Infants' symbolic comprehension of actions modeled with toy replicas
Abstract
While very young children's understanding of objects as symbols for other entities has been the focus of much investigation, very little is known concerning the emergence of comprehension for symbolic relations among actions modeled with toy replicas and their real counterparts. We used videotaped depictions of real actions in a preferential looking task to assess toddlers’ ability to comprehend such connections for action categories aligned with familiar object concepts. Across two experiments, 16- and 18-month-olds provided no evidence of understanding such relations, even when action categories were highlighted with verbal prompts. Among 24- and 26-month-olds, comprehension of relations between certain actions modeled with toys and videos of their real-world counterparts began to emerge, independent of expressive vocabulary size. Implications of our results for theoretical conclusions drawn from use of the generalized imitation procedure to study early conceptual development are discussed.
Infants and toddlers navigate through a remarkably complex world of objects. Not surprisingly, young children's first words frequently are names for things – objects that can be acted on, or objects that are particularly familiar (Gentner, 1978, Mervis, Mervis, Johnson & Bertrand, 1992; Nelson, 1973a). At the same time, objects that comprise the infants’ world provide opportunities for learning about action categories. Parents model such actions during play as they ‘gallop’ toy horses across rugs or put dolls to ‘sleep’ in toy cradles. While the objects engaged in such interesting events are named early and often by most toddlers, it is less clear how very young children interpret the actions that are modeled during play, and whether such understanding might mediate early verb learning. The goal of the present research was to begin to explore toddlers’ understanding of symbolic relations between actions modeled with toy replicas and videotaped instances of their ‘real-world’ counterparts.
Several theories of conceptual development highlight the centrality of object movement and relational action in infants’ developing representations of things. Nelson's theory of functional core concepts (Nelson, 1973b, 1974) and her subsequent account of slot-filler categories derived from event schemas (Lucariello, Kyratzis & Nelson, 1992; Nelson, 2000) both emphasize that categories of objects are constrained by functions that are characteristic of particular events. More recently, Mandler has argued that object concepts develop from image schemas– abstract spatial representations consisting of dynamic information about the paths objects take, their onsets and endpoints, and their relations with other objects (Mandler, 1992, 1998, 2000a). Image schemas are derived from a process of perceptual analysis that entails abstraction and redescription of information about how objects move and the events in which objects engage. Across these theories, infants’ attention to movement and action pave the way for learning about the objects engaged in dynamic events. Understanding what things do is the foundation for understanding what things are.
Despite the salience of dynamic information to theories of conceptual representation, researchers examining infants’ categorization of real-world objects typically have relied upon stimuli for which such information is absent. Visual familiarization and habituation tasks typically use two-dimensional drawings or photographs of static objects to represent category exemplars. When manipulable exemplars from domains such as animals or vehicles are selected for use in categorization tasks such as object examining and sequential touching, investigators typically rely on small models of their real-world counterparts given that such models are considerably less unwieldy (and more predictable) than their living or moving counterparts. Generalized imitation is another categorization task that uses small replicas of objects, and it provides much of the impetus for the present research. In generalized imitation, the infant's imitation of dynamic action features is central to evaluating the development of object concepts (e.g. Mandler, 2000a; Mandler & McDonough, 1996, 1998, 2000; Rakison & Poulin-Dubois, 2000).
The generalized imitation procedure is grounded in decades of research supporting the notion that infant imitation can be used to measure representational abilities (Fenson & Ramsay, 1981; Flavell, 1985; Hayne, Boniface & Barr, 2000; Herbert & Hayne, 2000; Killen & Uzgiris, 1981; Meltzoff, 1985, 1988). The task begins with the experimenter modeling a target action with a miniature object and prop (e.g. drinking modeled with a toy dog positioned as if to drink from a tiny cup). The experimenter generally produces vocalizations during modeling that are aligned with the action category (e.g. ‘Sip, sip, mmm good’). The toy object then is removed and a pair of test exemplars is placed simultaneously to the right and left of the infant. The infant is handed the prop (e.g. cup) and generalized imitation is assessed through observing which test exemplar the infant selects to imitate the modeled action. By experimentally manipulating the taxonomic relatedness of the pair of test exemplars, the scope of the category that constrains the infant's generalization of the action can be inferred. Mandler and McDonough (1996, 1998; McDonough & Mandler, 1998) report that 9- to 19-month-olds, (1) generalize domain-specific actions across basic-level category boundaries, though not across domain boundaries (e.g. rabbits, cats and birds – but not cars – will be given drinks with cups), (2) generalize neutral actions across domain boundaries (e.g. animals and vehicles will both be inserted into buildings) and (3) fail to imitate actions modeled with inappropriate objects (e.g. toy planes will not be ‘given drinks’, but see Rakison, 2003, for evidence that infants’ imitation is not always dependent on conceptual knowledge). Mandler and McDonough interpret their findings as providing evidence that infants implicitly constrain inferences based on global conceptual representations corresponding to broad categories such as animals and vehicles rather than the basic-level categories (e.g. dog, car) that constrain the inductions of 2- and 3-year-olds (Gelman & Coley, 1990; Johnson, Scott & Mervis, 1997).
To accept that infants’ imitation of modeled actions is guided by the activation of conceptual representations akin to animal and vehicle (particularly to the extent that such early concepts are based on analyses of animate versus inanimate motion patterns of real object kinds; Mandler, 1992, 1998), it seems important to assess the degree to which infants understand that actions modeled with inanimate toy replicas are related to corresponding real instances of those actions. If infants do not relate the activity of putting a toy cup to the mouth of a toy dog to model drinking to instances of real dogs drinking, it is not clear that we should interpret their efforts to make toy cups interact with other toy animals as reflecting their conceptual understanding of the fact that animate beings (animals) drink, but that inanimates (vehicles) do not. Rather, it may be that infants tend to imitate adults’ actions toward objects without necessarily recruiting relevant conceptual knowledge (Rakison, 2003; Rakoczy, Tomasello & Striano, in press; Striano, Tomasello & Rochat, 2001). A theoretical alternative intermediate between the ‘conceptual representation-or-nothing’ extremes is that infants’ conceptualization of actions or movement properties may be associated with salient perceptual features of objects (Rakison & Hahn, in press). Since miniature models of dogs, rabbits, fish and birds (like their real counterparts) possess salient mouths and/or faces, infants may willingly generalize the action of drinking to all such creatures, yielding a global-level property extension. Whereas Mandler (1992, 1998; Mandler & McDonough, 1996, 1998) maintains that generalized imitations are based on global conceptualizations of real object kinds, this alternative view suggests that infants extend (or refuse to extend) dynamic properties such as ‘offering a drink’ or ‘keying’ to inanimate objects with animal- or vehicle-like surface properties.
