Select your localized edition:

Close ×

More Ways to Connect

Discover one of our 28 local entrepreneurial communities »

Be the first to know as we launch in new countries and markets around the globe.

Interested in bringing MIT Technology Review to your local market?

MIT Technology ReviewMIT Technology Review - logo


Natural Language Processing

Karen Jensen

The 1968 film 2001: A Space Odyssey gave us a vision of the millennium based on the technological predictions of the day. One result: HAL 9000, a computer that conversed easily with its shipmates like any other crew member. The timing was off: In the real 2001, there’s not a computer in the solar system as articulate as HAL.

But maybe it wasn’t that far off. HAL’s modern-day counterparts are catching up fast (sans the homicidal tendencies, one hopes). Already we have commercial speech recognition software that can take dictation, speech generation equipment that can give mute people voices and software that can “understand” a plain-English query well enough to extract the right answers from a database.

Emerging from the laboratories, moreover, is a new generation of interfaces that will allow us to engage computers in extended conversation–an activity that requires a dauntingly complex integration of speech recognition, natural-language understanding, discourse analysis, world knowledge, reasoning ability and speech generation. It’s true that the existing prototypes can only talk about such well-defined topics as weather forecasts (MIT’s Jupiter), or local movie schedules (Carnegie Mellon’s Movieline). But the Defense Advanced Research Projects Agency (DARPA) is working on wide-ranging conversational interfaces that will ultimately include pointing, gesturing and other forms of visual communication as well.

Parallel efforts are under way at industry giants such as IBM and Microsoft, which see not only immediate applications for computer users who need to keep their hands and eyes free but also the rapid evolution of speech-enabled “intelligent environments.” The day is coming when every object big enough to hold a chip actually has one. We’d better be able to talk to these objects because very few of them will have room for a keyboard.

Getting there will be a huge challenge-but that’s exactly what attracts investigators like Karen Jensen, the gung-ho chief of the Natural Language Processing group at Microsoft Research. Says Jensen: “I can’t imagine anything that would be more thrilling, or carry more potential for the future, than to make it possible for us to truly interact with our computers. That would be so exciting!”

Such declarations are typical of Jensen, who at 62 remains as exuberant about technology’s promise as any teenager-and just as ready to keep hacker’s hours. Indeed, Jensen was one of the first people Microsoft hired when it opened its research lab in 1991. Along with colleagues Stephen Richardson and George Heidorn, she arrived at the Redmond, Wash., campus from IBM’s Thomas J. Watson Research Center, where they had worked on some of the earliest grammar-checking software, and immediately started building a group that now numbers some 40 people.

In Redmond, Jensen and her colleagues soon found themselves contributing to the natural-language query interface for Microsoft’s Encarta encyclopedia and to the grammar checker that first appeared in Word 97. And now, she says, they’ve begun to focus all their efforts on a unique technology known as MindNet. MindNet is a system for automatically extracting a massively hyperlinked web of concepts from, say, a standard dictionary. If a dictionary defines “motorist” as “a person who drives a car,” for example, MindNet will use its automatic parsing technology to find the definition’s underlying logical structure, identifying “motorist” as a kind of person, and “drives” as a verb taking motorist as a subject and car as an object. The result is a conceptual network that ties together all of human understanding in words, says Jensen.

The very act of putting this conceptual network into a computer takes the machine a long way toward “understanding” natural language. For example, to figure out that “Please arrange for a meeting with John at 11 o’clock” means the same thing as “Make an appointment with John at 11,” the computer simply has to parse the two sentences and show that they both map to the same logical structures in MindNet. “It’s not perfect grokking,” Jensen concedes. “But it’s a darn good first step.”

MindNet also promises to be a powerful tool for machine translation, Jensen says. The idea is to have MindNet create separate conceptual webs for English and another language, Spanish, for example, and then align the webs so that the English logical forms match their Spanish equivalents. MindNet then annotates these matched logical forms with data from the English-Spanish translation memory, so that translation can proceed smoothly in either direction.

Indeed, says Jensen, who is now in the process of passing on the leadership of the group to the younger generation, MindNet seems to tie together everything they’ve been doing for the past nine years: “All we see is doors opening. We don’t see any closing!”

Others in Language Processing

Organization Project

Victor Zue (MIT Laboratory for Computer Science)

Conversational interfaces

Alexander I. Rudnicky (Carnegie Mellon)

Verbal interaction with small computers

Ronald A. Cole (University of Colorado)

Domain-specific conversational systems

BBN Technologies (Cambridge, Mass.)

Dialog agent