Yandex Blog

One model is better than two. Yandex.Translate launches a hybrid machine translation system

Today Yandex.Translate launched a hybrid machine translation system that combines neural and statistical approaches to machine translation to deliver our users an even higher quality translation that utilizes the complementary strengths of both translation models. The new system first translates users’ queries using both a statistical and a neural machine translation model.  Next, CatBoost, our gradient boosting library ranks the outputs of each model, ultimately selecting the highest quality translation. 

There are several approaches to machine translation and over the years, a number of technological advances have improved the quality of machine translation.  Since its launch in 2011, Yandex.Translate has been powered by statistical machine translation, a widely used approach that works by comparing example translations to find statistical correspondences between words in the two languages.

With today’s launch, Yandex.Translate now also includes a neural machine translation component, a method that has led to more fluent, human-like translations in the last few years. The new Yandex.Translate system is unique in offering users a free machine translation service that combines these two methods.

Statistical translation and neural translation models each have different strengths that complement each other. When combined in our new hybrid machine translation system, they will produce higher quality results than either of the underlying models alone. 

Statistical models prove extremely efficient at memorizing example translations and can produce better translations of words or phrases that are seen less frequently in the training data.  However, statistical machine translation break sentences up into words or phrases during the translation process, which sometimes makes it challenging to construct fluent translations.  

Neural machine translation models, on the other hand, can process entire sentences at once.  Neural models choose a translation based on the full context of a query, often resulting in much more fluent, human-like translations. But, because the neural network uses context to understand how a word is translated, it often fails to learn reasonable translations for words that it saw very few times in the training data. By combining the two systems, which excel in different areas, we see significant improvements in translation quality over either of the individual methods. 

The hybrid system will initially be launched for the English and Russian language pair, which accounts for 80 percent of the tens of millions of daily Yandex.Translate requests. The Yandex.Translate team also hopes to add other language pairs in the near future. 

Yandex’s new Head of Machine Translation, David Talbot explains, “We are excited to launch our new hybrid system for Yandex.Translate users. Ultimately, we want to develop a deeper understanding of how we can better assist Yandex users with their language needs, be it communication, language learning or simply accessing the huge amounts of information on the web available in other languages.”

Currently, Yandex.Translate offers users text, speech, and image translation and supports 94 languages pairs. Start using Yandex.Translate today at https://translate.yandex.ru/ or download the mobile app for iOS and Android.

Yandex challenges language barriers

Yandex has another go at language barriers to make life more comprehensible for iPhone users all over the world. Our Yandex.Translate app available for free in Apple App Store now has the English user interface. Yay!

The app, which was first rolled out in December 2012 with the Russian user interface was primarily designed for mobile users in the Russian-speaking countries and offered translations between Russian and eight languages including English, Ukrainian and Turkish. But with the brand-new English user interface and more languages now any speaker of English in the world with an iPhone in their pocket can read emails, newspaper articles, instructions, signs or even novels translated from French, Italian, Spanish, German, Turkish, Ukrainian or Czech. And, conversely, a speaker of any of these languages can use the Yandex.Translate app to have English texts translated for them into their language.

There’s no need to manually pick the right language for translation in the app’s settings – Yandex.Translate detects the source language automatically. A rare feat for mobile translators.

Together with announcing the English user interface, Yandex releases a Ukrainian version. The speakers of Russian or Ukrainian have historically had a bit more choice. For them, the app could translate words, sentences and entire texts between these two languages, as well as between either of them and English, Polish, German, French, Spanish, Italian, Czech, Bulgarian, Romanian, Serbian and Turkish.

The Yandex.Translate app for iPhone answered popular demand for mobility and followed our web-based translation service at first launched in March 2011 and used the same machine translation technology based on statistical regularities rather sets of rules.

Just like the web-based service, the Yandex.Translate app can translate words simultaneously – as you type them in the input field, and also offers suggestions for quick and flawless input, which is especially convenient when you are on the go. Predictive typing is indispensable on the mobile phone, when the user has to solve their typing tasks on a tiny screen and with a touch keyboard. Yandex.Translate is the only automated translation app that implemented a proprietary predictive text technology to offer its users a painless input experience.

When translating individual words, Yandex.Translate displays both the definition and the full dictionary entry including comments and usage examples.

In addition, using a text-to-speech technology the app can sound words in the entry titles for dictionary entries in Russian and Turkish when translating English words, and for dictionary entries in English, Italian, Spanish, German, French, Polish and Turkish if the source word is Russian. Not only can you learn the translation, but also hear what the word sounds like in one of these languages. Could score you points or get you out of trouble at some point. You never know.

 

While the app can only function on the internet-enabled iPhone, translation history is stored on the device and can be accessed while offline.

With over 150,000 downloads since its launch in December 2012, the Yandex.Translate app sees on average almost 40,000 downloads every month and serves over 15,000 users making more than 100,000 translations every day.

The majority of these translations – over 85% – are between Russian and English. Although right now 98% of all downloads come from Russia and Ukraine, of course, there is considerable interest from mobile users in the US, China, France, the UK and Israel, which will hopefully only grow now that we can offer our international users an international user interface.

The Yandex.Translate App for iPhone is available for a free download from the App Store and does not require a Yandex user account.