������

Toward an exemplar-based context-sensitive distributional model of semantic memory

Phillip Hamrick, Ph.D., English, ������

A central question in understanding the neurocognitive underpinnings is how meanings of words are learned, stored, and processed in the mind/brain. One approach to these questions is to examine what can be learned about word meanings based on the statistical patterns of co-occurrence inherent in language and in the environment of the learner. This approach has been implemented in popular computational models, known as distributional models, of semantic memory. These distributional semantic models often operate by converting words to vectors and then computing word co-occurrences over numerous text documents. Although these models have had considerable success accounting for a range of semantic memory phenomena, some of their assumptions are problematic. For example, most models collapse semantic representations into a single composite representation, which is problematic in the case of things like homonymy (e.g., when a word has multiple unrelated meanings, such as bark or bank).

This project takes a novel approach to such problems by modeling semantic memory as an episodic memory system that stores exemplars (i.e., specific episodes of experience) with a retrieval process that averages over episodes. This contrasts with theories that propose strong dissociations between episodic and semantic memory systems. To examine whether such an exemplar-based distributional semantic model is plausible, this study will use large corpora (e.g., 500+ million words of natural language) to train several competing distributional semantic models, along with the exemplar-based model. The models’ “behavior” will then be probed in a variety of ways and compared with those of human participants in previously published studies of semantic cognition.

The outcomes of this study will be to (i) examine the validity of an exemplar-based account as a plausible account of semantic memory, and (ii) identify both unique predictions from the models as well as weaknesses in the available models that can be subsequently used for further research at the computational, behavioral, and possibly even neuroimaging levels.