Tree-structured decoding with doubly-recurrent neural networks

Abstract

We propose a neural network architecture for generating tree-structured objects from encoded representations. The core of the method is a doubly recurrent neural network model comprised of separate width and depth recurrences that are combined inside each cell (node) to generate an output. The topology of the tree is modeled explicitly together with the content. That is, in response to an encoded vector representation, co-evolving recurrences are used to realize the associated tree and the labels for the nodes in the tree. We test this architecture in an encoder-decoder framework, where we train a network to encode a sentence as a vector, and then generate a tree structure from it. The experimental results show the effectiveness of this architecture at recovering latent tree structure in sequences and at mapping sentences to simple functional programs.