Recent studies (e.g. Yu & Smith, 2007) show that both adults and young children can utilize cross-situational statistical information to build word-to-world mappings and solve the reference uncertainty problem in language learning. Our recent simulation work implemented several computational to take into account constraints of human learners, such as attentional limitations, embodied selection of visual information, and forgetting of stored information over time; all these models can fit behavioral data well. To distinguish between those models, this work reports the results from a new series of experiments in which adults are exposed to a rapid series of learning trials, wherein any given training trial contains uncertainty in sound-to-picture mappings, but in which this uncertainty is resolved across multiple trials. Several models of learning strategy are proposed and tested using variations of the original task (Yu and Smith, 2007) that provide more constraining data than merely percent of words learned, including (1) ranking results of options at testing, (2) effects of acquired knowledge to subsequent learning, and (3) effects of frequency variations and violations of the mutual exclusivity constraint. More importantly, we measure learners' moment-by-moment eye movements to (1) correlate eye movement patterns with learning results,(2) measure dynamic changes of eye movements from first trials to later trials, (3) encode this information in our models to infer underlying learning mechanisms based on the synchrony between eye movements and speech perception, and (4) evaluate which model is more cognitively plausible. |