Running Head: COMPETING ATTENTIONAL MECHANISMS

Two Competing Attentional Mechanisms in Category Learning

Alan W. Kersten

Robert L. Goldstone

Alexandra Schaffert

Indiana University

5/31/97

Abstract

This research provides evidence for two competing attentional mechanisms. Attentional persistence directs attention to attributes previously found to be predictive, whereas contrast directs attention to stimuli that have not already been associated with a category. Three experiments provide evidence for these mechanisms. Experiments 1 and 2 revealed increased attention to an attribute following training in which that attribute was relevant, providing evidence for persistence. These experiments also revealed increased attention to an attribute following training in which another, more salient attribute was relevant, providing evidence for contrast. Experiment 3 used a subtractive method to determine the contributions of persistence and contrast to changes in attention to an attribute. The results suggest that persistence operates primarily at the level of dimensions, whereas contrast operates at the level of dimension values.

Two Competing Attentional Mechanisms in Category Learning

Learning theorists have relied on the notion of attention to explain how learning one discrimination can be facilitated or retarded by the prior learning of another discrimination. In particular, a discrimination is easy to learn to the extent that one attends to the relevant dimension. Two seemingly contradictory mechanisms have been proposed, however, for the changes in attention following a successful discrimination. According to one mechanism, which we will call attentional persistence, increased attention is allocated to dimensions that have been found to be predictive of an outcome. This mechanism is involved in the analyser theory of Sutherland & Mackintosh (1971). According to analyser theory, when a discriminative stimulus in an operant conditioning paradigm is found to be predictive of reinforcement, greater attention is devoted to the dimension of the discriminative stimulus. Because the total pool of attentional resources is fixed, this requires a decrease in attention to other dimensions. As a result, a learner is better able to learn a second discrimination along the same dimension but is less able to learn a discrimination varying along a different dimension. Attentional persistence plays a role not only in theories of animal conditioning, but also in prominent theories of categorization (e.g., Kruschke, 1992; Nosofsky, 1986).

A second attentional mechanism, which we will call contrast, works in the opposite direction. In particular, once a stimulus has been associated with a category, attention is allocated to other dimensions when another category must be learned. This mechanism is involved in the Pearce-Hall model of learning (Pearce & Hall, 1980). According to this model, stimuli receive attention only to the extent that their consequences are unclear. Thus, once a stimulus has been found to reliably predict a given outcome, attention to that stimulus is reduced. This theory thus stands in direct opposition to analyser theory, according to which a consistent predictor of an outcome receives greater attention.

Evidence for Attentional Persistence

Early evidence for attentional persistence can be found in the classic study of Lawrence (1949). In the first stage of this study, rats were placed in a maze with one white arm and one black arm. They were reinforced for entering the black arm. Attentional persistence would thus predict increased attention to color following this discrimination. In contrast, rats in a control condition were trained in a gray maze with texture rather than color predicting reinforcement. After this first stage of learning, rats were placed in mazes that had either two white arms or two black arms. Rats were reinforced for entering the right arms of black mazes and the left arms of white mazes. Thus, specific prior associations learned in the first stage would not help rats learn these new contingencies. Consistent with attentional persistence, experimental rats learned these contingencies faster than control rats, presumably because of increased attention to color following the first stage of training.

Further evidence for persistence comes from studies of intradimensional versus extradimensional shifts in discrimination. For example, Mackintosh and Little (1969) presented pigeons with pairs of stimuli that varied on both the color and orientation of a stripe. One group of pigeons was reinforced for pecking a stripe of a particular color, whereas a second group was reinforced for pecking a stripe of a particular orientation. Following this initial training, both groups were transferred to a task in which only color was relevant. Two new colors and two new orientations were used so that prior associations would presumably not directly affect pigeons' performance. Pigeons that were subject to an intradimensional shift (i.e., those that were initially trained on color) performed better on this transfer task than did pigeons who were subject to an extradimensional shift. This finding may reflect attentional persistence to the dimension that was relevant in the first stage of learning.

Attentional persistence may also account for the results of studies of reversal learning (e.g., Mackintosh, 1969). This paradigm is similar to the shift paradigm except that the same discriminative stimuli are used in both stages of learning. The response contingencies are simply reversed following the first stage of learning. For example, if rats were only reinforced for responses to black in a black-white discrimination, they would only be reinforced for responses to white in the second stage of learning. Thus, prior associations between black and reinforcement and between white and no reinforcement would be expected to slow learning in the transfer task. Rats tend to perform better on the transfer task, however, when they have received a greater amount of training on the initial discrimination. This counterintuitive finding may reflect a gradual reallocation of attention to color during initial training. Thus, rats that receive a greater amount of initial training attend more to color, allowing them to more easily learn the reversal in the second stage of learning.

The evidence for attentional persistence has so far been confined to studies involving nonhuman animals. Evidence for attentional persistence in humans comes from the classic study of Shepard, Hovland, and Jenkins (1961). Shepard et al. presented participants with eight exemplars that varied on three binary dimensions. The assignment of exemplars to two categories varied in different conditions in order to examine the influences of category structure on difficulty of learning. Participants were found to have the least difficulty with categories that required attention to only one dimension and the most difficulty with categories that required equal attention to all three dimensions. Shepard et al. were unable to account for these results in terms of interstimulus similarities. They concluded that the notion of selective attention to dimensions was required to account for these data. This selective attention appears to be an example of attentional persistence. In particular, a participant who discovered a consistent relation between a particular dimension value and a particular category may have allocated extra attention to that dimension. This would facilitate the discovery of the relation between the other value of that dimension and the other category. This same persistence would retard the acquisition of categories that required attention to more than one dimension.

Further evidence for persistence in humans comes from a study by Goldstone (1994). In this study, undergraduates first learned to categorize squares that varied on size and brightness. For some participants, size was relevant to categorization, whereas for others, brightness was relevant. Participants were then transferred to a simple perceptual discrimination task in which they had to judge whether pairs of briefly presented squares were the same or different. Participants were found to be better able to discriminate squares that varied on the dimension that was relevant to categorization compared to controls that had received no categorization training. This finding may reflect attentional persistence following categorization training, allowing better discrimination of squares that varied on the attended dimension.

Evidence for Contrast

Evidence for contrast comes from word learning biases in children. Children tend to associate a novel word with the shape of an object, as indicated by their willingness to extend a novel word to objects with the same shape but not to objects with a different shape (Landau, Smith, & Jones, 1988). If a novel word is presented along with a shape that already has a label, however, children shift attention away from shape and toward other attributes such as the substance (Markman & Wachtel, 1988) or the motion of the object (Kersten & Smith, 1997). Markman & Wachtel (1988) interpret these findings as evidence for a mutual exclusivity assumption in children. According to this assumption, an object can have only one label, and thus if no unlabeled objects are present, a novel label must refer to something else. Consistent with attentional contrast, this assumption would direct attention away from a previously predictive dimension (i.e., shape) toward other, previously unpredictive attributes.

The mutual exclusivity assumption must be relaxed in order to acquire the full expressive power of language. The same object can be labeled by many different words in the adult vocabulary. One may argue that attentional contrast is only found in children, and that adults abandon this attentional mechanism in favor of others such as persistence. Other findings, however, provide evidence for contrast in adults.

Evidence for contrast in adults comes from the inverse base rate effect (e.g., Kruschke, 1996). In this research, adults learn two categories, one occurring several times more frequently than the other. Two cues are associated with each category. One is a perfect predictor of the category, whereas the other is imperfect, occurring with both categories. Because one category is more frequent than the other, however, the imperfect predictor occurs more often with the more frequent category than with the less frequent category. Kruschke (1996) found that adults learn the more frequent category first because of its greater number of presentations, associating this category with both its perfect and imperfect predictors. When participants then learned the less frequent category, they associated it only with its perfect predictor, because the imperfect predictor was already associated with the more frequent category. As a result, a novel test stimulus involving the perfect predictors of both categories was more likely to be categorized as an example of the less frequent category, because this stimulus possessed the only cue associated with the rare category whereas it possessed only one of the two cues associated with the common category.

Kruschke (1996) modeled this phenomenon using a connectionist network that employed an attentional mechanism consistent with contrast. In particular, if a feature that was already associated with a category was presented along with a different category label, attention was shifted toward other features to learn that new category. Thus, the association of the imperfect predictor with the more frequent category caused attention to be shifted away from that feature when learning the less frequent category.

