Memory & Cognition 1989. 17 (5), 563-571 Generation enhances semantic processing? The role ofdistinctiveness in the generation effect SACHIKO KINOSHITA University of New South Wales, Kensington, New South Wales, Australia The goal of the present study was to investigate the locus of the memory advantage for words that are generated according to a nonsemantic rule (letter transposition) over words that are presented intact (read words). In the first two experiments, a category instance generation task was used to test the possibility that the semantic features of generated words are more readily available than those of read words. This possibility was not supported. In Experiment 3, genera­ tion effects were found to depend on the level of meaningfulness of words in recall, but not in recognition. In Experiment 4, a between-list design eliminated the generation effect found in recall, but did not affect the generation effect in recognition. Taken together, these findings suggest that generating a target according to a letter transposition rule enhances the distinctiveness of the word along a nonsemantic dimension. The term generation effect refers to the memory ad­ of a biasing semantic context only the dominant meaning vantage found with verbal materials generated by a sub­ of a homograph is encoded (Simpson, 1981), McElroy ject, as opposed to verbal materials presented to the reasoned that only the dominant meaning of the homo­ subject intact by an experimenter. The subjects in ex­ graph target would be encoded in the presence of a rhyme perimental investigations of this effect typically are cue. Also, since the meaning of the word cannot be known presented with a rule and a cue for generating the target until the word is generated, any semantic processing (e.g., synonym: rapid-F for FAST), and the retention must occur after that. McElroy supposed that if no post­ ofthe generated items is compared with that for the items generation semantic processing occurred, then the that are presented intact. Generation effects have been dominant meaning of the target must be encoded to the reported with free recall, with cued recall and recogni­ same extent for read and generated targets, and that there­ tion, and with different rules (e.g., synonym, rhyme) for fore no generation effect should be found when recall is generating the target (Slamecka & Graf, 1978). cued with words that are related to the dominant mean­ Recently, McElroy (1987) suggested the possibility that ing. However, McElroy found a recall advantage for generated words may receive greater semantic process­ generated words over read words when recall was cued ing than read words even when the generation rule does with words related to the dominant meaning, yet not when not involve semantic processing. Such a view has an im­ recall was cued with words related to the subordinate portant implication for the finding that nonwords (which, meaning of the homographs. She concluded, therefore, by definition, must be generated according to a non­ that generated words received greater semantic process­ semantic rule) do not yield generation effects (e.g., see ing than read words after the (dominant) meaning was Nairne, Pusen, & Widner, 1985; Nairne & Widner, 1987; encoded. Payne, Neely, & Bums, 1986). That is, generation ef­ It may be pointed out that this is not the only inter­ fects may not be obtained with nonwords, because seman­ pretation open to the pattern of results reported by Me­ tic processing of nonwords is impossible. Elroy (1987). In particular, it can be suggested that the In McElroy's (1987) first experiment, subjects read or generation process in McElroy's experiment resulted in generated homograph targets in response to rhyme cues enhanced distinctiveness of the target, but that this en­ (e.g., luck-duck). Subsequently, the subjects were given hanced distinctiveness may not have been mediated by extralist cues that were related to either the dominant or greater semantic processing. This alternative interpreta­ the subordinate meaning of the homograph targets (e.g. , tion is described below. swan or avoid). On the assumption that in the absence It .is generally assumed that there are two stages in recall: an initial reconstruction process and a subsequent discrimination process (e.g., see Bahrick, 1970; Hunt & This research was supported by a University of New South Wales Elliott, 1980). In the reconstruction process, a set ofcan­ postdoctoral fellowship to the author. I am grateful to the editor, Alice didate items is generated on the basis of whatever retrieval Healy, and to Fergus Craik, Reed Hunt, and an anonymous reviewer, cues are available. The candidate items are then subjected for helpful comments on an earlier version of the article. Correspon­ dence should be addressed to Sachiko Kinoshita. Department of Psy­ to a discrimination process. For familiar items such as chology, UniversityofWollongong, P.O. Box 1144, Wollongong, NSW words, the discrimination