Implicit artificial grammar learning: effects of complexity and usefulness
of the structure
Bos, E.J. van den
Citation
Bos, E. J. van den. (2007, June 6). Implicit artificial grammar learning: effects of complexity and usefulness of the structure. Department of Cognitive Psychology, Leiden University Institute for Psychological Research, Faculty of Social Sciences, Leiden University. Retrieved from https://hdl.handle.net/1887/12037
Version: Corrected Publisher’s Version
License: Licence agreement concerning inclusion of doctoral thesis in the Institutional Repository of the University of Leiden
Downloaded from: https://hdl.handle.net/1887/12037
Note: To cite this publication please use the final published version (if applicable).
Chapter 5
The occurrence of implicit learning:
When useful structures draw attention
Abstract
The automatic abstraction view (Reber, 1989) and the episodic processing account (Wright & Whittlesea, 1998) make different predictions on the occurrence of implicit learning. Two artificial grammar learning experiments compared these predictions with an alternative proposal that implicit structure learning reliably occurs whenever the structure is useful to one’s current task. The results were in line with the latter hypothesis. Learning was observed when the structure was useful to the
participants’ task in the induction phase, but not when it was useless, which cannot be explained by the automatic abstraction view of implicit learning. A third experiment, using an adapted process dissociation procedure (Jacoby, 1991), showed that
incidental learning of useful structures can produce knowledge that is implicit in the sense that its application is difficult to control.
Introduction
Implicit learning is commonly defined as incidental acquisition of knowledge that is not fully accessible to consciousness (Reber, 1989; Seger, 1994). Two
experimental paradigms in which such learning is observed are Artificial Grammar Learning (AGL; Reber, 1967) and the Serial Reaction Time (SRT) task (Nissen &
Bullemer, 1987). In the induction phase of an AGL-experiment, participants are typically instructed to memorize letter strings. Subsequently, they are informed that the letter strings have been generated by an artificial grammar (see Figure 1 on page 66 for an example). In the test phase, knowledge of the grammar is indicated by above chance discrimination between new grammatical and ungrammatical strings.
Participants in an SRT-task have to respond to a stimulus appearing in one of several locations on a computer screen by pressing the key corresponding to that location.
Unknown to the participants, the stimulus follows a regular sequence of locations on most trials. Implicit learning is evidenced by participants’ reaction times, which decrease for stimuli that follow the sequence as compared to stimuli that violate it.
Reviews of the literature have identified some boundary conditions for the occurrence of implicit learning. For example, the process seems to be tuned to a certain kind of structure, described by Seger (1994) as “correlated features” and by Reber (1989) as “patterns of covariation in the environment”. Structures consisting of deterministic relationships between non-adjacent elements seem unsuitable for implicit learning (Matthews et al., 1989; Shanks, Johnstone & Staggs, 1997). In addition, implicit learning can only be observed for structures of sufficient
complexity. If the structure is too simple, people will become aware of it and learning will be explicit (Reber, 1989; Seger, 1994).
However, even when these boundary conditions are met, two influential views on implicit learning make different predictions regarding its occurrence. Some have claimed that implicit learning is an automatic consequence of processing structured stimuli (e.g. Reber, 1989; Servan-Schreiber & Anderson, 1990), whereas others have claimed that it is accidental, depending on a complex interplay of many factors (Whittlesea & Dorken, 1993; Whittlesea & Wright, 1997; Wright & Whittlesea, 1998). Below, we will discuss these contrasting positions in the literature as well as a third alternative, which suggests one type of situation in which the process should reliably occur.
The occurrence of implicit learning
In Reber’s (1989) view, implicit learning is an unintentional and even ineluctable process that automatically acquires abstract knowledge of patterns in the environment. Attending to the stimuli would be sufficient for implicit learning to occur. In support of this position, Reber and Allen (1978) demonstrated artificial grammar learning by participants who received no other instructions than to observe the stimuli. Similarly, implicit learning could be modeled as an automatic
consequence of processing structured materials (Servan-Schreiber & Anderson, 1990). The competitive chunking model was able to simulate the results of AGL- experiments by automatically forming and strengthening letter chunks in the process of reading. When the same chunks were used in reading a new string, this resulted in a high familiarity score, which the model could use to make a grammaticality judgment.
By contrast, Whittlesea and Wright (1997) reported an experiment in which participants failed to learn a structure despite studying structured items. Participants who studied non-words that differed from a prototype by one letter demonstrated structure learning by falsely recognizing the prototype. Participants who, in contrast,
studied words that differed from a prototype by one letter did not show structure learning. The authors proposed that accidental characteristics of the experimental situation affect the way participants process structured stimuli and, therefore, what they learn.
Several experiments provided evidence for this claim. The kind of knowledge acquired in implicit learning experiments was shown to be affected by the task participants perform (Whittlesea & Dorken, 1993), by characteristics of the stimuli, such as familiarity, similarity and salience (Whittlesea & Wright, 1997), and by the (spatial) context in which the stimuli are presented (Wright & Whittlesea, 1998). The authors concluded that structure has no special status (Whittlesea & Wright, 1997).
