Disentangling Neural Indices of Implicit vs. Explicit Morphosyntax Processing in an Artificial Language

Learning new languages is a complex task involving both explicit and implicit processes (i.e., that do/do not involve awareness). Understanding how these processes interact is essential to a full account of second language (L2) learning, but accounts vary as to whether explicit processes help (e.g., DeKeyser, 2007), hinder (e.g., Ellis & Sagarra, 2010), or have no effect on (e.g., Paradis, 2009) the successful acquisition of implicit processing routines. Studies using an artificial language paradigm have established that participants can learn L2 morphosyntactic regularities that they are unaware of (Leung & Williams, 2011, 2012), and a subsequent electroencephalography (EEG) experiment (Batterink et al., 2014) reported different distinct event-related potentials (ERPs) in participants with vs. without awareness of the covert regularity. However, the univariate nature of ERPs makes it impossible to determine whether/to what extent implicit processing occurred in rule-aware learners. Our study addresses this limitation using multivariate pattern analysis (MVPA; Fahrenfort et al., 2018) by training a decoder to detect neural indices of grammar processing at times in the experiment after behavioral measures indicated rule-learning but before participants became rule-aware, and subsequently testing this decoder after participants became rule-aware. We also conduct two follow-up analyses that shed light on the interplay between implicit and explicit grammar processing, asking whether EEG indices of semantic prediction vary between implicit vs. explicit learning as assessed by MVPA, and whether the timing of grammar processing at the neural level is correlated (and thus closely coupled) with behavioral response times. Following Batterink et al., 52 participants performed a word-classification task that covertly tests for grammar learning by comparing responses to words that follow vs. violate an underlying pattern. Rule-awareness was assessed via systematic debriefing halfway through the task, at which point the rule was revealed and a final block of trials was performed. Slower reaction times and lower accuracies for rule-violating trials indicated learning even in rule-unaware participants, replicating Batterink et al. However, we did not replicate Batterink et al.’s ERP findings, as we only found a negative deflection in rule-unaware participants and no significant ERP in rule-aware participants. This may be due to natural interindividual variability in ERPs during grammar processing (Tanner, 2019). Furthermore, our MVPA decoding did not show above-chance trial classification accuracy, providing no evidence for the co-occurrence of implicit processing during periods of rule awareness. We also found no evidence for semantic prediction at the neural level in either rule-aware or rule-unaware learners using MVPA. However, for both of these results, follow-up analyses suggested limited decoding sensitivity with the MVPA method on our data in the first place, which was not improved when using a host of alternate analysis parameters. Our ERP-to-reaction time correlation analyses showed evidence in favor of time locking between neural indices of grammar processing and behavioral responses, suggesting a link between the two; however, these results were tempered by the weak ERP effects we found. Overall, the results show strong behavioral learning effects but limited EEG effects. This, along with follow-up behavioral analyses, leads me to question the extent to which learning in this experiment is linguistic vs. non-linguistic. To the extent that our observed learning is linguistic, our results favor weak/no interface models in that rule-unaware and rule-aware participants showed equivalent behavioral performance and there was no MVPA evidence for implicit processing during rule awareness. However, qualitative inspection of the behavioral and debriefing data suggest possible downsides to explicit awareness on task performance. More broadly, this study demonstrates how alternate analysis methods may inform future research on the implicitness/explicitness of second language grammar learning.