When I ask about the weather, my smartphone chirps back an answer: 鈥淚t doesn鈥檛 look so nice tomorrow.鈥�
That鈥檚 doubly unsurprising. One, it鈥檚 February in Maine during a historic stretch of snowy weather. Two, a phone is supposed to answer questions.
But here鈥檚 a surprise. We tend to think that new ideas are the fuel for today鈥檚 technology. Yet voice recognition is one of the 鈥渕any things that people have tried before that now work,鈥� says Susan Dumais 鈥�75, a Distinguished Scientist with Microsoft Research in Redmond, Wash.
A member of the American Academy of Arts and Sciences and the National Academy of Engineering, Dumais visited campus in February as the College Key鈥檚 Distinguished Alumna in Residence, meeting with students and giving a talk on Feb. 1.
College Key Distinguished Alumna in Residence Susan Dumais ’75 gives a talk on Feb. 1 in the Olin Arts Center Concert Hall in front of an image showing a bygone way to retrieve information: the card catalog. (Phyllis Graber Jensen/乐播传媒 College)
Back in the 1980s, when Dumais was at Bell Labs, one ongoing project was to get a computer to recognize correctly whether 鈥測es鈥� or 鈥渘o鈥� was answered to the question, 鈥淲ill you accept this collect call?鈥�
鈥淵ears and years and years of work went into that,鈥� Dumais says. The project did succeed in the 1990s (though collect calls are going the way of the typewriter).
So, what makes today鈥檚 鈥渘ew鈥� technology possible? She points to better computational models, more computational horsepower, and access to much richer data.
In the case of voice recognition, it means access to vast amounts of spoken language with corresponding transcriptions, which makes it 鈥渆asier to map various acoustic features to the resulting words,鈥� she says.
鈥淪o when you search for ‘Who is the best teacher?’ the results include pages that use the word ‘professor’ even if they don’t use the word ‘teacher.'”
Another old idea whose time has come belongs to Dumais herself. In 1990 鈥� way before Google or Bing 鈥� she and fellow researchers developed an information retrieval process known as 鈥渓atent semantic indexing.鈥� In recent years, LSI-related ideas have “been a big splash,鈥� she says.
That鈥檚 because LSI enables search engines to retrieve the information that we鈥檙e actually looking for. 鈥淪o when you search for ‘Who is the best teacher?’ the results include pages that use the word ‘professor’ even if they don’t use the word ‘teacher,’ ” Dumais explains.
Susan Dumais ’75 laughs with Professor Emeritus of Psychology John Kelsey before her talk on Feb. 1. Dumais first met Kelsey in the 1970s when she was a doctoral student at the University of Indiana, where he was starting his faculty career. (Phyllis Graber Jensen/乐播传媒 College)
Simply put, LSI helps a search engine respond to queries by 鈥渄iscovering words that are semantically the same even though they aren鈥檛 literally the same,鈥� she says.
The idea existed 25 years ago but 鈥渢he time wasn鈥檛 right,鈥� she says. Today, computers can swiftly learn these more nuanced word relationships thanks to the availability of much larger collections of text and more powerful computational resources.
Though advances in computing power are turning yesterday鈥檚 ideas into today鈥檚 tools, the demands of today’s tools are also keeping those ideas viable.
“How we use this technology 鈥� such as everyone carrying a smartphone 鈥� makes these ideas even more necessary now,鈥� Dumais says.
Students do pre-computer information retrieval, circa 1974, using the library’s card catalog. (Muskie Archives and Special Collections Library)
Categories Academic Life Alumni
