Two-level Hypertext Models as an Underpinning for AHSs

James Mayfield

Computer Science and Electrical Engineering Department

University of Maryland Baltimore County

Baltimore MD 21228-5398 USA

mayfield@cs.umbc.edu

In recent years there has been an explosion of interest in models of hypertext that go beyond the explicit hard-wired links found in traditional hypertext models. In these new models, text is augmented with a secondary structure through which all links are routed. A variety of secondary structures have been proposed, including semantic nets [trigg86, collier87, rada90, kheirbek93, mayfield93], Petri nets [furuta89], and connectionist nets [hu93]. The purpose of the secondary structure also varies, although in each case it adds knowledge about the nodes and links of the hypertext.

I contend that two-level models form a natural underpinning for AHSs, because they provide a level of abstraction to hypertext links that is required if a hypertext system is to be adaptive. This abstraction allows the system to provide the appearance of an extremely malleable hyperspace which is nonetheless drawn out of a primarily static set of nodes and links.

SNITCH [mayfield93] is an example of a two-level hypertext model that provides this link abstraction capability. In the SNITCH model, a semantic net representation of a text is connected to the phrases of that text. A hypertext link in this model is a virtual connection between two pieces of text called anchors. Each anchor comprises one or more words in the text that express a single concept. A semantic net is a graph whose vertices correspond to concepts (e.g. "person" or "house"), and whose edges correspond to relationships between those concepts (e.g. "owns"). Semantic nets are used to represent knowledge about the world and to express the meaning of phrases and sentences of languages such as English. The underlying representation of a SNITCH hypertext link is a path that starts at an anchor in the text, moves across to the semantic net, traverses a sub-path through the semantic net, and then returns to another anchor in the text.

The original objectives of the SNITCH project were to support the automatic construction of hypertext links, and to endow those links with rich semantics. The augmentation of a set of documents with a semantic net has six significant advantages over existing hypertext construction methods:

  1. Links between documents are discovered dynamically. Links that connect different parts of the corpus can be discovered dynamically, as long as the components of the desired relationship have been expressed in the semantic net. This dynamic discovery of links means that the system designer need not anticipate every link type that a user might ultimately request.
  2. Knowledge can be reused. Because a single edge in the semantic net can be shared by many different paths, the knowledge implicit in a single semantic net edge is available for use in many distinct contexts. Furthermore, portions of the net can be developed independently, allowing system designers to use existing semantic nets (such as CYC [guha91] or WordNet [miller90]) that express facts about the domain of interest. The ongoing development of large, sharable ontologies and knowledge bases [neches91] makes this advantage particularly important.
  3. Texts can be incorporated individually. Using the SNITCH approach, input texts can be processed separately. Since each input text is connected to the same semantic net, paths between two related texts will be generated even if the texts are incorporated at separate times.
  4. Links derive meaning from the semantic net. Since each link between anchors is composed of edges, and since each edge has an associated semantics, the composition of these edge semantics give meaning to that link. Thus, the use of even a small number of semantic net edge types can result in great expressive power. This contrasts with the approach taken in many existing hypertext systems, whereby hypertext links indicate only an anonymous association between two pieces of text.
  5. Links between documents can be created automatically. The SNITCH approach facilitates the automatic creation of meaningful links, by using natural language processing techniques to translate a text into a semantic net representation of that text.
  6. New links can be created by automatic inference processes. Once an input text has been connected to its semantic net representation, automatic inference processes can manipulate and update the semantic net. New semantic net edges created by such a process can be interpreted as components of new hypertext links.
Of these advantages, the first and last are particularly important from the standpoint of AHSs. While many adaptive systems have a long-term component, it seems to me crucial in most contexts for the system to adapt to a user within a single session. By altering the shapes of the allowable paths through the semantic net, the SNITCH model can provide this kind of intrasession adaptation without requiring modification to the node/anchor/link structure of the hypertext. Furthermore, automatic inference processes can be tailored to particular users' needs, producing novel paths through the semantic net.

SNITCH is not the only two-level hypertext model; each such model presents slightly different advantages from the AHS viewpoint. The basic feature provided by all of them though is a level of abstraction between the appearance of a link and its underlying structure. Such abstraction is crucial for AHSs. I hope that thinking of AHSs as two-level models will help to clarify the nature of the requirements that we place on the hypertext half of AHSs.

References

collier87: George H. Collier, "Thoth-II: Hypertext with Explicit Semantics." In Hypertext '87 Proceedings, 269-289. 1987.

furuta89: Richard Furuta and P. David Stotts, "Programmable Browsing Semantics in Trellis." In Hypertext '89 Proceedings, pages 27-42. 1989.

guha91: R. V. Guha and D. B. Lenat, "CYC: A Mid-Term Report." Applied Artificial Intelligence 5:45-86. 1991.

hu93: Michael J. Hu and Peter Kirstein, "An Intelligent Hypertext System." In Proceedings of the Workshop on Intelligent Hypertext, James Mayfield and Charles Nicholas, eds. 1993.

kheirbek93: Ammar Kheirbek, "A Two-Level Hypermedia Model Based on Conceptual Graph Theory." In Proceedings of the Workshop on Intelligent Hypertext, James Mayfield and Charles Nicholas, eds. 1993.

mayfield93: James Mayfield and Charles Nicholas, "SNITCH: Augmenting Hypertext Documents with a Semantic Net." International Journal of Intelligent and Cooperative Information Systems 2(3):335-351, 1993.

miller90: George A. Miller, Richard Beckwith, Christiane Fellbaum, Derek Gross and Katherine Miller, "Five Papers on WordNet." Princeton University Cognitive Science Laboratory technical report #43, July 1990.

neches91: R. Neches, R. Fikes, T. Finin, T. Gruber, R. Patil, T. Senator and W. Swartout, "Enabling Technology for Knowledge Sharing." AI Magazine 12(3):36-56. 1991.

rada90: Roy Rada, "Hypertext Writing and Document Reuse: The Role of a Semantic Net." Electronic Publishing--Origination, Dissemination and Design 3(3):125-140. 1990.

trigg86: Randall H. Trigg and Mark Weiser, "TEXTNET: A Network-Based Approach to Text Handling." ACM Transactions on Office Information Systems 4(1):1-23. 1986.

JCM