隱示記憶的收錄–助益/抑制歷程

隱示記憶的收錄–助益/抑制歷程

中 華 民 國 94 年 5 月 13 日

Running Head: Positive and Negative Priming Effects on Memory

FACILITATIVE AND INHIBITORY EFFECTS OF STIMULUS ENCODING ON MEMORY IN IMPLICIT TESTS

Chao-Ming Cheng and Chia-Shan Tsai National Taiwan University

Correspondence: Chao-Ming Cheng

Department of Psychology National Taiwan University

1, Sec. 4, Roosevelt Road Taipei, Taiwan

Fax Number: 886-2-23629909 E-Mail Address: [email protected]

FACILITATIVE AND INHIBITORY EFFECTS OF STIMULUS ENCODING ON MEMORY IN IMPLICIT TESTS

Abstract

Processes of conscious and unconscious forms of human memory were explored in three implicit tests (stem completion, word association, and word identification) by examining how the two forms of memory within each test were differently affected by level-of-processing (LOP) and self-generation of study words. These two forms of memory were separated by a post-test dissociation (PTD) procedure, which was able to separate conscious from unconscious memory contributing to a test without being contaminated by memory and guessing effects produced by post-test judgments. Results show that both LOP and generation produced positive effects, associated with either a positive effect (in the tests of stem completion and word identification) or a null effect (in the test of word association) of repetition priming under shallow processing, on estimates of conscious memory. On the other hand, LOP produced null effects and generation produced reverse effects accompanied by a repetition-inhibition effect under generation conditions on

estimates of unconscious memory. Implications of this pattern of results for a componential versus holistic transfer approach to conscious and unconscious memories are discussed.

Key words

The Post-Test Dissociation Procedure, Repetition-Priming Effects, Repetition-Inhibition Effects, Componential versus Holistic Transfer of Processing, Effects of

A major focus of recent research in memory has been the distinction between two forms of human memory: conscious and unconscious (for reviews, see Roediger, 1990; Schacter, 1987). Conscious memory is referred to as memory that is characterized by intentional retrieval accompanied by awareness of memory. Unconscious memory is memory that operates automatically, without involving retrieval volition and memory awareness. A widely used method to illuminate these two forms of memory has been to contrast performance between explicit and implicit tests (the task-dissociation procedure). Explicit tests such as recall and recognition are assumed to tap conscious memory. Implicit tests such as word-stem completion and word identification, in which participants are not instructed to make reference to studied words, are thought to reflect unconscious memory. It has been claimed that memories measured by these two types of test are different in nature because they are differently sensitive to many experimental variables such as level-of-processing (LOP) and self-generation (e.g., Blaxton, 1989; Jacoby, 1983; Jacocby & Dallas, 1981; Shimamura, 1986; Tulving & Schacter, 1990; Weldon & Roediger, 1987; Winnick & Daniel, 1970). For example, Winnick and Daniel reported that compared with reading study words, generating study words increased memory in explicit tests but decreased memory in implicit tests. Jacoby and Dallas found that the variation in LOP exerted large effects in explicit tests but little or no effect in implicit tests.

Roediger and associates (e.g., Blaxton, 1989; Roediger, 1990; Roediger & Weldon, 1987) proposed a theoretical approach to performance on explicit and implicit tests based on transfer-appropriate processing (TAP). The approach is predicated on the notion that performance on a memory test benefits to the extent that the cognitive operations at test recapitulate those engaged during initial learning. According to this approach, explicit tests

are conceptually driven and implicit tests are perceptually driven. With regard to effects of generation, simple reading of study words involves perceptual encoding of the words, whereas generation of study words involves conceptual encoding of the words. Thus, performance on a subsequent explicit test would benefit from generation but not benefit from reading, resulting in a positive effect of generation on test performance. In contrast, performance on a subsequent implicit test would benefit from reading but not benefit from generation, leading to a reverse effect of generation on performance. The variation in LOP is also expected to produce positive effects in explicit tests for the same reason generation produces positive effects in the tests. However, deep processing involves not only

conceptual encoding but also perceptual encoding of study words (as participants see the study words while processing them deeply), performance on a subsequent implicit test would therefore benefit from perceptual encodings engaged in shallow and deep processing, resulting in a null effect of LOP in an implicit test.

However, it has also been shown that memories measured by the two types of test, particularly by implicit tests, are not consistently sensitive to experimental manipulations across studies. For example, despite initial evidence that LOP produces no effect in implicit tests (e.g., Graf, Mandler, & Hayden, 1982; Jacoby, 1983; Jacoby & Dallas, 1981), a large number of studies have shown positive effects of LOP in the tests (for reviews, see Brown & Mitchell, 1994; Challis & Brodbeck, 1992). Compared with simple reading, generation decreased memory (Blaxton, 1989; Java, 1994; MacLeod & Masson, 1997; Schwartz, 1989; Smith & Branscombe, 1988), increased memory (Gardiner, 1988, 1989; Gardiner, Dawson, & Sutton, 1989; Toth & Hunt, 1990), or resulted in the same amount of memory (Gardiner, 1988; MacLeod & Masson, 1997; Nicolas & Tardieu, 1996; Schwartz, 1989) in implicit

tests. These results suggest that LOP and generation affect performance on implicit tests in a way much more complicated than proposed by the TAP theory.

Conscious and Unconscious Memories Within a Test

The task distinction between conscious and unconscious memory is further

complicated by the current view that a memory test, whether explicit or implicit, does not provide a pure measure of either conscious or unconscious memory. Rather, it is more likely that any test involves the two forms of memory for normal participants. For example, it has been shown that explicit memory tests under the process-dissociation procedure (Jacoby, 1991, 1998; Jacoby, Toth, & Yonelinas, 1993; Toth, Reingold, & Jacoby, 1994) involve not only recollection (reflecting conscious memory) but also automatic uses of memory (including unconscious memory). Although no direct evidence so far has been shown to support the notion that implicit tests also involve both conscious and unconscious memories, indirect evidence for this notion has been given by neuropsychological studies (e.g., Hamann & Squire, 1996; Squire, Shimamura, & Graf, 1987) that show larger effects of LOP in implicit tests for normal than for amnesic participants. For example, Hamann and Squire found that, compared with normal participants, amnesic patients exhibited reduced or near-zero LOP effects on implicit word completion. Squire et al. investigated LOP effects in the implicit tests of word-stem and word-fragment completion. Their results showed that LOP effects were larger for normal control than for amnesic patients. Because explicit retrieval of memory is usually absent in amnesic participants and is positively sensitive to LOP in normal participants, these results suggest that implicit tests involve conscious memory as well as unconscious memory for normal participants.

