerts a clear influence on short-term sequence memory, with highly probable sequences better recalled than less probable ones. In addition to being interesting in its own right, this finding may also have important implications for understanding language acquisition (Storkel, 2001) and for constraining models

Matthew Botvinick and Lauren M. Bylsma, Department of Psychiatry and Center for Cognitive Neuroscience, University of Pennsylvania. The present work was supported by Grant MH16804 from the National Institutes of Mental Health to Matthew Botvinick. We thank Sara Clopton, Andrew Radu, and Sheriza Baksh for assistance in conducting the experiment. Correspondence concerning this article should be addressed to Matthew Botvinick, University of Pennsylvania, Center for Cognitive Neuroscience, 3720 Walnut Street, Philadelphia, PA 19104-6241. E-mail: mmb@mail .med.upenn.edu

351

OBSERVATIONS

352

responses are more frequent than would be expected by chance. More specifically, this means showing that regularizing responses are more common than would be observed if recall were not influenced by background knowledge concerning sequential regularities. We designed this experiment to allow for such a comparison. The approach involved having participants perform immediate serial recall on sequences of pseudowords. For one group of participants (to which we refer as the experimental group), these sequences were generated on the basis of an artificial grammar, meaning that some item-to-item transitions and by extension, some overall sequences, were more likely to occur than others. For a second (control) group, all sequences were equally probable. We analyzed errors in the experimental group in order to establish the frequency of regularizing responses, that is, errors where the response sequence was of higher probability, given the artificial grammar, than the stimulus sequence. We used the performance of the control group to infer the frequency of regularizations that could have been expected by chance, that is, the frequency that would be expected if performance were entirely unaffected by prior experience with the grammar. In order to test for a genuine regularization effect, we compared the performance of the experimental group with this chance baseline. Some aspects of the performance of the present experimental group were reported in a previous article (Botvinick, in press) written before we decided to run the additional, control group. This earlier article dealt with error rates for different classes of stimuli and did not consider error content. In the present article, in order to avoid redundancy, we focus exclusively on error content.

Method Participants Participants included 14 members of the University of Pennsylvania community who responded to an advertisement on an electronic bulletin board. Six individuals participated in the experimental condition, and 8 individuals (recruited several months later) participated in the control condition. Age ranged from 18 to 22 years. Four of the participants in each condition were men. All participants were fluent in English. Participants provided informed consent prior to the beginning of the experiment and were paid $10 per hour for their participation. The protocol was approved by the University of Pennsylvania Institutional Review Board.

Materials Stimuli were six-item lists, containing each of the pseudowords dah (pronounced to rhyme with hah), fie ( pie), poe (doe), kay (day), tee (tea), and noo (who). For the control group, throughout the experiment, each of the 720 permutations of these pseudowords was equally likely to be presented for recall. The target lists for the experimental group were generated on the basis of an artificial grammar, which made some permutations more likely to occur than others. According to the grammar, the six pseudowords were arbitrarily sorted into two groups of three, designated Group A and Group B (the assignment of specific pseudowords to these groups differed across participants). The first item in each target sequence was selected at random, and so could belong to either A or B. Thereafter, items were selected on the basis of the transition probabilities listed in Table 1. Once a full six-item sequence had

Table 1 Transition Probabilities Used in Constructing Stimulus Lists To From

A1

A2

A3

B1

B2

B3

A1 A2 A3 B1 B2 B3

0 .125 .125 .25 .25 .25

.125 0 .125 .25 .25 .25

.125 .125 0 .25 .25 .25

.25 .25 .25 0 .125 .125

.25 .25 .25 .125 0 .125

.25 .25 .25 .125 .125 0

Note. Row and column labels designate the members of Item Categories A and B, each mapping to a specific pseudoword (with different mapping across subjects). Six-item lists were generated based on these probabilities but were presented only if the sequence contained no repeated items.

been generated, it was discarded if any item appeared more than once, ensuring that every list constituted a permutation of the same six items. Note that the probabilities in Table 1 create a tendency for Group A items to be succeeded by Group B items, and vice versa. As a consequence, the most frequently occurring list structure is ABABAB or BABABA (30% of lists generated by the grammar). Lists with the structure AAABBB or BBBAAA are least likely (1.9% of lists generated). Table 2 presents the frequencies of each of the 10 possible list structures. Note that the probability of occurrence for each list type can be understood as reflecting its goodness of fit with the alteration constraint implicit in the grammar; the more violations of this soft constraint a list structure contains, the less probable that structure is to occur. In view of this, we will, in places, refer to the probability of a list type as its goodness of fit, or simply its goodness (see also Botvinick, in press). List types for which the locations of Group A and B items are simply reversed— e.g., ABBAAB and BAABBA—are equivalent with regard to both goodness and neighborhood relations. For brevity, we use only one of these pairs to refer to both.

