|Home | About | Journals | Submit | Contact Us | Français|
Child development is accompanied by a robust increase in immediate memory. This may be due to either an increase in the number of items (chunks) that can be maintained in working memory, or an increase in the size of those chunks. We tested these hypotheses by presenting younger and older children (7 and 12 years old) and adults with different types of lists of auditory sentences: four short sentences, eight such sentences, four long sentences, and four random word lists, each read with a sentence-like intonation. Young children accessed (recalled words from) fewer clauses than older children or adults, but no age differences were found in the proportion of words recalled from accessed clauses. We argue that the developmental increase in memory span was due to a growing number of chunks present in working memory, with little role of chunk size.
One can examine developmentally two different aspects of working memory, the small amount of information temporarily held in mind at any time. One can examine the number of separate units, or chunks, that are maintained (e.g., Miller, 1956; Broadbent, 1975; Cowan, 2001), as well as the size of each chunk. This chunk size depends on the participant's knowledge so that, for example, a chunk can be either a familiar word presented in isolation or a well-learned word pair (Cowan, Chen, & Rouder, 2004: Chen & Cowan, 2005, in press). Previous research documents a robust increase in working memory capacity as children develop (see Cowan & Alloway, 2009, for a review). Is this due to an increase in the number of chunks that can be maintained, the sizes of chunks, or both? There is long-standing debate regarding this question.
Gilchrist, Cowan, and Naveh-Benjamin (2008) used the coherence of words within clauses to examine working memory decline in adult aging. They found that aging was accompanied by a decline in the number of chunks that could be recalled, with no apparent decline in chunk size. Specifically, in memory for lists of spoken sentences, older adults recalled words from fewer clauses, but recalled the same amount as young adults from any clause that was at least partly recalled. This was referred to as diminished clause access with undiminished clause completion. This was interpreted as an aging decline in chunk capacity, with no change for these materials in chunk size. Here we examine whether the developmental difference between children and adults is similar in nature. To set the stage, we briefly review the empirical basis for a chunk limit in working memory, discuss competing theories of developmental increases in capacity, and discuss possible developmental influences on the formation and retention of chunks from linguistic materials.
Philosophers and early experimental psychologists proposed a limited working memory (e.g., James, 1890) leading to limits in immediate recall of about seven items (e.g., Miller, 1956). Other research has shown, however, that other factors affect the amount that can be recalled. Baddeley, Thomson, and Buchanan (1975) showed that adults typically could recall as much as they could pronounce in about 2 s in serial recall; this time limit was based on refreshment of a temporary representation through covert rehearsal. The “magical number seven” reported by Miller could reflect the typical number of monosyllabic words recalled when a basic capacity was supplemented by rehearsal or even on-line formation of multi-item chunks. Subsequent work suggested that, when the supplementary processes are eliminated, the limit in adults is in the range of three to five chunks (Broadbent, 1975; Cowan, 2001).
Given the 2-s limit for a phonological rehearsal loop (Baddeley et al., 1975; Baddeley, 1986), one way to examine the chunk limit is to use chunks so long that they preclude use of the loop, as in coherent linguistic materials. In the earliest such study, Tulving and Patkau (1962) presented series of words that varied in their approximation to English text, ranging from nonsensical to fully coherent. Free recall was examined for ‘adopted chunks’, sets of words recalled in their correct serial order. They found an invariant limit in recall of about four to six adopted chunks for all materials, but the size of adopted chunks (number of words per chunk) was dependent upon the level of approximation to English. Subsequent studies with linguistic materials (Glanzer & Razel, 1974; Simon, 1974) suggest slightly lower limits of 3 or 4 linguistic chunks on average, more in keeping with Broadbent (1975) and Cowan (2001).