process is expected to involve 2500, Australia. the context in which the item was encoded (i.e., from 563 Copyright 1989 Psychonomic Society, Inc. 564 KINOSHITA previousextraexperimental encodings). Given that in pre­ words may be more readily available, thus increasing the viousextraexperimental encounters, wordsare morelikely opportunity for generated words to be included in a can­ to have been read than generated, it seems reasonable didate set during recall. Experiment I was designed to to assume that the "generate" condition would form a test the possibility that generating a word according to smaller overlap with previous extraexperimental encod­ a nonsemantic rule resultsin greater availability of seman­ ings than the "read" condition. In this sense, the gener­ tic features of the word than does reading a word, using ation process may be said to enhance the distinctiveness an implicit test of memory. of the target. In an implicit memory test, noexplicitreferenceis made In McElroy's (1987) first experiment, recall was found to the specific episode in which the material has been to be better overall (i.e., for both read and generated tar­ learned (Schacter, 1987). For example, subjects may be gets) when recallcues were relatedto the dominantmean­ presented with a stem of a word (e.g., win-), and asked ings of target homographs than when they were related to write down the first word that comes to mind that con­ to the subordinate meanings. This effect of cue type may tains the stem (e.g., winter, window). Studieshave shown be suggested to have its locus in the reconstruction that implicit tests are sensitive to a recent encounter with process: that is, a target is more likely to be included in the stimulusmaterials: for example, a word is more likely the candidate set if the retrieval context matches the en­ to be produced in a word stem completion task if it has coding context (cf. Wickens, 1987). If in the absence of been presented recently. a biasing semanticcontext only the dominant meaning of Experiment I was madeto consistof a categoryinstance a homograph is encoded, the likelihood that a homograph generation task, in order to test the notion that generat­ target has been included in the set of candidate items ing a target according to a nonsemantic rule nevertheless should be greater when the recall cue is related to the involvesgreater semanticprocessing. In sucha task, sub­ dominant meaning than when it is related to the sub­ jects are presented with category names (e.g.,fumiture), ordinate meaning of the target. In addition, if generation and they are asked to produce instances of the category enhances distinctiveness of the target, as stated earlier, (e.g., chair). In a previous study that included this task this information can be used to facilitate the discrimina­ (Graf, Shimamura, & Squire, 1985), subjects (both am­ tion process. It is to be expected, therefore, that a gener­ nesicand normal) were found to be more likelyto respond ation effect should be observed when recall cues are with words that had been semantically processed earlier, related to the dominant meaning of the homograph tar­ relative to words that had not been presented (the sub­ gets. However, when recall cues are related to the sub­ jects rated how much they liked each word). This find­ ordinate meaning, the likelihood of targets being included ing may be taken to indicate that the category instance in the set of candidate items should be low. This should generation task is sensitive to semantic processing: be­ preempt the opportunity for distinctiveness to exert its ef­ cause the semantic features of words are more readily fect, and therefore little generation effect should be ex­ available whenthe wordshavejust been processed seman­ pected in this case. It can be seen that the interaction be­ tically, these words should be more likely to be produced tween the effects of generation and cue type reported by in response to category names. From this, it may be ex­ McElroy does not have to be interpreted in terms of post­ pected that if generated words receive greater semantic generation semantic processing. processing than read words, then generated words will The aim in the present study was to investigate the lo­ be more likely to be produced in the category instance cus of the memory advantage for words that are gener­ generation task. ated accordingto a nonsemantic rule, as opposedto words that are presented intact. The initial aim was to test Method whether the memory advantage for generated words is Subjects. Thirty introductory psychology students at the Univer­ due to a greater semantic processing, as suggested by sity of New South Wales participated in Experiment I, in return McElroy (1987). Experiment I was conducted to test for course credits. Half of the subjects were tested for recall, and the other half were tested in the category instance generation task. the possibility that generating words according to a non­ Materials.