COGNITIVE
PSYCHOLOGY
9, 1-30 (1977)
The Translation
Effect in Memory
CHARLES CLIFTON,JR., University
Search
AND PATRICIA SORCE
of Massachusetts,
Amherst
AND DONNA CRUSE Oregon State University
The function relating recognition reaction time to the size of a memorized set of items is steeper when the memorized items and the probe are in two different categories, related by a memorized translation scheme, than when they are in the same category. Experiment 1 demonstrates that this “translation effect” is obtained for both familiar and unfamiliar translation schemes and further demonstrates that the zero-intercepts of the functions are lower when the probe differs from the memorized items in category than when it does not. Experiment 2 demonstrates that the slopes of the functions relating negative RT to memorized set size when probe and set are the same in category are steeper than the slopes of the corresponding positive functions just in case subjects are aware that the probe and set categories may differ. Experiment 3 demonstrates that the translations between memorized items and a probe that differ in category are done during the rehearsal of the memorized set, not after the probe is presented. Arguments are presented that rehearsal strategy determines memory comparison time, presumably through a hypothetical memory strength variable, but that direct-access strength theories that deny a memory scanning process are inadequate to account for the data.
People are able to recognize and evaluate information that is not literally the same as information previously presented to them. We are able to judge the truth of an assertion which is, at best, a paraphrase of a previous assertion. We are able to recognize a visual scene as representing a place we have seen at a different time, under different circumstances, or, perhaps, which we have only heard about. In many ways, we must have the ability to translate, or recode, information from one form to another. The present research examines one manifestation of such an ability. It deals with the recoding of information which can be presumed to This research was supported in part by Grant MH-23939 from the National Institute of Mental Health. The authors gratefully acknowledge the support of the Department of Psychology, Stanford University, in the preparation of this report, thank Sally Gentry for her help in Experiment 3, and thank Jane Perlmutter for her valuable comments on an earlier version of this paper. Reprint requests should be sent to Charles Clifton, Jr., Department of Psychology, University of Massachusetts, Amherst, MA 01002. Copyright 0 1977 by Academic Press, Lnc. All rights of reproduction in any form reserved
2
CLIFTON,
SORCE AND CRUSE
exist in memory in a quite literal, nonabstract, way, and it deals with translation schemes which are arbitrary and applicable only to very limited classes of materials. Other, more powerful, recoding abilities have been considered extensively in recent years, primarily involving linguistic materials (e.g., Anderson & Bower, 1973; Clark & Chase, 1972; Miller, 1962; Schank, 1972). While some very interesting results have been obtained under these approaches, it has been difficult to make clear and precise conclusions about the dynamics of recoding, in part because there is no unequivocal way of specifying the forms of the representations which are recoded one into the other. The intent of the present research was to answer some questions about the nature of the recoding process in a simpler situation in which the recoded representations can be specified with reasonable confidence. The hope underlying the research was that answers to questions about recoding in such a simple situation would generalize to more complex, more interesting, situations. The research to be reported here studied the temporal characteristics of recoding, or translating, between elements of small memorized sets, according to a memorized scheme which maps the superficial form of an element of one set onto the superficial form of an element of the other set. A typical translation scheme is that used by Cruse and Clifton (1973, Experiment 1). Their scheme involved a set of letters and a set of digits, paired according to the list A = 1, B = 2, . . . , I = 9. After learning the translation scheme, subjects memorized a short list of items in one form (e.g., digits), and then were tested for recognition of an item in either the same form (a digit) or the other form (a letter). The proper response for the subject to make was “yes” if the recognition probe matched a memorized item or was a translation of a memorized item and “no” otherwise. Cruse and Clifton assumed that the contents of memory were literal representations of items presented because of the simplicity of materials, the rapid presentation rate of the items to be memorized, the short period of time items were held in memory, and the instructions given to the subjects to rehearse items in the form in which they were presented. In the case where the probe and the presented items did not match in form, the subject would have to translate between them in arriving at a decision. Any increase in decision time required by such translation could provide some information about the process of recoding. Several investigators have reported experiments of the type just described, measuring recognition reaction time (RT). Cruse and Clifton (1973) and Swanson, Johnsen, and Briggs (1972) studied translations between memorized pairs of items, Klatzky and Atkinson (1970) used translation between natural semantic categories and exemplars of those categories, and Chase and Calfee (1969) used translation between
TRANSLATION
IN MEMORY
SEARCH
3
auditorily and visually presented letters. All experimenters reported the same basic result, which will be illustrated by Cruse and Clifton’s (1973) Experiment 1. When the items from the memory set and the probe item presented on a trial were of the same category (all letters or all digits), the task is the character classification task first studied by Sternberg (1966). In this task, RT is found to be a linear increasing function of the number of items in the memorized set, with a slope of approximately 35 msec/item for both positive and negative responses. A simple interpretation of the finding is that subjects serially and exhaustively compare the probe against the memorized items, takingb msec for each comparison. Under this interpretation, RTNT =a +bS,
(1)
where the subscript NT indicates that no translation is needed between set and probe, S is the number of items in the memorized set, and a is the time taken to encode the probe and to select and execute the response (and may differ between positive and negative responses). Cruse and Clifton (1973) considered two primary hypotheses for the case where the memorized set and the probe differed in form. Both hypotheses assumed that a presented set was held in a form very similar to the form in which it was presented, e.g., a set of digits was held as a set of representations of digits. IJnder the “probe translation” hypothesis, a probe which differed in form from a memorized set was claimed to be translated from its own form into the form of the set, and the result of this translation was claimed to be compared against the members of the set. Thus, RT,=a+t+bS,
(2)
where the subscript T indicates that translation between probe and set is needed, and t is the time taken to translate the probe. The other hypothesis considered was the “list translation” hypothesis, under which subjects confronted with a probe which differed in form from the presented items would serially translate each presented item into the form of the probe, taking time t for each translations, and thus, RTT = a’ + (t + b)S,
(3) where a’ > a if time is needed to perform such an operation as calling up a translation program. The results of the Cruse and Clifton experiment clearly denied the probe translation hypothesis and were consistent with the list translation hypothesis. The zero-intercept of the function relating RT to set size (which will be referred to as the RT function) was no higher when translation was required than when it was not, and in fact, was nonsignificantly lower. Instead, the slope of the RT function increased
4
CLIFTON,
SORCE AND CRUSE
from an average of 37 msec/item when translation was not required to 94 msec/item when it was. Cruse and Clifton ruled out some alternatives to the list translation hypothesis, including the hypothesis that the act of translation displaced the memorized set from active memory, requiring that it be retrieved from a more permanent memory in a serial fashion before it could be examined. However, they offered no conclusive arguments against a class of alternative hypotheses which deny that presented items were stored in just their presented form and that any necessary translation was done after the probe was presented. Such hypotheses claim, instead, that the necessary translation was done when the items to be remembered were presented, or while they were being rehearsed, and that the translated or recoded items were held in a form which could be accessed only more slowly than the presented items. Some later research indicated that such hypotheses are, in fact, viable. Clifton, Cruse, and Gutschera (1973) manipulated the rate at which the members of the to-be-remembered set were presented, from 400 to 1600 msec/item. They found that this variation in presentation rate had no effect on the RT functions when no translation was required between set and probe, but that it did affect the functions obtained when translation was required. At the slower rates, these functions looked like those reported by Cruse and Clifton. However, at the 400 msec/item rate, the functions were markedly nonlinear, steeply increasing from one to three presented items, and then flattening out and becoming approximately parallel to the functions obtained when no translation was required as the set size increased from three to five items. Such an effect of presentation rate upon the translation RT functions (but not the no-translation RT functions) indicates that at least some of the recoding occurred at the presentation of the set, rather than at its test. Clifton et al. suggested that, under the rapid presentation conditions, subjects were able to recode only one or two items at presentation and were forced to translate the probe into the form of the presented items when a decision could not be made on the basis of direct comparisons between the probe and items in memory. However, the data were noisy enough to be somewhat inconclusive, and failed to specify the nature of the recoding process conclusively. The major purpose of the experiments to be reported here was to assess the correctness of various accounts of the steepening of the RT functions when translation is required between presented items and a probe (which will be referred to as the “translation effect”), and in particular, to determine whether any necessary translation is done before or after the probe is presented. If translation proves to be done after the presentation of a probe, and if the list-translation hypothesis (Equation 3) proves to be correct, then the magnitude of the translation effect can be used as a very sensitive measure of the time to translate a sinale item. t. The availabilitv of such a measure could be useful in