Some research suggests a developmental increase in the number of chunks that can be maintained. Pascual-Leone (1970, 2005) proposed that as children age, the number of available items that can be maintained increases by a specified, invariant number of age-related units. It was claimed that performance increases with age could not be explained by the development of strategic factors, such as the ability to form chunks, or by increases in knowledge (for a review, see Dempster, 1981). In one experiment, for example, Pascual-Leone (1970) taught different stimulus-response relations (e.g., raise your hand for a square; clap your hands for a red item) and then combined the cues into multi-dimensional signals (e.g., raise your hand and also clap hands for a red square). The number of dimensions that children could handle increased with age. Working in this tradition, Burtis (1982) provided sets of items that could form multi-item chunks (repetitions of the same letter; both letters in a pair colored red; pairs familiar through repetition or forming known acronyms) and found an increase in the number of chunks recalled. Similar results have been obtained in procedures in which each presented item is assumed to comprise a separate chunk because there is not enough free processing time devoted to the stimuli to allow them to be rehearsed or grouped (Cowan, Nugent, Elliott, Ponomarev, & Saults, 1999; Cowan et al., 2005; Cowan, Naveh-Benjamin, Kilb, & Saults, 2006).
Based on other studies, though, there are also reasons to expect that the average size of a chunk could increase with age. As children grow, they gain experiential knowledge and sophisticated linguistic ability (Chi, 1978; Ottem, Lian, & Karlsen, 2007). Also, children begin to use strategies to guide memory (Flavell, Beach, & Chinsky, 1966; Ornstein, Naus, & Liberty, 1975; Case et al., 1982; Kail, 1992; Harris & Burke, 1972; Towse et al., 1999). As these increase with development, it is assumed that size of chunks will also increase, as these can increase item meaningfulness. Recently, Ottem et al. (2007) proposed that the characteristic developmental increase in memory span was solely due to growth in linguistic abilities. Intuitively, language allows a person to better rehearse and encode items (Flavell et al., 1966), as well as to form associations that aid chunking of items, proposed to be a critical factor for the span increase. In contrast to Pascual-Leone and other neo-Piagetians, Ottem et al. (2007) maintained that it was the size, and not the number, of available chunks that increased with age.
Given this controversy, it is difficult to know whether the number of chunks in working memory, as opposed to chunk size, actually does change with age. We would be in an especially advantageous situation if chunk size remained constant across age groups, whereas the total amount recalled changed. That pattern of results could unequivocally be interpreted as a change in capacity in chunks. That was the result obtained by Gilchrist et al. (2008) for adult aging, using lists of simple, unrelated sentences to be recalled. Although the finding could be similar for child development, there are considerations of language development that theoretically could lead to a different outcome.
Some aspects of language improve throughout childhood (e.g., Chomsky, 1969). For our purposes, linguistic stimuli must be chosen with the aim of minimizing the differences as they pertain to working memory. There could be a developmental increase in the ability to use relatively simple linguistic structure to remember the details of short sentences, or in the ability to use more complex linguistic structure to remember the two halves of long sentences. Young children's processing of the structure under a working memory load might suffer from the unavailability of resources needed to carry out the linguistic processing, as suggested by some investigators (Daneman & Case, 1981; Kail & Hall, 2001). Reduced processing efficiency may correspond to less complete syntactic or semantic encoding of language in younger children. Consequently, what is a single chunk to an older participant could be encoded as multiple chunks in a young child.
Hopefully, though, the simplicity of stimulus materials will minimize such differences. The basic retention of sentences falling well within the individual's linguistic competence may occur automatically with few demands on attention (Allen & Baddeley, 2009; Caplan, Waters, & DeDe, 2007). If so, we can anticipate that participants in all age groups will have similar chunk sizes for the materials and will differ only in the number of unrelated linguistic units or chunks held in memory.
Our primary goal was to observe age differences in the number of chunks held in working memory but phonological storage and rehearsal could make a contribution to recall along with a chunk-capacity-limited mechanism. Chen and Cowan (2005) found a greater contribution of phonological processes for serial recall as opposed to free recall or free scoring of serial recall. We were therefore able to minimize the contribution of phonological rehearsal processes not only by using lists of sentences so as to exceed greatly the 2-s limit of phonological rehearsal in adults (e.g., Baddeley et al., 1975), but also by requiring free recall rather than serial recall.
We varied the number of unrelated sentences within a list, as well as their length. To examine the development of this capacity we presented children with the following: (1) lists of four short sentences, each with one independent clause; (2) lists of eight such short sentences; (3) lists of four long sentences, each composed of two meaningfully-conjoined clauses; and (4) lists of four random pseudo-sentences, made up of various words mixed together in a haphazard manner. We counterbalanced materials so that participants never received the same stimulus throughout the experiment. Every long sentence could be broken down into two comprising short sentences; participants in one group received a long sentence, and the other group received its comprising sentences (see Table 1).
