How machines learn languages... and how companies and customers can benefit from it
Tobias Jaroschek and his partner Björn Hanewinkel sell goods relating to home, household, decoration and gardening in all European marketplaces of Amazon. But honestly, who speaks more than two languages fluently and can whip out product descriptions in Spanish, French or Turkish?
Tobias Jaroschek and Björn Hanewinkel, the two managing directors of Mercatura Warenhandel GmbH in Moers, use Amazon’s machine translation so that they can serve customers in their respective native language.
"The translation service helps us offer our products quickly and efficiently abroad. Our international sales share has reached by now almost 40 percent," says Jaroschek, describing the benefits.
Another advantage is that machine translation makes it possible to test the demand for certain products in foreign markets without much effort.
"There are many products in our range that have specific lengths or sizes, such as cushions or blankets that we produce specifically for the German domestic market. The straightforward translation ensures that, for example, we can offer high-quality down duvets in France as well, so we can quickly see whether there is actually any demand at all," Jaroschek explains.
Amazon translates hundreds of millions of product pages every year on Marketplace. This gives even smaller and medium-sized companies the opportunity to enter international trade with little or no additional effort.
The translation service is a development of the Amazon Machine Learning team, in collaboration with experts from the in-house team in Luxembourg, which implements new technologies, and experts from Amazon Web Services (AWS) in the USA.
"Computer specialists have been dealing with machine translation for decades," says Hagen Fürstenau, manager of Applied Science at Amazon. But in recent years we have experienced groundbreaking leaps in development."
The Amazon Machine Learning team and its partners use neural network techniques. Put simply, this method is based on the functioning of the human brain. The result is that the translation system cannot only perform word-for-word translations, but "learns" to decrypt the meaning of a sentence.
In recent years we have seen groundbreaking advances in machine learning
A practical example in which the experts processed an English product description by an old and new translation system based on neural networks:
The initial sentence in English:
“Artwork is printed at high resolution with vivid color on thick high quality canvas to create the look and feel of the original nature and masterpiece.”
This is how the old system translates the text:
„Kunstwerk ist mit hoher Auflösung in lebendigen Farben auf hochwertiger Leinwand, sehen aus und fühlen sich an wie die Originale.“ ("Artwork is high resolution in vivid colours on high quality canvas, looks and feels like the originals.")
And that's the way it sounds when an intelligent translation system was at work:
„Das Kunstwerk wird in hoher Auflösung mit lebendiger Farbe auf dickem, hochwertigem Canvas gedruckt, um das Aussehen und die Haptik der ursprünglichen Natur und des Meisterwerks zu schaffen.“ ("The artwork is printed in high-resolution, vivid colour on thick, high quality canvas to create the look and feel of the original nature and masterpiece.")The practical example shows a clear improvement thanks to the new system," says Hagen Fürstenau. “Even though the translation is certainly not always perfect".
Slang or colloquial phrases can still cause problems for a machine translator. In German, the phrase "you are on the wrong track" is translated by the neural system as "You are on the woodway" and leaves English-speaking users rather lost.
"Such mistakes are part of the learning process and help us make the system a little bit better every day," said Alex Klementiev, Senior Manager, Applied Science at Amazon. Now Amazon is successfully offering Amazon Translate, which can translate not only product descriptions but any kind of text, to external customers. In the future, this offers completely new application possibilities wherever fast and good translation is important. Language is alive - the results are again and again crazy sentences that make our teams - and our customers – laugh. We experience “Day One” in developing machine translation literally every day!