Procedure The experiment was divided into 15 sessions of approximately 1 hr in length conducted on consecutive days. The first 7 sessions were intended to familiarize the participants with the artificial grammar. Participants were seated before a desktop computer and were asked to place the index, middle, and ring fingers of their left hands on the keys F, D, and S, respectively, and the corresponding fingers of their right hands on J, K, and L. On each trial during these sessions, an auditory target list was presented. Lists were presented at a rate translating to 120 words per minute in a woman’s voice simulated with text-to-speech software (2nd Speech Center, Zero2000 Software, Nanjing, China). The same prosody, involving a gradually falling pattern of emphasis, was used for all lists. Each stimulus list was followed immediately by a visual cue (a series of six horizontally arrayed underscores at the center of the computer monitor). Along with the cue, the six pseudowords were visually presented at the bottom of the monitor in a left-to-right order that varied randomly from trial to trial. Participants were asked to select these words in the order of their occurrence within the target list. They selected the words by pressing the key corresponding to each word’s position on the monitor (S for leftmost word, D for the second word from left, F for the third word from left; L for rightmost word, K for the second word from the right, and J for the third word from the right). On the selection of each word, a red strike-out would appear over that word, and pressing the corresponding key on subsequent steps of recall would produce no response. Thus, participants were required to enter, as their overall response,

OBSERVATIONS

Table 2 List Structures and Their Associated Probabilities

3.

Probability List structure ABABAB, ABABBA, ABBABA, AABBAB, ABBAAB, ABAABB, AABABB, ABBBAA, AABBBA, AAABBB,

BABABA BABAAB BAABAB BBAABA BAABBA BABBAA BBABAA BAAABB BBAAAB BBBAAA

Exemplars

Category

.0042 .0021 .0021 .0010 .0010 .0010 .0010 .0005 .0005 .0003

.302 .151 .151 .075 .075 .075 .075 .038 .038 .019

Note. Each structural category contains 72 specific stimulus lists. Probabilities listed are shown both for specific exemplars of each stimulus category and for the stimulus category as a whole. Values are rounded.

some permutation of the six pseudowords. Participants were asked to announce their responses aloud as they entered them via the keyboard, using the same prosodic pattern as heard in the stimuli. Participants were encouraged, in cases in which they could not confidently recall the target list, to make their best guesses. Performance was self-paced; participants pressed the space bar to trigger initiation of each trial. Participants performed 250 trials per session with one short break. Session 8 was identical to the preceding sessions, except that delay items were introduced between encoding and recall. After presentation of the target list, one, two, three, five, or seven individual Arabic numerals (the digits 1–9) were presented at the center of the screen at a pace of one numeral every 500 ms. Participants were asked to read these aloud as they appeared. Immediately following the final delay item, the usual recall cue appeared along with the six pseudowords, and responses were entered as in prior sessions. Participants were not required to recall the delay items. In the interest of avoiding floor or ceiling effects in subsequent sessions, we used performance data from Session 8 to select, for each participant, a number of delay items that would be likely to yield an overall response accuracy near 50%. This number of delay items was used on every trial in Session 9. On subsequent sessions, the number of delay items was increased by one if accuracy on the preceding session exceeded 60% and was reduced by one if accuracy fell below 40%. Beyond this, Sessions 9 –15 were conducted identically to Sessions 1– 8. We conducted the experiment using the E-Prime software package (Psychological Research Tools, Pittsburgh, Pennsylvania).

Data Analysis We performed analyses on data collected during Sessions 9 –15. We discarded the first five trials in each session. For the experimental group, we then sorted trials into 10 groups on the basis of the list types enumerated in Table 2. Within each group, we categorized trials on the basis of global response accuracy. For each error trial, we evaluated three aspects: 1.

The response permutation involved, that is, the mapping from items in the stimulus to items in the response. We coded this by ordering digits representing stimulus items, as those items appeared within the response. Thus, a response that transposed Items 2 and 3 would be denoted 132456.

2.

The list type of the response on the basis of the categorization in Table 2.

353 The probability of this list type occurring as a stimulus on the basis of the artificial grammar. In keeping with the terminology already established, we refer to this as the goodness of the response sequence.

