Google Says Latest AI Tweak Will Improve Search Results

Google says it has improved its search results by using artificial intelligence that can better understand natural language.

The change, announced Friday, will being phased in for English-language web searches in the U.S., with other languages expected to debut at a later date.

“Language understanding is key to everything we’re doing on search,” said Pandu Nayak, Google fellow and vice president of search. “This is the single, biggest, most positive change we’ve had in last five years.”

Google’s goal is to make it easier for users, who often don’t know how to enter queries for the information they want. Since its search engine debuted in 1997, Google has focused on getting its technology to better understand natural language to produce relevant results even in cases where users enter a misspelled word or a query that is off target.

With the latest change, Google will also now consider the sequential order in which words are placed in a search, instead of returning results based on a “mixed bag” of keywords. The search engine will also take into account smaller words like prepositions and articles in a query—words that it previously ignored.

The technology responsible for the change is an A.I. model Google unveiled last year called BERT, or Bidirectional Encoder Representations from Transformers. Incorporating it into search is the first time the company has applied it to a mainstream product.

Google says one in every 10 searches will be improved by BERT, while also admitting that some search results will actually get worse—a problem it’s still trying to fix.

For example, Google found through testing that when users searched for “What state is south of Nebraska?” they were led to a page about the Florida city of South Nebraska. Previously, Google would return a Wikipedia page about Nebraska that would include information on bordering states.

Google was unable to say, broadly, which searches would get worse.

“We’re really playing a statistical game here,” Nayak said. “We know in aggregate, there will be some wins and some losses.”

The wins, according to Google, will help some users find exactly what they’re looking for rather than information related to the keywords.

For example, users who previously searched for whether they can get medicine for someone at a pharmacy would’ve previously gotten results listing pharmacies but not the information they were seeking. BERT allows Google to take into account that the user is looking for someone, and is able to directly answer the question by providing links to pages with relevant information.

As Google researches how to better process natural language, it expects to fix some of the errors that BERT currently spits out. It has already fixed an error in which its technology returns English-language results for queries in a different language rather than in the same language as the query.

“BERT is not the magic bullet that solves all problems,” Nayak said. “There is still more work to be done here.”