Artificial Grammar Learning in Children With Developmental Language Disorder

Purpose: The purpose of the study was to evaluate implicit learning in children with developmental language disorder (DLD) by employing a visual artificial grammar learning task. Method: Thirteen children with DLD and 24 children with typical language development between the ages of 8 and 12 years completed a visual artificial grammar learning task. During the training phase of the task, participants were presented with strings of shapes that followed the underlying structure of a finite grammar. During the testing phase, participants were asked to judge whether new strings were grammatical or nongrammatical. Grammatical judgment of new strings served to measure generalization of the underlying grammatical structure. Endorsement based on chunk strength, or similarity to training exemplars, served to evaluate the extent to which children relied on surface features to guide their task performance. Results: As a group, children with typical development performed better on the artificial grammar learning task, compared with children with DLD, and accepted more grammatical strings regardless of their similarity to training exemplars. Task performance in both groups was not affected by surface features. Performance of children with DLD whose test accuracy exceeded the learning threshold of 0.5 was consistent with a generalization of the underlying grammatical structure that was unaffected by surface features. Conclusions: The study found group differences in learning outcomes between children with and without DLD. Consistent with previous reports, children with typical development correctly endorsed more grammatical strings than children with DLD, suggesting a better acquisition of the grammatical structure. However, there was no evidence to suggest that children in the DLD group (learners and nonlearners) relied on surface features (i.e., familiarity to training exemplars) in their grammatical judgment. These results refute our hypothesis that children in the DLD group would show judgment based on familiarity.

Literacy effects on artificial grammar learning (AGL) with letters and colors: evidence from preschool and primary school children

abstract Literacy affects many aspects of language and cognition, including the shift from a more holistic mode of processing to a more analytical part-based mode of processing. Here we examined whether this shift impacts the ability of preschool and primary school children to learn the rules underlying a finite-state grammar using an artificial grammar learning (AGL) paradigm implemented with either linguistic (letters) or non-linguistic (colors) materials to further examine if children’s AGL performance was modulated by type of stimuli. Both tasks involved a training phase in which half of the preschool children and half of the primary school children were exposed to a set of either letter or color strings without any information about the rules underlying the construction of those strings. Later, in the test phase, they were asked to decide whether a new set of letter or color strings conformed to those rules to test grammar learning. Results showed that only primary school children showed evidence of learning, and, importantly, only with colors. These findings seem to support the view that learning to read promotes reliance on smaller linguistic units that might hinder the ability of first-graders to learn the rules underlying finite-state grammars implemented with linguistic materials.

Desirable difficulties in statistical learning? How talker variability impacts artificial grammar learning

Contending with talker variability has been found to lead to processing costs but also benefits by focusing learners on invariant properties of the signal. These discrepant findings may indicate that talker variability acts as a desirable difficulty. That is, talker variability may lead to initial costs followed by long term benefits for retention and generalization. Adult participants learned an artificial grammar affording learning of multiple components by 1-, 2- or 8- talkers, tested at 3 time points. The 8-talker condition did not impact learning. The 2-talker condition negatively impacted some aspects of learning, but only under more difficult learning conditions. Across both experiments, generalization of the grammatical dependency was difficult. Taken together, we discovered that high and limited talker variability differentially impact artificial grammar learning. However, talker variability does not act as a desirable difficulty in the current paradigm, as the few evidenced costs were not related to long-term benefits.

Frequency as a Barrier to Learning Complex Nominal Morphology

Most languages with highly structed morphological systems show some degree of syncretism, where the same affix is used for multiple categories. The typology of syncretism has suggested that syncretism is most likely to occur for structurally and semantically marked categories. In two artificial grammar learning experiments, English-speaking adults were exposed to a 3-gender x 3-number nominal system, where one number category (Singular, Dual, or Plural) showed syncretism across gender. In the experiment, the frequency of the syncretic morpheme was equal to non-syncretic morphemes, but there were 3x fewer items containing the syncretic morpheme. Participants failed to learn the syncretic morpheme, with no biases for marked categories. These results suggest that low frequency of syncretic items significantly impairs learning syncretic categories. Suggestions for design

Comparing feedforward and recurrent neural network architectures with human behavior in artificial grammar learning

In recent years artificial neural networks achieved performance close to or better than humans in several domains: tasks that were previously human prerogatives, such as language processing, have witnessed remarkable improvements in state of the art models. One advantage of this technological boost is to facilitate comparison between different neural networks and human performance, in order to deepen our understanding of human cognition. Here, we investigate which neural network architecture (feedforward vs. recurrent) matches human behavior in artificial grammar learning, a crucial aspect of language acquisition. Prior experimental studies proved that artificial grammars can be learnt by human subjects after little exposure and often without explicit knowledge of the underlying rules. We tested four grammars with different complexity levels both in humans and in feedforward and recurrent networks. Our results show that both architectures can “learn” (via error back-propagation) the grammars after the same number of training sequences as humans do, but recurrent networks perform closer to humans than feedforward ones, irrespective of the grammar complexity level. Moreover, similar to visual processing, in which feedforward and recurrent architectures have been related to unconscious and conscious processes, the difference in performance between architectures over ten regular grammars shows that simpler and more explicit grammars are better learnt by recurrent architectures, supporting the hypothesis that explicit learning is best modeled by recurrent networks, whereas feedforward networks supposedly capture the dynamics involved in implicit learning.

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.