Google AI recently introduced a new feature called 'Proofread' to Gboard, the popular mobile keyboard application. With a single tap, this feature allows for seamless modifications and corrections at the sentence and paragraph levels, improving the overall typing experience.
Tackling the 'Fat Finger' Problem
Gboard uses statistical decoding as a means of addressing the "fat finger" issue, which is the natural inaccuracy of touch input on small screens. Research shows that the error rate per letter can reach 8 to 9 percent without decoding. Gboard comes with several error-correcting features to make typing easier and smoother. These features combine manual tasks involving human input with proactive, automated corrections.
Comprehensive Error Correction Tools
Typing is made easier collectively by Gboard's features, which include word completion, next-word predictions, active auto-correction (AC), and active key correction (KC). These tools include smart composition, error correction, and multiple-word candidates in the suggestion bar or inline. Post-correction (PC) also facilitates the rectification of errors in words that have already been committed. Despite these features, two key restrictions impair the user experience.
- While on-device correction models such as KC, AC, and PC are fast and compact, they have difficulty handling sophisticated and complex errors that need longer contexts. To prevent these models from being activated, users must type slowly and precisely.
- Typing speed decreases as a result of users having to manually correct committed words using grammar and spell checkers, which can be mentally and visually taxing.
The Need for High-Level Error Correction
Fast typers frequently disregard words they have already typed and concentrate just on the keyboard. Higher error rates may arise from this, necessitating the use of sentence or higher-level correction mechanisms.
Introducing Proofread
Proofread solves these frequent complaints by offering notable productivity boosts. It makes error correction simpler by providing one-tap sentence- and paragraph-level corrections.
The Technology Behind Proofread
The Proofread feature is made up of four primary components: data production, metrics design, model tweaking, and model serving. Together, all of these components guarantee effectiveness. To ensure that data distribution is in line with the Gboard domain, the system mimics frequent keyboard errors using a precisely and methodically built error synthetic architecture.
Advanced Metrics and Model Optimization
In order to assess the model, researchers have incorporated a number of metrics, with a particular emphasis on grammatical error presence checks and similar meaning checks derived from large language models (LLMs). Supervised fine-tuning and Reinforcement Learning (RL) tuning are used in the InstructGPT technique to improve the model. The model's proofreading performance is greatly improved by this procedure.
Deployment and Performance
Using 8-bit quantization, the medium-sized LLM PaLM2-XS is tailored to fit inside a single TPU v5 and serves as the foundation for the Proofread feature. The optimization lowers the cost of serving. Prior research has demonstrated that segmentation, speculative decoding, and bucket keys can reduce latency.
Real-World Impact
Tens of thousands of Pixel 8 users will profit from the high-quality edits provided by the Proofread feature, which is now available. A thorough generation of synthetic data and several iterations of fine-tuning have produced a model that substantially lowers grammatical errors, as seen by a 5.74 percent relative decrease in the PaLM2-XS model's Bad ratio. Additionally, optimizations have resulted in a 39.4% decrease in median latency.
Future Prospects
This study demonstrates how LLMs might enhance user experience and create intriguing possibilities for further investigation. The use of real-user data, multilingual support, tailored writing assistance, and device privacy solutions are a few examples of areas that could lead to further advancements in the sector.
Google AI is expanding the possibilities of mobile keyboard technology with the introduction of the Proofread feature, improving typing for people all across the world.