People ignore token frequency when deciding how widely to generalize

Many theoretical accounts of generalization suggest that with increasing data, people should tighten their generalizations. However, these accounts presume that the additional data points are all distinct. Other accounts, such as the adaptor grammar framework in linguistics (Johnson, Griffiths, & Goldwater, 2007), suggest that when the additional data points are identical, generalizations about grammaticality need not tighten appreciably: they may be made on the basis of type frequency rather than token frequency (although token frequency can affect other types of learning). We investigated what happens in this situation by presenting participants with identical data in both a linguistic and a non-linguistic context, some ten times as much as others, and asking them to generalize to novel exemplars. We find that people are insensitive to token frequencies when determining how far to generalize, though memory has a small mediating effect: generalizations tighten slightly more when people may rely on a memory aid.