Fast mapping

Abstract: By the time a child has learned a new word, he or she has gained many distinct kinds of information. To take a hypothetical example, consider the word "wolf" being learned by a child who already has a modest animal vocabulary. She must make a new lexical entry: she must note that "wolf" is an English word. She must learn its syntactic subcategorization, namely that it is a common noun. She must relate it to other English words, to its supernyms (such as "animal") and hyponyms (such as "Siberian wolf") and other words in the same lexical domain. She must also learn what "wolf" refers to_ And she must restructure the conceptual domain of animals, at least with respect to how they are named. Suppose, for example, that wolves were previously called "dog." Then learning a new word may be the occasion for learning a new concept, for differentiating dogs from wolves. At the very least, it is the occasion for learning that wolves have a different name from dogs. Clearly, then, learning even a single new word involves learning a great deal of information. In the past, word learning has been studied in several different ways-there have been vocabulary counts, diary studies, cross-sectional and longitudinal studies of the acquisition of organized lexical domains such as size, color or quantity. But in all of these paradigms, there is no control over the input to the child. It is not known how much exposure he or she has had to any given word, nor in what contexts. Thus, while much can be learned about word learning from these procedures, they are not ideal for achieving an understanding of the process itself. Surprisingly, the obvious technique of teaching children an unknown word has been little used (see Carey, 1978 for a review). None of these was designed to mimic the circumstances in which children naturally encounter new words and none was designed to probe for partial acquisition along the way. The present study is an attempt to fill this gap. The experimental procedures that we ultimately designed were intended to provide a strong test of the child's word learning skills. For us, that meant several things. First, we wanted to present the new word in a situation that would approximate a child's everyday word-learning experience at its most casual and undirected level. This meant using a situation in which there was not …

Abstract: Word learning as Bayesian inference Joshua B. Tenenbaum Department of Psychology Stanford University jbt@psych.stanford.edu Abstract We apply a computational theory of concept learning based on Bayesian inference (Tenenbaum, 1999) to the problem of learning words from examples. The theory provides a framework for understanding how people can generalize meaningfully from just one or a few positive examples of a novel word, without assuming that words are mutually exclusive or map only onto basic-level cat- egories. We also describe experiments with adults and children designed to evaluate the model. Introduction Learning even the simplest names for object categories presents a diAEcult inference problem (Quine, 1960). Given a typical example of the word \dog , e.g. Rover, a black labrador, the possible inferences a learner might make about the extension of \dog are endless: all (and only) dogs, all mammals, all animals, all labradors, all black labradors, all black things, all running things, this individual animal (Rover), all dogs plus the Lone Ranger's horse, and so on. Yet, even children under ve can often infer the approximate extension of words like \dog given only a few relevant examples of how they can be used, and no systematic evidence of how words are not to be used (Carey, 1978; Markman, 1989; Regier, 1996). How do they do it? One inuential proposal has been that people come to the task of word learning equipped with strong prior knowledge about the kinds of viable word meanings (Carey, 1978; Clark, 1987; Markman, 1989), allowing them to rule out a priori the many logically possible but unnatural extensions of a word. For learning nouns, one of the most basic constraints is the taxonomic as- sumption , that new words refer to taxonomic classes, typically in a tree-structured hierarchy of natural kind categories (Markman, 1989). Given the one example of \dog above, the taxonomic assumption would rule out the subsets of all black things, all running things, and all dogs plus the Lone Ranger's horse, but would still leave a great deal of ambiguity as to the appropriate level of generalization in the taxonomic tree that in- cludes labradors, dogs, mammals, animals, and so on. Other, stronger constraints try to reduce this ambiguity, at the cost of dramatically oversimplifying the possible meanings of words. Under the mutual exclusivity con- straint, the learner assumes that there is only one word that applies to each object (Markman, 1989). This helps to circumvent the problem of learning without negative evidence, by allowing the inference that each positive ex- ample of one word is a negative example of every other Fei Xu Department of Psychology Northeastern University fxu@neu.edu word. Having heard Sox called \cat as well as Rover called \dog , we can rule out any subset including both Rover and Sox (e.g. mammals, animals) as the exten- sion of \dog . But some uncertainty in how far to gen- eralize always remains: does \dog refer to all dogs, all labradors, all black labradors, or just Rover himself? Inspired by the work of Rosch et al. (1976), Markman (1989) suggested the even stronger assumption that a new word maps not to just any level in a taxonomy, but to an intermediate or basic level. Basic-level cate- gories are intermediate nodes in a taxonomic tree that maximize many dierent indices of category utility and are widely recognized throughout a culture (Rosch et al., 1976). Whether children really have a bias to map words onto basic-level kinds is controversial (Callanan et al., 1994), but it is certainly a plausible proposal. More- over, the basic-level constraint, together with the taxo- nomic constraint and mutual exclusivity, actually solves the induction problem, because each object belongs to one and only one basic-level category. However, this so- lution only works for basic-level words like \dog , and in fact is counterproductive for all the words that do not map to basic level categories. How do we learn all the other words we know at superordinate or subordinate levels? Some experimenters have found that seeing more than one labeled example of a word may help childern learn superordinates (Callanan, 1989), but there have been no systematic theoretical explanations for these ndings. Regier (1996) describes a neural network learn- ing algorithm capable of learning overlapping words from positive evidence only, using a weakened form of mutual exclusivity that is gradually strengthed over thousands of learning trials. However, this model does not address the phenomenon of \fast mapping (Carey, 1978) { the meaningful generalizations that people make from just one or a few examples of a novel word { that is arguably the most remarkable feat of human word learning. To sum up the problem: taking the taxonomic, mu- tual exclusivity, and basic-level assumptions literally as hard-and-fast constraints would solve the problem of in- duction for one important class of words, but at the cost of making the rest of language unlearnable. Admit- ting some kind of softer combination of these constraints seems like a reasonable alternative, but no one has of- fered a precise account of how these biases should inter- act with each other and with the observed examples of a novel word, in order to support meaningful generaliza- tions from just one or a few examples. This paper takes some rst steps in that direction, by describing one possi- ble learning theory that is up to the task of fast mapping

Abstract: During speech acquisition, children form quick and rough hypotheses about the meaning of a new word after only a single exposure—a process dubbed “fast mapping.” Here we provide evidence that a border collie, Rico, is able to fast map. Rico knew the labels of over 200 different items. He inferred the names of novel items by exclusion learning and correctly retrieved those items right away as well as 4 weeks after the initial exposure. Fast mapping thus appears to be mediated by general learning and memory mechanisms also found in other animals and not by a language acquisition device that is special to humans.