Learning to Categorize in the Context of Item Triples

We evaluate the relative contributions of two mechanisms of category learning: 1) abstraction across examples of the same category and; 2) differentiation between examples of different categories. A novel "triples" paradigm is introduced in which each classification target is presented with two different context items. Learners are informed of the structure of the triples so they may take advantage of knowledge about the relative category status of the items. We use feature-based categories with perceptually subtle variation among examples. The study is designed to advance a naturalistic yet controlled basis for the study of category learning by using multiply-instantiated feature values (Markman & Maddox, 2003) and three-way rather than binary classification decisions. In a control condition, items were presented one at a time. Learning was also tested under five experimental conditions based on the following triple structures: aAA (both context items match the category of the target 'a'), aAB (one matching and one mismatching context item), aBB (both context items mismatch the target, but the context items match one another), aBC (both context items mismatch the target and the context items also mismatch one another), and aXX (no systematic structure). One possible learning strategy is to locate common features between items known to belong to the same category and perform abstraction -in which case the aAA group should have an advantage. Another potential strategy is to identify contrasts between items known to belong to all three categories -in which case the aBC structure should be most beneficial. The aAB structure is least informative under either strategy because the learner is unable to know for certain whether any pair within the triple are in the same or different categories. The aBB group benefits from weaker forms of both abstraction and differentiation on each trial.