[expertise] While links within and among documents are critical to hypertext, the future of hypertext may depend on the addition of artificial intelligence capabilities to hypertext systems. One way to do this is to embed knowledge in links and to allow these links to precipitate arbitrary computations. With intelligent hypertext some of the expertise of people which is otherwise outside the computer system is moved into the computer.
[hypertext history]Hypertext systems of the 1970s already faced challenges of how to incorporate more flexibility or intelligence into the computer. The Dynabook system included navigation aids that introduced the possibility of the hypertext system responding dynamically to the user. Miniature page icons stamped with time and date could be selected. This not only took one to a page previously visited, but also to its status at that time. The model for Dynabook supported dynamic and not just static hypertext. The reader could cause different information to appear on the same page as a function of the traversal. Thus on two successive accesses a page might appear different through either hiding information or showing new information. The creators of Dynabook concluded that to use such dynamic facilities to best help the user, knowledge must be incorporated into the computer.
[definitions]Expert systems are built by knowledge engineers and experts who together translate expertise into a knowledge base plus inferencing mechanism [Haye83]. While the expert system may directly solve problems, it also communicates knowledge from an expert to a user. Similarly with a hypertext system a writer communicates with a reader. These parallels suggest improvements to both expert systems and hypertext systems.
[MYCIN]The development of hypertext systems has parallels with that of expert systems. One of the first expert systems, MYCIN, was developed in the mid-1970s. MYCIN employed a few hundred if-then rules about meningitis and bacteremia in order to deduce the proper treatment for a patient who presented with signs of either of those diseases. MYCIN, while provably expert within its domains of bacteremia and meningitis, did not cover enough of the information which physicians routinely need for them to spend valuable time accessing MYCIN. The developers of MYCIN have partly shifted their attention to the development of user-friendly or hypertext information systems that incorporate expert systems but primarily handle the day-to-day work flow of a physician in a hospital ward [Shor81]. The need to incorporate expert systems into the work flow has been a stimulus to the development of expertext. Other medical expert systems are following the hypertext paradigm in order that the computer's services be more attractive to the health-care professional [Timp86].
[INTERNIST]The INTERNIST expert system which was developed in the mid-1970s was the first expert system to handle a wide range of internal medicine problems. Over 4,000 signs and symptoms were connected with over 1,000 diseases for diagnosis and treatment purposes. Unfortunately, the user had to know how to phrase his patient's problem in the exact terms that the developers of the expert system had chosen this discouraged the user. Alternatives are being pursued which take advantage of hypertext strategies. For instance, INTERNIST has been converted into a teaching tool which allows the user to browse through the knowledge in INTERNIST in a hypertext fashion [Firs85]. The options of what to pursue at each point are made clear, as the user browses the options. The user doesn't have to guess at the phraseology of one expert. The expert system, modified for medical teaching, is expected to fit more easily into the work flow of students than it did the work flow of physicians.
[synergism]The rules on which an expert system is based often seem inadequate when the expert system attempts to explain its decisions to users. Hypertext can be integrated with an expert system so that the expert system can (synergistically) better explain its decisions. Textbook information can be offered to the user when the expert system is questioned. From the other side, the links in hypertext may be difficult to follow. Hypertext systems can include expert system features that help the user find relevant information. The combination of expert system and hypertext system is called an expertext system [Rada90]. More generally, the terms intelligent hypertext and expertext may be used interchangeably in this book.
[dynamic links]The links in hypertext can be exploited or extended so that the user can appreciate a dynamic, intelligent text. In one case, the links aren't more complicated but patterns of links are exploited. In another case, the graph has predicates at the nodes and implications on the links, and the graph can support inferencing. Procedures can be added to the links and support all kinds of computation.
[semantic patterns]The computer can exploit the pattern of links in a hypertext and give the user different perspectives on the hypertext. For instance, the user may express an interest in `causes', and the system could organize information so as to emphasize the causal links. The nodes and links of hypertext may be viewed as a semantic net. A link attributes meaning to the pair of nodes it connects, and a node may have more than one meaning, when it participates in relations of different types [Carl90]. Inheritance of properties along hierarchical links in a semantic net and spreading activation in a semantic net both take advantage of the patterns therein.
