Issue #79 -Merging Terminology into Neural Machine Translation

23 Apr20

Author: Dr. Patrik Lambert, Machine Translation Scientist @ Iconic

After several years being the state of the art in Machine Translation, neural MT still doesn’t have a convenient way to enforce the translation of custom terms according to a glossary. In issue #7, we reviewed several approaches to handle terminology in neural MT. Just adding the glossary to the training data is not effective. Replacing the source term by a placeholder, and then the placeholder by the glossary translation is easy. However, it is a hard decision, since it doesn’t allow the model to discard the glossary translation in some contexts. It also does not allow the model to fully decide the position of the term translation in the target sentence. Adding constraints in decoding significantly reduces translation speed, which makes it unsuitable for production engines. In this post we take a look at a paper which considers strategies that help enforce the custom terminology translation, without forcing the model to make any hard decisions, and with no impact on translation speed.

Merging approaches

The paper by Wang et al. (2019) proposes three strategies
To finish reading, please visit source site