conscious and unconscious memory by contrasting performance between explicit and implicit tests. Rather, it should be assessed by examining how the two forms of memory differ from each other in responding to experimental manipulations or study-test

interactions within a test. Such assessments have been largely reported in explicit tests (e.g., Bodner, Masson, & Caldwell, 2000; Jacoby, 1998; Toth et al., 1994) but seldom reported in implicit tests. The present authors argue that such assessments in explicit and implicit tests are not redundant to each other and that it is probably more desirable to assess the two forms of memory in implicit than in explicit tests, even granted that the nature of conscious and unconscious memories is not different in these two types of test. Explicit tests are primarily conceptually driven, whereas implicit tests can be perceptually driven,

conceptually driven, or perceptually and conceptually driven. For example, the implicit test of word association is conceptually driven because participants in the test are required, upon seeing a test word that is semantically related but visually dissimilar to a studied word, to provide the first associate that comes to mind based on semantic relatedness. The

word-identification test is perceptually driven or data-driven because the identification of a test word presented briefly in a noise background is based on the visual features detected. The stem-completion test is conventionally regarded as perceptually driven. However, the test is also conceptually driven because the completion of a stem by a word should be guided by lexical and semantic knowledge as well as by the stem. Thus, compared with explicit tests, implicit tests allow one to assess the two forms of memory in various study-test situations, which would lead to a better understanding of their underlying mechanisms

Separating Conscious From Unconscious Memory in Implicit Tests: A Post-Test Dissociation Procedure

The method widely used for separating conscious from unconscious component of memory within a test has been the process-dissociation procedure (Jacoby, 1991; Jacoby et al., 1993; Toth et al., 1994). However,the procedure is inadequate for implicit tests because it, whether adopting direct-retrieval, generate-recognize, or independence/remember-know instructions (Jacoby, 1998), is applicable to explicit tests only but not to implicit tests. For example, in the test of word-stem completion, the procedure separates recollection from automatic uses of memory by contrasting performance between inclusion and exclusion tests following a study phase. Participants in the two tests are required to complete stems by making reference to their own recollection of studied words. Thus, these two tests

correspond to explicit rather than implicit tests.

Conscious memory in an implicit test can be either based on recollection accompanied by memory awareness or based on unintentional retrieval immediately followed by memory awareness (i.e., involuntary conscious memory) (see

Richardson-Klavehn, Gardiner, & Java, 1996; Schacter, 1987) contrary to the intention of the experimenter. In contrast, its unconscious counterpart is memory without involving retrieval intention and awareness. Thus, the two components of memory are distinguished in terms of memory awareness that occurs at the time of test. Cheng (1999, 2003) proposed a post-test dissociation (PTD) procedure for separating these two components of memory within an implicit test. The procedure will be illustrated below in the implicit test of stem completion, but it can be applied as well to other implicit tests without modification. The procedure begins by administering an incidental study of words in which either LOP or

generation is varied (e.g., either complexity or pleasantness rating of study words), followed by a conventional implicit test of stem completion in which stems are presented for completion by the first word that comes to mind. Immediately after the test, participants are required to indicate studied words from among the words that had been used to

complete stems. Because the participants do not know the test to be a memory test until they are solicited to judge studied words, the test is implicit at the time of performing stem completion. However, the test up to this point may involve both conscious and unconscious memories. The PTD procedure is able to separate these two forms of memory, without being contaminated by guessing effects and recognition memory produced at the judgment stage.

Within the PTD procedure, the total number of trials for an experimental (either study or “nonstudy”) condition can be divided into four proportions of P(i,j). A P(i,j) denotes either the proportion of completion with experimental words (either the experimental words presented for study or those not presented for study) (i = 1) or that with other words (i = 0), with the words being either indicated as "studied” (j = 1) or not indicated (j = 0). For example, the P(1,1) denotes the proportion of completion with experimental words that were indicated as “studied”; the P(1,0) denotes that with experimental words not indicated as “studied”. The sum of P(1,1) and P(1,0), Po, following an experimental condition is then

the proportion of completion with experimental words under that condition. Components of P(1,1) Under a Study Condition.

The PTD procedure identifies two types of conscious memory that can contribute to P(1,1) in a study condition. One is conscious memory that contributes to completion with studied words at the time of test. Let the probability of this type of conscious memory be R.

The other is recognition memory, which is based on completion with studied words without involving memory awareness at the time of test, but later being recognized as “studied” as a result of making post-test judgments. Such recognition memory (hereafter termed

judgment-induced recognition) would be attributed to baseline performance if participants were not forced to make memory judgments. Let the probability of baseline performance be B, then the probability of judgment-induced recognition is part of B, B1. Because

judgment-induced recognition does not contribute to stem completion at test, it should be excluded from estimates of R. Note that judgment-induced recognition is different from involuntary conscious memory (Richardson-Klavehn et al., 1996; Schacter, 1987) in that the former is produced through post-test judgments whereas the latter is characterized by unintentional retrieval automatically followed by recognition of studied words at the time of completion. The latter is considered as part of R in the PTD procedure.

--- Insert Table 1 about here ---

It is assumed that unconscious memory by its nature cannot become conscious even when under memory judgment; it would otherwise be difficult to make a distinction

between conscious and unconscious memory if unconscious memory can become conscious under memory judgments. Further, guessing will occur at the judgment stage when the completion with studied words is either based on unconscious memory (the probability of unconscious memory is denoted by U) or based on the part of baseline completion without leading to judgment-induced recognition (the probability of this part of baseline completion is denoted by B2, where B2 is equal to B - B1). Thus, as Table 1 shows, the P(1,1) under a

study condition estimates (a) R (the probability of conscious memory contributing to stem completion at test), (b) B1 (the probability of judgment-induced recognition), when there is

a failure of either conscious or unconscious memory at test, and (c) g (the guessing rate in favor of a “studied” judgment), when there is a failure of either conscious memory or judgment-induced recognition. Namely,

P(1,1) = R + (1 - R)(1 - U)B1 + [(1 - R)U + (1 - R)(1 - U)B2]g, (1)

The probability g is to be estimated by the ratio of N(1,1)/[N(1,1) + N(1,0)], where the N(1,1) and N(1,0) are the P(1,1) and P(1,0) for “nonstudy” words (the experimental words not presented for study), respectively.

Components of P(1,0) Under a Study Condition.