TRANSLATION
IN MEMORY
SEARCH
5
studying the dynamics of various translation schemes. On the other hand, if translation proves to be done before the presentation of a probe, then a study of the speed with which decisions can be made about presented items vs their translations could yield interesting conclusions about the determinants of the rate at which memory is searched. The first experiment to be reported tested a prediction of the listtranslation hypothesis (as was also done, unsuccessfully, by Clifton, Gutschera, Brewer, & Cruse, Note l), and in so doing, tested whether similar translation effects are obtained for both an arbitrary, recently learned translation scheme and a familiar translation scheme. The second experiment to be reported explored some findings which seemed to be discrepant with the list-translation hypothesis and determined whether the possibility that translation would be required in the recognition task affected the memory search process on trials when translation between the presented items and the probe happened not to be necessary. The third experiment was a direct attack upon the question of when translation takes place, before or after the presentation of a probe. It is reported last because it provides a partial explanation of some puzzling phenomena noted in the earlier experiments, and because it suggests a need for a reconceptualization of the memory search process. EXPERIMENT
1
It was noted earlier that a translation effect has been observed by several researchers, using either arbitrary, recently learned translation schemes or more familiar translation schemes. The numerical size of the translation effect, however, varied greatly among the experiments, and it is possible that quite different phenomena take place when a translation scheme is familiar and when it is unfamiliar. The first experiment was designed to determine whether the magnitude, or possibly even the presence, of the translation scheme is affected by the familiarity of the scheme. Each translation scheme used 20 pairs of first and last names. In the familiar scheme, the pairs were names of famous people (e.g., Bob Hope), well known to the subjects at the time the experiment was conducted, while in the unfamiliar scheme, the pairs were randomly paired first and last names selected from a telephone book and learned by the subjects prior to the experiment. The items presented for memory on a trial were all first names or all last names, and the probe was either a first or a last name. When the memory set consisted of first names and the probe was a last name (or vice versa), translation was required. The names presented on a trial were all names of famous people, in which case translation would involve the familiar scheme, or all fictitious names, in which translation would involve the unfamiliar scheme. If the list-translation hypothesis, Equation 3, is correct, and if the
CLIFTON, SORCE AND CRUSE TABLE 1 MATERIALS USED IN EXPERIMENT 1
Familiar names
Fictitious names
First name
Last name
First name
Clark Jane Archie Pat Bob Janis Spiro Marilyn Abbie Helen
Kent Fonda Bunker Nixon Hope Joplin Agnew Monroe Hoffman Keller
Steve Betsy Lam-i Jim Emma Im Paul Anne Kate Hugo
Last name King Moran Duval Aiken Smith Lepley Dwyer Holt Yolan Burnett
familiar translation scheme is executed more rapidly than the unfamiliar scheme (t is smaller), then a numerically smaller translation effect should be obtained for the familiar than for the unfamiliar translation scheme. Alternatively, it is possible that a subject who is very familiar with the translation pairs will not store the presented items and perform the required translation at the time of test, but will store abstract representations of the pairs whose members occurred in the memorized list. If such abstract representations can be accessed as easily from one member of the pair as from the other, the translation effect should be nonexistent when familiar pairs are used. In this event, whatever conclusions are made about translation in an unfamiliar scheme will not apply directly to familiar translation schemes, such as translations between words in two familiar languages or between word categories and exemplars. Method Subjects. Six students at the University of Massachusetts were tested individually for two practice sessions and eight experimental sessions of approximately 1 hr each. They were paid a minimum of $1.75 per session, plus a bonus which depended upon a score that was increased as a linear function of mean responding speed and decreased 5e for each error. Procedure. Before the experiment, the subjects memorized the two lists of name pairs shown in Table 1. They were tested with each list on alternate days, half the subjects beginning with the list of famous people (the familiar translation scheme) and half beginning with the list of fictitious names (the unfamiliar translation scheme). Each daily session consisted of eight blocks of 32 trials. On each trial, one to four first names or one to four last names were presented at the same location on a computer-controlled video display for 1200 msec each, with a 300-msec blank period between names. Fifteen hundred milliseconds after the end of the last blank period, a 500-Hz tone sounded for 100 msec,
TRANSLATION
IN MEMORY SEARCH
7
and 400 msec later, a single probe item was presented. If the probe was a member of the presented set, or was the first or last name of a “person” whose last or first name was a member of the presented set, the subject indicated a positive response by pulling a lever with his right forefinger. Otherwise, he was to pull a lever with his left forefinger. His RT was measured to the nearest millisecond. Half the probes were first names and half were last names, and half the probes matched the set in category (first or last names, no-translation trials) while half differed in category (translation trials). All conditions were sequenced randomly, with the constraint that each occur once per block. Negative probes and the serial positions of positive probes were selected randomly. If the subject made an error, a ready button lighted, and remained lit until the subject pushed it. The next trial began 1500 msec later. At the end of each block of 32 trials, the subject was given a rest period, and his current earnings, mean RT, and total errors on the trial block were displayed on the video screen. The speed with which subjects could execute the translation scheme was measured before and after the experiment. On each of the two practice days, subjects were given 288 trials before recognition RT testing began, using the list of names being tested on that day. On each trial, either a first name or a last name was presented. If a first name was presented, the subject was to say the last name that went with it, while if a last name was presented, he was simply to pronounce it. After the last day, each subject was also given 288 trials on each list of names, with the task changed so that he said the last name that went with a presented first name, or the first name that went with a presented last name. Naming RT was measured with a computer-controlled voice key. The subject (monitored by the experimenter) signaled the computer when he made an error.
Results Naming time. At the start of the experiment, subjects were faster in correctly saying the last names of famous people when presented with their first names than they were for fictitious people, 625 vs 782 msec, t(5) = 4.8, SD, = 32.7, p < .Ol. There was no difference in time to pronounce presented last names, 560 vs 584 msec, t(5) = ~1, SD, = 44.1. However, the difference in translation speed had disappeared by the end of the experiment. The time to say a last name when presented with a first name was 688 msec for famous people and 709 msec for fictitious people, t(5) = Cl, SD, = 37.3, while the difference was reversed in saying a first name when presented with a last name, 727 vs 710 msec. Recognition pe$ormance. The correct recognition RTs, with RTs less than 100 or greater than 1500 msec eliminated (0.3%), appear in Fig. 1, together with the means of the parameters of straight lines fitted to each subject’s mean RTs using a least-squares criterion. The 95% confidence interval on slopes is 29.46 msec/item, and is r14.2 msec on intercepts. The squared correlation coefficients between set size and mean RT (taken as a measure of the goodness of fit of the linear functions) ranged from .95 to .98, and averaged .97. An analysis of variance performed upon the slopes indicated that the functions were steeper when translation was required than when it was not, 88 vs 39 msec/item, F(1,5) = 15.78, MS, = 3559, p < .025. The interaction between translation requirement and translation difficulty
CLIFTON, SORCE AND CRUSE
700
-
azs.357,‘mu /
l ,/‘,,‘a’
” I
POSITIVE
1 1234
NEGATIVE
SET SIZE
FIG. 1. Reaction time as a function of memorized set size, Experiment 1. Expressions adjacent to lines indicate mean RT as a function of set size, S. F, familiar translation scheme; U, unfamiliar translation scheme.
