Authors
Alexander Clark
Year
2021
Abstract
The problem of learning syntactic structure is notoriously difficult, especially with mildly context-sensitive grammars. Existing approaches to learning these types of grammars are limited – they are not guaranteed to learn the correct grammars or structures, only the correct derived structures whether these are trees or strings. Here we present some progress towards strong learning of these formalisms, extending a recent result on strong learning of probabilistic context-free grammars to a class of probabilistic context-free tree grammars that is weakly equivalent to Tree-Adjoining Grammars. Given a sufficiently large sample of derived trees, the algorithm is guaranteed to converge to a structurally identical grammar to the grammar generating the trees. Combined with the approach to learning PCFGs from strings, this presents a possible model of the acquisition of these types of grammars.
Comment
Extended abstract