Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Slide Session C: Syntax, Semantics, & Network Plasticity
Friday, October 2, 10:00 - 11:00 am, Amphitheatre D
Talk 1: Two Words Are Faster Than One: Neural Dynamics of Accelerated Meaning in Phrases
Shaonan Wang1, Songhee Kim2, Jeffrey R. Binder2, Nan Lin3, Liina Pylkkänen4; 1Hong Kong Polytechnic University, 2Medical College of Wisconsin, 3Chinese Academy of Sciences, 4New York University
Understanding how the brain constructs complex meanings from individual words is a central challenge in cognitive neuroscience. Although substantial progress has been made in identifying the neural networks that support phrase-level composition, the role of semantic content—specifically, the internal structure of conceptual representations—remains insufficiently characterized. Recent componential models grounded in experiential features provide a principled framework for linking conceptual structure to neurobiology by representing concepts as high-dimensional vectors spanning perceptual, motor, spatial, temporal, affective, and social domains. In the present study, we used magnetoencephalography (MEG) to investigate whether experiential semantic features are retrieved and integrated during two-word phrase comprehension, and whether lexical and phrasal representations unfold in strictly serial or temporally overlapping fashion. Participants read two-word phrases and their constituent words presented in a rapid parallel visual presentation (RPVP) paradigm, equating temporal input across stimulus types. The stimulus set comprised 216 phrases spanning six distinct linguistic relation types and 107 component words. All words and phrases were independently rated on 65 experiential features across 14 experiential domains. We applied time-resolved representational similarity analysis (RSA) to quantify the alignment between neural activation patterns and experiential semantic structure over the course of processing. Contrary to strictly serial accounts, phrase-level semantic representations emerged earlier than single-word representations presented in isolation. Neural alignment with phrase-level experiential structure outpaced that of isolated words, indicating rapid construction of integrated meaning. Moreover, phrasal context accelerated lexical-semantic access: for example, the word “cake” showed semantic activation at approximately 312 ms when presented alone, but at approximately 236 ms when embedded in the phrase “green cake.” These findings demonstrate that combinatory context facilitates, rather than delays, lexical-semantic retrieval, even when the context is not temporally in the past. Critically, experiential feature representations corresponding to both constituent words were detectable during phrase processing, with partially overlapping temporal windows. This temporal overlap is inconsistent with models positing that lexical representations are fully resolved and then replaced by integrated phrase meanings. Instead, the data support interactive or cascaded architectures in which multiple levels of representation remain concurrently available during composition. Furthermore, the magnitude and temporal profile of composition-related effects varied systematically as a function of phrasal relation type and experiential domain. Certain relations elicited stronger or earlier integration of specific feature domains, indicating that semantic composition is sensitive to both relational structure and content type. Together, these results provide temporally resolved evidence that experiential semantic features are dynamically retrieved and integrated during real-time language comprehension. Composition does not merely operate over abstracted lexical tokens; rather, it recruits structured experiential knowledge in a relation-dependent manner. By demonstrating accelerated semantic access in phrasal contexts and overlapping lexical and combinatorial representations, this study advances our understanding of how distributed semantic features are orchestrated to support the rapid construction of complex meaning.
