We use historical change to explore whether children filter their input for language learning. Although others (e.g., Rohde & Plaut, 1999) have proposed filtering based on string length, we explore two types of filters that assume richer linguistic structure. One presupposes that linguistic utterances are structurally highly ambiguous and focuses learning on unambiguous data (Dresher, 1999; Fodor, 1998b; Lightfoot, 1999). The second claims that children learn only from matrix clauses (Lightfoot, 1991), defining simplicity in a structural manner. We assume that certain language changes occur via mismatches during acquisition. This allows us to use patterns of change to demonstrate that filtering restrictions are necessary to model language learning. Viewing language change as a result of mismatches during learning thus constrains the learning algorithm itself.

Language involves knowledge of multiple complex linguistic systems, such as phonology, morphology, and syntax. For each system, the process of acquisition requires children to discover the underlying components of that system, which native speakers use to generate the observable data. Theoretical research traditionally provides a description of the knowledge to be acquired, identifying