Does the generation effect occur for pictures?

The generation effect is the finding that self-generated stimuli are recalled and recognized better than read stimuli. The effect has been demonstrated primarily with words. This article examines the effect for pictures in two experiments: Subjects named complete pictures (name condition) and fragmented pictures (generation condition). In Experiment 1, memory was tested in 3 explicit tasks: free recall, yes/no recognition, and a source-monitoring task on whether each picture was complete or fragmented (the complete/incomplete task). The generation effect was found for all 3 tasks. However, in the recognition and source-monitoring tasks, the generation effect was observed only in the generation condition. We hypothesized that absence of the effect in the name condition was due to the sensory or process match effect between study and test pictures and the superior identification of pictures in the name condition. Therefore, stimuli were changed from pictures to their names in Experiment 2. Memory was tested in the recognition task, complete/incomplete task, and second source-monitoring task (success/failure) on whether each picture had been identified successfully. The generation effect was observed for all 3 tasks. These results suggest that memory of structural and semantic characteristics and of success in identification of generated pictures may contribute to the generation effect.

The phenomenon in which self-generated words are recalled and recognized better than read words is widely known as the generation effect (Slamecka & Graf, 1978). Across five experiments, Slamecka and Graf (1978) consistently demonstrated that a generated word was remembered better than a read word under a variety of generation conditions (e.g., generating a semantic associate, a categorical associate, an antonym, a synonym, or a rhyme to a stimulus word), and under a variety of test conditions (cued and uncued recognition, free and cued recall, and confidence ratings in recognition). Since then, a number of studies have replicated the phenomenon with additional generation conditions. Some studies asked subjects to fill in missing letters to generate words with various generating rules: word fragment completion with semantic cues (e.g., "political killer ASSA___N" Gardiner, 1988), word fragment completion with and without cue words (Rabinowitz, 1990), and word stem completion with high or low information (e.g. , "pursue-av_ _d" for high-information condition vs. "pursue-a_ _ _ _"for low-information condition; Slamecka & Fervreiski, 1983). McFarland, Frey, and Rhodes (1980) found the effect by asking subjects to fill in a missing word to generate a sentence to make sense and to not make sense. Snodgrass and Kinjo (1998) found equal generation effects in free recall when subjects attempted to name either very incomplete or moderately incomplete words during study, compared to when they named to identify complete words.

Although words have been the dominant stimulus medium, the generation effect has been demonstrated for other media. For example, Crutcher and Healy (1989) and Gardiner and Rowley (1984) tested memory for simple mathematical operations (multiplication). They found better memory for self-generated answers in both generating and answer verification conditions than for experimenter-provided answers. Although pictures have been one of the most dominant stimuli in experimental psychology for decades, surprisingly only two studies have tested the generation effect for pictures. Peynircioglu (1989) was the first to report a generation effect for pictures. She found that drawing to-be-remembered pictures of objects or scenes from descriptions during study led to better recall than either copying down objects or scenes or looking at them. She also found superior memory for drawn (generated) nonsense pictures than for copied or simply observed nonsense figures. Pring, Freestone, and Katan (1990) tested a generation eff ect with blind and sighted children. In the generation condition, subjects were asked to touch and identify raised shape pictures and were given an associated cue word auditorily. In the name condition, subjects were asked to touch each raised shape picture but were told the name of the picture along with the cue word. Because children with normal sight were asked to wear a blindfold, there was no visual information for either group. They found a generation effect for sighted children (and the reverse effect for blind children).

There are several explanations for the generation effect and these are not mutually exclusive: a semantic activation theory, a cognitive operation theory, and factor theories. According to the semantic activation theory, the generation effect is attributed to the greater activation of an item's semantic attributes under generation than under reading. In support of this view, some studies have found that nonsemantic stimuli do not produce the effect. For example, no generation effect was found for nonwords that were generated by the use of formal letter-transposition rules (McElroy & Slamecka, 1982), for anomalous sentences (Graf, 1980), and for meaningless bigrams (Gardiner & Hampton, 1985). In another line of evidence supporting this view, Slamecka and Fervreiski (1983) found a generation effect even when subjects failed to complete a word (the try-to-generate effect) and attributed their result to retention enhancement for both generated and try-to-generate items due to deeper semantic processing. However, other studies have demonstrated the generation effect for nonwords (Johns & Swanson, 1988; Nairne & Widner, 1987) and nonsense pictures (Peynircioglu, 1989). Furthermore, both Slamecka and Graf (1978) and McFarland et al. (1980) found equivalent generation effects across various levels of encoding (semantic vs. phonemic). One would expect a bigger generation effect for items that are encoded semantically than for items that are encoded phonemically if semantic activation is the main source of the generation effect. Thus, these latter studies have called into question the semantic activation theory.

According to the cognitive operation theory, the generation effect is attributed not specifically to semantic activation but to the process of generation itself. In support of this view, Jacoby (1978) reported that increasing the effort required to solve a problem enhances later memory performance. He found that a difficult generation condition in which subjects were asked to fill in two missing letters of each stimulus word given a cue word produced better cued-recall performance than an easy generation condition in which subjects were asked to fill in a single missing letter of each stimulus word given a cue word. Griffith (1976) and McFarland et al. (1980) also argued that the generation effect is caused by a greater number of cognitive or mental operations. However, these studies did not specify which types of cognitive operations were important to the generation effect. In addition, they did not exclude the possibility that extra semantic activation could contribute to the generation effect. In fact, Cr utcher and Healy (1989) argued that the two theories can be seen as components of the same theory. That is, both extra semantic activation and the effort to generate items are parts of the auxiliary mental processes carried out by subjects.

These two theories are not compelling explanations for the generation effect. In addition, it seems impossible to explain the generation effect with only one factor, such as semantic activation. Some researchers have proposed multiple-factor theories for the phenomenon (Hirshman & Bjork, 1988; McDaniel, Riegler, & Waddill, 1990). In Hirshman and Bjork's two-factor theory, the generation effect is attributed to two factors: the activation or strengthening of item-specific features in memory (item-specific factor) and the activation or strengthening of the relationship between a stimulus and a response (relational factor). The former factor is required mainly to explain the generation effect for free recall and the latter factor is required to explain the generation effect for cued recall. McDaniel et al. (1990) proposed a three-factor theory for the generation effect. According to their theory, Factor 1 is item-specific information and Factor 2 is relational information between a stimulus and its response. Th ese two factors are comparable to Hirshman and Bjork's (1988) item-specific and relational factors, respectively. The theory also has a third factor, whole-list information; that is, contextual information created by items in the list, as when the list is composed of targets from the same category. In light of this theory, the semantic activation theory could relate to the activation of item-specific features or the activation of whole-list information. McDaniel et al.'s theory seems to be the most comprehensive theory of the generation effect.

This study aimed to investigate two questions. One was whether a generation effect occurs for pictures. Based on previous reports of the generation effect for pictorial stimuli (Peynircioglu, 1989; Pring et al., 1990) and a pilot study, we predicted that a generation effect would occur for pictures. The next question was what factors affect the generation effect for pictures and whether these factors differ for pictures compared to words. Pictures are both recalled and recognized better than words, and researchers have attributed this picture superiority effect mainly to two sources: the superior sensory code of pictures (Nelson, Reed, & McEvoy, 1977; Paivio, 1971) and the superior conceptual code of pictures (Roediger & Blaxton, 1987; Roediger, Weldon, & Challis, 1989). Thus, it is possible that the generation effect with pictures could be produced by greater activation of the item's sensory features or by greater activation of the item's semantic features.