Cross‐Situational Learning: An Experimental Study of Word‐Learning Mechanisms

Abstract

Cross-situational learning is a mechanism for learning the meaning of words across multiple exposures, despite exposure-by-exposure uncertainty as to the word's true meaning. We present experimental evidence showing that humans learn words effectively using cross-situational learning, even at high levels of referential uncertainty. Both overall success rates and the time taken to learn words are affected by the degree of referential uncertainty, with greater referential uncertainty leading to less reliable, slower learning. Words are also learned less successfully and more slowly if they are presented interleaved with occurrences of other words, although this effect is relatively weak. We present additional analyses of participants' trial-by-trial behavior showing that participants make use of various cross-situational learning strategies, depending on the difficulty of the word-learning task. When referential uncertainty is low, participants generally apply a rigorous eliminative approach to cross-situational learning. When referential uncertainty is high, or exposures to different words are interleaved, participants apply a frequentist approximation to this eliminative approach. We further suggest that these two ways of exploiting cross-situational information reside on a continuum of learning strategies, underpinned by a single simple associative learning mechanism.

Finding meaning in a noisy world: exploring the effects of referential ambiguity and competition on 2·5-year-olds’ cross-situational word learning

“…However, as noted above, children in the long-competition condition performed worse than would be expected if they tracked the co-occurrence probabilities for the entire set of candidate referents. Our results are thus most consistent with recent models that offer a middle ground between accumulative and conjecture-based approaches (e.g., Smith et al, 2011;Yurovsky & Frank, 2015). Although adults and children are sometimes capable of accumulating information about multiple potential referents for a word (e.g., Vouloumanos & Werker, 2009;Yurovsky et al, 2014), their ability to do so depends on the difficulty of the learning situation and the demands imposed on attention and memory (e.g., Smith et al, 2011;Yu & Smith, 2011;Yurovsky & Frank, 2015).…”

“…In addition, these four observations of the novel noun occurred without any other novel words interleaved between them. Recent evidence suggests that both children and adults have greater difficulty aggregating cross-situational information across observations of a given word when those observations are interleaved with observations of other novel words (e.g., Smith et al, 2011;Vlach & Johnson, 2013), in part because of the greater memory burden imposed by encoding, retrieving, and comparing multiple sets of candidate referents. The fact that children in the long-competition condition failed even with consecutive observations that occurred close together in time indicates that the burden imposed by the high-probability competitor referents was substantial.…”

“…Recent evidence suggests that under at least some circumstances, both adults and children can use cross-situational information to identify the referents of new words (e.g., Akhtar & Montague, 1999;Childers, 2011;Childers & Paik, 2009;Dautriche & Chemla, 2014;Gillette, Gleitman, Gleitman, & Lederer, 1999;Scott & Fisher, 2012;Smith, Smith, & Blythe, 2011;Smith & Yu, 2008;Yu & Smith, 2007, 2011. For instance, Smith and Yu (2008) presented 12-and 14-month-olds with a series of training trials in which pairs of novel objects were accompanied by two novel labels.…”

A communicative approach to early word learning

“…While infants demonstrate sensitivity to the co-occurrence information between words and objects, their memory for this information is quite fragile even under low levels of ambiguity (Vlach & Johnson, 2013;Vouloumanos & Werker, 2009). Even for adults, this process of statistical inference appears to be highly constrained 90 by limits on memory and attention (Smith et al, 2011;Trueswell et al, 2013;Yurovsky & Frank, 2015). In contrast to domains like low-level vision, where human performance is often quite well described by ideal observer models (e.g., Najemnik & Geisler, 2005), human statistical word learning is markedly less efficient than should be expected from a system that made optimal use of the 95 available information Yu & Smith, 2012b;Yurovsky & Frank, 2015).…”

“…Because this computational framework makes learning the lexicon the goal of language acquisition, many models are tested by comparing the lexicons they have inferred to a "gold standard" lexicon (e.g., Frank et al, 2009;Fazly 270 et al, 2010;Yu, 2008). And similarly, experimental participants are tested for their ability to select the right referent when cued with the right word (e.g., Smith et al, 2011;Yu & Smith, 2007;Yurovsky et al, 2014). But this is drastically different from the way that real children's language knowledge is tested outside the laboratory.…”