According to their episodic processing account, sensitivity to this attribute (at test) is due to accidental overlap in information processing with earlier (learning) situations (Wright & Whittlesea, 1998). This implies that the occurrence of implicit structure learning is unpredictable.
Contrary to the assumptions underlying this view, however, research in the contextual cueing paradigm has provided some evidence that people learn to systematically attend to useful information. In this paradigm, participants are presented with search displays containing one target and several distracters. Half of the displays have configurations of targets and distracters that are repeated during the experiment. For these configurations, participants become increasingly faster at localizing the targets, because they learn where to attend. However, they are unable to distinguish between old and new configurations on a subsequent recognition test (Chun & Jiang, 1998).
Endo and Takeda (2004) found that participants learned to attend to the information that was most useful to the search task. In their experiments, the location of the target could be predicted by both the configuration and the identity of the distracters. When both were predictive of target location, attention was guided by the strongest predictor and only distracter configuration was learned. However, when each relationship predicted the target location on half of the trials, both were learned.
When distracter configuration was made less informative than distracter identity, only the identity predictor was learned.
These findings suggest an interactive relationship between attention and implicit learning. We propose that people implicitly learn to attend to information that is useful to their current task, while increased attention, in turn, enhances encoding of
this information. In this view, because attention is drawn to the aspect of a stimulus that is most useful to one’s current task, processing structured stimuli does not automatically result in implicit structure learning. Nevertheless, the phenomenon is not accidental, as the structure will draw attention whenever it is useful. Two experiments were performed to test the hypothesis that implicit artificial grammar learning occurs more reliably when the structure is useful to the participants’ task than when it is useless. A third experiment investigated whether implicit learning of a useful structure indeed produces implicit knowledge.
Implicit knowledge
As noted above, implicit learning is thought to produce knowledge that is in some way different from explicit knowledge. Traditionally, the two types of
knowledge have been proposed to differ in their accessibility to awareness (Reber, 1989). Although numerous studies have demonstrated that incidental learning can result in knowledge that is accessible to awareness (e.g. Dulaney, Carlson & Dewey, 1984; Gomez, 1997; Perruchet & Pacteau, 1990), there is also evidence that such knowledge is insufficient to explain performance (Mathews et al., 1989) and that it may not actually be used in making grammaticality judgments (Meulemans & Van der Linden, 2003). Knowledge gained from implicit as opposed to explicit learning is therefore generally considered to be not completely accessible to awareness (e.g.
Gomez, 1997; Reber, 1989; Seger, 1994).
A second way in which knowledge gained from incidental learning may differ from explicit knowledge is that people may not have conscious control over its application. In a study of recognition memory, Jacoby (1991) developed a procedure to estimate the contributions of controlled and automatic influences to a task. This process dissociation procedure contrasted performance on a recognition task when participants were instructed to treat a category of items as old (inclusion condition) with performance when they were instructed to treat that category as new (exclusion condition). In the inclusion condition, both controlled and automatic influences of previously processing an item lead to an ‘old’-response. In the exclusion condition, however, ‘old’-responses to items of the target category can only be due to automatic influences.
A few studies have applied process dissociation procedures to implicit learning. In the SRT-paradigm, generation tasks under inclusion and exclusion instructions have revealed that participants have some degree of control over their
knowledge. Participants produced more fragments of their training sequence in the inclusion condition than in the exclusion condition (Shanks, Rowland & Ranger, 2005). However, there was also evidence of automatic knowledge application:
participants could not entirely prevent themselves from producing old fragments in the exclusion condition (Destrebecqz & Cleeremans, 2001).
In the AGL-paradigm, Dienes, Altmann, Kwan and Goode (1995) only found evidence of controlled knowledge application. Participants first studied exemplars from one grammar and then studied exemplars from a second grammar. Subsequently, they were able to selectively apply one of the grammars to a grammaticality judgment test, without automatically accepting exemplars from the other. In the present study, the process dissociation procedure was adapted for use with a single artificial grammar. Instead of assessing whether participants had control over which grammar to apply, it was tested whether participants were able to refrain from applying knowledge of the only grammar they had been exposed to.
The aim of Experiments 1 and 2 was to identify situations in which an artificial grammar could be learned incidentally. It was hypothesized that artificial grammar learning would reliably occur if the structure was useful to the participants’
task, but not if it was useless. In Experiment 3, a modified process dissociation procedure was used to explore whether incidental learning of useful structures is different from explicit learning in the sense that the two learning modes produce knowledge over which participants have different degrees of control.
Experiment 1
In most experiments demonstrating successful artificial grammar learning, participants were instructed to memorize letter strings in the induction phase. This can be considered as a task to which the structure is useful, as Reber (1967) has
demonstrated that participants are better at memorizing letter strings that have been generated by an artificial grammar than letter strings that have been generated randomly. Experiment 1 tested the hypothesis that artificial grammar learning would not occur if the grammar was useless with respect to the task in the induction phase.