was significant, F(1,5) = 52.25, MS, = 47, p < .Ol. The translation RT functions were significantly (p < .Ol, t test) steeper on days testing fictitious names than on days testing famous names, 95 vs 81 msec/item, while the no-translation functions were nonsignificantly flatter for fictitious than for famous names, 36 vs 42 msec/item. The RT functions were steeper for negative than for positive responses, F(1,5) = 10.78, MS, = 702, p < .05, and the interaction between translation difficulty and response was significant, F(1,5) = 15.12, MS, = 42, p < .025. The effect of translation difficulty, pooled over conditions in which translation was and was not required, was significant only for positive responses. The interaction between translation requirement and response was nonsignificant, F(1,5) = 2.84, MS, = 329, although the difference between positive and negative RT function slopes reached the .05 level of significance only when translation was not required. The analysis of the zero-intercepts indicated that they were higher for negative than for positive responses, 437 vs 408 msec, F( 1,5) = 14.72, MS, = 1351, p < .025, and that they were higher when translation was not required than when it was, 451 vs 394 msec,F( 1,5) = 8.09, MS, = 9583, p < .05. In addition, the four-way interaction involving all factors in the experiment was significant, F(1,5) = 39.9, MS, = 39, p < .Ol, for reasons which are not clear. The error rates appear in Table 2. When translation is not required, they seem within the usual range reported for memory scanning tasks. When translation is required, they are larger and increase with set size. This positive covariation between error rate and RT indicates that the translation effect in RT is not simply due to a speed-accuracy tradeoff.
TRANSLATION
IN MEMORY TABLE
SEARCH
2
ERRORPERCENTAGES,EXPERIMENT1 Negative Probe
Positive Probe Set Size 1
2
3
4
1
2
3
4
Familiar names No translation Translation
2.5 1.3
1.3 5.9
1.0 8.3
0.8 9.4
2.5 1.5
0.8 3.0
2.8 5.5
3.0 5.7
Unfamiliar names No translation Translation
2.0 2.3
1.5 7.0
1.3 6.5
1.5 8.8
2.1 0.5
1.3 3.9
2.1 4.4
3.0 10.4
Condition
The translation effect was obtained for both the familiar and the unfamiliar translation schemes. The magnitude of the effect was greater for the unfamiliar than for the familiar scheme, which could be taken as evidence in favor of the list-translation hypothesis (Equation 3). There are reasons, however, to be cautious in accepting the listtranslation hypothesis. First, the Clifton, Cruse, and Gutschera (1973) data, described earlier, argue against it. Second, there seem to be some inconsistencies between the effect of familiarity on translation time as measured in the recognition task and in the naming task. The effect of familiarity disappeared in the latter task from the beginning to the end of the experiment, while an analysis of variance with four-day blocks as a factor conducted upon recognition RTs indicated no interactions involving blocks. The effect of familiarity upon the translation effect in recognition was as great at the end of the experiment as at the start. Further, the magnitude of this latter familiarity effect (estimated as the difference in slope between familiar and unfamiliar translation functions) was 15 msec/item in the recognition task, while the difference between accessing familiar and unfamiliar last names in the naming task was 157 msec at the beginning of the experiment. The comparison of these two estimates of translation time is complicated by the fact that the magnitude of the second estimate changed with practice and by the further fact that translation involved preparation of a motor response in the naming task but not in the recognition task. Still, the great disparity between the estimates must give one pause in concluding that they both reflect differences in time to translate between a pair of items, as claimed by the list-translation hypothesis. The third reason to question the list-translation hypothesis involves
10
CLIFTON, SORCE AND CRUSE
some puzzling aspects of the data reported here. The zero-intercepts of the RT functions were significantly lower when translation was required than when it was not. Further, while the interaction between translation requirement and response (positive vs negative) was not significant, the negative function was nearly twice as steep as the positive function when translation was not required (an 8% difference in slope, 51 vs 27 msec/item), while the difference was only 15% when translation was required (94 vs 82 msec/item). These two effects have been obtained, at least in part, in some previous research on the translation effect in memory search (Cruse & Clifton, 1973, Experiment 1; Clifton, Cruse, & Gutschera, 1973; Swanson ef al., 1972; Klatzky & Atkinson, 1970; Klatzky , Juola, & Atkinson, 1971)) but have not previously received serious consideration (and, it must be pointed out, have not always been obtained (Chase & Calfee, 1969; Juola & Atkinson, 1971; Cruse 8z Clifton, 1973, Experiment 2; Clifton et al., Note 1, Experiment 1). Both these findings are somewhat anomalous under the list-translation hypothesis. While the nearly 2:l slope difference between negative and positive response functions suggests a self-terminating scan process when translation is not required, it is difficult to imagine why the scan would be self-terminating in this condition but not in the slower and harder translation condition.’ The slope differences may indicate the need for a more radical revision of the list-translation hypothesis. The fact that zero-intercepts are lower when translation is required than when it is not is similarly puzzling. Under some memory-search hypotheses, including the list-translation hypothesis (Equation 3), the zerointercept estimates the duration of all RT components other than those of translating or comparing the remembered items. Introducing extra translation operations during the comparison process should not reduce the durations of other components of the decision process. The theoretical interpretation of these two effects will be considered further under Conclusions. First, though, Experiment 2 will be reported as an attempt to replicate the effects in question and to assessthe relevance of the slope difference between positive and negative RT functions to 1 An attempt was made to evaluate the possibility of self-termination, using a technique suggested by Stemberg (Note 2). This technique compares the cumulative positive RT distributions for different set sixes. For two set sizes m and n, m
TRANSLATION
IN MEMORY SEARCH
11
the translation effect. While the fact that negative no-translation slopes are steeper than positive no-translation slopes may be related to the possibility that translation is required and may thus serve as a clue as to when and how subjects are doing the necessary translation, it is alternatively possible that some unidentified aspect of the present experimental technique produces nonparallel positive and negative RT functions in a memory search task. Experiment 2 tested this latter possibility by introducing some blocks of trials on which translation was never required. If the slope difference observed in Experiment 1 between positive and negative RT functions is due to the possibility that translation will be required, rather than to some other aspect of the present experimental procedure, then it should disappear on these blocks of trials. EXPERIMENT
2
Method Sub&~ Six students at the University of Massachusetts were tested for two practice sessions and six experimental sessions. They were paid in the same manner as in Experiment 1. Procedure. Subjects learned the list of letter-digit translations A = 1, B = 2, . . . , I = 9 before the first practice session, so that they could fluently translate from letters to digits and vice versa. Each daily session consisted of 18 blocks of 16 trials each. On each trial, one to four digits or one to four letters were presented sequentially on an alphanumeric Nixie tube for 900 msec each, with a 300-msec blank period between items. Fifteen hundred milliseconds after the end of the last blank period, a 500-Hz tone sounded for 100 msec, and 400 msec later a single probe item was presented and remained on until the subject responded. On six of the trial blocks, each probe item was the same in form as the set which it tested. Trials from these blocks will be referred to as “pure” trials, never requiring translation. On the remaining 12 trial blocks, a randomly selected half of the probes were of the same form as the set which they tested (resulting in the “no-translation” trials), while the remaining probes differed in form (“translation” trials). Half of the sets in each type of trial block contained digits and half contained letters, and half of the probes required a positive response and half required a negative response. Subjects were informed of the type of each trial block. They responded and received feedback as in Experiment 1.
