9 Working memory and short-term memory storage

What does backward recall tell us?

Gerald Tehan

Kaye Mills

One of the recent issues for those interested in working memory is the distinction between what have become to be known as simple span and complex span tasks. This distinction is made because some have argued that the different tasks rely upon different storage mechanisms (Baddeley 1986; Brainerd and Kingma 1985, Swanson 1996). The current research continues the exploration of the grounds for such a distinction. It also deals with a second issue, that of the direction of recall. Almost all of the research into working memory has involved forward serial recall and many sophisticated formal models of serial recall exist (Burgess and Hitch 1999; Henson 1998; Farrell and Lewandowsky 2002, Page and Norris 1998). In contrast, no such models of backward recall exist. Moreover, there is no agreement as how to conceptualize backward recall: while there are some clear similarities between forward and backward recall in simple span tasks, there are also many differences. Those who focus upon the differences argue for different retrieval mechanisms in backward recall to those in forward recall. The current research combines the common storage and direction of recall issues by examining backward recall of simple, complex and delayed memory tasks.

Common storage

The distinction between short-term memory and working memory has generally been based upon the assumption that short-term memory tasks are primarily tasks that focus upon storage capacity. In working memory tasks, however, processing issues are seen as just as important as storage issues. To that extent a common definition of a working memory task is one in which small amounts of information must be kept highly active and available for later use during cognitive processing of some other information. For example, in the operations span task devised by Engle and his colleagues (Engle, Kane and Tuholski 1999), words must be kept in memory for future recall while simultaneously processing of maths problems. Clearly, both simple and complex tasks require the storage of items for later recall. The dispute has been about whether different stores underlie the different tasks.

Exploring differences between short-term memory (STM) and working memory tasks has largely been conducted using individual differences methodologies. The basis for the exploration is that working memory tasks appear to be better predictors of other complex cognitive tasks (e.g. measures of fluid intelligence) than short-term memory tasks, and it is often argued that this relationship reflects the common processing demands of working memory and complex cognitive tasks rather than storage issues. The individual differences approach, however, is complicated by the fact that short-term memory tasks often correlate with the complex cognitive (p.154) tasks (although not as strongly as the working memory tasks) and more importantly short-term memory and working memory tasks correlate quite highly with each other. This high correlation is the basis for the common storage assumption. In short, an individual differences approach highlights the fact that short-term memory and working memory tasks have much in common. In fact, the collinearity problem has often made it extremely difficult to tease apart storage and processing components of the two types of tasks.

A less frequent approach to the common storage issue has involved experimental techniques. Here factors that are known to affect performance on a simple span task are also applied to working memory tasks. For present purposes, an experiment by Tehan, Hendry and Koscinski (2001) illustrates this. In their experiment they examined word length and phonological similarity effects across immediate serial recall, operation span and Brown–Peterson tasks. It is well documented that in immediate serial recall (simple span, STM task) short words are better recalled than long words, and phonologically similar items are less well recalled than phonologically dissimilar items. Tehan et al. wanted to know if these same factors would influence recall in a working memory task (operations span) and a long-term memory task (Brown–Peterson) task. It emerged that all three tasks produced robust word length and phonological similarity effects. In spite of theoretical reasons why one might expect different outcomes across the three tasks, the only thing that differentiated performance on the three tasks was absolute levels of recall. There was little support for the notion that the underlying processes for the three tasks differed.

In sum, a distinction between short-term memory, working memory and to a lesser extent long-term memory appears consistently in the literature. The empirical evidence for such a distinction is not quite so compelling. Instead, both the individual differences and experimental literature indicate that most short-term retention tasks, be they simple span, operations span or delayed recall, share much in common and that it is extremely difficult to distinguish between such tasks.

Backward recall

Backward serial recall has not been as extensively or as systematically studied as forward recall. In contrast to forward recall there are no formal mathematical models of how the task is done. Where the issue has been addressed, the assumption is often made that backward recall is simply repeated serial forward recall (Page and Norris 1998). That is, in order to recall the last word in the list participants use forward recall to get to the last item, they output it, and then start forward serial recall through the list again until they get to the second from last item, they output it and then commence the same process again and again until only the first item has to be recalled. This explanation does have some empirical support in that response latencies differ for backward and forward recall, and that the pattern of latencies for backward recall is consistent with the notion of repeated cycles of forward recall (Thomas, Milner and Haberlandt 2003).

There are two interesting corollaries of such an explanation. First, describing simple backward recall in such a way seems to qualify the task as a working memory task. That is, participants need to continually keep hold of where they are in a list while forward recalling to that point. Secondly, if backward recall is nothing more than repeated forward recall, the factors that are known to influence forward recall should have a similar impact upon backward recall. As it turns out, there are factors that have similar influences on backward and forward recall, but there are others that have differential effects.

The current research centers on two factors that are know to have robust influences forward recall: phonological similarity and word length. In immediate forward serial recall similar-sounding words are less well recalled than dissimilar-sounding words and short words are better (p.155) recalled than long words. As mentioned above, these effects also generalize to working memory and delayed memory tasks (Tehan et al. 2000). While the effects are readily observed in forward recall, there is some contention as to their effects in backward recall.