We conducted two experiments to investigate whether infants consider modeled actions of the sort typically used in studies involving generalized imitation (Mandler & McDonough, 1996, 1998) to be symbolically related to videotaped depictions of real-world action sequences that correspond to those modeled actions. Mandler (2000b) has argued that infants must understand the models used in the generalized imitation task to represent real-world objects, since infants refuse to imitate representationally incorrect events with those models. She states, ‘This does not mean that the representational understanding is deep or sophisticated, as it must become to solve problems such as those posed by DeLoache,1 but it surely is evidence for the beginnings of later symbolic understanding’ (2000b, p. 71). In the present research, we focus on the origins of such symbolic comprehension for the action information inherent in the properties that are modeled during generalized imitation tasks.
Both experiments involved a preferential looking task in which children watched as the experimenter modeled a target action with toy replicas and then were presented with a pair of test videos, one depicting an agent engaging in a corresponding real-world action and the other featuring the same agent engaged in a qualitatively different action. Children were considered to comprehend the symbolic relation between the modeled and videotaped real actions if they looked significantly longer at the matching videotaped action. In Experiment 1, we included a range of modeled actions, some involving human agents, others not. Some actions were directed toward a goal, others were not. Some of the modeled actions were equivalent to those used by Mandler and McDonough in studies of generalized imitation, others were intended to be representative of the types of actions parents might model with toy replicas in the context of play with their infants or toddlers. In previous research utilizing preferential looking procedures, Younger and Johnson (2004) found that 18-month-olds were beginning to demonstrate symbolic competence with toy model–real exemplar object relations. By 26 months, performance in the object task was essentially ‘at ceiling’ (i.e. toddlers consistently directed their attention toward a video of the corresponding real object upon viewing a target toy replica). Thus, in Experiment 1, we selected these age groups to begin to examine toddlers’ symbolic understanding of modeled actions. Experiment 2 focused specifically on the subset of actions used to assess 14-month-olds’ generalized imitation of four properties (drinking, sleeping, starting, riding) said to be constrained by global concepts of animal and vehicle (Mandler & McDonough, 1996, 1998). Since toddlers in Experiment 1 demonstrated relatively poor comprehension of the action categories aligned with these particular four properties, we systematically tested in Experiment 2 whether or not comprehension would be facilitated by the presence of verbal prompts of the sort that accompany the modeled actions in studies using the generalized imitation procedure.
Because it was not feasible to bring live instances of the real-world action categories (e.g. drinking and sleeping dogs) into the laboratory, we used videotaped actions in our preferential looking task. Thus, an important issue to consider is toddlers’ comprehension of actions conveyed through television or videotape. Task difficulty and age are important moderators of infants’ ability to understand information portrayed through televised models (Barr & Hayne, 1999). While 24-month-olds have difficulty using information presented on television to guide their search behavior for a hidden object (Troseth, 2003; Troseth & DeLoache, 1998), infants as young as age 14 to 15 months are successful in imitating behaviors toward novel objects when presented with a videotape of an adult modeling the target behavior (Meltzoff, 1988).
The preferential looking procedure adopted in the current experiments does not require that toddlers be able to use video representations of objects or events to plan their actions. Rather, toddlers need only recognize actions presented on videotape as being related to those that occur in the real world. Several lines of evidence converge to suggest that young children respond similarly to objects and events presented ‘live’ and those presented on video. Pierroutsakos and Troseth (2003) recently reported that 9-month-olds attempt to hit and grasp objects depicted on the screen in a similar manner as might be exhibited with ‘real’ objects. Such manual exploration declined with age and was replaced by pointing and vocalizations toward the video by 15- and 19-month-olds. Mumme and Fernald (2003) reported that 12-month-olds used videotaped depictions of an actor's emotional reaction toward objects to guide their own behavior toward a set of corresponding real objects. Very young infants also react similarly when viewing contingent video of their mothers as they do during live interactions with their mothers (Murray & Trevarthen, 1985). Based on such findings, we were reasonably confident that performance on the preferential looking task could be used to draw conclusions about toddlers’ understanding of symbolic relations between actions modeled with toy replicas and their real-world counterparts.
Experiment 1
Method
Participants
Participants included 24 toddlers at each of two ages: 18 months (12 girls, 12 boys, M= 18.1 months, range = 17.5–18.6 months), and 26 months (13 girls, 11 boys, M= 26.1 months, range = 25.6–26.5 months). No children were excluded from participating. However, one 18-month-old and four 26-month-olds did not look toward either the matching or nonmatching screen on at least one test trial and their scores were based only on the trials for which there were looking data. Recruitment and testing were conducted at two Midwestern university campuses; in Indianapolis, IN (52% of sample), parents were contacted through brief articles placed in local newspapers, and by letters sent to individuals identified through a commercially available database. In West Lafayette, IN, parents were contacted first by mail and then by phone after being identified through published birth announcements. All children were full-term at birth and were reported to be developing normally, with no impairments in vision, hearing or language. The majority were Caucasian and middle class. An additional 12 undergraduate psychology majors (7 females, 5 males) participated in rating tasks for stimuli construction. The students participated for course credit in Indianapolis.
