Pages: pp. 19-21
Abstract—AI-assisted browsing combines knowledge-based systems with searching and browsing to find something for the first time, find changes in something, or refind something. To facilitate search or refinding, some systems allow or require users to establish markers indicating items of interest. In other settings, systems are less obtrusive, using data generated from browsing to learn a user's interests.
AI-assisted browsing can be helpful in information overload or underload situations:
AI-assisted browsing helps users generate or choose knowledge from the available virtual knowledge warehouses. As a result, from one point of view, AI-assisted browsing is a part of knowledge management. Knowledge management is the formal management of knowledge using a wide range of advanced technologies, including, for example, intranet technology, Lotus Notes, and intelligent agents. Typically, knowledge-management systems are promulgated by an organization or community.
Generally, either individuals or organizations organize AI-assisted browsing. Individuals have particular knowledge interests and generate individual views. For example, individuals generate and maintain a set of bookmarks. Firms such as Andersen Consulting and Price Waterhouse have experimented with intelligent agents designed to support and generate individual views.
At the organization level, firms attempt to provide a single promulgated view or multiple promulgated views of knowledge. David King and I have suggested that any number of users can just share the same agent, with the organizational-agent-gathered information going to all interested parties. 5 This approach can provide a corporate view, ensuring that everyone on the list receives the same set of information. Such a view can provide a basis for sharing important knowledge across an organization and generating a shared view of the world.
Organizations can use a number of approaches to generate a corporate view. For example, using empirical analyses of users, organizations such as Coopers & Lybrand have established multiple search paths by which their employees can access available knowledge. Alternatively, there can be an aggregation of individual views. Aggregation can occur using multiple approaches, including subscribing, cascading, and averaging. Unfortunately, each approach has limitations.
Arthur Andersen uses "grapeVine" as a group-decision tool to generate cascaded aggregation. Certain individuals make evaluations as to the level of importance (for example, "urgent") of pieces of information as the pieces are received. Other individuals then subscribe to those assessments, choosing a level that the information must exceed before they are interested in receiving information. Accordingly, information makes its way along a virtual grapevine to those who indicated an interest. However, behavioral incentives might influence the assessments that individuals attribute to information or that they make to receive information. For example, if others are depending on an individual's judgment, he or she might overstate the level to ensure that the subscribing individuals' needs are met. Further, to ensure that a conservative assessor misses nothing, subscribers might ask for a lower level than needed—for example, "important," not just "urgent."
Individual agents can be cascaded into the corporate agents. Corporate agents can be generated by gathering information directly from individual home pages so that information in the corporate agent is a cascaded summary of all the individual agents. Individual agents are added to the corporate agent by including each agent's lists of URLs to be examined. For example, if an individual agent provides information to the corporate agent, the agent's home page is on the list of pages. 1 Unfortunately, the cascaded approach's usefulness is somewhat limited because the aggregations might cause information overload. Further, the interpretation of summaries of this type potentially are limited by conflicting needs or theories.
Cascaded organizational views can be generated through other tools besides agents. For example, Nathalie Mathé and her colleagues suggest that individual markers set up in reference manuals can be aggregated for use in a shared-profile database. Unfortunately, knowledge might be context- and user-dependent, as researchers such as Guy Boy suggest. 2 Further, as Mathé and her colleagues note, users must constantly revise their knowledge. So, the usefulness of cascaded aggregated cognitive hypertext maps is also somewhat limited.
I've discussed a system that averaged the expertise of individual agents, in the form of probability assessments. Unfortunately, averaging ignores that different sources of information might come from different schools of thought. For example, if the information sources are physicians, one might be a surgeon and the other a chemotherapist. In the case of trying to cure cancer, the two would have widely different views. As a result, averaging would generate a view likely to be different than that of either expert.
Each of these approaches can be useful in some settings but not in others. Therefore, the problem of generating a corporate view needs additional research.
In addition, individual agents might have incentives to not disclose their interests to corporate view agents. If the organizational culture is collaborative, the incentive will be to have the corporate view capture and include individual information. If the culture is competitive, individuals might not be inclined to share information. To gain a competitive advantage, the individual would use his or her agent without having that agent disclose information to any organization agent.
These issues have received only limited attention. Most of the previous research has been at workshops and has not been generally available. So, as part of a focus on AI and knowledge navigation, IEEE Expert is featuring one of these articles per issue, starting last issue:
Robin D. Burke, Kristian J. Hammond, and Benjamin C. Young discuss assisted browsing using knowledge-based retrieval strategies in a series of what they label "FindMe" systems. These systems include Car Navigator (selecting a new car), Video Navigator (choosing a rental video), RentMe (finding an apartment), Entree (selecting a restaurant), and Kenwood (configuring a home audio system). Each system searches a specific database, retrieves information that meets certain criteria, and ranks the results. (This article appeared on pages 32-40 of the July/Aug. Expert.)
Bruce Krulwich and Chad Burkey developed InfoFinder, an intelligent agent that learns a user's interests from texts such as messages or documents. The system can review a large number of documents and messages and make recommendations without the user spending substantial time searching the document set. In particular, InfoFinder extracts semantically significant phrases from each document; learns decision tress for each category, based on phrases extracted from the documents; and transforms the decision trees into Boolean search query strings. (This article, "The InfoFinder Agent: Learning User Interests through Heuristic Phrase Extraction" is in this issue.)
Nathalie Mathé, Joshua Rabinowitz, James R. Chen, and Shawn R. Wolfe are concerned with facilitating information access to technical and operational manuals. Using sophisticated annotations and hyperlinking capabilities, Adaptive HyperMan (AHP) helps users organize and access large amounts of information. AHP has a number of capabilities, such as indexing markers by topics, retrieving markers by topics, giving feedback to modify the relevance of markers, and combining markers to facilitate sharing.
Nicholas Caldwell, Bernie Breton, and David Holburn integrate an expert system into a web browser to facilitate remote use of an electron microscope over the World Wide Web. Although this study is not directly concerned with AI-assisted browsing, it has at least two implications for using AI and expert systems in a web environment. First, the expert system facilitates the use of a remote test facility. Whenever a user comes to a rule that includes a remote test, the user is offered the chance to perform the test manually or semiautomatically. The semiautomatic mode lets an engineer see the commands and the sequence in which they will be invoked. In addition, the engineer can provide any or all portions of the test before providing the expert system with an answer. Second, the web-based approach facilitates online help. Help files are maintained separately from the expert system, using HTML to facilitate hypertext links to local and centrally secure files.
Mark S. Ackerman, Brian Starr, and Michael J. Pazzani developed the Do-I-Care agent to help users discover interesting changes on the World Wide Web. The DICA has six functions:
The agent uses machine learning to determine what changes are interesting to the user and how often they occur.