Phonological Similarity 1 Dissimilar Items Benefit From Phonological Similarity in Serial Recall: The Dissimilar Immunity Effect Revisited

In short-term serial recall, similar-sounding items are remembered less well than items that do not sound alike. This phonological similarity effect has been observed with lists composed only of similar items, and also with lists that mix together similar and dissimilar items. An additional consistent finding with mixed lists has been an effect we refer to as “dissimilar immunity”. Dissimilar immunity refers to the finding that ordered recall of dissimilar items is the same whether these items occur in pure lists of dissimilar items or mixed lists. We present three experiments that contradict these previous findings by showing that phonological similarity enhances memory for order of dissimilar items on the same list. Previous studies failed to detect this effect because they did not appropriately examine or control the distribution of different error types. The memory benefit for dissimilar items cannot be accommodated by several current models but is compatible with a dynamic network theory of memory and a distinctiveness account of serial recall. Phonological Similarity 3 Dissimilar Items Benefit From Phonological Similarity in Serial Recall: The Dissimilar Immunity Effect Revisited The standard phonological similarity effect in serial recall refers to the well-replicated finding that lists composed of similar-sounding items are less accurately recalled than lists in which items do not sound alike (e.g., Baddeley, 1966, 1968; Conrad, 1964; Henson, Norris, Page, & Baddeley, 1996; Wickelgren, 1965a, b). This effect is of considerable generality, occuring with lists of letters (Baddeley, 1968) as well as with word lists (Baddeley, 1966; Coltheart, 1993; Henry, 1991), and it has had unparalleled theoretical impact, being considered one of the most crucial features of short-term serial recall (Farrell, 2001). The phonological similarity effect also arises when phonologically similar (e.g., B, P, T) and dissimilar (e.g., K, Q, R) items are presented together in a mixed list. Mixed lists also give rise to an additional, highly diagnostic finding: The recall of dissimilar items in mixed lists is not affected by the presence of similar items. That is, the recall of dissimilar items in mixed lists is the same as that of dissimilar items when they appear in pure dissimilar lists (e.g., Baddeley, 1968; Bjork & Healy, 1974; Henson et al., 1996). This immunity of dissimilar items to the list environment has led numerous theorists to propose that serial recall involves two independent stages of processing, with order errors occurring between positional tokens in a primary stage, and with item-based similarity confusions occurring in a separate secondary stage (e.g., Burgess & Hitch, 1999; Henson, 1998; Page & Norris, 1998a, b). This article critically re-examines the mixed list effect and reports three experiments that show that—contrary to previous reports—phonological similarity enhances memory for order of dissimilar items on the same list. Previous studies have failed to detect this effect because they did not appropriately examine or control the distribution of different types of recall errors, in particular the trade-off between transposition and item (intrusion and omission) errors. When Phonological Similarity 4 item errors are left uncontrolled, the enhanced memory for dissimilar items is expressed in the transposition rates (Experiment 1). When item errors are eliminated by using a reconstruction task (Experiment 2) or by controlling guessing strategies (Experiment 3), absolute serial recall of dissimilar items is boosted. These results are shown to be compatible with the predictions of two novel theoretical views of serial order memory: a dynamic network model that is based on similarity sensitive encoding (Farrell & Lewandowsky, in press) and a computational distinctiveness account (Brown & Chater, 2001; Brown, Neath, & Chater, 2002). Mixed List Phonological Similarity in Serial Recall Baddeley (1968, Experiment 5) published the first study that used lists in which dissimilar items were interleaved with similar items. In his experiment, similar items were similar to each other but not to the dissimilar items, which in turn were dissimilar to each other. For example, a list in Baddeley’s (1868) study might have contained the letters M V K D R T, where V, D, and T are phonologically similar, and M, K, and R each bear little phonological similarity to any other item on the list. This mixed list would be coded abstractly as DSDSDS, where S and D represent, respectively, a similar and dissimilar item. Previous research on mixed-lists has used a variety of list types, typically including the cases SSSSSS (pure similar), DDDDDD (pure dissimilar), and the alternating cases DSDSDS and SDSDSD (e.g., Baddeley, 1968; Henson et al., 1996; but see Bjork & Healy, 1974). The primary theoretical impetus for mixed-list research has been the assessment of chaining models of serial recall, which posit that lists are represented as a chain of items, each pair of items being joined by an association (e.g., Lewandowsky & Murdock, 1989; Wickelgren, 1965b). Serial