On the basis of these conditions we could observe whether there are age differences in the access and completion of clauses, indexing the number and size of chunks recalled, respectively. The 4-short-sentence condition was included to allow a sensitive age comparison given previous research suggesting that most young adults have an immediate capacity limit of around four chunks (e.g., Cowan, 2001; Chen & Cowan, 2005). The list conditions, summarized in Table 2, permit several critical comparisons. First, the lists of 4 short and 8 short sentences differ in list length, but not in the amount of coherence within a sentence. If there is a constant-capacity mechanism and no effect of list length in this procedure, the result should be access to the same number of sentences in both conditions. This should differ according to the individual's capacity. Second, the lists of 8 short and 4 long sentences differ in the amount of coherence between clauses, but not in the list length. There should be no difference between trial types in the usefulness of a phonological-length-based retention mechanism, but the 4-long-sentence condition should produce higher performance if the two clauses within a long sentence are sometimes combined to form a single long chunk. We can determine if all age groups benefit similarly from this extra structure in the 4-long-sentence condition. Third, the lists of 4 short sentences and 4 random pseudo-sentences also are similar in phonological length, but differ in coherence inasmuch as the syntactic and semantic structure that combines words into clause units is absent from the random pseudo-sentences. We can determine whether all age groups benefit similarly from this extra structure in lists of coherent sentences.
The method serves as an opportunity for evidence convergent with what Pascual-Leone (1970) obtained, using a very different method to determine the units or chunks in memory. Pascual-Leone assumed that each stimulus cue-response association taught to a child served as a unit. In contrast, we used pre-existing linguistic knowledge (cf Burtis, 1982; Tulving & Patkau, 1962) in a way that allowed us to verify the coherence of a chunk by examining the mean amount of completion of a clause. If the completion remains the same across age groups, then the difference in performance can be attributed to a difference in the number of units accessed, i.e., chunks in working memory.
Participants included 25 first-grade children (mean age 7.73 years, SD=0.21), 26 sixth-grade children (mean age 12.43 years, SD=0.39), and 24 adults (mean age 18.37 years, SD=0.49). Children were recruited from local public schools and adult data were from Gilchrist et al. (2008). All participants reported normal or corrected-to-normal vision and hearing. Children received $10 and a book as compensation, and adults received course credit.
Four different sentence conditions were presented to participants, with each condition comprised of one of three possible sentence types (i.e., short sentences, long sentences, or random pseudo-sentences). Short sentences were simple, one-clause sentences (e.g., Thieves took the painting). Long sentences were composed of two short sentences that were meaningfully conjoined (e.g., Our neighbor sells vegetables but he also makes fruit juice). Finally, random pseudo-sentences were composed of words presented with little syntactic structure (e.g., a close football your cheese). Random pseudo-sentences were equivalent to short sentences in terms of length, and were presented using a sentence intonation.
The sentence conditions were formed by varying the types of sentences presented, as well as their number in some cases. A given list of sentences to be remembered could contain: (1) 4 short sentences, (2) 4 long sentences, (3) 8 short sentences, or (4) 4 random pseudo-sentences. Two trials were presented for each sentence condition and the order of trials was randomized across participants.
The stimuli were the same as those used in Gilchrist et al. (2008). Spoken sentences were presented in a female voice between 45 and 70 dB. Content words within these sentences had age of acquisition norms with ratings between 100 and 350, found using the MRC psycholinguistic database (Wilson, 1988). Short sentences and random pseudo-sentences ranged from 3 to 5 words in length and long sentences ranged from 8 to 11 words in length.
Participants were tested one at a time, in a sound-attenuated booth equipped with a computer, headphones to listen to presented stimuli, and a microphone to record participant recall. The participant was instructed to listen carefully to the presented sentences, and to recall verbally what was just heard in any order when cued to do so. Trials began with the word “Ready” on the computer screen for 1000 ms. Spoken stimuli were then presented via headphones.
