Inductive and deductive learning of grammar: dealing with incomplete theories

A framework is given for learning plausible unification-based natural language grammars. The authors assume that the system will have some initial unification-based grammar and they use learning to overcome the incompleteness of this grammar (its undergeneration, i.e. where it fails to generate strings that humans would regard as grammatical). The authors use both model-driven (deductive) and data-driven (inductive) learning. The framework requires no a priori decision to be made about the balance between being model-driven or data-driven. One can experiment using anything from being purely model-driven to being purely data-driven. The authors expect the data-driven learning to compensate for weaknesses of the model-driven learning (most notably the problems of incompleteness of the model), and they expect the model-driven learning to compensate for weaknesses of the data-driven learning (most notably the likelihood that data-driven learning will learn a linguistically implausible grammar)