recall in chaining models proceeds by using each recalled item to cue for its successor. Computational instantiations of chaining, such as TODAM (Lewandowsky & Murdock, 1989), furthermore specify that an item will be elicited to the extent that a cue is similar to the stimulus Phonological Similarity 5 with which the response was originally associated. One implication of this mechanism is that in the case of mixed lists, cueing with any of the similar items should lead to interference on the following dissimilar items. This in turn leads to the prediction that recall of the dissimilar items should be adversely affected by the presence of similar items, in comparison to a pure list consisting only of dissimilar items (see Baddeley, Papagno, & Norris, 1991, for a computational demonstration of the predictions of TODAM). This prediction by chaining models was first called into question by Baddeley’s (1968) study, which found that the accuracy with which D items were recalled from pure lists (DDDDDD) was identical to that of D items in the corresponding serial positions from mixed lists (e.g., SDSDSD). However, one limitation of the Baddeley study was its use of unconditionalised serial position curves, whose diagnosticity for chaining models is limited because they do not consider whether recall of a given item was preceded by a correct or incorrect retrieval. This limitation was overcome in the study by Henson et al. (1996, Experiments 2 and 3), who used conditionalised serial position curves where performance was scored up to the first error. Even with this refined measure, Henson et al. (1996) replicated Baddeley’s finding of a non-effect of similar items on surrounding dissimilar items. This immunity of dissimilar items to the surrounding list environment is now considered a benchmark result in serial recall, and has contributed to the demise of chaining accounts. The finding has posed a considerable challenge to non-chaining theorists as well, and in handling the mixed-list dissimilar immunity effect, several models have made common assumptions about the interaction between phonological similarity and mechanisms for serial order. Theoretical Accounts of The Mixed List Effect The Primacy model (Page & Norris, 1998a, b) accounts for mixed-list dissimilar immunity through addition of an extra “confusion” stage, where confusion of items occurs based only on Phonological Similarity 6 their phonological similarity. Thus, the first stage of the model is thought to store items and their order without regard to their phonological similarity. The output of the first stage is passed to the second stage, where phonological confusions are assumed to take place. Critically, these confusions only occur between similar items, such that items dissimilar to other list items will pass through the second stage unaffected. Similar items, by contrast, will be confused during the second stage, resulting in more frequent transpositions among these items, but without disruption of the recall of dissimilar items. Henson’s (1998) Start-End Model (SEM) accounts for mixed-list effects in a similar manner, by assuming a separate stage of phonological confusions after normal retrieval operation of the model (see Henson, 1998, for details). The Burgess and Hitch (1999) model also accounts for the mixed-list dissimilar immunity effect through multiple stages of item selection and output. In their model, item selection, which is driven by a representation of encoded context, is unaffected by phonological similarity. (This stage loosely corresponds to the first stage in the Primacy model and SEM.) Similarity effects instead arise due to phonemic feedback. The model contains two layers of phonemes, one at input and one at output, that are mutually interconnected. When the output phonemes are cued by a candidate item selected in the first stage, feedback is unavoidably sent to the input phonemic layer, which in turn “... increases the likelihood of a (nearby) similar item being recalled in its place” (Burgess & Hitch, 1999, p. 569). In summary, to accommodate the mixed-list dissimilar immunity effect, these three major current models postulate two (or more) stages of processing, with order errors occurring between positional tokens at an early stage, and with item-based similarity confusions occurring in a separate, later stage. The order representations involved in the early stage are not affected by phonological confusability. The tight and seemingly inseparable link between the empirical finding of dissimilar immunity and the theoretical need for separate stages was underscored by Phonological Similarity 7 Page and Norris (1998b), who reported that explorations of several single-stage models to handle the mixed-list results were unsuccessful. Page and Norris therefore concluded that “... the data alone appear to force us to accept a two-stage model...” (p. 243). 