Based only on the evidence from children and adult humans, one may believe that attentional contrast is a language-specific phenomenon that allows people to learn labels for categories. There is also evidence for contrast in rats, however, which clearly do not have language. This evidence comes from the latent inhibition paradigm (Lubow & Moore, 1959). In this paradigm, animals in the experimental condition are first presented with a conditioned stimulus (CS) a number of times in isolation. This same CS is then paired with an unconditioned stimulus (US) via classical conditioning. The animals in the experimental condition tend to show slower learning of the US-CS association than do animals in a control condition that receive no preexposure to the original CS. Pearce and Hall (1980) proposed that animals in the experimental group learn that the CS is associated with a consistent outcome during preexposure (i.e., no event). As a result, the Pearce-Hall learning model predicts that they will start to attend to other dimensions. This attentional contrast reduces their ability to learn subsequent associations involving the CS. Thus, although it does not involve language, latent inhibition is similar to mutual exclusivity and the inverse base rate effect in that attention is directed away from previously predictive dimensions toward other dimensions.

An additional finding that may reflect contrast comes from Schyns and Rodet (1997). Participants learned two categories of "Martian cells," one defined by a single novel feature (X) and one defined by this feature contiguous with a second novel feature (Y). The X-XY group learned the X category prior to the XY category. This group was predicted to represent the XY category as the conjunction of the X and Y features. In contrast, the XY-X group learned the XY category before the X category. This group was expected to represent the XY category as a single conjoined feature because they had never seen X in isolation before learning the XY category.

Participants were tested with stimuli involving the same X and Y features separated in space. The XY - X group tended to categorize these as examples of the X category, because the conjoined XY feature was not present whereas the X feature was. In contrast, the X - XY group tended to categorize them as XY, because both the X and Y features were present. This finding seems to provide evidence for contrast. In particular, participants in the X - XY group may have attended primarily to Y in the second phase of learning in order to distinguish XY from X. As a result, the presence of the Y feature in the test stimuli may have encouraged these participants to categorize these stimuli as examples of the XY category.

Schyns and Rodet (1997) interpreted their results as evidence for feature creation. In particular, Schyns and Rodet proposed that the creation of an X feature in the first phase of learning allowed participants in the X - XY condition to extract the X feature in the second phase of learning and thus to create an independent Y feature. A feature delineation task, as well as other experiments, produced evidence consistent with this interpretation. Attentional mechanisms such as contrast, however, may also have contributed to the results of these experiments. Attentional processes are not incompatible with feature creation. In fact, feature creation may be a necessary prerequisite for featural attention weighting.

The Interaction of Attentional Persistence and Contrast

Findings consistent with contrast provide evidence that learners direct attention away from previously predictive dimensions, whereas findings consistent with persistence provide evidence that learners direct attention toward these predictive dimensions. The distinction between dimensions and dimension values may be key in resolving this apparent inconsistency. Persistence may operate primarily at the level of dimensions (although there may also be some persistence to particular values; see Goldstone, 1994). Thus, when a dimension is found to be predictive of an outcome, persistence results in increased attention to that dimension regardless of the values involved.

Contrast may operate primarily at the level of dimension values. Thus, when a value along a dimension is associated with a response, contrast results in increased attention to other values along that dimension or to values along other dimensions. For example, when the shape of one object is already associated with a label, children associate a new label with the shape of a different object (Markman & Wachtel, 1988). Persistence may direct attention to shape as a dimension even as contrast directs attention away from particular shapes that already have a label. If no novel shapes are present, contrast results in increased attention to values along other dimensions such as substance (Markman & Wachtel, 1988). This analysis suggests that persistence and contrast together determine the amount of attention directed to a stimulus. Persistence will dominate in experimental designs in which identical dimensions are used, but the values along these dimensions are different. Contrast effects will reduce or possibly overwhelm these persistence effects when identical or perhaps similar values of a dimension are present but the categories or outcomes involved are different.

The present experiments were designed to provide evidence for the attentional mechanisms of persistence and contrast using Schyns and Rodet's (1997) experimental method. The first experiment examined the learning of categories that were similar to those of Schyns and Rodet (1997) in that they were thought to require the creation of functional features in order to be learned. The second experiment compared the learning of these categories to the learning of categories defined by features that were thought to be perceptually specified, obviating the need for functional feature creation. Any effects of learning history were thus presumed to reflect attentional processes rather than feature creation. The third experiment employed a subtractive method in order to determine the contributions of persistence and contrast to the overall attention allocated to a feature.

Experiment 1

Experiment 1 was designed to provide evidence for attentional persistence and contrast in a task similar to that employed by Schyns and Rodet (1997). The features that defined the categories were components of simple animated events. Novel verbs accompanied the events, serving as labels for the event categories. Two features of the events were related to the accompanying verbs. These were the path and manner of motion of one of the characters appearing in each event. For example, one verb corresponded to a character moving along a path toward a second character while along the way displaying a manner of motion involving zig-zagging to the left and right. As with Schyns and Rodet's (1997) stimuli, these two features can be interpreted either as separate or as a conjoined feature, such as a zig-zagged path (see Figure 1).

The order of learning the different categories was varied in order to influence the amount of attention allocated to each feature. One group of participants first learned categories defined by a conjunction of a path feature and a manner feature, as in the above example. People have previously been found to attend more to path than to manner of motion in the context of a verb learning task (Kersten, in press; 1997; Kersten & Billman, 1995), so this group of participants was expected to attend primarily to path. As a result, they were expected to later have difficulty distinguishing categories that shared identical paths, namely conjunctive categories and categories defined solely in terms of path.

Other participants learned categories defined by the component features before learning categories involving conjunctions of those features. Participants who first learned categories involving manner of motion were predicted to later attend more to manner of motion when learning conjunctive categories than participants who learned conjunctive categories first. This finding would provide evidence for attentional persistence. Participants who first learned categories involving path were predicted to later attend more to manner of motion when learning conjunctive categories than participants who learned conjunctive categories first. This finding would provide evidence for contrast. Similar effects were expected for path, but were expected to be lessened because of the high baseline attention to path. Small effects of persistence and contrast on attention to path were thus expected to be obscured by larger effects on attention to manner of motion.

Although we predicted effects of learning order on attention, learning order has also been found to influence feature creation in a similar experiment by Schyns and Rodet (1997). Two types of test items were constructed in order to distinguish effects of attention from effects of feature creation. One type of test, the conjunctive trials, examined the allocation of attention to path and manner of motion. The stimuli for this type of test were identical to those seen when learning conjunctive categories. Because participants had also learned categories that could only be defined in terms of path, a participant who had failed to attend to manner of motion during conjunctive learning would sometimes be expected to categorize conjunctive stimuli as examples of path categories.

A second test type, the carpet trials, was analogous to Schyns and Rodet's (1997) test involving spatially separated features, designed to test for effects of feature creation. Because event stimuli were employed in the present study, a simple spatial separation of features was not possible. Instead, the two features were separated by displaying them as being carried by different objects. In particular, a smoothly moving flying carpet (representing path) picked up an oscillating character (representing manner of motion) and carried it along with it. If participants create separate path and manner of motion features, they should categorize this test stimulus as a conjunction of these features, because both were present. In contrast, if participants represent conjunctive categories in terms of conjoined path and manner features, they should not categorize this stimulus as a conjunction, because each object carried only one of these features.

Attention was expected to influence the categorization of both test types, whereas a participant's featural representation of a category was only expected to influence performance on the carpet trials. Thus, any effects of learning order that were common to the two tasks were assumed to reflect attentional processes, whereas any order effects that were unique to the carpet trials were assumed to reflect feature creation. We predicted that, because of attention to manner, participants would on both tests make more correct conjunctive choices when either path or manner categories had been learned first. If learning order only influenced feature creation, more conjunctive choices would only be seen on the carpet trials.

Method

Participants

One hundred forty-four undergraduates at Indiana University participated in this experiment in partial fulfillment of course requirements for Introductory Psychology.

