How do we learn?

Barner, Brooks, and Bale explore how kids generally cannot calculate scalar implicatures because they don’t have access to scalar alternatives with which to compare. Kids are sensitive to so many other linguistic cues like eyegaze or the motivations of the speaker, yet cannot perform the task of scalar implicature. This involves four steps of finding the literal meaning of an expression, generating alternative sentences with different scalar words, removing those that are less informative, and reinforcing the interpretation of the sentence by negating the alternatives. The paper argues against the idea that kids fail to compute scalar implicatures because of processing costs or incapability to reason pragmatically. Instead, it argues that the main issue is a lack of knowledge of which items are scales, therefore affecting kids’ abilities to compute alternatives when interpreting a sentence. It is interesting that kids are extremely good at interpreting numerals. This may be due to a brute-memorization process of the counting system, while kids never explicitly memorize quantifiers. The experiments revealed that the word “only” didn’t affect context-independent alternatives like “some” and “all,” but had a significant effect on contextual alternatives. For example, they were equally likely to say yes when asked if some of the animals were sleeping compared to if only some of the animals were sleeping. Even with the addition of “only,” kids weren’t good at interpreting sentences with context-independent scales.

Stiller, Goodman, and Frank built upon the previous paper by testing another experiment to prove that kids do indeed have the ability to reason pragmatically, but still fail at scalar implicature tasks. They assert that lexical items affect the kids’ abilities to perform well in scalar implicature tasks. They argue against both the Gricean counterfactual theory and the linguistic alternatives theory and claim that linguistic and social factors combine with an understanding of world knowledge. This extra knowledge about the world and language affects our ability to interpret sentences with quantifier words. For example, statistical knowledge of the rarity of a particular description or characteristic affects how we understand sentences.

Kids must learn to more generally reject under-informative descriptions before they can compute scalar implicatures. For example, kids learn that “animal” is an ill-suited word to refer to a dog since it is too general. This learning process is gradual and calls to mind Bayesian reasoning versus learning by deduction. Bayesian reasoning argues that the most restrictive hypothesis is always preferred. In this case, kids would favor “dog” to “animal” given certain examples of dogs, because the latter is more general, despite its accuracy. In contrast, deductive reasoning doesn’t help the child get any closer to narrowing in on “dog” instead of “animal” since both are still compatible with the given evidence and deductive reasoning says nothing about which hypothesis should be weighted more heavily. Inductive reasoning thus explains how kids can learn with a lack of negative evidence; parents don’t have to point to tables, houses, cats, etc. and tell their kids that those are “not dogs.”