We used the performance of the control group to infer what would be expected under the null hypothesis, that is, under the assumption that errors were entirely unaffected by experience with the artificial grammar. As in the experimental group, errors were coded according to permutation, and for each participant, the frequency of every permutation was established. Using the resulting distribution for each control participant, we computed what the goodness of errors would have been for each category in Table 2, had experimental participants produced the same distribution of permutations for that category. Again, this provided an estimate of what performance in the experimental group would have been like if experimental participants had made errors that were unaffected by experience with the grammar. We refer to this inferred pattern of performance as the baseline pattern and to the pattern produced by the experimental group as the observed pattern. Statistical analyses, which we describe in detail in conjunction with results, involved comparisons between these two patterns.

Results Three participants in the control group did not consistently attain accuracies above 40% during Sessions 8 –15 and were therefore excluded from further analysis. The median number of delay items presented during Sessions 9 –15 was three for the experimental group (range 1–5) and four for the control group (range 1–7).

Response Accuracy As indicated earlier, accuracy data for the experimental group are presented in detail in a separate article (Botvinick, in press). For present purposes, we note only that response accuracy was positively correlated with stimulus goodness, in keeping with previous studies of recall accuracy in sequentially structured domains (see Figure 1). Accuracy for nonexcluded control participants ranged between 51% and 56% (with a median of 54%).

Error Analysis The mean goodness of incorrect responses for the experimental group (the observed goodness, following the terminology introduced above) was .14 (SD ⫽ .018). The baseline value, derived from the performance of the control group (weighted to reflect differences in stimulus class frequency) was .11 (SD ⫽ .002). A t test based on individual participant data indicated a significant difference between these two values ( p ⬍ .01). In order to obtain a more detailed picture, we repeated the same analysis for each level of stimulus goodness (see Table 2). Results are shown in Figure 2. A two-way repeated measures analysis of variance (ANOVA) was performed on participant means for response goodness, with factors for stimulus goodness (five levels) and error pattern (observed vs. baseline). This yielded a significant main effect of stimulus goodness, F(4, 36) ⫽ 19.0, p ⬍ .0001, but, more important, a main effect of error pattern, F(1, 9) ⫽ 14.6, p ⬍ .005. No statistically significant interaction was observed between these two factors, F(4, 36) ⫽ 1.5, ns. Individual t tests for each level of stimulus goodness indicated a statistically significant difference in every case ( p ⬍ .04).

354

OBSERVATIONS

Figure 1. Mean accuracy for experimental group across Sessions 9 –15 for each level of stimulus goodness. Bars indicate standard errors, computed in a fashion appropriate to the reported repeated measures analysis as recommended by G. R. Loftus and Masson (1994).

In order to test for a bias toward regularizing responses, we calculated, for each level of stimulus goodness, the proportion of errors that yielded a sequence higher in goodness than the stimulus (see Figure 3). (We excluded stimuli with structure ABABAB, given that it was not logically possible to regularize stimuli in this class.) Once again, we calculated baseline values using the performance of the control participants. We conducted an ANOVA on the proportions obtained, with stimulus goodness (four levels) as a within-subject factor and error pattern (observed vs. baseline) as a between-subjects factor. This yielded a significant main effect of error pattern, F(1, 9) ⫽ 12.9, p ⬍ .01, as well as a main effect of stimulus goodness, F(3, 27) ⫽ 359.2, p ⬍ .0001. An Error Pattern ⫻ Stimulus Goodness interaction was also present, F(3, 27) ⫽ 3.2, p ⬍ .05. Individual t tests indicated that the proportion of regularizing errors was higher in the experimental group than the control group for all levels of stimulus goodness ( p ⬍ .02), except for the highest level, where a trend in the same direction was observed ( p ⫽ .07). Note that different response permutations are needed to regularize the various stimulus classes enumerated in Table 2. Inspection of the error patterns for specific stimulus classes revealed a consistent weighting toward those permutation patterns that had the effect of regularizing the stimulus. The three stimulus classes AABBAB, ABBAAB, and ABAABB provided a particularly

striking example. In addition to having the same level of goodness, these classes’ structures can be regularized to ABABAB with the simple exchange of two adjacent items. It is important to note, however, that the relevant items differ across the three stimulus classes. For stimuli with the structure AABBAB, the items in Positions 2 and 3 must be exchanged, yielding the response permutation 132456; for ABBAAB stimuli, it is Items 3 and 4 that must be exchanged (response permutation 124356); and for ABAABB stimuli, it is Items 4 and 5 (permutation 123546). As shown in Figure 4, the pattern of errors for each of these three stimulus classes was weighted toward the regularizing permutation. The most common error for stimulus class AABBAB was 132456; for ABBAAB, it was 124356; and for ABAABB, it was 123546. We found the difference in error patterns across the three stimulus types to be statistically significant on the basis of a repeated measures ANOVA on error proportions with permutation (three levels) and stimulus class (three levels) as within-subject factors (interaction term), F(4, 20) ⫽ 9.4, p ⬍ .0002.