Stimuli

All Events. Events were displayed on Macintosh IIsi computers using MacroMind Director 3.1. An example event is depicted in Figure 2. Each event involved two bug-like characters. One character, the agent, moved throughout the event, while the other, the patient, remained motionless. The static appearance of each character varied on three attributes, its head, body, and legs. Each of these attributes had two possible values, chosen randomly for each participant from the four values used by Kersten (in press). An agent's motion also varied on three attributes. These were the path, manner of motion, and leg motion of the agent. The two possible values of each of these attributes are depicted in Figure 3. The only other attribute that varied was the environment in which the events took place, represented by a static line-drawing in an unused corner of the screen.

Each event began after a black screen faded away to reveal the starting positions of the two characters. The agent started each event at a point randomly chosen from a region near the center of the screen. The patient started each event at a varying distance from the agent along one of eight directions, either N, S, E, W, NE, NW, SE, or SW. Each event lasted approximately 8 seconds.

Learning Events. There were three phases of learning. Each phase involved between eight and 24 learning events, depending on test performance following each set of learning events (see below). Each event was labeled by a novel verb appearing in an unused corner of the screen, either gupping, morping, spogging, wunking, yimming, or zelling. Only two of the six verbs were presented in any one phase of learning. In the path phase of learning, the two verbs were perfectly correlated with the path of the agent. Manner of motion was not represented in this phase, with agents moving smoothly throughout. In the manner phase, the two verbs were related to the manner of motion of the agent. Path was not represented in this phase, with agents oscillating in place either front and back or left and right to represent their manners of motion. In the conjunctive phase, the two verbs were related to both the path and manner of motion of the agent. All other attributes varied randomly throughout.

Criterion Test Trials. Following each set of eight learning events, participants were tested on their knowledge of the two verbs they just saw. There were six test trials, three testing for knowledge of each verb. Participants were presented with pairs of events, one after the other, each labeled by the same verb. One event in each pair was consistent with the accompanying verb, whereas the other mismatched the verb on either path (in the path phase), manner of motion (in the manner phase), or both (in the conjunctive phase). The correct event was presented first in half of the trials and second in the other half. The participant's task was to choose which event was the better example of the accompanying verb. The criterion for moving on to the next part of the experiment was 100% correct performance on the set of six test trials. Participants were also allowed to move on if they had not achieved criterion following three sets of learning events.

Final Test Events. Following the third phase of learning, participants were presented with 20 events testing for knowledge of all six verbs seen earlier in the experiment. After each event, participants were asked to choose which of the six verbs was the best label for that event. There were four events each of five different test types. The first three types, path, manner of motion, and conjunctive trials, involved events that were generated in the same way as events seen during learning. One of the two verbs from the corresponding phase of learning was thus the correct answer in these trials. The fourth type, mismatch trials, involved events that depicted a combination of path and manner of motion that had not been presented during learning. Each event thus mismatched each conjunctive verb on either path or manner of motion, whereas it was consistent with one path verb and one manner verb except for the fact that the other attribute was also represented.

The fifth test type, carpet trials, was designed to be analogous to Schyns and Rodet's (1997) critical test type in that the two attributes of interest were presented as separate entities. Whereas the attributes were separated spatially in Schyns and Rodet's research, they were separated in the present work by employing two different objects to carry the two different types of motion. In particular, the agent carried the manner of motion attribute, just as it had throughout learning. The agent began the event oscillating in place, just as during the manner phase of learning. While it did this, however, a flying carpet, represented by a red square, flew underneath the agent and then carried the agent along on its path. The carpet flew along one of the two paths taken by the agent during learning. The agent started the event about halfway along the path from the carpet to the patient, thus allowing the carpet to start in motion independently of the agent and then pick it up. This was done to encourage the interpretation that the agent was not responsible for its own motion, but rather was being carried by the carpet. The agent, however, continued to oscillate throughout the event while the carpet moved smoothly. The carpet thus represented the path attribute, whereas the agent represented manner of motion. The path and manner of motion were paired as they had been during the conjunctive phase of learning.

Procedure

Participants were instructed that they would be learning verbs from a language spoken on another planet. They were then presented with the first phase of learning. After each event, a button labeled "Next Event" appeared in the lower right-hand corner of the screen, allowing participants to continue on to the next event. After eight learning events, participants were given six criterion test trials. After the first event in each trial, participants clicked on the "Next Event" button to see the second event. After the second event, participants clicked on one of two buttons labeled "First Event" and "Second Event" to indicate which event they thought was a better example of the verb accompanying those events. A third button labeled "Repeat" allowed participants to review the two events as many times as they wished. If participants were correct on all six events, they were allowed to continue on to the next phase of learning. If participants failed to achieve this criterion, they repeated the learning/test cycle up to two more times. This same procedure was followed for all three phases of learning. Different participants were presented with the three phases of learning in six different orders, representing all possible permutations of the path, manner, and conjunctive phases of learning. After completing the three phases of learning, participants were shown 20 final test trials. At the end of each event, participants chose one of six buttons, each labeled by an individual verb, to indicate which verb they thought was the best label for that event. The "Repeat" button also appeared, allowing participants to review an event as often as they wished.

Design

The primary dependent variable in this experiment was the percentage of correct conjunctive responses to the conjunctive and carpet trials. The independent variables for this analysis were the order in which participants received the three phases of learning, manipulated between-participants, and the type of test (conjunctive vs. carpet), manipulated within-participants. Additional dependent variables were the percentages of path and manner responses to the path, manner, and mismatch trials. The independent variable for these measures was learning order. A final dependent variable was the percent correct on criterion test events. The independent variables for this dependent variable were learning order and the type of verb being tested (i.e., path vs. manner vs. conjunctive).

Results

Criterion Test Trials

The results of the criterion test events are displayed in Figure 4. Because all participants completed at least one block of trials for each verb type (i.e., path, manner, and conjunctive), an ANOVA could be performed on criterion test scores following the first block of learning with each verb type. This analysis revealed a main effect of verb type, F (2,276) = 17.07, p < .001, MSE = 437.66. Post-hoc t-tests revealed that participants performed worse on tests of manner verbs than on tests of either path verbs, t (143) = 4.45, p < .001, or conjunctive verbs, t (143) = 4.19, p < .001. Participants did not perform significantly differently on tests of path verbs and conjunctive verbs, t (143) = 0.57, p > .10.

There was also a main effect of learning order, F (5, 138) = 4.95, p < .001. To understand the cause of this main effect, each participant's scores for the three verb types were averaged to create a composite criterion test score. These composite scores were then entered into a Fisher's LSD post-hoc test. This analysis revealed that participants in the Path - Manner - Conjunctive (PMC) condition performed significantly better than participants in the MPC, CMP, and MCP conditions, whereas participants in the PCM and CPM conditions performed better than participants in the MPC and CMP conditions. It appears that participants performed better in conditions in which path verbs were learned before manner verbs, regardless of when conjunctive verbs were learned.

There was also a significant interaction of learning order with verb type, F (10,276) = 3.60, p < .001, MSE = 437.66. To understand the cause of this interaction, post-hoc LSD analyses were conducted on each verb type in isolation. These analyses revealed no significant differences between the three conditions on path and conjunctive verbs. There were, however, significant differences on manner verbs. In particular, the PMC and PCM conditions learned manner verbs more quickly than the MPC, MCP, and CMP conditions, the CPM condition performed better than the MPC and MCP conditions, and the CMP condition performed better than the MPC condition. Thus, participants appear to have had difficulty learning manner verbs if they had not already learned path verbs, particularly if they learned manner verbs first of all (i.e., before path and conjunctive verbs).

Final Test Trials

Participants were selected for analysis of final test scores based on performance in the criterion test trials. In particular, participants had to perform with 100% accuracy on at least one block of criterion test trials for each verb type in order to be included in analyses of final test scores. The numbers of participants in each condition who achieved this criterion are as follows: 24 in the PMC condition, 17 in the MPC condition, 24 in the PCM condition, 18 in the MCP condition, 24 in the CPM condition, and 18 in the CMP condition.

The results of the conjunctive and carpet trials are depicted in Figure 5. An ANOVA was performed on the percentage of correct conjunctive choices to these two trial types, with trial type and learning order as independent variables. This analysis revealed a main effect of trial type, F (1,119) = 62.79, p < .001, MSE = 822.45. As can be seen in Figure 5, participants tended to make more correct conjunctive choices in the conjunctive trials than in the carpet trials. There was also a main effect of learning order, F (5, 119) = 4.98, p < .001, MSE = 1163.01. The interaction of trial type and learning order did not approach significance, however, F (5, 119) = 0.33, p > .10, MSE = 822.45, indicating that the effects of learning order were similar for the two trial types.

