Information Today
Volume 16, Number 6 • June 1999
• Report from the Field •
Search Engines: The 1999 Conference
Visualization was the ‘star of the show’ at the 
recent Infonortics-sponsored meeting in Boston
by Susan Feldman

Search Engines is a small conference that packs a wallop. Now in its fourth year, this Infonortics conference crosses the lines between professions and disciplines, and the mix is heady. Search engine designers from major Web and non-Web search engines, information professionals who design intranets, professional searchers, researchers from the information retrieval world, and researchers in visualization and user studies all crossed paths, and occasionally crossed swords. This year, the attendees left stuffed with ideas, and suffering from writer’s cramp. At least I did.

Three themes emerged during these two intense days: visualization, metadata and categorization, and pursuit of the elusive user.

Visualization was the star of this show. As James Wise noted in his half-day seminar on the subject, people are visual animals. We can process more visual information and process it more quickly if it’s in the form of graphs, charts, or pictures than if it is text. Those of us in information-intense fields are burdened with more text than we 1must be able to explore the information we receive in some sort of intuitive spatial format. Color, shape, and proximity to other shapes can convey information quickly. Landscapes and galaxies of stars are good approaches because they are familiar forms. 
Figure 1
All kinds of visual representations turned up at this conference. Cartia’s ( virtual maps of topics retrieved by a search look like a geological survey map. They map a subject terrain in understandable, easy-to-view terms—if  a hill is high, it has lots of documents. A valley indicates that there is not much there. Spotfire ( is a clever way to map several topics using colors and dots. It can also graph topics on a timeline. Wise’s own galaxy approach shows clusters of documents. The closer they are, the more similar their meaning. For those of you wondering how this is done, usually similarity between documents is calculated using some variation of the vector space model that Gerald Salton pioneered.

Many of the presenters at this conference were experimenting with visualizations to help the user navigate through large sets of documents, or improve a query. New data mining systems are also experimenting with visual displays. KNOW-IT from TextWise ( shows two concepts linked by the kind of relationship that binds them. (See Figure 1.) Imagine that a system could show what caused an event to occur, in a visual format. The user would be able to grasp the significance of that relationship much more readily than if he had to plow through 10 documents himself.  The InXight Hyperbolic Browser, developed at Xerox PARC, uses hierarchies to break up large collections of things into small usable chunks ( (See Figure 2.) Both InXight’s and TextWise’s tools invite interaction. You are presented with top-level categories. Clicking on topics lets you drill down to more specific subjects, or, at the end of the road, to a list of documents in that category. How will people use this novel approach to browse for and explore information?

Figure 2
Metadata and categorization was by far the most surprising trend to those of us from the library world. Ev Brenner, who is one of the conference organizers and who is astonished at this return to the past, chaired a panel that examined whether the need for categorization was valid. It seems that the automatic categorization in use today is an attempt to solve two problems: multiple meanings of terms, and the user’s need to understand the contents of either a search or a large collection of Web pages. Automatic categorization, which occurs at the document processing stage, rather than at the searching stage, makes sense in this context. Users need any tool they can get to discriminate among the many documents returned in a search. They also need to have a sense of what is in the collection, as well as in the search results. In contrast, manual categorization, a process that was questioned in the ’50s and ’60s, is a labor-intensive activity that adds significant costs and delays to putting information online. Northern Light’s Marc Krellenstein stated that they apply terms automatically based on rules developed by humans. This facilitates browsing and searching, and it also improves Northern Light’s relevance ranking. Since it is applied automatically when the Web page is added to the NL index, it does not delay access to the information. This categorization also produces Northern Light’s custom folders, a discovery and navigation tool I like a lot.

In contrast, LookSmart is using people to review sites for appropriateness, and also to assign them to a category. Its target audience is new arrivals on the Web. It is aimed at family usage, a “G-rated” search engine. LookSmart maintains that it needs humans to make these subtle judgments of quality and suitability. Note that this is one of several tactics Web search engines are choosing to distinguish themselves from the rest. On the downside, Peter Tomassi noted, you have to feed and entertain humans, but computers just keep going. For this reason, they are considering adding automatic categorization.

