2026
EMNLP
EMNLP 2026
Universal-Guided Grammar Induction
Abstract
AbstractWe study whether Universal Grammar syntactic principles can serve as explicit inductive biases for unsupervised grammar induction. We introduce a neuro-symbolic PCFG framework that converts hypothesized linguistic universals into rule schemata, and then learns rule probabilities from child-directed speech. We investigate three assumptions: 1) extended projection, 2) prototype theory of part-of-speech, and 3) the final-over-final constraint. Experiments show that with a much smaller hypothesis space, the universal-priored model achieves superior performance to state-of-the-art neural grammar induction models, and yields more interpretable functional category assignment.