Learning individual words and learning about words simultaneously

Learning individual words and learning about words simultaneously Sylvia Yuan (shyuan@berkeley.edu) Department of Psychology, University of California, Berkeley Berkeley, CA 94720 USA Amy Perfors (amy.perfors@adelaide.edu.au) School of Psychology, The University of Adelaide Adelaide, SA 5005 Australia Josh Tenenbaum (jbt@mit.edu) Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology Cambridge, MA 02139 USA Fei Xu (fei_xu@berkeley.edu) Department of Psychology, University of California, Berkeley Berkeley, CA 94720 USA Abstract Children are guided by constraints and biases in word learning. In the case of the shape bias—the tendency to extend count nouns by similarity in shape—previous findings have revealed that learning plays an important role in its development (e.g., Smith et al., 2002). Some have proposed that children acquire inductive constraints like the shape bias by making inferences about observed data on multiple levels of abstraction (e.g., Smith et al. 2002; Kemp et al., 2007). The current study provides support for this hypothesis by demonstrating that preschoolers can rapidly and flexibly form overhypotheses about the role of arbitrary features, not just shape, in determining how words refer to object categories. This work suggests that when learning individual words, children are also learning about words simultaneously. Children’s ability to “learn to learn” may have implications for the origins of learning biases in many different cognitive domains, not just in language learning. Keywords: word learning; shape bias; inductive constraints; overhypothesis; rational inference Introduction Much of what we know about the world depends on inductive learning—inferring an underlying general principle based on limited data. Induction is not a trivial problem: in principle, there is an infinite set of possible generalizations that can be made from the same observed examples (Quine, 1960). For example, in the domain of word learning, from hearing the word ‘dog’ while observing a situation involving the presence of a dog, a learner could hypothesize that the word refers to fur, cute, tail, dog that is alive, or undetached dog parts, among a potentially infinite set of possible meanings. Learning must therefore be constrained by biases of some sort (Keil, 1981; Markman, 1989). Children rely on inductive constraints in many cognitive domains, such as word learning (e.g., Landau, Smith, & Jones, 1998; Markman, 1989) and physical and psychological reasoning (Baillargeon, 2008; Carey & Spelke, 1996; Gergely & Csibra, 2003). Given the early appearance of these constraints, it seems conceivable that some may be innately given. It is possible, however, that some of the early cognitive biases might be learned. For example, 1.5-year- olds can be trained to exhibit a shape bias in word learning, extending a newly-learned label to a similarly-shaped object (e.g., Smith et al., 2002; see also Samuelson, 2002). The acquisition of inductive biases continues through childhood and adulthood and takes place not only in the domain of word learning. Based on only a few examples, toddlers and preschoolers can learn the dimension used in classification and apply this knowledge to sort new objects into new categories (Macario, Shipley, & Billman, 1990; Mareschal & Tan, 2007). From observing causal relations of an initial set of objects, preschoolers and adults form abstract causal schemata and use them to make inferences about the behaviors of newly-encountered objects (Kemp, Goodman, & Tenenbaum, 2010; Lucas, Gopnik, & Griffiths, 2010). In all of these cases, learners rapidly acquire abstract knowledge of some form that helps them readily learn about novel items or situations based on sparse data. How do learners acquire knowledge that guides subsequent learning? How do they “learn to learn”? A constraint on learning, whether in the form of perceptual biases (e.g., shape bias) or principles or systems of relations (e.g., causal schemata), is a type of abstract knowledge specifying how things work in general, going beyond individual instances. Having such a constraint thus requires learners to represent knowledge on multiple levels; the constraint itself is a form of higher-level knowledge, or overhypothesis (Goodman, 1955). In the case of the shape bias, the learner may have some lower-level knowledge that objects labeled as ‘ball’ are all spherical, but do not seem to all be white or plastic; this first-order knowledge is about each individual category. Having the shape bias means that the learner also has some higher-level knowledge that objects that share the same name share the same shape, but

[1]  Willard Van Orman Quine,et al.  Word and Object , 1960 .

[2]  F. Keil Constraints on knowledge and cognitive development. , 1981 .

[3]  Linda B. Smith,et al.  The importance of shape in early lexical learning , 1988 .

[4]  E. Markman Categorization and naming in children , 1989 .

[5]  D. Billman,et al.  Induction from a single instance: formation of a novel category. , 1990, Journal of experimental child psychology.

[6]  E. Spelke,et al.  Science and Core Knowledge , 1996, Philosophy of Science.

[7]  Larissa K. Samuelson,et al.  Statistical regularities in vocabulary guide language acquisition in connectionist models and 15-20-month-olds. , 2002, Developmental psychology.

[8]  Linda B. Smith,et al.  Object name Learning Provides On-the-Job Training for Attention , 2002, Psychological science.

[9]  G. Csibra,et al.  Teleological reasoning in infancy: the naı̈ve theory of rational action , 2003, Trends in Cognitive Sciences.

[10]  Linda B. Smith,et al.  From the lexicon to expectations about kinds: a role for associative learning. , 2005, Psychological review.

[11]  J. Tenenbaum,et al.  Sensitivity to Sampling in Bayesian Word Learning We Thank Members of the Ubc Baby Cognition Lab for Their Help with Data Collection, And , 2022 .

[12]  J. Tenenbaum,et al.  Bayesian Special Section Learning Overhypotheses with Hierarchical Bayesian Models , 2022 .

[13]  D. Mareschal,et al.  Flexible and context-dependent categorization by eighteen-month-olds. , 2007, Child development.

[14]  R. Baillargeon Innate Ideas Revisited: For a Principle of Persistence in Infants' Physical Reasoning , 2008, Perspectives on psychological science : a journal of the Association for Psychological Science.

[15]  Fei Xu,et al.  Induction, overhypotheses, and the shape bias: Some arguments and evidence for rational constructivism , 2009 .

[16]  J. Tenenbaum,et al.  Learning to learn categories , 2009 .

[17]  Fei Xu,et al.  Induction, Overhypothesis, and the Origin of Abstract Knowledge , 2010, Psychological science.

[18]  Joshua B. Tenenbaum,et al.  Learning to Learn Causal Models , 2010, Cogn. Sci..

[19]  J. Tenenbaum,et al.  A tutorial introduction to Bayesian models of cognitive development , 2011, Cognition.