There were two conditions in which the structure was useless. In one
condition, participants had to compute the total value of a string from the values of its elements. The grammar was irrelevant to this task. In the other condition, participants had to rate how much they liked each string. Structure was useless, because there are
no objective criteria for performance; the grammar could not facilitate the affective task. Performance in the structure useless computing and rating conditions was compared with performance in a structure useful memorize condition.
Methods
Participants. Sixty-four undergraduate students of Leiden University (13 male, 51 female; 18-27 years of age) participated in the experiment. They received either course credits or money for their participation. The reward depended on the duration of the experiment; experimental participants were paid € 4 and control participants were paid € 2.50.
0 0
1
2
3
4 DIR
DIR
DIR BOG
BOG
BOG
TAF
TAF
NUP
KES KES
Figure 1. The artificial grammar used in the present study (based on Howard & Ballas, 1980).
Materials. The stimuli in the experiment were sentences of non-words (in Dutch) generated by a finite-state grammar adapted from Howard and Ballas (1980), see Figure 1 and Appendix B. The grammar could generate 59 unique sentences of three to eight non-words. Four sentences were presented as examples in the cover- story and as stimuli on the practice trials. Fifteen sentences were assigned to the induction phase and 40 to the test phase, so that the paths of the grammar were equally represented in both phases of the experiment. One half of the stimuli in the test phase were unaltered grammatical sentences. The other half was made
ungrammatical by switching two adjacent non-words in the middle of each sentence (excluding the first and last).
All stimuli were displayed on a computer monitor as black text (Courier New 18, bold) against a white background. The instructions were accompanied by a picture of an ice-cream van and by four pictures of ice-creams. In the induction phase, a picture of a customer was presented above each sentence; no ice-creams were shown.
In the test phase, the sentences were presented in isolation and followed by a picture of the customer, who either paid or walked away. Participants were seated in front of the computer monitor at a distance of about 50 cm. They reacted by pressing keys on a keyboard.
Procedure. Participants were tested individually in a dimly lit room. The experimenter stayed with them to check that they carried out the instruction of reading aloud, but remained out of the participants’ sight. The experiment was introduced in a cover-story, informing participants that they would be employed at an ice-cream van on an imaginary planet where people spoke an unfamiliar language. The 48
experimental participants were presented with four examples of sentences from this language and pictures of the corresponding ice-creams.
Then, the experimental participants received their instructions for the induction phase. They were informed that the boss of the ice-cream van was aware that they did not speak the language yet and had set them a task to prepare for their new job. All experimental participants were instructed to read aloud the sentences of 30 customers. In addition, they had to perform one of three tasks: memorizing the sentences (structure useful), rating how much they liked each sentence describing an ice-cream (structure useless) or computing the price of each ice-cream (structure useless).
The experimental participants completed 2 practice trials. Then they were presented with 2 consecutive blocks of 15 experimental trials. The order of the trials was random within each block. Each trial started with a picture of a customer. After 1 second, a non-word sentence was added to the display. Participants in the memorize- condition had to press the space-bar to end the trial when they thought they knew the sentence. Participants who had to indicate for each sentence how good they thought the ice-cream was were, in addition, presented with a rating scale, ranging from 1 (very bad) to 5 (very good). The trial ended when they pressed any number in this range. Participants in the compute-condition were presented with a price-list instead
of a rating scale, associating each non-word with a value. Although the sentence appeared in its complete form, participants had to perform a stepwise computation of the price of the ice-cream, entering subtotals for every non-word, to make sure that the non-words were processed in the order in which they appeared in the sentence.
After entering the total price of the ice-cream, they were shown the correct answer for 2 seconds, accompanied by the word “good!” if their answer was right.
After the induction phase, the experimental participants were informed that, so far, they had only seen sentences referring to ice-creams that were for sale, but that not all ice-creams were for sale on the planet. It was announced that they would be presented with sentences from new customers. Some would order ice-creams that were for sale (highly similar to the sentences they had seen before) and some would order ice-creams that were not for sale (less similar to the sentences they had seen before). The participants’ task was to decide whether or not the sentence referred to an ice-cream that was for sale, by pressing ‘j’ for yes or ‘n’ for no. They were informed that each time they pressed ‘j’, they would see the customer pay and each time they pressed ‘n’, they would see the customer walk away. The 16 control participants entered the test phase without having been exposed to any non-word sentences. Their instructions were the same as for the experimental participants, except that no
reference was made to the induction phase.
Participants completed 40 test trials. Grammatical and ungrammatical sentences were presented in random order. On each trial, a non-word sentence was presented in the middle of the screen until the participant pressed either the ‘j’ or the
‘n’ key. After the participant had pressed a key, one of two pictures was shown for one second, depending on the response (but not on its accuracy). After the test phase, participants were thanked for their participation. The experiment took about 20 minutes.