Table 1 also shows that the P(1,0) under a study condition estimates the probability of completion with a studied word either based on unconscious memory or based on the part of baseline completion without leading to B1, with the word being not indicated as

“studied” under guessing. Namely,

P(1,0) = [(1 - R)U + (1 - R)(1 - U)B2](1 - g). (2)

Equations 1 and 2 were formulated based on the assumption that events for R, U, and B are independent of one another. It is immediate from Equations 1 and 2 that Po [the sum

of P(1,1) and P(1,0)] estimates R + (1 - R)U + (1 - R)(1 - U)B. Let P’(1,1) and P’(1,0) be P(1,1) and P(1,0) after correction for guessing, respectively, then

P’(1,1) = P(1,1) - [P(1,0)N(1,1)/N(1,0)] = R + (1 – R)(1 – U)B1, (3)

and

P’(1,0) = P(1,0)[N(1,1) + N(1,0)]/N(1,0) = (1 – R)U + (1 – R)(1 – U)B2. (4)

Estimates of R and U

Solving R and U as parameters from Equations 3 and 4 (see Appendix A), one obtains

R = P’(1,1) – B1(1 – Po)/(1 – B), (5)

and

U = [P’(1,0) – B2(1 – Po)/(1 – B)]/(1 – R). (6)

One way to justify Equations 5 and 6 is to rewrite these two equations as

R = [(Po – B)/(1 – B) – U]/(1 – U), (7)

and

respectively (see Appendix B). The value of (Po – B)/(1 – B) in Equations 7 and 8 can be

understood as much as the proportion of completion with studied words that increases from baseline performance (B) by both conscious and unconscious memories in a test, Po – B,

cannot exceed the limit that remains to be increased from baseline performance to 1.0 in the test, 1 – B. The magnitude of this limit is to be determined by the difficulty of the memory test used, among others. The extent to which Po – B approaches this limit is determined by

the ability to remember; the higher the ability is, the better the observed memory

approaches this limit. It is therefore desirable to indicate the contribution of memory to the test by the relative magnitude of the observed amount of memory to its maximal memory, (Po – B)/(1 – B). The value of (Po – B)/(1 – B) ranges from 0 to 1.0, as Po ranges from B to

1.0 [i.e., the sum of P’(1,1) and P’(1,0) is equal to 1.0, indicating that the contribution of conscious and unconscious memories to performance results in 100 percent correct]. Subtracting either U in Equation 7 or R in Equation 8 from (Po – B)/(1 – B) then denotes

either the proportion contributed by conscious memory or that by unconscious memory under the independence assumption.

The probability of B1 (judgment-induced recognition) in Equation 5 is estimated

based on the assumption that the probability of conscious memory relative to a proportion of completion with studied words is the same across the testing and judgment stages (i.e., R/Po = B1/B). It follows that B1 is equal to BR/Po and B2 in Equation 6 is equal to B(1 –

R/Po) [i.e., B - BR/Po].

By using the PTD procedure, Cheng (2003) assessed conscious and unconscious memories within an implicit test of stem completion by examining how they were affected

by LOP and self-generation. His results showed that LOP and generation produced positive effects on estimates of conscious memory. On the other hand, LOP produced no effect, whereas generation produced a reverse effect associated with a repetition-inhibition effect under generation, on estimates of unconscious memory. This pattern of results is not paradoxical because similar results are also found in previous studies separating conscious from unconscious memory in explicit stem-completion tests (e.g., Bodner et al., 2000; Cheng, 2005; Toth et al., 1994).

Purpose of the Present Study ---

Insert Table 2 about here ---

It was argued earlier that it is more desirable to assess conscious and unconscious memories in implicit than in explicit tests. Following this argument, the present study assessed the two forms of memory in three implicit tests: stem completion (Experiment 1), word association (Experiment 2), and word identification (Experiments 3 and 4). In particular, we examined how these two forms of memory in each test were differently affected by level-of-processing (LOP) and self-generation varied at study. The TAP theory (e.g., Blaxton, 1989; Roediger, 1990) is not explicit about LOP and generation effects in the case in which an implicit test involves both conscious and unconscious memories. However, the theory may be modified to accommodate this case as shown in Table 2. In Table 2, the “0” and “+”signs stand for zero and positive repetition-priming effects, respectively. The component of conscious memory within an implicit test is in fact based on explicit retrieval

and, hence, is conceptually driven. According to the TAP theory, this component of memory would therefore be better under both deep and generation than under shallow conditions. On the other hand, its unconscious counterpart, that is perceptually driven, would be equal for shallow and deep conditions and better under shallow than under generation conditions. Thus, LOP is expected to produce a positive effect on the conscious component and no effect on the unconscious component of memory and a positive effect on overall memory across the two components of memory in an implicit test. On the other hand, generation in the test is expected to produce a positive effect on conscious memory, a reverse effect on unconscious memory, and a positive, null, or reverse effect on overall memory, depending on the net value of priming effects across the two components of memory for each study condition. Another prediction is that effects of LOP and generation on conscious and unconscious memories would be invariant across the three implicit tests because the extent to which the cognitive operations of conscious and unconscious memories recapitulate those engaged at study will remain the same across the three tests.

Alternatively, effects of LOP and generation on conscious and unconscious memories may be different in the three implicit tests for the reason that the three tests are differently driven and, hence, the extent to which the cognitive operations at test recapitulate those engaged at study will be different for different tests.

Experiment 1: Effects of LOP and Generation on Implicit Stem Completion Method

Participants. Participants were 48 volunteers at National Taiwan University, who were native speakers of Chinese Mandarin. They have learned English as a second

extra credit for an introductory psychology course.

Stimulus materials. Experimental stimuli consisted of 108 Chinese two-character words [each word consisted of two characters such as 地毯 (carpet); a Chinese character

corresponds to a stem in an English word] with each word paired with its English translation and 108 stems created from the Chinese words. The Chinese words were selected from the Liu, Chuang, and Wang (1975) corpus, with their frequency ranging from 1 to 89. Each Chinese word was selected such that it ranked from third to 34th according to the frequency of use in a word group sharing the same first character. The Chinese words were also selected such that none shared the first constituent character with the others, and each could be translated from its English counterpart without ambiguity by a group of university students of the participants’ level in a pilot study. The words in Chinese and English were prepared in a personal computer for a visual presentation on the monitor. All words were grouped into six blocks of words, with the blocks matched in word complexity, frequency, and ranking order. The words in each block were randomly permuted and were then fixed for all participants. The 108 stems were created by replacing the second

constituent character of each experimental Chinese word with a dash (for example, the stem for the experimental word 地毯 is 地__). The phonology of each stem was the same in the

group of words beginning with the stem.

Design and procedure. The participants were tested individually. The experiment was conducted in thee phases: study, test, and memory judgment. In the study phase, the

participants were first given instructions to perform the task and 20 trials for practice. The practice was then followed by an experimental session in which the six blocks of

experimental words were rotated through two presentation conditions, with three blocks presented for study and the other three not presented for study. For a participant, one of the three study blocks was presented for rating character complexity of each Chinese word by asking the participant, upon seeing a word on the monitor, to judge roughly whether the first constituent character was more complex than the second one with respect to number of strokes. The participant was to respond as quickly as possible by pressing one of the three response keys on the computer keyboard corresponding to three response alternatives: "yes", "no", and "about equal".