Results Figure 2 presents the mean correct RTs, after eliminating all RTs less than 200 msec or greater than 1400msec (O&t%), with the parameters of straight lines fitted to each subject’s mean RTs. The 95% confidence intervals were k5.7 msec/item for slopes, and k15.7 msec for zerointercepts, and the squared correlation coefficients ranged from .98 to 1.O. The only significant effects in an analysis of variance performed upon slopes were translation requirement, F(2,lO) = 15.96, MS, = 774, p < .Ol, and the interaction of this effect with response, F(2,lO) = 27.73,
12
CLIFTON,
SORCE AND CRUSE NEGATIVE
POSITIVE 700 -
ao+374
*, 0
300 .
r L 1 1
I 2
3
I 4
SET SIZE
FIG. 2. Reaction time as a function of memorized set size, Experiment 2. Expressions adjacent to lines indicate mean RT as a function of set size, S. P, pure trial blocks; T, no-translation trials; T, translation trials.
MS, = 58.6, p c .Ol. A Scheffe test indicated that the positive notranslation function (31 msec/item) and the pure functions (39 and 33 msec/item for positive and negative responses) were significantly (p < .05) flatter than the negative no-translation function (59 msec/item) and that all of these functions were significantly flatter than the translation functions (76 and 82 mseclitem for positive and negative responses), while no other differences were significant. The zero-intercepts were higher for the no-translation condition (410 msec) than for the translation or the pure condition (350 and 338 msec), F(2,lO) = 19.04, MS, = 1898, p < .Ol, while the latter two did not differ. The zero intercepts for negative responses were higher than for positive responses, 384 vs 348 msec, F(1,5) = 19.78, MS, = 1203, p < .Ol. The interaction between these two variables was also significant, F(2,lO) = 12.34, MS, = 445,~ < .Ol. The effect of translation upon the intercepts was greater for positive than for negative responses. Finally, the zerointercepts were higher for letter than for digit probes, 377 vs 355 msec, F(1,5) = 24.3, MS, = 356, p < .Ol, but this factor interacted with no other. The error rates appear in Table 3. They vary little among conditions, except in the translation conditions, where they increase with set size. Discussion
The translation effect was again found. Further, the data obtained on blocks of trials on which translation was never required (the pure condition) replicated the basic memory scanning phenomenon reported by
TRANSLATION
13
IN MEMORY SEARCH TABLE 3
ERROR PERCENTAGES, EXPERIMENT 2
Positive Probe
Negative Probe Set Size
Condition
1
2
3
4
1
2
3
4
Pure No translation Translation
1.1 2.2 2.8
2.0 2.0 5.8
2.7 1.8 8.2
3.2 2.5 8.3
0.4 2.5 1.6
1.3 2.1 1.0
0.2 1.6 4.3
1.3 2.5 7.7
Sternberg (1966) and many others; the positive and negative functions were approximately parallel, and the zero-intercept for negative functions was greater than for positive functions. Experiment 2 obtained the findings of Experiment 1 which seemed to pose difficulty for the serial scanning models that have been considered. The negative RT function in the no-translation condition was steeper than the positive function, while no difference was obtained between negative and positive translation functions. Neither was there a difference between positive and negative pure functions, indicating that the difference for no-translation functions resulted from the expectation of translation rather than from some other aspect of the experimental procedure. Comparison of the slope values from the no-translation and the pure conditions indicates that the effect results from a steepening of the negative no-translation function when there is a possibility of translation, and, perhaps, a (nonsignificant) flattening of the positive function. Further, the zero-intercepts of the RT functions in the translation condition were lower than in the no-translation condition, and in fact, did not differ from those in the pure condition (which would be expected to be low because of the greater ease of encoding an item whose category is known; Clifton & Brewer, in press; Marcel & Forrin, 1974). This effect was greater for positive than for negative responses, a difference which was not significant in Experiment 1 although a tendency in the same direction can be noted in Fig. 1. Since these effects seem incompatible with the serial-scanning models considered thus far, in which the translation operations are done at the time of test, the third experiment will explore models which postulate that any necessary translation is done at the time of encoding. Two models will be considered which can account for the translation effect under this claim, and will be contrasted with the list-translation model (Equation 3) considered previously. While the models, as initially developed, do not account for the problematic effects, they do motivate a strong test of
14
CLIFTON,
SORCE AND CRUSE
when translation is done, at input or at test. Some suggestions will be made for how they may be extended to account for the effects considered. EXPERIMENT
3
The list translation hypothesis (Equation 3) basically claims that the steepening of the RT function when translation is required between set and probe is due to extra translation operations performed upon the members of the set in the process of comparing them to the probe. Hypotheses that claim that translation is done when the set is originally presented must explain the steepening of the RT function in terms of a longer time to compare a probe with a translation than with a presented item, or a longer time to access the representation of a translation than that of a presented item. Two specific versions of such hypotheses will be considered. The first will be termed the “dual buffer” hypothesis and appeals to an ability of a subject to access presented items or translations selectively. It claims that a subject stores the members of a memory set upon presentation and also generates and stores their translations in a separate buffer. When a probe is presented, the subject identifies its category, and then selectively compares it against just the appropriate representations. Since the demands of the task emphasize the presented items, their representations are assumed to be stronger or more distinct in memory than their translations, and hence are compared more rapidly against the probe item. The second hypothesis will be called the “long buffer” hypothesis and does not assume selective search. It postulates that the subject stores the presented items and their translations separately, and that the presented items are accessed first, and then their translations. When a probe is presented, it is first compared with the several presented items, and simultaneously, its category is identified and compared with the category of these presented items. If the categories match, a response is made upon the completion of these comparisons. If the probe does not match the presented items in category, then the translations of the presented items are compared against the probe. Under this hypothesis, the RT function is steep when translation is required because more items must be compared with the probe than when no translation is required. Available data on selective search make the long buffer hypothesis more likely. The long buffer hypothesis is a version of a hypothesis used by Naus, Glucksberg, and Omstein (1972) and Naus (1974) to account for the search of semantically categorized lists, and by Clifton and Brewer (in press) to account for the search of mixed lists of presented letters and digits. These authors obtained data which were inconsistent with the selective search posited by the dual buffer hypothesis, but
TRANSLATION
IN MEMORY
SEARCH
15
which were consistent with the partially selective search assumed by the long buffer hypothesis. On the other hand, evidence does exist (Mohs, Wescourt, & Atkinson, 1973) which indicates that selective search may in fact be possible when the classes of items to be searched differ in their status in memory, one in short-term storage and the other in long-term storage, and it is possible that the difference in strength or clarity assumed by the dual buffer hypothesis would encourage selective search of memory in the translation task. The three hypotheses under consideration could be distinguished by an experiment that independently varied the number of presented items and the number of translated items. In the present experiment, N items were presented for memory on each trial, of which some subset M (0 5 M 5 N) had translations. All three hypotheses predict that RT when the probe matches the set in form would be described by RTNT= a+bN,
(4)
which is equivalent to Equation 1. The hypotheses differ in their predictions for RT when translation is required. An unelaborated version of the list translation hypothesis would claim RTT = a’ + (t + b)N,
(W
which is equivalent to Equation 3. An elaboration is possible, in which different amounts of time are required to deal with presented items which have translations, and those which do not: RT,=a’+eN+(t+b)M,
Vb)
