Taking Development Seriously: Modeling the Interactions in the Emergence of Different Word Learning Biases.

Taking Development Seriously: Modeling the Interactions in the Emergence of Different Word Learning Biases Savannah M. Schilling (Savannah.Schilling@Colorado.Edu) Department of Electrical, Computer, & Energy Engineering, 425 UCB Boulder, CO 80309-0425 USA Clare E. Sims (Clare.Holtpatrick@Colorado.Edu) Department of Psychology & Neuroscience, 345 UCB Boulder, CO 80309-0345 USA Eliana Colunga (Eliana.Colunga@Colorado.Edu) Department of Psychology & Neuroscience, 345 UCB Boulder, CO 80309-0345 USA Abstract knowledge? Our results indicate that different biases do interact, and the emergence of one bias can either strengthen or weaken other biases depending on the strength of cues provided in the learning context . These results allow us to make predictions about the timing of children’s word learning and generalization as biases emerge. Development is about change over time. Computational models have provided insights into the developmental changes seen in different cognitive phenomena, including within the domain of word learning. The present paper uses a computational model to investigate the interdependencies between the emergence of different word learning biases. This model allows investigation of how the emergence of the shape bias influences novel noun generalization to two other types of items. The results suggest that the emerging shape bias for solids can either strengthen or weaken other types of biases depending on the strength of the cues to solidity or non- solidity; further, these results make predictions about children’s biased word learning over time. Word Learning Biases Keywords: Computational models; neural networks; trajectories; word learning; shape bias. Introduction Computational models have proven to be an important tool for investigating many issues within cognitive development (e.g., Munakata & McClelland, 2003). Such models can provide insights about the mechanisms that underlie learning patterns seen across childhood. In the domain of word learning, various models have been used to investigate fast mapping and the taxonomic shift (Mayor & Plunkett, 2010), variability effects in learning phonetically similar words (Apfelbaum & McMurray, 2011), task effects in novel noun generalization (Samuelson, Schutte, & Horst, 2009), and word learning at different levels of abstraction (Xu & Tenenbaum, 2007). In this paper we focus on using connectionist models to examine developmental trajectories in word learning. This approach has the potential to guide novel and testable predictions about children’s language development. This paper employs a computational modeling approach to investigate the emergence of word learning biases that support early language acquisition. This approach allows us to analyze in detail how different word learning biases interact and influence one another over the course of word learning. For example, does a later emerging bias build onto and benefit from an earlier bias, or is there a period of conflict as new knowledge is assimilated with prior One of the reasons that children are such skilled language learners is because of biases. In the context of word learning, biases are constraints on the range of things that children will consider in deciding what a new word refers to. Rather than assuming that any word can be used to label any item, children exhibit principled patterns of behavior in the ways in which they learn words. The main constraint we will focus on in this paper is found within the domain of noun learning: the shape bias. The shape bias refers to young children’s tendency to generalize newly learned nouns to other objects based on similarity in shape (Landau, Smith, & Jones, 1988). That is, if a child is taught a novel name for a novel solid object, he or she will extend that name to other objects that match the original in shape, even if that shape match differs in texture, color, or size. Children show a reliable shape bias by 2 years of age (Samuelson & Smith, 1999). A related phenomenon in noun learning is the material bias. While the shape bias is seen in children’s generalization of labels to solid objects, the material bias concerns the labeling of non-solid substances. The material bias has been found using the same novel noun generalization (NNG) paradigm typically used in studies of the shape bias. Children taught a novel name for a novel non-solid substance tend to generalize that name to other non-solids that match the original in material rather than to non-solids matching in features like shape and size but made out of a different material (e.g., Soja, 1992; Soja, Carey, & Spelke, 1991). The material bias is typically seen slightly later than the shape bias, at 3 years of age (Yoshida & Smith, 2005). Altogether, the evidence suggests that over the first years of life children develop preferential attention to different