Friday, March 29, 2024

Google explains how machine learning improved its Gboard mobile keyboard

Share

Why it matters to you

Google’s machine learning techniques predict how you type in Gboard, it’s third-party mobile keyboard.

Gboard, Google’s mobile keyboard for iOS and Android, is almost like a digital secretary. It can predict what you’re going to type before you type it, perform Google searches, translates in real time between more than 15 different languages, and boasts a growing library of handy emojis. And on Thursday, Google researchers pulled back the curtain on efforts to make it better.

“Most people spend a significant amount of time each day using mobile-device keyboards: Composing emails, texting, engaging in social media, and more. Yet, mobile keyboards are still cumbersome to handle,” Google said. “The average user is roughly 35 percent slower typing on a mobile device than on a physical keyboard. To change that, we recently provided many exciting improvements to Gboard for Android, working towards our vision of creating an intelligent mechanism that enables faster input while offering suggestions and correcting mistakes, in any language you choose.”

In a post on the Google Research Blog, Google explained how Gboard is using machine learning to cut down on errors. Using neural networks — computer systems modeled on the human brain and nervous system — researchers were able to correct for for mistyped letters, misspellings, character insertions, and deletions. “[Gboard] address[es] these errors at the character level, mapping the touch points on the screen to actual keys,” Google said.

In addition, Google employed TensorFlow, its hardware-accelerated machine learning platform, to train hundreds of neural networks and optimize them for suggestions, completions, and other keyboard-specific features. After a year of work, the models were six times faster and 10 times smaller than the initial versions, Google said, and showed a 15 percent reduction in bad autocorrects and 10 percent reduction in “wrongly decoded gestures.”

Those changes were just the beginning. Google factored lexicon, which tells which words occur in a language, and probabilistic grammar, which tells what words are likely to follow other words, into Gboard’s machine intelligence. The result was what Google calls Finite-State Transducers, which power the app’s natural language processing and supply word completions and predictions.

Google used the same Finite-State Transducers to map sequences of Latin keys to symbol sequences in Indic languages. The result? Gboard lets you type a word according to its phonetic pronunciation — the Hindi “daanth” for “दांत” (teeth) — and automatically transliterates it. Compared to Gboard’s old system, the new models are 50 percent faster, and reduce the fraction of words users have to manually correct by more than 10 percent. “Some languages have multiple writing systems […] so between transliterated and native layouts, we built 57 new input methods in just a few months,” Google said.

“While we hope that these recent changes improve your typing experience, we recognize that on-device typing is by no means solved. Gboard can still make suggestions that seem nonintuitive or of low utility and gestures can still be decoded to words a human would never pick, Google said. “However, our shift towards powerful, machine-intelligence algorithms has opened new spaces that we’re actively exploring to make more useful tools and products for our users worldwide.”




Read more

More News