Elsevier

Cognition

Volume 114, Issue 3, March 2010, Pages 372-388
Cognition

Fluency does not express implicit knowledge of artificial grammars

https://doi.org/10.1016/j.cognition.2009.10.010Get rights and content

Abstract

It is commonly held that implicit knowledge expresses itself as fluency. A perceptual clarification task was used to examine the relationship between perceptual processing fluency, subjective familiarity, and grammaticality judgments in a task frequently used to produce implicit knowledge, artificial grammar learning (AGL). Four experiments examined the effects of naturally occurring differences and manipulated differences in perceptual fluency, where decisions were based on a brief exposure to test-strings (during the clarification task only) or normal exposure. When perceptual fluency was not manipulated, it was weakly related to familiarity and grammaticality judgments, but unrelated to grammatical status and hence not a source of accuracy. Counterbalanced grammatical and ungrammatical strings did not differ in perceptual fluency but differed substantially in subjective familiarity. When fluency was manipulated, faster clarifying strings were rated as more familiar and were more often endorsed as grammatical but only where exposure was brief. Results indicate that subjective familiarity derived from a source other than perceptual fluency, is the primary basis for accuracy in AGL. Perceptual fluency is found to be a dumb heuristic influencing responding only in the absence of actual implicit knowledge.

Introduction

There is substantial evidence that the knowledge acquired in implicit learning – especially of artificial grammars – is expressed largely as familiarity, defined as the subjective feeling of oldness elicited by a stimulus (e.g. Higham, 1997, Johnstone and Shanks, 2001, Kinder and Assmann, 2000, Norman et al., 2007, Scott and Dienes, 2008, Servan Schreiber and Anderson, 1990). The question now arises as to the basis of that familiarity. In the memory literature familiarity has been proposed to result from perceptual or conceptual processing fluency (Jacoby & Dallas, 1981), or surprising fluency (Whittlesea & Williams, 2000). It has been logical to infer that perceptual fluency accounts for familiarity in artificial grammar learning (AGL) and that it is thus the means by which implicit knowledge affects classification either generally or when employing certain strategies (Kinder et al., 2003, Whittlesea and Price, 2001). However, evidence that perceptual fluency contributes to familiarity is far from conclusive in either recognition memory generally (e.g. Kinoshita, 2002, Levy et al., 2004, Stark and Squire, 2000, Wagner et al., 1997) or in AGL in particular (e.g. Chang and Knowlton, 2004, Lieberman et al., 2004, Newell and Bright, 2001, Zizak and Reber, 2004). The current study examines the role of perceptual fluency in AGL, evaluating its influence on subjective ratings of familiarity and grammaticality judgments, and how this influence differs when people can or cannot freely use veridical implicit knowledge.

AGL has been one of the most commonly employed paradigms for the study of implicit learning (Pothos, 2007, Reber, 1989). In a typical AGL experiment participants are exposed to letter strings conforming to a complex set of rules referred to as a grammar. The strings are commonly presented under the guise of a short-term memory task with participants unaware of their rule-based nature. At test, participants are informed of the existence of rules and asked to judge which of a new set of strings are grammatical. Participants are typically able to discriminate the grammatical strings with above-chance accuracy despite believing they are guessing or using intuition and despite being unable to verbalise the rules of the grammar (e.g. Allwood et al., 2000, Channon et al., 2002, Dienes and Altmann, 1997, Dienes et al., 1995, Dienes and Longuet Higgins, 2004, Dienes and Scott, 2005, Reber, 1967, Tunney and Altmann, 2001). A.S. Reber (1967) originally proposed that the ability to discriminate grammatical strings resulted from the implicit acquisition of regularities encountered during learning. Since that time research has proceeded to examine the nature of the regularities acquired. These are now known to include commonly recurring fragments or chunks of the training-strings (Dulany et al., 1984, Knowlton and Squire, 1994, Perruchet and Pacteau, 1990, Servan Schreiber and Anderson, 1990), the pattern of repetitions within training-strings (Brooks and Vokey, 1991, Vokey and Higham, 2005), and knowledge of whole training exemplars (Vokey & Brooks, 1992). Similarity between training-strings and test-strings arising from any of these features could in principle result in familiarity. Servan Schreiber and Anderson were the first to characterise the knowledge acquired in this way. The resulting familiarity account holds that grammatical strings, by virtue of conforming to the grammar, are more likely to have properties seen in training and will consequently feel more familiar. Discrimination performance then results from more familiar strings being endorsed as grammatical.

