Information theory and artificial grammar learning: inferring grammaticality from redundancy

In artificial grammar learning experiments, participants study strings of letters constructed using a grammar and then sort novel grammatical test exemplars from novel ungrammatical ones. The ability to distinguish grammatical from ungrammatical strings is often taken as evidence that the participants have induced the rules of the grammar. We show that judgements of grammaticality are predicted by the local redundancy of the test strings, not by grammaticality itself. The prediction holds in a transfer test in which test strings involve different letters than the training strings. Local redundancy is usually confounded with grammaticality in stimuli widely used in the literature. The confounding explains why the ability to distinguish grammatical from ungrammatical strings has popularized the idea that participants have induced the rules of the grammar, when they have not. We discuss the judgement of grammaticality task in terms of attribute substitution and pattern goodness. When asked to judge grammaticality (an inaccessible attribute), participants answer an easier question about pattern goodness (an accessible attribute).     

Prior familiarity with components enhances unconscious learning of relations

The influence of prior familiarity with components on the implicit learning of relations was examined using artificial grammar learning. Prior to training on grammar strings, participants were familiarised with either the novel symbols used to construct the strings or with irrelevant geometric shapes. Participants familiarised with the relevant symbols showed greater accuracy when judging the correctness of new grammar strings. Familiarity with elemental components did not increase conscious awareness of the basis for discriminations (structural knowledge) but increased accuracy even in its absence. The subjective familiarity of test strings predicted grammaticality judgments. However, prior exposure to relevant symbols did not increase overall test string familiarity or reliance on familiarity when making grammaticality judgments. Familiarity with the symbols increased the learning of relations between them (bigrams and trigrams) thus resulting in greater familiarity for grammatical versus ungrammatical strings. The results have important implications for models of implicit learning.     

Fluency does not express implicit knowledge of artificial grammars

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.     

Sources of confidence judgments in implicit cognition

Subjective reports of confidence are frequently used as a measure of awareness in a variety of fields, including artificial grammar learning. However, little is known about what information is used to make confidence judgments and whether there are any possible sources of information used to discriminate between items that are unrelated to confidence. The data reported here replicate an earlier experiment by Vokey and Brooks (1992) and show that grammaticality decisions are based on both the grammatical status of items and their similarity to study exemplars. The key finding is that confidence ratings made on a continuous scale (50%—100%) are closely related to grammaticality but are unrelated to all of the measures of similarity that were tested. By contrast, confidence ratings made on a binary scale (high vs. low) are related to both grammaticality and similarity. The data confirm an earlier finding (Tunney & Shanks, 2003) that binary confidence ratings are more sensitive to low levels of awareness than continuous ratings are and suggest that participants are conscious of all the information acquired in artificial grammar learning.     

Abstraction Processes in Artificial Grammar Learning

Four experiments explored the extent to which abstract knowledge may underlie subjects' performance when asked to judge the grammaticality of letter strings generated from an artificial grammar. In Experiments 1 and 2 subjects studied grammatical strings instantiated with one set of letters and were then tested on grammatical and ungrammatical strings formed either from the same or a changedletter-set.Evenwith a change ofletter-set, subjects were found to be sensitive to a variety of violations of the grammar. In Experiments 3 and 4, the critical manipulation involved the way in which the training strings were studied: an incidental learning procedure was used for some subjects, and others engaged in an explicit code-breaking task to try to learn the rules of the grammar. When strings were generated from a biconditional (Experiment 4) but not from a standard finite-state grammar (Experiment 3), grammaticality judgements for test strings were independent of their surface similarity to specific studied strings. Overall, the results suggest that transfer in this simple memory task is mediated at least to some extent by abstract knowledge.     

The subliminal mere exposure effect does not generalize to structurally related stimuli.

