Redundancy in Perceptual and Linguistic Experience: Comparing Feature-Based and Distributional Models of Semantic Representation

摘要: Since their inception, distributional models of semantics have been criticized as inadequate cognitive theories human semantic learning and representation. A principal challenge is that the representations derived by are purely symbolic not grounded in perception action; this has led many to favor feature-based We argue amount perceptual other information can be learned from statistics underappreciated. compare three nine using a clustering task. Several demonstrated comparable with clustering-based on representations. Furthermore, when trained child-directed speech, same perform well sensorimotor-based feature children's lexical knowledge. These results suggest that, large extent, relevant for extracting categories redundantly coded linguistic experience. Detailed analyses clusters also reveal make use complementary cues organization two data streams. Rather than conceptualizing competing theories, we future focus should understanding mechanisms humans integrate sources.