Monday, August 8

Google wants to take its autocorrect to another level: this is its technology to correct our grammar

The autocorrect has been with us for years, showing suggestions or replacing those words that we have misspelled. With the arrival of the Google Pixel 6 and the new neural engine of your chip tensor, Google wants to take this autocorrect even further, analyzing the grammar of the sentences we write.

A grammar correction instantly and locally

Google wants to improve the Android autocorrect (on the Gboard keyboard) and, for this, it has trained a neural network that takes an input phrase and executes its grammatically correct version. In the event that the input phrase is correct, the model does not make any correction and offers an identical output to that input.

To train this model, data is required in the form of text pairs (original text – corrected text), for which they have used a huge database with millions of phrases from all over the web. This first set of sentence pair training data is then used to train the final model itself, capable of correcting entire sentences.

In the same way they indicate that, so that the model can correct the grammar in real time, it is necessary to analyze the prefixes of the phrases, since these determine a good part of how the rest is to be written.

“This raises the question of how to decide if a given phrase prefix is ​​grammatically correct. We use a heuristic to solve this: if a given phrase prefix can be completed to form a grammatically correct phrase, we consider it correct. If it is not, it is assumed which is incorrect. ” Google.

Tg Image 4272129444

To achieve this, Google you have created a second dataset centered on the prefixes themselves. In this case, they give us an example in English of two pairs of sentences: the misspelled phrase ‘She puts a lot of effort yesterday afternoon’ and the corrected phrase ‘She put in a lot of effort yesterday afternoon’. Although it is an example in English, this prefix analysis will be applied in the different languages.

According to Google, the key to this model is that everything is done locally. It is programmed on the device itself and, when we use Gboard, the app sends a request to the grammar model itself every time the user types more than three words. This model occupies 22 megabytes of storage and is managed by the phone’s own CPU.

More information | Google