Collaborative, programmable intelligent agents. (e-mail computer programs)

Extracting semantics from everyday documents, intelligent agents, illustrated by Apple Data Detectors, infer high-level goals from simple user actions.

We began our research on intelligent agents with the same romantic imagery that has always fueled interest in agents: Robbie the Robot from "Forbidden Planet," HAL from "2001: A Space Odyssey," the Star Trek computers (including Data). The images they conveyed of intelligent machines and the ways people would interact with them - by talking as if to an old friend who knows you so well they could finish your sentences for you - were so attractive that a whole generation viewed them as proven technology. They also became the models for the intelligent interface community that emerged out of artificial intelligence research in the 1980s, even for early work on AI itself.

But building anything approximating real intelligence into a computer has proved to be a painfully difficult task, and the powers of Robbie and HAL have remained beyond our grasp. We need to step back a bit, think carefully about what people and computers are each good at, understand how they can complement each other, and where we as system designers can do some good.

Shneiderman [12] observed that claims about intelligent software agents are vague, dreamy, and unrealized. As Apple Computer researchers, we started from a simple but focused approach to agents: That they should have the ability to infer appropriate high-level goals from user actions and requests and take action to achieve these goals. Further, based on a study of reference librarians as exemplary human agents [9], we wanted to build a system in which the user would not have to state goals explicitly and in detail. We learned from librarians that a large part of their value to clients is in working with imprecise requests. Beyond this concern, our general design strategy was to keep the most basic user question in front of us at all times: Will this software do something useful for users in an intelligent way that makes them more productive? The system we describe here - called Apple Data Detectors - meets our criteria of being unobtrusive, being able to infer user needs, and doing useful work. Apple Data Detectors shipped as a product in 1997.

Earlier work on intelligent agents was multi-faceted, to the point where it is difficult to find a consensus among researchers on exactly what constitutes an "agent" or even "intelligence." However, in nearly all cases, systems described as "agent-based" rely on some explicitly represented knowledge about relevant aspects of the world - the objects or concepts being addressed by the software, the tasks relevant to the user, and the user's own knowledge about the world. Researchers have used machine learning techniques to track user actions and construct models of user preferences [7], create explicit models of user knowledge and skill levels in an attempt to anticipate user actions, misconceptions, and information needs [2], and implement planning systems to leap from a user's stated intention to the specific actions required to achieve that intention [3]. The locality of agents also varies across different agent-based systems; some act only within one's own machine, find others autonomously crawl the Web, searching for interesting content [4]. We tried to find a middle ground by using explicit representations of user-relevant information as a means of identifying actions users might wish to take but to leave the choice of these actions to users.

Working with Information Inside User Documents

Our first step was to find a user problem that needed solving in which intelligent agents would add value. In an investigation of how people file information on their computer desktops [1], we discovered that a common user complaint is that they cannot easily take action on the structured information found in everyday documents (structured information being data-recognizable by a grammar). Ordinary documents are full of such structured information: phone numbers, fax numbers, street addresses, email addresses, email signatures, abstracts, tables of contents, lists of references, tables, figures, captions, meeting announcements, Web addresses, and more. In addition, there are countless domain-specific structures, such as ISBN numbers, stock symbols, chemical structures, and mathematical equations. These structures are not only relevant to users, but because of their structure, are also recognizable by parsing technologies. Once identified, the structure's type can be used to identify appropriate actions that might be carried out, like placing a meeting on a calendar, adding an address to an address book, dialing a phone number, …