This section presents the design of WebWatcher through a scenario of its use. WebWatcher is an information search agent that is ``invoked'' by following a web hyperlink to its web page, then filling out a Mosaic form to indicate what information is sought (e.g., a publication by some author). WebWatcher then returns the user to (a copy of) the web page from which he or she came, and assists the user as they follow hyperlinks forward through the web in search of the target information. As the user traverses the web, WebWatcher uses its learned knowledge to recommend especially promising hyperlinks to the user by highlighting these links on the user's display. At any point, the user may dismiss WebWatcher, by clicking one of two indicators on the WebWatcher icon, indicating either that the search has succeeded, or that the user wishes to give up on this search.
Figure: Copy of original page with WebWatcher advice
Figure: Next page (user has followed WebWatcher's advice)
The sequence of web pages visited by the user in a typical scenario is
illustrated in figures 1 through 5. The first screen
shows a typical web page,
providing information about Machine Learning. Notice in the
third paragraph, this page invites the user to try out WebWatcher. If the
user clicks on this link, he or she arrives at the front door WebWatcher page
(Figure
2), which allows the user to identify the type of information he
seeks. In this scenario the user indicates that the goal is to locate a
paper, so he is shown a new screen (Figure
3) with a form to elaborate this information request. Once
completed, the user is returned to the original page (Figure 4),
with WebWatcher now ``looking over his shoulder''. Notice the WebWatcher icon
at the top of the screen, and the highlighted link (bracketed by the
WebWatcher eyes icon) halfway down the screen. This highlighted link
indicates WebWatcher's advice that the user follow the link to the
University of Illinois / Urbana (UIUC) AI / ML Page. The user decides to
select this recommended link, and arrives at the new web page shown in Figure
5, which contains new advice from WebWatcher. The search
continues in this way, with the user directing the search and WebWatcher
highlighting recommended links, until the user dismisses WebWatcher by
clicking on ``I found it'' or ``I give up''.
From the user's perspective WebWatcher is an agent with specialized knowledge about how to search outward from the page on which it was invoked. While WebWatcher suggests which hyperlink the user should take, the user remains firmly in control, and may ignore the system's advice at any step. We feel it is important for the user to remain in control, because WebWatcher's knowledge may provide imperfect advice, and because WebWatcher might not perfectly understand the user's information seeking goal.
From WebWatcher's perspective, the above scenario looks somewhat different. When first invoked it accepts an argument, encoded in the URL that accesses it, which contains the user's ``return address.'' The return address is the URL of the web page from which the user came. Once the user fills out the form specifying his or her information seeking goal, WebWatcher sends the user back to a copy of this original page, after making three changes. First, the WebWatcher banner is added to the top of the page. Second, each hyperlink URL in the original page is replaced by a new URL that points back to the WebWatcher. Third, if the WebWatcher finds that any of the hyperlinks on this page are strongly recommended by its search control knowledge, then it highlights the most promising links in order to suggest them to the user. It sends this modified copy of the return page to the user, and opens a file to begin logging this user's information search as training data. While it waits for the user's next step, it prefetches any web pages it has just recommended to the user, and begins to process these pages to determine their most promising outgoing hyperlink. When the user clicks on the next hyperlink, WebWatcher updates the log for this search, retrieves the page (unless it has already been prefetched), performs similar substitutions, and returns the copy to the user.
This process continues, with WebWatcher tracking the user's search across the Web, providing advice at each step, until the user elects to dismiss the agent. At this point, the WebWatcher closes the log file for this session (indicating either success or failure in the search, depending on which button the user selected when dismissing WebWatcher), and returns the user to the original, unsubstituted copy of the web page he is currently at.
The above scenario describes a typical interaction with the current WebWatcher. We plan to extend the initial system in several ways. For example, WebWatcher could be made to search several pages ahead, by following its own advice while waiting for the user's next input, in order to improve upon the quality of advice it provides. In addition, if it encounters an especially promising page while searching ahead, it might suggest that the user jump directly to this page rather than follow tediously along the path that the agent has already traversed.