The second study block was presented for rating pleasantness of each Chinese word by asking the participant to press a response key corresponding to one of three alternatives: "pleasant", "unpleasant", and "neutral". The third study block presented English words and the participants were asked to translate verbally each English word into a two-character Chinese word. Because English was learned as a second language, the participants’ English-Chinese translation skills are considered as consciously controlled rather than automatically controlled. The translation was thus regarded as a sort of semantic generation.

The assignment of the three study blocks to the three study conditions and the sequence of block presentation were balanced across participants. Each of the study words was kept exposed on the monitor until the participant either pressed a response key or verbally generated a Chinese translation. Its termination was immediately replaced by the onset of a succeeding word. However, the maximum exposure duration for each word was set at 3 sec. The experiment was thus carried out in a completely within-participants design, with the processing of study words varied at three levels: shallow, deep, and generation.

Immediately after the ratings, the participants were given instructions for a

stem-completion test. In the test, the participants were given the 108 stems for completion. Half of them were originated from the Chinese studied words (those presented for

complexity and pleasantness rating and generated from English) and the other half were from the Chinese words in the three “nonstudy” blocks (the experimental words not presented for study). All stems were randomly permuted and printed on answer sheets, and were presented to the participants one by one for completion. The participants were to complete each stem with the first appropriate word that came to mind, and to write down the word on the answer sheet provided. For all tests, the participants were encouraged to complete as many stems as possible with the completion words being two characters long. No proper names were allowed.

Immediately after all stems had been completed, the participants further indicated on the answer sheets those response words that were studied words.

Statistical analyses. The proportion of stems completed with experimental words in each condition (shall, deep, generation, or “nonstudy”), Po, was collected for each

participant. The proportion data were then subject to ANOVAs. Multiple comparison tests were performed if necessary. The same procedure was also applied to testing the

significance of the difference in estimates of R or U among study conditions. In the case in which only two study conditions were to be compared in an experiment, a paired-t test was used to test the difference in estimates of R or U between the two conditions. For all statistical tests, the level of significance was set at 0.05 unless otherwise specified. Results and Discussion

and P(1,0) scores after correction for guessing, P’(1,1), and P’(1,0)] for the three study and “nonstudy” conditions in Experiment 1. The shallow and deep conditions stand for the conditions for character-complexity and pleasantness ratings, respectively. The generation condition stands for the condition for English-Chinese translation. The “nonstudy”

condition is the condition without a prior study phase. Each Po score is the mean proportion

of completion with experimental words under an experimental (either study or “nonstudy”) condition. Its components P(1,1) and P(1,0) are the proportions of completion with

experimental words with the words being indicated as "studied" and not indicated, respectively. The number of trials was 48 (participants) × 18 (stems) = 864 for each study condition and 48 × 54 = 2592 for the “nonstudy” condition.

--- Insert Table 3 about here ---

Table 3 shows relatively low completion with studied words and baseline

performance. A possibility is that in general a Chinese character serving as a stem has a large number of two-character word completions. The stems used in Experiment 1 had the number of completions ranging from 3 to 130 two-character words with a mean completion of 28.5 words. This large number of potential completions may have decreased baseline performance and, hence, Po, P(1,1), and P(1,0) scores. On the contrary, English stems used

in previous studies (e.g., Jacoby, 1998; Jacoby et al., 1993; Toth et al., 1994) had a smaller number of completions restricted by a criterion, which resulted in high rates of completion with experimental words. For example, the stems used by Jacoby (1998) had the number of five-letter word completions ranging from two to nine words.

Statistical analyses of Po scores [which estimates R + (1 - R)U + (1 - R)(1 - U)B]

showed a significant difference among the four experimental conditions, F(3, 141) = 30.7, MSE = 111.6. The proportion of completion with experimental words was larger under the three study conditions than under the “nonstudy” condition, F(1, 141) = 79.1, MSE = 111.6, under the deep than under the shallow condition, F(1, 141) = 5.2, MSE = 111.6, and under the deep than under the generation condition, F(1, 141) = 12.1, MSE = 111.6. The

difference in completion between the shallow and the generation condition did not reach the .05 level of significance, F(1, 141) = 1.45. These results indicate a positive effect of LOP and a null effect of generation on overall performance in the test. Further, these effects were associated with a repetition-priming effect under each study condition, F(1, 141) = 47.7, MSE = 111.6, for the shallow, F(1, 141) = 84.4, MSE = 111.6, for the deep, and F(1, 141) = 32.5, MSE = 111.6, for the generation condition.

Table 3 also presents mean estimates of R and U for the three study conditions. Equation 5 was used to estimate R for each participant under each condition. Results of this estimation show that the estimate of R is substantial under each study condition, indicating that the present implicit test did not measure a pure form of unconscious memory; it also involved conscious memory. There was a main effect of study condition on this estimate, F(2, 94) = 7.54, MSE = 66.1, due to a larger estimate of R following the deep than

following the shallow condition, F(1, 94) = 11.3, MSE = 66.1, and following the generation than following the shallow condition, F(1, 94) = 6.10, MSE = 66.1. The difference in this estimate between the deep and the generation condition was not significant, F(1, 94) = 1.92. These results indicate positive LOP and generation effects on the R estimate.

Results show that study condition produced a main effect on the estimate of U, F(2, 94) = 13.0, MSE = 73.6. The estimate was lager under the shallow than under the generation condition, F(1, 94) = 20.3, MSE = 73.6. There was no difference in this estimate between the shallow and the deep condition, F(1, 94) < 1.0. These results indicate a reverse effect of generation and a null effect of LOP on the U estimate in the test.

Taken together, the results of the experiment show that the variation in LOP produced a positive effect on overall memory and the R estimate but no effect on the U estimate. The results also show that compared with the shallow condition, the generation condition did not increase overall memory, but it increased the R estimate and decreased the U estimate. Similar results are also found in previous studies including those using the

process-dissociation procedure to estimate recollection and automatic memory in explicit tests. For example, Richardson-Klavehn and Bjork (1988), Roediger and McDermott (1993), Shimamura (1986), and Tulving and Schacter (1990) reported positive effects of LOP on explicit recall. Toth et al. (1994, Experiment 1) reported positive LOP effects on recollection and null effects of LOP on automatic memory in explicit stem-completion tests. The reverse generation effect on the U estimate parallels reverse generation effects on

automatic memory in explicit stem-completion tests (Bodner et al., 2000; Toth et al. 1994, Experiment 2) despite the fact that the to-be-studied words in the present study were generated from their English counterpart, whereas those in previous studies were generated from various semantic cues. This result suggests that the reverse effect of generation on unconscious memory is relatively robust across various manipulations of generation. However, a comparison of the present results with those of Bodner et al. and of Toth et al.

also suggests that the variation in LOP and generation results in more conscious memory under deep and generation conditions in explicit than in implicit tests.