Results
Induction phase. A one-way analysis of variance (ANOVA) showed that the duration of the induction phase was different for the three experimental conditions (F(2,45) = 24.11, p < .001). The rating task took less time than the task to memorize each sentence (p = .009) and the task to compute the price of each ice cream (p <
.001). In the compute condition, the mean proportion of correct prices was .935 (SD = .055).
Test phase. A one-way ANOVA showed that the proportion of correct
grammaticality judgments was significantly different in the four conditions (F(3,60) = 4.11, p = .01). Orthogonal contrasts showed that proportion correct was higher in the memorize condition (M = .592, SD = .085) than in the other conditions (t(60) = 3.45, p = .001). In addition, they showed that the structure-useless rating (M = .527, SD = .069) and compute (M = .523, SD = .065) conditions did not differ from each other (t(60) = .123, p = .903) or from the control condition (M = .511, SD = .068, t(60) = .637, p = .527).
Discussion
The results of Experiment 1 were in line with the hypothesis that artificial grammar learning would not occur if the grammar was useless with respect to one’s current task. Participants who had performed a task to which the structure of the non- word sentences was useless were not better at judging the grammaticality of new sentences than participants who had no prior experience with sentences generated by an artificial grammar. Only participants whose task of memorizing non-word
sentences could be facilitated by the common structure were successful at the grammaticality judgment test.
The results for the rating condition were highly similar to those of a previous study in which the induction phase consisted of an affective task. McAndrews and Moscovitch (1985) presented participants with pairs of grammatical letter strings and instructed them to make a preference judgment. As in the present study, this led to a proportion correct of .53 on a subsequent grammaticality judgment test, which was probably not significant, as the authors reported high variability in their data.
Although liking judgments can be used in the test phase to reveal structure knowledge (Helman & Berry, 2003; Manza & Bornstein, 1995), these findings suggest that affective rating of structured stimuli does not reliably induce artificial grammar learning.
Experiment 1 provides a first indication that the occurrence of implicit structure learning depends on the structure’s usefulness. However, successful
performance in the standard memorize condition could be due to other characteristics of the task as well. Similarly, high demands on attentional resources or shallow processing could be responsible for the lack of structure learning in the useless conditions. Although computing the price of each ice-cream only involved adding values in the range of 0 to 3 and the outcome was never more than 11, one might still
argue that the task was quite attention demanding. Unsuccessful performance in the compute-condition could then be attributed to a lack of attentional resources.
Likewise, performance in the rating condition could be explained by a ‘levels of processing’-account. Craik and Lockhart (1972) claimed that stronger memory traces are formed as stimuli are processed at deeper levels of analysis. Processing stimuli for meaning results in stronger memory traces than analyzing abstract patterns, which produces stronger traces than analyzing sensory features. As participants in the rating condition took less time to process the sentences than participants in the memorize condition, they may have restricted their analyses to the sensory level, forming traces that were too weak to affect later grammaticality judgments. The purpose of Experiment 2 was to rule out these alternative explanations.
Experiment 2
Experiment 2 aimed at providing a more stringent test of the hypothesis that implicit learning occurs if the structure is useful to the participants’ current task, but not if it is useless. A sentence decoding task, to which learning the structure would be useful, was developed to rule out the possibility that structure learning depended on other characteristics of the memorize task and to explore the generality of the findings from Experiment 1. This task was compared with a decoding task that could not be facilitated by knowledge of the grammar. To decrease the likelihood that a lack of structure learning in this condition could be attributed to other factors, such as higher demands on attentional resources or more shallow levels of processing, the structure- useful and structure-useless tasks were kept as similar as possible.
In both conditions, participants were presented with a non-word sentence and three pictures of ice-creams. Their task was to guess which ice-cream the sentence referred to. In the structure-useless condition, participants could perform this task by learning the meanings of the individual non-words. In the structure-useful condition, participants had to learn both the meanings of the individual non-words and whether two of the non-words related to the preceding or to the following non-word. It was predicted that only participants for whom the order of the words had been useful to perform their task would be able to subsequently discriminate between grammatical and ungrammatical sentences.
Methods
Participants. Thirty undergraduate students of Leiden University (3 male, 27 female; 17-33 years of age) participated in the experiment. None of them had participated in Experiment 1. They received either course credits or € 4.50 for their participation.
Materials. The non-word sentences used in Experiment 2 were the same as in Experiment 1. In the induction phase, each sentence was now accompanied by three pictures of ice-creams. In the structure-useless condition, each non-word referred to a color (BOG = yellow, DIR = red, KES = green, NUP = brown, TAF = pink). For the ice-cream the sentence referred to, the color of the balls corresponded with the meaning of the non-words. The sentences described all ice-creams from left to right and from bottom to top. One incorrect alternative was created by using a new, randomly assigned, mapping of colors to the non-words in the sentence. A second incorrect alternative was created by substituting the color of one ball by one of the other colors.
