On Chomsky and the Two Cultures of Statistical Learning

Statistical and probabilistic models have achieved overwhelming dominance in computational linguistics, contrary to Chomsky's criticisms. While Chomsky argued that probabilistic models lack insight and that "probability of a sentence" is useless, modern statistical approaches (like PCFGs, smoothing techniques, and algorithmic modeling) successfully handle language ambiguity, gradual language change, and real-world linguistic variation. The essay defends the "algorithmic modeling culture" that prioritizes accurate prediction over simple interpretable models, arguing that natural language is inherently probabilistic and contingent rather than categorical and eternal. Statistical models not only achieve state-of-the-art engineering performance but also better represent linguistic reality than Chomsky's idealized, parameter-based Universal Grammar framework.