It is a Google tool that we use on a daily basis and perhaps that is why it goes unnoticed, but it is really useful. We are talking about predictive text, a functionality that has been in smartphones for more than a decade thanks to its incorporation into the different keyboards of almost all operating systems.
And now Google explains to us how its model works. The selective text prediction (Smart Text Selection) allows us to write faster avoiding having to type words that come next. For that, the system is in charge of carrying out a study of our habits, something that Google has enhanced through machine learning that serves to feed an algorithm thanks to the data sent by all our devices from the Google Keyboard. They call it federated learning and thanks to it and according to Google, the accuracy of the model has been improved by up to 20%.
It’s not magic, it’s thanks to an algorithm
All of our phones and tablets form a huge network that Google’s system draws on to improve automatic text prediction. At the time of writing we see how the Google keyboard anticipates words that we are going to use later and it is not magic. It is the result of the work of an algorithm that analyzes the data of millions of devices.
Now Google explains in its blog dedicated to Artificial Intelligence how this system works to make text auto-completion even faster and more efficient. An algorithm is responsible for analyzing all the words that correspond to addresses or numbers, to be clear where the text selection prediction should cut.
The improvements achieved by Google come thanks to federated learning, a model in which users and our devices serve as sparring for the system to learn. All data generated by Google keyboard devices is sent ensuring privacy and integrity at all times: on the one hand, no raw data is sent, only small updates on the model and also “the data of this network is protected by policies that restrict their use. “
The data that the keyboard collects from users’ day-to-day lives are sent to Google’s servers when we connect the phone, preferably at night. This is how Google sums up how it works:
- The server begins by initializing the model.
- Devices are selected.
- The selected devices improve the model using their local data and then send only the improved model to the server, which returns the information to the device, although data used for training from other phones is not returned.
According to Google, federated learning allows user privacy, because the raw data is not exposed to a server. Instead, the server only receives updated model weights. In addition, according to the company, thanks to this system it is possible to improve the selection process of several words between 5 and 7%, and between 8 and 20% improvement in the specific case of the selection of addresses.