The sum R + (1 – R)U yields a score of 16.5 for the shallow, 22.6 for the deep, and 14.1 for the generation condition. This pattern of results is the same as shown by Po scores,

suggesting that effects of LOP and generation in an implicit test can be understood in terms of their effects on the R and U components within the test. These effect and the results showing a positive effect of LOP and a null effect of generation on overall performance can be successfully accommodated by a variant of of the TAP theory shown in Table 2.

However, the present results also showed a substantial repetition-priming effect on the R estimate following the shallow condition, despite the fact that the operation of conscious memory did not match the shallow condition in perceptual processing. This result is not predicted by the TAP theory. In addition, the reverse effect of generation on the estimate of U was resulted mainly by a repetition-priming effect under the shallow condition and a repetition-inhibition effect under the generation condition on U estimates. According to the present independence assumption, the repetition-inhibition effect is evidenced by the P’(1,0) score (which estimates both unconscious memory and baseline performance, see Equation 4), 4.8, being lower than baseline performance [(1 – R)(1 – U)N(1,0) where N(1,0) is the P(1,0) for “nonstudy” words], 10.1, under the generation condition. Cheng (2003, 2005) also found similar repetition-inhibition effects under generation on estimates of automatic memory in an explicit stem-completion test and on estimates of unconscious memory in an implicit stem-completion test. Such inhibition effects on U estimates cannot be

produces no repetition-priming rather than a repetition-inhibition effect on estimates of U. A possibility for such effects is that generation involves perceptual and conceptual

encoding of generation cues and conceptual encoding of study words, without

simultaneously involving perceptual encoding of the study words. This pattern of stimulus encoding may have created a perceptual environment that is not friendly for a copy of a studied word in the perceptual mode of information to be cued by perceptual information triggered by its stem at test. A possibility for the repetition-inhibition effect is that when information encoding of a word under generation does not result in conscious memory of the word, it may create the perceptual environment that pushes down, rather than pushes up, the ranking of the word in the potential word group to be triggered by a test item. For

example, a Chinese stem can be completed by a number of potential two-character word completions; the probability of stem completion with the studied word in this case would be lower when it is at the new ranking than when it is at the original ranking (which can be estimated from baseline performance), resulting in a repetition-inhibition effect on U.

Experiment 2: Effects of LOP and Generation on Implicit Word Association Method

Participants. Participants were 39 volunteers from introductory psychology courses at National Taiwan University, who were native speakers of Chinese Mandarin. They have learned English as a second language at least for seven years. They participated in the experiment in return for extra course credit.

Stimulus materials, design, procedure, and statistical analyses. Stimulus materials, design, procedure, and statistical analyses were the same as used in Experiment 1, with the following exceptions:

Stimuli consisted of 108 Chinese two-character words with each word paired with its English translation, which were used as experimental words, and 108 Chinese

two-character words, which were presented at test for association. The test words were obtained from performance of another group of participants on an association test. Each test word was selected such that it was a moderate associate of an experimental word and shared no constituent character with the word. In addition, the test word was visually

dissimilar to its corresponding experimental word. For example, the word 沙發 (sofa) was

used as a test word because it was semantically related but visually dissimilar to the experimental word 地毯 (carpet).

In the test session, the participants were given the 108 test words one by one for association. Half of them were associates of the studied words (those presented for

complexity rating and pleasantness rating and generated from English) in the three “study” blocks and the other half were associates of the experimental words in the three “nonstudy” blocks. All test words were randomly permuted and printed on answer sheets. The

participants were told, upon seeing a test word, to provide the first associate (a Chinese two-character word) that came to mind, and to write down the word on the answer sheet provided.

Results and Discussion

--- Insert Table 4 about here ---

score in Table 4 is the mean proportion of using experimental words as associates in a condition and its component P(1,j) is the proportion with the experimental words being either indicated as "studied" (j = 1) or not indicated (j = 0). It appears that the P(1,1) and P(1,0) scores and estimates of R and U were lower in Experiment 2 than in Experiment 1. There was a significant main effect of processing condition on Po scores, F(3, 69) = 10.4,

MSE = 55.4. The proportion of using experimental words for association at test was larger under the deep and generation conditions than under the “nonstudy” condition, F(1, 69) = 33.11, MSE = 55.4, and under these two study conditions than under the shallow condition, F(1, 69) = 15.8, MSE = 55.4. These results indicate a repetition-priming effect under the deep and generation conditions and positive effects of LOP and generation on overall memory in the test. The Po score tends to be larger under the shallow than under the

“nonstudy” condition, but the difference between the two failed to reach the .05 level of significance, F(1, 69) = 3.1. This result indicates the shallow condition in the experiment produced no repetition-priming effect on overall memory.

No correction for guessing was performed for further data analyses because the guessing rate was practically zero in the experiment. For this reason, Equations 5 and 6 were applied to estimate R and U in the experiment by replacing P’(1,1) and P’(1,0) scores by P(1,1) and P(1,0) scores, respectively. Results of estimating R showed a significant difference among the three study conditions, F(2, 46) = 5.8, MSE = 52.4. The difference was due to a larger estimate of R following the deep than following the shallow condition, F(1, 46) = 6.6, MSE = 52.4, and following the generation than following the shallow condition, F(1, 46) = 18.2, MSE = 52.4. The difference in this estimate between the deep and the generation condition was not significant, F(1, 46) = 2.9. This pattern of results

indicates a positive effect of LOP and of generation on the estimate of R. This result is the same as found in Experiment 1. However, the estimate of R following shallow processing was substantial in Experiment 1 but negligible in Experiment 2.

There was also a significant effect of study condition on the estimate of U, F(2, 46) = 4.43, MSE = 13.6. However, the only difference in the estimate between two study

conditions to reach the .05 level of significance was that between the deep and the generation condition, F(1, 46) = 12.1, MSE = 13.6. This result indicates that neither LOP nor generation produced a significant effect on the estimate of U in the experiment. Superficially, the null LOP effect on the U estimate is the same as found in Experiment 1. However, the null effect was caused by the two estimates of U under shallow and deep conditions being practically nil in Experiment 2, but being equally substantial in

Experiment 1. The null generation effect on the estimate of U was associated with a null repetition-priming effect under the shallow condition and a repetition-inhibition effect under the generation condition on this estimate. The two U estimates under the two study conditions were too small to be different from each other. The repetition-inhibition effect is evidenced by the P(1,0) score under the generation condition being lower than baseline performance (see Table 4). This inhibition effect is the same as found in Experiment 1 and provides further evidence that the pattern of stimulus encoding under generation produces an inhibition effect on unconscious memory. The magnitude of R + (1 – R)U under each study condition is 3.2 for the shallow, 10.2 for the deep, and 10.2 for the generation

condition. This pattern of results is the same as indicated by Po scores, suggesting that LOP

conscious and unconscious components of memory in the test.

