Emergent Learning

             This week’s readings examined the idea of scalar implicature and why children have difficulty developing scales whereas adults do not.

In Accessing the Unsaid, Barner explores reasons for the failure of children to derive scalar implicatures in conversation, postulating that their inability to process differences between words like “some” and “only some” depends on whether the question asked is context-dependent. In the study, Barner asked children questions using context-independent words like “some” and context-dependent words like “the dog and the cat” and found that the context-dependent words changed responses significantly, whereas “only” did not change responses significantly at all. By focusing on children, the study shows that accessing scalar alternatives requires additional learning.

            Similarly, Stiller’s paper compares the counterfactual and linguistic alternatives hypotheses that explain how people derive scalar implicatures. By running 3 experiments with context-dependent scale, no scale, and scale with items of differing rarity, Stiller shows that linguistic and social factors must be integrated with world knowledge (like knowing that rarer features are more informative) to make lexical inferences possible.

            It’s interesting that children respond so well to context-specific uses of only (“only the cat and the dog”) but not to more general terms (“only some”) when the interpretations are highly similar. The fact that adults judge these statements the same way while children do not means that, at some point, children learn to generalize from specific cases like “dog and cat” to mean “some.” This is similar to Stiller’s finding that people infer that more rare features are more informative, a conclusion based on shared statistical knowledge. How much exposure to contextual examples is required for a generalized pattern to be learned? At what point to people learn the idea that rareness indicates informative-ness? At what point does an understanding of “some” and its scalar alternatives emerge?

            The idea of statistical world knowledge accumulating after a certain age to allow people to make certain inferences is a common theme in learning, not just in linguistics. Even the simple learning task of labeling items (furry four-legged mammal as “cat”) after being exposed to just a tiny subset of all the possibilities of what form that label can take is one that humans do particularly well. In thinking about machine learning, the accumulative learning tasks that humans struggle with, like differentiating between context-specific scales and lexical scales, might be more informative than the ones that we do easily. Maybe understanding how humans learn to generalize context-dependent information into lexical information can provide insight into learning from repeated exposure and the amount of exposure needed for learning to emerge, whether in humans or machines.