Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Readers Extract Some Grammatical Information In A Single Fixation, Across Sentence Structures
Poster Session C, Friday, October 25, 4:30 - 6:00 pm, Great Hall 3 and 4
This poster is part of the Sandbox Series.
Dustin A. Chacón1, Donald G. Dunagan2, Tyson Jordan2; 1UC Santa Cruz, 2University of Georgia
The visual system can extract detailed information extremely rapidly (e.g., Huang & Staub 2024). Short sentences seen for 200ms ('the man can run') are recalled more accurately than ungrammatical, scrambled counterparts ('run can man the') (Snell & Grainger 2017); and evoke greater activity in language regions 200–400ms (EEG: Wen et al. 2019; Dunagan et al. 2024; MEG: Dufau et al. 2024; Flower & Pylkkänen 2024). This 'Sentence-Superiority Effect' (SSE) may diagnose rapid, parallel language processing. But, these experiments involve many trials, often comparing a small number of sentence types against many ungrammatical contrasts. Our question: Does the SSE depend on habituation to specific visual stimuli (i.e., 'the noun aux verb') in the experimental context? Or, does the SSE arise even if participants don't know which structures to expect? [METHODS] [Materials] We built a context-free grammar (CFG) that generated 17 5-word sentence types, with a vocabulary of 10 words (3–6 letters) per lexical category. We generated every lexicalization that did not repeat words, then generated every 2-word transposition and reversal of each sentence, excluding permutations that resulted in grammatical sentences. From this distribution (~950 million stimuli), we sampled without replacement 170 grammatical (10/type), 147 transposed, and 160 reversed sentences per subject. [Procedure] Sentences were displayed for 300ms, centered on a screen, followed by a 500ms blank screen, then followed by a probe sentence. Participants judged whether the sentences matched. EEG data were recorded with a 128ch BrainVision actiChamp+ system. [RESULTS] [Behavioral] Participants were more accurate for Grammatical vs. Reversed trials (p < 0.001; 85.4% vs. 81.0%), but not for Grammatical vs. Transposed trials (p = 0.19; 85.4% vs. 83.3%). [EEG] Spatio-temporal cluster-based permutation tests on sensor data (800ms epochs), comparing Grammatical vs. Reversed and Grammatical vs. Reversed vs. Transposed did not yield any clusters. We then conducted RSA analyses, constructing representational dissimilarity matrices (RDMs) for surprisal values and syntactic categories of each word, frequency and semantic association for each bigram, and internal parser states of a bottom-up chart parser using the CFG, quantifying the length of each phrase in the most complete analysis. We conducted an RSA searchlight analysis (50ms, 30mm) for each predictor for each subject, then conducted group-level t-tests to determine which correlations differed from 0, in a broad 'SSE' time window (100–500ms). Significant correlations were identified between the RDM for word 2-word 3 semantic association (p = 0.02), 100–217ms; word 2 surprisal (p < 0.05), 190–500ms in right parietal sensors and word 3 surprisal (p = 0.03), 320–500ms in frontal sensors; and VP length, 120–388ms in posterior sensors. Marginal correlations were observed for syntactic categories of word 1, 227–500ms (p=0.11), and word 2, 187–419ms (p=0.08), in right lateral sensors, and object NP length, 128–346ms (p = 0.08) in posterior sensors. [CONCLUSION] With only 300ms presentation time, readers' brain activity suggests they extract semantic and syntactic features of words 1–3, which likely fall near the foveal center. We provide further evidence of rapid activation of grammatical information in short fixations, with unanticipated sentence structures, and without a neural 'SSE'.
Topic Areas: Syntax and Combinatorial Semantics,