Artificial Grammar Learning Experiments Research Articles

ENIGMA: A Web Application for Running Online Artificial Grammar Learning Experiments.

Artificial grammar learning (AGL) is an experimental paradigm frequently adopted to investigate the unconscious and conscious learning and application of linguistic knowledge. This paper will introduce ENIGMA ( https://enigma-lang.org ) as a free, flexible, and lightweight Web-based tool for running online AGL experiments. The application is optimized for desktop and mobile devices with a user-friendly interface, which can present visual and aural stimuli and elicit judgment responses with RT measures. Without limits in time and space, ENIGMA could help collect more data from participants with diverse personal and language backgrounds and variable cognitive skills. Such data are essential to explain complex factors influencing learners' performance in AGL experiments and answer various research questions regarding L1/L2 acquisition. The introduction of the core features in ENIGMA is followed by an example study that partially replicated Chen (Lang Acquis 27(3):331-361, 2020) to illustrate possible experimental designs and examine the quality of the collected data.

The "Starting-Small" Effect in Phonology: Evidence From Biased Learning of Opaque and Transparent Vowel Harmony.

The starting-small effect is a cognitive advantage in language acquisition when learners begin by generalizing on regularities from structurally simple and shorter tokens in a skewed input distribution. Our study explored this effect as a potential explanation for the biased learning of opaque and transparent vowel harmony. In opaque vowel harmony, feature agreement occurs strictly between adjacent vowels, and an intervening "neutral vowel" blocks long-distance vowel harmony. Thus, opaque vowel harmony could be acquired even if learners start with structurally simpler and more frequent disyllabic tokens. Alternatively, transparent vowel harmony can only be observed in longer tokens demonstrating long-distance agreement by skipping a neutral vowel. Opaque vowel harmony is predicted to be learned more efficiently due to its compatibility with local dependency acquired via starting-small learning. In two artificial grammar learning experiments, learners were exposed to both vowel harmony patterns embedded in an equal number of disyllabic and trisyllabic tokens or a skewed distribution with twice as many disyllabic tokens. In Exp I, learners' test performance suggests the consistently biased learning of local and opaque vowel harmony with starting-small learning. Furthermore, in Exp II, the acquired vowel harmony patterns varied significantly by working memory capacity with a balanced but not skewed input distribution, presumably because of the ease of cognitive demand with starting-small learning.

Instance theory predicts categorization decisions in the absence of categorical structure: A computational analysis of artificial grammar learning without a grammar.

Theories of categorization have historically focused on the stimulus characteristics to which people are sensitive. Artificial grammar learning (AGL) provides a clear example of this phenomenon, with theorists debating between knowledge of rules, fragments, whole strings, and so on as the basis of categorization decisions (i.e., stimulus-driven explanations). We argue that this focus loses sight of the more important question of how participants make categorization decisions on a mechanistic level (i.e., process-driven explanations). To address the problem, we derived predictions from an instance-based model of human memory in a pseudo-AGL task in which all study and test strings were generated randomly, a task that stimulus-driven explanations of AGL would have difficulty accommodating. We conducted a standard AGL experiment with human participants using the same strings. The model's predictions corresponded to participants' decisions well, even in the absence of any experimenter-generated structure and regardless of whether test stimuli contained any incidental structure. We argue that theories of categorization ought to continue shifting towards the goal of modeling categorization at the level of cognitive processes rather than primarily attempting to identify the stimulus characteristics to which participants are drawn.

Formal language hierarchy reflects different levels of cognitive complexity.

Formal language hierarchy describes levels of increasing syntactic complexity (adjacent dependencies, nonadjacent nested, nonadjacent crossed) of which the transcription into a hierarchy of cognitive complexity remains under debate. The cognitive foundations of formal language hierarchy have been contradicted by two types of evidence: First, adjacent dependencies are not easier to learn compared to nonadjacent; second, crossed nonadjacent dependencies may be easier than nested. However, studies providing these findings may have engaged confounds: Repetition monitoring strategies may have accounted for participants' high performance in nonadjacent dependencies, and linguistic experience may have accounted for the advantage of crossed dependencies. We conducted two artificial grammar learning experiments where we addressed these confounds by manipulating reliance on repetition monitoring and by testing participants inexperienced with crossed dependencies. Results showed relevant differences in learning adjacent versus nonadjacent dependencies and advantages of nested over crossed, suggesting that formal language hierarchy may indeed translate into a hierarchy of cognitive complexity. (PsycInfo Database Record (c) 2022 APA, all rights reserved).