R.F. Bornstein (1994) questioned whether subliminal mere exposure effects might generalize to structurally related stimuli, thereby providing evidence for the existence of implicit learning. Two experiments examined this claim using letter string stimuli constructed according to the rules of an artificial grammar. Experiment 1 demonstrated that brief, masked exposure to grammatical strings impaired recognition but failed to produce a mere exposure effect on novel structurally related strings seen at test. Experiment 2 replicated this result but also demonstrated that a reliable mere exposure effect could be obtained, provided the same grammatical strings were presented at test. The results suggest that the structural relationship between training and test items prevents the mere exposure effect when participants are unaware of the exposure status of stimuli, and therefore provide no evidence for the existence of implicit learning.     

Frequency discrimination: assessing global-level and element-level units in memory

Subjects' knowledge of how often various events occur was used to assess the retention of memory units for word-like strings of letters. A series of strings was presented at one of three exposure durations. Within the series, the frequencies of occurrence of different strings and of the letters composing the strings were varied orthogonally. At relatively long exposure durations, subjects could discriminate the frequency of occurrence for both strings and their constituent letters. The formation of global-level (string) memory units was indicated by judgments of string frequency being unaffected by either the frequencies of their component letters or experimental conditions (brief exposures) that prohibited accurate judgment of letter frequency. Although judgments of letter frequency were sometimes biased by the frequency of the strings containing the letters, the success with which the judgments discriminated different levels of letter frequency did not depend on the activation of string-level memory units. Furthermore, subjects' frequency judgments for letters were not predictable from their recall of the strings containing the letters. These results, which could not be explained by Tversky and Kahneman's (1973) "availability heuristic," provided evidence for the formation of element-level (letter) memory units. A converging experiment established that element-level frequency information could be abstracted from words as well as nonwords, and further, that this information was stored in long-term memory.     

Visual feature learning in artificial grammar classification

The Artificial Grammar Learning task has been used extensively to assess individuals' implicit learning capabilities. Previous work suggests that participants implicitly acquire rule-based knowledge as well as exemplar-specific knowledge in this task. This study investigated whether exemplar-specific knowledge acquired in this task is based on the visual features of the exemplars. When a change in the font and case occurred between study and test, there was no effect on sensitivity to grammatical rules in classification judgments. However, such a change did virtually eliminate sensitivity to training frequencies of letter bigrams and trigrams (chunk strength) in classification judgments. Performance of a secondary task during study eliminated this font sensitivity and generally reduced the contribution of chunk strength knowledge. The results are consistent with the idea that perceptual fluency makes a contribution to artificial grammar judgments.     

Beyond dissociation logic: evidence for controlled and automatic influences in artificial grammar learning

Evidence for unconscious learning has typically been based on dissociations between direct and indirect tests of learning. Because of some inherent problems with dissociation logic, we applied the logic of opposition to 2 artificial grammar learning experiments. In Experiment 1, participants were exposed to 2 different sets of letter strings, generated from 2 different grammars, and later rated test strings for grammaticality with either in-concert (rate grammatical strings consistent with either structure) or opposition (rate grammatical only strings from 1 of the structures) instructions. Manipulating response deadline affected controlled, but not automatic influences. In Experiment 2, after similar training, a source-monitoring test was administered from which the in-concert and opposition conditions were derived. The test indicated that varying the retention interval affected controlled, but not automatic, influences. The results are discussed in terms of awareness, knowledge representation, and metacognitive processing.     

AGSuite: Software to conduct feature analysis of artificial grammar learning performance

To simplify the problem of studying how people learn natural language, researchers use the artificial grammar learning (AGL) task. In this task, participants study letter strings constructed according to the rules of an artificial grammar and subsequently attempt to discriminate grammatical from ungrammatical test strings. Although the data from these experiments are usually analyzed by comparing the mean discrimination performance between experimental conditions, this practice discards information about the individual items and participants that could otherwise help uncover the particular features of strings associated with grammaticality judgments. However, feature analysis is tedious to compute, often complicated, and ill-defined in the literature. Moreover, the data violate the assumption of independence underlying standard linear regression models, leading to Type I error inflation. To solve these problems, we present AGSuite, a free Shiny application for researchers studying AGL. The suite’s intuitive Web-based user interface allows researchers to generate strings from a database of published grammars, compute feature measures (e.g., Levenshtein distance) for each letter string, and conduct a feature analysis on the strings using linear mixed effects (LME) analyses. The LME analysis solves the inflation of Type I errors that afflicts more common methods of repeated measures regression analysis. Finally, the software can generate a number of graphical representations of the data to support an accurate interpretation of results. We hope the ease and availability of these tools will encourage researchers to take full advantage of item-level variance in their datasets in the study of AGL. We moreover discuss the broader applicability of the tools for researchers looking to conduct feature analysis in any field.     