In the structure-useful condition, BOG, DIR and NUP referred to the same colors as in the structure-useless condition. KES and TAF, however, had different referents: KES meant ‘with sprinkles’ and referred to the word preceding it, TAF meant ‘extra large’ and referred to the word following it. One incorrect alternative was created by either putting sprinkles on the ball following (instead of preceding) KES or enlarging the ball preceding (instead of following) TAF. The other incorrect alternative was again created by substituting the color of one ball by one of the other colors. For one sentence, that contained neither KES nor TAF, two incorrect
alternatives were created in this way.
Procedure. The procedure for the induction phase was the same as in Experiment 1, except that the participants’ task was to guess which of three ice- creams each sentence referred to. Participants responded by pressing 1, 2 or 3 on the keyboard. The location of the correct ice-cream (left, center, right) was balanced within each of the two blocks. The location where each picture appeared was balanced across blocks. After pressing a key, participants were shown the correct ice-cream for 2 seconds, accompanied by the word ‘good!’ if they had chosen that ice-cream. The procedure for the test phase was identical to that of Experiment 1. Experiment 2 took about 25 minutes.
Results
Induction phase. The number of sentences for which the correct ice-cream was chosen was marginally higher in the structure-useless than in the structure-useful condition (t(28) = 1.899, p = .068). Separate analyses per block showed that
participants in the structure-useless condition (M = 10.46, SD = 3.56) were more often correct than participants in the structure-useful condition (M = 7.27, SD = 1.75) in block 1 (t(20.4) = 3.122, p = .005). In block 2, however, performance in the structure- useful condition (M = 11.67, SD = 3.74) had reached the same level as in the
structure-useless condition (M = 12.13, SD = 2.59, t(28) < 1). There was no difference in reaction times between the two conditions (t(28) = 1.623, p = .116).
Test phase. A one-way ANOVA was performed on the proportion of correct grammaticality judgments in three conditions: the structure-useful and structure- useless conditions of the present experiment and the control condition from Experiment 1. The effect of condition was significant (F(2,43) = 9.043, p = .001).
Post hoc tests with Bonferroni correction showed that the proportion correct in the structure-useful condition (M = .607, SD = .059) was higher than in the structure useless condition (M = .533, SD = .070, p = .010) and the control condition (M = .511, SD = .068, p = .001). The structure-useless condition did not differ from the control condition (p < 1).
Discussion
In Experiment 2, all participants had to perform the same task on the same set of non-word sentences: they had to indicate which of three ice-creams each sentence referred to. In both conditions, participants had to analyze the stimuli for meaning, requiring a relatively deep level of processing (Craik & Lockhart, 1972). In addition, the data from the induction phase suggest that the two conditions were equally attention demanding; there was no difference in the amount of time spent on analyzing the stimuli. The accuracy data from the first block show that, if anything, the task of choosing the right ice-cream was easier in the structure-useless than the structure-useful condition.
Performance on the grammaticality judgment task, however, was significantly better in the structure-useful than in the structure-useless condition. Participants acquired knowledge of the artificial grammar when learning the structure of the non- word sentences was useful in determining which ball should be extra large and which ball should have sprinkles. In contrast, when each non-word referred to a differently colored ball and learning the meaning of the individual non-words was sufficient to
identify the right ice-cream, participants did not acquire any knowledge of the grammar. This suggests that people learn to attend to structure that is useful to their current task, which, in turn, leads to encoding of the structure.
Experiment 3
Experiment 3 investigated whether the knowledge acquired in the structure- useful tasks of Experiments 1 and 2 was implicit in the sense that participants could not control its application. Previous studies applying a process dissociation procedure to AGL exposed participants to two grammars and investigated whether they could control which one to apply at a subsequent grammaticality judgment test (Dienes et al., 1995; Higham, Vokey & Pritchard, 2000). In contrast, the present experiment investigated whether participants exposed to only one grammar could refrain from applying any knowledge of this grammar. Therefore, unlike the previous studies, this experiment was based on a word-stem completion design by Jacoby, Toth and Yonelinas (1993) rather than on Jacoby’s (1991) recognition design.
Participants were presented with non-word sentences from one artificial grammar in the induction phase and were subsequently instructed to complete new sentences, from which one non-word was missing, in either a grammatical or an ungrammatical way. Controlled and automatic knowledge application were compared for one explicit and two incidental learning conditions. In the incidental conditions, participants were instructed either to memorize or to decode the sentences.
Participants in the explicit condition were instructed to look for rules underlying the sentences. This condition was included as a manipulation check. It was expected that participants who learned intentionally would apply their knowledge in a controlled way.
Methods
Participants. Fifty-four undergraduate students of Leiden University (10 male, 44 female; 17-37 years of age) participated in this experiment. None of them had participated in the previous experiments. They received either course credits or € 4.50 for their participation.
Design. The experiment consisted of an induction phase and a test phase. The task in the induction phase was a between-subjects variable: 18 participants had to memorize non-word sentences, 18 had to decode the sentences by identifying the ice- creams they referred to (structure useful version) and 18 had to look for rules. The test
task was a within-subjects variable. All participants had to complete sentences by filling in a non-word. On one half of the trials, indicated by a green background, they had to choose a non-word that made the sentence grammatical. As in the inclusion condition of Jacoby et al. (1993), these desired grammatical completions could be produced by controlled as well as automatic knowledge application. On the other trials, indicated by a red background, participants had to choose a non-word that made the sentence ungrammatical. As in the exclusion condition of Jacoby et al. (1993), undesired grammatical completions could only result from knowledge application that was not controlled.
