Organized semantic representations encoding across- and within-domain distinctions are a hallmark of mature cognition, and understanding how they change with experience and learning is a key endeavor in developmental science. Existing computational modeling studies provide a mechanistic framework for understanding how structured semantic representations emerge as a result of development and learning. However, their predictions remain largely untested in young children, with the existing evidence providing only indirect tests of these predictions. Across two experiments, we provide the first direct examination of a key prediction derived from these computational models -- that early in development, broad across-domain distinctions should generally be more strongly represented relative to finer-grained within-domain distinctions. The results support this hypothesis, being consistent with the exploitation of patterns of covariation among entities as a mechanism supporting the acquisition of structured semantic representations.