Bing has reported that it has been using BERT since April so much so that natural language processing capabilities are now applied to all queries.
Bing has been using BERT for months to improve the quality of search results and to report it directly to Microsoft, so much so that the transformer models have been applied to all the various types of requests made by users to the search engine.
In the example above, as you can read, Bing’s best natural language processing capabilities allow the search engine to understand what the user is looking for online. Specifically, Bing was asked, “what can aggravate a concussion”.
Before the addition of the new transformer models, the results to the question provided answers regarding the symptoms of a concussion, while now they provide more comprehensive and specific answers.
If you look at the before and after April you can see how Bing has started to give more and more precise answers according to the research made by the user.
A better understanding of natural language should lead to more relevant search results for users.
For a few months now, we have known that both Bing and Google are using these advanced algorithms to deliver better search results, especially those involving longer queries.
Today we can say that Bing’s implementation of BERT, to improve its search results, precedes Google’s BERT announcement by six months and Bing says this has led to further improvements in search quality in the last year.
How Bing’s BERT works
“Unlike previous architectures deep neural networks (deep neural network or DNN) that elaborated the words individually in order, the new transformer model include the context and the relationship between each word and all the words that surround it in a sentence.”
This is said by Jeffrey Zhu, program manager of Bing Platform, who adds that “starting in April of this year, we have been using large transformer models to provide our customers with further quality improvements over last year”.
BERT builds on the deep learning capabilities used in Bing’s intelligent search features, such as intelligent answers that tap into multiple sources, intelligent image search with object recognition, and hover-over definitions for uncommon words.
These improvements can help Bing maintain or increase its search market share, which will continue to make it a viable platform for both organic and paid campaigns.
Microsoft implemented the transformer models using Azure N-series Virtual Machines with GPU accelerators and then made further optimizations to perform parallel computing on a web-search scale.