Materials. The materials for the induction phase were the same as in
Experiment 2 for the sentence decoding task and the same as in Experiment 1 for the tasks to memorize and to look for rules. The 40 test sentences from the previous experiments were all used in their grammatical form in this experiment, but one non- word was deleted from each sentence. No other non-word could occur at the position of this deletion. All sentences were displayed as black text (Courier New 18, bold) in a white rectangle. One half of the sentences appeared against a green background; the other half appeared against a red background. The frequency and the position of the deleted non-word were balanced over the two background colors.
Procedure. In the induction phase, the procedures for the memorize- and decoding tasks were the same as in the previous experiments. Participants in the look for rules condition were informed that a complex set of rules determined which ice- creams were for sale on the planet and which were not. Their task was to read each of the customers’ sentences aloud and to find out which words could follow each other.
In all other respects, the procedure was the same as in the memorize condition.
Before the test phase, participants in the memorize and decoding conditions were told that only ice-creams that followed a complex set of rules were for sale on the planet and that all ice-creams they had seen so far had followed the rules. All participants were informed that they would now be presented with new sentences from which one word had been deleted. They were instructed to press ‘B’, ‘D’, ‘K’,
‘N’ or ‘T’ to complete each sentence with BOG, DIR, KES, NUP or TAF, respectively.
It was announced that one half of the sentences would appear against a green background and the other half against a red background. If the background was green, participants had to complete the sentence in accordance with the rules. If the
background was red, they had to complete the sentence so that it would violate the rules. Sentences with green and red backgrounds were presented in random order to prevent participants from developing strategies for responding on the exclusion trials.
The sentences appeared in the middle of the screen and remained visible until the participant pressed the initial letter of one of the non-words. Then, the next sentence appeared. Although participants were informed that reaction times were not the main measure in this study, they were encouraged to respond as quickly as possible. The experiment took about 25 minutes.
Analyses. The proportion grammatical completions on the two kinds of test trials was used to determine the contributions of controlled and automatic knowledge application resulting from each task in the induction phase. The proportion of
controlled grammatical completions was computed by subtracting the proportion of undesired grammatical completions (red) from the proportion of desired grammatical completions (green). The proportion of automatic grammatical completions was computed by dividing the proportion of undesired grammatical completions (red) by 1 minus the proportion of controlled grammatical completions1.
A 3 x 2 mixed-model ANOVA, with task in the induction phase (memorize vs.
decode vs. look for rules) as between-subjects variable and knowledge application mode (controlled vs. automatic) as within-subjects factor was performed on the proportions of grammatical completions. Paired samples t-tests compared controlled and automatic knowledge application at test for each task in the induction phase. One sample t-tests were performed to test whether the proportions of controlled and automatic grammatical completions were significantly above chance. With both green and red backgrounds, participants had a chance of one out of five to provide a
grammatical completion by guessing. The proportion of automatic grammatical completions was therefore compared to a chance level of .20. The proportion of controlled grammatical completions was compared to a chance level of zero, as the contribution of guessing was already removed by subtracting the proportion of
1 This computation of automatic knowledge application is based on the assumption that the two modes are independent: knowledge application can be automatic, controlled, or both. With a red background, grammatical completions exclusively arise from automatically applied knowledge that is not controlled.
Therefore, the proportion of undesired grammatical completions underestimates automatic knowledge application and has to be divided by 1 minus the contribution of controlled knowledge (Jacoby, Toth &
Yonelinas, 1993). Several alternatives to this assumption have been proposed in the memory literature (see Jacoby, Yonelinas & Jennings, 1997, for review and reply). Our results are not restricted to the independence assumption. Taking the (uncorrected) proportion of undesired grammatical completions as a measure of automatically applied knowledge produces the same pattern of results.
undesired grammatical completions (red) from the proportion of desired grammatical completions (green).
Results
Induction phase. A one-way ANOVA showed that the duration of the
induction phase was different for the three tasks (F(2,51) = 6.620, p = .003). Looking for rules took less time than decoding which ice-cream each sentence referred to (p <
.001). On average, participants chose the correct ice-cream for 17.72 sentences (SD = 6.71).
Test phase. The ANOVA showed a main effect of knowledge application mode (F(1,51) = 5.517, p = .023), which was modified by an interaction with induction phase task (F(2,51) = 3.676, p = .032) as can be seen in Figure 2.