Most of the results regarding LOP and generation effects on estimates of conscious and unconscious memories support the prediction of the modified version of the TAP theory shown in Table 2. However, the findings of no repetition-priming effect under the shallow and deep conditions and of a repetition-inhibition effect under the generation condition on the U estimate cannot be accommodated by the modified version of the TAP theory, which should predict a repetition-priming effect under the shallow and deep conditions and a null repetition-priming effect under the generation condition on U estimates.

Experiment 3: Effects of LOP on Implicit Word Identification Method

Participants. Forty-seven introductory psychology students from National Taiwan University served as participants of the experiment for extra course credit. The participants were test individually.

Stimulus materials. Stimuli consisted of 112 Chinese two-character words. The words were selected from the Liu, Chuang, and Wang (1975) corpus in the same way as used in Experiment 1 except that the words in Experiment 3 were not paired with their English counterpart. All words were grouped into four blocks of words, with the four blocks matched in word complexity, frequency, and ranking order.

Design, procedure, and statistical analyses. Design, procedure, and statistical analyses were the same as used in Experiment 1 except the following:

In the study phase, the participants were first given 20 trials for practice, followed by an experimental session in which the four blocks of experimental words were rotated

through two presentation conditions, with two blocks presented for study and the other two not presented for study. For a participant, one of the two study blocks was presented for shallow processing and the other for deep processing, which were the same as used in Experiment 1.

--- Insert Table 5 about here ---

Immediately after the study phase, the participants were given the 112 experimental words as targets for identification, half of which were studied words and the other half were “nonstudy” words. All target words were randomly mixed to determine their orders of presentation. The words were presented by embedding each word in a mask of random dots (density 2/15). When the participant was ready for a trial, she pressed the space bar of the keyboard to initiate a fixation point that lasted for 2.5 sec, which was then replaced by a test word. The exposure duration of each test word was set at 68 msec, which, according to results of a pilot study, resulted in roughly 30 percent correct identifications of words embedded in the same mask without a prior study. The participants were simply to identify each word and then to write down the answer on a sheet of paper provided. Participants were encouraged to guess each word when under uncertainty. Immediately after all target words had been tested, the participants were further instructed to indicate on the answer sheet those response words that were known as studied words.

Results and Discussion

Table 5 presents Po, P(1,1), and P(1,0) scores and estimates of R and U in the

targets and each P(1,j) is its component with the experimental words being either indicated as “studied” [P(1,1)] or not indicated [P(1,0)]. The number of trials was 47 (participants) × 28 (stimuli) = 1316 for each study condition, and was 47 × 56 = 2632 for the “nonstudy” condition. Baseline performance was higher in the experiment than in Experiments 1 and 2, suggesting that the number of potential alternatives from which a word can be identified based on the visual features detected is much lower than the number of potential

completions given a stem or the number of potential associates given a word for association. Processing condition produced a significant effect on Po scores in the test, F(2, 92) = 201.5,

MSE = 101.4. The identification of experimental words as targets was better under the shallow than under the “nonstudy” condition, F(1, 92) = 85.4, MSE = 101.4, and under the deep than under the “nonstudy” condition, F(1, 92) = 103.2, MSE = 101.4. The difference in identification between the shallow and the deep condition did not reach a level of significance, F(1, 92) < 1.0. Thus, although there was a repetition-priming effect under each study condition, effects of LOP on overall performance was practically nil in the experiment. This result is not the same as found in Experiment 1, primarily caused by a higher P(1,1) [or P’(1,1)] score and a lower P(1,0) [or P’(1,0)] score under the deep than under the shallow condition.

Results of estimating R were the same as shown in Experiment 1; the estimate of R was substantial under each study condition and the estimate was larger under the deep than under the shallow condition, t(46) = 4.6. Results of estimating U showed a reverse pattern; the U estimate was larger under the shallow than under the deep condition, t(46) = 5.0. This result was primarily due to a repetition-priming effect under the shallow condition,

baseline performance, and a repetition-inhibition effect under the deep condition, evidenced by the P’(1,0) scores under this condition being lower than its baseline performance. The positive repetition-priming effect on the R estimate under the shallow condition and the repetition-inhibition effect on the U estimate under the deep condition did not support the prediction of the modified version of the TAP theory that there should be no repetition effect on R estimates under shallow conditions and a positive repetition effect on U estimates under deep conditions (which also involve perceptual encoding of study words). The score of R + (1 – R)U was 36.4 under the deep condition, which is roughly equal to the score of 30.1 under the shallow condition. This suggests that the positive effect of LOP on the estimate of R and the reverse effect of LOP on the estimate of U may have offset each other to result in the null effect of LOP on overall memory.

Experiment 4: Effects of Generation Implicit Word Identification Method

Participants. Twenty-eight introductory psychology students from National Taiwan University served as participants of the experiment for extra course credit. They were native speakers of Mandarin Chinese and they have learned English as a second language. The participants were test individually.

Stimulus materials. Stimuli consisted of 112 Chinese two-character words with each paired with its English translation. The procedure for selecting these words was the same as used in Experiment 1. All words were grouped into four blocks of words, with the four blocks matched in word complexity, frequency, and ranking order of Chinese words.

Design, procedure, and statistical analyses. Design, procedure, and statistical analyses were the same as used in Experiment 1, except the following:

In the study phase, one of the two study blocks was presented for shallow processing. The other study block presented English words for English-Chinese translation.

Immediately after the study phase, the participants were given the 112 Chinese words (either presented for shallow processing or generated from English) as targets for

identification, half of them were studied words and the other half were “nonstudy” words. Results and Discussion

--- Insert Table 6 about here ---

Results of Experiment 4 are presented in Table 6. Main effects of processing

condition were significant, F(2, 54) = 57.0, MSE = 104.1. The shallow condition resulted in better identification of experimental words as targets than did the “nonstudy” condition, F(1, 54) = 17.8, MSE = 104.1; so did the generation condition, F(1, 54) = 16.9, MSE = 104.1. There was no difference in identification between the two study conditions, F(1, 54) < 1.0. These results indicate a positive repetition-priming effect under each study condition and a null generation effect on overall memory in the experiment. These results are the same as found in Experiment 1.

Estimates of R were substantial under the two study conditions, with the estimate of R. being larger under the generation than under the shallow condition, t(27) = 4.6. In contrast, effects of generation on the estimate of U show a reverse pattern; the estimate was larger under the shallow than under the generation condition, t(27) = 4.3. The reverse generation effect was due to a repetition-priming effect under the shallow condition and a

by the P’(1,0) scores being larger under the shallow condition and being smaller under generation than their corresponding baseline performance, respectively (see Table 6). The inhibition effect under the generation condition is further indicated by a negative value of the estimate of U under this condition, though a repetition-inhibition effect does not necessarily imply a negative value of memory estimate.