Concurrent visual sequence learning

Many researchers in the field of implicit statistical learning agree that there does not exist one general implicit learning mechanism, but rather, that implicit learning takes place in highly specialized encapsulated modules. However, the exact representational content of these modules is still under debate. While there is ample evidence for a distinction between modalities (e.g., visual, auditory perception), the representational content of the modules might even be distinguished by features within the same modalities (e.g., location, color, and shape within the visual modality). In implicit sequence learning, there is evidence for the latter hypothesis, as a stimulus-color sequence can be learned concurrently with a stimulus-location sequence. Our aim was to test whether this also holds true for non-spatial features within the visual modality. This has been shown in artificial grammar learning, but not yet in implicit sequence learning. Hence, in Experiment 1, we replicated an artificial grammar learning experiment of Conway and Christiansen (2006) in which participants were supposed to learn color and shape grammars concurrently. In Experiment 2, we investigated concurrent learning of sequences with an implicit sequence learning paradigm: the serial reaction time task. Here, we found evidence for concurrent learning of two sequences, a color and shape sequence. Overall, the findings converge to the assumption that implicit learning might be based on features.

Substantive bias and the positional extension of major place contrasts

The role of phonetic naturalness in biasing the learning of sound patterns remains an unsettled question in phonology. The present study investigates naturalness bias in phonotactic learning using a novel experimental paradigm that tests whether learners reproduce a phonetically-motivated phonotactic implicational about the distribution of major place contrasts in stops. Stops differing in place of articulation are easier to distinguish word-initially than word-finally, so place contrasts in word-final position should also exist in word-initial position. The reverse is not necessarily true. This implicational is typologically supported as well as motivated by perceptual naturalness. In two artificial grammar learning experiments, I exposed participants to place contrasts in stops either word-initially or word-finally and tested whether they extended the contrasts to the other word-edge position. Participants successfully learned to recognize novel words that fit the phonotactic pattern they had been trained on, but they were equally willing to extend the place contrasts in both directions, yielding no evidence for naturalness bias. These results contrast with those of a similar study that found asymmetric extension of the stop voicing contrast, supporting an effect of naturalness bias. Confusion data suggests that the reduction in perceptibility from word-initial to word-final position may be greater for stop voicing than for stop place of articulation. This difference may underlie the divergent results of the two studies, leading to the hypothesis that the strength of a substantive bias depends on the magnitude of its phonetic precursor.

Cumulative markedness effects and (non-)linearity in phonotactics

How do grammars assess the well-formedness of words with multiple phonotactic violations? Certain models predict that as the strength of phonotactic restrictions decrease, forms that violate multiple restrictions should be less acceptable than expected, in a pattern we term super-linear cumulativity. We test this prediction using a series of Artificial Grammar Learning experiments, in which we vary the number of exceptions to phonotactic patterns in artificial languages. We find that super-linear cumulativity is indeed observed in the conditions with the weakest restrictions. Strikingly, participants exhibit super-linear cumulativity even when the trained language does not contain evidence for it.

Are People Able to Disentangle Perceptual and Conceptual Fluency? Evidence from Artificial Grammar Learning Experiment

The heuristic of information processing fluency plays an important role in making judgments. Some sources of processing fluency can be relevant or irrelevant to the content of a judgment. In this study, we aim to check whether individuals can distinguish different sources of fluency or fluency has a general effect on judgments. We used an artificial grammar learning paradigm (AGL) and tested the effects of different fluency sources (grammaticality and perceptual noise) on the judgment of grammaticality or of subjective ease of reading. It was found that both grammaticality and perceptual noise affected grammaticality judgements: the grammatical and the less noisy strings were evaluated more often as grammatical. However, only the perceptual noise affected judgments of subjective ease of reading. The results obtained provide evidence that fluency may contribute to the effects of implicit learning. It is possible that the processing fluency heuristic is the additional factor of judgement in the lack of explicit knowledge. Perhaps, perceptual noise provided almost complete explicit information for judgment of ease of reading; hence there was no need for additional heuristics. Another possible explanation is that perceptual fluency sources affect the early stages of information processing in a mandatory manner, unlike the conceptual ones. Overall, results are better explained by the non-specificity fluency hypothesis supporting the impossibility to distinguish between different fluency sources.