Participants who had looked for rules provided more controlled (M = .328, SD = .181) than automatic (M = .178, SD = .097) grammatical completions (t(17) = 2.907, p = .010) and only the proportion of controlled grammatical completions was above chance (t(17) = 7.69, p < .001). Participants who had memorized the sentences showed the same pattern: more controlled (M = .250, SD = .107) than automatic (M = .165, SD = .087) grammatical completions (t(17) = 2.450, p = .025), with only the proportion of controlled grammatical completions exceeding chance (t(17) = 9.903, p
< .001). In the sentence decoding condition, the difference between controlled and automatic grammatical completions was not significant. Overall, knowledge application was as often automatic as it was controlled. A significant negative correlation (r = -.565, p = .015) between the proportions of controlled and automatic grammatical completions suggested that this result was due to the presence of two subgroups (cf. Creele, Newport & Aslin, 2004). Presumably, one group of participants had control over their knowledge, while the other applied it automatically.
To examine this possibility, the sentence decoding condition was split at the median proportion of controlled grammatical completions (median = .20). Participants who scored above the median, demonstrating relatively strong control over their knowledge, showed the same pattern as the other conditions. They provided more controlled (M = .361, SD = .086) than automatic (M = .202, SD = .111) grammatical completions (t(8) = -2.802, p = .023) and only the proportion of controlled
grammatical completions was above chance (t(8) = 12.627, p < .001). In contrast, participants who scored below the median, demonstrating relatively little control over their knowledge, provided more automatic (M = .269, SD = .055) than controlled (M =
.039, SD = .070) grammatical completions (t(8) = 5.68, p < .001) and only the
proportion of automatic grammatical completions was above chance (t(8) = 3.717, p = .006).
Figure 2. Mean proportions of controlled and automatic grammatical completions with 95%
confidence intervals by instruction in the induction phase.
Discussion
The present experiment used an adaptation of the process dissociation
procedure (Jacoby, Toth & Yonelinas, 1993) to investigate whether artificial grammar learning under useful, incidental instructions produces knowledge that is implicit in the sense that its application cannot be controlled. We compared the incidental
structure learning instructions to memorize and to decode the sentences of an artificial grammar with the explicit instruction to look for rules underlying the sentences. As expected, participants who looked for rules acquired knowledge that could be applied in a controlled way. These participants completed sentences in accordance with the grammar when they were asked to do so, but not when they were asked to provide ungrammatical completions. This finding can be taken as a first indication that our adaptation produces valid results.
In addition, the results for the instruction to memorize are in line with a previous finding that participants who had memorized exemplars of two artificial grammars could control which one they applied in a subsequent grammaticality judgment test (Dienes et al., 1995). In the present study, when asked to complete sentences in an ungrammatical way, participants who had memorized sentences generated by one artificial grammar did not automatically provide grammatical completions. These studies provide converging evidence that memorizing sentences of an artificial grammar does not produce knowledge that is implicit in the sense that it cannot be controlled.
However, such findings do not indicate that artificial grammar learning necessarily leads to explicit knowledge. Dienes et al. (1995) found that, although participants could apply their knowledge in a controlled way, it was implicit in the sense that they were unaware of possessing it. Furthermore, the present study
demonstrated that incidental artificial grammar learning can result in knowledge that is implicit in the sense that it cannot be controlled. When participants who had decoded sentences to identify the right ice-cream were asked to complete non-word sentences in an ungrammatical way, they could not refrain from providing
grammatical completions. For a subgroup of these participants, knowledge application was even predominantly automatic.
A possible explanation for the discrepant results in the two incidental conditions may be that the instruction to memorize non-word sentences invokes a wider variety of strategies than the instruction to decode them. In addition to repeatedly processing groups of non-words, producing implicit knowledge, participants in the memorize condition may engage in activities producing explicit knowledge. For example, they may generate sentences (Schacter, 1987) to test their recollection of previous trials or make explicit comparisons to individuate successive sentences (Verdolini-Marston & Balota, 1994). The task to identify which ice-cream each sentence refers to is less likely to involve such activities. Participants in this condition are absorbed in the current trial and only process groups of non-words to decode the meaning of the sentence they are currently presented with.
General discussion
The present study tested the implications of three views on the nature of the implicit learning process for the occurrence of implicit structure learning. According
to the first view, implicit learning is an automatic process that occurs whenever structured stimuli are processed (Reber, 1989; Servan-Schreiber & Anderson, 1990).
The second view states that structure learning is only demonstrated when structural aspects of the stimuli are attended in both the induction phase and the test phase. As attention is considered to be guided by accidental situational characteristics
(Whittlesea & Dorken, 1993; Whittlesea & Wright, 1997; Wright & Whittlesea, 1998), this view implies that the occurrence of implicit structure learning is
unpredictable. The third view, proposed here, acknowledges that attention determines what is learned, but adds that people implicitly learn to attend to the information that is most useful to the task they are performing, as suggested by the findings of Endo and Takeda (2004). According to this view, implicit learning occurs more reliably when a structure is useful to one’s current task than when it is useless.
The results of Experiments 1 and 2 were in line with this usefulness hypothesis. Participants failed to acquire knowledge of an artificial grammar when they performed computing-, affective rating- and decoding tasks to which the
common structure of the non-word sentences was useless. Participants who performed memorize and decoding tasks to which the common structure was useful, in contrast, were successful at judging the grammaticality of new non-word sentences.