in which e is the time needed to examine a presented item to determine whether it has a translation, and in which translation and comparison operations are performed only on those items which do have translations. This elaborated hypothesis will be considered only briefly later, because it proves to make unreasonable assumptions in accounting for the data obtained. The dual buffer hypothesis predicts RT,=a’+cM,
(6)
in which c > b, and the long buffer hypothesis predicts RT,=a’+b(N+M)
(74 or, in an elaborated version in which the translations are claimed to be compared more slowly against the probe than the presented items, RT,=a’+bN+cM,
(7b)
in which c > b. Briefly then, the list translation hypothesis predicts that RT when
16
CLIFTON, SORCE AND CRUSE
translatibn is required will be a function of IV, the dual buffer hypothesis predicts that it will be a function of M, and the long buffer hypothesis predicts that it will be a function of N + M. Method Subjects. Twelve subjects, undergraduates, graduate students, and faculty at the University of Massachusetts, most of whom had had prior experience in memory scanning tasks, were tested for two practice and eight experimental sessions of approximately 1 hr each. Some subjects were paid as in the earlier experiments, while others volunteered their time. Marerials. Sixteen names of common animals and sixteen names of articles of clothing were selected from the Battig and Montague (1969) norms. All words began with a consonant, had one syllable of three to five letters, and appeared at least one time in the KuEera and Francis (1967) table of normative frequency. Each class of items was divided into two subclasses of eight words, approximately matched for normative frequency of occurrence. For six subjects, the words in one subclass of animal names were randomly paired with the words in one subclass of articles of clothing. These subjects learned the paired-associates list thus formed, together with the list of eight unpaired animal names and the list of eight unpaired articles of clothing. The same procedure was followed for the remaining six subjects, except for exchanging the paired and the unpaired words. Each subject was given the materials to learn the day before the experiment began, with the paired-associate list typed twice, once with animals as the left-hand member of the pair and once with articles of clothing, and with each set of unpaired words typed once, and was instructed to learn the materials fluently. Before the first practice session, the experimenter informally tested the subject by presenting both animal names and names of clothing, requiring the subject to name the associate of the presented word or to indicate it had no associate, until he or she was satisfied that the subject knew the lists. Procedure and appnrutus. On each day, each subject was tested individually for 12 blocks of 26 trials each. On each trial, a set of one to four words, all animal names or all articles of clothing, was presented once. The words appeared sequentially at the same place on a computer-controlled video display, and each word appeared for 1200 msec with a 300msec blank period following. Fifteen hundred milliseconds after the end of the blank period after the last word on the list, a lOO-msectone of 500 Hz was sounded, and 400 msec later, a probe word appeared on the same place on the video display and remained until the subject responded. Response requirements and feedback were as in the earlier experiments. There were 104 conditions in the experiment. The presented sets were determined by the combination of number of items presented (N), 1 c N =G4; the number of the presented items which had a memorized paired associate (M), 0 s M G N; and the category of words presented, animals or clothes, yielding a total of 28 types of lists. Probes were determined by the combination of no translation vs translation (matched the set in category or differed in category) and positive vs negative, which would have yielded a total of 112conditions. However, eight of these conditions were impossible (those involving a positive translation probe of a list in which M = 0, that is, in which no presented items had translations), resulting in the 104 conditions. Conditions were randomly sequenced, with the constraint that each occur once in every four trial blocks, and the items presented in each set, their serial position, and the probes were randomly chosen. Subjects were instructed to respond as quickly as possible, while always keeping their error rate below 5%. They were further carefully instructed in how to rehearse the remembered sets. They were told to concentrate their attention upon the actually presented items, rehearsing them so that they were clear in mind, but were told that they could use
TRANSLATION 65Or
r
NEGATIVE
POSITIVE g
17
IN MEMORY SEARCH
*oo750 700 650 600 550 500 450 400f N(items
, 1
2
3
I 4
presented)
FIG. 3. Reaction time as a function of the number of presented items, N, experiment 3. The number of items which had a translation, M, is the parameter. Expressions adjacent to dotted lines indicate mean RT as a function of the number of presented items, N, when all presented items had a translation. any spare time to think of the translations of the presented items. Three subjects reported difficulty in following these procedures, in that they tended to devote nearly equal attention to the presented items and their translations. These subjects yielded data which differed from the data provided by the remaining subjects, but were included in the formal data analysis to be presented. The nature of their deviation will, however, be commented upon under Conclusions.
Results and Discussion Correct RT. Figure 3 presents the mean correct RTs obtained in the experiment, with RTs less than 200 or greater than 1800 msec discarded (0.17%). The data were averaged over animal vs clothing lists, which had no interesting effects. The first aspect of Fig. 3 to examine is the dotted lines, which connect the points where N = M and which amount to a replication of the translation effect demonstrated in the earlier experimerits. The parameters of these dotted lines appear in Fig. 3 and show that the translation RT functions were steeper and had lower zero-intercepts than the no-translation functions. In addition, the RT function when
18
CLIFTON, SORCE AND CRUSE TABLE 4 SLOPES OF STRAIGHT LINES RELATING RT TOM,
EXPERIMENT 3
M
Condition
0
1
2
3
Positive responses No translation Translation
39.2 -
31.7 34.8
27.2 46.0
24.2 34.1
Negative responses No translation Translation
41.3 43.1
46.2 39.0
48.7 58.1
19.6 27.5
no translation is required was steeper for negative than for positive responses. The solid lines in Fig. 3 indicate RT as a function of N, with M as a parameter. Examination of these functions indicates that they are much more consistent with the long buffer hypothesis (Equation 7b) than with the other hypotheses. Reaction time in the no-translation conditions was largely a function of N, not of M, except in the special M = 0 case, which had a lower zero-intercept. (There are some small additional effects of M which will be discussed later.) Reaction time when translation is required, however, seems to be a function of both N and M. The data from the M = 0 case in the translation, negative response condition further support the long buffer hypothesis over the dual buffer hypothesis. This RT function had approximately the same slopes as the remaining functions in its panel, 43.1 vs 41.5 msec/item. The responses which contribute to this function could logically be made without comparing the probe with the individual items in the memorized set, but simply by noting that the probe differed in category from any possible positive probe. A subject who performed a selective search of memory as assumed by the dual buffer hypothesis should have been able to use such a decision strategy, producing a RT function with zero slope. The RT data (excluding the M = 0 case) were statistically analyzed in several ways. First, slopes of the functions relating RT to N were calculated for each subject in each condition, separately for each value of M (M = 1,2, or 3). These slopes were subjected to an analysis of variance with set category, M, translation requirement, response, and subjects as factors. The mean values of these slopes, together with the slopes for M = 0, appear in Table 4. The analysis indicated that two higher-order interactions involving memory set category were significant, but uninterpretable at present and that the effect of M was marginally significant, F(2,22) = 4.11, MS, = 2065, p < .05. In this effect, the slopes of the
TRANSLATION
19
IN MEMORY SEARCH TABLE 5
PARAMETERS OF EQUATION TB, FITTED TO MEAN RTs, EXPERIMENT 3
Parameter Condition
a’
b
C
Positive responses No translation Translation
438.8 306.1
29.9 37.6
-9.6 58.1
Negative responses No translation Translation
407.1 374.4
45.0 43.4
8.8 60.9
Note. The 95% confidence intervals are k29.3 msec, 9.0 msec/item, and 9.9 msec/item for parameters n’, b, and c, respectively.