Forward and backward recall seem to be differentially affected by verbal and visual distractor activity, and this has led to the proposition that forward recall relies upon phonological coding, but backward recall does not. The genesis of this debate can be found in Li and Lewandowsky’s (1993, 1995) findings that verbal distractors disrupted forward but not backward recall, whereas visuospatial activity adversely affected backward but not forward recall. Hulme et al. (1997) also suggested that participants did not rely upon a phonological strategy in order to explain their finding that word frequency effects had little impact on backward recall. Farrand and Jones (1996) provided counter evidence to the Li and Lewandowsky conclusion, but probably the most direct test of the phonological coding assumption is in a study by Rosen and Engle (1997) where they directly examined the phonological similarity effect in immediate forward and backward recall. Phonological similarity effects were observed for each direction. The logical conclusion from this study was that, at least for immediate recall, both forward and backward recall were supported by phonological representations.

With respect to word length, there have been a number of studies that have explored word length effects in backward recall with the general outcome that word length effects are present but weaker than with forward recall (Cowan, Wood and Borne 1994; Walker and Hulme 1999). For example, Cowan et al. (1994) looked at differences between immediate recall and continuous distractor tasks on backward recall. The study involved mixed and pure lists of short and long words, that is lists varied as to whether short and long words occurred early or late in a sequence, with some beginning with long words and ending in short words (LS) or vice versa (SL), and others in which all were short (SS) or all were long (LL). Results showed that the typical short word advantage was present on the immediate backward recall task. However, on the continuous distractor task, a long word advantage emerged. Cowan et al. concluded that this result was prima facie evidence for the need to distinguish a short-term store from a long-term store.

The current study

The current study is motivated primarily by the finding that when it comes to phonological similarity and word length effects in forward recall, it is not possible to distinguish between short-term memory, working memory or delayed recall tasks. The above review suggests that the pattern may be different with backward recall. With respect to word length, Cowan’s results lead to the conclusion that differences in word length effects can be observed between immediate recall and long-term memory tasks. Word length effects on backward recall of complex span tasks remains untested. With respect to phonological similarity effects, the similarity decrement can be observed with immediate recall, but given Li and Lewandowsky’s results when interlist distraction is employed as is the case with complex span and delayed memory tests, the potential is there for the similarity effects to be attenuated.

Experiment 1

In the first experiment word length effects were examined in three tasks that required the backward recall of four-word lists. The simple version of the task involved the recall of the four words in reverse order immediately after presentation. The complex span version of the task was similar to that used by Tehan et al. (2001, Experiment 1B) in that four maths equations that required shadowing were interleaved with four words on each trial. In the delayed task the four words (p.156) were presented and followed by 12 seconds of distractor activity prior to recall. If we are to replicate Cowan et al., word length effects should be present on the immediate test, they should be attenuated or reversed on the delayed memory task. Given that the complex span task is a mini version of the continuous distractor task used by Cowan et al., one might expect that word length effect would be attenuated or reversed.

Method

Participants. Twenty adults volunteered to participate in the experiment. All participants were tested individually in sessions averaging 30 minutes’ duration.

Materials. The long and short words used in this experiment were taken from the MRC linguistic database (Coltheart 1981). A pool of 96 words made up of seven phonemes (two and three-syllable words) and 96 words made up of three phonemes (one-syllable words) was selected. The word pools were matched for word frequency using the Kucera and Francis (1967) norms. The mean frequency of the short words was 8.22 words per million (SD = 8.14) and that of the long words was 7.79 words per million (SD = 10.43). The words were also matched for concreteness. The mean for short words was 564.86 (SD = 64.84) and the mean for long words was 550.40 (SD = 52.74). The word pool was supplemented by a pool of 128 maths equations constructed following the procedures used by LaPointe and Engle (1990). They consisted of either a product or dividend, followed by either a simple addition or subtraction, then an answer that was either true or false (e.g., 10/2 −1 = 4). Half the problems had correct answers and half had incorrect answers. These were randomly assigned to blocks of four for each trial excluding the immediate serial recall trials.

Participants studied 48 trials and on each trial were requested to remember the four words in reverse order. For 24 of the trials the four words were randomly sampled from the pool of short words and for the other 24 trials they were randomly sampled from the long word pool. Eight each of the short and long lists were randomly allocated to the immediate recall condition (e.g., thong dew maze vine), eight each were randomly allocated to the operations span condition and eight each were randomly allocated to the delayed recall condition. Each trial was created by following the procedures used by Tehan et al. (2001). To create each trial for the operations span task, four maths equations were randomly selected from the pool and these were interleaved with the four words (e.g., stair: (12/4) +1 = 4 peg: (2 × 5) + 3 = 13 bib: (8/4) + 3 = 4 ham: (8/4) + 3 = 4. To create each delayed trial, four maths equations were randomly selected from the pool and placed after the four words (e.g., keg lung babe lute (2 × 5) −3 = 7 (3 × 4) −2 = 10 (2 × 4) −5 = 4 (8/2) − 1 = 1. The order of the 48 trials was then randomized. This procedure was conducted for each participant and 20 different sets of trials were generated.

