Fresh off of a record setting quarterly earnings release, Google is heating up competition in the neural machine translation (NMT) race with AutoML Translation, a new capability of its AutoML machine learning suite of products currently released in beta. AutoML essentially allows Google’s cloud platform customers to employ their own training data to customize and train domain-specific machine learning models.
Or, as Google AI Chief Scientist Fei-Fei Li put it in a blog post, “with AutoML Translation you can upload translated language pairs to train your own custom translation model.”
Google announced during its first day of the Google Cloud Next conference on July 24 that its previously alpha-only AutoML Vision was going into beta stage. Two new capabilities are being launched in beta as well: AutoML Natural Language and AutoML Translation.
The move is another step forward in the rapidly evolving machine translation product offering by big tech and follows Amazon’s market entry in April 2018.
AutoML Translation Already Integrated
AutoML allows Google cloud service users to upload translated language pairs to train their own custom, domain-specific NMT model.
Google has already updated its Cloud Translation landing page to explain that its users can now use the Google Translation API along with AutoML Translation. This means that not only can Google customers use the Translation API to use Google’s NMT engines, they can now also employ their own training data to improve those engines.
Google’s Cloud Translation pricing is separate from AutoML Translation. The former charges USD 20 per million characters for up to a billion characters monthly, while the latter charges USD 76 per hour of training time after the first two hours. Additionally, AutoML Translation’s “prediction” capability charges USD 80 per million characters after the first 500,000 characters.
As of this writing, the AutoML Translation beta supports 17 language pairs, most of them bidirectionally.
Google is essentially allowing customers to improve their Translation API output with an integrated capability that lets them train the NMT engine without machine learning expertise. This will likely put pressure on more niche machine translation providers, whose unique selling point often revolves around customization of machine translation engines.
It potentially presents yet another business opportunity for language data brokers, who can now pitch their data troves to what’s likely going to be an increasing number of companies looking to build custom engines.
Expert commentary from other media and from those who Slator follows, however, points out a few caveats. One concern is that Google is not very transparent when it comes to the underlying processes used in the technology or how its output compares to competitors, as the offering targets more generalist business customers instead of expert programmers.
Meanwhile, others point out that NMT output quality also relies on the hardware, and users who take advantage of Google’s cloud platform Translation API and AutoML Translation combo will be forced to rely on Google’s hardware as well.
Download the Slator 2019 Neural Machine Translation Report for the latest insights on the state-of-the art in neural machine translation and its deployment.