To ensure that any possible differences across age groups were not due to the selection of materials, short and long sentences were reassigned to different conditions across participants as a means of counterbalancing. Long sentences presented to half of the participants were broken into two short sentences presented in different conditions to the other half (Table 1). For example, participants in the first group might be presented with the long sentence I upset my mother when I lied about the money. Participants in the second group would be presented with the short sentences I upset my mother and I lied about the money, with each sentence placed into different trials of a given condition (here, either the four- or eight-short-sentences condition) to reduce any potential sentence-specific effects. In a similar manner, long sentences presented to the second group were deconstructed into short sentences that were presented to the first group of participants.
The four or eight sentences in a given condition were separated by 1000-ms pauses. After the last sentence in a given condition was presented and was followed by a final 1000-ms pause, a 500-ms, 400-Hz tone was presented to cue recall. The participant then provided responses by speaking into the microphone. In each condition, the participant was given 1 minute to recall as many of the words as possible, in any order, but he or she was free to terminate recall earlier via a keypress. Another keypress signaled when the participant was ready to move to the next list of sentences.
All verbal responses were converted into sound files and saved for later transcription by the experimenter.
We used three different measures to examine how different aspects of immediate memory for sentences changed with development: words recalled per trial, clause access, and clause completion. Words recalled per trial was measured as the total number of words recalled from each sentence condition. As more than one instance of a word could be present within a trial (e.g., the often appeared more than once), each occurrence of that word in recall was scored as a separate word recalled.
Two special measures were used also by Gilchrist et al. (2008), adapted from Naveh-Benjamin et al. (2007). Clause access was taken as a measure of the number of independent groups presented on a trial that were successfully retrieved. We measured this as the number of clauses from which at least one content word was retrieved. This measure was based on the assumption that words recalled from one particular clause or one-clause sentence could be presumed to form part of the same chunk present within working memory. Previous research suggests that words recalled from a given clause are not typically recalled in isolation; instead, access to a clause typically entails the recall of multiple words from the clause (Gilchrist et al., 2008).
Designating the unit of analysis as a clause meant that short sentences had one such unit and long sentences had two. It is possible that young children are less able than older participants to join two short sentences into an overarching, long sentence, in which case there would be more separate units to be recalled in young children, producing especially poor clause access in the 4-long-sentences condition. For the random condition each random pseudo-sentence was counted as a clause, allowing this condition to serve as a control to observe the effect of linguistic coherence in the other conditions.
Clause completion was defined as the proportion of word recalled from a clause, contingent on at least one word having been recalled. This measure was assumed to reflect the amount of inter-word association between words in a clause, i.e., the coherence of the clause as a chunk in memory.
Words recalled yields a holistic view of developmental growth in working memory, and the access and completion measures yield a more analytic understanding of that developmental growth.
Developmental increases in mean words recalled per trial can be observed in Figure 1. A repeated-measures ANOVA of this variable included age as a between-subjects factor and sentence condition as a within-subjects factor. There was a significant effect of age, F(2,72)=12.26, ηp2=.254, p<.0001, with fewer words recalled in younger children (M=7.52, SD=2.90) than in older children (M=11.22, SD=2.90) or adults (M=10.81, SD=2.48). Post-hoc Newman-Keuls tests showed significant differences between first-grade children and the two older age groups, which did not differ. There was also a significant effect of condition, F(3,216)=135.95, ηp2=.654, p<.0001, with the greatest number of words recalled for four long sentences (M=14.30, SD=5.82), followed by eight short sentences (M=10.51, SD=3.78), four short sentences (M=9.66, SD=2.60), and four random pseudo-sentences (M=4.93, SD=1.82). Performance levels on all conditions were significantly different from all others in the post-hoc tests, except that performance on four short sentences and eight short sentences did not differ. The overall pattern of performance replicates what is found in a separate analysis of the young adults: four long > eight short = four short > four random.
The main effects were qualified by a significant age group × condition interaction, F(6,216)=5.06, ηp2=.123, p<.0001. Newman-Keuls tests indicated that the first-grade pattern was the same as the adult pattern (above). Interestingly, the pattern in the intermediate, sixth-grade group was more differentiated: four long > eight short > four short > four random. The discrepancy occurred because the sixth-grade children did better than adults on 8 short, unrelated sentences. The reason for this counter-intuitive difference is unclear but it can be said, at least, that performance on our task is adult-like by the sixth grade.