Preferential looking apparatus
The child sat on the parent's lap in a semi-darkened three-sided booth (1.8 × 1.8 × 1.2 m) facing a wooden panel concealing all but the screens of two Mitsubishi Diamond Pro 56 cm video monitors (see left side of Figure 2). There was a 56 × 61 cm opening between the two screens for a 56 × 35 cm stage that could be illuminated from above by a concealed light (controlled by a dimmer switch). The tester (wearing black) sat in a cloth-covered enclosure behind the stage. A 2.5 cm hole centered beneath the stage concealed a closed-circuit camera used to monitor the child's visual responses. The chair on which the parent sat was positioned 70 cm away from the stage – near enough for the child to be able to see all aspects of the modeled events, yet far enough to deter the child's reaching toward the objects. The parent wore a visor that prevented her from seeing the modeled actions and video screens to ensure that she did not inadvertently influence the child's looking. The experiment was controlled by Habit 2000 (Cohen, Atkinson & Chaput, 2000) software run on a Macintosh G4 system. The coder monitored the child's looking on a television monitor in the control area adjacent to the testing room and recorded the duration and direction of gaze with preprogrammed keystrokes. A different key was used to initiate the presentation of videos at the onset of each trial segment. An intercom allowed the coder to monitor verbal cues from the tester as to the timing of stimulus presentation. The coder was blind to the left–right positioning of the two video images and to the modeled target action. The tester was blind to the left–right position of the matching image across all test trials.
Schematic overview of preferential looking procedure.
Stimuli
Stimuli consisted of the objects and props used to demonstrate the target actions, and videotaped depictions of real-world action sequences corresponding to each of the modeled actions. The three-dimensional miniature replicas of objects used to model the actions (and props, if needed) are listed in the Appendix. Replicas were iconic representations of the target category, and ranged in length from 6.02 to 13.35 cm, with an average size of 9.87 cm × 5.72 cm. Props were comparable in size to the miniature objects and thus were oversized in relation to the objects (see Figure 1) to facilitate modeling.
Modeled action of feeding and still images from corresponding matching and nonmatching videotapes of ‘real’ actions.
Videos were created for the present study using a Sony Digital Handycam (DCR-TRV11). For each of four pairs of target actions used on test trials, and for two pairs of target actions used on training trials, 6-s Quicktime movies were created. Each pair of videos featured naturalistic scenes of the same real object (e.g. dog) engaged in each of the two target actions within a pair (the same dog was filmed drinking water from a bowl and lying down to sleep). The same props (e.g. the water bowl) appeared in both movies, regardless of whether they were acted upon. Three actions (drinking, rolling/bouncing training stimuli) were digitally slowed by 50% to enhance the interpretability of actions that otherwise would have occurred too quickly for toddlers to encode during a 6-s clip, and to better match pairs of videotapes in terms of their dynamic features.2 Still images from the feeding trial are presented in Figure 1.
Adult ratings of stimuli
To confirm the validity of the video stimuli, adults were tested individually or in pairs in the infant testing area. Adults viewed pairs of videotaped modeled actions and videotapes of either the matching or nonmatching counterparts and were asked to rate the similarity of each pair to verify that the intended matching videotape was indeed more perceptually similar to the modeled event than was the nonmatching video. Two quasi-random orders of the videotape pairs were created, with the constraints that (a) no more than two ‘matching’ or ‘nonmatching’ pairs were presented consecutively within the series, and (b) consecutive videos never featured events from the same action pair. Six participants were randomly assigned to each of the two orders. Participants were presented with an answer form containing a space for each of the pairs, with a Likert scale at the top of the page ranging from 1 (extremely similar) to 7 (extremely dissimilar). Participants were instructed to rate how similar they considered the modeled action and the corresponding real-word action to be. They were told not to consider how similar the objects in the videos were to each other, but to focus only on the action depicted in the video (e.g. how similar modeled feeding was either to the real-world instances of feeding or stirring).
Preferential looking procedure
Toddlers participated in two training trials and four test trials, each of which included three familiarization segments (left video only, right video only, and both videos) and two test segments (both videos). Two 6-s test segments were presented during each trial block to yield a more stable measure of comprehension, given that younger toddlers might be more easily distracted during testing or need more time to encode the modeled actions. A schematic overview of the procedure is presented in the right side of Figure 2. The left–right position of each action video in the pair was constant across the five trial segments shown to a given infant, though left–right position for each pair was counterbalanced across infants. The tester began each trial segment by turning on the light above the stage to orient the toddler's attention to the midpoint between the monitors saying, ‘What's going on here?’ while tilting her head toward the monitor to the child's left. As the light over the stage went off, the coder hit a preprogrammed key (upon hearing the cue word ‘here’) to initiate the video presentation for the left-only familiarization segment. The tester then initiated the right-only segment by saying, ‘What's going on there?’ while tilting her head toward the monitor on the child's right. The paired saliency/familiarization segment was then administered as the tester turned the light on and said, ‘What are they doing?’ Following paired familiarization, the tester placed the appropriate toy replica (and prop, if needed) on the stage and then turned on the light. The tester pointed to the replica saying, ‘Hey, watch this!’ She then modeled the target action two times (or for approximately 2 s for the continuous actions of stirring and running), pausing briefly between demonstrations. As she began to model the target action a third time she asked, ‘Where's the one doing this?’ while simultaneously dimming the light (by approximately 50%), and modeling a single, less exaggerated instance of the target action (e.g. a single rotation of the spoon in the bowl for stirring) as a memory cue during the preferential looking test segment. At the completion of the 6-s test segment, the light was turned up and the modeled action was repeated, followed by a second test segment. At the completion of the 5-segment trial, the light in the testing booth was turned off, the tester put the previous modeling object(s) away and picked up the next one(s), and the entire procedure was repeated until all test trials had been presented.
Training trials were intended to familiarize toddlers with the sequence of events involved in the procedure and to reinforce their looking toward the matching screen. The order of the two training trials (ball rolling and plane flying) was counterbalanced across children. The positions of the matching and nonmatching videos (ball bouncing and plane taxiing) were fixed within each training trial; rolling always appeared to the child's right, and flying to the left. On both training trials, the tester provided verbal reinforcement when the child looked to the matching video, or encouragement to look to the appropriate screen in the event the child failed to do so spontaneously. For the first training trial only, the tester also labeled the target action (‘Watch this! Where's the one doing this? Where's rolling/flying?’) to further increase the likelihood that the child would look to the matching action.