Distinguishing cognitive from historical influences in phonology

Distinguishing cognitive influences from historical influences on human behavior has long been a disputed topic in behavioral sciences, including linguistics. The discussion is often complicated due to empirical evidence being consistent with both the cognitive and the historical approach. This article argues that phonology offers a unique test case for distinguishing historical and cognitive influences on grammar, and it proposes an experimental technique for testing the cognitive factor which controls for the historical factor. The article outlines a model called catalysis for explaining how learnability influences phonological typology and presents experiments that simulate this process. Central to this discussion are unnatural phonological processes, that is, those that operate against universal phonetic tendencies and require complex historical trajectories in order to arise. By using statistical methods for estimating historical influences, mismatches in predictions between the cognitive and historical approaches to typology can be identified. By conducting artificial grammar learning experiments on processes for which the historical approach makes predictions that differ from those of the cognitive approach, the experimental technique proposed in this article controls for historical influences while testing cognitive factors. Results of online and fieldwork experiments on two languages, English and Slovenian, show that subjects prefer postnasal devoicing over postnasal fricative occlusion and devoicing in at least a subset of places of articulation, which aligns with the observed typology. The advantage of the proposed approach over existing experimental work is that it experimentally confirms a link between synchronic preferences and typology that is most likely not influenced by historical biases. Results suggest that complexity avoidance is the primary influence cognitive bias has on phonological systems in human languages. Applying this technique to further alternations should yield new information about those cognitive properties of phonological grammar that are not conflated with historical influences.

Local and non-local dependency learning and emergence of rule-like representations in speech data by deep convolutional generative adversarial networks

This paper argues that training Generative Adversarial Networks (GANs) on local and non-local dependencies in speech data offers insights into how deep neural networks discretize continuous data and how symbolic-like rule-based morphophonological processes emerge in a deep convolutional architecture. Acquisition of speech has recently been modeled as a dependency between latent space and data generated by GANs in Beguš (2020b), who models learning of a simple local allophonic distribution. We extend this approach to test learning of local and non-local phonological processes that include approximations of morphological processes. We further parallel outputs of the model to results of a behavioral experiment where human subjects are trained on the data used for training the GAN network. Four main conclusions emerge: (i) the networks provide useful information for computational models of speech acquisition even if trained on a comparatively small dataset of an artificial grammar learning experiment; (ii) local processes are easier to learn than non-local processes, which matches both behavioral data in human subjects and typology in the world’s languages. This paper also proposes (iii) how we can actively observe the network’s progress in learning and explore the effect of training steps on learning representations by keeping latent space constant across different training steps. Finally, this paper shows that (iv) the network learns to encode the presence of a prefix with a single latent variable; by interpolating this variable, we can actively observe the operation of a non-local phonological process. The proposed technique for retrieving learning representations has general implications for our understanding of how GANs discretize continuous speech data and suggests that rule-like generalizations in the training data are represented as an interaction between variables in the network’s latent space.

The Effect of Phonotactics on Alternation Learning

This study investigates whether alternation learning is facilitated by a matching phonotactic generalization. In a series of artificial grammar learning experiments, English learners were trained on artificial languages evincing categorical vowel harmony alternations across morpheme boundaries. These languages differed in the degree of harmony within stems (disharmonic, semiharmonic, and harmonic), and thus the degree of phonotactic support for the alternation. Results indicate that alternation learning was best when supported by matching stem phonotactics (harmonic language; experiment 1). Learners, however, were reluctant to extend a learned phonotactic constraint to novel unseen alternations (experiments 2 and 3). Taken together, the results are consistent with the hypothesis that alternation learning is facilitated by a matching static phonotactic generalization, but that learners are conservative in positing alternations in the absence of overt evidence for them.

