Steven Levy, the author of the book “Hackers” that made me want to be a computer scientist at age 12, was invited into Building 43 at Google headquarters to talk to Amit Singhal, who re-wrote their search function in 2001. There’s a fascinating quote on their use of semantics that gives more evidence that they are very closely studying factors that, when combined with their massive amounts of data in all languages and a growing selection of TMs, would lead to the first workable MT:
Take, for instance, the way Google’s engine learns which words are synonyms. “We discovered a nifty thing very early on,” Singhal says. “People change words in their queries. So someone would say, ‘pictures of dogs,’ and then they’d say, ‘pictures of puppies.’ So that told us that maybe ‘dogs’ and ‘puppies’ were interchangeable. We also learned that when you boil water, it’s hot water. We were relearning semantics from humans, and that was a great advance.”
But there were obstacles. Google’s synonym system understood that a dog was similar to a puppy and that boiling water was hot. But it also concluded that a hot dog was the same as a boiling puppy. The problem was fixed in late 2002 by a breakthrough based on philosopher Ludwig Wittgenstein’s theories about how words are defined by context. As Google crawled and archived billions of documents and Web pages, it analyzed what words were close to each other. “Hot dog” would be found in searches that also contained “bread” and “mustard” and “baseball games” — not poached pooches. That helped the algorithm understand what “hot dog” — and millions of other terms — meant. “Today, if you type ‘Gandhi bio,’ we know that bio means biography,” Singhal says. “And if you type ‘bio warfare,’ it means biological.”
And later, when talking about search algorithm failures, Google discovers they have problems from language mistakes:
Google is famously creative at encouraging these breakthroughs; every year, it holds an internal demo fair called CSI — Crazy Search Ideas — in an attempt to spark offbeat but productive approaches. But for the most part, the improvement process is a relentless slog, grinding through bad results to determine what isn’t working. One unsuccessful search became a legend: Sometime in 2001, Singhal learned of poor results when people typed the name “audrey fino” into the search box. Google kept returning Italian sites praising Audrey Hepburn. (Fino means fine in Italian.) “We realized that this is actually a person’s name,” Singhal says. “But we didn’t have the smarts in the system.”
The Audrey Fino failure led Singhal on a multiyear quest to improve the way the system deals with names — which account for 8 percent of all searches. To crack it, he had to master the black art of “bi-gram breakage” — that is, separating multiple words into discrete units. For instance, “new york” represents two words that go together (a bi-gram). But so would the three words in “new york times,” which clearly indicate a different kind of search. And everything changes when the query is “new york times square.” Humans can make these distinctions instantly, but Google does not have a Brazil-like back room with hundreds of thousands of cubicle jockeys. It relies on algorithms.
…the hard-won realization from inside the Google search engine, culled from the data generated by billions of searches: a rock is a rock. It’s also a stone, and it could be a boulder. Spell it “rokc” and it’s still a rock. But put “little” in front of it and it’s the capital of Arkansas. Which is not an ark. Unless Noah is around. “The holy grail of search is to understand what the user wants,” Singhal says. “Then you are not matching words; you are actually trying to match meaning.”
The whole article is more technical, although easy to read. Will be fascinated to see how long it takes until they finally get something working as a human replacement MT.