KEYWORDS: Automated linking, typed links, structured hypermedia system
ASK systems are a class of structured hypermedia system developed at the ILS. ILS has built over a dozen ASK systems in domains as diverse as business process reengineering, recent American history, and social services for Mexican immigrants. In contrast to most hypermedia systems, ASK systems are based on the metaphor of a conversation with an expert . In particular, they are designed to simulate a question-answer dialog in which the user asks the questions and the system provides the answers. An interaction with an ASK system consists of two phases: first a user zooms to a story relevant to his interests, then he browses the story, asking follow-up questions as desired and retrieving additional stories in response. If the user has additional top-level questions, he may return to the top-level of the system to begin another round of zooming and browsing.
To support browsing, ASK systems contain a rich network of links, each of which joins a source story which raises a question to a target story which answers it. The browsing interface in an ASK system surrounds the current story with specific questions it might raise for the user, grouped by question type. The specific questions mark browsing links; clicking on one takes the user to a target story which provides an answer. If a user has a specific question in mind, the question types enable him or her to quickly find it. If the user has only a vague idea of what question to ask, he or she can look under a generic question that looks promising.
Hence, in ASK systems, generic questions serve as link types, though links are also labeled with specific questions. The set of link types used in ASK systems is inspired by a simple theory of conversation which argues that at any point in a conversation, there are only a few general categories of follow-up statements that constitute a natural continuation rather than a topic shift . These "conversational associative categories" (CACs) can also be thought of as the general classes of questions a person is likely to formulate in conversation. The particular link types used in ASK systems are based on a set of CACs tailored for conversations about problem-solving. Figure 1 shows the set most commonly used.
Refocusing: Adjustments to the specificity of topic under consideration. 1. Context: What is the big picture within which the current topic fits? 2. Specifics: What are the details of this situation or an example of it? Comparison: Related topics at the same level of abstraction as the current one. 3. Analogies: When have other similar situations occurred? 4. Alternatives: What other approaches exist, or what did other experts say? Causality: Explanations and outcomes3 5. Causes (or earlier events): How did this situation develop? 6. Results (or later events): What is the outcome of this situation?. Advice: Planning knowledge for use in the problem solver's situation. 7. Opportunities: How can I capitalize on this situation? 8. Warnings: What should I watch for that might go wrong?Figure 1: The eight CACs most commonly used in ASK systems
The current manual process, called "question-based linking," has proven reliable. In this process, professional indexers attach to each story a list of the questions it raises and those it answers. They then create links by attaching questions raised to semantically equivalent questions answers . This process (or simple variations on it) has been used to build over a dozen ASK systems. As a sign of its reliability, we are now able to accurately predict how long building a new system will take. However, the method has several drawbacks:
* Time-consuming: It requires approximately 5 person-weeks of effort to link each hour of video stories or 100 text stories.
* Expertise-intensive: Novice indexers do not pose the same breadth and quality of questions as experienced ones. The need for specialized indexing expertise becomes particularly nettlesome as we move towards the goal of having subject matter experts build ASK systems in the course of doing their jobs (e.g., a team of designers who collaboratively construct a system to capture design rationale).
* Indexer habituation: As indexers work with a body of content over time, they stop posing many common novice questions (such as questions about the definitions of terms) in contexts in which they are relevant.
* Limits on "density" of content: When a system grows to have many stories in a single topic area, the questions they answer are frequently similar, and question-based indexing becomes cumbersome.
Three criteria are critical when evaluating the results produced by an automated linker. Thoroughness measures how completely the linker supports the three tasks in the linking process. Recall rate measures the percentage of the links which an ASK system should have that the linker actually generates. Precision rate measures the percentage of links that a linker actually generates which are links that an ASK system should have. An additional criterion is useful to judge the input a linker requires. Ease of use judges how hard is it for an indexer to learn and use the representational scheme a linker uses.
Note that the relative importance of these criteria depends on the target audience of indexers. Subject matter experts who build ASK systems as job aids require a linker that is easy to use even if it performs only moderately well on the results criteria. Professional indexers can invest the time to learn and use a more complicated linker if it produces better results.
Unfortunately, global similarity metrics are not suitable for building browsing links for a structured hypermedia system because they rely on lexical features rather than a semantic analysis of stories. The most critical impact of this shortcoming is on thoroughness. When a global similarity metric suggests that two stories should be paired, it cannot determine which relationship holds between them. For instance, a global similarity metric might determine that the two stories in Figure 2 should be linked since they both mention the lexical term "simulator." However, the metric could not determine that the second story provided specifics about the first. Additionally, global similarity metrics will have limited recall because some stories which share few terms in common should be linked. For example, two different plans for solving a problem should be linked as alternatives, though the stories themselves may not share any terms.
