Word learning emerges from the interaction of online referent selection and slow associative learning.

Abstract

Classic approaches to word learning emphasize the problem of referential ambiguity: in any naming situation the referent of a novel word must be selected from many possible objects, properties, actions, etc. To solve this problem, researchers have posited numerous constraints, and inference strategies, but assume that determining the referent of a novel word is isomorphic to learning. We present an alternative model in which referent selection is an online process that is independent of long-term learning. This two timescale approach creates significant power in the developing system. We illustrate this with a dynamic associative model in which referent selection is simulated as dynamic competition between competing referents, and learning is simulated using associative (Hebbian) learning. This model can account for a range of findings including the delay in expressive vocabulary relative to receptive vocabulary, learning under high degrees of referential ambiguity using cross-situational statistics, accelerating (vocabulary explosion) and decelerating (power-law) learning rates, fast-mapping by mutual exclusivity (and differences in bilinguals), improvements in familiar word recognition with development, and correlations between individual differences in speed of processing and learning. Five theoretical points are illustrated. 1) Word learning does not require specialized processes – general association learning buttressed by dynamic competition can account for much of the literature. 2) The processes of recognizing familiar words are not different than those that support novel words (e.g., fast-mapping). 3) Online competition may allow the network (or child) to leverage information available in the task to augment performance or behavior despite what might be relatively slow learning or poor representations. 4) Even associative learning is more complex than previously thought – a major contributor to performance is the pruning of incorrect associations between words and referents. 5) Finally, the model illustrates that learning and referent selection/word recognition, though logically distinct, can be deeply and subtly related as phenomena like speed of processing and mutual exclusivity may derive in part from the way learning shapes the system. As a whole, this suggests more sophisticated ways of describing the interaction between situation- and developmental-time processes and points to the need for considering such interactions as a primary determinant of development and processing in children.

Learned Labels Shape Pre‐speech Infants’ Object Representations

“…Nonetheless, not all infants responded correctly to the label, and some shifted their attention away from the target. We do not therefore claim that these prespeech infants learned a new word during training, but rather we interpret these results as supporting accounts of early word learning in which infants learn label‐object associations incrementally (Bion, Borovsky, & Fernald, 2013; McMurray, Horst, & Samuelson, 2012; Yurovsky, Fricker, Yu, & Smith, 2014). That is, these infants learned something about the object, something about the label, and something about the mapping between the two.…”

When the daffodat flew to the intergalactic zoo: Off-line consolidation is critical for word learning from stories.

“…N recognise the novel words was typically better than their ability to recall and produce the novel words (Brown et al, 2012;Henderson et al, 2012Henderson et al, , 2013aDumay & Gaskell, 2007;McMurray et al, 2012); however, children only recalled 11% of novel words and adults recalled only 29% of novel words immediately after training. Children and adults of a similar age range in Henderson et al (2013a) showed mean cued recall rates of 18% and 61%…”

Do as I say, not as I do: A lexical distributional account of English locative verb class acquisition

“…Because the word layer used a winner-take-all mechanism, a small difference in activation could lead one unit to be selected over another. Thus, this part of the architecture instantiated the competition between referents and learned concept-word associations specified by both theories of word learning (McMurray, Horst, & Samuelson, 2012) and word production (Gordon & Dell, 2003).…”