When indicated by the weighted sum of R + (1 – R)U, the overall memory was 16.9 under the shallow condition and 16.3 under the deep condition, suggesting that the positive generation effect on the estimate of R and the reverse generation effect on the estimate of U may have offset each other to result in the null generation effect on overall memory in the test.

General Discussion

This study was motivated by a need to assess conscious and unconscious memories in implicit tests because compared with explicit tests, implicit tests are able to provide various study-test situations, across which these two forms of memory can better be understood by examining how they are influenced by stimulus encoding. Following this way of thinking, we examined in this study influences of LOP and self-generation on conscious and

unconscious memories in the three different implicit tests. The two components of memory in each test were separated by the PTD procedure which began by administering a

conventional implicit test followed by a judgment task in which participants were required to identify studied words from among the words that had been generated in responding to test cues. The PTD procedure was developed by assuming that events responsible for R, U, and B are independent of one another. The independence assumption was expressed in the equations by letting the probability of intersection of two or three of the events being the

product of the probabilities of individual events. Logically, this stochastic independence implies independent sensitivities of R and U to experimental manipulations; an independent variable can yield a positive, a reverse, or a null effect on unconscious memory,

independent of its effect on conscious memory. In this view, the PTD procedure is to reveal rather than to affect effects of an experimental variable on estimates of R and U. Conscious memory in an implicit test measured by this procedure was either based on intentional recollection accompanied by memory awareness or based on unintentional retrieval immediately followed by memory awareness at test. Its unconscious counterpart was memory without involving retrieval intention and awareness at test and judgment stages.

The PTD procedure estimated R and U following a study condition by Equations 5 and 6. The two equations were able to separate conscious from unconscious memory that contributed to test, without being contaminated by recognition memory and guessing effects produced at the judgment stage. One way to justify these two equations for

estimation is to rewrite them as Equations 7 and 8, in which the component of (Po – B)/(1 – B) can be understood as the relative magnitude of the proportion of response with studied words that increases from baseline performance by both conscious and unconscious memories in a test to the maximal proportion that remains to be increased from baseline performance to 1.0 in the test. Subtracting either U in Equation 7 or R in Equation 8 from (Po – B)/(1 – B) is thus either the proportion contributed by conscious memory or that by unconscious memory. The use of Equations 5 and 6 was also justified by showing that the result of Experiment 1 using the implicit stem-completion test revealed a pattern of LOP and generation effects on the two components of memory similar to that found in previous

studies using explicit stem-completion tests (e.g., Bodner et al., 2000; Cheng, 2005; Toth et al., 1994, Experiment 2).

Effects of LOP

The variation in LOP consistently produced positive effects on estimates of R across the three implicit tests. However, in spite of the consistency, the positive LOP effect in Experiments 1 and 3 was associated with a repetition-priming effect on estimates of R following shallow processing, but that in Experiment 2 was not associated with such effect. These results suggest that whether or not conscious memory of studied words benefits from perceptual encoding of the words depends on the test used.

Superficially, LOP consistently produced no effect on U estimates in Experiments 1 and 2. However, the null LOP effect on U estimates in Experiment 1 was due to an equal amount of repetition-priming effects under shallow and deep conditions, whereas that in Experiment 2 was due to the absence of a repetition-priming effect under each of the two study conditions. These results suggest that whether or not estimates of U benefit from perceptual encoding in shallow and deep conditions is also determined by the test used. Effects of LOP on U estimates in the word-identification test (Experiment 3) showed a reverse pattern; the estimate under the shallow condition was larger than that under the deep condition. This reverse effect was a combined result of a repetition-priming effect under the shallow condition and a repetition-inhibition effect under the deep condition (evidenced by a lower performance under deep than under “nonstudy” conditions).

LOP produced a positive effect on overall memory either indicated by Po scores or indicated by the sum of R + (1 – R)U in the tests of stem completion and word association but no effect on the overall memory in the test of word identification. This result suggests

that influences of LOP on memory in an implicit test can be understood in terms of its componential effects on the conscious and unconscious components of memory in the test. In Experiments 1 and 2, the positive effect of LOP on overall performance can be regarded as the union of a positive LOP effect on R estimates and a null LOP effect on U estimates. In Experiment 3, the positive LOP effect on R estimates and the reverse LOP effect on U estimates may have offset each other to result in the overall null effect of LOP in the word-identification test.

Effects of Generation

Experiments 1, 2, and 4 consistently showed positive effects of generation on estimates of R, whether or not accompanied by a repetition-priming effect under shallow conditions. Thus, the positive generation effect on R estimates was robust across the three implicit tests. On the other hand, generation produced reverse effects in Experiments 1 and 4 but no effect in Experiment 2 on estimates of U. In spite of this inconsistency, the three experiments consistently showed a repetition-inhibition effect on estimates of U under generation conditions, evidenced by a lower performance under generation than under “nonstudy” conditions. The reverse generation effect on U estimates shown in Experiments 1 and 4 was due to a combination of a repetition-inhibition effect under generation and a repetition-priming effect under shallow conditions on U estimates in the two experiments. The null effect of generation on U estimates in Experiment 2 was due to a

repetition-inhibition effect under generation and a null repetition-priming effect under shallow processing, which produced two U estimates being too small to reveal a difference between the two.

was found in Bodner et al. (2000), who compared performance on an explicit

stem-completion test following three encoding tasks: read, association, and generation. In the read task, participants simply read study words aloud. In the association task, the participants read each study word aloud and then reported aloud the first associated word that came to mind. In the generation task, the participants generated study words upon seeing generation cues and read the words aloud. By using the process-dissociation procedure, they found that estimates of conscious memory were larger, but those of automatic memory (including unconscious memory) were smaller, under the association and generation tasks than under the read task. Thus, both association and generation produced a reverse effect on automatic memory. They offered an interpretation for these reverse effects as follows: If conceptual encoding makes awareness of prior occurrence more likely and if participants use a generate-recognize strategy even when given

direct-retrieval instructions, then adding a conceptual processing component to an encoding task that is primarily data-driven in association and generation conditions should attenuate automatic memory because participants would become more sensitive to prior occurrence of studied words. However, as shown in Experiment 1 and in Toth et al. (1994), deep processing of study words, that also added a conceptual processing component to a perceptual encoding task, augmented rather than attenuated automatic or unconscious memory. Further, Bodner et al. failed to note that these reverse effects of generation and association were in fact largely contributed by a repetition-inhibition effect on automatic memory under generation and association conditions shown in their study.

--- Insert Table 7 about here

---

Generation produced a positive effect in Experiment 2 but no effect in Experiments 1 and 4 on overall memory indicated either by Po or by R + (1 – R)U scores. Thus, like effects of LOP, effects of generation on overall memory in a test are also predictable in terms of its effects on estimates of R and U in the test. In Experiments 1 and 4, the positive generation effect on estimates of R and the reverse generation effect on estimates of U may have offset each other to result in the null effect of generation on overall memory. In Experiment 2, the positive generation effect on estimates of R and the null generation effect on estimates of U may have joined to result in the positive generation effect on overall memory.