Two nonorthogonal planned comparisons were used in order to better understand the main effect of learning order on participants' performance in the conjunctive and carpet trials. In particular, one comparison pitted the PMC and MCP conditions against the PCM and CMP conditions. Each of the former conditions is identical to one of the latter conditions except that the order of learning manner and conjunctive verbs is reversed. This comparison thus examines the evidence for persistence to manner in the conjunctive phase as a result of prior manner training. The CPM and MPC conditions were not included in this comparison because path learning intervened between manner and conjunctive learning. As a result, it would be impossible to determine whether changes in attention to manner were due to persistence, contrast, or both.

The second comparison pitted the MPC and PCM conditions against the MCP and CPM conditions. Each of the former conditions is identical to one of the latter conditions except that the order of learning path and conjunctive verbs is reversed. This comparison thus examines the evidence for influences of contrast on attention to manner in the conjunctive phase as a result of prior path training. The CMP and PMC conditions were not included in this comparison because manner learning intervened between path and conjunctive learning.

An alpha level of .025 was adopted for each of these comparisons in order to maintain an overall alpha level of .05. The comparison that tested for evidence of persistence was significant, t (119) = 2.43, p < .025, indicating that participants who learned manner verbs prior to conjunctive verbs made more correct conjunctive choices in the conjunctive and carpet trials than did participants who learned conjunctive verbs before manner verbs. The comparison that tested for evidence of contrast was also significant, t (119) = 2.34, p < .025, indicating that participants who learned path verbs prior to conjunctive verbs also performed better.

Participants who failed to choose correct conjunctive verbs in these trials most often chose a path verb. These participants thus chose a verb that matched the test event on the path component but mismatched on the manner component. The percentages of matching path verb choices in the conjunctive and carpet trials are depicted in Figure 6. These graphs reveal trends in learning order opposite to those for conjunctive verb choices. To illustrate this point, an ANOVA identical to the one above was carried out with matching path verb choices as the dependent measure. This analysis again revealed a significant main effect of test type, F (1, 119) = 5.98, p < .05, MSE = 481.44, with more path choices in the conjunctive trials than in the carpet trials. There was also a significant main effect of learning order, F (5, 119) = 4.93, p < .001, MSE = 805.25, but test type again did not interact with learning order, F (5, 119) = 0.70, p > .10, MSE = 481.44.

The same two planned comparisons as in the previous analysis were performed with the percentage of matching path choices as the dependent variable. The contrast testing for evidence of persistence approached significance, t (119) = 1.71, p < .10, indicating that participants who learned conjunctive verbs prior to manner verbs tended to make more path choices than did participants who learned manner verbs before conjunctive verbs. The contrast testing for evidence of contrast was significant, t (119) = 2.94, p < .01, indicating that participants who learned conjunctive verbs prior to path verbs made more path choices. Thus, participants who learned conjunctive verbs before path and/or manner verbs showed less differentiation of path and conjunctive verbs than did participants who learned these other verb types first, apparently failing to notice the manner component of conjunctive verbs.

Although the results of the conjunctive and carpet trials are consistent with greater differentiation of path and conjunctive verbs for participants who learned path and/or manner verbs before conjunctive verbs, an alternative explanation is that these data simply reflect a recency effect. In particular, participants who performed better in conjunctive and carpet trials had learned conjunctive verbs more recently, whereas participants who performed worse had learned path verbs more recently. The results of the path trials are inconsistent with this account, however, as can be seen in Figure 7. A oneway ANOVA on performance on the path trials revealed a significant main effect of learning order, F (5, 119) = 3.40, p < .01, MSE = 2952.90. Post-hoc Fisher's LSD tests revealed that the PMC condition performed better than the MCP, CPM, MPC, and CMP conditions, whereas the PCM condition performed better than the CMP condition. Thus, the two groups that performed best on the path trials learned path verbs first. This finding is inconsistent with a recency account, but consistent with contrast. In particular, participants who learned path verbs first attended more to manner in subsequent learning, allowing them to distinguish path verbs from conjunctive verbs.

The results of the manner trials are depicted in Figure 8. An ANOVA on the percentage of correct choices on these trials revealed a significant main effect of learning order, F (5, 119) = 2.86, p < .05, MSE = 1306.21. Post-hoc Fisher's LSD tests revealed that the CPM, CMP, PCM, and PMC conditions performed better than the MCP condition, whereas the CPM, CMP, and PCM conditions performed better than the MPC condition. Thus, participants who learned manner verbs first performed worse on the manner trials than did participants who learned path and/or conjunctive verbs first. These results are consistent with those of the criterion test trials, which also revealed difficulty in learning manner verbs prior to other verb types. Although all of the participants in the present analysis successfully completed the criterion test trials, their level of proficiency with these verbs may still have been less than that of participants who learned other verb types first. There were no significant effects on the mismatch trials.

Discussion

The results of this experiment are consistent with the hypothesis that both attentional persistence and contrast can act to draw attention to an otherwise nonsalient attribute. The results of the criterion trials indicated that manner of motion was indeed less salient than path. In particular, participants had more difficulty learning to differentiate two verbs when manner was the only diagnostic attribute than when path was diagnostic. Participants had particular difficulty learning manner verbs when they were presented in the first block of learning. Presumably, participants looked first to path when trying to differentiate verbs, and thus only proceeded to look at manner of motion when path was found to be nondiagnostic. The finding that participants performed better with manner verbs following either path or conjunctive training can be interpreted as providing evidence for contrast. In particular, after associating verbs with paths in the first block of learning, they may have been more likely to look to other attributes when learning new categories in subsequent blocks of learning, facilitating the learning of manner verbs.

The results of the final test trials also provide evidence for contrast, as well as for attentional persistence. Participants who had learned either path verbs or manner verbs prior to conjunctive verb learning were better able to distinguish path verbs from conjunctive verbs in the conjunctive and carpet trials. These participants thus attended more to manner of motion when learning conjunctive verbs than did participants who learned conjunctive verbs first. The greater attention to manner of motion following manner verb learning is consistent with attentional persistence, whereas the greater attention to manner of motion following path verb learning is consistent with contrast.

Experiment 2

The finding that learning order had similar effects on performance in the conjunctive and carpet trials does not support a feature learning account for the results of Experiment 1. The effects of prior path and manner learning were evident not only on trials that portrayed path and manner of motion as separate features (i.e., carpet trials) but also on trials in which path and manner of motion were represented together in the motion of a single object, just as during learning. It is possible, however, that the carpet trials were not in fact successful in separating the path and manner components of motion, and thus are not a true test of feature learning. Given the dynamic nature of the present stimuli, it would be difficult to construct a test that was exactly analogous to the separated test stimuli of Schyns and Rodet (1997), which involved a purely spatial separation of features.

Experiment 2 employed a different approach to look for evidence of feature learning. In particular, we presented some participants with motions that were perceptually specified by different parts of an object, thus presumably not requiring feature learning to be represented as separate components of motion. In particular, one component of motion involved the object as a whole, namely path, whereas the other component involved independent motions of the legs relative to the rest of the object. These two motions were considered likely to be perceived as independent features upon first exposure to them, and thus to be represented independently even when perfectly correlated, as in the conjunctive phase of learning. We contrasted this condition with a condition like that of Experiment 1, in which the two components of motion were carried by the same parts of the object, potentially requiring feature learning in order to represent these two components separately.

Experiment 2 involved only two different learning orders: Path -> Manner of motion (involving either leg motion or whole body motion) -> Conjunctive and Conjunctive -> Manner -> Path. These two orders were used because the key determinant of performance in Experiment 1 was whether conjunctive learning came first or was preceded by learning of the component motions. The other independent variable was the type of motion used to operationalize manner of motion, either body motion, as in Experiment 1, or leg motion. Prior experiments involving similar values of leg motion and path have revealed that leg motion, similar to body motion, receives less attention than path in the context of verb learning (Kersten, in press; Kersten & Billman, 1995). If feature learning contributed to the effects found in Experiment 1, learning order would be predicted to have more pronounced effects when manner of motion was operationalized as body motion than when operationalized as leg motion. In contrast, if the results of Experiment 1 reflect attentional processes, the effects of learning order should be similar regardless of how manner of motion is operationalized, providing a replication of the results of Experiment 1 using a different manner of motion.