Discussion It is well established that background knowledge concerning domain-specific regularities in sequential structure can affect immediate recall for serial order. Previous work demonstrating this

OBSERVATIONS

355

Figure 2. Mean goodness of error responses for each level of stimulus goodness. Solid line ⫽ observed values; dashed line ⫽ baseline values. Bars indicate standard errors.

point has focused almost exclusively on recall accuracy. This experiment provides clear evidence that the content of errors is also influenced by background knowledge. Specifically, errors show a tendency toward regularization. In the present data, such a tendency was evident in two interrelated effects. First, errors showed a bias toward high-goodness (highly probable) sequences, evident in the fact that the mean goodness of error responses was higher than would be expected if error patterns had been unaffected by the presence of sequential regularities in the stimuli. Second, regularizing responses— error responses higher in goodness than the to-be-recalled stimulus list—were more frequent than would be expected, on the basis again of the assumption that the structure of the stimuli did not affect error patterns. Consistent with this, different stimulus structures elicited different patterns of error, with errors weighted toward rearrangements of the stimulus that yielded responses higher in goodness. The finding of regularization in short-term recall for serial order is consistent with results pertaining to other domains of memory. Since the pioneering work of Bartlett (1932), it has been repeatedly observed that episodic memory errors tend to be shaped by background knowledge, or, as it is sometimes expressed, that memory errors tend to be schema-based (e.g., Hannigan & Reinitz, 2001; see also Bower, Black, & Turner, 1979; E. F. Loftus & Palmer, 1974). Bartlett explained this effect in terms of a reconstructive process whereby background knowledge is used to disambiguate degraded memory traces. This basic idea remains central to many

theories of memory (see, e.g., Shiffrin, 2003). In recent work, it has been applied to the domain of immediate serial recall. Specifically, a number of researchers have suggested that long-term knowledge might be used to reconstruct or disambiguate degraded memory traces for sequences, a process sometimes referred to in this context as trace redintegration (Hulme et al., 1997; Lewandowsky, 1999; Nairne, 1990; Neath, 2000; Schweickert, 1993). Although this idea has been applied primarily to the issue of item recall (e.g., to explain effects of lexicality), initial efforts have been made to extend it to cover phenomena pertaining more specifically to serial order. Specifically, Gathercole et al. (1999) have extended an earlier model by Schweickert (1993) in order to explain the phonotactic frequency effect, and Botvinick (in press; see also Botvinick & Plaut, 2003) has presented an account of the bigram frequency effect, which casts trace redintegration as a process of Bayesian inference. These models, in keeping with the available empirical data, have focused on the impact of serial structure on recall accuracy. However, the idea of trace redintegration also fits nicely with the data on error content that we have reported here. According to trace redintegration accounts, and reconstructive memory accounts more generally, the recall process is biased toward sequences that fit well with established schemata or (equivalently) sequences that are judged to have a high prior probability of occurrence. This bias explains accuracy-based phenomena, such as the bigram and phonotactic frequency effects, because it works in favor of accurate

356

OBSERVATIONS

Figure 3. Proportion of regularizing responses for each level of stimulus goodness observed (solid line) and baseline (dashed line). Bars indicate standard errors based on individual subject data.

recall for highly probable stimuli. However, on a trace redintegration account, the same bias should be evident in patterns of error. That is, the same principle that gives rise to better recall for highly probable stimuli should also give rise to a regularizing tendency among errors. The findings of the present experiment confirm this implicit prediction of trace redintegration accounts, adding further support to their relevance for immediate serial recall. At the same time, the current findings present a substantive challenge for many general models of serial recall. This is because the vast majority of recent computational work on serial recall (e.g., Brown, Preece, & Hulme, 2000; Burgess & Hitch, 1999; Henson, 1998; Houghton, 1990; Page & Norris, 1998) has posited mechanisms that are insensitive to regularities of serial structure. Because these models provide no account for how background knowledge may bias the recall process, they cannot be directly applied to such phenomena as the bigram and phonotactic frequency effects. The present findings, indicating regularization in serial recall errors, are equally difficult for such models to address, given that they, too, reflect recall biases driven by background knowledge for domain structure. It is, of course, possible that existing models could be extended to account for both accuracy and error patterns in structured domains. In one attempt along these lines, Hartley and Houghton (1996) proposed a model that superimposed a procedural memory mechanism upon an otherwise independent short-term memory system, allowing the selection of each item during recall to be influenced by the other items recalled so far. Such an arrangement

