Chomsky vs. statistical language models: the argument over whether they count as science

At MIT's 150th anniversary Brains, Minds, and Machines symposium, Chomsky criticized purely statistical methods as "approximating unanalyzed data" with no scientific insight. An article on the Hacker News front page on November 6 answers these points one by one. The author argues that engineering success is not the only standard for science but can serve as evidence a model works, and that probabilistic models are the simplest tool for describing stochastic phenomena like language. On theoretical limits to language learning, the article says Gold's theorem's limits do not apply to natural language learning, and that the learnability of probabilistic context-free grammars (PCFG) was proven long ago.

Written by AI from the coverage · updated 1 hour ago

Not enough continuous observations to draw a trend yet.