Google AI introduces proofreading: a new Gboard feature that enables seamless corrections at the sentence and paragraph level with a single tap

Gboard, Google's mobile keyboard app, works on the principle of statistical decoding. This approach is necessary because of the inherent inaccuracy of touch input, often called the “fat finger” problem, on small screens. Studies have shown that without decoding, the error rate for each letter can be as high as 8 or 9 percent. To ensure a smooth typing experience, Gboard incorporates a variety of error correction features. Some of these features are active and automatic, while others require the user to perform additional manual actions and selections.

Word completion, next word predictions, active autocorrect (AC) and active key correction (KC) work together to make typing easier for the user by correcting errors and offering multiple candidate words in the suggestion bar or online . as well as intelligent writing. Correction of errors in the last or more committed words is supported by post-correction (PC).

When it comes to user experience, current rectification methods on Gboard have two distinct limitations. First, on-device correction models, such as active key correction (KC), active self-correction (AC), and post-correction (PC), are compact and fast, but have problems with more complex errors that require contexts of longer duration. As a result, users still need to type slowly and accurately to avoid triggering these models. Additionally, users must systematically repair words they make using grammar and spelling checkers, two of the multi-step passive correction capabilities. This process can be mentally and visually demanding, as users must carefully monitor their words and correct mistakes sequentially after making them. This may cause a decrease in writing speed. A common strategy among Gboard users who type quickly is to ignore the words they've already typed and focus solely on the keyboard. People who are “quick and sloppy” when writing and then move on to higher-level bug fixes sometimes ask for a sentence or a higher-level fix function to help them.

A new feature called Correction was introduced in a recent Google study. This feature is designed to address the most common complaints of fast typists, providing a significant boost to your productivity. It offers one-touch sentence and paragraph level problem repair, making it easy for users to fix errors in their text. The field of grammatical error correction (GEC), which includes proofreading, has a rich history of study spanning rule-based solutions, statistical methods, and neural network models. Large language models (LLMs) have incredible growth capacity, presenting a new opportunity to find high-quality fixes for sentence-level grammar.

The system behind the Proofread feature is made up of four main components: data production, metric design, model tuning, and model serving. These components work together to ensure the effectiveness of the function. Various procedures are performed to ensure that data distribution is as close to the Gboard domain as possible. This is achieved through a meticulously constructed error synthetic architecture that incorporates commonly made keyboard errors to mimic user input. The researchers have included several measures covering different aspects to further evaluate the model. Since responses are never truly unique, especially in long examples, the metric is considered the most important statistic for comparing model quality, along with checking for existence of grammatical errors and checking for same meaning based on LLMs. Finally, to make the LLM dedicated to the review function, they applied the InstructGPT approach of using supervised fine-tuning followed by reinforcement learning (RL) tuning. The proposed formula for reinforcing learning and adapting rewriting tasks was found to greatly improve the revision performance of the basic models. They build their feature on the medium-sized LLM PaLM2-XS, which can be accommodated in a single TPU v5 after 8-bit quantization to reduce the cost of serving.

Previous studies show that latency is further improved by using segmentation, speculative decoding, and escrow keys. Now that the proposed model is available, tens of thousands of Pixel 8 consumers will reap the benefits. Careful production of synthetic data, many phases of supervised fine-tuning and RL tuning allow us to achieve a high-quality model. Researchers suggest Global Reward and Direct Reward in the RL adjustment stage, which greatly improves the model. The results demonstrate that RL adjustment can effectively decrease grammatical errors, leading to a 5.74 percent relative reduction in the Bad index of the PaLM2-XS model. After optimizing the model through quantization, buckets, input segmentation, and speculative decoding, they deploy it on TPU v5 in the cloud with highly optimized latency. According to the findings, speculative decoding reduced the average latency by 39.4 percent.

This study not only demonstrates the innovative nature of LLMs for improving UX, but also opens up a world of interesting possibilities for future research. Using data from real users, adapting to multiple languages, providing personalized support for different writing styles, and developing solutions that protect privacy on devices are areas that could be explored, generating new ideas and innovations in the field.

Review the Paper. All credit for this research goes to the researchers of this project. Also, don't forget to follow us on twitter.com/Marktechpost”>twitter. Join our Telegram channel, Discord Channeland LinkedIn Grabove.

If you like our work, you will love our Newsletter..

Don't forget to join our 44k+ ML SubReddit