would presumably produce regularization errors of the kind observed in the present experiment. However, further observations from the present experimental paradigm appear to contradict the Hartley and Houghton account. Specifically, as we reported elsewhere (Botvinick, in press; Botvinick & Plaut, 2004), the recall of items in structured sequences may be influenced not only by which items have been recalled so far but also by which items occurred at later positions in the target sequence. It is conceivable that the Hartley and Houghton paradigm could somehow be modified to accommodate such an effect. However, there is reason to suspect that the issue of domain structure will ultimately demand a fundamentally different computational approach to immediate serial recall (Botvinick & Plaut, 2003, 2004). In addition to yielding results with significant theoretical implications, the present experiment also introduced a potentially useful methodological innovation. Previous studies of immediate serial recall in structured domains have exploited preexperimental knowledge, such as knowledge for the orthographic or phonotactic structure of a native language. Our approach focused instead on knowledge acquired in the laboratory. This allowed the sequential structure of the stimulus domain to be both precisely characterized and directly controlled and allowed performance to be compared with that of a control group, whose members dealt with precisely the same stimulus items sequenced randomly. Such an approach has been used with great success in the study of procedural memory (Reber, 1967), in work on language acquisition (Saffran, 2001) and production (Dell, Reed, Adams, & Meyer, 2000), and in

OBSERVATIONS

357

Figure 4. Average proportion of errors involving reversals of Items 2 and 3, 3 and 4, and 5 and 6 for three stimulus categories. Bars indicate standard errors based on individual subject data.

some studies of rote serial learning (Cumming, Page, & Norris, 2003; Miller, 1958). The work we have presented demonstrates that the approach can be fruitfully applied to the study of immediate serial recall in structured domains. One limitation of the paradigm applied in this work is that any response sequence produced was very likely to have occurred as a stimulus earlier in the experiment. In view of this point, one could interpret the observed regularization effect as reflecting a recall procedure based on the retrieval of whole lists rather than one influenced by finer grained sequencing constraints. The interpretation is challenged, if not ruled out, by experiments demonstrating the bigram and phonotactic frequency effects (as reviewed in the introduction to this article), because in those experiments, recall was performed on novel material. Nevertheless, here, as in general, further research will be needed before we can clearly adjudicate among the various mechanisms by which background knowledge may impact serial recall.

References Baddeley, A. D., Conrad, R., & Hull, A. J. (1965). Predictability and immediate memory for consonant sequences. Quarterly Journal of Experimental Psychology, 17, 175–177. Bartlett, F. C. (1932). Remembering. Cambridge, England: Cambridge University Press. Botvinick, M. (in press). Effects of domain specific knowledge on recall for serial order. Cognition. Botvinick, M., & Plaut, D. C. (2003). Constructive processes in immediate serial recall: A recurrent network model of the bigram frequency effect.

In B. Kokinov & W. Hirst (Eds.), Constructive memory (pp. 129 –137). Sofia, Bulgaria: New Bulgarian University. Botvinick, M., & Plaut, D. C. (2004). Short-term memory for serial order: A recurrent neural network model. Manuscript submitted for publication. Bower, G. H., Black, J. B., & Turner, T. J. (1979). Scripts in memory of text. Cognitive Psychology, 11, 177–220. Brown, G. D. A., Preece, T., & Hulme, C. (2000). Oscillator-based memory for serial order. Psychological Review, 107, 127–181. Burgess, N., & Hitch, G. J. (1999). Memory for serial order: A network model of the phonological loop and its timing. Psychological Review, 106, 551–581. Cumming, N., Page, M., & Norris, D. (2003). Testing a positional model of the Hebb effect. Memory, 11, 43– 63. Dell, G. S., Reed, K. D., Adams, D. R., & Meyer, A. S. (2000). Speech errors, phonotactic constraints, and implicit learning: A study of the role of experience in language production. Journal of Experimental Psychology: Learning, Memory, and Cognition, 26, 1355– 1367. Gathercole, S. E. (1995). Is nonword repetition a test of phonological memory or long-term knowledge? It all depends on the nonwords. Memory & Cognition, 23, 83–94. Gathercole, S. E., Frankish, C. R., Pickering, S. J., & Peaker, S. (1999). Phonotactic influences on short-term memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 25, 84 –95. Gathercole, S. E., Willis, C., Emslie, H., & Baddeley, A. D. (1991). The influences of syllables and wordlikeness on children’s repetition of nonwords. Applied Psycholinguistics, 12, 349 –367. Grant, J., Karmiloff-Smith, A., Gathercole, S. E., Paterson, S., Howlin, P., Davies, M., & Udwin, O. (1997). Phonological short-term memory and

