Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Poster Slams
Neural decoding of grammatical number within and across languages
Poster B11 in Poster Session B and Reception, Thursday, October 6, 6:30 - 8:30 pm EDT, Millennium Hall
Jeonghwa Cho1, Jonathan R. Brennan1; 1University of Michigan
BACKGROUND Grammatical features are an essential part of human language that are used to express numerosity, tense, aspect, and more. Nevertheless, how they are processed in the brain has only recently begun to be investigated (Dunagan et al. 2022; Papageorgiou et al. 2020). A related issue is how these grammatical features are represented across multiple languages in bilingual brains. Multi-voxel pattern analysis (MVPA) can be a useful tool in this domain to investigate common patterns in brain activity when processing the same grammatical feature across languages (cf. Correia et al. 2015 for lexical semantics.) This study uses MVPA to decode EEG data for grammatical number as well as lexical semantics both within and across different languages. METHODS EEG data from six Korean-English bilinguals (all females, age: 18–29) living in the United States are analyzed in the study (32 channels sampled at 500 Hz; filtered from 1–30 Hz after artifact rejection). Experimental stimuli consist of four nouns (dog, rat, swan, and lion) in singular and plural forms and four verbs (lean, own, fill, and chop) in past and present tense in English and Korean, recorded by three different female speakers. Participants listened to the stimuli during EEG recording. Across ten runs we presented an English block a Korean block alternating in order; In each block we played 24 animal noun stimuli (8 words × 3 speakers) and 24 action verb stimuli (8 words × 3 speakers). Participants decided whether the word they heard is a noun or a verb for 10% of the trials (these were excluded from analysis.) For within-language classification, a Support Vector Machine was trained on 80% of the EEG data (0–500 ms after onset for lexical semantics; 300–800 ms after onset for number which is only available word-finally) per participant to classify between different nouns (e.g. dog(s) vs rat(s)) or between singular versus plural nouns, and tested on the 20% of held-out test data. For between-languages classification, the classifier trained on all noun data from one language was tested on the other language. RESULTS The classifier reached an above-chance accuracy (0.5) in decoding individual nouns for all participants in English (range: 0.52–0.63) and for five participants in Korean (range: 0.52–0.66). Cross-linguistic decoding accuracy was above-chance for four participants (range: 0.51–0.55). For grammatical number decoding, accuracy was above-chance for four participants in both English (range: 0.55–0.68) and Korean (range: 0.52–0.61). Across-languages classification of grammatical number showed an above-chance accuracy for four participants (range: 0.51 - 0.53). DISCUSSION The results of lexical concept decoding replicate Correia et al. (2015), suggesting that processing the same lexical concept in multiple languages yields a common pattern of brain activation that supports neural decoding cross-linguistically as well as within language. Crucially, the results from grammatical number decoding indicate that the neural responses to singular versus plural may also be shared to some extent between languages.
Topic Areas: Meaning: Lexical Semantics, Morphology