Additionally, we examined whether qualitative patterns in recall differed across age groups in three ways. First, certain items within a sentence may have been more likely to be recalled over others. Specifically, we were interested in comparing the recall of key nouns or verbs within each clause to items that were less critical for overall comprehension (e.g., articles, prepositions). Indeed, critical items within clauses were more likely to be recalled over non-critical items, F(1, 72) = 183.84, ηp2 = .72, p<.001 (critical, M=.41, SD=0.11; non-critical, M=.34, SD=0.11). Although all age groups showed a recall benefit for critical items, a significant age × item type interaction, F (2, 72) = 6.30, ηp2 =.15, p<.005, revealed that this advantage was smaller for young children relative to the older age groups, who did not differ from each other. Means for critical and non-critical items were .32 and .28 for first-graders, .47 and .38 for sixth-graders, and .45 and .37 for adults; SD = 0.11 in each case. Thus, items are not equiprobable in their likelihood of being recalled and young children rely somewhat less than older participants on the items that are critical for general comprehension and gist.
Second, we investigated any instances of synonyms for the correct words recalled for each trial. In the prior analysis, synonyms were not included among words recalled. If young children were more likely than older participants to recall synonyms of words instead of the correct words, our scoring could be viewed as underestimating capacity in young children and therefore overestimating the age difference in capacity. Overall, instances of recalled synonyms were extremely rare, with mean recall of synonyms per trial being less than 1. There was an effect of age in terms of mean synonyms recalled, F(2,72)=3.53, ηp2=.09, p<.05, and Newman-Keuls tests indicated that more synonyms were recalled by adults (M=.22, SD=.15) than by first-grade (M =.14, SD=.15) or sixth-grade (M =.12, SD=.15) children, who did not differ. Given that synonym recall increased with age, this finding amplifies, rather than nullifies, the age difference in words recalled verbatim.
The greatest number of synonyms recalled in place of presented words occurred for four long sentences (M=.31, SD=.38), with increasingly fewer recalled for eight short (M =.21, SD=.31), four short (M =.12, SD=.24), and four random (M =.006, SD=.06) sentence conditions. This difference was statistically significant, F(3,216)=17.10, ηp2=.191, p< .0001. Post-hoc Newman-Keuls tests indicated that all conditions were significantly different from each other. The interaction between age and sentence condition did not approach significance. Thus, the conditions and groups producing the most verbatim words also produced the most synonyms.
Finally, it was possible that different patterns of recall across serial positions could have been present in the different age groups. For this reason, we examined number of words recalled per clause serial position for each age group and sentence condition. It was assumed that if young children differed from older age groups in serial position effects, this would be manifest as an age × clause position interaction for each sentence condition. However, for all conditions, repeated-measures ANOVAs produced no significant interactions of age × clause position [four long: F(14, 504) =1.18, p = .29; four random: F(6, 216) < 1, p = .56; four short: F(6, 216) < 1, p = .71; eight short: F(14, 504) = 1.27, p = .22]. In general, regardless of age group, words contained in list-final clauses were most likely to be recalled (see Figure 2), with words recalled from four short sentences as the sole exception [four long: F(7, 504) = 32.12, ηp2=.308, p< .0001; four random: F(3, 216) = 118.47, ηp2=.622, p< .0001; four short: F(3, 216) = 2.21, p = .09; eight short: F(7, 504) = 43.38, ηp2=.376, p< .0001]. This suggests that the significant age differences that we found were not influenced by changes in serial position effects with age.
As in the previous study, we operationally defined the chunk as the most coherent unit presented, the clause or short sentence. In the case of random pseudo-sentences, we defined a chunk as a presented pseudo-sentence, equivalent in length to a one-clause, short sentence. Figure 3 shows that there were age differences in the number of clauses accessed. The ANOVA showed a significant effect of age group, F(2,72)=13.08, ηp2=.267, p< .0001. Newman-Keuls tests indicated that first-grade children accessed fewer clauses (M=2.34, SD=.79) than sixth-grade children (M=3.37, SD=.79) or adults (M=3.28, SD=.79), who did not differ.
There was also a significant effect of sentence condition, F(3,216)=117.08, ηp2 = .619, p<.0001, with the greatest number of clauses accessed from four long sentences (M=4.16, SD=1.49), followed by eight short sentences (M=3.19, SD=1.12), four short sentences (M=2.83, SD=.67), and four random pseudo-sentences (M=1.80, SD=.65). The post-hoc tests showed that all conditions were significantly different from each other.