Procedure. Participants were given printed instructions detailing the different tasks and the requirement to recall the four words verbally in reverse order They then completed standard consent forms and practice trials for each type of task. Participants’ questions were answered and they were seated in front of a computer with a monochrome screen. Each trial started with an auditory beep and ended with a row of question marks. The stimulus items were presented one at a time in black lowercase type in the centre of the screen. When the row of question marks appeared on the screen the participants were requested to recall the four words in reverse order and to say “something” or “pass” if they could not recall a word, in order to preserve the serial order of the remaining recalled words. In all trials participants had 15 seconds to recall the four items in reverse order before the next trial began.

Immediate recall task. The immediate recall lists contained word stimuli presented at a rate of one word per second. Participants were instructed to read each word aloud as it appeared (p.157) on the screen and to recall the four words in reverse order immediately after the row of question marks.

Delayed recall task. The delayed lists contained words that were again presented at a rate of one word per second. The four words were immediately followed by four maths equations presented at a rate of one equation every three seconds. Participants were instructed to read each word and the digits of each maths equation aloud as they appeared on the screen. Participants did not have to solve the maths problems: All they had to do was to read the four or five digits involved in the problem. Verbal shadowing is a typical means of distractor activity in the Brown–Peterson task, upon which the current task is modelled. A row of question marks again appeared three seconds after presentation of the final equation as a cue to commence reverse recall.

Operations span task. On the operations span trials, a word and a maths equation appeared simultaneously on the screen. The four word-equation pairs were presented at a rate of one pair every four seconds. Participants were instructed to read each word and the digits in the corresponding maths equation aloud as they appeared on the screen. The recall cue was presented four seconds after the final list item. The request to read the digits, rather than solve the maths problems, deviates from the typical working memory task where solution is often requested. Tehan et al. (2001) demonstrated that shadowing led to exactly the same outcomes as solving the problems, which is not all that surprising given that both are designed to prevent the rehearsal of the to-be-remembered items.

Results

Scoring. Recall performance was scored in three ways. A response was scored as correct if the correct word was reported in the appropriate serial position. This is the traditional measure used in short-term memory experiments. Secondly, an item was scored correct if it appeared somewhere in the recall protocol regardless of what position it was recalled in. This form of scoring has been widely used in complex span tasks (La Pointe and Engle 1990; Tehan et al. 2001). Thirdly, recall accuracy was scored by dividing the number of items recalled correctly in position by the total number of items recalled (Fallon, Groves and Tehan 1999; Poirier and Saint-Aubin 1995). Unless stated, an alpha level of 0.05 was used to determine statistical reliability.

Correct in position. The mean number of words correctly recalled in their appropriate serial position is shown in the upper section of Table 9.1. A 3 × 2 repeated measures ANOVA with word length and type of task as factors revealed a significant main effect for type of task, F (2,38) = 88, Mse = 24.44. Performance on the immediate recall task differed significantly from the operations span task, t (19) = 2.96, and performance on both tasks differed significantly from the delayed recall task, t (19) = 12.98, and t (19) = 9.57, respectively. The main effect of word length and the interaction between word length and task were not significant, F (1,19) = 1.37, Mse = 19.72, and, F (2,38) = 0.40, Mse = 9.37.

Item recall. The mean number of words correctly recalled irrespective of the order they were recalled is shown in the middle panel of Table 9.1. A repeated measures ANOVA again revealed a main effect for task type, F (2,38) = 134.40, Mse = 8.35; performance on the immediate recall task differed significantly from the operations span task, t (19) = 5.88, and performance on both tasks differed significantly from the delayed recall task, t (19) = 13.88, and t (19) = 11.13. The main effect of word length and the interaction between word length and task was not significant, F (1,19) = 4.02, Mse = 10.17, and, F (2,38) = 0.11, Mse = 5.94.

Order accuracy. Order accuracy is shown at the bottom of Table 9.1. A repeated measures ANOVA again revealed a main effect of task type, F (2,38) = 44.81, Mse = 0.03; performance on the immediate recall task did not differ significantly from the operations span task, t (19) = 1.06, (p.158)

Table 9.1 Mean backward serial recall performance (and SD) for short and long words across recall tasks

	Simple span	Task Operation span	Delayed recall
Correct in position
Short	24.70 (5.81)	21.30 (4.64)	10.10 (6.63)
Long	23.30 (4.49)	20.10 (5.20)	9.85 (6.45)
Item recall
Short	29.95 (1.82)	26.60 (2.66)	19.45 (4.88)
Long	28.50 (2.28)	25.65 (2.76)	18.35 (4.52)
Order accuracy
Short	0.82 (0.18)	0.80 (0.13)	0.49 (0.23)
Long	0.81 (0.12)	0.78 (0.15)	0.52 (0.23)