Poster Presentation

Search Abstracts | Symposia | Slide Sessions | Poster Sessions

Temporal Dissociation of Syntactic Disambiguation and Memory Retrieval during Sentence Processing: Naturalistic MEG Evidence from Interpretable Models

Poster Session C, Thursday, October 1, 10:45 am - 12:45 pm, Wangari Maathai

Dylan Scott Low¹, Donald Dunagan¹, Shisen Yue¹, Lars Meyer^2,3,4, John Hale¹; ¹Johns Hopkins University, ²Johannes Gutenberg University Mainz, ³Max Planck Institute for Human Cognitive and Brain Sciences, ⁴Clinic for Phoniatrics and Pediatric Audiology, University Hospital Münster, Münster, Germany

Human sentence comprehension proceeds word-by-word (Marslen-Wilson, 1973; Tanenhaus et al., 1995). At some level it must involve working memory, for instance when connecting information contributed by previously-heard words to the current word. Likewise, there is a consensus that some sort of expectancy-based disambiguation process allows people to decide how words fit into the prior context (see Frazier, 2013, for a review). It remains to be seen what the joint effect of these two hypothesized cognitive operations might be in naturalistic listening. To investigate this, we re-analyzed the MEG data of Brodbeck et al. (2022). This re-analysis suggests that the two operations temporally dissociate. The MEG data come from 12 people listening to an audiobook. The re-analysis applies multivariate temporal response functions (mTRFs). Word-by-word metrics of processing complexity for the audiobook were derived from two cognitive models: the Grammar Bilinear Asymmetric word2vec model of Yue (2026) and the Incremental Left-Corner Generative Dependency Parser of Dunagan (2025). Yue’s model implements cue-based retrieval theory (Lewis et al., 2006) wherein the current word serves as a cue for the retrieval of word-related information from the prior context. In this model, working memory is used every time a person forms a grammatical dependency (for example, Subjecthood or Direct Objecthood). Following Ryu & Lewis (2025) and Oh & Schuler (2022), this model quantifies retrieval difficulty via a diffuseness measure called normalized attention entropy. The other model is the dependency parser introduced in Dunagan (2025). It disambiguates, word by word, between partial syntactic analyses that sometimes include as-yet-unheard words (see Jurafsky & Martin, 2026, Chapter 19, for a pedagogical introduction to dependency parsing). This model quantifies disambiguation cost via the surprisal linking hypothesis. Surprisal and normalized attention entropy values from the two cognitive models serve as predictors in mTRFs of source-localized MEG data, alongside acoustic and word frequency predictors of non-interest for this analysis. In a spatiotemporal analysis of the estimated TRFs, we analyzed the dorsal and ventral streams of the bilateral language network. Results suggest two processing stages — an initial stage well-modeled by surprisal from Dunagan’s dependency parser between 300–500 ms, and a subsequent stage between 750–850 ms that is well-modeled by normalized attention entropy values from Yue’s memory retrieval model. These findings are roughly consistent with existing theories. It accords with the Retrieval- Integration theory of Brouwer et al. (2012; 2017) which distinguishes between an early stage involving lexical operations and a later stage involving the formation of grammatical dependencies. Our memory finding maps on to phase 3 of Friederici’s (2002) Neurocognitive Model. It is also consistent with Martin et al. (2012; 2014) who associate memory retrieval difficulty with late effects. In virtue of sharing the same set of grammatical dependency types across Dunagan’s model and Yue’s model, the overall proposal offers a consistent high-level linguistic interpretation of Brodbeck et al.’s (2022) MEG data.

Topic Areas: Computational Approaches,