“…For underspecification-based systems, it can be assumed that children look for properties that the various environments in which exponents with the same form occur have in common; i.e., they learn underspecified feature structures of exponents by intersecting the sets of the different (fully specified) environments; see Harley (2001) and Pertsova (2007) for proposals along these lines (essentially, this is what Pertsova's NoHomonymy learner mentioned above does). On this view, the child assumes a syncretism to be systematic (i.e., going back to a single entry) whenever possible (see Pertsova (2007, 135)), and postulates two separate entries only as a last resort (e.g., when the interaction of (i) the Subset Principle and (ii) a system of decomposed features that is assumed as given fail to permit a coherent underspecified feature structure underlying two occurrences of one exponent form); this is essentially the meta-grammatical Avoid Accidental Homonymy condition assumed in Embick (2003), or the Syncretism Principle argued for in Müller (2007a) and Alexiadou & Müller (2008). Evidently, such an approach is not available in an underspecification-free approach such as the one developed here: Intersection invariably leads to underspecification.…”