Locus of control and the learning of an artificial language

An experiment was conducted to explore the cognitive processes associated with locus of control beliefs. Internal and external subjects were instructed to remember as much as they can from an array of 21 letter strings generated from a finite state grammar. Although internals and externals did not significantly differ in their ability to recall these items, internals were better able to discriminate grammaticality and nongrammaticality in a new set of letter strings. Thus, although both groups attended to the exemplars, internals extracted more invariance and hence learned more about the underlying structure than externals. These results were discussed in relation to those of Wolk and DuCette (1974) on I-E and incidental learning.     

人工语法范式下的无意识知识探析

人工语法范式不仅是内隐学习领域广泛运用的范式,同时也是研究无意识知识的主要方法。文章通过对人工语法范式组成材料中的刺激频率、组块等方面进行深入探讨,不仅使人们对无意识知识有了更全面的认识,而且也有利于对内隐学习获得规则知识的假设提出更确定的回答;并通过对无意识知识的神经机制研究以及考察镜射规则学习过程的脑神经激活状况,获得了规则学习对应更为准确的脑神经活动,推动了人工语法范式下无意识知识的深入研究。     

Memory and display search in binary classification reaction time

Attention, Perception, & Psychophysics - Character classification tasks and “same” “different” judgments of letter strings were examined with reference to major...     

Realism in confidence judgements of performance based on implicit learning

In two experiments we employed calibration methods to investigate the realism of participants' confidence ratings of their own classification performance based on knowledge acquired after training on an artificial grammar. In Experiment 1 participants showed good realism (but overconfidence) for grammatical strings but very poor realism for non-grammatical strings. Method of training (string repetition in writing or mere exposure) did not affect the realism. Furthermore, the participants underestimated their overall performance. In Experiment 2, using a more complex grammar and controlling for two types of associative chunk-strength, participants showed good realism (but still overconfidence) for both letter and symbol strings, irrespective of grammaticality. Together, these experiments show that implicit learning can give rise to knowledge products that are associated with fairly realistic meta-knowledge. It is argued that both the zero-correlation criterion and the guessing criterion are misplaced when used to define implicit knowledge; two reasons being that confidence judgements may be affected both by implicit knowledge and by inferences.     

Artificial grammar learning by 1-year-olds leads to specific and abstract knowledge.

Four experiments used the head-turn preference procedure to assess whether infants could extract and remember information from auditory strings produced by a miniature artificial grammar. In all four experiments, infants generalized to new structure by discriminating new grammatical strings from ungrammatical ones after less than 2 min exposure to the grammar. Infants acquired specific information about the grammar as demonstrated by the ability to discriminate new grammatical strings from those with illegal endpoints (Experiment 1). Infants also discriminated new grammatical strings from those with string-internal pairwise violations (Experiments 2 and 3). Infants in Experiment 4 abstracted beyond specific word order as demonstrated by the ability to discriminate new strings produced by their training grammar from strings produced by another grammar despite a change in vocabulary between training and test. We discuss the implications of these findings for the study of language acquisition.     

