A Comparative Study of Experiential and Distributional Representations for Concrete and Abstract Concepts

We investigate the semantic distinction between concrete and abstract concepts byexamining the relative contribution of experiential (grounded) features and distributional(language-based) features in modeling their representation via data driven analysis usingSpearman’s correlation between human similarity ratings, representational similarityanalysis between various feature spaces, multidimensional scaling, and category similarity.Our results reveal a robust double dissociation between feature types and concept classes.Grounded features better align with human judgements for concrete concepts than forabstract concepts. Concrete concepts exhibit higher representational similarity acrossdifferent independently obtained grounded feature representations. In contrast, languagebased features better align with human judgements for abstract concepts than for concreteconcepts. Abstract concepts exhibit higher representational similarity across differentindependently obtained distributional feature representations. Multidimensional scalinganalyses further show that abstract concepts occupy a compressed and highly clusteredregion in experiential space, with clustering driven more by degree of abstractness thansemantic similarity. Concrete concepts in contrast are highly correlated with sensorimotorfeatures in experiential space. Further, better category similarity for concrete concepts isexhibited by experiential features rather than distributional features.