Guided NMT

Abstract

Guided Neural Machine Translation
Guided Neural Machine Translation Zoom

The area of machine translation (MT) studies the automatic translation from one natural language to another. Count-based statistical models have been dominating open domain MT for the last decades. For the last few years, however, the field is experiencing a clear shift towards neural machine translation which carries out the translation by using a single neural network.

This report introduces traditional MT briefly, and describes neural MT in great detail with a comprehensive review of the latest neural MT literature. We show that neural MT is in stark contrast to traditional statistical MT, and that both paradigms have complementary strengths and weaknesses. Therefore, we advocate combining both approaches. This report contains two system-level combination schemes which were developed in my first year of the PhD program and resulted in two conference publications. My goal is to continue working on bridging the gap between traditional and neural MT even beyond system-level combination. For example, by using neural models for word alignments which are essential to count-based models, or designing specialized neural network architectures, neural and traditional MT could benefit from each other on a more profound level. My plan for a successful PhD thesis is to develop an efficient and effective approach to MT which incorporates ideas from both methods.

Downloads