Modeling Scalar Categorization: Comparing Threshold-Based and Prototype-Based Representations

Understanding how people categorize along continuous dimensions is central to theories of conceptual representation and language meaning. In the case of gradable adjectives, two major representational accounts have been proposed: prototype- based models (e.g., Gärdenfors and Williams, 2001; Rosch et al., 1976), which rely on similarity to typical instances, and threshold-based models (Hampton, 2007), which posit context-sensitive decision boundaries. This study investigates which of these representations better predicts human categorization behavior, focusing on English-language speakers. Participants (N = 98) completed binary and continuous categorization tasks for four adjective pairs, followed by the elicitation of their internal thresholds and prototypes. Using these elicited values, we simulated model predictions and compared them with behavioral responses using both mean-squared error and likelihood-based analyses. Both models captured participants’ behavior well, but the threshold model consistently provided a better fit, particularly in the continuous task and for most individuals. Bayesian analysis further revealed that, for three of the four adjective pairs, threshold-based representations more accurately predicted categorization behavior for the majority of the population (0.61 < θ < 0.77). At the same time, model preference varied both across and within individuals, suggesting that categorization with gradable adjectives engages flexible representational strategies that may shift depending on context or conceptual domain. These findings highlight the central role of threshold-based reasoning in scalar categorization, while also affirming that prototype-based representationsremain part of the cognitive resources individuals draw on in this domain.