Talk 2: Cross-modal decoding of syntactic frames from word comprehension to sentence production
Adam Morgan1,2, Orrin Devinsky1, Werner Doyle1, Patricia Dugan1, Daniel Friedman1, Analia Arevalo2, Guilherme Lepski2, Adeen Flinker1,3; 1NYU Grossman School of Medicine, 2University of São Paulo School of Medicine, 3NYU Tandon School of Engineering
Syntax, the abstract hierarchical structure of language, remains poorly understood in the brain. Most studies compare conditions that engage syntax more or less strongly. These contrasts have identified syntax-sensitive regions and networks, but they operate at a relatively coarse grain, obscuring representations and their encoding. They also typically rely on multi-word stimuli, where syntax is confounded with combinatorial processing. Here, we take a different approach. Rather than asking where syntactic processing is engaged, we ask whether we can detect the neural patterns associated with individual syntactic structures. And rather than relying on within-modality and within-task analyses, subject to overfitting, we ask whether neural patterns associated with particular structures during comprehension can be used to track which structures speakers are planning during sentence production in real time. We focused on verb frames: the hierarchical structures coordinating a verb’s arguments (AKA argument structure). For example, verbs take different frames, which are often independent of their meaning. For instance, "dine" (intransitive) and "devour" (transitive) have similar semantics but different syntax. Similarly, "donate" is associated with “NP-PP" syntax (donate [money]NP [to charity]PP), whereas "give" more commonly occurs with “NP-NP" syntax (give [me]NP [that]NP). Prior work has shown that verbs automatically activate associated syntactic frames, even in isolation. We therefore used single-verb comprehension as a neural localizer for specific syntactic frames, removing combinatorial confounds inherent to sentence-level stimuli. We recorded stereo-electroencephalography (sEEG) from neurosurgical patients while they completed two tasks. In the single-verb comprehension task, participants simultaneously saw and heard isolated verbs, selected to be strongly biased toward one of four syntactic frames: intransitive, transitive, NP-NP, or NP-PP. In a separate overt sentence-production task, participants described visual scenes using sentences that instantiated these frames; for example, “The seal slept” or “The knight was cooking [steak]NP [for the dog]PP.” For each patient, we trained a machine learning classifier to distinguish the four frames using neural activity (high-gamma) in the single-verb comprehension data. We then applied these classifiers, without retraining, to sentence-production trials to test whether the neural patterns identified during verb comprehension generalized to syntactic frame planning during overt production. Significance was assessed against shuffled-label distributions. In the first analyzed participant, classifiers trained only on single-verb comprehension successfully decoded the syntactic frame of the sentence being produced, starting before the onset of sentence articulation and persisting throughout the first several words of the sentences (analyzed in word-onset-locked epochs). Thus, neural patterns distinguishing syntactic frames in isolated verb comprehension generalized across modality (from comprehension to production) and task (isolated word processing to sentence processing). Data collection/analysis is ongoing; we expect two more patients analyzed by SNL2026, including spatial analyses to identify the regions driving cross-task generalization. This novel approach to studying syntax not only demonstrates shared syntactic representations across tasks and modalities, but also introduces a promising new approach to understanding the neural code for syntax by characterizing the neural patterns associated with particular syntactic representations rather than entire regions recruited by syntactic processing.
Talk 3: Neural Encoding of Verbs in Natural Sign Language Narrative
Maria Zimmermann1, Joanna Huczyńska1, Monika Kozub1, Piotr Tomaszewski1, Annemarie Kocab2, Marina Bedny2; 1University of Warsaw, 2Johns Hopkins University
The noun verb distinction is a putative linguistic universal. Across signed and spoken languages, nouns typically refer to entities, whereas verbs denote actions, events, and relations (Talmy, 1985; Langacker, 1987; Croft, 2025; Abner et al., 2019). Verbs also play a special role in sentence grammar, specifying relations between arguments in a sentence (Gleitman, 1990). A distinctive feature of many verbs in sign-languages is their iconicity (Padden, et al. 2015) The form of so called ‘depicting’ or ‘classifier verbs,’ such as ‘grab,’ ‘jump’ and ‘run’ includes a pantomime or showing the movement of the action denoted by the verb (Newman et al., 2015). For spoken languages, verbs and nouns have partially overlapping and yet distinctive neural signatures (Caramazza & Hillis, 1991; Perani et al., 1999; Marti, et al., 1995). Verb comprehension preferentially activates left lateral