There is considerable evidence supporting this account of AGL. Signal detection analyses of implicit learning tasks are consistent with decisions based on a continuous underlying dimension, such as familiarity, but not with certain rule-based accounts e.g. where a limited number of rules lead to black and white decisions (Kinder and Assmann, 2000, Lotz and Kinder, 2006). Successful computational models of AGL, and implicit learning generally, also assume a continuous output from the network that reflects similarity (for a review see Cleeremans & Dienes, 2008). More directly, Johnstone and Shanks (2001) showed that the objective similarity of training and test-strings strongly predicts grammaticality judgements. Finally, direct evidence has been provided by Scott and Dienes (2008) who showed that subjective ratings of the familiarity of test-strings were reliably predicted by structural similarity measures (mean R = .45), and that those familiarity ratings themselves reliably predicted grammaticality judgments (Mean r = .64).

Jacoby and Dallas (1981) proposed that when processing an item with relative ease, or fluently, people may attribute this to the item having been seen before and experience it as familiarity. This notion was developed further by Whittlesea and Williams (2000) who demonstrated that familiarity arises from a discrepancy with expected fluency. In AGL perceptual fluency could result from repetition priming during training; the elements most commonly observed in training would subsequently be processed more fluently at test. Given that grammatical test-strings have more in common with training-strings than do ungrammatical test-strings, the resulting difference could, in principle, be a source of accurate responding. Buchner (1994) found evidence supporting grammaticality as a source of differential perceptual fluency in AGL. Employing a perceptual clarification task to measure naturally occurring differences in the perceptual fluency of test-strings, Buchner found grammatical strings to be identified on average 200 ms faster than ungrammatical strings. This is an important and widely cited result. The implication for fluency as a potential source of implicit knowledge both in AGL and implicit learning generally make replication an imperative. The need to explore the generalisability of the effect is particularly acute in light of potential alternative explanations for the differences observed.

Fluency is known to be affected by a range of factors, most obviously repetition. Repetition priming has been demonstrated to enhance perceptual fluency in a range of experimental contexts (e.g. Jacoby and Dallas, 1981, Tulving et al., 1982). When parsing a string, if a letter is the same as the previous letter then within-string repetition priming will result in that letter being perceived more fluently. For grammar A of Buchner (1994), the only grammar used in Experiment 1, grammatical test-strings contained more repetitions than ungrammatical strings e.g. TXXTVV vs. TVXTVV. Based on this difference alone, grammatical strings would be expected to be perceived more fluently. However, letter repetition is only one feature known to influence fluency, others include the repetition of larger elements (e.g. bigrams) and the presence or absence of symmetry (R. Reber, Schwarz, & Winkielman, 2004). These superficial features are features that a string has intrinsically, i.e. can be determined from the string alone because they are not a relation between the string and training-strings. All such possible superficial test-string features will be controlled only when grammatical and non-grammatical strings are counterbalanced.

In addition to controlling for alternative sources of fluency, where fluency is assessed using a reaction-time task other influences on response times must also be avoided. In Buchner (1994) Experiment 1 the perceptual clarification task was not followed by any other decision. In Experiment 2, however, participants were required to make grammaticality and recognition judgements after completing the clarification task. Crucially, this was done with the test-string no longer available for reference. Under these circumstances participants might be expected to delay their response to the clarification task until arriving at a decision for the subsequent judgment. Consistent with this influence, the average identification time was 1700 ms longer and the difference between identification times for grammatical and ungrammatical strings 117 ms (66%) greater in Experiment 2 than for the same materials in Experiment 1. Where identification times reflect decision processes, theories from the categorization literature make clear predictions regarding how identification times will be affected. The RT-Distance Hypothesis, based on decision bound theory, holds that reaction time decreases with the distance between the perceptual effect and the decision boundary (Ashby, Boynton, & Lee, 1994). Scott and Dienes (2008) found evidence that in AGL the decision boundary lies approximately at the mean familiarity. Strings with rated familiarity greater than the mean were more likely to be categorised as grammatical while those with familiarity less than the mean were more likely to be categorised as ungrammatical. Furthermore, participants’ confidence in their judgments increased with the extremity of familiarity i.e. the further a string’s familiarity was from the mean. This arrangement predicts that the more extreme the familiarity – either high or low – the quicker the string will be identified. This prediction is readily distinguishable from that of the fluency hypothesis which predicts faster identification times only for higher familiarity.

