A Psychologically Motivated Model of Word Learning

We present an on-line computational model that aims to bridge the gap between formal and psychological studies of word learning. Like the model of Fazly et al. (2010), our model incrementally processes input data rather than using iterative optimization to derive the lexicon over the entire corpus of learning instances (Yu & Ballard 2007, Frank et al. 2009). Unlike Fazly et al. (2010), our model incorporates recent experimental findings (Medina et al. 2010, 2011) which suggest a simpler learning mechanism than that assumed by most recent computational studies beginning with Siskind (1996). Our approach models word learning with a general class of mechanisms in which linguistic hypotheses (word-to-meaning mappings) are probabilistically assessed and evaluated against input data, similar to the variational model of parameter setting (Yang 2002). Following experimental results, our learner only considers one semantic hypothesis at a time for each word (partial cross-situational knowledge) rather than keeping track of all possibilities (full cross-situational knowledge). We show that this simpler mechanism can produce competitive results when tested on annotated child-directed English input. Motivation It is usually assumed that children learn the meanings of their first words by storing a large number of situations during which a word is uttered in memory and then finding the commonalities across these experiences. Two recent studies have challenged this idea. Instead, they suggest that learners consider one candidate meaning at a time, and converge to the target by making and testing hypotheses of word meaning against sequences of learning instances. Medina et al. (2010) used a procedure in which subjects heard nonsense words accompanying objects on a screen and were asked, for each word, to click on the object they thought it referred to (henceforth " clicking experiment "). Each word occurred with five different sets of objects, with an object representing the target meaning always present. Subjects could receive a highly informative (HI) instance (with only two objects to choose from rather than five) either as the first (HI first), third (HI middle), or last (HI last) instance for each word. If word learners were really accumulating full contextual data with each instance, one would expect performance on later instances to be significantly better than performance on earlier instances. However, subjects perform at chance level (50%) for all HI instances (see the left side of Fig. 1). Having a HI instance in the middle of the experiment was no better than having …