[ Back ]
Michael N. Louka, March 1994
HTML conversion, June 1995
Nodes in a hypermedia network are objects that normally contain a single concept or idea. One of the best introductions to hypertext is that of Conklin (1987). He describes the hallmark of hypertext as being a "single coherent interface to the database". What the reader sees on the computer screen usually corresponds to a single node in the information space, offering a level of integration of information and interaction that other information retrieval systems do not.
Clearly, semantic networks (Quillian, 1968) and hypermedia information structures have a lot in common in that they are both (ordinarily) networks of related concepts, representing knowledge. The main difference between the two concepts is that a semantic network is usually a knowledge representation structure that the computer uses to perform some task, whereas a hypermedia structure is created with the intention that a human user will traverse it to perform tasks using natural intelligence. In a time when the amount of information that managers and scientists are having to cope with is increasing rapidly, techniques from the fields of AI combined with hypermedia navigation techniques show some promise as tools for locating and recording information.
AI techniques have been applied to hypermedia systems for two purposes -- structuring information and improving information retrieval. Tools and systems have been devised that can be used to generate hypermedia information structures using AI techniques, and AI techniques have also been used to create 'intelligent' search mechanisms that attempt to locate information in hypermedia structures, using, for example, fuzzy searches (search mechanisms based on fuzzy logic). A number of authors (e.g. Waterworth, 1992:184, and Nielson, 1990:190) predict that personal agents that can traverse hypermedia information structures looking for information will eventually be common. Agents could work by either monitoring an information structure while the user is doing other things, and alerting the user if some important new information becomes available, or by guiding the user while he or she manually navigates through an information space. If the agent knows what kind of information the user is looking for, it can give hints as to which routes through the structure might be most appropriate to the reader.
This paper examines how AI techniques are being applied to hypermedia systems to create the underlying knowledge representation. The first set of techniques discussed look at how natural language processing can be used to generate nodes and links from raw information, usually to create a static hypermedia structure. A system called StrathTutor is then examined, which uses a simulation model of human memory to create dynamic links between nodes.
Knopik and Ryser (1989) have written a text analyser that parses a textual document looking for nouns that can be used as candidate 'concepts'. Potential candidates are compared with existing concepts in a hypertext network after converting them into a standard form using heuristic methods.
The goal for the tool is to build up part of a network depending on the content and type of the nodes linked, and to create new nodes depending on the semantics of the given type of links and the contents of existing nodes.
Candidates are processed in the following ways:
The system is limited to three kinds of links, which are 'super-concept-of', 'is-defined-by', and 'uses-concept' links. These are used to determine whether a new link is appropriate. It also contains number of rules for processing potential candidates. The tool works well for texts that don't challenge it's limitations, but it's significance lies in that it demonstrates that the application of natural language processing to the task of creating hypertext information structures does have potential. The authors believe that more sophisticated natural language processing is required to make the tool more accurate.
Another tool is HEFTI (Hypertext Extraction From Text Incrementally, Chignell et al.,1990). It has seven sequential modules through which a text is passed in order to generate a hypertext information structure. The modules perform text preparation, node preparation, indexing, link creation, organisation and link refinement. The modules can be altered individually in order to evaluate different techniques for each process and so the system can easily be improved as better techniques become available. Unfortunately the authors do not describe the actual language processing techniques that are used to create and refine links.
Rada (1992) has created a tool that converts textbooks into hypertext by parsing a text in a markup language format (typical document markup languages include LaTeX and SGML) to create a semantic net of link objects. These link objects point to text blocks. The text blocks are sequentially numbered, since a textbook is usually a linear sequence of text blocks separated by section headings. Bibliographic citations are then located and links to the book's bibliography are generated. Links to cross-references and footnotes that are explicit in the markup of the original document are also created appropriately. The index defined in the original document is then placed in an index node and links between items in the index and their corresponding text blocks are established. Rada calls this process, which directly reflects the markup of the original document, 'first-order hypertext'. 'Second-order hypertext' is defined as hypertext that also contains links that are not explicit in the markup of the original document.
Second-order hypertext is created after first-order hypertext by examining word patterns in order to identify new links, which are then used to generate a derived index. The quality of the second-order hypertext links is dependent on the computer's ability to generate links that are relevant, a task that probably needs more advanced AI techniques than those employed by Rada, who, for this tool, simply uses word frequencies and patterns.
