Hypothesis Testing and Associative Learning in Cross-Situational Word Learning: Are They One and the Same?

Hypothesis Testing and Associative Learning in Cross-Situational Word Learning: Are They One and the Same? Chen Yu, Linda B. Smith, Krystal A. Klein and Richard M. Shiffrin ({chenyu}@indiana.edu) Department of Psychological and Brain Sciences, and Cognitive Science Program, Indiana University Bloomington, IN 47405 USA cluttered than this with many candidate objects for a word and many candidate words for an object, and in the discourse context many shifts in attention among the candidate words and referents. These highly ambiguous learning environments with many words and many objects may significantly limit the plausibility of fast-mapping solutions. There is, however, an alternative way that learners might solve the indeterminacy problem, not in a single encounter with a word its referent, but across many trials and simultaneously for many words and referents. This solution is possible if learners can accumulate the statistical evidence across multiple learning situations. A learner who is unable to unambiguously decide the referent of a word on any single learning trial could nonetheless store possible word- referent pairings across trials, evaluate the statistical evidence, and ultimately map individual words to the right referents through this cross-trial evidence. For example, a young learner (an infant perhaps) might hear the words “bat” and “ball” in the context of seeing a BAT and BALL. Without other information, the learner cannot know whether the word form “ball” refers to one or the other visual object. However, if subsequently, while viewing a scene with the potential referents of a BALL and a DOG, the learner hears the words “ball” and “dog” and if the learner can combine the conditional probabilities of co-occurrences from two streams of data across trials, the learner could correctly map “ball” to BALL. This mechanism seems to be quite straightforward. However, until recently, there was no evidence as to whether human learners perform these kinds of statistical computations. In a series of recent experiments, we showed that both adults (Yu & Smith, in press) and 12 month-old infants (Smith & Yu, submitted) do calculate cross-trial statistics and find correct word-referent mappings amidst highly ambiguous learning contexts, and they do so with impressive accuracy over relatively few trials. Cross- situational statistical learning is clearly within the repertoire of human learners. This paper intends to go beyond demonstrating what language learners can do, and focus on investigating the internal learning mechanisms that may underlie their powerful statistical learning capabilities. What is the nature of the underlying learning processes? Can we provide a formal account of cross-situational learning? Traditionally, two classes of cross-situational learning mechanisms have been considered. One is associative learning. Across trials, the learner could accrue associations between words and their potential referents by strengthening and weakening associative links between experiences of Abstract Recent studies (e.g. Yu & Smith, in press; Smith & Yu, submitted) show that both adults and young children possess powerful statistical computation capabilities -- they can infer the referent of a word from highly ambiguous contexts involving many words and many referents. This paper goes beyond demonstrating empirical behavioral evidence -- we seek to systematically investigate the nature of the underlying learning mechanisms. Toward this goal, we propose and implement a set of computational models based on three mechanisms: (1) hypothesis testing; (2) dumb associative learning; and (3) advanced associative learning. By applying these models to the same materials used in learning studies with adults and children, we first conclude that all the models can fit behavioral data reasonably well. The implication is that these mechanisms – despite their seeming difference -- may be fundamentally (or formally) the same. In light of this, we propose a formal unified view of learning principles that is based on the shared ground between them. By doing so, we suggest that the traditional controversy between hypothesis testing and associative learning as two distinct learning machineries may not exist. Keywords: language computational modeling. acquisition, word learning, Introduction There are an infinite number of possible word-to-world pairings in naturalistic learning environments. Quine (1960) illustrated this indeterminancy problem with this example: Imagine an anthropologist who goes to a foreign country and observes a native speaker saying “gavagai” while pointing in the general direction of a field with a rabbit in it. The intended referent (rabbit, grass, the field, or rabbit ears, etc.) is indeterminate from this experience. Hard as it seems to be to infer referents correctly from such data, typically developing children have no problem using data of this sort to learn their native vocabulary smoothly and effortlessly. For 30 years, research on the indeterminacy problem has concentrated on single trial learning such that a language learner – despite the logical ambiguity pointed out by Quine – nonetheless correctly and rapidly maps the word to the intended referent on that trial and by most accounts does so on the basis of social, linguistic and/or representational constraints (e.g. Gleitman, 1990; Tomasello, 2000). However, most previous experiments showing such fast- mapping of a word to a referent were conducted in highly constrained laboratory environments. A typical scenario is like this: an experimenter presents one or two objects to young subjects and utters a simple phrase, such as “look, this is a toma!” Everyday learning contexts are much more