A Componential Versus Holistic Transfer Approach to Conscious and Unconscious Memories

Table 7 summarizes results of the four experiments with respect to repetition effects on conscious and unconscious memories following shallow, deep, and generation

conditions. The “+” and “0”signs were the same as used in Table 2, denoting a positive and a null repetition-priming effect, respectively. The “-” sign stands for a repetition-inhibition effect. The results cannot be fully accommodated by the TAP theory or its variant shown in Table 2, particularly those showing (1) different effects of LOP and generation on

conscious and unconscious memories in different implicit tests, (2) positive

repetition-priming effects on R estimates following shallow processing in the tests of stem completion and word identification, (3) zero repetition-priming effects on U estimates following shallow processing in the test of word association, and (4) repetition-inhibition effects on U estimates following generation across the three implicit tests. Experiment 1 showed that conscious memory in the stem-completion test benefited from shallow

conditions despite the fact that conceptual processing involved in conscious memory did not match perceptual processing involved in shallow conditions. Experiment 2 showed that unconscious memory, which is perceptually driven, did not benefit from perceptual

encoding in shallow conditions. In addition, the repetition-inhibition effect on unconscious memory under generation across the three implicit tests and under deep conditions in the word-identification test do not support the prediction of the TAP theory that unconscious memory would benefit from perceptual encoding involved in deep processing and would not benefit from generation. These results and others shown in the present study suggest a need to approach conscious and unconscious memories based on different study-test interactions in the three implicit tests.

Cheng (2003; 2005) proposed a componential versus holistic transfer approach to conscious and unconscious memories within a test (either explicit or implicit) based on study-test interaction. According to the approach, shallow processing of study words is regarded as involving perceptual encoding of the words (Ps). Deep processing of the words involves both perceptual and conceptual encoding of the words (Ps-Cs) because

participants view the words while processing them deeply. Self-generation of study words involves perceptual and conceptual encoding of generation cues and conceptual encoding of the study words (Pg-Cg-Cs), without simultaneously involving perceptual encoding of the study words. The first assumption underlying the approach is concerned with the operations of conscious and unconscious memories within a test. Conscious memory of a studied word is either perceptually or conceptually cued because certain visual (or auditory) stimuli presented at test can be used as cues to trigger both perceptually and semantically stored information, and conceptual cues can be used to trigger semantically stored

information, about the studied word under intentional retrieval. On the other hand, unconscious memory of a studied word is cued by an information gestalt formed by perceptual and/or conceptual information provided at test. For example, a visual stem presented for completion under intentional retrieval can be used as a cue to trigger either perceptually or semantically stored information about the early occurrence of the study word from which the stem was originated, whereas the same visual stem triggers certain perceptual and conceptual information, which form an information gestalt for cueing the information Gestalt formed by encoded information at study. This type of cueing processes can occur mechanically, resulting in unconscious memory of studied words.

The second assumption is that being componentially cued by perceptual and

conceptual information, conscious memory within a test benefits from either perceptual or conceptual encoding of study words depending upon the number of processing modes matched between the test and the study. For example, the test of stem completion, which is regarded as perceptually and conceptually driven, matches shallow conditions on the component of perceptual processing, matches deep conditions on the components of perceptual and conceptual processing, and matches generation on the component of conceptual processing. Thus, conscious memory within the test benefits from perceptual encoding in shallow conditions, from both perceptual and conceptual encoding in deep conditions, and from conceptual encoding in generation. These different beneficial effects are taken as bases for the result shown in Cheng (2003; 2005) that LOP and generation produced positive effects on estimates of conscious memory within a stem-completion test (either explicit or implicit).

unconscious memory of a studied word either benefits from stimulus encoding at study to the extent that the stimulus encoding at test and at study form the same information Gestalt or suffers inhibition from stimulus encoding at study to the extent that the stimulus

encoding at study and at test form different information Gestalten. The inhibition aspect of this assumption is based on the reasoning that if stimulus encoding at study and at test form different information Gestalten, then nominally identical information triggered at test provides wrong access to, and hence results in an inhibition to cue, the early occurrence of its copy at study. This idea is expanded from the encoding-specificity hypothesis proposed by Tulving and Thomson (Thomson & Tulving, 1970; Tulving & Thomson, 1971), who proposed, “Nominally identical input and test items may sometimes be encoded differently because of their different cognitive environments, with the consequence that the ‘old’ test item as a retrieval cue fails to provide access to the stored information about the earlier occurrence of its copy”. However, the inhibition aspect of the third assumption applies only to mechanically-cued unconscious memory but not to conscious memory because we failed to find a repetition-inhibition effect on conscious memory either within an explicit test (Cheng, 2005) or within an implicit test (Cheng, 2003), including the present experiments.

According to the third assumption, unconscious memory in the test of stem

completion benefits from stimulus encoding in shallow and deep conditions, because the test is both perceptually and conceptually driven, perceptual and conceptual information triggered by a visual stem form an information Gestalt, which is said to be the same Gestalt formed by either perceptual or perceptual-conceptual information about the studied word from which the stem was originated. On the contrary, unconscious memory in the test suffers repetition-inhibition effects from generation, because generation involves perceptual

and conceptual encoding of generation cues and conceptual encoding of study words, without simultaneously involving perceptual encoding of the study words, which constitute an information Gestalt quite different from the one formed at test. This accounts for the finding of Cheng (2003; 2005) that LOP produced no effect and generation produced a reverse effect, associated with a repetition-priming effect following shallow and deep processing and a repetition-inhibition effect following generation, on estimates of

unconscious (or automatic) memory within a test. In the same vein, the repetition-inhibition effect on stem completion following an association task (in which participants read each study word aloud and then reported aloud the first associated word that came to mind) found in Bodner et al (2000) can also be understood with reference to the inhibition aspect of the third assumption. The association task involves both perceptual and conceptual encoding of study words and conceptual encoding of an associated word, which would also constitute an information Gestalt different from the one provided at the test. Compared with simple reading of study words that produces no repetition effect on unconscious memory, generation of study words decreases unconscious memory.

The approach mentioned above is also able to explain the results shown in

Experiment 1 because the results of Experiment 1 are the same as found in Cheng (2003; 2005). The approach can be expanded to account for the results shown in Experiments 2, 3, and 4 by further assuming that different implicit tests are differently driven (Assumption 4). The implicit test of stem completion is both perceptually and conceptually driven (Ps’-Cs’) because the generation of words for stem completion should be guided by lexical-semantic knowledge as well as by the visual stems provided. The implicit test of word association is conceptually driven (Ca) because the generation of a word in the test for association is