GOOG Alphabet Inc

By Rob Copeland 

MOUNTAIN VIEW, Calif. -- Google rarely talks about its secretive search algorithm. This week, the tech giant took a stab toward transparency, unveiling changes that it says will surface more accurate and intelligent responses to hundreds of millions of queries a day.

Top Google executives, in a media briefing Thursday, said they had harnessed advanced machine learning and mathematical modeling to produce better answers to complex search queries and questions that often confound its current algorithm. They summed up the changes under a new sequence of code called "Bert," as in "Bidirectional Encoder Representations from Transformers," as among the largest improvements to search in half a decade.

Google, a unit of Alphabet Inc., said one in 10 queries would see better results. While Google doesn't disclose the exact volume of activity on its ubiquitous search platform, outside searchers peg daily search volume around 3.5 billion, which would mean some 350 million daily searches would see changes.

The new algorithm will also be used on advertising, one executive said.

Thursday's briefing offered a rare window into Google's continuous effort to fine-tune the software at the core of its $136 billion-a-year advertising business. Here's an example of how the changes work, per Google: Take the search for "do estheticians stand a lot at work." Searches like these have long proved difficult for a computer to parse, because "stand" carries multiple meanings, and "at work" is essentially slang. Under Google's old system, the top search result bordered on irrelevant: It was a local newspaper article on free-standing beauty schools.

Bert, however, is able to strip through the complexity, in part by ignoring transition words and other noise that may be grammatically necessary in speech but tangential to the actual substance of the query. Indeed, in the estheticism example, the new system produced a highly relevant result on the physical demands of such jobs.

"You have to figure out how to transform what the user has typed in, into the actual documents that they were looking for," said Google search head Ben Gomes.

The new system isn't perfect. Take the search "What is south of the state of Nebraska?" Google's old algorithm came close to the right answer, but Bert overthinks it, serving up a link to the tiny community of South Nebraska, Florida.

Google search executive Pandu Nayak said the company will continue to employ thousands of manual testers to ferret out results like that.

The search changes are being rolled out now, though they may take some time to go live, depending on a user's location, Google said.

Google, the world's biggest online advertising platform, by far, wasn't eager to discuss the new system's implications for its bread-and-butter business. At the briefing, executives showed sample screenshots of potential searches that didn't include the multiple lines of advertising that appear above so-called organic search results for myriad queries on Google's actual website.

Mr. Nayak said he wasn't able to explain exactly how the advertising team uses the Google-developed machine-learning methods. "I'm sure they must be applying it, I just don't know" the details, he said.

A spokeswoman then interjected, "We can follow up." A spokeswoman later said in a statement: "We're exploring how BERT can be useful for our various products independently. We have not used BERT for any Search ads."

Google executives did address a persistent criticism, from lawmakers and advertisers alike, that the company has steadily transformed its search-engine over time to keep more traffic on Google's own pages.

For English queries, Google says, the new algorithm won't change the proportion of queries that are answered directly on Google.

For foreign languages, Google may yield more search answers because the improved algorithm boosts the company's ability to answer complex topics in a variety of tongues.

Write to Rob Copeland at rob.copeland@wsj.com

(END) Dow Jones Newswires

October 25, 2019 03:15 ET (07:15 GMT)

Copyright (c) 2019 Dow Jones & Company, Inc.