Comparison of Semantic Space Models for Neuroimaging with Abstract and Concrete Words

Open Access
Dimercurio, Dominick
Graduate Program:
Master of Science
Document Type:
Master Thesis
Date of Defense:
June 12, 2017
Committee Members:
  • Chaleece Sandberg, Thesis Advisor
  • Ping Li, Committee Member
  • Diane Williams, Committee Member
  • semantic space models
  • semantic processing
  • neuroimaging
  • concreteness effect
  • abstract words
  • concrete words
  • concreteness
  • Different Representational Frameworks hypothesis
  • computational modeling
  • neurolinguistics
  • neuroscience
  • psycholinguistics
  • predictive models
Psycholinguists have long noted a distinction between abstract and concrete words, especially in measures of performance thereof in lexical tasks. In addition, neuroscientists discovered differences in the localization of abstract and concrete word processing and in behavioral responses regarding abstract and concrete words from people with brain injury or impairment. Various psycho- and neurolinguistic theories have been put forward to explain these differences, including the Different Representational Frameworks (DRF) hypothesis, which states that abstract words are represented more paradigmatically and concrete words are represented more syntagmatically. Meanwhile, computational models are becoming increasingly attractive in neuroscience; in particular, previous researchers (Mitchell et al., 2008) have investigated the ability of a type of semantic space model to predict neuroimaging data from corpus data. The present study expands the previous research by applying the methods to both abstract and concrete words within different semantic space models in a 2 × 2 design of word type and model type to test the DRF hypothesis. Additionally, a third model type inspired by latent semantic indexing was developed and included in the analysis. The present study found that 3 out of 3 model types predicted stimuli significantly better than chance with abstract words, and that 2 out of 3 model types predicted stimuli significantly worse than chance with concrete words; however, these prediction accuracies deviated only slightly from chance. Furthermore, analysis of variance (ANOVA) revealed no significant effect of word type, model type, or interaction between the two. The present thesis discusses possible limitations to the methodology, considerations of deviations from previous research, and suggested trajectories for future study. Improvements to the models are highly desired as semantic space modeling represents a promising avenue of research for improving neuroscientific understanding, language-processing technologies, and speech therapies.