The current system introduced in February is using machine translation, and the process can be considered as “translating from the language of incorrect grammar into the language of correct grammar.”
A significant improvement leverages neural machine translation to “correct many more of the grammar errors you can create when writing.” For example, tension can be corrected and homonyms can be identified. A squiggly blue line highlights mistakes that you can right-click to accept changes.
Google uses Tensor Processing Units behind the scenes to run two Grammatical Error Correction models:
- The first method takes good sentences and makes them worse by automatically translating them to some other language and then back to English.
- The second method extracts source-target pairs from Wikipedia edit histories with a minimal amount of filtration.
These grammar improvements to Google Docs will be available in the weeks to come. Meanwhile, Docs will soon add auto-correct spelling with suggestions tailored to words that may be company-specific. This capability works in real time in Gmail.
According to Google, every week 100 million suggestions for grammar are flagged, with the company positioning these changes to help boost productivity through AI. Smart Compose has also saved users every week from typing more than 2 billion characters.