Each infant received four test trials with the following action pairs: feeding–stirring, running–jumping, drinking–sleeping, starting–riding. Order of presentation of the action pairs was counterbalanced across the four test trials, using a Latin Square design. The target actions were grouped into two sets, with half of the toddlers in each age group assigned randomly to each set. Thus, the target actions for 12 children at each age were feeding, running, drinking and starting, and the target actions for the remaining children were stirring, jumping, sleeping and riding. Descriptions of the means by which target actions were modeled appear in the Appendix. Motion paths always were depicted in a right-to-left direction (ending in the center of the stage). All other actions were modeled equidistant between the two video monitors.
At the end of the procedure, parents completed a brief demographic survey and the vocabulary section of the MacArthur Communicative Development Inventory (CDI, Toddler Version; Fenson, Dale, Reznick, Thal, Bates, Hartung, Pethick & Reilly, 1993).3 A score reflecting the relative proportion of verbs in the child's productive vocabulary also was derived by dividing the number of words from the Action Words category by the total number of words produced.
Coding
The question of interest was whether infants could make use of actions modeled with toy replicas to guide their attention to videos of the corresponding real actions. For online coding, the coder depressed one of two preprogrammed computer keys when the child was looking to the video image on the left or right screens, and released it when the child looked away. A second coder assessed reliability by rescoring 20% of the children's videotaped responses. For each coder, looking to the right was subtracted from looking to the left for each paired trial. Pearson correlations computed on the difference scores across trials for the two coders ranged from .95 to 1.00 (M= .98).
Results
We first contrasted adults’ similarity ratings for each of the four pairs of matching videos and each of the four pairs of nonmatching videos. Ratings were compared for each target action (modeled action vs. matching real action) vs. (modeled action vs. nonmatching real action) through paired t-tests. The results are depicted at the top of Table 1. For each of the eight target actions, ‘matching’ pairs were rated as significantly more similar than nonmatching pairs, verifying that adults clearly recognized which of the videotaped real actions corresponded to the modeled action.
| Experiment | Modeled action | Matching video mean (SD) | Nonmatching video mean (SD) | Comparison |
|---|---|---|---|---|
| 1 | Drinking | 2.42 (0.90) | 5.48 (1.88) | t(11) = 4.17, p < .01 |
| Sleeping | 4.17 (1.90) | 5.92 (1.56) | t(11) = 4.26, p < .01 | |
| Stirring | 1.42 (0.67) | 4.67 (1.83) | t(11) = 5.49, p < .001 | |
| Feeding | 1.83 (1.53) | 3.92 (1.78) | t(11) = 3.57, p < .01 | |
| Jumping | 2.33 (1.37) | 3.50 (1.83) | t(11) = 2.55, p < .05 | |
| Running | 2.33 (0.89) | 3.83 (1.40) | t(11) = 3.00, p < .05 | |
| Starting | 2.25 (1.14) | 5.67 (1.88) | t(11) = 5.73, p < .001 | |
| Riding | 2.67 (1.30) | 5.50 (1.73) | t(11) = 6.99, p < .001 | |
| 2 | Drinking | 2.92 (1.62) | 5.08 (1.62) | t(11) = 5.12, p < .001 |
| Sleeping | 3.75 (1.60) | 5.50 (1.83) | t(11) = 3.66, p < .01 | |
| Starting | 2.33 (1.67) | 5.67 (1.67) | t(11) = 5.86, p < .001 | |
| Riding | 1.92 (1.24) | 4.92 (1.56) | t(11) = 6.51, p < .001 |
- Note: 1 = extremely similar; 7 = extremely dissimilar.
We next calculated the amount of time that each child spent looking to the matching video during test segments for each trial, and subtracted from that the time that the child spent looking to the nonmatching video. These difference scores were averaged across the four test trials and compared in a 2 (age) × 2 (gender) × 2 (site) univariate analysis of variance (ANOVA). The only significant effect was a main effect of age, F(1, 40) = 6.55, p= .01, due to 26-month-olds looking longer at the matching test videos than 18-month-olds (mean difference scores= 1.52 s and −0.02 s for the 26- and 18-month-olds, respectively). Because testing site did not emerge as a significant factor, it was omitted from subsequent analyses. Overall, boys tended to produce slightly higher difference scores than girls, although the effect of gender was not significant (p= .12). Because of the trend, gender was retained as a factor in subsequent analyses.
We next compared the amount of looking to the matching versus nonmatching videos in the test segments to looking towards the would-be match versus nonmatch during paired familiarization trials. The rationale for this comparison was to determine whether the pattern of looking following the experimenter's modeling of the target action differed from the pattern of looking that occurred spontaneously during familiarization. Prior to the paired familiarization segment, the child had already seen each video presented in isolation in the same location (in the left-only then right-only segments). Thus, the direction in which the child preferred to look during paired familiarization potentially reflected the child's a priori preference for particular real-world actions. To the degree that looking to the matching video on test differed from looking to the would-be match during familiarization, we were confident that the child was selecting the video on test based on its relation to the modeled action.
Because of the age effect reported above, separate 2 (gender) × 4 (action pair) × 2 (trial segment: familiarization versus test) mixed ANOVAS, with both action pair and trial segment as repeated measures, were conducted for each age group. For 18-month-olds, no significant effects emerged, suggesting that looking on test segments did not differ appreciably from looking on paired familiarization trials for any of the action pairs. For 26-month-olds, there was only a significant main effect of trial segment, F(1, 18) = 10.94, p < .01, indicating that the difference between children's looking to the matching versus nonmatching video was greater following modeling of the target actions than it had been during paired familiarization trials. In the remainder of the Results, we focus only on the 26-month-olds’ data since it is clear that 18-month-olds showed no evidence of comprehending relations between instances of action categories that were modeled with toy replicas and videos of their real-world counterparts.