The effect of phonotactics on alternation learning

This study investigates whether alternation learning is facilitated by a matching phonotactic generalization. In a series of artificial grammar learning experiments, English learners were trained on artificial languages evincing categorical vowel harmony alternations across morpheme boundaries. These languages differed in the degree of harmony within stems (disharmonic, semiharmonic, and harmonic), and thus the degree of phonotactic support for the alternation. Results indicate that alternation learning was best when supported by matching stem phonotactics (harmonic language; experiment 1). Learners, however, were reluctant to extend a learned phonotactic constraint to novel unseen alternations (experiments 2 and 3). Taken together, the results are consistent with the hypothesis that alternation learning is facilitated by a matching static phonotactic generalization, but that learners are conservative in positing alternations in the absence of overt evidence for them.

Constraint cumulativity in phonotactics: evidence from artificial grammar learning studies

An ongoing debate in phonology concerns the treatment of cumulative constraint interactions, or ‘gang effects’, and by extension the question of which phonological frameworks are suitable models of the grammar. This paper uses a series of artificial grammar learning experiments to examine the inferences that learners draw about cumulative constraint violations in phonotactics in the absence of a confounding natural-language lexicon. I find that learners consistently infer linear counting and ganging cumulativity across a range of phonotactic violations.

Unlearnable phonotactics

The Subregular Hypothesis (Heinz 2010) states that only patterns with specific subregular computational properties are phonologically learnable. Lai (2015) provided the initial laboratory support for this hypothesis. The current study aimed to replicate and extend the earlier findings by using a different experimental paradigm (oddball task) and a different measure of learning (sensitivity index, d′). Specifically, we compared the learnability of two phonotactic patterns that differ computationally and typologically: a simple rule (“First-Last Assimilation”) that requires agreement between the first and last segment of a word (predicted to be unlearnable), and a harmony rule (“Sibilant Harmony”) that requires the agreement of features throughout the word (predicted to be learnable). The First-Last Assimilation rule was tested under two experimental conditions: one where the training data were also consistent with the Sibilant Harmony rule, and one where the training data were only consistent with the First-Last rule. As in Lai (2015), we found that participants were significantly more sensitive to violations of the Sibilant Harmony (SH) rule than to the First-Last Assimilation (FL) rules. However, unlike Lai (2015), we also found that participants showed some residual sensitivity to the First-Last rule, but that sensitivity interacted with rule type so that participants were significantly more sensitive to SH rule violations. We conclude that participants in Artificial Grammar Learning experiments exhibit evidence of Universal Grammar constraining their learning, but patterns predicted to be unlearnable as a linguistic system can still be learned to some degree, due to non-linguistic learning mechanisms.

An inductive learning bias toward phonetically driven tonal phonotactics

ABSTRACT In two artificial grammar learning experiments, we tested the learnability of tonal phonotactics forbidding non-domain-final rising tones (*NonFinalR) against the phonotactics banning non-domain-final high-level tones (*NonFinalH). We propose that a firm phonetic ground drives a presumably innate inductive bias favoring *NonFinalR and against *NonFinalH. In Exp. I, we trained two groups of participants with an artificial language conforming to either *NonFinalR or *NonFinalH and tested them with the same set of novel items violating either tonal constraint in an acceptability judgment task. In two separate test sessions, *NonFinalR learners demonstrated a significantly higher consistency in making correct judgments than *NonFinalH learners. In Exp. II, learners participated in the same acceptability judgment task without being exposed to inputs in an a priori training session; participants had to learn from the immediate explicit feedback given to their judgments on every test item. Results suggest that only *NonFinalR learners demonstrated signs of converging on the target tonal phonotactics. In addition, both experiments found that *NonFinalR learners, but not *NonFinalH learners, acquired the tonal phonotactics and a baseline segmental phonotactics prohibiting retroflex consonants similarly. Altogether, the experimental results support the hypothesis of an inductive learning bias toward *NonFinalR and against *NonFinalH.

Word Frequency Is a Cue to Lexical Category for 8-Month-Old Infants.

The linguistic distinction between function words (functors) (e.g., the, he, that, on…), signaling grammatical structure, and content words (e.g., house, blue, carry…), carrying meaning, is universal across the languages of the world. These two lexical categories also differ in their phonological makeup (functors being shorter and more minimal) and frequency of occurrence (individual functors being much more frequent than most content words). The frequency-based discrimination of the two categories could constitute a powerful mechanism for infants to acquire the basic building blocks of language. As functors constitute closed classes and content words come in open classes, we examined whether 8-month-old monolingual infants relied on word frequency to categorize and track functors and content words. In six artificial grammar-learning experiments, we have found that infants process frequent words as belonging to closed classes, and infrequent words as belonging to open classes, and they map the relative order of these categories following the basic word order of their native language. These findings provide the earliest evidence that infants use word frequency as a cue to lexical categories and combine them to build rudimentary representations of grammar.