function relating RT to N were steepest when M = 2, next steepest when M = 1, and flattest when M = 3. Although some complex accounts of this ordering could be offered, its nonmonotonicity together with the marginal level of significance of the effect indicates that one might be justified in concluding that the effect of N upon RT is largely independent of the value of M. To illustrate that M does have an effect in the translation conditions, Equation 7b was fitted to each subject’s data (IV = 1 to 4, M = 1 to N) using a least-squares criterion (RMSD = 11 msec). Estimates of a’, b, and c for each combination of translation requirement and response were evaluated in separate analyses. The means of the three parameters appear in Table 5. The value of u’, which can be interpreted as the N = 0, M = 0 intercept and thus as the sum of encoding and response times, was lower when translation was required than when it was not, 340 vs 423 msec, F(l,ll) = 13.86, MS, = 5923, p < .Ol, and this effect was larger for positive than for negative responses, F(l,ll) = 14.15, MS, = 2122, p < .Ol. The value of b, which can be interpreted as the linear effect of N upon RT, was greater for negative than for positive responses, 44.2 vs 33.8 msec/item, F(l,ll) = 9.56, MS, = 136,~ < .05. The effect of translation requirement yielded an F of less than 1. Finally, the analysis of c, the linear effect of M upon slope, resulted in a significant effect only of translation requirement, F( 1,11) = 36.1, MS, = 730, p < .Ol. The effect of M was large when translation was required, 59.5 msec/item (substantially larger, incidentally, than the effect of N), but averaged only -.4 msec/item when no translation was required. No other effects were significant in these analyses. The analyses thus far support the long buffer hypothesis as a model
20
CLIFTON. SORCE AND CRUSE
for the translation task. The number of presented items, N, has an equal effect when translation is and is not required; the number of translations, M, has an effect only when translation is required; and the effects of N and M are approximately additive.2 The only inconsistent result noted thus far is that the N = 0, M = 0 intercept is lower when translation is required than when it is not, especially when the correct response is positive. One must ask, though, what happened to the deviant steepness of the negative no-translation functions observed in the earlier experiments and in the A4 = N conditions of Experiment 3. In the analysis of b, the effect of N (independent of M) was greater for negative than for positive responses, 44.2 vs 33.7 msec/item, but this effect is too small to account fully for the previously found steepness of the negative no-translation function. In addition, the effect of response upon b does not differ between translation and no-translation trials. Close examination of the effect of M upon RT when translation is not required provides a partial answer. As seen in Fig. 3, positive RT actually decreases as M increases. This is reflected in the numerically negative value of c, -9.6 msec/item, in the positive no-translation condition. A f test conducted upon c indicates that it was significantly less than zero, t(l1) = 7.23, SD,,, = 1.32, p < .OOl, and in fact, all 12 subjects had negative values of c in this condition. On the other hand, negative no-translation RT tended to increase with M. The value of c in this condition was 8.8 msec/item, which was not significantly greater than zero, t(l1) = 1.98, SD, = 4.45, .lO >p > 05, but was significantly greater than the value of c in the positive no-translation condition, t(l1) = 4.22, SD, = 4.36, p < .Ol. The analysis above, together with the assumption that translation is done when a list is memorized, leads one to a most peculiar conclusion: Positive no-translation RT decreases as the memory load imposed by potential translation probes (M) increases. A further post hoc analysis of the data which more fully examines the effect of M in the no-translation conditions shows that this conclusion is unwarranted.3 In this experiment, a positive set consisted of two types of items: items which have potential translations (M in number), and items which do not have potential translations (N -M). Since positive no-translation probes were selected 2 Equation 5b, the liberalized list-translation model, can also be fit to the data and provides the same fit to the translation functions as Equation 7b does. However, in doing SO, the value of (t + b) must be estimated as 59.5 msec/item, the previous estimate for c. Since it is reasonable to assume that b in the sum ( f + b) is the same as b estimated in the no-ttanslation functions, 37.5 msec/item, t must approximate 22 msec. It is not likely that translation between the memorized paired-associates can take place in such a short time (cf. the naming time data of Experiment 1). 3 The authors would like to thank Jane Perlmutter and E. E. Smith for independently suggesting this analysis.
TRANSLATION
1
2
3 N ( ITEMS
21
IN MEMORY SEARCH
4
1
2
3
4
PRESENTED)
FIG. 4. Reaction time on no-translation trials as a function of the number of items presented, separated by probes which had and which did not have a potential translation, Experiment 3.
randomly from the presented items, the probability that such a probe would be an item which had a potential translation increased as A4 increased. Thus, A4 was partially confounded with type of positive notranslation probe, so that any difference in RT to probes which had and did not have potential translations could appear as an artifactual effect of M. In particular, if positive responses to no-translation probes were faster when the probe had a potential translation than when it did not, mean positive RT would decrease as A4 increases. For negative no-translation responses also, probe type and M were partially confounded, although to a lesser extent than and in the opposite direction from the confounding for positive responses. To escape these confoundings, the data from the no-translation conditions were further analyzed by separating them into RTs to probes which had a potential translation and probes which did not. The mean RTs to these reclassified data appear in Fig. 4, plotted as a function of N with A4 as a parameter. Several remarkable points can be noted. First, the positive functions, except for the A4 = 0 function, were very flat, averaging 13.6 msec/item in slope. Second, positive RTs were much faster when a probe had a potential translation than when it did not. Third, no such effect was obtained for negative RTs. However, negative RT did increase with A4 in the case where the probe had a potential translation, an effect which is similar to that observed for translation probes in the main analysis.
22
CLIFTON,
SORCE AND CRUSE
Analyses of variance performed on estimates of a ‘, b , and c for each subject in each condition for these reclassified data confirm the observations just made. The analysis of a’ revealed a significant interaction between response and presence vs absence of a potential translation, F(1,ll) = 18.57, MS, = 2261, p < .Ol (and signilkant main effects of response and the presence of a potential translation). Zero-intercepts were higher for positive probes which had no potential translation than they were for the other conditions. The analysis of b revealed that the mean slope for negative responses, 42.6 msec/item, was greater than the mean slope for positive responses, 13.6 msec/item, F(l,ll) = 39.57, MS, = 511, p c .Ol. The analysis of c indicated a significant interaction between response and potential translation, F(l,ll) = 8.33, MS, = 648, p < .Ol (and a significant main effect of the presence of a potential translation), such that the value of c did not differ significantly from zero for the positive conditions or for negative probes without a potential translation, while it was greater than zero for negative probes which had a potential translation. The value of c in this latter case, 32.3 msec/ item, was smaller than the value of c observed when translation between probe and list was required, 60.9 msec/item. The post hoc analysis thus reveals that positive RT to no-translation probes was not affected by the number of presented items which had translations. Positive RT, however, was faster to probes which had a translation than to probes which did not, and this effect was independent of the number of items presented for memory. Negative no-translation RT did increase with the number of presented items with translations, just in case the probe was an item which had a potential translation. Finally, the effect of the total number of presented items was greater upon negative than upon positive RT. Error rates. Percentages of errors are presented in Table 6. The overall experimental error rate was 3.78%. Error rate was generally related in a positive fashion to correct RT, typically being greater when translation was required than when it was not, and increasing with both N and M when translation was required. In addition, error rate was quite high when positive no-translation probes which had no potential translation were presented. Since these conditions also had high RTs, neither the effect upon RT or error rate can be attributed in a simple way to a speed-accuracy tradeoff. It appears that both measures reflect some source of difficulty in dealing with such probes. Although the number of observations of error RT is too small to yield very stable data, it is interesting to note that the mean RT of “miss” errors to positive no-translation probes which had no potential translation (548 msec) was not much less than the mean RT of correct responses in that condition (562 msec) or of correct responses in the negative no-translation condition when the probe had no potential translation (539 msec), while the mean RT of “false
TRANSLATION
IN MEMORY TABLE
23
SEARCH
6
PERCENTAGES OF ERRORS,EXPERIMENT3 Negative probe
Positive probe
M N
0
1
2
3
4
0
1
2
3
4
No translation, probe has no potential translation 0.7 2.6 2.9 3.3
9.9 11.9 7.6
13.6
11.0
8.7
2.3 0.5 0.9
0.4 0 0.5
0.4 1.3
1.2
1.1
3.8
1.9 3.0
0.4
2.2 1.5 2.9
3.9 7.3
8.1
3.2 4.1 4.1
6.2 8.7
12.9
No translation, probe has a potential translation 1.5 1.9 3.2 2.0
8.0
1.0
2.1
1.4 2.9
1.4 3.1
0.9 3.3
1.2
Translation 4.3 2.7 5.0 4.8
6.7 8.7 9.0
9.7 9.4
8.1
0.9 0.2 0.7 0.7
1.8 1.8 2.7 3.9
Note. Empty cells represent conditions that could not occur in the experiment.
