Is argument structure processing syntactic or semantic in nature? Evidence from computational modeling on naturalistic fMRI

Background: A long-standing theoretical debate exists in linguistics concerning argument structure processing, with separationism focusing on syntactic structure and projectionism on semantic properties[1]. To investigate whether argument structure processing is primarily influenced by syntactic structure or semantic properties, this study employed integrative neurocomputational modeling[2, 3] to link brain functions with explicitly defined computational models. Method: We analyzed naturalistic functional magnetic resonance imaging (fMRI) data from participants listening to a story[4], with a focus on subject noun phrase + verb chunks. The methodological framework integrated a general linear model (GLM) analysis of the fMRI data with computational modeling using natural language processing algorithms. These components were integrated using representational similarity analysis (RSA), allowing us to assess the relatedness of two symbolic computational models—one relying on syntactic information[5] from parse trees and the other based on semantic selectional preference information of verbs[6]—to brain activities. Results: The GLM analysis identified significant neural correlates of argument structure processing largely consistent with previous findings, including the precuneus, the right superior temporal gyrus, and the right middle temporal gyrus. Some deviations from previous studies likely reflect the naturalistic nature of the stimuli and our contrast design. The RSA results favored the model utilizing semantic information—a finding further supported by effects observed in brain regions associated with argument structure processing in the literature and by an additional RSA comparing constructions with varying levels of transitivity. Discussion: These findings suggest that during naturalistic story listening, humans rely heavily on semantic information to interpret argument structure. This study demonstrates an alternative method to engage with the debate on argument structure, highlighting a collaborative effort between theoretical neuroscientific, and computational linguistics.