Learnability of CFLs: inferring syntactic models from constituent structure

Abstract

Inductive inference is a learning process based on discovering models for bodies of knowledge, given sample information. The inference process we discuss here is concerned with inductive acquisition of syntactic models for context-free languages (CFLs), given appropriate language samples. The knowledge to be modeled in this case is any CFL L, with the model to be determined a recognitive or generative characterization of L's syntactic structure. L will be learned syntactically once a machine M recognizing L, or a context-free grammar (CFG) G generating L, is inductively inferred from a sentence sample. The capability of distinguishing between L and its complement, or of generating all and only L 's sentences, is the knowledge acquired, with the learner (inference process) gaining this knowledge by acquiring M or G. An observer (informant, teacher, or oracle) has such knowledge of L and can provide the learner with appropriate sample information to ensure that M or G is correctly identified.