A second change from Experiment 1 was that all attributes that were not relevant to the meaning of a verb varied randomly during learning. In Experiment 1, manner of motion did not vary when only path was relevant and path did not vary when only manner was relevant. This change to the procedure of Experiment 2 was designed to rule out the possibility that the effects of contrast in Experiment 1 simply reflected the fact that new variation in a dimension was introduced, potentially drawing attention to that dimension.

Method

Participants

Eighty-eight undergraduates at Indiana University participated in this experiment in partial fulfillment of course requirements for Introductory Psychology.

Stimuli

Learning Events. The learning events of Experiment 2 differed from those of Experiment 1 in that leg motion was related to verb meaning in the manner and conjunctive phases of learning for half of the participants. The body motion of the agent as a whole was relevant for the other half of the participants, as in Experiment 1. Also in contrast to Experiment 1, path, body motion, and leg motion varied randomly when they were not relevant to verb meaning, whereas in Experiment 1 manner of motion and path did not vary when only the other was relevant.

Criterion Test Trials. These differed from those of Experiment 1 in that there were only four criterion test trials after each block of learning events. When only one of the motion attributes (e.g., path) was related to verb meaning in a particular block, the other attribute that would be relevant to verb meaning during the experiment (e.g., body motion or leg motion) held a constant value across the two events in a test trial. Each such value was displayed on half of the criterion test trials. All other attributes varied randomly.

Final Test Events. The final test events were identical to those of Experiment 1 except that leg motion played the role of manner of motion for half of the participants. The body motion played this role for the other half of the participants. Although path had varied randomly when body motion or leg motion was relevant during learning, path was not displayed in the manner trials during the final test, as in Experiment 1. Similarly, during the path trials, body motion was not displayed to participants for whom body motion was relevant during learning, whereas leg motion was not displayed to participants for whom leg motion was relevant during learning.

Procedure

The procedure was identical to that of Experiment 1 except that only four criterion test trials followed each block of learning events.

Design

The primary dependent variable in this experiment was the percentage of correct conjunctive responses to the conjunctive and carpet trials. The independent variables for this analysis were the order in which participants received the three phases of learning (PMC vs. CMP) and the attribute used to operationalize manner of motion (leg motion vs. body motion), both manipulated between-participants, and the type of test (conjunctive vs. carpet), manipulated within-participants. Additional dependent variables were the percentages of path and manner responses to the path, manner, and mismatch trials. The independent variables for these measures were learning order and the attribute used to operationalize manner of motion. A final dependent variable was the percent correct on criterion test events. The independent variables for this dependent variable were learning order, the attribute used to operationalize manner of motion, and the type of verb being tested (i.e., path vs. manner vs. conjunctive).

Results

Criterion Test Trials

The results of the criterion test trials are depicted in Figure 9. As in Experiment 1, the results of trials following the first block of each phase of learning were analyzed in an ANOVA. This analysis revealed a main effect of the type of verb being tested, F (2,168) = 20.86, p <.001, MSE = 699.97. Post-hoc t-tests revealed that participants performed worse with manner verbs than with either path verbs, t (87) = 5.24, p < .001, or conjunctive verbs, t (87) = 5.55, p < .001. There was no significant difference in performance between path verbs and conjunctive verbs, t (87) = 0.22, p > .10.

The only other significant effect revealed by the ANOVA was an interaction of verb type with learning order, F (2,168) = 6.13, p < .01, MSE = 699.97. Post-hoc t-tests revealed that participants performed better on manner verbs following conjunctive verb learning than following path verb learning, t (86) = 2.60, p < .05. This provides evidence for attentional persistence, in that manner was relevant to conjunctive verbs but not to path verbs. A second t-test revealed that participants performed better on path verbs when they were learned first compared to when they were learned last, t (86) = 2.04, p < .05. This may again provide evidence for attentional persistence if one assumes that persistence to one attribute draws attention away from other attributes. In particular, participants who learned path verbs last had just completed manner verb learning, possibly drawing attention away from path. There was no difference between conditions on conjunctive verb learning, t (86) = 0.73, p > .10.

Final Test Events

As in Experiment 1, participants were selected for analysis of final test scores if they performed with 100% accuracy on at least one block of criterion test trials for each verb type. The numbers of participants in each condition who achieved this criterion are as follows: 11 in the PMC condition with whole body manner of motion, 15 in the PMC condition with leg motion, 8 in the CMP condition with whole body manner of motion, and 13 in the CMP condition with leg motion.

The results of the conjunctive and carpet trials are depicted in Figure 10. An ANOVA on these scores revealed only a main effect of learning order, F (1,43) = 11.86, p < .001, MSE = 1154.47. Participants made more correct conjunctive responses when conjunctive learning followed path and manner learning than when conjunctive verbs were learned first. Notably, there were no significant main effects or interactions involving the attribute used to operationalize manner of motion, all Fs < 2.

Participants who learned conjunctive verbs first made more path verb choices in the conjunctive and carpet trials than did participants who learned path and manner verbs before learning conjunctive verbs. The percentages of path choices on conjunctive and carpet trials are depicted in Figure 11. An ANOVA on conjunctive and carpet trials with the percentage of path choices as the dependent variable also revealed only a main effect of learning order, F (1,43) = 21.94, p < .001, MSE = 710.71. Again, there were no significant main effects or interactions involving the attribute used to operationalize manner of motion, all Fs < 1.

The results of the path trials are depicted in Figure 12. An ANOVA on these scores also revealed a main effect of learning order, F (1,43) = 4.55, p < .05, MSE = 1006.67. Participants made more correct path verb choices when path and manner verbs were learned prior to conjunctive verbs compared to when conjunctive verbs were learned first. The interaction of learning order with the attribute used to operationalize manner of motion approached significance, F (1,43) = 3.70, p < .07, MSE = 1006.67. This interaction reflects a trend toward higher performance on path trials in the PMC condition when leg motion was used to operationalize manner compared to when whole body manner of motion was used, t (24) = 1.96, p < .07. There were no differences between the two CMP conditions, t (19) = 0.82, p > .10.

Participants who learned conjunctive verbs first made more conjunctive choices on path verb trials than did participants who learned path and manner verbs first. The percentages of conjunctive verb choices in these trials are displayed in Figure 13. An ANOVA with the percentage of conjunctive verb choices on path trials as the dependent measure revealed only a main effect of learning order, F (1,43) = 7.70, p < .01, MSE = 578.292. There were no significant main effects or interactions in the manner or mismatch trials.

Discussion

The results of Experiment 2 revealed that participants were better able to distinguish path and conjunctive verbs when conjunctive verbs were learned after path and manner verbs, in replication of Experiment 1. Participants who learned conjunctive verbs first made more path responses in the conjunctive and carpet trials and more conjunctive responses in the path trials than did participants who learned path and manner verbs first. This finding is consistent with an attentional account of category learning, with path and manner learning drawing more attention to the manner component of conjunctive verbs as a result of contrast and persistence, respectively. Of course, because both path and manner learning either preceded or followed conjunctive learning, the results of this experiment could be explained entirely in terms of one of these mechanisms. The similarity of the results of Experiments 1 and 2, however, suggests that both mechanisms played a role in this experiment, just as in Experiment 1.

The present results provide no evidence for feature learning, in that participants performed no differently when manner of motion was specified independently of path (i.e., via the legs of the agent) than when manner of motion was specified by the same parts of the agent as was path (i.e., the whole body of the agent). Only one effect involving the attribute used to instantiate manner of motion even approached significance, namely the interaction of this attribute with learning order on correct path choices in the path trials.

This interaction was actually in the opposite direction to that predicted by feature learning. In particular, the difference between leg motion and manner of motion was greater in the PMC conditions than in the CMP conditions. If feature learning contributed to the effects found in this experiment and in Experiment 1, one would expect the type of manner of motion to matter more in the CMP condition. Participants in the CMP condition would have no opportunity to learn the separate identity of manner of motion while learning conjunctive verbs, making conjunctive verbs and path verbs more difficult to distinguish when the whole body of the agent was used to operationalize manner of motion. In contrast, the separate identity of leg motion from path would be perceptually specified, and thus conjunctive verbs and path verbs would be easier to distinguish when leg motion was used to operationalize manner of motion. These differences would be expected to be reduced in the PMC condition, where there was opportunity for feature learning. Yet the advantage was greater in the PMC condition than in the CMP condition. The results of this experiment provide no evidence for feature learning, but they are consistent with an attentional account.

