Sensorimotor Norms in Word Embeddings: SENSE Model Predicts Human Associations

A paper published on arXiv presents SENSE (Sensorimotor Embedding Norm Scoring Engine), a model designed to predict Lancaster sensorimotor norms based on word lexical embeddings. The research involved a behavioral experiment with 281 participants who chose nonce words that triggered particular sensorimotor associations. Notable correlations were observed between the selection rates of participants and SENSE ratings in 6 out of 11 modalities. An analysis of sublexical elements indicated consistent phonosthemic patterns related to the interoceptive norm, offering a computational approach to suggest potential phonosthemes derived from text. This study connects the fields of computational linguistics and cognitive science, linking abstract representations of words to sensory and motor experiences.