Can we ground nonwords? A first data-driven insight on the potential sensorimotor features of novel words.

Nonwords have long been considered devoid of any meaning. However, recent studies have called this assumption into question, demonstrating that they can indeed elicit activation in semantic memory. Despite their inherently symbolic nature, here we question whether nonwords can also evoke experiential features across various sensorimotor modalities. To investigate this, we trained a model aimed at predicting the sensorimotor patterns elicited by nonwords on the basis of their estimated semantic activation. Such mapping model was trained on a set of 39,707 English words, for which semantic vector representations were obtained using fastText, a Distributional Semantic Model that can also be used to approximate the meaning of nonwords via their sub-lexical units. Sensorimotor norms were obtained from the Lancaster Sensorimotor Norms (Lynott et al., 2019). Then, eleven separate ridge regression models were trained to predict individual ratings across six perceptual modalities (touch, hearing, smell, taste, vision, and interoception) and five action effectors (mouth/throat, hand/arm, foot/leg, head, and torso), on the basis of fastText embeddings (encoding the distributional history of both the whole word and its sub-lexical units). Sensorimotor ratings for 27,136 nonwords were hence predicted by feeding the model their corresponding fastText embeddings (encoding the distributional history of their sub-lexical units). Moving from these predictions, perceptual strength measures (i.e., Maximum perceptual strength and Minkowski 3 distance) were computed for each nonword and considered as predictors of reaction times to the nonwords in a lexical decision task extracted from the British Lexicon Project (Keuleers et al., 2012). Both measures significantly predicted responses, with Minkowski3 outperforming Maximum perceptual strength. Nonwords with higher estimated sensorimotor strength take longer to be rejected. This suggests that nonwords can be associated to experiential traces via the distributional properties of their sub-lexical units. We provide a first data-driven insight into the possible sensorimotor features of nonwords. Our findings challenge the traditional assumption whereby nonwords lack semantic content, suggesting instead that they can evoke experiential features. This highlights the importance of further investigating the grounded nature of nonwords, in order to understand the information humans would rely on when assigning perceptual features to novel words.