IRMACS: The Interdisciplinary Colloquium: "Pareto optimal machine translation: multiple objectives are better than one"

Abstract

If machines are to learn how to translate from one language to another (say, Chinese to English) we first need to devise a measure that can score the output of a machine translation system against a reference translation produced by a human expert. We can then aim to iteratively get better at translation by adjusting our translation model parameters to minimize error as per this measure.

The problem with natural language is that there is no single way to translate a sentence (there could be multiple reference translations) and so a good measure of translation quality is difficult to obtain. As a result, many translation quality measures have been proposed that address different aspects of translation quality. Current statistical machine translation systems (including Google) pick one measure and use it for optimization.

In this talk, I present an approach that uses multiple measures as independent objectives when training a machine translation model. We train a machine translation system that produces Pareto optimal translations, high scoring simultaneously in multiple objectives.

Rather than the conventional approach of choosing the right trade off solution from all solutions that are Pareto optimal, we combine them all using a novel algorithm that uses an ensemble of experts from the Pareto frontier to produce translations. Our experimental results show that our approach obtains better translations as measured by different objectives and we show that humans find it easier to edit our output into a fluent target language sentence.