Name: Robert T. Ashwill

School of Psychology – Ph.D. Dissertation Defense Meeting

Date: Friday, September 12, 2025

Time: 10:00 A.M.-12:00 P.M.

Location:  J.S. Coon Room 148 (virtual option available)

Thesis Chair/Advisor:

Rick Thomas, Ph.D. (Georgia Tech)

Thesis Committee Members:

Rick Thomas, Ph.D. (Georgia Tech)

Elizabeth Whitaker, Ph.D. (Georgia Tech Research Institute) 

Scott Moffat, Ph.D. (Georgia Tech)

James Roberts, Ph.D. (Georgia Tech)

Sashank Varma, Ph.D. (Georgia Tech)

Title: Exploring Semantic Grounding in Word Embeddings

Abstract: Current grounded theories of human cognition propose that concepts develop through both grounded sensorimotor experiences and language use. Dual coding theories posit that concepts are initially rooted in modal systems corresponding to the sense in which they are experienced. Language then allows associations between concepts beyond direct sensorimotor experiences, encoding relationships linguistically through which are reflected in language statistics and captured by distributional semantic models (DSMs).

A basic semantic distinction among words is word concreteness: concrete words directly reference physical entities experienced through our senses, while abstract words lack such direct physical referents. Examining how concreteness is reflected in sensorimotor versus linguistic representations can shed light on how these sources of information are integrated within human cognition.

This dissertation investigates how linguistic and sensorimotor dimensions uniquely and jointly encode semantic information relevant to word concreteness. Correlational analyses indicate meaningful relationships between representations derived from language statistics, sensorimotor dimensions, and word concreteness. Mediation analyses identified unique and overlapping contributions from linguistic and grounded information in predicting concreteness, suggesting both sources play key roles in representing word concreteness. Additionally, representational similarity across DSMs indicates consistent grounded content across several model architectures.