Dissociating Sources of Knowledge in Artificial Grammar Learning - eScholarship

Dissociating Sources of Knowledge in Artificial Grammar Learning Michelle A. Hendricks (mpatte20@slu.edu), Christopher M. Conway (cconway6@slu.edu), and Ronald T. Kellogg (rkellogg@slu.edu) Department of Psychology, Saint Louis University St. Louis, MO 63108 USA Abstract Previous studies have suggested that individuals use both implicit and explicit, as well as rule and exemplar-based knowledge, to make grammaticality judgments in artificial grammar learning (AGL) tasks. Experiment 1 explored the importance of explicit mechanisms in the learning of exemplar and rule-based information by using a dual-task during AGL training. We utilized a balanced chunk strength grammar, assuring an equal proportion of explicit exemplar-based cues (i.e. chunks) between grammatical and non- grammatical test items. Experiment 2 explored the importance of perceptual cues by changing letters between AGL training and test, while still incorporating the dual-task design and balanced chunk strength grammar used in Experiment 1. Results indicated that participants with a working memory load learned the grammar in Introduction There is widespread agreement that there exist two distinct forms of learning, explicit and implicit. Explicit learning refers to learning that happens actively, consciously, and with effort, such as the type of learning that occurs during much of formal education. Implicit learning, on the other hand, occurs passively, unconsciously, and without effort. Implicit learning is theorized to be involved in procedural motor activities such as riding a bike or typing, as well as in more complex phenomena such as social interaction and language learning (Reber, 1993). Artificial grammar learning (AGL) has been a useful paradigm for the study of implicit learning. In the typical artificial grammar learning (AGL) paradigm, individuals are shown (or asked to memorize) letter strings that, unknown to them, conform to rules instantiated by an artificial grammar. Following presentation of the training exemplars, participants are able to reliably determine whether a newly presented letter string is grammatical according to the artificial grammar, without being able to explicitly verbalize the rules of the grammar. Originally, it was theorized that individuals rely on an implicit abstract rule-learning system during AGL tasks, with participants’ failures to verbalize the rules as evidence that the rules were unconscious (Reber, 1989). Additional support for implicit rule-based learning in AGL was provided by what are now referred to as “transfer” experiments. In an AGL transfer experiment, the surface features (e.g. letters) of the training exemplars are changed during the test phase, though the underlying grammar stays the same. Clearly, this would make grammaticality decisions based solely on item similarity difficult, if not impossible. Thus, the transfer manipulation is meant to increase reliance on (presumably implicit) rules divorced from the surface details of the exemplars. Impressively, results from multiple studies have indicated that individuals still successfully demonstrate above-chance classification Experiment 1 just as well as the single-task no-load group, presumably by relying solely on implicit learning mechanisms. However, changing the letters from training to test resulted in no significant learning for dual-task participants in Experiment 2, suggesting that exemplar-based perceptual cues may the major contributor to implicit knowledge. Overall, the results suggest that implicit and explicit mechanisms for learning rule-based and exemplar-based information may both contribute to AGL via four independent, parallel routes, providing a new framework for understanding the complex dynamic of learning in AGL tasks. Keywords: artificial grammar learning; implicit learning; working memory; dual-task performance, though the learning is often attenuated (Reber, 1989, Knowlton & Squire, 1996). In addition to the transfer studies, multiple studies have shown that amnesic subjects, who putatively cannot rely on explicit forms of learning, demonstrate artificial grammar learning similarly to non-brain damaged controls (Knowlton, Ramus, and Squire, 1992; (Knowlton & Squire, 1996). The evidence from both the transfer and the amnesic studies suggest that AGL is mediated by implicit rule- learning mechanisms. Under this view, given that implicit learning is theorized to happen automatically and without effort, executive functions such as working memory (an explicit mechanism, by definition) should have a minimal impact on artificial grammar learning. Although studies with amnesic patients strongly suggest that AGL can occur without explicit memory, research with non-brain damaged subjects suggests that under normal conditions, explicit processes are also recruited. For instance, test phase classification judgments have been found to be sensitive to the similarity between test and training items, specifically in terms of chunk strength (Chang & Knowlton, 2004; Knowlton & Squire, 1996). Chunks are bigrams and trigrams that are encountered frequently in an artificial grammar due to repetitions in the underlying structure. Studies have shown that individuals do retain some explicit information regarding the chunks of the training items (Dienes, Broadbent, & Berry, 1991; Dulany, Carlson, & Dewey, 1984), and that participants studying only training bigrams can classify the grammaticality of test items correctly at rates similar to controls (Perruchet & Pacteau, 1990). In addition, fMRI studies of AGL tasks have suggested some involvement of the medial temporal lobe (MTL; Fletcher, Buchel, Josephs, Friston, & Dolan, 1999; Opitz & Friederici, 2004). These findings suggest that individuals may rely on a combination of both implicit rule-based knowledge and explicit exemplar-based chunk knowledge to make grammaticality judgments (Vokey & Brooks, 1992; Knowlton & Squire,