alarms” to negative no-translation probes which had no potential translation did seem to be less (453 msec). Thus, the tendency to respond slowly and with a high error rate to positive no-translation probes which had no potential translation is not simply due to a response bias to make fast negative responses to them. CONCLUSIONS
Some conclusions are clear. The translation effect in memory search holds for a variety of translation schemes. The apparent speed with which memory is searched decreases when translation, or recoding, is required between presented items and a probe item, whether the translation scheme involves letters and digits, arbitrarily paired words of different classes, or well-known names. The translation effect, however, does not arise because of translation operations performed after the probe is presented. Rather, the necessary translations are done when the to-be-remembered items are presented or rehearsed. Experiment 3 provides the primary evidence for this conclusion. Here, RT to a probe
24
CLIFTON,SORCEANDCRUSE
which differed in category from the presented items was affected in an additive fashion by the number of items presented and the number of these items which had translations. It appeared that the memory set which actually determined such translation RTs was the combined set of the presented items and their translations, a set which could only have been constructed by a subject who performed all necessary translations before the probe was presented. However, to a first approximation, RT to a probe that matched the presented items in category was affected only by the number of presented items, not by the number of their translations. Here, the effective memory set comprised just the presented items. A “long buffer” model was proposed (Equation 7b) for these findings. It claimed that subjects in fact stored the presented items plus their translations, and when a probe was presented, compared it first with the presented items and then, if it differed in category from them, with their translations. Several phenomena in the data, however, were inconsistent with such a model. First, the zero-intercepts of the translation RT functions were lower than those of the no-translation RT functions. Second, the slope of the RT function was greater for negative than for positive no-translation probes. Third, Experiment 3 made it clear that the steepness of the negative no-translation functions can be attributed to the existence of potential translations of presented items. Reaction time to no-translation probes which had potential translations was affected by the number of presented items which had translations, while RT to negative probes which had no potential translations was not. Fourth, the number of translations presumed to be in memory did not influence RT to positive no-translation probes. However, positive no-translation RT was slower to probes which had no potential translation than to probes which did have one, and this effect was additive with effects of memory set size. Fifth, and finally, the effect of the number of presented items on positive no-translation RT was very small in Experiment 3. Rehearsal Strategies The difference in time to respond to probes which do or do not require translation can be attributed to differences in how presented items and their translations are encoded and rehearsed. Subjects in the present experiments were instructed to concentrate their rehearsal upon the presented items rather than their translations, and they generally found it natural to do so because of the greater availability of the presented items. It is reasonable to suggest that items which are rehearsed more are, perhaps because of their greater strength in memory, searched before and more rapidly than items which are rehearsed less.
TRANSLATION
IN MEMORY
SEARCH
25
Some incidental data from the present experiments indicate the importance of rehearsal strategies as determiners of recognition RT. Three subjects in Experiment 3 failed to show the translation effect, but instead yielded parallel RT functions for translation and no-translation probes. Before examinipg their RT data, it had been determined that they failed to follow instructions but devoted equal amounts of rehearsal to the presented items and their translations. One subject rehearsed both types of items overtly, an equal amount, and two subjects reported that they were unable to follow instructions and subvocally rehearsed all items equally. Two of these three subjects had been subjects (and experimenters!) in other experiments on the translation effect, and examination of their data in these other experiments indicated that they failed to show the translation effect there, too. Three additional subjects have been observed in other, unreported, translation experiments to devote equal rehearsal to presented items and to their translations, and none of these subjects showed the translation effect. The translation effect thus seems to require differential rehearsal of presented and translated items. Some of the phenomena which seemed inconsistent with the long buffer hypothesis, Equation 7b, may be understood in terms of rehearsal strategies. Consider the fact that the translation zero-intercepts were consistently lower than the no-translation intercepts. One common report by subjects was that they sometimes rehearsed presented items and their translations nearly equally when one, or perhaps two, items were presented, even though they followed instructions and rehearsed primarily the presented items for larger set sizes. If there was a shift from rehearsing all items equally to concentrating rehearsal upon the presented items as set size increased beyond one, and if RT to items of a class depends upon how much that class of items had been rehearsed, then the relative advantage of presented items and relative disadvantage of their translations would increase as set size increased beyond one. In effect, the RT functions would pivot around a point near set size one, the no-translation function downward and the translation function upward, which would increase the zero-intercept of the no-translation function and decrease that of the translation function. Consider, also, the fact that positive RT was faster to presented items with potential translations than without. Some subjects in Experiment 3 indicated that they may have concentrated their rehearsal upon the former items, which in turn may have speeded RT to them. (It is not obvious, however, that the speculation could be developed to account for the fact that the positive no-translation functions for probes with and without potential translations had similar, small, slopes.) Further, the effect of familiarity of the translation scheme upon the slope of the translation RT function (Experiment 1) may be understood
26
CLIFTON,
SORCE AND CRUSE
in terms of the greater ease of executing the familiar scheme, facilitating the generation and rehearsal of the familiar translated names. Memory Strength and Memory Search These speculation about the importance of rehearsal strategies rest critically upon the notion that differences in amount of rehearsal result in differences in decision time, presumably mediated by differences in “memory strength.” Some arguments have been advanced against the suggestion that memory strength affects decision time. For instance, Stemberg (1966) found approximately equal apparent memory search rates for lists that were presented and tested just once and lists that were tested 180 times. However, other theorists (e.g., Baddeley & Ecob, 1973; Cavanagh, 1976; Nickerson, 1972; cf. Stemberg, 1975) have presented arguments for theories of recognition reaction time in which the relevant memory strength is less a quantity that accumulates with rehearsal and testing than a variable which represents the proportion of some fixed memory or attentional capacity which is devoted to a given item in memory. These memory strength theories propose that a recognition probe is not compared against the representations of the set of items presented for memory, but instead, that its memory strength is assessed against some decision criterion (cf. Clifton & Cruse, in press, for discussion of some alternative decision rules). Is such a direct access theory able to account for the present data on the translation effect? We think that some of the data involving negative RT rule out such a theory and demand instead a process in which a probe item is compared in some fashion against several items in memory. A direct access memory strength theory accounts for differences in negative RT either by appealing to differences in the memory strength of nonpresented items (where strength is presumably accrued by generalization from presented items) or by differences in the criterion against which memory strength is assessed. In either case, negative RT is slower as the criterion is nearer the strength of nonpresented items. In Experiment 3, negative RT and the relation between negative RT and the numbers of presented items and translations differed among negative translation probes, negative no-translation probes with potential translations, and negative no-translation probes without potential translations. Only the first two types of probes were influenced by the number of translations represented in the memorized list, the first type of probe more than the second type. It does not seem reasonable to attribute these differences to differences in generalized memory strength of nonpresented items; to do so, one would have to claim that the presentation of items of one category induces more generalized strength in non-
TRANSLATION
IN MEMORY
SEARCH
27
