Slide Slam B16
Generalized additive mixed modeling of EEG supports dual-route account of morphosyntax in finding no word frequency effects on grammar processing
David Abugaber1, Irene Finestrat1, Alicia Luque2, Kara Morgan-Short1; 1University of Illinois - Chicago, 2UiT The Arctic University of Norway
A critical component of understanding how our minds process language is accounting for inflected forms, e.g., the difference between "eat" and "eats" in a sentence like "They eat/*eats." Single-route models of morphosyntax posit that processing of regular inflections involves associative memory-based storage (e.g., McClelland & Patterson, 2002), whereas dual-route models alternately propose rule-governed composition (e.g., Pinker & Ullman, 2002). These accounts can be tested by their divergent predictions regarding whether word frequency—an indirect approximation of the strength of stored representations—affects processing of regular morphosyntactic inflections (as in the single-route model) or not (dual-route model). To our knowledge, the only study to test this using electroencephalography (EEG) comes from Allen, Badecker, and Osterhout (2003), who examined whether grammatical processing of verb tense violations ("will work/*worked") was affected by word frequency (low-/high-frequency). No interaction between grammaticality and word frequency was found for regular forms, consistent with the dual-route model’s predictions. However, reproduction of these results for a different regularly inflected morphosyntactic form would increase the generalizability and validity of the findings. Furthermore, applying a newer statistical analytical approach can address limits in statistical power from their event-related potential-based analysis (arising from time-window averaging of EEG amplitudes [Fields & Kuperberg, 2020] and from dichotomization into low-/high-frequency bins [Baayen, 2004]), as well as theoretical issues related to the assumption that the time course of language processing is identical across words/participants (see Sassenhagen, Schlesewsky, & Bornkessel-Schlesewsky, 2014). Thus, we extend Allen et al. (2003) in a conceptual replication experiment by using generalized additive mixed models (GAMM), which retain per-trial and per-time sample information for improved statistical power while allowing the time course of language processing to vary idiosyncratically across participants and words. Our data come from 51 English native speakers who completed a grammaticality judgment task (GJT) while reading 124 experimental sentences that either did or did not contain a determiner-noun agreement violation (e.g., "this/*these school"). We follow the GAMM-fitting procedure based on Wieling (2018), with word frequency taken from the British National Corpus as a continuous predictor. Our pre-planned analysis involved sequentially fitting models to predict mean EEG amplitude in an eight-electrode region of interest (based on Tanner, 2019) by (a) adding fixed linear terms for word grammatical/ungrammatical status and frequency while allowing amplitude to vary across time; (b) allowing effects of grammaticality and frequency to vary non-linearly across time; and (c) testing whether the grammaticality-by-frequency interaction significantly improved model fit as per the Akaike Information Criterion (AIC). Our results reproduced previously-reported main effects of frequency and grammaticality, with significant effects for these parametric coefficient terms at p < .05. Critically, we found no significant improvement to model fit when including the grammaticality-by-frequency interaction, as per the significance criterion of a minimum reduction threshold of 2 AIC units for selecting a more complex model (following Wieling, 2018). Our results are consistent with findings from Allen et al. (2003) and align with the dual-route model’s account of rule-based composition as a mechanism that underlies processing of regularly-inflected morphosyntactic forms.