Dan Miller from Ask Jeeves described their human-centered process. Ask Jeeves tries to answer questions with the single best answer it can find. To do this, it is manually building a knowledge base of answer templates and question templates, rather than a collection of Web pages. Miller maintains that despite this being a manual process, it is quite scalable, since the process happens at the input stage, rather than at the time of searching. Ask Jeeves is getting faster as it adds more questions, since there are similarities between many questions. (If you want a review of a Ford Explorer, it’s easy to use the same source as an answer to other car reviews.) One promising application for this approach is in a corporate customer service department, which answers a finite number of questions within a defined domain. Dell Computers uses Ask Jeeves for this purpose.

In contrast, James Callan, of the University of Massachusetts, pointed out that it is difficult to create good categories because they overlap. Clear distinctions are hard to define. They require labor and insert lag time in the process. A list of 30,000 categories is difficult to navigate. Full-text searching is an attractive alternative. Most of the Web search engines are based on older search technologies. If they incorporated newer approaches, such as full natural language processing, better statistical models, or hub and link technologies like Clever, the search results might improve.  It may be more useful to make documents easier to find, and to understand how people search than to return to categorization.

The value of this surprising return to an old library approach will not be resolved soon. The trend we will see in information systems of the future, I predict, is that they will combine as many entry points, views, and sources of information as they can about a set of documents. The reason is that different people need the same information presented to match their particular need for information, as well as their own styles of searching and learning. The University of Tennessee’s Carol Tenopir has done recent studies that underline this need. She reports that experience, technical aptitude, age, cognitive and learning styles, and personality type all distinguish how people seek information. Of particular interest are Tenopir’s studies on the influence personality traits and emotional factors have on how people search. What she found is that the affective domain—emotions such as stress, satisfaction, or frustration—influences searching behavior as much as cognitive or sensorimotor factors. She has also studied the differences in searching between novices and information intermediaries. Both novices and professionals create search strategies based on their personality traits, which is not a surprise. What we didn’t expect is that both of them also alter strategies based on emotional factors, not cold hard logic—rather an unpleasantly revealing fact for those of us who regard ourselves as rational beings. Novices, however, are satisfied if they find just one answer. Presumably, professionals seek additional confirmation of accuracy, as well as other points of view. She expects that new research in how people seek information, as well as new input/output devices such as voice interaction, games, or wearable computers, may change how we interact with computers. This research is critical if we are to design easy-to-use information systems that can serve a spectrum of information needs and users.

Other Highlights
Danny Sullivan, of Search Engine Watch fame, talked about “portalmania”—the “shift to serving information instead of Web pages”
—and other Web search engine trends. Portals have “sticky features”—features that attract users so that they won’t leave the site. Anyone who has used Web search engines recently must have noted the addition of directories, chat areas, free e-mail, shopping, and content that resides on the search engine site—to say nothing of ad banners. Search, says Sullivan, is becoming much less prominent. However, Web search engines are also trying to improve the relevance of results for popular queries. Since most users do not use advanced search features, or enter complex queries, the Web search engines are trying to direct them to the most popular sites for that query, or they are creating directories (hence the interest in categorization) to help the user find the right ballpark so that he can browse productively. Sullivan predicted that we will see more use of nontraditional ranking criteria (like popularity, or number of links to or from a site). He also expects continued growth of directories, and more specialized collections with less emphasis on comprehensive Web crawling.

