The Future of Neural Machine Translation
Improvements to Neural Machine Translation (NMT) engines fall into one of two buckets. The first bucket has to do with accuracy. These problems are largely met with engineering solutions that offer incremental improvements. This applies to high-resource language pairs (e.g., English to Spanish), but especially to low-resource language pairs (e.g., Hindi to Hebrew) where parallel training data is scarce.
Google’s solution is to train a massively multilingual NMT to transfer knowledge about one language to another. In this way, the learnings from the high-resource languages could be applied to the low-resource languages. The goal is to achieve the same level of quality between a language pair with 10,000 training examples as one with 10 million. This is still very much a work in progress and one issue Google has faced is called “negative transfer,” where the quality of the high-resource language translations declined as they added more languages to the model. There is a need for continued fine-tuning.
The second bucket is around understanding meaning.
Accuracy is mostly a data/math problem; meaning requires a design thinking approach that is empathy-driven.
One challenge is helping users find their voice in another language. The one-size-fits-all nature of NMTs today erases the individual’s personality and hinders their ability to engage authentically in another language. There are B2B players like Unbabel who offer a level of personalization to enterprise clients. Still, currently, no one has been able to figure out how to do that for individual consumers.
Part of solving for meaning will involve developing feedback loops with users. Asking users generally, “Is this how people would say this?” or specifically, “Is this how you would say this?” can help ensure that the meaning is not being lost in translation. What those feedback mechanisms will look like is unclear, but the general idea is that human insight paired with NMTs will result in a more nuanced understanding of language.
The long-term goal is to move NMTs towards something closer to the way human interpreters work.
Acting as an interpreter involves taking in the source text, parsing out the meaning, and delivering the resulting message in the way that most closely aligns with the same sentiment and purpose. This may require swapping out idioms or cultural references. We still have a long way to achieve this goal, but advancements in the field of natural language processing offer some hope.