Experiment 3

Whereas the results of the first two experiments are consistent with the combined effects of two opposing attentional processes, attentional persistence and contrast, path verb learning has provided evidence only for contrast whereas manner verb learning has provided evidence only for persistence. In particular, each attentional mechanism has resulted in extra attention to manner of motion. More convincing evidence for these two attentional mechanisms would come from a demonstration of how their competitive interaction results in increments or decrements in attention to an attribute.

Experiment 3 attempted to provide such a demonstration by decomposing the changes in attention to an individual attribute into persistence and contrast components. Three conditions were employed to this end (see Table 1). In the control condition, conjunctive verbs were learned first, each labeling a particular combination of values of path and leg motion. In the persistence condition, prior to conjunctive verb learning participants learned two verbs labeling leg motions that were different from those labeled by conjunctive verbs. Differences in performance between the control and persistence conditions provide a measure of the effects of persistence on attention to leg motion. Prior learning of leg motion verbs would heighten attention to leg motion, whereas differences in the particular values of leg motion labeled by the different verbs would minimize effects of contrast. In the contrast condition, however, the same values of leg motion were involved in both of these phases of learning. Thus, the difference between the contrast condition and the control condition provides a measure of the combined effects of persistence and contrast, whereas the difference between the contrast condition and the persistence condition provides a measure of contrast in isolation.

Several different measures of attention to leg motion and path were examined. We predicted that participants would attend more to leg motion in the persistence condition than in the control condition as a result of attentional persistence. We further predicted that participants would attend less to leg motion in the contrast condition than in the persistence condition as a result of contrast. Based on the results of Experiment 1, however, we predicted that participants would attend more to leg motion in the contrast condition than in the control condition. In particular, we predicted that persistence to the dimension of leg motion would dominate over contrast effects to particular values of leg motion in determining the overall amount of attention to leg motion.

Method

Stimuli

Learning Events. There were three phases of learning. In the conjunctive phase, each verb labeled a particular combination of path and leg motion. In the identical leg motion phase, each verb labeled a leg motion that was identical to the leg motions presented in the conjunctive phase. In the different leg motion phase, each verb labeled a leg motion different from those presented in the conjunctive phase. Path varied orthogonally to verb meaning in both leg motion phases of learning.

Criterion Test Trials. These were identical to those of Experiment 2 except for differences in the meanings of the verbs tested in each phase.

Final Test Events. These were different from those of Experiment 2 in that only two verbs were presented as possible labels for each event. There were seven different test types (see Table 2 for a schematic description). Three tested for attention to leg motion. In the first type (Leg Motion 1), path was not represented, with the agent sitting in place moving its legs in a particular leg motion. Participants chose between the two conjunctive verbs, one consistent with the presented leg motion and one inconsistent. In the second type (Leg Motion 2), a path and leg motion from the conjunctive phase were presented, but paired differently than in the conjunctive phase. Participants chose between a conjunctive verb consistent with the presented path but inconsistent with the presented leg motion and a verb from the identical leg motion phase that was consistent with both the leg motion and path. Making this second choice would indicate attention to leg motion. In the third type (Leg Motion 3), a path from the conjunctive phase and a leg motion from the different leg motion phase were presented. Participants chose between a verb from the conjunctive phase consistent with the path but inconsistent with the leg motion and a verb from the different leg motion phase that was consistent with both leg motion and path. Making this second choice would indicate attention to leg motion.

Four test types measured attention to path. In the first type (Path 1), leg motion was not represented, with the agent moving along its path while holding its legs stationary. Participants chose between two conjunctive verbs, one consistent with the presented path and one inconsistent. In the second type (Path 2), a path and leg motion from the conjunctive phase were presented, but paired differently than in the conjunctive phase. Participants chose between a conjunctive verb consistent with leg motion but inconsistent with path and a leg motion verb that was consistent with both path and leg motion. Making this second choice would indicate attention to path. In the third type (Path 3), path was not represented, as in the Leg Motion 1 trials. Participants chose between a conjunctive verb that was consistent with the presented leg motion but also made predictions for path and a leg motion verb that was consistent with both leg motion and path. Making this second choice would indicate attention to path. In the fourth type (Path 4), a path and leg motion from the conjunctive phase were presented, paired in the same way as in the conjunctive phase. Participants chose between a conjunctive verb and an identical leg motion verb, both of which were consistent with both path and leg motion. The conjunctive verb specifically predicted the presented path, however, whereas the leg motion verb made no predictions for path. Thus, participants were expected to choose the conjunctive verb to the extent that they attended to the path components of these verbs.

Procedure

The procedure was identical to that of Experiments 1 and 2 except for the three learning orders. Participants in the control condition received the conjunctive phase first, followed by the identical leg motion phase and then the different leg motion phase. Participants in the persistence condition received the different leg motion phase first, followed by the conjunctive phase and then the identical leg motion phase. Participants in the contrast condition received the identical leg motion phase first, followed by the conjunctive phase and then the different leg motion phase.

Design

The primary dependent measures were the percentages of correct responses to the 7 different final test types. The independent variable for these analyses was learning order (control vs. persistence vs. contrast). A secondary dependent variable was performance in the criterion test trials. The independent variables for this analysis were learning order and the type of verb being tested (conjunctive vs. different leg motion vs. identical leg motion).

Results

Criterion Test Trials

The results of the criterion test trials are displayed in Figure 14. As in Experiments 1 and 2, the results of trials following the first block of each phase of learning were analyzed in an ANOVA. This analysis revealed a main effect of verb type, F (2,330) = 14.18, p < .001, MSE = 526.63. Post-hoc t-tests revealed that participants performed better with conjunctive verbs than with identical leg motion verbs, t (167) = 4.79, p < .01, whereas there were no other significant differences among verb types. There was no main effect of learning order, F (2,165) = 0.79, p > .10, MSE = 605.02. There was, however, a significant interaction between verb type and learning order, F (4,330) = 8.52, p < .001, MSE = 526.63. Post-hoc LSD tests revealed that the persistence condition performed better than the other two conditions on identical leg motion verbs, whereas the persistence condition performed worse than the other two conditions on different leg motion verbs. There were no differences between the three conditions on conjunctive verbs. Thus, participants had difficulty with the first set of leg motion verbs they encountered, whereas they performed better with the second set of leg motion verbs and conjunctive verbs.

Final Test Trials

As in Experiments 1 and 2, participants were selected for analysis of final test scores if they performed with 100% accuracy on at least one block of criterion test trials for each verb type. The numbers of participants in each condition who achieved this criterion are as follows: 43 in the control condition, 50 in the persistence condition, and 46 in the contrast condition.

The effects of learning order on performance in the final test trials were analyzed using two orthogonal planned comparisons. The first comparison tested for attentional persistence. This comparison pitted the control condition against the persistence and contrast conditions. These last two conditions were grouped together in this analysis because they had in common a leg motion phase of learning prior to conjunctive learning, which would be expected to draw attention to leg motion as a dimension. The second comparison tested for contrast. This comparison pitted the persistence condition against the contrast condition. These two conditions varied in the particular values of leg motion that were employed. The use of the same leg motions in the first phase of learning and the conjunctive phase of learning was expected to result in reduced attention to leg motion during conjunctive learning in the contrast condition compared to the persistence condition.

The results of three different tests of attention to leg motion are depicted in Figure 15. A contrast analysis on Leg Motion 1 trials with learning order analyzed using the above planned comparisons revealed significant persistence, t (136) = 3.41, p < .001, with less attention to leg motion in the control conditions compared to the other two conditions. There was also evidence for contrast, t (136) = 2.04, p < .05, with greater attention to leg motion in the persistence condition than in the contrast condition. Despite this effect of contrast, there was a trend toward greater attention to leg motion in the contrast condition than in the control condition, t (87) = 1.84, p < .07. The analysis of Leg Motion 2 trials revealed no evidence for persistence, t (136) = 0.80, p > .10, although the trend was in the predicted direction. The effect of contrast approached significance, t (136) = 1.90, p < .06. The analysis of Leg Motion 3 trials revealed no significant effects.