The general approach underlying HieNet is appealing. A linker capable of making links analogous to those found in a training corpus could potentially construct a wide variety of links. However, HieNet as implemented has a significant disadvantage. Because it fails to abstract why stories were linked, it offers little ability to transfer. If an analogy-maker is to achieve transfer, it must either determine the abstract structure which underlies links or be given access to a representation which explicitly labels this structure.
Nanard & Nanard's system can serve as the foundation for a useful automated linker and the four linking methods we discuss below share several aspects with it. In particular, we require indexers to create hand-coded conceptual representations of stories and two of our methods employ linking rules. Furthermore, one of our methods using a representation quite similar to Nanard & Nanard's qualified
Source Story: Different Simulators for Different Skills Tools such as the flight simulator pave the way for a natural, effective way to learn physical skills. To teach social skills in a similarly effective way, we need to build social simulators.... Target Story: Dustin We have built a number of social simulations that provide learning-by-doing environments. Dustin is an example of a simple simulator that helps students learn a foreign language through using it....
Linking Task Output for Example Link 1) Determine that the two stories are They are both about social simulation related 2) Determine how they are related The target story provides specifics (i.e., which link type joins them) about the source story 3) Build a question to bridge from the "What is an example of a social source story to the target story simulation?"Figure 2: The three tasks a linker must perform
concepts. However, our work differs in two important ways from theirs. Their method requires a daunting amount of theory-building from its users. In contrast, we wish to provide indexers with predefined theories of linking (i.e., sets of linking rules) so that they do not need to specify how the linker should go about joining stories. Also, two of our methods use representational frameworks which offer significantly more expressive power than qualified concepts.
To construct links between stories, the automated linker employed linking rules each of which inferred links for a specialized sense of one CAC. Conceptually, a rule performs a pairwise comparison between the frames representing two stories. For example, a rule might specify that two stories whose agents who have the same Goal but employ different Plans should be linked as alternatives.
We ran an informal experiment to see whether narrative linking would be effective when used by novice indexers. A team of two Northwestern University undergraduates were given a corpus of 47 stories in the domain of military logistics (a subject about which they knew nothing) and the empty narrative frame. They created taxonomies of fillers, instantiated the frame for each story, and defined inference rules for four of the eight CACs. We then ran the rules on the frames, thereby producing approximately 120 story-to-story links. Of these, approximately 80 were judged to be correct by expert indexers. Twenty more were judged to join two stories that were actually related, but not in exactly the way the proposed link indicated (e.g., the direction of causality was wrong). The remaining 20 were judged to be incorrect.
These results indicated that novices could build narrative frames and that an automated linker could use them to construct links with high precision (i.e., 20 incorrect links out of 120 proposed). Further, narrative linking was thorough, supporting each of the three linking tasks: relating stories, assigning conversational categories, and generating questions from a representational template associated with each linking rule.
Because we did not manually construct a normative set of links that a linker should create among the sample stories, the experiment did not provide a measure for recall. However, narrative linking's recall rate is limited because narrative frames cannot conveniently encode information which does not revolve around intentionality. For example, making the link in Figure 1 requires knowing that "Dustin contains a social simulation." Because it cannot capture this information, the narrative linker will be unable to build this link.
Nevertheless, the primary problem with narrative linking was not in the output it produced, but rather in the input it required. The method was not easy to use. The students worked a total of approximately 160 person-hours to represent 40 stories. An experienced indexer could link the stories by hand in roughly half the time it took the students to engineer the representation! Furthermore, the students had difficulty distilling stories down into narrative frames which were consistent with those they developed for similar stories.
Our work with narrative linking taught us that an automated linker can successfully use structured representations, but that they are difficult for novice indexers to create. It is particularly difficult to create representations which are global (i.e., which integrate a story's representation into a single overarching structure) and yet which also contain enough detail to enable an automated linker to relate stories appropriately.
As an example, consider the link in Figure 2. To create this link, an indexer would first need to assign concept sets to each story that shared at least one concept (e.g., "social simulation"). The linker would then suggest the stories be paired because of the shared concept. To complete the link, the indexer would need to add the link type, specifics, and the question "What is an example of a social simulation?".
Simple concept linking requires the indexer to construct a taxonomy of "important" concepts, which is unstructured but is otherwise roughly analogous to the taxonomies of fillers used in narrative linking. In theory, "important" concepts are those which underlie the relations between stories which should be linked. In practice, it is easier for indexers to think of them as ones which capture key distinctions in a system's content. In our experience, an indexer can produce a workable taxonomy in a single pass over a set of stories.