Another highlight, for me, was an informal lunchtime get-together of Web search engine staff and several industry observers. Sullivan invited us all to discuss the possibility of establishing standards for search syntax, and also other topics of mutual concern. For more information on this, see At present, Sullivan has two proposals up for discussion: that all search engines be able to narrow a search by site, and that they all have the ability to locate an exact URL within their indexes. Participants also discussed the problems they are all having with the spamming that is so prevalent on the Web. The interest in this discussion is indicative of the need for working together on some common problems. As of April 29, the participants in the group are AltaVista; Excite; Fireball; Google; GoTo; HotBot; Infoseek; Inktomi; LookSmart; MSN Search; Netscape; Northern Light; Search UK; Snap; Yahoo!; Luis Gravano, Columbia University; Sue Feldman, Datasearch; Jakob Nielsen, User Advocate; Greg Notess, Search Engine Showdown; Avi Rappaport,; Lou Rosenfeld, Argus Associates; Chris Sherman, Mining Co. Web Search Guide; Danny Sullivan, Search Engine Watch; and Roy Tennant, Web4Lib ( A more extensive account of this discussion will be published in Searcher magazine by Avi Rappaport.

Steve Arnold, in a sweeping overview of the state of the online/information industry, noted a number of trends, some of which seem to work against each other. The consolidation of companies to create a vertical market works against the fragmentation of information sources. Computationally intensive technologies, such as visualization or multiple relevance-ranking techniques, are becoming mainstream as bandwidth and desktop computing power increase. There is a shift from fee to free software, which will make software business models scramble for income. Products and services are coming bundled together like a Russian matryushka doll, with search bundled with shopping or portals. XML and commercial XML (cXML) are more and more prevalent, improving display and search options.

Our online community often loses sight of the role government funding has played in developing new technologies. Ellen Voorhees from NIST and Terry Firmin from NCSC reviewed the TIPSTER, TREC, and SUMMAC programs. Together, these programs—which fund development and compare and test information technologies—have helped to create the excitement and ferment of the information retrieval field. They fund development of new statistical and probabilistic techniques, natural language processing, cross-language retrieval, filtering, relevance feedback, retrieval from spoken text, and question-answering systems. Their new Web track will examine whether Web documents are inherently different from other types of documents. It will also test how well search algorithms perform on large collections of text—100 GB of Web documents. NIST is also examining user behavior, queries, and multimedia information sources through new competitive tracks. The SUMMAC program examines methods for creating summaries automatically; this technology is in its infancy. Not only are questions of what constitutes a “good” summary dependent on the purpose to which it will be ut, but issues of copyright have yet to be settled, since a summary or an extract may be considered a derivative work. Nevertheless, experiments by various researchers have shown that some systems have achieved a reasonable rate of accuracy. Wouldn’t it be nice to have the “executive summary” of your monthly report written for you?

New Technologies
The plethora of new technologies we saw makes it impossible to highlight them all. Here are a few that have left me dreaming of the very near future:

Predictions for the Future
David Evans of Claritech did a masterful job of wrapping up the conference by giving direction and perspective to these many mind-boggling technologies. He suggests that we are heading towards decision-support systems. These information management systems would incorporate text-mining tools to integrate information into the decision process. Data mining exploits patterns and regularities, particularly in relational databases. Text mining will integrate both relational and free-text sources. It will automatically analyze document structure to discover fields, attributes, and values. It will use natural language processing to parse text and determine lexical content. If we can develop robust fact extraction, summarization, filtering, visualization, agent, and learning technologies, then we will be able to use decision-support systems to detect and track events such as the arrival of a new competitor, or a new strain of disease-resistant bacteria. A new user interface that frees the user from the underlying process is critical to the success of this goal.

Many of the speakers offered their best guesses about the future. Here are some of the most tantalizing:

 This last prediction is technically feasible. I wonder, though, if it is commercially and politically possible. With publishers unbundling their publications from large aggregators and indexing services, and with separate Web sites requiring separate passwords and accounts springing up like mushrooms, will the mere technical ability to search among many sources seamlessly make it happen? Interesting times.

Next year’s Search Engines conference is scheduled for April 10-11, 2000, in Boston. For more information, check This year’s presentations will all be posted at that site.

Susan Feldman is president of Datasearch and a principal owner of Datasearch Labs, a new independent usability testing company. Her e-mail address is

Table of Contents Information Today Home Page