Google announced a large-scale update of the Google Translate service: the basis of the translator will be driven by neural networks. For some languages, neural translation began to be used from November 2016 while others, including Russian, will switch to new technology in the near future. The developers argue that neural networks significantly improve the quality of translation, because machines can analyze not merely individual words and phrases but complete sentences and context.
How the new technology differs from the previous one? How to measure the quality of translation, and whether the machines can translate from one language to another as well as people?
Konstantin Benyumov, the journalist of Russian online publication Medusa, spoke with Barak Turovsky, Product Lead of Google Translate.Barak Turowski has been with Google since 2012, for the last three years managing Google Translate, including user interaction algorithms and design. Barak Turowsky was born in the USSR, lived and worked in Israel before moving to the United States. At his request, the conversation was conducted in English, but during the interview Mr. Turowsky frequently switched to Russian.
Since I read the interview in Russian, the translation of an excerpt of the interview below is mine, not Google Translate. Although Turowsky’s words are given in quotes, these might not be the exact verbiage he used when he spoke in English. The likelihood that Mr. Turovsky will check my translation for accuracy is rather small. Should it happen, however, I can always put a blame on… Google Translate. So there.
– What is the weight of Google Translate in the company?
“The Translator is a very important product for us, and Google allocated significant resources to its development. The main reason for this is that almost 50% of the Internet content is in English. However, only about 20% of the people proficient in English. This is a very serious obstacle for effective use of the Internet. In fact, there are two obstacles. First, there is an infrastructure barrier: a huge number of people, for example in China and India, do not have access to the Internet at all. But this situation is gradually changing, thanks to the development of mobile Internet and smartphones. Thus nowadays people in these countries often go directly to the mobile Internet. However, they immediately encounter a language barrier, and this effectively limits their usage of the Internet.
Therefore, Google Translate is a priority product for us. The Translator is an immensely popular product — nearly a half billion people per month use it, we translate about 140 billion words per day.”
– Is Google Translate main goal mostly ideological — to unite the world, or mostly commercial — to attract more users?
“For our users, the Translator is free. My main task as the Product Lead is to provide services to as many people as possible. Another important goal is to simplify access to the Internet, remove barriers that hinder its effective use. Personally, I’m not concerned about monetization whatsoever.
– Let’s talk about the design for which you are also responsible. Design in Google Translate clearly does not play a leading role. At least, it looks rather simple and over the years the project has undergone a minimum of changes. Is is so?
“You are absolutely right. Google Translator is a tool, its task is to be useful and convenient, and the design should therefore be minimalistic in order not to interfere with its effective use. But we constantly introduce new mechanisms and ways of interaction with the Translator, for example, using a mobile phone or camera. There is the Word Lens — a function that allows to use the technology of augmented reality to aim the camera at the text and translate it.”
– Has the translator learned to work with pairs of languages directly, bypassing English?
“Not quite yet. The popularity of English on the Internet means that the vast majority of training data is a translation from English into other languages and back. And if a user needs to translate from Russian into Japanese, in most cases we have to translate first from Russian into English and then into Japanese. Of course, direct translation would be better, but we support translation from 103 languages, that is, the number of their combinations is 103 square, it is more than ten thousand language pairs. For training it is very difficult.
But neural translation allows machines to learn to work with multiple languages at the same time. For example, languages can be combined into related groups, which greatly simplifies our work. In addition, within the related groups, we can create working models that do not need English language as an intermediary.”
Further in the interview, Turovsky answers questions about putting a Neural Machine Translation into action. Instead of me competing with Google Translator, I suggest you to read Barak Turovsky’s related blog articles Google’s new translation software is powered by brainlike artificial intelligence and Found in translation: More accurate, fluent sentences in Google Translate.