Yandex is the first to use neural networks to improve swipe text input. The technology allows to speed up typing and reduce the number of typos and grammatical errors by 30% compared to the previous version of the swipe.
When using a swipe, a person does not click on individual letters but swipes a finger over the virtual buttons without lifting it from the screen. Previously, the word that is hidden under the resulting curve was recognized using human-made formulas. Now these formulas create and constantly improve neural networks.
As a result, the keyboard simultaneously reads the drawn curve, selects the corresponding words, ranks them and determines the only suitable one, taking into account all possible probabilities. It also focuses on the previously typed text, the theme of the application in which everything happens, and knows how to choose the right endings and cases.
Head of the Yandex Keyboard team, Alexander Afanasiev says:
Swiping is one of the easiest ways to type really quickly on a touch screen. According to our data, experienced users create messages with gestures 60% faster than with spelling. Neuroswipe does not change the usual process of working with the keyboard, but at the same time it significantly increases the speed and accuracy of typing. Now it works in Russian, but over time we will add support for other popular languages.
To train Yandex Keyboard neural networks, the developers used 7 million examples of real swipes collected using the company’s internal crowdsourcing platform.