The above effects were qualified by a significant interaction, F(3,216)=5.28, ηp2=.128, p<.0001. Newman-Keuls tests indicated an interesting pattern. In the older two groups, the pattern was the same as for words recalled (adults, four long > eight short = four short > four random; sixth-grade children, four long > eight short > four short > four random). In first-grade, children, though, there was little difference between sentence types: four long = eight short = four short > four random. The absence of a difference between four long sentences and eight short sentences indicate that the first-grade children did not benefit significantly in clauses accessed from the association between clauses in the four-long-sentence condition.
As mentioned above, we defined clause completion as the proportion of words recalled from a given clause, given that it had already been accessed (i.e., conditional upon at least one content word having been recalled). Importantly, as Figure 4 illustrates, we found no significant effect of age upon clause completion, F(2,70)=1.22, ηp2 = .034, p =.30. Power analyses showed that we were able to detect an effect of ηp2 = .08, a small effect, with a power of .8.
There was a significant effect of condition, F(3,210)=12.97, ηp2 =.156, p<.0001. Completion rates were highest for four short sentences (M=.81, SD=.12) and eight short (M=.81, SD=.13), followed by four long (M =.74, SD=.13), and four random (M=.71, SD=.18) sentences. Shorter sentences, by virtue of their simplicity as well as brevity, are easier to complete than are long sentences that contain additional words or random pseudo-sentences that are incoherent in their meaning and structure. Newman-Keuls test showed only that the lists of shorter sentences (either 4 or 8 short sentences) had higher levels of completion than lists of either long or random, incoherent sentences.
It might seem remarkable that four random pseudo-sentences could be completed about as well as four long sentences. The way that this occurred is that it was usually the most recent pseudo-sentence from which most words were recalled, presumably making use of phonological memory. The equivalence of four short and eight short sentences at a high rate of completion (.81) strengthens the assumption that these short sentences function as integrated units or chunks in working memory.
This effect was qualified by a significant age × condition interaction, F(6,210)=3.50, ηp2 =.091, p<.01, in contrast to Gilchrist et al. (2008), who found no significant differences in patterns of completion between young and older adults. Newman-Keuls analyses showed that, to a first approximation, both young adults and sixth-grade children conformed to the pattern, four short = eight short > four long = four random. (The groups differed slightly in that, in adults, the four long versus four short comparison did not reach significance; whereas in sixth-grade children, the four long versus eight short comparison did not reach significance.) In first-grade children, in contrast, there was no difference between conditions (see Figure 4). Linguistic structure did not aid completion rate in first-graders.
We examined whether the increase in working memory performance accompanying development was driven by an increased ability to retain more chunks, or to form larger chunks. The number of chunks was estimated as the number of unrelated short, 1-clause sentences at least partly recalled (i.e., accessed) from lists of such sentences. The size of each chunk was estimated as the proportion of words recalled from an accessed clause. The results indicated a developmental increase in the number of independent chunks recalled. In contrast, chunk completion, the proportion of words recalled for a given accessed clause, remained at a similar level across the age groups tested. In our further discussion we first examine the effects of our manipulations of the stimulus type on three dependent measures, in order to understand what they can tell us about the memory processes for our materials. Then we examine the basis of developmental changes that were obtained.
It is clear that increasing the amount of information present within a given word sequence facilitates recall (Figure 1). Across all age groups, when sequences contained no linguistic coherence, as in the four-random-pseudo-sentences condition, the number of words recalled was small and roughly corresponded to the limits of working memory capacity to be expected given previous work (Baddeley, 1986; Broadbent, 1975; Cowan, 2001; Cowan et al., 2005; Miller, 1956), under the assumption that each word functioned as a separate chunk. Specifically, the number recalled was about 4 in young children and increased to about 6 in adults, which approximates the usual limit if phonological memory for sequences can make a contribution.
