Tim O’Donnell was in Leiden last week for the The Comparative Biology of Language Learning workshop, held at the Lorentz Center April 3–7. He gave a talk Thursday, title and abstract below:
Bayesian Program Learning of Morphophonological RulesBoth children and linguists confront a similar problem of inference:given utterances produced by speakers, together with aspects of themeaning of those utterances, discover the grammatical principles thatrelate form to meaning. We study this abstract computational problemwithin the domain of morphophonology, contributing a computationalmodel that learns phenomena from many natural languages andgeneralizes in humanlike ways from data used in behavioral studies ofartificial grammar learning.Our work draws on two analogies. The child-as-linguist analogy holdsthat both children and linguists must solve the same abstractinductive reasoning problem, even though the nature of the input dataand underlying mental algorithms are surely different in precisedetail. Accordingly we isolate the problem of learningmorphophonological systems, and show that a single solution to thisproblem can capture both linguistic analyses from natural languagesand infant rule learning of artificial languages. We adopt theframework of “Bayesian Program learning” (BPL) – in which learning isformulated a synthesizing a program which compactly describes theinput data. This learning-as-programming analogy lets us exploitrecent techniques from the field of program synthesis to inducemorphophonological rules from data. While child-as-linguist poses thecomputational problem, learning-as-programming offers a solution.