Improved neural machine translation systems for low resource correction tasks

Abstract

Recent advances in Neural Machine Translation (NMT) systems have achieved impressive results on language translation tasks. However, the success of these systems has been limited when applied to similar low-resource tasks, such as language correction. In these cases, datasets are often small whilst still containing long sequences, leading to significant overfitting and poor generalization. In this thesis we study issues preventing widespread adoption of NMT systems into low resource tasks, with a special focus on sequence correction for both code and language. We propose two novel techniques for handling these low-resource tasks. The first uses Generative Adversarial Networks to handle datasets without paired data. This technique allows the use of available unpaired datasets which are typically much larger than paired datasets since they do not require manual annotation. We first develop a methodology for generation of discrete sequences using a Wasserstein Generative Adversarial Network, and then use this methodology to train a NMT system on unpaired data. Our second technique converts sequences into a tree-structured representation, and performs translation from tree-to-tree. This improves the handling of very long sequences since it reduces the distance between nodes in the network, and allows the network to take advantage of information contained in the tree structure to reduce overfitting.