temporal cortices (LTC) and patterns of activity in the LTC are sensitive to verb meaning, whereas noun-responsive regions in ventral temporal and parietal networks are more sensitive to object noun meaning (Elli et al., 2019; Wurm & Carmazza, 2019). Here we use naturalistic fMRI, combining feature-rich linguistic annotations with encoding models to test whether the neural signatures of verbs/nouns are modality invariant i.e., observed for sign-languages in the context of naturalistic discourse comprehension. We applied an encoding model approach to study neural responses to verbs and nouns in a continuous naturalistic Polish Sign Language (PJM) narrative (Naselaris, et al., 2011; Huth et al 2016)) Twenty congenitally Deaf native PJM signers watched a naturalistic PJM narrative during fMRI. Native PJM signers annotated the narrative for part of speech: verbs (plain, directional, depicting/classifier) and nouns (plain, depicting/classifier). To dissociate lexical category (verb/noun) from potentially confounded visual motion and form properties, we used spatiotemporal Gabor filters to extracted motion-energy, direction and speed (Dupré La Tour et al., 2025), as well as mid-level kinematic features including derived with OpenPose-based body-pose estimation (Cao et al., 2021). Additionally, we extracted semantic features from verbs and nouns using word2vec embeddings. We modeled BOLD responses as a function of linguistic and perceptual features using voxelwise encoding models with ridge regression and nested cross-validation (Dupré La Tour et al., 2022). Consistent with prior findings with spoken languages, verbs reliably explained variance in the left lateral temporal cortex, whereas nouns did not. Responses to nouns were preferentially observed in ventral temporal, parietal and precuneus. Model comparison and winner analyses further showed that verb-based models outperformed noun models specifically within left posterior temporal cortex and these effects could not be explained by low-level motion-energy or kinematic features. LTC verb responses were observed across verb types and robust for depicting/classifier (iconic) verbs. Representational similarity analysis further demonstrated that LTC activity reflected semantic similarity among verbs rather than low-level visual or kinematic similarity. We find that previously identified neural signatures of verbs vs. nouns are modality invariant, iconicity invariant, and identifiable in naturalistic sign language discourse.
Talk 4: Reorganisation of Functional Connectivity Gradients in Post-Stroke Aphasia
Ramya Balakrishnan1, Tirso Rene del Jesus Gonzalez Alam2, Robert Leech3, Cathy J Price4, Elizabeth Jefferies1; 1University of York, 2Bangor University, 3Kings College London, 4University College London
Aphasia after stroke arises from focal damage to language regions as well as widespread changes in connectivity, yet conventional approaches often struggle to capture these distributed effects. Here, we introduce whole brain functional connectivity gradients as a novel framework to characterize the cortical impact of stroke and its consequences for language outcomes. Connectivity gradients are continuous axes describing functional organisation across the cortex, with different dimensions capturing distinctions between regions based on their patterns of connectivity. In 60 individuals with chronic aphasia from the PLORAS (Predict Language Outcome and Recovery After Stroke ) dataset, we combined structural lesion–symptom mapping with a lesion–gradient symptom mapping approach. Language outcomes, derived from principal component analysis of the Comprehensive Aphasia Test, were linked both to lesion sites and to alterations in the brain’s connectivity gradients derived from resting state fMRI. Structural lesion–symptom mapping identifies which damaged regions are associated with specific language impairments but does not capture how surviving cortex is reconfigured within the brain’s large scale architecture. In contrast, lesion–gradient mapping characterises whole brain connectivity gradients – data driven axes that summarise the principal patterns of variation in cortical connectivity – and quantifies how behavioural variation relates to displacement of structurally intact regions along these intrinsic axes. Stroke was associated with altered positioning of cortical regions along the gradient separating default mode and control networks. Notably, the position of the left inferior frontal gyrus along this axis predicted dissociable language outcomes: displacement toward default mode connectivity patterns was associated with better speech but poorer writing performance. This opposing behavioural profile suggests that inferior frontal cortex contributes to language through flexible large scale coupling, with distinct coupling regimes differentially supporting spoken and written production. These findings indicate that gradient based mapping provides a mechanistic link between focal tissue damage and distributed behavioural consequences by revealing systematic reorganisation of macroscale functional architecture beyond the lesion site.