1 It follows that any empirical disconfirmation of the dissimilar immunity effect would have considerable theoretical significance. Specifically, should it be shown that dissimilar items are not always immune to the similarity of surrounding list items, the principal empirical support underpinning the separate confusion stage postulated by the above models would be removed. Despite the seemingly clear empirical picture, there are several reasons why a further exploration of the dissimilar immunity effect is warranted. First, data analysis in previous studies focused primarily on the serial position curves, perhaps at the expense of a more detailed analysis of error patterns and the possible trade-off between different types of errors (e.g., Baddeley, 1968; Henson et al., 1996). Specifically, although Henson et al. (1996) reported transposition gradients for one of their experiments, they did not break down intrusion and omission rates between different list types although those errors together accounted for up to 23% of all responses (their Experiment 2). Second, the immunity of dissimilar items may be tied to the relatively weak similarity manipulations used in research to date. For example, the mixed lists used by Henson et al. (1996) contained an equal number of similar and dissimilar items, which may have limited the potential distinctiveness of dissimilar items (Newman & Jennette, 1975). Perhaps the immunity of dissimilar items can be shattered by a stronger similarity manipulation, for example when only a single dissimilar item is embedded in a list of similar items. Indeed, there is much evidence in other areas of memory research to suggest that a single item that is dissimilar to the remaining homogeneous list items should be recalled more accurately. Phonological Similarity 8 The Isolation Effect The isolation effect—often called the “von Restorff effect” after its initial investigator— refers to the ubiquitous finding that free recall of an item is facilitated if it is dissimilar from a homogeneous set of surrounding list items. For example, the word tiger (referred to as the “isolate”) will be remembered better if it is presented in a list of vegetables than if it is presented in a list of totally unrelated items (for a review see Hunt, 1995; Wallace, 1965). Translated to the current context, the isolation effect implies that a single dissimilar item on a list of similar items should be recalled better than an item in the same serial position on a pure dissimilar list. Thus, the isolation effect lies in empirical opposition to the dissimilar immunity effect. Research on the isolation effect has typically involved free recall tests (e.g., Dunlosky, Hunt, & Clark, 2000; Fabiani & Donchin, 1995; Hirshman & Jackson, 1997; Winters & Hoats, 1989). Only a few studies have been directed towards order memory, these investigations using dimensions other than phonological similarity to isolate items. For example, isolation effects have been found in a serial judgement task (Cimbalo, Nowak, & Soderstrom, 1981; Lippman, 1980) and for the serial learning paradigm (Bone & Goulet, 1968). Of particular interest is the fact that isolation effects have also been observed with retrieval tasks that are considered relatively “pure” measures of order memory: Lippman and Lippman (1978) found enhanced re-ordering of an item isolated by colour on a reconstruction task; similar effects were reported by Kelley and Nairne (2001), also with a reconstruction task. In a reconstruction task, participants are provided with the identity of all list items at recall, and the task is to re-arrange them into the order in which they were presented at study. This task is considered a relatively pure measure of order memory because people do not have to remember the identity of items, only their positions. These related precedents call for a further examination of possible phonological isolation effects in serial recall. Demonstrating the existence of a phonological isolation effect would have Phonological Similarity 9 considerable theoretical implications: First, as discussed earlier, such an isolation effect could not be accommodated by the Primacy model, SEM, or the Burgess and Hitch model. Second, the existence of this isolation effect would support two alternative theoretical approaches to serial recall: one based on similarity sensitive encoding and the other one based on distinctiveness. Similarity Sensitive Encoding Virtually all current models of serial recall assume that the quality of information available for retrieval of an item decreases across serial positions. In some models this primacy gradient is incorporated to enhance recall of initial items, thus accounting for the primacy effect (e.g., Brown et al, 2000; Lewandowsky & Murdock, 1989). Other models go further by relying on this assumption as the sole vehicle for ordering of items (e.g., Farrell & Lewandowsky, in press; Page & Norris, 1998a). In those models, the strongest or most active item is reported at each recall attempt, and once an item is recalled, it is suppressed and becomes unavailable for further report. This mechanism, known as “competitive cueing” because items compete for report on the basis of their encoding strengths, permits forward recall of a list via retrieval of the most active item at each step. The ubiquity of the primacy gradients suggests it to be an important facet of serial recall. Indeed, Brown et al. (2000, p. 151) justified the primacy gradient in their model by suggesting that an adaptively rational organism might pay successively less attention