Simple concept linking has several attractive features. Because it uses a minimal representation, it is easy to use. It does require that indexers build a taxonomy of concepts, but to the extent that new hypermedia systems deal with idiosyncratic topics, any linking method must at least require indexers to define a customized vocabulary. To consider some examples, it would be unrealistic to expect a predefined universal vocabulary to include concepts such as Dustin (required for Engines for Education), Roll-On, Roll-Off Ship (required for TransASK), and privatization (required for ASK NorthWest Water).
Simple concept linking also has a high recall rate. To measure this rate, we encoded each story in Engines using a hierarchy of 180 concepts. We then compared the pairs of stories which indexers linked manually to those which simple concept linking suggested. The automated method duplicated the pairings of 70% of the manual links.
Since we suspected that simple concept linking might be better able to create links for some link types (such as specifics) and than others (such as alternatives), we broke down the analysis by link type. We found that the method performed consistently across the link types.8 Its lowest recall rate was on warnings (67%) and its best was on specifics (78%). We believe analogies links to be the most difficult to create of the link types listed in Figure 1, both for human indexers and automatic linkers. Good analogies links are often abstract, joining stories which do not share low-level features. Unfortunately, the Engines system did not use the analogies link type, so the experiment did not provide a recall rate for it.
Unfortunately, the method's poor performance on the remaining two evaluation criteria make it an unacceptable choice for automated linking. Its most important shortcoming is that it wildly overgenerates potential links. Only 12% of the links it suggested were duplicated by manual links. Although this is a conservative measure of precision (as discussed above) and although it is well above chance (manual links represented 3% of possible links), this rate is unacceptably low. Furthermore, simple concept linking is not thorough. After it determines that two stories are related, it relies on the indexer to manually indicate how they are related and build a bridging question between them. This shortcoming is particularly acute given that concept linking produces so many poor links.
Armed with what we have learned from our experiments with concept and narrative linking, our current strategy is to develop two different linking methods for two different populations of human indexers. For novice indexers (e.g., subject matter experts who are constructing an ASK system as an adjunct to their primary job), we have created elaborated concept linking. This method is based on concept linking (henceforth called "simple concept linking") but adds a small amount of additional representation to address its shortcomings. For professional indexers, we have created point linking. This method is analogous to narrative linking, but introduces a new representational scheme which we expect will be easier to learn and use.
Elaborated concept linking adds a straightforward extension to simple concept linking: it requires indexers to divide the concepts they associate with stories into two groups, concepts mentioned and concepts elaborated. The indexer must label each concept in a story's concept set as either a concept mentioned or a concept elaborated (i.e., one the story says something significant about). Furthermore, elaborated concept linking requires the indexer to specify what the story says about each concept elaborated by attaching a link type and question answered to it.
For example, when representing the top story in Figure 2 ( "Different Simulators for Different Skills" ), the indexer might add social simulation as a concept elaborated, attaching the link type context to it and the question answered "Why are social simulations valuable?" Likewise, when representing the bottom story ("Dustin"), the indexer might also add social simulation as a concept elaborated, but attach the link type specifics to it and the question answered "What is an example of a social simulation?"
Elaborated concept linking proposes links from source stories which mention some concept only to those target stories which elaborate on it (or from one story which elaborates a concept to another that also elaborates it). Hence, if the stories in Figure 2 were represented as described above, the linker would suggest that they be paired by two links: a specifics link from the top story to the bottom one with the question "What is an example of a social simulation?" and a context link from the bottom story to the top one with the question "Why are social simulations valuable?"
This extension dramatically reduces the number of suggested links. Without it, simple concept linking proposed over 20,000 links in Engines. With it, elaborated concept linking proposed less than half as many. Furthermore, the extension predicted reasonably reliably which links to eliminate. An algorithm which eliminated links at random would have eliminated good links as well as bad in the same proportions that they occurred in the set of links proposed by simple concept linking (specifically, 12%). Since only 6% of the suggested links that were eliminated duplicated manual links, the extension proved to be a useful filter.
Nevertheless, this extension only raises elaborated concept linking's precision to 16%. Although this is a large jump relative to simple concept linking's 12%, it still leaves significant room for improvement. Accordingly, we are investigating a series of additional extensions which further raise precision but do not compromise the ease of use offered by concept linking methods. The next extension we are planning adds linking rules which consider not just whether stories elaborate some concept, but also how they do. It will enable the linker to determine, for example, that it makes sense to link a story which describes a problem to one which describes its solution, but not to link a story which only gives a detail about a problem to one which describes an analogy to it. If this extension does not raise precision sufficiently, we are also planning a third which views the task of building a set of links as a constraint satisfaction problem. Instead of deciding whether to include each potential link in isolation, this extension will collect the best overall set of links for each story, trading off the strength of each potential link against others attached to the same story through the same link type.