A negative relationship between identification times and the extremity of familiarity ratings would indicate that identification times are being influenced by decision processes. Under such circumstances, if the extremity of familiarity ratings is greater for grammatical than ungrammatical strings then identification times would be shorter for grammatical strings independent of differences in fluency. Analysis of the data from Scott and Dienes (2008) revealed precisely this pattern; the familiarity ratings for grammatical test-strings were further from the mean test-string familiarity (were more extreme) than the ratings for ungrammatical test-strings.1 A reliable replication of the finding that grammaticality is related to fluency therefore requires adequate control both over alternative sources of fluency and over the effect of decision processes on the reaction-time task.

Assuming perceptual fluency is related to grammaticality it would still need to influence responding, either by affecting familiarity or by some other means, in order to be a source of accuracy. A relationship between perceptual fluency and grammaticality judgments was not observed by Buchner (1994) but has been observed where fluency has been artificially manipulated. Kinder et al. (2003) employed a perceptual clarification task where the rate at which strings clarified (appearing pixel by pixel) was varied in order to artificially manipulate perceptual fluency. They found that strings clarifying more quickly had an increased chance of being classified as grammatical, and concluded that people exploit perceptual fluency to make their grammaticality judgments. Such an effect has not been universally observed however. More generally, manipulating perceptual fluency has been found to significantly influence the rated liking of test-strings but not judgments of grammaticality (Newell and Bright, 2001, Zizak and Reber, 2004). However, even a small influence of perceptual fluency on classification judgments could be of crucial importance in understanding the nature of implicit knowledge. Much of what is learnt in AGL is amenable to conscious report, what is of enduring interest however, is the presence of above-chance accuracy in the absence of verbalizable knowledge. Perceptual fluency, if derived from grammaticality, could account for that important subset of responses.

If perceptual fluency were to contribute to a subset of responses it might be apparent depending on the type of strategy adopted. Whittlesea and Price (2001) proposed that the use of perceptual fluency in decision making varies depending on whether processing is analytical or non-analytical. Kinder et al. (2003) developed a related idea in the context of AGL. They found that manipulating fluency influenced grammaticality judgments but left recognition judgments unaffected except when all test-strings were new – preventing the use of recollection. They postulate that participants exploit either a fluency heuristic or a recollection heuristic depending on the task, with the latter insensitive to fluency. When making grammaticality decisions based on a non-analytical approach, such as familiarity, people are thought to adopt processing fluency as the default strategy. In contrast when making more explicit recognition judgments they are thought to rely less on fluency and more on recollection processes. However, other research findings provide a possible alternative account for this pattern of results. Whittlesea and Leboe (2000) showed that when participants can exploit either fluency or structural similarity to make grammaticality judgments that they reliably favour the latter. Willems, van der Linden, and Bastin (2007) similarly found that in preference and recognition judgments the influence of processing fluency depended on the amount of information contained in the stimuli. And most recently, Johansson (2009) found that fluency manipulations influenced grammaticality judgments made under response deadlines but not those made under free response. Together these findings are more consistent with fluency being exploited where other sources of judgment are restricted. In Kinder et al.’s study the test-strings were presented only briefly during a perceptual clarification task, with participants required to make their grammaticality judgments based on that momentary exposure. This contrasts with the standard AGL protocol where test-strings are available for reference while participants judge their grammaticality. The effect of fluency on recognition judgments was also only observed where the usual basis for that judgment was restricted i.e. where none of the test-strings had in fact been seen before. As such, it is feasible that the observed difference in sensitivity to the fluency manipulation resulted not from the use of different heuristics but simply from the presence or absence of alternative bases for decision.