The derived index is generated by identifying key terms in each text block and examining the distribution of these terms across all of the text blocks. Terms can be added to the index, and similar terms can be linked together, forming a semantic network. Quillian's (1968) idea of spreading activation in a semantic network can be applied to implicitly create links, which can be explicitly stored in the hypertext information structure.
The authors considered that "since the human memory is the largest, most complex and interconnected knowledge base that we know" it was reasonable to assume that successful modelling of that system would yield valuable insights into how to represent and retrieve information effectively from large hypermedia systems.
The following assumptions have been made:
The basic idea of Hintzman's simulation model of human intelligence is based on work in the field of memory theory described in 1923 by Richard Semon. His fundamental idea was that all activated memory traces respond in parallel during retrieval. Consequently, "[a]bstract ideas will arise when individual traces having common properties are activated by the same retrieval cue. Semon argued that the contents of consciousness are produced by a kind of resonant state (homophony), in which distinctive features of the activated memory traces mutually interfere. Thus the shared properties of all such traces will stand out or 'resonate'" (Kibby & Mayes, 1993:140). Hintzman took Semon's theory and created a simple, elegant, computational model.
An experience or event is represented as a trace or vector, which is an
ordered list of attributes. A three-state attribute system is used:
An index of similarity is computed between the probe and each trace for each retrieval and each trace is then activated by a weighted function of its similarities to the probe, wi. The echo, e, is computed by summing all weighted traces:
The echo comprises a kind of profile across attributes encoded in the traces. It has both intensity and content, partly reflecting the number of traces matching the probe and the pattern of attributes that is returned after the matching of the probe against all individual traces.
In StrathTutor, Hintzman's memory traces become multiple areas in each frame, which the author's call 'hotspots', with each hotspot having a subset of attributes attached to it. "Each frame is represented as a profile of attributes, summed across all hotspots in that frame. The probe is generated by selecting a hotspot or a menu item" (Kibby and Mayes, 1993:142).
As a hypermedia system, StrathTutor is particularly elegant because it allows for authoring without programming, without links between nodes being explicitly represented, as they are computed at run-time using pattern-matching heuristics according to the type of interaction the user initiates.
A major problem with StrathTutor is that it requires an enormous amount of computation. The complexity of the computation increases as the information structure gets larger, making the approach impractical for large dynamic hypertexts using conventional non-parallel computer technology. Ideally, simultaneous comparison of a probe with everyframe (node) in the structure is desirable. Hintzman's model is ideally structured for parallel computing as each trace is required to react to the probe simultaneously, making a weighted contribution to the echo. The creators of StrathTutor claim that advances in the area of parallel computing will make their approach more attractive.
AI techniques have been demonstrated to improve the creation of certain types of hypertext structures, but there is clearly room for improvement, especially to tools using techniques based on word frequency rather than semantic analysis. New discoveries, particularly in the field of natural language processing, will improve the relatively simple methods currently in use.
Conklin, J. (1987) Hypertext: A Survey and Introduction. IEEE Computer, 20 (9). pp. 17-41.
Hintzman, D. L. (1986) 'Schema abstraction' in a multiple-trace memory model. Psychological Review 93. pp. 411-428.
Jonassen, D.H. (1989) Semantic network elicitation: tools for structuring hypertext. In R. McAleese and C. Green (eds.) Hypertext: State of the Art. Oxford: Intellect.
Jonassen, D.H. (1993) Effects of Semantically Structured Hypertext Knowledge Bases on Users' Knowledge Structures. In C. McKnight, A. Dillon, and J. Richardson (eds.) Hypertext: A Psychological Perspective. Chichester, England: Ellis Horwood Limited. pp. 153-168.
Kibby, M. R. and Mayes, J. T. (1993) Towards Intelligent Hypertext. In R. McAleese (ed.) Hypertext: Theory into Practice. Oxford: Intellect. pp.138-144.
Knopik, T. and Ryser, S. (1989) AI methods for structuring hypertext information. In R. McAleese and C. Green (eds.) Hypertext: State of the Art. Oxford: Intellect. pp.224-230.
Nielsen, J. (1990) Hypertext and Hypermedia. New York: Academic Press.
Quillian, M. R. (1968) Semantic Memory. In M. Minsky (ed.) Semantic information Processing. Cambridge, MA: MIT Press. pp.227-270.
Rada, R. (1992) Converting a Textbook to Hypertext. ACM Transactions on Information Systems (10) 3. pp. 294-315.
Waterworth, J.A. (1992) Multimedia Interaction with Computers. Chichester, England: Ellis Horwood Ltd.