Susan T. Dumais
Bellcore
Interfaces matter. Although this may be obvious to this group, it is not well-reflected in information science conferences. In the information science community, a good deal of research is focused on what might be called "system" issues - matching algortihms (Boolean, vector, probabalisitic), data structures, term weighting, etc. Only recently have we begun to see a general interest in evaluating the larger human-system environment. Good user interfaces consistently provide retrieval advantages of 25% or more in retrieval accuracy and speed, and the differences are often substantially larger. By contrast, 25% differences between retrieval engines are seldom found. Most of the standard "enhancements" (stemming, automatic or manual phrase detection, thesaurus classes, differential term weighting) have effects of 1-10%; only relevance feedback has effects that consistently approach 25% ... and this, of course, depends on user feedback!
The only sure way to good interface design is user-centered design (UCD), and ongoing evaluation. Human-computer interaction design principles and guidelines may give you a reasonable starting place, but there is no replacement for testing. Interface design is especially difficult when the users and tasks vary as widely as they do in digital libraries, because selecting representative users and tasks may be difficult. To make matters even worse there are probably many interactions. The appropriate interface may depend on the underlying retrieval engines, task or user characteristics. This means that it is not necessarily enough to find a good interface in one domain and import it to a new domain or task. Luckily, large interface problems will be evident with only a few users. Among the design keys are: details matter; get the common tasks right; design don't delegate; and, of course, evaluate.
While the challenges are tough, there are also several opportunities. The next-generation of retrieval interfaces will, I believe, go beyond today's keyword retrieval systems and hand-crafted organiztions. The two main access methods today are search and navigation. Both have well documented problems. While the appeal of point-and-click hypertext interfaces is alluringly, it is all too obvious that they do not solve access problems. It is surprisingly difficult to know which links to follow. How many of us would have stumbled on "computer programming" under the heading "Generalities / Knowledge and Its Extensions / Data Processing" in the Dewey Decimal Classification? The difficulties that end-users have with standard search engines, especially Boolean systems are also well-documented. Combining the two methods can be quite powerful, and may help overcome the disadvantages of either alone. Interactive systems and iterative query formulation are here to stay, and can be used to good advantage. Another new and exciting possibility is social-augmented finding. In the real world, we often find information not by running to the library or connecting to the WWW but rather by asking our friends. Similar kinds of social recommending are now being tried with virtual electronic communities with good success.
Challenges:
Opportunities: