Science, about DLI-1, 266:20-22, October 1994

Turning an Info-Glut into a Library

An exploding amount of information has made the Internet into the digital equivalent of a used book store.
A new NSF program aims to bring some order to it.

In Rome, the Vatican is digitizing its library, capturing rare and one-of-a-kind materials in electronic form. Soon any scholar with a personal computer and a modem-not just the favored few with Vatican library privileges-will be able to pore over such papal treasures as the four oldest known manuscripts of Virgil's poems or a magnificently illustrated copy of The Divine Comedy. Meanwhile, in Ithaca, New York, Cornell University is saving thousands of crumbling books by recording them on optical disks. Now they will survive as facsimiles, available to be viewed on com-puter screens or resurrected as hard copies.

Around the globe, information that was once analog -- text, audio recordings, videos, maps, photographs, diagrams, scientific data -- is turning digital, and much of it is flooding out into the vast public space created by electronic networks such as the Internet. 'Almost overnight, we've gone from a high-mass, paper-based world to an almost massless, paperless world," says computer scientist Dan Atkins at the University of Michigan. But this digital cornucopia has a serious downside: The information has been growing much faster than the ability to keep up with it.

The Internet has come to resemble enormous used book store with volumes stacked on shelves and tables and overflowing onto the floor, and a continuous stream of new books being added helter-skelter to the piles. Recently some relief has appeared, such as the Mosaic software program that makes it easy to browse through much of the information available on the Internet by creating "links" between related items at separate sites (Science, 12 August, p.895). But even with Mosaic, finding what you're looking for is hit-or-miss. The solution, say Atkins and other experts, is to transform this used book store into one or several "digital libraries," complete with the electronic equivalents of neatly ordered shelves, a catalog, and a helpful staff.

Last week, the most ambitious effort so far to develop these virtual libraries got under way, with the announcement of major grants to six university-led consortia. Over 4 years, the National Science Foundation (NSF), together with the Advanced Research Projects Agency and the National Aeronautics and Space Administration, will spend $24.4 million to develop systems for collecting, storing, and organizing digital information and making it easily available to anyone who wishes to use it, from high school students to research scientists. Some of the consortia focus mainly on the problem of finding information scattered throughout cyberspace; others will concentrate on developing automated means to catalog new kinds of digital information, such as images and videos And by including libraries, museums, publishers, schools, and computer and communications companies in the consortia, NSF hopes to lay the foundation of an infrastructure that will serve all parts of society.

"We view building partnerships between researchers, applications developers, and users as essential to achieving success," explains Y. T. Chien, director of NSF's division of Information, Robotics, and Intelligent Systems. At the end of the 4 years, each of the six teams should have created a working digital library that serves a large public while also acting as a testbed for further research and development. And after that, NSF hopes to scale up the libraries, adding more information, more advanced information-handling tools, and more users.

NSF is looking well into the future with this ambitious project, but it is following the lead of other digital library projects, aimed at a more immediate payoff. One of the more innovative was announced in August by IBM and the Institute for Scientific Information (ISI). The companies plan to create a prototype digital library based on 1350 scientific journals in the life sciences, complete with a catalog that allows searching by jour-nal, author, title, subject matter, key words, and so on. That much is relatively standard on existing databases, but IBM is adding a twist: Instead of having every customer deal directly with one central library -- which would demand tremendous communications and computing capacity -- ISI will set up "branch libraries" for clients, such as universities or corporate research labs.

Each branch will be run on a server corn-putet attached to a local area network The server will hold a subcollection of the available works -- customized to the needs of that customer-plus a catalog of everything that can be found in the main collection, which would be kept at one central location. Anyone on the network can search the entire catalog and request copies of any article. If the article is held in the branch collection, it will appear on-screen immediately. If not, the user can ask the server to request a copy from a computer attached to the main collection.

The IBM-ISI group faces some tricky challenges. One is the problem -- common to all digital information sources -- of protecting intellectual-property rights. As IBM's Robert J. T. Morris notes, a customer who requests a copy of a journal article can be charged for it, but there is little to prevent that customer from distributing the articles electronically to dozens of colleagues who pay nothing. The library's developers will also have to tackle difficult technical jobs, such as devising a means of quickly searching through tens of thousands of abstracts.

The center cannot hold

But in one important respect those challenges are easier than the one facing the six NSF-funded consortia. The contents of the IBM-ISI library will all be in one location -- the main collection -- making it simple to keep an up-to-date catalog for distribution to the various branches. That's not possible for any digital library that takes advantage of the riches of the Internet, where pertinent information may be scattered across dozens or even thousands of locations.

The obvious solution might seem to be to keep a central catalog and have each unit -- the Vatican, Cornell, and every other information source encompassed by the library -- report its contents there. But it's not that easy. "The problem with centralization is that it doesn't scale," explains Robert Wilensky, a computer scientist at the University of California, Berkeley, and principal investigator for one of the six groups that won NSF grants. Keeping track of information from several locations or even several hundred might be workable, but the resources needed to maintain a central record grow much faster than the number of locations.

As a result, Wilensky's group at Berkeley and several of the other groups will be exploring an alternative strategy: inventing electronic helpers that can visit the locations and report on what's there. With no middle-man -- no central catalog or clearinghouse -- between the user and the information source, the result will be radically different from a traditional library, Wilensky says. Perhaps a better term for what he and the others are developing, he suggests, would be "digital nonlibraries."

One example of this kind of decentralization is the three-pronged strategy mapped out by a consortium centered at the University of Michigan. Atkins, the principal investigator, says his team will develop three types of software "agents," which will collaborate. The first will work with individual users, asking questions to determine their needs -- not just the information they seek, but how much detail they want and how much they're willing to pay for it. A second group will take users' requests and range over the Internet to find the

locations that can fill them. And the third group will be attached to individual information providers. Each of these agents can search its particular location for requested information and will know how much the information costs and whether it is restricted to certain users.

That approach brings the challenge of writing software that can -- in a reasonable amount of time and with acceptable accuracy -- search hundreds or perhaps thousands of locations, each with thousands of individual items or more. Making the job more difficult is the great variety of items within each collection. Many will be text or numerical data, but an increasing number will be photographs, drawings and other images, maps, and audio and video recordings. Each of these types of nontext documents will demand its own specialized searching methods.

The NSF-funded group at the University of California, Santa Barbara, for example, is at work on search techniques for digital images --i n this case maps, satellite photos, and other geographical information. Terence Smith, a co-leader of the group, explains that people who use such a library will generally want to search by three criteria: location ("Find all maps and photos containing 340N, 1200W"); content ("Find a map of Amelia Earhart's last flight"); and time) -- when the photo was taken or the map was drawn. The time and spatial coordinates are easy to catalog, but content is not. Conceivably, specially trained librarians could examine each map and image as it was digitized and enter an exhaustive list of the contents, but that would take too long and cost too much for the millions of images Smith and his colleague Michael Goodchild are eventually aiming for. Instead they will try to develop pattern-recognition software that can auto-matically characterize each image as it is added to the collection.

Multimedia libraries

While the Santa Barbara team is learning to put maps and images into a digital library, researchers at Carnegie Mellon University will be taking on an equally difficult task as part of another NSF project: assembling a digital library of video clips -- a total of 1000 hours, to start. Video, with its moving images, demands innovative search techniques. To determine the content of each video, the Carnegie Mellon team will run the soundtrack through Sphinx 2, a speech-recognition system developed at the school, in tandem with a natural-language processor, a system that analyzes grammar and meaning. The library will also keep tabs on the images in the collection. Given an image from one video clip -- of the Eiffel Tower, say -- the search program will be able to seek other clips with the same structure, even if it's shown from a different perspective. This capability will rely on software that constructs a three -- dimensional representation of an object and calculates what it would look like from different angles.

Assuming that such search techniques can be peffected, the ultimate goal will be to weave together many different kinds of libraries containing both text and nontext items into giant, multimedia libraries. Although the search software for different kinds of items may vary, says Terry Winograd, principal investigator for a consortium based at Stanford University, the user shouldn't have to worry about that. A student should be able to put in a single request for, say, information on the assassination of John F. Kennedy and receive a video of the shooting, a map of the final route JFK took through Dallas, text from newspapers, maga-zines, and books about the event, photos of Lee Harvey Oswald, and so on -- all from dozens of sources around the Internet.

Besides the sheer power of having access to so much information, digital libraries offer another, potentially greater capability, says Bruce Schatz of the National Center for Supercomputing Applications and the University of Illinois, who is principal investigator for the Illinois project. In the past, libraries have provided for only a one-way flow of information -- from the authors and other creators of information to the users. At most, a reader might write in the margins of a library book and thus offer something to future readers, but otherwise the information was fixed. Digital libraries offer a way for the users to play an active role.

The simplest way will be to offer annotating capabilities. Like the scribbler in a library book, users will be able to write notes to themselves and offer comments and extra information to any later user who wants to see what predecessors have added. Both the Illinois and Berkeley teams plan to offer this capability in their libraries. And Schatz and his colleagues are planning to equip their test library -- a collection of thousands of ar-ticles on science and engineering -- with an even more powerful interactive tool derived from Mosaic. A user who discovers an interesting relation between items will be able to set up links between them so that later users who called up one article would be referred to the others.

Ultimately, Schatz says, users should be able to add to the library themselves. A scientist might, for instance, take an equation from a journal article, use it to analyze data from another place on the Internet, use a visualization program from a third source to view the results, and then leave a record of the work for others to see and work with.

With this new suite of capabilities, the Internet would become something else altogether -- a space of connected information instead of a collection of machines. Schatz calls it the "Interspace." Others may offer other names, but certainly no one will call it a used book store.

-Robert Pool