The research findings relating to the role of fluency in AGL raise three crucial questions: (1) Can the relationship between grammaticality and perceptual fluency observed by Buchner (1994) be replicated where the potential effects of other sources of fluency and decision processes on identification times are eliminated? Confirming the relationship between fluency and grammaticality is essential to establishing whether fluency has the potential to contribute to accuracy and be a source of the non-verbalizable knowledge observed in AGL. (2) To what extent does the influence of perceptual fluency in AGL differ with more or less opportunity to process the grammar strings and with the adoption of different decision strategies? Understanding the basis of knowledge in AGL requires that we establish the contribution of fluency where the usual sources of judgment are not restricted. (3) What is the relationship between the subjective experience of familiarity and perceptual fluency in AGL? Scott and Dienes (2008) demonstrated that subjective familiarity can largely account for the accuracy of responding in AGL but there has been no investigation of how subjective familiarity relates to perceptual fluency in this paradigm. The experiments in the current study were devised to address each of these questions.

Examining naturally occurring differences in perceptual fluency is the only means to confirm whether fluency is related to grammaticality but cannot establish whether a relationship between fluency and responding is due to correlation or cause. Examining the effects of manipulated fluency can establish whether a relationship is causal, but because the fluency variations are artificial they may not be representative of those occurring naturally. We therefore exploited both approaches: Experiments 1 and 2 examined naturally occurring differences in perceptual fluency, and Experiments 3 and 4 examined the effect of manipulating perceptual fluency. We explore the effect of restricting exposure to the test-strings under each approach; in Experiments 1 and 3 strings are available for reference while making grammaticality judgments while in Experiments 2 and 4 judgments must be made after only the brief presentation occurring during the clarification task. The restricted exposure in Experiments 2 and 4 also permits us to examine whether decision processes compromised the assessment of fluency under such circumstances; this would reveal itself as shorter identification times for more extreme familiarity ratings.

The effect that superficial test-string features, such as letter repetition, had on fluency estimates was counterbalanced across grammaticality in all experiments. This was achieved by employing the two grammar design of Dienes and Altmann (1997). This involves training half the participants on one grammar, and half on another, and having all participants classify the same set of test-strings exactly half of which conform to each. In this way the ungrammatical test-strings for half the participants are grammatical for the other half and analysis collapsed across grammar ensures counterbalancing. This approach has the additional benefit that while the effects are counterbalanced we can still examine whether they have the capacity to influence fluency estimates.

In addition to grammaticality judgments participants were required to rate the familiarity of each test-string and to report the decision strategy used for each judgment. The reporting of familiarity ratings permitted the relationship between fluency and subjective familiarity to be examined. The reporting of decision strategy permitted the influence of fluency on grammaticality judgments to be contrasted according to the type of strategy employed. Previous research has demonstrated that the nature of participants’ responses is quantifiably different depending on the strategy they report using (Dienes and Scott, 2005, Scott and Dienes, 2008).

In sum, Experiment 1 examined the relationship between fluency, grammaticality, and familiarity where superficial sources of fluency were counterbalanced and the influence of decision processes were avoided by having strings available for reference when grammaticality judgments were made. Experiment 2 had participants make their grammaticality judgments having only seen the test-strings during the clarification task. This was done to test the prediction that under those circumstances decision processes would influence responses times, resulting in shorter identification times for more extreme familiarity and hence for grammatical strings. Experiments 3 examined whether the influence of manipulating perceptual fluency on grammaticality judgments observed by Kinder et al. (2003) would be eliminated if participants were able to reference the test-strings when making their grammaticality judgments. Experiment 4 aimed to replicate Kinder et al.’s original result by limiting test-string exposure to that obtained during the clarification task. Experiments 3 and 4 also provided the opportunity to replicate key findings from Experiments 1 and 2 respectively.