To further examine the effect of trial segment, separate mixed 2 (target action) × 2 (trial segment: familiarization versus test) ANOVAs, with trial segment as a repeated measure, were conducted for each action pair. Inclusion of target action as a between-subjects factor enabled us to test whether the trial segment effects were driven by particular modeled actions that were perhaps easier to map between the toy and real instances, or by preferences for certain target videos (assessed during familiarization trials). Comparisons to chance (defined as zero, since equal looking to the matching and nonmatching screens would result in a difference score of zero) are reported below in conjunction with the results of the ANOVA for each action pair. Means and standard deviations associated with these chance comparisons are also depicted in Table 2 (along with comparable results from the 18-month-olds).
| Action pair | 18-month-olds | 26-month-olds | ||
|---|---|---|---|---|
| Familiarization | Test | Familiarization | Test | |
| Stir/Feed | −0.07 (3.29) | 0.53 (4.24) | −0.86 (2.62) | 1.84 (2.90) |
| Run/Jump | 0.40 (3.07) | 0.84 (4.77) | 0.19 (2.91) | 1.69 (2.44) |
| Drink/Sleep | 0.03 (2.18) | −0.44 (3.00) | −0.05 (1.98) | 0.33 (2.53) |
| Start/Ride | −0.38 (4.18) | −1.03 (4.41) | −0.24 (3.60) | 1.76 (4.59) |
- Note: Mean difference scores represent looking time to matching video minus looking time to nonmatching video.
For starting–riding, there was only a significant effect of target action, F(1, 21) = 7.33, p= .01, reflecting the fact that 26-month-olds tended to look longer towards the starting video than the riding video across both familiarization and test segments.4 Difference scores at test did not differ significantly from chance, and the trend toward positive looking scores on test trials depicted in Table 2 is due to saliency effects. For stirring–feeding there was a significant effect of trial segment, F(1, 22) = 9.27, p < .01. Difference scores at test were also significantly greater than chance, t(47) = 2.25, p < .05. For running–jumping there was a nonsignificant trend for 26-month-olds to match the modeled action to the ‘real’ counterpart on both of these action pairs (trial effect: F(1, 19) = 2.49, p= .13). Difference scores at test for running–jumping were, however, significantly greater than chance, t(44) = 2.15, p < .05. Finally, for drinking–sleeping there was a significant trial segment × target action interaction, F(1, 21) = 4.19, p= .05, attributable to toddlers tending to show a greater difference in looking between familiarization and test trials for drinking than for sleeping. However, difference scores at test for both actions were at chance, suggesting that this was a spurious result.
While 26-month-olds clearly were more adept at comprehending the symbolic relations between actions modeled with toy replicas and videos of their real-world counterparts than were 18-month-olds, their ability was fragile and apt to fluctuate considerably across trials. Finally, there were no significant correlations between the difference in children's looking time on test trials and either the number of words in their productive vocabulary (r= .17) or the relative proportion of action words (calculated by dividing the number of action words checked on the CDI by the total number of words produced; r= .15), suggesting that the ability to match videotaped real-world instances of actions with their modeled counterparts is not initially mediated by language.
Experiment 2
In the first experiment, we saw little evidence that the modeled actions included in Mandler and McDonough's (1996, 1998) generalized imitation studies were comprehended in relation to videotaped real-world referents even by 26-month-olds, despite the fact that these children were 12 months older than the 14-month-olds they studied. It is possible that younger infants’ performance on generalized imitation tasks is largely supported by verbalizations that the experimenter produces in conjunction with the modeled events (e.g. ‘night-night’ said as the dog is patted while lying in the toy bed). Such verbalizations did not accompany the modeled actions in our preferential looking task. It also may have been the case that toddlers in Experiment 1 were overwhelmed by the presentation of four distinct modeled actions over the course of the procedure, or that the delay between the initial training trials and the last test trials may have been too long. Experiment 2 was conducted to address these possibilities. First, the presence or absence of supporting verbalizations was treated as a randomized between-groups factor. Second, we focused exclusively on the two pairs of actions drawn from Mandler and McDonough's research (drinking–sleeping and starting–riding). During testing, each pair was repeated twice, in separate trial blocks. Trial blocks were each preceded by a training trial with feedback (feeding, accompanied by supportive verbalizations) in order to remind children that they were to look toward the matching video. Because these modifications were expected to simplify the procedures in relation to those used in Experiment 1, slightly younger groups of infants (16- and 24-month-olds) were recruited.
Method
Participants
Sixty-three children (27 males, 36 females) participated. Thirty-two 24-month-olds (M= 24.1 months, range = 23.5–24.8 months, 13 males) and 31 16-month-olds (M= 16.1 months, range = 15.6–16.5 months, 14 males) were recruited through the same procedures described for Experiment 1 (61% of the sample was tested in Indianapolis). All participants were full-term at birth and were developing normally, with no known impairments to hearing, vision or language. Most were Caucasian and from middle-class homes. Four additional 16-month-olds and two 24-month-olds participated but were dropped due to experimenter error or equipment problems. Three additional 24-month-olds became fussy and did not complete the session. One additional 16-month-old and three 24-month-olds were excluded based on their failure to look at the videos on any of the training or test trials.5 Two 16-month-olds and two 24-month-olds failed to look at all on one test trial and their data were based only on the looking times from the other test trial. The same 12 adult raters who participated in Experiment 1 generated ratings for stimuli developed for this experiment.
Stimuli and apparatus
The preferential looking apparatus was identical to that described for Experiment 1. The stirring–feeding action pair from Experiment 1 was used during training trials. In efforts both to reduce mismatches in the saliency of paired videos and to enhance the degree of similarity between the modeled events and the videotaped actions, we created new videos for drinking–sleeping and starting–riding. The new drinking video depicted a side-view of a Welsh Terrier drinking from a dog dish, with ‘splashing’ water more clearly visible than in the original video. In sleeping, the same dog was viewed lying on its side on a sofa, stretching slightly as it slept. The new riding video (filmed on a road adjacent to a grassy field) featured a woman seated high on top of the back seat of a white convertible, waving as the car moved slowly along the road. The starting video featured the same woman using a key to unlock and open the driver's side door of a white convertible. For riding and starting, the change in video was intended to heighten the similarity of the modeled action (i.e. seating a doll in a waving pose on top of the toy car and putting a key to the side of a toy car) to the video of its ‘real’ counterpart. Still images from the original and revised riding videos are presented in Figure 3. Care was taken to match the size of the objects (dog and car) across each pair of videos, while at the same time making the actions depicted as salient as possible.