358

OBSERVATIONS

its relationship to language in Williams syndrome. Cognitive Neuropsychiatry, 2, 81–99. Hannigan, S. L., & Reinitz, M. T. (2001). A demonstration and comparison of two types of inference-based memory errors. Journal of Experimental Psychology: Learning, Memory, and Cognition, 27, 931–940. Hartley, T., & Houghton, G. (1996). A linguistically constrained model of short-term memory. Journal of Memory and Language, 35, 1–31. Henson, R. N. A. (1998). Short-term memory for serial order: The start-end model. Cognitive Psychology, 36, 73–137. Houghton, G. (1990). The problem of serial order: A neural network model of sequence learning and recall. In R. Dale, C. Mellish, & M. Zock (Eds.), Current research in natural language generation (pp. 287–319). London: Academic Press. Hulme, C., Roodenrys, S., Schweickert, R., Brown, G. D. A., Martin, S., & Stuart, G. (1997). Word-frequency effects on short-term memory tasks: Evidence for a redintegration process in immediate serial recall. Journal of Experimental Psychology: Learning, Memory, and Cognition, 23, 1217–1232. Kantowitz, B. H., Ornstein, P. A., & Schwartz, M. (1972). Encoding and immediate serial recall of consonant strings. Journal of Experimental Psychology, 93, 105–110. Lewandowsky, S. (1999). Redintegration and response suppression in serial recall: A dynamic network model. International Journal of Psychology, 34, 434 – 446. Loftus, E. F., & Palmer, J. C. (1974). Reconstruction of automobile destruction: An example of the interaction between language and memory. Journal of Learning and Verbal Behavior, 13, 585–589. Loftus, G. R., & Masson, M. E. J. (1994). Using confidence intervals in within-subject designs. Psychonomic Bulletin & Review, 1, 476 – 490. Mayzner, M. S., & Schoenberg, K. M. (1964). Single-letter and diagram frequency effects in immediate serial recall. Journal of Verbal Learning and Verbal Behavior, 3, 397– 400.

Miller, G. A. (1958). Free recall of redundant strings of letters. Journal of Experimental Psychology, 56, 485– 491. Miller, G. A., & Selfridge, J. A. (1951). Verbal context and the recall of meaningful material. American Journal of Psychology, 63, 176 –185. Nairne, J. S. (1990). A feature model of immediate memory. Memory & Cognition, 18, 251–269. Neath, I. (2000). Modeling the effects of irrelevant speech on memory. Psychonomic Bulletin & Review, 7, 403– 423. Page, M. P. A., & Norris, D. G. (1998). The primacy model: A new model of immediate serial recall. Psychological Review, 105, 761–781. Reber, A. S. (1967). Implicit learning of artificial grammars. Journal of Verbal Learning and Verbal Behavior, 6, 855– 863. Roodenrys, S., & Hinton, M. (2002). Sublexical or lexical effects on serial recall of nonwords? Journal of Experimental Psychology: Learning, Memory, and Cognition, 28, 29 –33. Saffran, J. R. (2001). The use of predictive dependencies in language learning. Journal of Memory and Language, 44, 493–515. Schweickert, R. (1993). A multinomial processing tree model for degradation and redintegration in immediate recall. Memory & Cognition, 21, 167–175. Shiffrin, R. M. (2003). Modeling memory and perception. Cognitive Science, 27, 341–378. Storkel, H. L. (2001). Learning new words: Phonotactic probability in language development. Journal of Speech, Language, and Hearing Research, 44, 1321–1337. van Bon, W. H. J., & van der Pijl, J. M. L. (1997). Effects of word length and wordlikeness on pseudoword repetition by poor and normal readers. Applied Psycholinguistics, 18, 101–114.

Received April 15, 2004 Revision received August 2, 2004 Accepted August 16, 2004 䡲