Slide Slam B13
Robust effects of working memory load during naturalistic language comprehension in language-selective cortex
Cory Shain1, Blank Idan2, Gibson Edward1, Fedorenko Evelina1, Schuler William3; 1MIT, 2UCLA, 3The Ohio State University
Some theories of human language comprehension posit processes that retrieve and update representations in working memory (WM; Gibson, 2000; Lewis & Vasishth, 2005). Others hypothesize prediction processes that might underlie comprehension (Levy, 2008). Prior behavioral and neuroimaging experiments have supported effects of memory (Fiebach et al., 2001) and prediction (Bonhage et al., 2015) in isolation using carefully constructed stimuli. Empirically distinguishing these two classes of accounts is generally challenging (Levy & Gibson, 2013), and many previously reported effects of syntactic processing (e.g., syntactic anomalies; Osterhout & Holcomb, 1992) are consistent with both theories. Further complicating matters are poorly understood effects of experimental design in studies of human language comprehension (Hasson et al, 2018). Although studies overwhelmingly investigate syntax using carefully selected or hand-crafted sentences presented out of context, linguistic input in typical conditions is richly varied and contextualized, and growing neuroscientific evidence indicates that artificial stimuli and tasks may engage cognitive mechanisms that are not central to language processing (Diachek, Blank, & Siegelman et al., 2020). Importantly, recent naturalistic behavioral studies have generally not shown strong evidence of memory retrieval costs (Demberg & Keller, 2008), casting doubt on theories that hypothesize syntax-driven memory operations as central to comprehension. However, it is possible that WM effects are underlyingly present without registering strongly in behavioral measures like reading time, which may aggregate over a range of mental processes (Rayner, 1998). Furthermore, granting a core role of WM in naturalistic language comprehension, debate exists about the extent of overlap between neural mechanisms that support WM for language processing vs. WM as required by other domains of cognition. Some have argued for language-specific working memory resources (Caplan & Waters, 1999), whereas others have argued that language comprehension relies primarily on domain-general WM resources (Amici et al., 2007). This study jointly investigates the existence and functional specificity of WM load during language comprehension using data from a large-scale naturalistic fMRI study (Shain, Blank et al., 2020) under rigorous controls for word predictability. We first explore multiple plausible implementations of prior theories of WM in language processing, then evaluate the most robust of these on unseen data. In so doing, we examine neural responses in two candidate brain networks, each functionally localized in individual participants: the language-selective network (LANG; Fedorenko et al., 2010) and the domain-general multiple-demand (MD) network implicated in executive functions, including WM (Duncan, 2010). Responses in the MD network show no evidence of an association with any of the estimators of working memory load explored here. Responses in the language network show clear and generalizable WM effects, especially as described by the Dependency Locality Theory (Gibson, 200). Based on these results, we argue (1) that a core function of the human language processor is to compose representations in working memory based on syntactic cues, even in naturalistic settings, (2) that dependency locality best describes the cost of these composition operations among extant broad-coverage theories of WM in language processing, and (3) that these operations are implemented locally within the language-selective cortical network.