Section snippets

Experiment 1

This experiment sought to replicate Buchner’s (1994) finding, that grammaticality is related to perceptual fluency, while counterbalancing superficial test-string features and avoiding the potential influence of decision processes on identification times. It was predicted that, controlling for these influences, the relationship would be reduced or eliminated. We further evaluated the extent to which naturally occurring differences in perceptual fluency predicted familiarity and grammaticality

Experiment 2

The current experiment sought to establish whether an influence of decision processes on identification times may have contributed to the apparent relationship between fluency and grammaticality observed in Buchner’s (1994) Experiment 2. Replicating that experiment, grammaticality judgments are made based solely on the test-string exposure received during the clarification task. We then examine if the resulting identification times reflect decision processes and whether this results in shorter

Experiment 3

Experiment 3 sought to establish the extent to which artificially enhancing perceptual fluency increases the subjective familiarity of test-strings and their likelihood of being endorsed as grammatical. Kinder et al. (2003) observed a substantial effect on grammaticality judgments; however, reference to the test-strings in that study was restricted to brief exposure during the clarification task. Here we examine whether that effect occurs when the test-strings remain available for reference

Experiment 4

Experiment 4 sought to establish if manipulated perceptual fluency would influence familiarity and grammaticality judgments when exposure to the test-strings was limited. The experiment was the same as Experiment 3 with just one exception, consistent with Kinder et al. (2003) the test-strings were only seen during the perceptual clarification task.

General discussion

This study sought to establish the extent to which perceptual fluency reflects the grammaticality of test-strings in artificial grammar learning, the degree to which it might influence grammaticality judgments, and its relationship with subjective feelings of familiarity. Perceptual fluency was found to be unrelated to grammaticality and as such is not a source of implicit knowledge in this paradigm. However, perceptual fluency derived from sources unrelated to grammaticality was found to

References (63)

  • T.M. Bailey et al.

    AGL StimSelect: Software for automated selection of stimuli for artificial grammar learning

    Behavior Research Methods

    (2008)
  • R.M. Baron et al.

    The moderator-mediator variable distinction in social psychological research: Conceptual, strategic and statistical considerations

    Journal of Personality and Social Psychology

    (1986)
  • C.J. Berry et al.

    A single-system account of the relationship between priming, recognition, and fluency

    Journal of Experimental Psychology: Learning Memory and Cognition

    (2008)
  • L.R. Brooks et al.

    Abstract analogies and abstracted grammars: Comments on Reber (1989) and Mathews et al. (1989)

    Journal of Experimental Psychology: General

    (1991)
  • A. Buchner

    Indirect effects of synthetic grammar learning in an identification task

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (1994)
  • G.Y. Chang et al.

    Visual feature learning in artificial grammar classification

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (2004)
  • A. Cleeremans et al.

    Computational models of implicit learning

  • S.A. Dewhurst et al.

    Effects of exact and category repetition in true and false recognition memory

    Memory and Cognition

    (1999)
  • Z. Dienes

    Assumptions of subjective measures of unconscious mental states: Higher order thoughts and bias

    Journal of Consciousness Studies

    (2004)
  • Z. Dienes

    Subjective measures of unconscious knowledge

    Progress in Brain Research

    (2008)
  • Z. Dienes et al.

    Transfer of implicit knowledge across domains: How implicit and how abstract?

  • Z. Dienes et al.

    Unconscious knowledge of artificial grammars is applied strategically

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (1995)
  • Z. Dienes et al.

    Measuring unconscious knowledge: Distinguishing structural knowledge and judgment knowledge

    Psychological Research

    (2005)
  • D.E. Dulany et al.

    A case of syntactical learning and judgment: How conscious and how abstract?

    Journal of Experimental Psychology: General

    (1984)
  • J.M. Gardiner

    Functional aspects of recollective experience

    Memory and Cognition

    (1988)
  • P.A. Higham

    Dissociations of grammaticality and specific similarity effects in artificial grammar learning

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (1997)
  • Higham, P. A. (unpublished manuscript). Dumb use of the fluency heuristic: Reliance on perceptual fluency in the...
  • L.L. Jacoby et al.

    On the relationship between autobiographical memory and perceptual learning

    Journal of Experimental Psychology: General

    (1981)
  • L.L. Jacoby et al.

    An illusion of memory: False recognition influenced by unconscious perception

    Journal of Experimental Psychology: General

    (1989)
  • T. Johansson

    In the fast lane toward structure in implicit learning: Non-analytic processing and fluency in artificial grammar learning

    European Journal of Cognitive Psychology

    (2009)
  • A. Kinder et al.

    Learning artificial grammars: No evidence for the acquisition of rules

    Memory and Cognition

    (2000)
  • Cited by (0)

    View full text