Modeling disambiguation in word learning via multiple probabilistic constraints

摘要: Modeling disambiguation in word learning via multiple probabilistic constraints Molly Lewis Michael C. Frank mll@stanford.edu Department of Psychology Stanford University mcfrank@stanford.edu Abstract the grocery store. There are an infinite number possi- ble meanings this given referential context, but child’s ability to correctly disambiguate would lead her rule out all for which she already had a name. With restricted hypothesis space, child is more likely identify correct referent than if objects context were considered as possible referents. What cognitive processes underlying effect? broadly two proposals literature. Under one proposal, Markman and colleagues (1988; 2003) suggest that children have constraint on types lexicons when meaning new — “mutual exclu- sivity constraint.” constraint, biased consider only those one-to-one mapping between words objects. Importantly, can be overcome cases where it incorrect (e.g. adjectives), nonetheless serves restrict set initially entertained novel word. Un- der view, then, effect emerges from structure lexicons. second ar- gued result online inferences made within refer- ential (Clark, 1987; Diesendruck & Markson, 2001). Clark suggests due pragmatic assumptions held by speakers. The first assump- tion speakers same speech community use refer (“Principle Conventionality”). assumption different linguistic forms Contrast”). In task described above, might reason (implicitly) follows: You used I’ve never heard before. Since, presumably we both call ball “ball” you’d meant you said “ball,” must object. Thus, un- account, not higher- order lexicons, instead in-the-moment using general principles. These traditionally been viewed competing explanations effect. Re- search area has consequently focused identifying empirical tests distinguish these theo- ries. For example, Markson (2001) com- pare performance told fact about object relative ref- erential label. They found disambiguated conditions argued grounds parsimony mechanism was responsible inferences. More recent evidence contradicts Young tend map even presence familiar competitors, finding dubbed “disambiguation” Theoretical accounts debated whether initial children’s principle mutual exclusivity) or situation-specific We could true. present hierarchical Bayesian model implements situation-level inference, show contribute with levels strength depending differ- ences situation language experience learner. additionally data testing prediction view disambiguation. Keywords: Word learning; exclusivity; mod- els. Introduction A central property each lexicon maps unique concept, concept 1987). Like other important regularities lan- guage grammatical categories), cannot directly observe property. Instead, they learn way consistent generalization basis specific word-object pairs. Even very young behave consis- tent con- cepts language. Evidence claim comes what known typical demon- stration Wachtel, 1988), chil- dren presented whisk ball), asked (“show me dax”). Children choose referent, behaving word-concept regularity language, across wide range ages ex- perimental paradigms (Mervis, Golinkoff, Bertrand, 1994; Mervis, Hirsh-Pasek, et al., Markman, Wa- sow, Hansen, 2003; Halberda, Bion, Borovsky, Fernald, 2013). This received much attention learn- ing literature because ambiguous contexts is, essence, core prob- lem learning. That any underdetermined (Quine, 1960), challenge world learner context. Critically, infer makes problem easier solve. suppose hears “kumquat” while produce aisle