Improving Neural Machine Translation with External Information Thesis Proposal

Results of previous studies in the field suggest that neural machine translation (NMT) models can achieve better performance by exploiting information from external sources. In this thesis proposal, we categorize a variety of types of the enhanced models into four classes – multimodal, multilingual, linguistically cosupervised, and linguistically inspired. We summarize some of the approaches in each of these categories found in the literature. A part of this work is a development of a sequence-to-sequence learning toolkit designed for fast prototyping of various kinds of experiments. We report results of experiments proposed by us in the past, which were mainly based on the multimodal models. In the future, we plan to focus on the linguistically cosupervised models, which could make use of the abundance of linguistic annotation available.