We have developed and empirically evaluated a method of information seeking (called Adaptive Search) that combines automatic document clustering and user feedback in a novel way. In this approach, the user starts with a natural text description of the needed information and goes through a sequence of interactions with the system in order to find documents of interest. Adaptive Search utilizes Kohonen Self-Organizing maps and acts as a layer between the user and a commercial search engine. In a laboratory experiment, subjects searched the World Wide Web for answers to a given set of questions. Our results indicate that the subjects spent less time finding correct answers using Adaptive Search than using the search engine directly. In addition, the Adaptive Search-suggested documents contained answers that were positioned consistently higher in the rank-ordered lists than those suggested by the Internet search engine. This suggests that document clustering can be integrated into an interactive search system in such a way that it substantially helps information seekers.

Summarization and visualization tools are believed to be helpful in navigating through large volumes of data since a visual representation may elicit more deliberate query reformulation and better feedback to the retrieval system. Results from improved feedback explain the growing interest in visualization tools in academic and industrial research (for example, Scatter/Gather by Xerox, ThemeMedia by Batelle, Live Topics by AltaVista).Most summarization and visualization tools rely on the ability of a computer to cluster documents or terms and visualize relationships among them.Prior work has shown that automatically generated concepts and clusters sometimes fall short of user expectations and do not reliably facilitate information access.We demonstrate our prototype system with novel features to overcome these problems. In addition to navigation of clusters of documents and concepts, our prototype adds customization of concepts, forming new clusters adapted to the particular user and task.Our prototype acts as a visualizing layer between the user and a commercial Web search engine. We currently connect to AltaVista, but remain independent of any specific search engine features. Our system summarizes search results and suggests additional terms for query modification. Based on user feedback, the system can create rank ordered lists of found URLs.Our prototype system uses Kohonen's self-organizing map (SOM), an unsupervised two-layer neural network. Our current system extends the one we presented at ACM SIGIR 97. Figure 1 shows the user/system/search engine interaction. The search follows these steps:1. The user enters a query on an HTML form. 2. Our system, ASOM, implemented as a CGI script, routes the query to an external search engine (AltaVista in our prototype).3. The search engine returns a ranked list of urls and their short summaries (called snippets) to ASOM.