Automated generation of psycholinguistic stimuli
Tags:
Experiments in psycholinguistics and the cognitive neuroscience of language rely on linguistic stimuli (sentences) with either specific linguistic properties or which target specific cognitive processes. Such stimuli are generally assembled using manual or semi-manual methods, limiting their quality, quantity, and diversity. We propose GOLI - a gradient-based optimization method that transforms a random sentence into a novel linguistic stimulus which fulfills an experimenter’s goal. We apply GOLI to two deliberately different tasks—creating minimal pairs of counterfactual sentences, and deriving constrained stimuli that predict specific responses in human brain regions. We demonstrate how GOLI supports diverse experiment goals and efficiently generates stimuli that are not subject to experimenter biases which may arise from manual methods.
ACL 2023 - [Reviews] [Rebuttals]