The Translation: Google Goes Neural

On Wednesday, the biggest machine translation service in the world quietly got dramatically better, and the improvement was the sound of a decade of research arriving in production. Google announced this month that it has switched Google Translate to a neural network system for eight major languages, replacing the statistical approach that had powered the service for years. The company says the new system cuts translation errors by around sixty percent. The change is the November 2016 story, and the story is the lesson: the advances that change the world often arrive not with a bang, but with a better translation.

The Translation is the subject of this article: what neural machine translation is, why it is better, how Google built it, and what it means for the future of language and artificial intelligence. The new system began rolling out to users this month, for translations to and from English with French, German, Spanish, Portuguese, Chinese, Japanese, Korean, and Turkish. This is the story of the translation, and the story is about the moment the machines started to understand sentences, not just words.

1. The Service

Google Translate is the most used translation service in the world, and the use is the context of the change. The service translates more text every day than all the human translators in the world produce in a year, and it does it in more than a hundred languages. The service is free, and the freeness is the point: Google built the tool that the world uses to talk to itself, and the tool is the bridge between the languages.

The service had been running on the same technology for a decade, and the technology was showing its age. The old system, phrase-based statistical machine translation, translated by matching phrases against a giant database of translated texts and stitching the matches together. The system worked, and the work was clumsy: the translations were often grammatical enough and wrong enough to be recognizable as machine output. The service was the bridge, and the bridge was in need of repair.

2. The Problem

The problem with the old approach was the unit of translation, and the unit was the phrase. The old system looked at a few words at a time, matched them against the database, and moved on, and the moving on ignored the context. The word "bank" could be a river bank or a money bank, and the old system had no way to know which one the sentence meant. The sentence was the real unit of meaning, and the old system could not see the sentence.

The problem was the reason the translations sounded mechanical. The machines translated the words and missed the meaning, and the missing was the gap between the machine output and the human expectation. The problem had been known for decades, and the solution had been known in theory: a system that could see the whole sentence, understand its structure, and generate a translation in the target language. The theory was the goal, and the goal was the neural network.

3. The Network

The new system is built on a neural network, and the network is the heart of the change. The system is an end-to-end deep learning model, a sequence-to-sequence architecture that reads the whole source sentence and writes the whole target sentence. The network has an encoder, which turns the source sentence into a representation of its meaning, and a decoder, which turns the representation into the target language. The attention mechanism lets the decoder focus on the right parts of the source sentence as it writes.

The network was trained on millions of sentence pairs, and the training taught it the patterns of translation. The network does not match phrases; it learns the mapping between languages at the level of meaning, and the learning is the difference. The network can handle the words that depend on context, and it can produce translations that sound like a person wrote them. The network is the method, and the method is the breakthrough.

4. The Paper

The change was previewed in the research, and the preview was the roadmap. In September, a team of Google researchers published a paper describing the system, called the Google Neural Machine Translation system, and the paper showed the approach in detail. The paper was the research version of the announcement, and the research had been running for years before the paper appeared.

The paper mattered because it showed the gap between the old and the new. The neural system did not just match the quality of the old system; it beat it by a wide margin, and the margin was measured by human raters who preferred the neural translations in the large majority of cases. The paper was the proof, and the proof was the preparation for the production launch. The paper was the beginning, and the beginning was the rollout.

5. The Launch

The launch this month was the arrival of the research in the product, and the arrival was the hard part. The neural network had to be fast enough to serve billions of requests, and the serving was an engineering problem as much as a research problem. The network had to run on Google's infrastructure, and the infrastructure had to translate in real time, and the real time was the constraint.

The launch covered eight language pairs, to and from English, and the eight were the beginning. The languages were chosen for the volume of traffic and the maturity of the training data, and the choice was the strategy: prove the system on the biggest languages, and expand from there. The launch was the first step, and the first step was the plan. The launch was the moment the research became a product, and the product was the translation.

6. The Quality

The quality of the new translations is the reason for the excitement, and the quality is the difference between a translation and a machine translation. The neural system produces translations that read naturally, that preserve the meaning of idioms and metaphors, that understand the context of words like "bank" and "light". The errors are no longer the errors of the old system; they are the errors of a system that is close to understanding.

The improvement is measured in the numbers, and the numbers are the story. Google says the neural system reduces translation errors by around sixty percent compared with the phrase-based system, and the reduction is the largest single improvement in the service's history. The numbers are the claim, and the claim is the demonstration: the machines are learning to translate the way people translate, and the way is the future.

7. The Future

The future of the change is the expansion, and the expansion is the roadmap. The eight languages are the beginning, and the system will spread to more languages as the training data and the infrastructure allow. The same neural methods will improve the voice translation, the camera translation, and the other products that use Google Translate underneath. The future is the spread, and the spread is the scale.

The future is also the bigger story of artificial intelligence. The translation system is a demonstration of deep learning in the most human of tasks, the task of language, and the demonstration follows the other milestones of the year: the AlphaGo match in March, the assistants in the spring, and now the translation in the autumn. The future is the pattern, and the pattern is the point: the machines are learning the skills that were thought to be uniquely human, and the learning is just beginning.

8. The Lesson

The lesson of the translation is about the quiet power of incremental improvement, and the power is easy to miss. The change was announced without a keynote and without a fanfare, in a blog post and a new version of the app, and the change was one of the biggest improvements in the history of the service. The lesson is that the advances that matter are often the ones that arrive quietly, and the quiet arrival is the sign that the work is solid.

The lesson is also about the value of the long view. The neural translation system was the product of years of research, and the research did not look like progress at the time; it looked like papers and experiments and dead ends. The progress arrived when the pieces came together, and the coming together was the payoff for the patience. The translation is the November 2016 story, and the story is the lesson: the machines started to understand sentences, not just words, and the understanding arrived with a better translation. The bridge between the languages just got stronger, and the strengthening is the future. The same neural methods are spreading through every product that touches language, and the spreading is the pattern of the decade. The engineers who learn to work with the models, to train them, to test them, and to know their limits, will be the engineers who build the next generation of tools. The translators of the world are watching, and the watching is the measurement. The change was also the answer to a question the industry had asked for years: when would the machines learn to translate the way people translate? The answer arrived this month, and the answer was the network.

Tags

#technology #ai