With the linguistic coherence provided by sentence structure in the other conditions, the numbers of words recalled per trial were increased substantially, as one would expect under the assumption that each chunk unit in working memory then includes multiple words, such as the words within a clause. Linguistic structure in the short sentences roughly doubled the number of words recalled, compared to random pseudo-sentences. The reason it did not increase the number of words recalled even more may be that the phonological memory trace makes little contribution for stimulus materials exceeding a spoken duration of about 2 s (Baddeley, 1986).
Another indication of the benefit of linguistic coherence across all age groups is that the number of words recalled was highest when the words were presented within long sentences. The eight-short-sentences condition contained the same number of clauses as the four-long-sentence condition, but did not provide a similar advantage in recall. This difference between conditions shows that meaningfully connecting information across clauses helps to increase the number of words that can be remembered. However, the facilitation of recall that a long sentence provided is less than one would expect if each long sentence were encoded as a single chunk. In particular, although the four-long-sentences condition contains double the number of words present within the four-short-sentences condition, there was less than a two-to-one ratio of words recalled in these two conditions, for every age group. It appears that the association between clauses was sometimes, but not always, of use, a conclusion further supported below.
Below, clause completion will provide evidence for the clause as a functional unit or chunk. Assuming that to be the case for the time being, results from the proportion of clauses accessed (i.e., clauses from which at least one substantive word was recalled) can be taken as evidence for a constant capacity in working memory. Between 2 and 4 clauses were accessed in the four- and eight-short-sentence conditions (Figure 2), as would be expected according to previous developmental evidence on capacity limits (e.g., Cowan, 2001; Cowan et al., 2005). The finding that the number of clauses accessed was independent of the number of sentences is an important confirmation that a mnemonic process limited by the number of chunks is at work here. The long length of the sentence lists puts them out of range for there to be much of a contribution of phonological memory and the resulting number of clauses accessed is similar to the core verbal working memory capacity that can be obtained by requiring articulatory suppression during recall (Chen & Cowan, in press).
In the case of random pseudo-sentences, individual words may function as separate chunks so the method of scoring each pseudo-sentence as a single clause resulted in very few (1 to 2) such pseudo-sentences being accessed. This condition therefore effectively served as a control to observe the effect of removing syntactic and semantic associations between items in a clause.
As was found in the number of words recalled, access to clauses was facilitated by additional semantic and syntactic structure. Thus, access for clauses within four long sentences was greater than those within eight short sentences, despite both conditions containing the same number of independent clauses. We believe that, in this case, pairs of clauses within a long sentence sometimes function as a single chunk.
Results from clause completion suggested that a one-clause sentence generally did serve as a functional chunk in working memory. In the four- and eight-short-sentences conditions, in which each sentence had one clause, we found high levels of clause completion (~80%) for all age groups (Figure 4), despite proportions of clause access well below 80%. This suggests strong associations between words within a sentence and much weaker associations between words in different unrelated sentences, a pattern that validates the idea that the short sentence generally serves as a single chunk in this experiment (cf Cowan, 2001).
In the older groups, the level of clause completion was somewhat lower for the remaining conditions but this age-related trend will be discussed below in the developmental differences section. Notice that the overall pattern is one in which the number of units accessed varies widely across conditions whereas the completion of accessed units varies only slightly across conditions. This organization of recall into large sequences could be driven at least in part by syntactic and semantic coherence in the linguistic conditions, versus only phonological memory and prosodic coherence in the four-random-pseudo-sentences condition.
Previous research suggests that, relative to adults, children have reduced working memory capacity, as well as a reduced immediate memory span (e.g., Cowan et al., 2005). As mentioned above, these developmental differences may either be due to young children having a smaller number of chunks that can be held in memory, or to them forming smaller chunks. For the lists of simple linguistic materials that we used, the results provide support for the former factor (number of chunks retained) over the latter. The meaning of the developmental change in each of three dependent measures will be discussed in turn.
We found evidence for a developmental increase in the number of words recalled in all sentence conditions (Figure 1), consistent with the well-known increase in span with development. Older children and adults generally recalled more words per condition than did young children.
Young children were able to take advantage of semantic or syntactic structure. Like the older groups, young children recalled about twice as many words from short sentences as they did from random pseudo-sentences. Structure provided by long sentences also was used by all age groups, who recalled more words from four long sentences than from either short-sentence condition. The advantage for long sentences, however, was somewhat smaller in young children than in the older groups, suggesting that the young children were somewhat less likely than older participants to combine the clauses of a long sentence into a single chunk.
