Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks
Tracking word meanings across contexts in MEG with Representational Similarity Analysis
There is a Poster PDF for this presentation, but you must be a current member or registered to attend SNL 2023 to view it. Please go to your Account Home page to register.
Poster A19 in Poster Session A, Tuesday, October 24, 10:15 am - 12:00 pm CEST, Espace Vieux-Port
Aline-Priscillia Messi1, Liina Pylkkanen1; 1New York University
How does context influence the neural representations of lexical items during online processing? Considering the inherent ambiguity and context-dependent nature of most word forms, this study aims to investigate the representational effects of context dependency. While existing neuro-computational models propose three hypotheses (local, distributed, hybrid) regarding the neural instantiation of semantic space, the impact of context dependency on these representations remains unexplored. In this MEG study, we used representational similarity analysis (RSA) to query the representations of three types of noun-verb ambiguous stems: (i) stems whose meanings are unambiguous or at least very consistent across their noun and verb uses (e.g., dream), (ii) stems whose noun and verb meanings are related though different, that is, a relation of polysemy (a fly/to fly) and (iii) homonym stems whose noun and verb meaning are unrelated to each other (a spell/ to spell). Stems were categorized into these semantic types using Wordsmyth and Wordnet, an online sense-norming experiment as well as a computational measure of contextual dispersion. In addition to addressing the effect of syntactic category and meaning consistency, we also manipulated the size of the local syntactic context of each target word, resulting in a 3 (semantic type) x 2 (syntactic category) x 3 (syntactic context) design. Finally, all stimuli were presented after a context sentence that disambiguated the target item: for example, a sentence such as “Gary is a flight attendant” would be used to disambiguate “fly” towards flying as opposed to the insect sense. Our goal was to determine the extent to which evidence for shared representations could be uncovered across the various contextual manipulations and what the effect of the different types of context factors would be. We performed univariate analyses of source-localized MEG signals to provide a basic profile of the effects of our three factors and then correlated theoretical models of dissimilarity with single-trial MEG activity to characterize representational similarity across contexts. Univariate results showed consistent, wide-spread activity increases for larger syntactic contexts, indicating a higher compositional load, as well as higher signals in fronto-temporal cortex for verbs than nouns. In the RSA, a relatedness model grouping unambiguous and polysemous words together and distinguishing them from homonyms was significant in occipital regions starting around 100ms and then evolved into a more temporal cluster. This could reflect repetition priming under the assumption that the unambiguous and polysemous items share a stem morpheme across the noun and verb uses whereas the homonyms do not. Thus the different instances of the shared stem across the different contexts would serve to prime each other for unambiguous and polysemous items, but not for homonyms. We also identified a large frontal cluster at 284-558ms that was sensitive to the semantic distance between the noun and verb senses of target items, as assessed by our sense-norming experiment. Overall, our results suggest an early emerging correlate of shared representations for unambiguous and polysemous items and a later general sensitivity to sense distance in frontal cortex.
Topic Areas: Meaning: Lexical Semantics,