Google search receives Update: Better search results, thanks to BERT?

Alex Alex 28 October 2019
Google search receives Update: Better search results, thanks to BERT?

The Google search is now to lead in ten percent of cases to fit their results with the help of BERT, Google's Pre-Training method for NLP. The Update relates, however, to date only English-language search queries in the United States.

Google has improved the accuracy of search results or how it is called in the Google-AI-blog entry “We make in the history of the search, one of the biggest steps forward.“ To do this, BERT was used, a method for Pre-Training of Natural Language Processing (NLP), which was developed last year by Google. So far is supported by the new method, only the English search in the USA, but other languages and places to follow.

More Natural search queries

looking Particularly for longer sentences with prepositions is the new Google-to better deliver results and thus, a more natural sentence input. Have been revised, both the Ranking and Featured Snippets in highlighted boxes.

As an example for such a search Google calls "2019 Brazil traveler to the USA need a visa". The human Intention is whether a Person is from Brazil requires a visa to enter the US. This context is expressed by "to".

In the "old" search, to see to the left in the picture, it was understood that the search input is wrong, because the first result is related to entry requirements for U.S. citizens in Brazil. In the BERT-based search were spent, however, as desired, the entry requirements for Brazilian US visitors:

Google search without and with BERT; source: Google AI Blog

Context-making by BERT

The new Google search based on BERT (Bidirectional Encoder Representations from Transformers), a method for Natural Language Processing (NLP) Pre-Training. BERT was a year ago Google's Open Source and should enable one to develop systems for question answering.

The peculiarity lies in the fact that BERT to execute the complete word context of the preceding and the following word includes instead of the words according to their order. Due to the complexity of BERT was, according to Google, even new Hardware is required: For the first Time, the latest Cloud TPUs have been used in the Google for Machine Learning developed.

More Details to BERT and the new Google search, there are in the Google-AI-blog entry Pandu Nayak, Google Fellow and Vice President at Search.



Comments (0)

    No comments yet

You must be logged in to comment.