MOUNTAIN VIEW, California — Google is making a big change in the way it presents search results. Most people won’t notice, but, Google executives said, the new technology represents a leap forward in the company’s ability to understand what people are asking for when they search.
Google’s search engine uses natural language processing to analyze queries. For example, for the entry “get medicine at pharmacy,” Google will home in on the words “medicine” and “pharmacy” and look for pages related to those two keywords.
Starting today, the company is rolling out a new machine learning–powered language processing method called Bidirectional Encoder Representations From Transformers, or BERT for short. BERT looks at the sequences of words in searches — not just the words themselves — to glean more information on the intent behind them.
For example, results for “can you get medicine for someone pharmacy” would previously have served a link to a 2017 MedlinePlus article about getting a prescription filled, and missed the point that the search was looking for information on how to pick up a prescription for someone else. Using BERT, Google’s search engine now shows a 2002 article from the Department of Health and Human Services about how to have a friend or family member pick up the medicine on your behalf.
Pandu Nayak, Google’s vice president of search, gave another example at a press event yesterday, using the query “How old was Taylor Swift when Kanye went on stage?” Before BERT, Google surfaced videos of the 2009 event during which the rapper interrupted the pop star’s acceptance speech at the MTV Video Music Awards. After BERT, Google presents as its first result a snippet from a BBC article, which states: “A 19-year-old Swift had just defeated Beyoncé to win Best Female Video for her country-pop teen anthem You Belong With Me.” Google's search returns automatically highlighted “19-year-old” for emphasis.
"On the ranking front, this is the single biggest change we’ve had in the last five years — and one of the biggest from the beginning," said Nayak.
While most people will probably not realize a new natural language processing technology is at work, behind the scenes, the change will most likely impact websites that rely on Google for traffic. The tech giant controls 90.8% of the search market across all its products, including YouTube and Google Maps. Google, through its Search and News tools, drives the majority of traffic to articles on news sites, especially those related to politics and technology.
In response to a reporter’s question on whether Google saw significant changes in the traffic it sent to any particular websites in tests of BERT, Ben Gomes, vice president of core search, said, “We did not see any particular pattern there. But if you have niche questions, then a niche publisher might be surfaced for that.”
The Google executives didn’t say whether websites should expect to see more or less traffic. Gomes did add, however, that he expected improving the feature for users would lead to more searches, which would bring more traffic to all websites. “As we answer more exotic questions, hopefully that will lead to people asking more and more exotic questions,” he said.
BERT, Nayak said, was released to improve more complex searches, and would affect just 1 in 10 searches in the United States. Still, with billions of searches a day, that’s hundreds of millions of results that have changed overnight.
There are areas that BERT doesn’t handle well, Nayak said. In one example, for the search “tartan,” BERT promoted dictionary results because it’s a technology that focuses solely on text. Pre-BERT searches showed images of tartan fabric, which is a more compelling result. A Google spokesperson said the tartan issue was corrected before BERT's launch.
In another case, when one searches for “What state is south of Nebraska?” BERT surfaces the Wikipedia page for South Nebraska, as opposed to the ideal result, which would be the Wikipedia page for Kansas.
People who use Google won’t know whether their results are powered by BERT, and can’t revert to non-BERT results. Moving forward, BERT will be applied to all US searches. Nayak said he is confident that, in general, BERT will improve results for more complicated queries: “We’re playing a statistical game here. In aggregate, we know that any change will have some gains and some losses. ... I think the payoff [of BERT] is very much there.”
Initially, BERT will only be applied to searches in English in the US.