Kompositionsbildung bei Symbolfolgen und Bediensequenzen: Empirische Befunde und die Theorie des „Competitive Chunking”

In the paradigm of artificial grammar learning subjects have to reproduce regularly structured sequences of symbols. In a subsequent and unexpected test, they have to discriminate sequences that correspond to those of thememorization task from sequences that do not. Discriminability of grammatical and nongrammatical strings is judged as evidence for unconscious learning. This effect has been explained by the creation of chunks: Memorization of strings is facilitated by composing elements to chunks. The hierarchical organization of these units allows the representation of grammatical strings by a few high level chunks. The fewer chunks are necessary to represent a given string the higher is its familiarity and the probability to be judged grammatical. For the incidental acquisition of knowledge about operating a device the creation of chunks could also be shown. In contrast to letter series, two peculiarities were found: (a) discriminability was rather low, (b) irregularities at the end of input sequences were most effective. These peculiarities dissappeared, if — similar to letter strings — the information about a single operation of a sequence was permanently available. In the present simulations it was investigated whether these effects are consistent with the theory of competitive chunking. The temporal grain within sequences was varied to investigate the impact of the availability of sequential information on discriminability. Three different modes of chunk creation were simulated in the test phase to investigate whether the creation of chunks is obligatory for perceptual integration or requires a memorization task. The variation of temporal grain had only a marginal impact on discriminability. The simulation of the empirical data required that grammatical chunks are created within the test phase. Implications of these findings are discussed.     

Judgments of grammaticality of sentences with a differing number of arguments: a comparison of English and Japanese speakers

This study explored native speakers' linguistic intuition as revealed in judging the grammaticality of sentences. Speakers of English (n = 36) and Japanese (n = 57) judged the relative grammaticality of sentences involving a verb which occurred together with one, two or three arguments. Findings showed that English speakers were more affected by the number of arguments in sentences. They judged sentences having three arguments as grammatical most often and ones having one or two arguments as grammatical less often. However, Japanese speakers gave rather similar and more grammatical judgments regardless of the number of arguments in sentences. The findings indicate a difference in tightness of argument structure in the two languages even when the sentences judged are simple sentences and they are given without any sentential context.     

Measuring strategic control in artificial grammar learning

In response to concerns with existing procedures for measuring strategic control over implicit knowledge in artificial grammar learning (AGL), we introduce a more stringent measurement procedure. After two separate training blocks which each consisted of letter strings derived from a different grammar, participants either judged the grammaticality of novel letter strings with respect to only one of these two grammars (pure-block condition), or had the target grammar varying randomly from trial to trial (novel mixed-block condition) which required a higher degree of conscious flexible control. Random variation in the colour and font of letters was introduced to disguise the nature of the rule and reduce explicit learning. Strategic control was observed both in the pure-block and mixed-block conditions, and even among participants who did not realise the rule was based on letter identity. This indicated detailed strategic control in the absence of explicit learning.     

Two Routes to Grammatical Gender: Evidence from Hebrew

Access to grammatical gender in Hebrew was examined using gender decisions and grammaticality judgments. In gender decisions, and in rejecting agreement violations, the role of gender-marking regularity was robust; irregular feminine nouns were classified more slowly and produced more errors than did regular masculine nouns. These same exception nouns, however, produced radically different results when they appeared as part of a grammatical phrase. In this context, irregular nouns were processed as quickly and accurately as were regular nouns. These findings are interpreted within a model containing two routes to grammatical gender: one that involves an abstract gender node, and another that is form-based and is assumed to play a greater role in recovery from agreement errors. We also argue that there are cross-linguistic variations in the availability and speed of the form-based route to gender. Finally, we caution that under some circumstances grammatical gender may also be retrieved through a variety of heuristics that do not necessarily reflect the normal mechanisms of access to gender.     

Adult age differences in the perception and learning of artistic style categories

Adult age differences in conceptual behavior were studied using informationally complex stimuli from real-world categories: paintings by two impressionist artists. In Experiment 1 we examined perceptions of category structure by having subjects sort paintings according to style similarity. Young adults were observed to depend more on abstract information in making style judgments, whereas older adults relied more on similarity in content. This resulted in different category structures between age groups, but similarity judgments in both groups appeared to correspond to actual style differences between the two artists. In Experiment 2, learning efficiency was shown to increase with a painting's category centrality, but older adults had particular trouble learning noncentral items. At transfer, both age groups were able to use abstracted central tendency information to categorize new paintings, although young adults appeared to have better access to information about specific category exemplars from acquisition. The results are generally consistent with those from studies using simpler artificial stimuli.