Mutual Exclusivity in Cross-Situational Statistical Learning

Mutual Exclusivity in Cross-Situational Statistical Learning Daniel Yurovsky (dyurovsk@indiana.edu) Department of Psychological and Brain Science, and Cognitive Science Program 1101 East 10 th Street Bloomington, IN 47408 USA Chen Yu (chenyu@indiana.edu) Department of Psychological and Brain Science, and Cognitive Science Program 1101 East 10 th Street, Bloomington, IN 47408 USA ultimately discover correct word-referent associations. With respect to the above example, a learner may hear the words toma and bosa in the ambiguous context of seeing both a loofah and gyroscope without any information as to which word refers to which object. Although the mappings cannot be determined from this single situation, the learner could nonetheless solve the problem by keeping track of co- occurrences and non-occurrences across situations. Using this example, if the learner viewed a second scene with a spatula and gyroscope while hearing diti and toma, and could combine co-occurrence probabilities across these two situations, the learner could correctly infer that toma maps to gyroscope, “bosa” to loofah, and “diti” to spatula. Recent empirical evidence has shown successful cross-trial learning in ambiguous situations in both adults (Yu & Smith, 2007) and children (Smith & Yu, 2008). Yu & Smith (2007) exposed adults to a series of 27 ambiguous trials, each containing 4 unknown words and 4 possible referents. Their participants learned 9.5 out of 18 mappings in less than 6 minutes. Thus, humans can use statistical information across multiple situations to learn word-object mappings by storing, computing, and continuously reducing a set of possible referents over time. The present study intends to investigate the ME constraint in the cross-situational learning paradigm. Can adults learn multiple one-to-many word-referent pairings across a number of ambiguous learning trials? Since most previous ME experiments have focused on one-trial learning, and previous cross-situational language learning experiments used one-to- one stimuli, it is unclear what to expect. There are two potential outcomes. One is that human learners rely on mutual exclusivity in cross-situational learning in much the same way as in single-trial learning. As a result, they may still build one-to-one mappings while ignoring additional statistical regularities in the training data, or they may fail to learn at all due to the one-to-many confusion. This would suggest that ME is an additional constraint in cross-situational language learning. Alternatively, learners may break the mutual exclusivity constraint across multiple learning trials, demonstrating one-to-many mapping at test. This result would provide support for cross-situational learning as a fundamental mechanism supporting human learning in general, and word learning in particular. Abstract The Mutual Exclusivity (ME) constraint – a preference for mapping one word to one object – has been shown to be a powerful aid to children learning new words. We ask whether cross-situational language learning, in which word meanings are learned through computation of word-object co-occurrences across a series of highly ambiguous trials, is subject to the ME constraint. Our results show that participants can break the constraint to learn one-to-two word-referent mappings both when the referents are separated across time and when they are interleaved. This demonstrates the robustness of cross-situational statistical learning. We then use participants’ ratings of their knowledge after individual trials to shed light on the underlying learning mechanism. Our results suggest that the ME constraint may be applied at multiple points along learning – within a single trial, across trials, and at test – which may explain one of its residual test effects found in the traditional language literature. Keywords: mutual exclusivity; statistical learning; word learning; language acquisition Introduction The Mutual Exclusivity (ME) constraint – a preference for building one-to-one word-referent mappings – plays a critical role in human word learning. In Markman and Wachtel’s (1988) classic experiment, a child is presented with a known object (ball) and an unknown object (gyroscope) and is asked by the experimenter to bring the “toma”. Having never before heard “toma,” the child will select the novel object gyroscope as its referent, but not the object ball, which already has a name. Since this original experiment, mutual exlcusivity has been reliably demonstrated in various studies. Subsequent research has also proposed several potential explanations for children’s behavior, such as the Principle of Contrast (Clark, 1983), the Novel-Name-Nameless-Category Principle (Golinkoff, Hirsh-Pasek, Bailey, & Wenger, 1992), and a Pragmatic Account (Diesendruck & Markson, 2001). Further, the ME constraint provides a powerful and reliable mechanism to deal with the reference uncertainty problem (Quine, 1960) in a single learning trial. Given a novel word and multiple potential referents, learners can apply mutual exclusivity to rule out those referents with known names and determine the correct referent on a single trial. Recently, an alternative solution has been proposed to deal with the uncertainty in word learning: cross-situational statistics (Yu & Smith, 2007). A learner who is unable to determine the correct word-referent mappings in a single learning trial can nonetheless accumulate cross-trial statistics from a number of individually ambiguous training trials to Experiment 1 Following the cross-situational learning paradigm presented by Yu and Smith (2007), we asked participants to