presented items of the other category than in nonpresented items of the presented category. However, attributing these differences to the use of different decision criteria for the different types of negative probes entails the most implausible conclusion that subjects recognize the category of a probe and select from at least three different decision criteria (one for translation probes, one for no-translation probes with a potential translation, one for no-translation probes without a potential translation) between the time the probe is presented and a response is initiated. We know of no evidence that subjects can choose criteria quickly enough and consider it most unlikely that they are able to do so. The natural account of the negative translation RT data is given by the long buffer hypothesis, Equation 7b, which claims that the probe is compared first against the presented items and then, if necessary, against their translations. A simple generalization of this hypothesis can be made to account for the finding that the number of translations presumed to be in memory influences RT to a negative no-translation probe with a potential translation. Subjects, at least on some trials, use the information that a probe has a potential translation, together with the information that no match was found in comparing the probe with the presented items, to occasion a further search of the translations of the presented items. That is, they sometimes treat a negative notranslation probe with a potential translation (but not one without a potential translation) like they do a negative translation probe, resulting in the steepening of the RT functions for negative no-translation probes observed in Experiments 1 and 2. The data on the translation effect in memory search thus point to a process in which a probe item is compared against some or all of the items in memory, but in which the order and rate of comparison is determined by the relative amount of encoding and rehearsal capacity devoted to different items. Much of the argument for this conclusion has been, admittedly, speculative and imprecise and certainly does not allow detailed conclusions to be made about the nature of the memory search process-whether it is serial or parallel, exhaustive or self-terminating, the functional relation between amount of rehearsal and comparision time, etc. Fortunately, some theorists (particularly Townsend, 1974) have begun to consider memory search models which permit variations in individual item comparison time, and others (particularly Cavanagh, 1976) have theorized about the relation between rehearsal, memory strength, and decision time. Further such theorizing can be expected to correct errors made in the speculations advanced in the present paper. An additional, and equally important, approach would be to bring subjects’ rehearsal strategies under experimental control and directly assess the relation between amount of rehearsal and RT, rather than relying on subjects’ reports and speculations about task demands as has been done
28
CLIFTON, SORCE AND CRUSE
here. Some initial attempts have been made in this direction (Cavanagh, 1976; Seamon & Wright, 1976; Sorce & Clifton, Note 3). Recoding Information
in Memory:
Final Conclusions
The original intent of the research that has been reported was to measure the time taken to decode information from memory. The initial model of the translation effect in memory scanning suggested that the time to decode an item from the form in which it was held in memory to another form could be measured by the increase in slope of the memory scanning function (the value oft in Equation 3). If the initial model stood up to experimental test, the time taken to execute different types of recoding schemes could be measured. The model did not stand the test. Instead, it appears that, in the situations that have been considered, any necessary recoding is done at the time information is originally input into memory. Still, some interesting conclusions can be made and some interesting further questions raised. It seems that the time to make a positive recognition decision about an item may depend upon the extent to which the item has been rehearsed. Even more, the time to make a negative recognition decision about an item belonging to a given class may depend upon the extent to which other items of that class have been rehearsed. Thus, it may prove possible to make inferences from recognition RT to the extent to which rehearsal, or encoding, capacity is devoted to various types of recodings of various types of presented items. An interesting question for further study is whether the encoding process which influences recognition RT is limited to conscious, deliberate rehearsal (as has been assumed in discussing the present data), or can include automatic processes such as the concept of spreading activation between semantic concepts discussed by such researchers as Collins and Loftus (1975) and Meyer and Schvaneveldt (1971). REFERENCES Anderson, J. R., & Bower, G. H. Human associative memory. Washington, D.C.: Winston, 1973. Baddeley, A. D., & Ecob, J. R. Reaction time and short-term memory: Implications of repetition effects for the high-speed exhaustive scan hypothesis. Quarterly Journal of Experimental Psychology, 1973, 25, 229-240. Battig, W. F., & Montague, W. C. Category norms for verbal items in 56 categories: A replication and extension of the Connecticut Category Norms. Journal of Experimental Psychology Monograph, 1%9, SO,(3, Part 2). Cavanagh, P. Holographic and trace strength models of rehearsal effects in the item recognition task. Memory & Cognition, 1976, 4, 186-199. Chase, W. G., & Calfee, R. C. Modality and similarity effects in short-term recognition memory. Journal of Experimental Psychology, 1969, 81, 510-514.
TRANSLATION
IN MEMORY SEARCH
29
Clark, H. H., & Chase, W. G. On the process of comparing sentences against pictures. Cognitive Psychology, 1972, 3, 472-517. Clifton, C., Jr., & Birenbaum, S. Effects of serial position and delay of probe in a memory scan task. Journal of Experimental Psychology, 1970, 86; 69-76. Clifton, C., Jr., & Brewer, E. Partially-selective search of memory for letters and digits. Memory and Cognition, in press. Ciifton, C., Jr., & Cruse, D. Time to recognize tones: Memory scanning or memory strength? Quarterly Journal of Experimental Psychology, in press. Clifton, C., Jr., Cruse, D., & Gutschera, K. D. Recoding processes in recognition: Some effects of presentation rate. Memory and Cognition, 1973, 1, 387-394. Collins, A. M., & Loftus, E. F. A spreading-activation theory of semantic processing. Psychological Review, 1975, 82, 407-428. Cruse, D., & Clifton, C., Jr. Recoding strategies and the retrieval of information from memory. Cognitive Psychology, 1973, 4, 157-193. Darley, C. F., Klatzky, R. L., & Atkinson, R. C. Effects of memory load on reaction time. Journal of Experimental Psychology, 1972, %, 232-234. Juola, J. F., & Atkinson, R. C. Memory scanning for words versus categories. Journal of Verbal Learning and Verbal Behavior, 1971, 10, 522-527. Klatzky, R. L., & Atkinson, R. C. Memory scans based on alternative test-stimulus representations. Perception & Psychophysics, 1970, 8, 113- 117. Klatzky, R. L., Juola, J. F., & Atkinson, R. C. Test stimulus representation and experimental context effects in memory scanning. Journal of Experimental Psychology, 1971, 87, 281-288. KuEeva, H., & Francis. W. N. Computational analysis of present-day American English. Providence, RI: Brown University Press, 1967. Marcel, T., & Forrin, B. Naming latency and the repetition of stimulus categories. Journal of JTxperimental Psychology, 1974, 103,450-460. Meyer, D. E., & Schvaneveldt, R. W. Facilitation in recognizing pairs of words: Evidence of a dependence between retrieval operations. Journal of Experimental Psychology, 1971, 90, 227-234. Miller, G. A. Some psychological studies of grammar. American Psychologist, 1962, 17, 748-762. Mohs, R. C., Wescourt, K. T., & Atkinson, R. C. Effects of short-term memory contents on short- and long-term memory searches. Memory and Cognition, 1973, 1, 443-448. Naus, M. J. Memory search of categorized lists: A consideration of alternative selfterminating search strategies. Journal of Experimental Psychology, 1974, 102, 9921000. Naus, M. J., Glucksberg, S., & Omstein, P. A. Taxonomic word categories and memory search. Cognitive Psychology, 1972, 3, 643-654. Nickerson, R. S. Binary-classification reaction time: A review of some studies of human information-processing capabilities. Psychonomic Monograph Supplements, 1972,4, 275-318 (whole number 65). Schank, R. Conceptual dependency: A theory of natural language understanding. Cognitive Psychology, 1972, 3, 552-63 1. Seamon, J. G., & Wright, C. E. Generative processes in character classification: Evidence for a probe encoding set. Memory & Cognition, 1976, 4, 96-102. Stemberg, S. High-speed scanning in human memory. Science, 1966, 153,652-654. Stemberg, S. Memory-scanning: Mental processes revealed by reaction-time experiments. American Scientist, 1969, 57, 421-457. Stemberg, S. Memory scanning: New findings and current controversies. Quarterly Journal of Experimental Psychology, 1975, 27, l-32.
30
CLIFTON, SORCE AND CRUSE
Swanson, J. M., Johnsen, A. M., & Briggs, G. E. Recoding in a memory search task. Journal of Experimental Psychology, 1972, 93, l-9. Thomas, E. A. C., & Myers, J. L. Implications of latency data for threshold and nonthreshold models of signal detection. Journal of Mathematical Psychology, 1972, 9, 253-285. Townsend, J. T. Issues and models concerning the processing of a finite number of inputs. In B. H. Kantowitz (Ed.), Human information processing: Tutorials in performance and cognition. Hillsdale, NJ: Erlbaum Associates, 1974.
REFERENCE
NOTES
1. Clifton, C., Jr., Gutschera, K. D., Brewer, E., & Cruse, D. Recoding in a character classification task: Some inconsistent effects of recoding difficulty. Report 73-3, Cognitive Processes Laboratory, University of Massachusetts, Amherst, June 1973. 2. Stemberg, S. Evidence against self-terminating memory search from properties of RT distributions. Paper presented at meetings of Psychonomic Society, St. Louis, November 1973. 3. Sorce, P., & Clifton, C., Jr. Effect of frequency of presentation in a memory search task. Report 76-5, Cognitive Processes Laboratory, University of Massachusetts, Amherst, June, 1976. (Accepted July 20, 1976)