The results of four different tests of attention to path are depicted in Figure 16. A contrast analysis on Path 1 trials revealed significant persistence to leg motion, t (136) = 4.25, p < .001, with less attention paid to path in the persistence and contrast conditions than in the control condition. The effect of contrast did not reach significance, t (136) = 0.83, p > .10. The analysis of Path 2 trials revealed significant persistence, t (136) = 2.73, p < .01, as well as significant contrast, t (136) = 1.98, p < .05, with greater attention to path in the contrast condition than in the persistence condition. The analysis of Path 3 trials revealed significant persistence, t (136) = 3.38, p < .001, whereas there was not a significant effect of contrast, t (136) = 0.02, p > .10. Despite the effects of contrast, attention to path was significantly greater in the control condition than in the contrast condition in both the Path 1 trials, t (87) = 2.97, p < .01, and the Path 2 trials, t (87) = 2.84, p < .01, whereas this effect was not significant in the Path 3 trials, t (87) = 1.42, p > .10. The analysis of Path 4 trials revealed no significant effects.

Discussion

The results of Experiment 3 revealed that attentional persistence and contrast together determine the total amount of attention paid to an attribute. In particular, persistence following leg motion training resulted in increased attention to leg motion and decreased attention to path compared to a control condition that was given no prior leg motion training. If the same values of leg motion were seen in the first two phases of learning, however, contrast resulted in decreased attention to leg motion and increased attention to path compared to a condition in which different values of leg motion were seen in the first two phases. Thus, attentional persistence seems to operate at the level of attributes or dimensions, whereas contrast effects appear to be realized primarily at the level of values.

Even when the same values of leg motion were involved in leg motion training and conjunctive training, attention to leg motion was increased and attention to path was decreased when leg motion training preceded conjunctive training compared to a control condition in which conjunctive training came first. Thus, persistence dominated over contrast in determining the total attention directed to leg motion and path. This finding may reflect the baseline salience of the two attributes. Attention to leg motion was initially low, so mechanisms that served to increase attention to leg motion (i.e., persistence) were able to exert a large influence on attention. Attention to path was initially high, so mechanisms that served to further increase attention to path (i.e., contrast) exerted little influence.

The results of Experiment 3 are for the most part consistent with a model in which the total pool of attention is fixed (e.g., Sutherland & Mackintosh, 1971). In such a model, increases in attention to one attribute necessarily result in decreased attention to other attributes. In Experiment 3, increases in attention to leg motion were apparently accompanied by decreases in attention to path. The most straightforward comparison of attention to leg motion and path involves the Leg Motion 1 and Path 1 trials, which were highly analogous. Such a comparison reveals that the total amount of attention to leg motion and path is quite similar across the three conditions, even though the distribution of attention over these two attributes varies considerably. This null finding, however, does not disconfirm models that allow changes in the total pool of attention (e.g., Mackintosh, 1975).

General Discussion

The present experiments provide evidence for two competing attentional mechanisms. Attentional persistence serves to increase attention to dimensions previously found to be predictive. Contrast serves to increase attention to stimuli whose predictive power has yet to be discovered. Experiment 1 demonstrated that attention to manner of motion, an attribute whose salience was initially low, could be increased as a result of prior training in which manner of motion was relevant. This finding provides evidence for attentional persistence. Attention to manner was also increased following path training, however, providing evidence for contrast. Experiment 2 replicated these findings using path and leg motion as attributes. Unlike path and manner of motion, path and leg motion are likely to be perceptually specified as separate features. Thus, these results cannot be explained in terms of functional feature creation. Experiment 3 decomposed changes in attention to an attribute into persistence and contrast components, providing evidence for a competitive interaction between these two attentional mechanisms.

Feature Creation and Dimensional Attention

The results of the present experiments appear to be better explained by the operations of attentional mechanisms than by feature creation. Feature creation and dimensional attention may play intimately related roles in categorization, however. When confronted with a novel category label, a categorizer may look first to his or her existing featural repertoire in order to determine whether any of those features can be found in the sensory environment. Dimensional attention may determine which features are most likely to be sampled in this way. If no matching features are discovered, a new feature may be created based on some salient characteristic of the present situation. Dimensional attention may again play a role in this process of feature creation, determining which novel features are most salient. For example, after learning a number of words that refer to shape, a child may be more likely to create a new shape feature to categorize a novel object, as a result of attentional persistence. An association can then be forged between this new feature and its corresponding category.

If a match is found between an existing feature and a characteristic of the environment, the category associated with this feature can be compared to the presented category label. If there is a match between these two sources of information, the association between the feature and its category is increased. In addition, attention to the dimension of that feature may be increased because of attentional persistence. If the category associated with a feature is different from the presented category, the association between the feature and its category may be decreased. In addition, attention to the dimension of that feature may be decreased because of contrast. As a result, other dimensions of the environment would become more likely to be sampled, possibly resulting in the discovery of new predictive features from these other dimensions.

This account of categorization becomes more complex if one views the identification of features as a categorization problem in its own right, as do Schyns, Goldstone, & Thibaut (in press). The confidence with which one categorizes a feature may influence the magnitude of the shifts in attention associated with persistence and contrast. For example, the two different path features in the present experiments were quite different, with one path involving a straight line and the other involving a 90 degree turn. As a result, when a novel category label was presented along with a path that was already associated with a different category, participants were likely to shift attention to other attributes because of contrast. This tendency may have been weaker if a novel category label had been presented along with a path that was similar but not identical to a path that was already associated with a category. In particular, participants may have been less likely to shift attention because they were less confident that this new path was an example of the previously-presented category of paths. Further research involving quantitatively varying features could be used to examine this issue.

The Role of Contrast in the Acquisition of Relational Terms

The present research is consistent with the work of Merriman, Marzita, and Jarvis (1993) and Merriman, Evey-Burkey, Marazita, and Jarvis (1996) in revealing a bias to associate novel verbs with previously unnamed referents. There is an important difference in the methodologies of the present research and of this previous research, however. Merriman and colleagues found that children were more likely to associate a novel verb with an event that was entirely novel than with an event that already had a label. The size of this effect was weaker than that generally found with nouns. Merriman et al. (1993) proposed that the weaker novelty effect associated with verbs may be a result of differences in organization between noun and verb meanings. Whereas nouns tend to be organized hierarchically, verbs tend to be organized in a matrix, with elements of verb meaning (e.g., path, manner of motion) combining interchangeably (Huttenlocher & Lui, 1979). As a result, the same event can have many different labels depending on which aspect of the event is being focused on (e.g., "entering" vs. "walking").

The present research revealed evidence for contrast at the level of dimension values rather than at the level of entire events. In particular, participants associated verbs with a previously unnamed value of a dimension rather than with a previously named value of another dimension. This attentional mechanism is particularly well-suited for the learning of verbs because verbs can refer to so many different aspects of an event. Thus, if a familiar event is encountered along with a novel verb, it is a particularly useful strategy to shift attention away from previously labeled aspects of the event because other aspects are likely to be relevant.

Contrast may be particularly useful in learning English verbs because such verbs often refer to the manner of motion of an object (Talmy, 1985). As has been demonstrated in this and previous research (Kersten, in press; Kersten, 1997), manner of motion is less salient than path in the context of verb learning. Indeed, children learning English as well as other quite different languages such as Korean tend to learn path terms before manner of motion terms (Choi & Bowerman, 1991). The present research suggests that this learning of path terms may in fact facilitate the later acquisition of manner of motion terms through attentional contrast.

Conclusions

This research provides evidence for two competing attentional mechanisms. Persistence seems to operate primarily at the level of dimensions, serving to draw attention to dimensions that have previously been predictive of an outcome, regardless of the values involved on that dimension. Contrast seems to operate primarily at the level of dimension values, serving to divert attention away from values that are already associated with a category or response. The presence of these two attentional processes suggests that if one wants to encourage attention to a particular dimension, the best strategy is to give a learner prior training on the same dimension, but with different values on the dimension.

References

Choi, S., & Bowerman, M. (1991). Learning to express motion events in English and Korean: The influence of language-specific lexicalization patterns. Cognition, 41, 83-121.

Goldstone, R.L. (1994). Influences of categorization on perceptual discrimination. Journal of Experimental Psychology: General, 123, 178-200.

Huttenlocher, J., & Lui, F. (1979). The semantic organization of some simple nouns and verbs. Journal of Verbal Learning and Verbal Behavior, 18, 141-162.

