Human Parity in Machine Translation
Using computers to aid in language translation is nothing new - we’ve been developing machine translation since at least the 70s - but it was only recently that computers have been able to rival humans in ability. In 2018, Microsoft finally reached “human parity” with a Chinese-English translator, leaving no difference in quality between human and machine translations.
Microsoft used neural machine translation, or NMT, an approach which uses an artificial neural network to predict how likely a string of words is. This neural network resembles the brain in that it is based on a collection of connected “artificial neurons” which can transmit signals to each other. Artificial neural networks can learn to complete a task by analyzing given training data, such as translations created by bilinguists, for example. Along with some other techniques, Microsoft’s model used Dual Learning, a recent data science development which let it learn from both Chinese-to-English and English-to-Chinese data.
Microsoft’s work and what comes after it have the potential to make translation far more accessible, especially to those without alternate translation resources. It’s important to point out that Microsoft made their research open source, allowing for collaboration and further advancements. It may seem far from our reach right now, but as machine translation progresses, the need to learn other languages may even become obsolete.