Poster C129, Sunday, March 25, 1:00-3:00 pm, Exhibit Hall C
Discrimination and Prediction of Concreteness from Neuroimaging and Corpus Data
Dominick DiMercurio1, Chaleece Sandberg1; 1Pennsylvania State University
The mechanisms underlying the semantic system remain a mystery in cognitive neuroscience. Particularly, concreteness is a remarkable phenomenon in semantics due to suggestions from previous literature that abstract and concrete words might be represented distinctly in the brain. Methods in neuroimaging, behavior, and corpora can be employed to probe concreteness and related effects; specifically, discrimination and prediction techniques with machine learning are rapidly emerging for doing so. The present study utilized neuroimaging data from participants who performed a concreteness judgment task while undergoing fMRI. The neuroimaging data, in conjunction with document or dependency corpus data in the form of word-specific semantic vectors, were analyzed with support vector machines (SVMs) and semantic space models (SSMs) to determine: i.) the performance of word type discrimination between neuroimaging and corpus data, and ii.) whether an interaction between type of semantic vectors and type of word exists to support the hypothesis of distinct representations. The performance of SVMs trained with neuroimaging data exceeds that of SVMs trained with corpus data; meanwhile, the comparison of SSMs defined with either documents or dependencies failed to reproduce better-than-chance performance with the neuroimaging data. The latter limits any conclusion that can be made regarding the existence of distinct representations for abstract and concrete words. These findings highlight the challenges in using corpus data to discriminate concreteness and the importance of task selection for the creation of reliable semantic space models. Future research will investigate how to improve these techniques and how to incorporate real or simulated clinical data.
Topic Area: LANGUAGE: Semantic