In general, developmental studies of memory that have documented an increase in span with age have typically used word recall as a measure of interest. Although our results provide additional confirmation of these prior findings, examining recalled words provides insufficient understanding of potential mechanisms that underlie the characteristic increase in memory span. For this reason, we also examined developmental differences in clause access and completion. These measures can help determine whether any age differences observed reflect changes in the number or size of chunks maintained in working memory.
We found a developmental increase in the number of independent units that could be retrieved. With access to a clause taken as memory of one chunk (even though it was often an imperfect memory of the presented chunk), the number of remembered chunks increased with age (Figure 2). The increase in the number of recalled chunks from about 2.5 in young children to 3 or 4 in older children and adults accords well with previous observations (e.g., Chen & Cowan, in press; Cowan et al., 2005).
The increase in clause access with age was not dependent on a certain level of structure; it occurred in all four conditions. Nevertheless, there is evidence that the use of information from longer structures increased with age. The increase in the number of clauses accessed in long sentences compared to short sentences was significant by post-hoc tests in the two older age groups, but not in the young children. This is perhaps further evidence that the appreciation of subtle details of linguistic structure continue to develop in the elementary school years (e.g., Chomsky, 1969).
Clause completion in older children and adults changed across sentence conditions (Figure 4), implying sensitivity to the linguistic structure. Compared to short sentences, there was a diminished completion of “clauses” in the form of random pseudo-sentences, as one might expect if more than one slot in working memory had to be used to retain the words that were recalled from a single pseudo-sentence. There was also a diminished completion of clauses from long sentences. Perhaps these older participants attempted to use the associations between clauses within a long sentence to retain more clauses, which could occur at the expense of some within-clause information.
Interestingly, as shown in Figure 4, the younger children showed no sensitivity to linguistic structure in their clause completion. As noted previously, there were no age differences in serial position effects that might account for this difference in the use of linguistic structure. It is possible that the intonation structure was more important than high-level information for chunk formation in young children (cf Cowan, 1989). Their focusing on intonation structure typically led them to recall most of the words from a single pseudo-sentence, whereas older participants were more likely to sacrifice words from a pseudo-sentence in order to recall something from a larger number of pseudo-sentences, or to sacrifice words from clauses within a long sentence in order to recall something from a larger number of such clauses. Given that the sentences were well within young children's ability, the absence of an effect of sentence type on clause completion is an interesting and new finding.
Despite these somewhat subtle age differences in the effects of conditions, there was no age difference in the mean level of clause completion across conditions. This result extends the findings of Gilchrist et al. (2008), who obtained no effect of adult aging on clause completion, only on clause access.
As children get older, the amount of information that can be held in memory increases. This finding is robust, and has been observed in many developmental studies. Yet, it has often been unclear exactly what factors underlie developmental differences in memory span. Using measures not only to examine words recalled in spoken sentences, but also to examine how many independent one-sentence clauses came to mind (clause access) and how much of each such clause was recalled (clause completion), we found large developmental differences in the level of clause access, but not in clause completion. From these results, we propose that an increase in the number of chunks that can be stored in working memory underlies developmental differences in immediate recall, converging with the developmental increase proposed by several others (e.g., Burtis, 1982; Cowan et al., 2005; Pascual-Leone, 1970, 2005).
We wish to emphasize the importance of these findings, as the results converge with findings from previous studies, despite very different experimental methodologies and assumptions. Studies by neo-Piagetians showing changes with age in M-space often relied on a priori assumptions of how chunks were formed based on the number of units that were presented in the stimuli (e.g., Burtis, 1982; Pascual-Leone, 1970). In contrast, our clause completion measure allows us to verify that the mean size of the units in memory did not differ by age group. Despite these methodological differences, our clause access measure confirms the conclusion that there are developmental improvements in immediate memory coming from an increase in the number of units that can be maintained. Contrary to some other results (e.g., Ottem et al., 2007), we found no evidence for an increase in chunk size with age. This outcome might, of course, be different for more complex materials.
This research was conducted with support from NIH Grant R01-HD21338. We thank Angela AuBuchon, Melissa Knipe, Caleb O'Brien, and Christopher Zwilling for assistance, and J. Scott Saults for programming.
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.