Lila Gleitman, Tamara Medina, Jesse Snedeker, and John Trueswell
"Everybody" "knows" that vocabulary acquisition starts as follows: The learners try to match up the sound segments they isolate from the stream of speech with things and events in view. However, because mothers do not produce conversation that is a faithful running commentary on the stream of events , this procedure should be wildly errorful. But it is not. According to Everybody, the learners succeed by storing multiple sound-situation pairs and then parsing out that which is common to all - or most - of them. So the correlation between hearing "rabbit" and seeing a rabbit will be better, across scenes observed, than the correlation between hearing "rabbit" and seeing a mongoose. This theory goes back at least to John Locke, is formally modeled by many contemporary computer scientists (e.g., Siskind, 1996), and is documented as well in lots of experiments by psychologists (e.g., Yu & Smith, 2007). In this talk, I will argue that such a procedure (a) doesn't, and (b) couldn't, work. The evidence is from further experiments with adults, who can be assumed to have the requisite cognitive categories and information-handling capacities for the items tested, and who are trying to learn them by observing situational contexts. Based on the findings, we suggest that observational vocabulary learning is characteristically from one encounter ("fast mapping, per Carey, 1978). Or, more modestly, from one-and-a-bit encounters.
References:
Carey S (1978) in Linguistic Theory and Psychological Reality, eds Halle M, Bresnan J, & Miller GA (MIT Press, Cambridge), pp 264-293.
Yu C, Smith LB (2007) Rapid word learning under uncertainty via cross-situational statistics. Psychol Sci 18(5):414-420.