Kersten, A.W. (in press). A division of labor between nouns and verbs in the representation of motion. Journal of Experimental Psychology: General.

Kersten, A.W. (1997). Decomposing a motion event into path and manner of motion components: Evidence from differential association with nouns and verbs. Unpublished manuscript.

Kersten, A.W., & Billman, D. (1995). The roles of motion and moving parts in noun and verb meanings. In J.D. Moore & J.F. Lehman (Eds.), Proceedings of the seventeenth annual conference of the cognitive science society (pp. 625-630). Hillsdale, NJ: Erlbaum.

Kersten, A.W., & Smith, L.B. (1997). Different noun and verb learning biases in 3 1/2- to 4-year-old children. Manuscript submitted for publication.

Kruschke, J.K. (1992). ALCOVE: An exemplar-based connectionist model of category learning. Psychological Review, 99, 22-44.

Kruschke, J.K. (1996). Base rates in category learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 3-26.

Landau, B., Smith, L.B., & Jones, S.S. (1988). The importance of shape in early lexical learning. Cognitive Development, 3, 299-321.

Lawrence, D.H. (1949). Acquired distinctiveness of cues: I. Transfer between discriminations on the basis of familiarity with the stimulus. Journal of Experimental Psychology, 39, 770-784.

Lubow, R.E., & Moore, A.U. (1959). Latent inhibition: The effect of nonreinforced preexposure to the conditional stimulus. Journal of Comparative and Physiological Psychology, 52, 415-419.

Mackintosh, N.J. (1969). Further analysis of the overtraining reversal effect. Journal of Comparative and Physiological Psychology, (Monograph supplement) 67, No. 2, part 2.

Mackintosh, N.J. (1975). Blocking of conditioned suppression: Role of the first compound trial. Journal of Experimental Psychology: Animal Behavior Processes, 104, 335-345.

Mackintosh, N. J., & Little, L. (1969). Intradimensional and extradimensional shift learning by pigeons. Psychonomics Science, 14, 5-6.

Markman, E.M., & Wachtel, G.F. (1988). Children's use of mutual exclusivity to constrain the meanings of words. Cognitive Psychology, 20, 121-157.

Merriman, W.E., Evey-Burkey, J.A., Marazita, J.M., & Jarvis, L.H. (1996). Young two-year-olds' tendency to map novel verbs onto novel actions. Journal of Experimental Child Psychology, 63, 466-498.

Merriman, W.E., Marazita, J., & Jarvis, L.H. (1993). Four-year-olds' disambiguation of action and object word reference. Journal of Experimental Child Psychology, 56, 412-430.

Nosofsky, R.M. (1986). Attention, similarity, and the identification-categorization relationship. Journal of Experimental Psychology: General, 115, 39-57.

Pearce, J.M., & Hall, G. (1980). A model for Pavlovian learning: Variations in the effectiveness of conditioned but not of unconditioned stimuli. Psychological Review, 87, 532-552.

Schyns, P.G., Goldstone, R.L., & Thibaut, J.P. (in press). The development of features in object concepts. Behavioral and Brain Sciences.

Schyns, P.G., & Rodet, L. (1997). Categorization creates functional features. Journal of Experimental Psychology: Learning, Memory, and Cognition, 23, 681-696.

Shepard, R.N., Hovland, C.L., & Jenkins, H.M. (1961). Learning and memorization of classifications. Psychological Monographs, 75 (13, Whole No. 517).

Sutherland, N.S., & Mackintosh, N. J. (1971). Mechanisms of animal discrimination learning. New York: Academic Press.

Talmy, L. (1985). Lexicalization patterns: Semantic structure in lexical forms. In T. Shopen (Ed.), Language typology and linguistic description: Vol. 3. Grammatical categories and the lexicon (pp. 57-144). New York: Cambridge University Press.

Author Note

Please address correspondence to: Alan Kersten, Department of Psychology, Indiana University, Bloomington, IN 47405-1301, email: akersten@indiana.edu, fax (812) 855-4691. Examples of the events used as stimuli in this research are available over the World Wide Web at http://php.ucs.indiana.edu/~akersten.

Table 1

Schematic description of the three learning conditions in Experiment 3

__________________________________________________________

Condition Phase 1 Phase 2 Phase 3

Control P1LM1 vs. P2LM2 LM1 vs. LM2 LM3 vs. LM4

Persistence LM3 vs. LM4 P1LM1 vs. P2LM2 LM1 vs. LM2

Contrast LM1 vs. LM2 P1LM1 vs. P2LM2 LM3 vs. LM4

__________________________________________________________

Notes. Subscripted numbers represent values of leg motion and path. An attribute that is not represented for a given phase of learning varied randomly during that phase.

Table 2

Schematic description of the test stimuli in Experiment 3

__________________________________________________________

Learning Stimuli Verbs Test Type Stimuli Verb Choices

P1LM1 morping Leg Motion 1 P0LM1 morping or wunking

P2LM2 wunking Leg Motion 1 P0LM2 morping or wunking

P1 or 2LM1 spogging Leg Motion 2 P1LM2 morping or zelling

P1 or 2LM2 zelling Leg Motion 2 P2LM1 wunking or spogging

P1 or 2LM3 yimming Leg Motion 3 P1LM3 morping or yimming

P1 or 2LM4 gupping Leg Motion 3 P1LM4 morping or gupping

Leg Motion 3 P2LM3 wunking or yimming

Leg Motion 3 P2LM4 wunking or gupping

Path 1 P1LM0 morping or wunking

Path 1 P2LM0 morping or wunking

Path 2 P1LM2 wunking or zelling

Path 2 P2LM1 morping or spogging

Path 3 P0LM1 morping or spogging

Path 3 P0LM2 wunking or zelling

Path 4 P1LM1 morping or spogging

Path 4 P2LM2 wunking or zelling

__________________________________________________________

Notes. Subscripted numbers represent values of leg motion and path. A value of 0 indicates that the value of that attribute was not represented in the stimuli.

Figure Captions

Figure 1. Three possible features that could be derived from the learning events of Experiment 1. The first two are independent path and manner of motion features. The third is a unitary, conjoined representation of both path and manner of motion.

Figure 2. Three frames from an example event in Experiment 1. The agent, on the right, moves along a straight path to the patient, on the left. The event is labeled by the verb "gupping," and takes place against a mountainous background.

Figure 3. Values of the motion attributes in Experiment 1.

Figure 4. Results of the criterion test trials in Experiment 1. The six conditions reflect the six orders of learning (e.g., P -> M -> C = Path -> Manner -> Conjunctive). The numbers on the x-axis refer to the block of learning for each verb type. Participants who were 100% correct following a given block of learning trials proceeded to the next phase of the experiment without further blocks of trials on that verb type. Thus, the only data points that reflect the results of all participants are the first block of trials with each verb type. The results of the second and third blocks are less stable because of the smaller number of participants in those blocks.

Figure 5. Percent correct conjunctive choices in the conjunctive and carpet trials of Experiment 1. Error bars reflect standard errors.

Figure 6. Percent of choices in the conjunctive and carpet trials of Experiment 1 involving path verbs that matched the test event on path but failed to predict the manner of motion displayed in the event.

Figure 7. Percent correct path choices in the path trials of Experiment 1.

Figure 8. Percent correct manner choices in the manner trials of Experiment 1.

Figure 9. Results of the Criterion Test Trials in Experiment 2. BM = body motion. LM = leg motion.

Figure 10. Percent correct conjunctive choices in the conjunctive and carpet trials of Experiment 2.

Figure 11. Percent of choices in the conjunctive and carpet trials of Experiment 2 involving path verbs that matched the test event on path but failed to predict the manner of motion evident in the event.

Figure 12. Percent correct path choices in the path trials of Experiment 2.

Figure 13. Percent of choices in the path trials of Experiment 2 involving conjunctive verbs that matched the test event on path but also predicted a manner of motion that was not evident in the test event.

Figure 14. Results of the Criterion Test Trials of Experiment 3.

Figure 15. Results of the Leg Motion trials in Experiment 3.

Figure 16. Results of the Path trials in Experiment 3.

Path

Manner of Motion

Conjoined Path and Manner

______________________________________________________________________

______________________________________________________________________

Path


Manner of Motion


Leg Motion