Modeled action of riding (doll waving) and matching videotapes of real-world instances of riding from Experiment 1 (child waving from inside car) and Experiment 2 (woman waving from position on car that is more similar to that depicted in modeled action).
Following the procedure specified in Experiment 1, adult raters assessed the relative degree of similarity between videos of each modeled action and videos of both the matching and nonmatching real-world actions. As depicted at the bottom of Table 1, results of the adult ratings for the revised videos were comparable to those obtained in Experiment 1. Adults clearly recognized that each modeled action was significantly more similar to the matching real action than to the nonmatching real action. As reported below in the Results, data obtained from the infant participants indicated that the new video pairs were better matched for saliency.
An additional questionnaire was created for parents to fill out in conjunction with the CDI in order to gauge more precisely children's relative levels of familiarity with the action categories. Parents indicated whether or not they owned a dog or a cat (assuming children with pets could easily observe animals drinking and sleeping). Parents also indicated whether children comprehended verbs for each of the target actions through a comprehension checklist. Finally, parents rated the child's relative level of familiarity with each of the four target actions using a 7-point Likert scale (1 = very unfamiliar; 7 = very familiar).
Procedure
Each child was tested on two test pairs –drinking–sleeping and starting–riding. Each pair was presented twice, in each of two trial blocks. Within each block of trials, order of presentation of the two action pairs was counterbalanced. The action modeled within each pair was constant across trial blocks for a given infant, though modeled action was counterbalanced across infants. Left–right position of the matching test video was counterbalanced on the first trial block and then reversed for each infant on the second trial block.
Each block of two test trials was preceded by a training trial involving feeding as the modeled action. Left–right position of the matching (feeding) and nonmatching (stirring) videos was counterbalanced on the first trial and then reversed for the second trial. Modeling of feeding always was accompanied by supporting verbalizations. The tester raised the spoon to the doll's lips and ‘fed’ the doll once, then repeated that action a second time saying, ‘Open wide . . . yum yum!’ The tester then asked, ‘Where's the one doing this?’ as she ‘fed’ the doll with the spoon a third time and again said ‘Yum yum!’ If the child looked to the matching video the tester made comments like, ‘Yeah! You found it! That's the one!’ If the child failed to look in the right direction, the tester said, ‘Look, here it is! Here's the one doing it’, while pointing toward the appropriate screen, then back and forth between the spoon held to the doll's lips and the screen.
Toddlers were randomly assigned to either the verbal support condition (16 16-month-olds; 16 24-month-olds) or to the no-support condition. In the verbal support condition, toddlers were provided with verbalizations similar to those used by Mandler and McDonough (1996, 1998; McDonough & Mandler, 1998) in their generalized imitation procedure.6 The procedure for the no-support condition was identical to that described for Experiment 1.
Coding
A second observer coded videotaped responses for 20% of the toddlers tested at each site. Reliability between on-line and off-line coding ranged from .93 to .99 (M= .97).
Results and discussion
To determine whether the revised videos had alleviated the problem of preferences for particular videos (starting) reported in Experiment 1, paired t-tests were conducted to compare looking times to each of the two videos presented during familiarization trials within each trial block. No significant differences were found, suggesting that the revised videos effectively dampened saliency effects for starting.
Preliminary analyses revealed no significant effects of testing site, gender, pet ownership (whether or not a dog or a cat lived in the home) or test order on looking times to the matching versus nonmatching videos. These variables were collapsed in all subsequent analyses. The initial set of analyses, with looking to the matching minus nonmatching screens considered across the target or modeled actions for each pair, revealed no reliable effects of the manipulated variables. A 2 (age) × 2 (condition) ANOVA performed on the mean difference scores averaged across the four test trials revealed no significant main effects or interactions. Separate 2 (age) × 2 (condition) × 2 (trial block) mixed analyses of variance, with trial block as a within-group factor, indicated that children's looking to the matching versus nonmatching screen did not change across the two trial blocks for either of the modeled actions. One-sample t-tests conducted separately within each condition for each modeled action indicated that all mean difference scores were at chance levels.
In the next set of analyses, we included target action as a between-subject variable. Separate 2 (age) × 2 (condition) × 2 (target action) × 2 (trial segment: familiarization versus test) mixed ANOVAs with trial segment as a within-group factor were conducted for each pair of actions. For starting–riding, there was a significant interaction between trial segment and target action, F(1, 53) = 5.74, p < .05. This was attributable to a slightly greater difference in looking between familiarization and test trials for starting than for riding (the tendency, however, was for slightly greater looking to the riding video following modeled starting; M=– 0.70 s). A parallel analysis performed on data from the drinking–sleeping pair revealed a similar trial segment × target action interaction, F(1, 52) = 4.18, p < .05, which was tempered by a three-way interaction among trial segment, target action, and condition, F(1, 52) = 4.97, p < .05. To examine the three-way interaction, separate mixed 2 (target action) × 2 (trial segment) ANOVAs, with trial type included again as a within-group factor, were conducted for the verbal support and no-support conditions. In the no-support condition, no significant effects emerged. However, in the verbal support condition, there was a significant interaction between target action and trial segment, F(1, 28) = 8.17, p < .01, for which relevant means and standard deviations are presented in Table 3. Interestingly, the verbalizations accompanying the modeling of sleeping (‘night-night!’) supported a shift in toddlers’ looking from the drinking video during familiarization toward the matching video of the real dog sleeping. When this analysis was conducted separately for each age group, we found that a significant interaction between target action and trial segment emerged only for 24-month-olds, F(1, 14) = 11.32, p < .01, suggesting that the facilitative effect of the ‘night-night’ vocalization was specific to the older age group (mean difference scores =−1.29 and 1.17 for familiarization and test, respectively). Of all the verbalizations provided, ‘night-night’ was perhaps most tightly aligned with the particular action it accompanied (the sipping noises accompanying drinking and the ‘vroom’ and ‘whee’ noises accompanying starting and riding, respectively, could have been equally well-suited to other actions aligned with animal and vehicle properties). Therefore it is not surprising that the fragile facilitative effects of supportive verbalizations were specific to older toddlers’ comprehension of sleeping.
| Action | 16-month-olds | 24-month-olds | ||
|---|---|---|---|---|
| Familiarization | Test | Familiarization | Test | |
| Drinking | 0.34 (2.16) | −0.61 (0.96) | 0.69 (1.10) | −0.50 (0.57) |
| Sleeping | 0.14 (2.59) | 0.77 (2.66) | −1.29 (1.78) | 1.17 (1.87) |
- Note: Mean difference scores represent looking time to matching video minus looking time to nonmatching video.
Although the majority of toddlers performed at chance levels on the preferential looking task, it seemed possible that variations in performance might be related either to expressive vocabulary or to parents’ assessment of the degree to which each of the four target actions was familiar to the child. However, none of the correlations emerged as significant between children's mean difference in looking time to matching versus nonmatching videos and three relevant measures: raw number of words produced as assessed through the CDI (r= .12), comprehension of action words aligned with the target actions (r= .02) and relative familiarity with the target actions (all rs < .24). In contrast to prior findings indicating relations between expressive vocabulary and both verb generalization (Forbes & Poulin-Dubois, 1997) and comprehension of objects as symbols (Younger & Johnson, 2004), we found no evidence that language initially mediates children's performance on the modeled action preferential looking task.
It is clear that the effects of additional verbal information on children's performance on the preferential looking task were fragile and limited to older children's performance on the sleeping trial. Because 26-month-olds in Experiment 1 also demonstrated virtually no comprehension of relations between modeled and videotaped ‘real’ depictions of sleeping in the absence of supportive verbalizations, it seems likely that 24-month-olds’ looking during test trials in this experiment was driven by their comprehension of the phrase, ‘night-night!’ Despite our best efforts to increase the correspondence between the actions modeled with toy replicas and videotaped depictions of their real-world counterparts and to eliminate problems associated with the heightened salience of the original starting video, we found no evidence that 24- or 26-month-old children are able to relate modeled actions aligned with the vehicle domain (starting, riding) in Mandler and McDonough's generalized imitation studies to videotaped depictions of their real-world referents.
General discussion
As infants acquire knowledge about objects in their world, their discovery of properties, actions and events associated with those objects provides a fertile repository of concepts upon which adjectives and verbs may be mapped. Negative effects must obviously be interpreted with caution, and it is possible that some of toddlers’ difficulty in the preferential looking task stemmed directly from our reliance on videotaped depictions of real-world actions, rather than live instances of those actions. Nevertheless, our results suggest that actions modeled with miniature toy replicas of objects may not be transparently related to corresponding real-world actions, at least as they are depicted on videotape. We repeatedly failed to find any evidence that 16- or 18-month-olds comprehended such relations, despite the fact that 18-month-olds are beginning to master similar relations between model and real objects (Younger & Johnson, 2004). While 26-month-olds reliably comprehended relations between modeled and videotaped real events for intransitive actions (jumping, running) and for familiar actions involving a human agent and/or patient (stirring, feeding), very little evidence was found to support 24- or 26-month-olds’ comprehension of the more complicated actions frequently modeled in the generalized imitation procedure, despite efforts to increase the correspondence between the modeled actions and videos of their real-world counterparts. Supportive verbalizations enhanced 24-month-olds’ comprehension only for sleeping, presumably because toddlers (who have a wealth of experience with napping and sleeping) have come to associate the phrase ‘night-night’ with beds and being ‘patted to sleep’. Otherwise, performance generally did not improve in Experiment 2 when supportive verbalizations were provided in conjunction with modeled actions, suggesting that toddlers’ capacity for comprehending relations between real and modeled events is quite fragile. Below, we first consider possible bases for 16- and 18-month-olds’ difficulty with the present tasks. We then consider theoretical implications for conclusions that can be drawn from use of the generalized imitation task in assessing infants’ conceptual development.
Developmental impediments
Sixteen- and 18-month-olds showed no evidence of comprehending relations between instances of action categories that were modeled with toy replicas and videos of their real-world counterparts. This was somewhat surprising, given that 16-month-olds typically are beginning to comprehend verbs (Golinkoff, Hirsh-Pasek, Cauley & Gordon, 1987) and are apt to imitate renditions of such events modeled by adults during symbolic play (Fenson, 1984). Furthermore, infants are adept at establishing categories based on correlations among form attributes (e.g. shape) and the functional affordances with which they are linked (Mervis, 1985, 1987), and particular theories of conceptual development emphasize infants’ attention to dynamic or functional information as the foundation for learning about objects (Mandler, 2000a; Nelson, 1973a, 1973b, 2000). Functional properties (e.g. flying, rolling) associated with familiar object concepts for which labels are known also would seem to provide optimal support for verb learning. First, invariants associated with the action categories aligned with particular functions should be easier to detect when associated with familiar object concepts. Second, lexical principles such as mutual exclusivity (Markman, 1989) or Novel-Name-Nameless Category (N3C; Golinkoff, Mervis & Hirsh-Pasek, 1994; Mervis & Bertrand, 1994) support the acquisition of nonobject words only after object words are known. In the present experiments, actions were modeled in conjunction with highly familiar objects, and the actions themselves were selected based on their familiarity or because of their being included in generalized imitation tasks with younger infants. It seems unlikely that younger toddlers’ difficulty with the present tasks stemmed from a general deficit in attending to and individuating action categories.
A more plausible explanation for young toddlers’ poor performance is related to symbolic comprehension of objects and props used to model target actions. Across experiments, successful comprehension of relations between videotaped depictions of real actions and their counterparts modeled with toy replicas depended (perhaps critically) on successful symbolic comprehension of the props used to model the actions. As discussed previously, we recently have found that at 18 months, babies demonstrate only a fragile understanding of such relations (Younger & Johnson, 2004). Mastery of such symbolic relations would appear to be necessary, but not sufficient, for success in the present experiments. Although 26-month-olds’ comprehension of objects as symbols was essentially at ceiling in our earlier experiments, 26-month-olds in the present study continued to have some difficulty (although not as much as 18-month-olds) in performing a parallel set of tasks involving action categories. If this explanation for younger toddlers’ poor performance is correct, it suggests a novel hypothesis that could be subjected to empirical testing: verbs learned in conjunction with actions modeled during symbolic play are apt to be underextended to include only modeled (but not real-world) instances of the relevant action.
