Thursday, May 25, 2017

Google Talks About the Machine Learning Technology Behind Gboard

Google has been heavily investing in machine learning and neural network technology for a few years now. Once it was good enough to be announced, we started seeing Google apply various machine learning techniques to many of their applications and services.

So as Google's machine learning technology gets better, we will continue to see them figure out ways to implement it into their products. One such area that they've seen a lot of success in is their language translation services. Google Translate is now more accurate than ever before, and it's all thanks to their machine learning technology. Recently, they started to experiment with machine learning for their Gboard keyboard application and they're happy with the results so far.

In a new post over on the Google Research Blog, Google spoke about how they're using machine learning to improve the user experience of people using Gboard. With Neural Spatial Models, Google is able to address the 'fat finger' issue that some people experience when typing on a smartphone keyboard. They've been able to detect these errors and predict the intended words rapidly and accurately at the character level as they map the touch points on the screen to actual keys.

Google also talks about Finite-State Transducers and how they're using a lexicon to figure out what words are likely to come next in a sentence. The company is using these FSTs to represent various probabilistic models (so lexicons, grammars, and normalizers for example) that are used in natural language processing. They put this together with a mathematical framework they've come up with to manipulate, optimize, combine and search the models.

Google talks about these things and additional techniques in more detail in the blog post linked below. They've been able to cut the decoding latency by 50%, reduced the number of words people have to manually correct by 10%, and they're working to enhance Gboard in new and unique ways.

Source: Google Research Blog



from xda-developers http://ift.tt/2qju1wR
via IFTTT

No comments:

Post a Comment