Poster D2, Wednesday, August 21, 2019, 5:15 – 7:00 pm, Restaurant Hall
Modelling a full-size grounded conceptual system: Categorical structure emerges spontaneously from the latent structure of sensorimotor experience
Louise Connell1, James Brand2, James Carney3, Marc Brysbaert4, Dermot Lynott1;1Lancaster University, 2University of Canterbury, 3Brunel University, 4Ghent University
Many theories of semantic memory assume that categories spontaneously emerge from commonalities in the way we perceive and interact with the world around us. However, efforts to test this assumption computationally have been hampered by a number of issues, including the use of abstracted features without clear sensorimotor grounding and over-reliance on small samples of concepts from a limited number of categories. As such, even though theories of emergent category structure may assume grounded conceptual representations, current models do not adequately instantiate or test this assumption. In the present work, we take a radically different approach by creating a fully-grounded, multidimensional sensorimotor model at the scale of a full-size human conceptual system and examining whether categorical structure emerges from the latent structure of sensorimotor experience. The data underlying the model come from our newly developed set of sensorimotor strength norms, where each of 40,000 English lemmas is rated on the extent to which the referent concept is experienced via 11 separate sensorimotor dimensions: six perceptual modalities (auditory, gustatory, haptic, interoceptive, olfactory, visual) and five action effectors (foot/leg, hand/arm, head excluding mouth, mouth/throat, torso). Each concept is therefore represented as a single point within 11-dimension sensorimotor space, where similar concepts are located close together. To model latent structure within this space, we ordered the concepts by age of acquisition (AoA) to provide a developmentally plausible trajectory of cluster formation, and then used two-step cluster analysis (preclustering ordered by AoA, followed by agglomerative hierarchical clustering) to extract the optimal cluster solution. We found evidence for (a) a high-level separation of abstract and concrete categories (that was not enhanced by the inclusion of affective information); (b) a hierarchical structure of concrete concepts that separated categories commonly impaired in double dissociations, such as fruit/vegetables, animals, tools, and musical instruments; and (c) a flatter hierarchy of abstract concepts that separated categories such as negative emotions, units of time, social relationships, and political systems. These findings demonstrate that sophisticated categorical structure can emerge spontaneously from grounded sensorimotor representations alone, without positing high-level, abstracted features. Moreover, the findings support theoretical claims that sensorimotor information is fundamental to the representation of all conceptual knowledge, including abstract domains where it has traditionally been assumed to play a minimal role.
Themes: Computational Approaches, Meaning: Lexical Semantics
Method: Computational Modeling