Exploring Variation Between Artificial Grammar Learning Experiments: Outlining a Meta-Analysis Approach.

Artificial grammar learning (AGL) has become an important tool used to understand aspects of human language learning and whether the abilities underlying learning may be unique to humans or found in other species. Successful learning is typically assumed when human or animal participants are able to distinguish stimuli generated by the grammar from those that are not at a level better than chance. However, the question remains as to what subjects actually learn in these experiments. Previous studies of AGL have frequently introduced multiple potential contributors to performance in the training and testing stimuli, but meta‐analysis techniques now enable us to consider these multiple information sources for their contribution to learning—enabling intended and unintended structures to be assessed simultaneously. We present a blueprint for meta‐analysis approaches to appraise the effect of learning in human and other animal studies for a series of artificial grammar learning experiments, focusing on studies that examine auditory and visual modalities. We identify a series of variables that differ across these studies, focusing on both structural and surface properties of the grammar, and characteristics of training and test regimes, and provide a first step in assessing the relative contribution of these design features of artificial grammars as well as species‐specific effects for learning.

Word Frequency is a Cue to Lexical Category at 8 Months

The distinction between function words, signalling grammatical structure, and content words, carrying meaning, is universal. The two lexical categories also differ in their phonological makeup and frequency of occurrence. Thus, their frequency-based discrimination could constitute a powerful initial mechanism for infants to acquire the basic building blocks of language. As functors constitute closed classes, while content words come in open classes, we examined whether 8-month-old monolingual infants relied on word frequency to categorize and track functors and content words. In four artificial grammar-learning experiments we have found that infants treat frequent words as belonging to closed classes, and infrequent words as belonging to open classes and they map the relative order of these categories following the basic word order of their native language. These findings provide the earliest evidence that infants use word frequency as a cue to lexical categories and combine them to build rudimentary representations of grammar.

Inductive learning of locality relations in segmental phonology

This paper reports on a series of artificial grammar learning experiments focused on locality relations in patterns of long-distance consonant agreement (harmony) and disagreement (dissimilation). Participants in experimental conditions were exposed to dependencies affecting stem-suffix pairs of liquids at either a short-range (transvocalic, CVCVLV-LV) or medium-range (beyond-transvocalic, CVLVCV-LV) distance. Two experiments used a poverty of stimulus paradigm, offering no information about the other distance level; participants interpreted short-range interaction as a strictly transvocalic dependency but medium-range interaction as unbounded, generalizing to other distances. Two experiments employed a ‘rich stimulus’ paradigm, where training data unambiguously indicated the absence of any dependency at the other distance; this enabled probing of specific locality patterns, in particular strictly beyond-transvocalic dependencies. The constraint-based Agreement by Correspondence model of non-adjacent consonant interactions predicts such patterns to be possible for dissimilation but not harmony. The results do not support this hypothesis: Participants seem to have serious difficulty learning strictly beyond-transvocalic dependencies of either kind. Our findings are more consistent with recent proposals that the space of learnable phonotactic restrictions is delimited by the Tier-based Strictly 2-Local class of formal languages. Strictly transvocalic and unbounded dependencies lie within this region, whereas strictly beyond-transvocalic dependencies are more complex, falling outside the learner’s hypothesis space.

Is phonological consonant epenthesis possible? A series of artificial grammar learning experiments

Consonant epenthesis is typically assumed to be part of the basic repertoire of phonological grammars. This implies that there exists some set of linguistic data for which epenthesis is the best analysis. However, a series of artificial grammar learning experiments found no evidence that learners ever selected an epenthesis analysis. Instead, phonetic and morphological biases were revealed, along with individual variation in how learners generalised and regularised their input. These results, in combination with previous work, suggest that synchronic consonant epenthesis may only emerge very rarely, as a result of a gradual accumulation of changes over time. It is argued that stored stem–allomorph pairs should be adopted as the null hypothesis in cases of potential epenthesis, and that a universal criterion is needed for rejecting this hypothesis. Such a criterion requires a formal theory of exceptions, a necessity for falsifiability in phonological theory.