Hierarchical Structural Learning for Language Generation

We develop novel generative methods for language modeling that leverage graphical structure instead of traditional, strict left-to-right generation. Language generation is viewed as a topdown process through a parse tree. Our models are applicable to language modeling and representation as well as neural language parsing. We also propose a novel distribution for language modeling that enables these models. We both experiment with methods using supervised parse trees and methods that infer the most useful parse tree in an unsupervised manner.