Exposure to information about properties of objects or discovery of those properties during play is the primary avenue through which conceptual knowledge accrues. Thus, it is important to consider the role of existing conceptual knowledge in helping children to make sense of the instances of action categories to which they are exposed. In the present experiment, we found no facilitative effects on performance of either pet ownership (which presumably enables access to knowledge concerning what pets look like when they are drinking or sleeping) or parents’ assessments of relative familiarity with the target actions. This suggests that young children may not necessarily recruit available conceptual knowledge when processing information related to actions. If such knowledge is recruited, it may not be sufficient to override problems with interpreting symbolic connections between real objects depicted on videotape and small model replicas of those objects.
Methodological implications for studying conceptual development
Children's failure to comprehend relations between real and modeled actions raises questions concerning what infants are generalizing when they extend actions modeled with toy replicas to novel toy referents in the generalized imitation procedure. Across the two experiments, performance was poor for the two pairs of target properties consistently used by Mandler and McDonough (1996, 1998; McDonough & Mandler, 1998), despite attempts to render the videotaped real-world and modeled instances of these four actions as perceptually similar as possible. Post-hoc analyses confirmed that toddlers looked randomly at test videos regardless of the particular actions modeled. It was not the case that children failed to find the videos interesting. Rather, looking typically was accompanied by labeling of the objects involved in each action (or enthusiastic waving when the human in the video waved toward the camera). At best, infants participating in generalized imitation tasks may be responding to the affordances aligned with the props provided and/or drawing on conceptual representations linking surface features (e.g. parts that are shared between the models and real object kinds) with action- or motion-based properties associated with the modeled actions (Rakison & Hahn, in press). For example, both real animals and animal models have heads and faces (configured in relation to an elongated body) which may be associated with drinking. This part-based associative learning model could also account for infants’ failure to imitate counterconventional behaviors using inappropriate objects (Mandler & McDonough, 1996). Alternatively, the modeled actions used in the generalized imitation procedure may function as empty predicates, rendering the task highly comparable to studies of inductive inference in older children (Gelman & Markman, 1986; Johnson, Scott & Mervis, 1997). But if such imitations are not accompanied by a clear understanding of how modeled properties align with conceptual representations associated with real kinds (e.g. that animate kinds drink and sleep), the degree to which results from such tasks can inform theories of conceptual development seems limited.
Surprisingly, verbalizations provided in conjunction with modeled actions exerted little measurable impact on preferential looking. We had hypothesized that the presence of such verbalizations in the generalized imitation procedure helped to orient infants’ attention to the target action, perhaps even scaffolding their imitation. We found little evidence that such input was beneficial in our task, at any age. In hindsight, it seems possible that such language may have been confusing to infants. For example, the ‘vroom! vroom!’ used in conjunction with starting is a sound frequently used by adults and children when pretending to make toy vehicles move and it probably did not help to focus children's attention on the critical features of starting. To complicate matters further, the only video to feature a moving vehicle was riding. These factors may have accounted for the general lack of improved performance in Experiment 2, despite the simplified procedure and revision of videos to heighten the perceptual similarity between the real and modeled actions.
It is likely that there is variability among children in terms of which aspects of modeled and real events are attended to in both our own study and in studies involving generalized imitation. It also seems possible that action categories that involve humans as agents may be disambiguated from the dynamic flow of everyday events earlier than action categories that do not involve human agents. Results from several veins of infancy research have converged in suggesting that humans are processed qualitatively differently from other objects (Bonatti, Frot, Zangl & Mehler, 2002; Morton & Johnson, 1991; Pauen, 2000; Quinn & Eimas, 1998). An interesting question for future research is whether the action categories typically used in the generalized imitation procedure to convey properties of animates (e.g. sleeping, drinking) would be extended similarly if modeled with dolls, rather than with replicas of nonhuman animals. In sum, the results from the present experiments suggest that children's comprehension of symbolic relations between videotaped depictions of real actions and their modeled counterparts emerges well after they begin to understand similar relations involving real and model objects. The results also suggest many directions for continuing to explore the mechanisms that underlie developments in performance on generalized imitation tasks throughout the first two years of life.
Footnotes
Acknowledgements
This research was supported by grants 0091809 and 0115860 from the National Science Foundation and was presented at the 2003 meeting of the Society for Research in Child Development in Tampa, FL. We are very grateful to the infants and their families who have participated in these studies, and we thank Karyn Bartnicki, Carrie McGinnis, Raven Cuellar and Kathleen McRee for their assistance.
Appendix
| Target Action | Modeling Objects | Modeled Actiona | Videotaped Actionsb |
|---|---|---|---|
| Starting | Car/Key | Turn key against car door [vroom, vroom] | Person opens car door with key, starts car |
| Riding | Car/Doll | Push car with doll (arm held up) on top [whee!] | Person rides in car, smiling and waving |
| Sleeping | Dog/Bed | Put toy dog in a bed, pat dog [night-night] | Dog lying down to sleep |
| Drinking | Dog/Cup | Hold cup to dog's face [umm good, sip, sip] | Dog drinking from a bowl |
| Jumping | Horse | Horse takes two steps, jumps (15 cm arc), two more steps | Horse jumping over gate |
| Running | Horse | Horse moves rapidly in short steps across table | Horse galloping past gate |
| Feeding | Spoon/Doll | Place toy spoon to doll's mouth | Spoon dipping into bowl, feeding child |
| Stirring | Spoon/Bowl | Spoon stirring repeatedly inside of toy bowl | Spoon stirring in bowl |
- a The italicized verbal prompts [in brackets] were provided to half of the toddlers in each age group in Experiment 2.
- b Both videotaped actions were presented during test trials. The target actions and the positions of the matching and nonmatching videos (left versus right) were counterbalanced across toddlers (the same exemplar appeared in both videos).
