Discovery Procedure

A discovery procedure is a methodological framework in linguistics and formal science that specifies explicit, algorithmic steps for deriving the structure of a system from observation of its behavior, without prior theoretical assumptions about what that structure must be. In American structuralism, Leonard Bloomfield and Zellig Harris insisted that linguistic analysis should be conducted through discovery procedures — mechanical operations on a corpus of utterances that would yield the phonemes, morphemes, and syntactic classes of a language without appeal to meaning, intuition, or mental states.

The appeal of discovery procedures was epistemological: they promised an objective, replicable foundation for linguistic description that would free the field from the subjectivity of philological judgment. The problem was computational. Harris's 1951 procedures were computable in principle but intractable in practice for any but the smallest corpora. The dream of a fully automatic discovery procedure — a machine that would read a corpus and output a grammar — was not realized until the advent of modern machine learning and probabilistic grammar induction, which replaced Harris's deterministic algorithms with stochastic optimization. The discovery procedure thus lives a double life: as a philosophical ideal of empirical objectivity, and as a technical problem whose modern solutions bear little resemblance to the original vision.