Early research on categorisation suggested that verbalizable and nonverbalizable category-learning are qualitativelydifferent. Toward this end, the implementational-level model (COVIS–COmpetition between Verbal and Implicit Systems) ofcategorisation assumes that category-learning involves separate but parallel sub-systems. Specifically, for verbalizable tasksabstract category-labels are learned by a hypothesis-testing sub-system, while for nonvertbalizable tasks response position islearned by a procedural-learning sub-system. However, recent research has revealed that: 1) regardless of category structure,reversal learning is easier than learning novel categories; 2) qualitative difference between verbalizable and nonverbalizabletasks disappears when automaticity has developed; and 3) control of automatic categorisation is different from both proposedsub-systems. These challenges suggest a fundamental revision of the mechanisms of categorisation. Contrary to the separate,parallel-processing sub-systems theory, we argue that categorisation involves hierarchical-processing sub-systems of response-production and category-label association. This framework, when combined with Supervisory Attentional System theory, mayfacilitate the unification of human categorisation.

perceptual symbol systems hypothesis describes how semantic knowledge is grounded insensorimotor experience. According to the theory, knowledge is acquired through sensorimotor simulations. This challengesthe classical view supported by the disembodied cognition hypothesis, which generally favours an abstract and symbolic sys-tem. We propose a unified perspective, in which, the embodied cognition hypothesis, with a particular focus on the semanticdomain, is provided with a mechanistically tractable computational framework based on the parallel distributed processing(PDP) paradigm. A critical difference between the current approach and previous mechanistic accounts of embodied cognitionis that the current approach avoids using hand-coded representations and instead, relies on an agent-based simulation withenvironmental interaction for the creation of situated inputs and outputs, supplemented with supervised and unsupervised deeplearning mechanisms, from which semantic cognition emerges.