Collaborative Research: CompCog: Modeling Search within the Mental Lexicon

Bowdoin College, Brunswick ME

Investigators

Abstract

Searching for the right word in our mental dictionary - sometimes called our mental lexicon - is something that humans do frequently and mostly with great ease. The current research is designed to better understand the mechanisms that underlie memory search for words, which helps us better understand how words are organized and retrieved in the mind. Furthermore, failures of this search process, which show up as word-finding difficulties or having words on the tip of your tongue, can be symptomatic of certain clinical disorders, and a stronger theory of how this kind of memory search works can inform our understanding of why it sometimes fails. The proposed research leverages recent advances in machine learning, network science, and gamification. Word search is conceptualized within a memory system that includes both information about a word's meaning (e.g. that a dog is a type of animal) and information about a word's sounds (e.g., that dog and dot start with the same sounds). This research informs the use and interpretation of clinical measures of memory search as well as promote participation from underrepresented identities in STEM generally and cognitive science specifically. This collaborative proposal brings together researchers in psycholinguistics, network science, and language modeling to develop a unified computational framework to study lexical search and retrieval using a task that is widely used in both basic science and clinical settings: the fluency task. In this task, participants are asked to name as many words as they can in a certain amount of time that all fit with a category (e.g. types of animals). The proposed work tests the central hypothesis that memory search unfolds via complex interactions between different sources of information as well as individual variation based on our unique experiences in the world. Our approach involves designing a gamified, web-based fluency experiment to generate a large behavioral dataset of verbal fluency and using novel computational approaches based on two fundamental search mechanisms (optimal foraging and random walks) to account for which words are produced in what order. With this approach, we will evaluate the interaction of meaning and sound information, and account for the variance observed in how different individuals perform on these tasks. The proposed gamified web-based fluency task, combined with theory-driven, bottom-up computational models of knowledge representation and memory search have the potential to generate an unprecedented amount of fluency data, uncover novel insights about memory search, and open up new lines of inquiry that cut across fields of psychology and linguistics, thereby leading to a more robust and unified cognitive theory of memory search. This project is jointly funded by the Perception, Action and Cognition program and the Established Program to Stimulate Competitive Research (EPSCoR). This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.