The results are clearly at odds with the view that implicit learning is an automatic consequence of processing structured stimuli. Participants in all conditions were required to pay sufficient attention to the non-word sentences to read them aloud. Yet, in three out of five conditions they were unable to discriminate between new grammatical and ungrammatical sentences on a subsequent test. Although the structure-useless computing task may have been attention demanding, a lack of attentional resources is unlikely to account for the results in the structure-useless decoding condition, which was highly similar to the structure-useful version. In particular, the prediction of the competitive chunking model (Servan-Schreiber &
Anderson, 1990), that implicit artificial grammar learning is an automatic
consequence of reading grammatical sentences, is not supported by the present results.
Perruchet, Vinter, Pacteau and Gallego (2002) also proposed that artificial grammar learning may be a product of elementary processes involved in reading. In their account, however, attention plays a selective role in implicit learning. By focusing attention, people initially segment grammatical letter strings into random chunks. The representations of those chunks will weaken because of decay and
interference, but, as long as they are active, repetitions of the chunks will draw attention. This, in turn, causes the representations to be strengthened. As the grammar determines which chunks are often repeated, learning will necessarily reflect this structure (Perruchet et al., 2002). An explanation of the present results in line with this view may be that the chunks (because of their uselessness) received insufficient attention for the representations to be strengthened.
The results of the present study extend the findings of Whittlesea and
colleagues that structure learning does not occur unless selective attention is directed at structural aspects of the stimuli (Whittlesea & Dorken, 1993; Whittlesea & Wright, 1997; Wright & Whittlesea, 1998). As proposed by the episodic processing account, attention may have been guided to this aspect by accidental characteristics of the experimental situation. However, the pattern of results suggests that implicit learning did not occur accidentally in the present study. The structure was learned in those conditions in which it was useful to the participants’ task and not in other conditions.
Although the evidence is still limited, this suggests that useful structures reliably draw attention (cf. Endo & Takeda, 2004).
The view that people implicitly learn to attend to information that is useful to their current task, while increased attention, in turn, enhances encoding of this information has the advantage that it specifies a type of situation in which implicit structure learning reliably occurs. Usefulness of the structure to a given task can be established independently from the occurrence of structure learning. For example, a structure is useful to a task if performance is higher with structured than with unstructured materials and if informing participants of the structure enhances performance on the task.
According to these criteria, structured materials have been demonstrated to be useful to the two tasks most commonly used in implicit learning research. As noted before, participants needed fewer learning trials to memorize letter strings generated by an artificial grammar than to memorize randomly composed letter strings (Reber, 1967). Similarly, informing participants of the sequence of locations in which a stimulus would appear and allowing them to memorize this sequence produced faster responses on the SRT-task (Destrebecqz, 2004).
In addition, implicit learning may be invoked by useful structures in natural situations. Ellis (2005) proposed that constructions in natural language are implicitly strengthened while they are used to comprehend and produce sentences. He pointed
out that second language learners tend to rely more heavily on temporal adverbs (e.g.
yesterday) than on tense markings on the verb to comprehend sentences and therefore often fail to learn conjugations implicitly. As with the common structure of non-word sentences describing ice-creams in the present study, selectively attending to and learning structural aspects of a natural language may depend on the structures’
usefulness in communication.
Although situations in which a structure is useful seem to reliably produce implicit learning, the third experiment of the present study indicated some variability in the process. An adaptation of the process dissociation procedure (Jacoby, 1991) showed that participants who had memorized non-word sentences could control the application of their knowledge of the grammar, while participants who had decoded sentences also applied some knowledge automatically. The latter finding is in line with a demonstration that knowledge acquired in the SRT-task can be implicit in the sense that it cannot be controlled (Destrebecqz & Cleeremans, 2001).
Interestingly, the study by Destrebecqz and Cleeremans (2001) demonstrated automatic knowledge application in a condition in which the response to one stimulus was immediately succeeded by presentation of the next, but not in a condition with a short delay. Like participants in the sentence decoding condition of the present study, who seemed to be absorbed in the current trial, participants in the SRT-task without a delay presumably refrained from comparing current transitions with previous ones.
Future studies may clarify the effect of inter-trial comparisons on the type of knowledge acquired. In general, more research using the process dissociation procedure (Jacoby, 1991) is needed to investigate how controllable and automatic knowledge comes about.
In conclusion, the present study identified a type of situation in which implicit learning reliably occurs: the process takes place when a structure is useful to one’s current task, but it may not when the structure is useless. Simply processing structured stimuli does not necessarily lead to structure learning. The results are in line with an interactive view on the relationship between attention and implicit learning: people implicitly learn to attend to information that is useful to their task (Endo & Takeda, 2004) and attending to useful information causes it to be learned. In addition,
incidental learning of useful structures can, under some circumstances, be implicit in the sense that it produces knowledge that is applied automatically.