to items in a homogeneous stream of information when tasked with predicting environmental changes. However, most models to date have relied on one or two free parameters to reduce encoding strengths across successive items. One exception to this is a process account of the primacy gradient that was recently provided by Farrell and Lewandowsky (in press; see also Farrell, 2001). Their serial-order-in-a-box (SOB) model, named after the Brain-State-In-A-Box (BSB) algorithm (e.g., Anderson, Silverstein, Ritz, & Jones, 1977) on which it is based, endogenously Phonological Similarity 10 generates a primacy gradient by assuming that the storage of earlier items affects the encoding strengths of subsequent items. Specifically, incoming list items are compared to the contents of memory (i.e., the composite matrix of connection weights that represents all previous items), and the encoding strength of each new item is determined by its relationship to the memory contents. In Farrell’s (2001) version of SOB, the encoding strength of an incoming item is inversely related to its similarity to the memory contents. This assumption gives rise to a primacy gradient because, by chance alone, the similarity between an incoming random item and the contents of memory will increase with the number of items already encoded, thus decreasing the encoding strength of the incoming item. Additionally, if an incoming item is similar to one previously presented, its encoding strength will be further reduced compared to the case in which all preceding items are dissimilar. To understand how this similarity sensitive encoding predicts that inter-item similarity will affect the ordering of all items, consider the expected gradients of encoding strengths for two three-item lists: one consisting only of dissimilar items (i.e., DDD) and one that has a dissimilar item placed between two similar items (SDS). The first item is stored with the same strength on both lists, as there are no previous items to which it could be similar. The second item is also stored with the same strength in both cases, as it is equally dissimilar to the first item. However, the third item on the SDS list receives a smaller weight than its sibling on the DDD list, because its similarity to the first item is detected during the pre-encoding comparison to the contents of memory. In combination with competitive cueing at retrieval, this encoding scheme predicts that the isolated dissimilar item will be recalled better than its counterpart on the DDD list. This is because once the first item is recalled and suppressed, and thus removed from the competition of to-be-recalled items at the second output position, it is the relative difference of encoding Phonological Similarity 11 strengths between the second and third item that determines recall performance. Because the third item on the SDS list, having been encoded with less strength, will not offer as much competition as its counterpart on the DDD list, the isolated dissimilar item is predicted to be recalled more accurately. This prediction runs counter to those of the Primacy model, SEM, and the Burgess and Hitch model, thus providing a strong alternative hypothesis. SIMPLE: A Computational Distinctiveness Account of Serial Recall The same alternative hypothesis can be derived from another recent approach to serial recall, albeit for very different reasons. This approach, known as SIMPLE (Brown & Chater, 2001; Brown et al., 2002), is based on a computational implementation of local distinctiveness. According to SIMPLE, items are retrieved on the basis of their distance, in psychological space, from their nearest neighbors. Items that are isolated from their nearest neighbors are recalled better than items in close proximity to others. In the case of mixed lists, psychological space is considered to be two-dimensional, with one dimension representing temporal distance (between encoding and recall of an item), and the other separating items according to their phonological similarity. Given this representation, the possible distances between any two S and D items along the phonological similarity dimension are, by definition, uniformly large. By contrast, the set of possible distances between any two D items will be more variable—some pairs of D items may be closer together along the phonological similarity dimension than others. This representation gives rise to an isolation effect because on mixed lists, each D item is temporally adjacent to two S items—and hence maximally distant from its neighbors in two-dimensional space—whereas on a pure list, the random D items that happen to be temporally adjacent are not guaranteed to be maximally distant in two-dimensional space. Phonological Similarity 12 Because of the core assumption that the proximity of immediate neighbors contributes more to distinctiveness than that of distant items, a D item on a mixed list is expected to be more distinct than the same item on a pure list. Although this prediction is identical to that of SOB, it is obtained for very different reasons: Whereas the similarity-sensitive encoding in SOB differentiates between items at study, the local distinctiveness computation in SIMPLE operates exclusively at retrieval.