PROJECT SUMMARY/ABSTRACT
Approximately two million people in the U.S. suffer from acquired language disorders, known as
aphasias. These patients can present with strikingly different patterns of semantic language impairments.
For example, they can be more impaired at processing concrete than abstract concepts or vice-versa, and
particular categories of concepts such as animals or artifacts can be differentially affected. One major obstacle
to understanding the mechanisms underlying semantic language disorders is a larger theoretical gap in our
understanding of semantic word processing in the healthy human brain, particularly in terms of how it
relates to perceptual, motor, and memory processes.
The main goal of this project is to develop a generative model of how semantic representations are encoded
in the brain based on interpretable, neurologically plausible dimensions. Our approach is based in part on
human ratings of the relevance of several neurally-grounded features of phenomenal experience to the meaning
of each word, including perceptual, motor, and affective components. Starting with a set of 48 features, we will
identify the subset of features (and their respective weights) that best predicts human performance in
semantic tasks, as well as the similarity structure of neural activity patterns associated with individual word
meanings. Aim 1 will generate empirical norms of pairwise semantic similarity for a set of 640 English nouns
and for a set of 300 English verbs, including both implicit (priming) and explicit (ratings) measures, and will
identify which model features are essential for predicting those similarity measures for each word set. It will
also determine whether taxonomic and distributional information contribute to semantic representation
independently of experiential information. Aim 2 will characterize the similarity structure and information
content of the neural activation patterns encoding the meanings of individual nouns and verbs. Finally, Aim 3
will use resting-state connectivity, at the individual participant level, to identify the cortical subnetworks that
make up the semantic representation system and characterize their information content. The project will result
in a quantitative, feature-based model of word meaning that will provide a theoretically principled approach
to decoding semantic representations directly from brain activity, enabling the development of accurate and
computationally efficient speech prosthetic devices via brain-computer interface.