[Electronic Yellow Pages]When submitting a query to an Electronic Yellow Pages, the consumer wants a certain type of product or service, a phone number, and a location. The consumer is constrained by nearby location, degree of specificity in product-service hierarchy, time available for search, and amount of information wanted. Responding to a consumer's query with a list of hundreds of businesses would not be reasonable. At the other extreme, returning without the name of at least one business which could satisfy the consumer's needs would not be very helpful. In each case, whether reducing the solution space to present the consumer with a manageable set of alternatives or enlarging the solution space to present the user with at least a few alternatives, the semantic network behind the Electronic Yellow Pages can be used as a guide. For example, grocery stores and restaurants are related in that they share the common ancestor `business' and both sell `food'. If a consumer was looking for a restaurant in a specific locale and none existed, the system might present the consumer with names of nearby grocery stores.
[connectivity]The connections among nodes in a semantic net are critical to its functionality. In spreading activation a concept is identified in a semantic net and then adjacent concepts in the net are visited in the course of solving some problem. The significance of the traversal of the net depends on the net and the way the traversal is used to infer a solution to a problem. In one of its simplest forms, spreading activation starts from two nodes and activates all the nodes connected to each of them. Then all of the nodes connected to each of those are activated, forming an expanding sphere of activation around each of the original concepts. When some concept has been activated simultaneously from two directions, a conclusion is drawn [Fike77].
[inheritance]One key feature of the semantic net representation is that important associations can be made explicitly and succinctly. Relevant facts about a concept can be inferred from the nodes to which they are directly linked. Given that Tweety is a canary and that a canary is a bird, one can use the property of inheritance hierarchies to infer that Tweety is a bird. Semantic nets facilitate representation and reasoning with inheritance hierarchies, and such hierarchies are particularly germane to hypertext.
[analogical structure]To say that a concept inherits the attributes of its parent concept is to only characterize the attributes which the two concepts have in common. There is a structure-function relationship in semantic nets which allows a special kind of inheritance called analogical inheritance. Assume that every concept has both hierarchical and non-hierarchical relations. In one example of analogical inheritance each attribute f(x) of a concept x would have a child relationship to the attribute f(y) when x is the child of y. More generally, analogical inheritance means that a concept x which is a child of a concept y will have an attribute x' which is related in some systematic way to the corresponding attribute y' of y (see Exercise "Isomorphism").
[insertion]If an author has developed part of a semantic net and has a fragment but doesn't know where to connect that fragment to the rest of the network, analogical inheritance might provide a guide. An example of how this insertion might occur is in the placement of the `Chairman of the Mathematics Department' into the hierarchy about the administration of the university. If one knows that the `Chairman of the Mathematics Department' has an office in the `School of Natural Sciences', then one might expect that the Chairman reports to the `Dean of the School of Natural Sciences' rather than to the `Dean of the School of Engineering'. The computer can keep track of the semantic net that has been created and make suggestions to the author about patterns.
[outline]To what extent does a document outline constitute a semantic net (see Exercise "Outline")? Imagine an outline as a special binary tree. For every pair of siblings the first entry is meant to relate to the second entry by implication (see Figure "Patterns in Outline"). For instance, if the first section of a document is called `Assumptions' and the second section is called `Demonstration', then the `Assumptions' imply the `Demonstration'. Likewise, the `Assumptions' section may be divided into a `Problem' and `Hypothesis' section in which the `Problem' implies the `Hypothesis'. These patterns demonstrate analogical inheritance.
[models of books]Non-fiction writers appeal to a number of models of their areas of discourse. Such models are reflected in the structure of their books. Inherent to the outline of a book are two kinds of relationships: precedence relationships as reflected by the sequence of subsections and hierarchical relationships of textual containment between a section of text and its subsections [Mili90].