This expedient assumes that the link type and question answered assigned to an elaborated concept will be appropriate to use in suggested links regardless of the relationship between the source and target stories. Although generally true, this assumption is not flawless. In particular, it mislabels alternatives links, since such links typically join stories which elaborate the same concept in the same way. For instance, the same corpus of stories that includes the Dustin story of Figure 2 also includes a story about Yello, another computer program which contains a social simulation. An indexer who attached the concept elaborated social simulation to the Dustin story, marking it with the specifics question "What is an example of a social simulation?" would likely do the same to the Yello story. Hence, the linker would pair these stories with dual specifics links. However, they should be linked by alternatives links (e.g., "What is another example of a social simulation?") Practically speaking, we believe that this limitation is manageable since the alternatives case appears to account for the bulk of mislabeled links and this case may be caught automatically.
In elaborated concept linking, indexers must not only attach concepts to stories, they must attach link types and questions to elaborated concepts. To enable the linker to also provide guidance about appropriate link types and questions, we are developing a taxonomy of common questions, categorized by link type, for each predefined concept. When an indexer enters an elaborated concept into the linker, this taxonomy will enable the linker to offer suggestions of already-typed questions in response.
This reduction in recall is significant enough to represent a new shortcoming which we desired to address. To investigate how to improve recall, we analyzed a sample of 100 manually-created links which elaborated concept linking failed to duplicate to determine the cause of the failure.
AttributeTable 1: The predefined concept hierarchy
This analysis isolated three reasons why elaborated concept linking failed. However, the first two shed little light on how to improve it. Thirty of the missed links were caused by incomplete concept sets (i.e., concept sets which failed to include a concept needed to make the link). Although these links would have been made given perfect concept sets, it is unrealistic to expect humans to produce flawless representations, even if they are given automated assistance.
For another thirteen of the supposedly "missed" links, elaborated concept linking actually behaved correctly. Upon review, the manually-constructed links which the method
"missed" were themselves found to be flawed; they joined source stories to targets that were not relevant to them.
However, the remaining (and most common) cause of missed links does suggest a way to improve recall -- adding more powerful forms of inference to the linking algorithm. Elaborated concept linking currently performs only a single form of inference, intersecting concept sets to infer that two
Type of MissedTable 2: Types of inference required to pair stories
Example Inference Links Taxonomic 5% A story which discusses Dustin might be related to a story which discusses educational software because Dustin is a piece of educational software. Encapsulated 5% A story which discusses educational software might topics be related to a story which discusses education since the topic educational software encapsulates the topic education. Alternatives 4% A story which discusses learning-by-doing might be related to a story which discusses memorization since they are alternative ways of learning. Planning/
ca 3% A story which discusses educational software might usality be related to a story which discusses learning because educational software can cause learning. Bridging 10% A story which discusses educational software might bridge to a story which discusses teaching.
stories are related. When pairing stories requires more complex inference, as it did in 57 of the missed links, we found that five additional types of inference, displayed in Table 2, were sufficient to account for the missed links.
The last category, bridging inference, merits further discussion. Most browsing links in ASK systems are based on questions users are likely to raise when they view a source story (such as "Why are social simulations valuable?" when reading the Dustin story). In contrast, bridging links introduce content that users might not have known to look for, but which might draw their interest when presented to them (such as "Why is it important to let students learn by doing?"). Users who wish to broadly sample a system's content may use bridging links to navigate between "clusters" of tightly related stories. However, bridging links are difficult to construct automatically because the relationship between the stories they join is often somewhat tenuous and therefore difficult to infer
Point linking is our attempt to develop such a method. It addresses three shortcomings of narrative linking:
* Representational approach: Narrative linking attempted to capture the global structure of a story. Point linking uses local "snippets" of representation which attempt to capture just those aspects of a story which would cause an indexer to create a link involving it.
* Representational guidance: Narrative linking requires the indexer to build a custom taxonomy from scratch for each slot in its representational frame. Point linking provides predefined taxonomies for its slots. Only one of these taxonomies needs to be extended by the indexer.
* Domain coverage: Point linking can be used with any type of content, not just intentional chains.
One source of motivation for point linking comes from observing how experienced indexers perform manual question-based linking. When indexers argue about why two stories should be linked in a particular way, they often frame the argument in terms of terse statements about the contents of those stories. An indexer might say, for instance, "We should add an alternatives link because the first story states that `necessity leads to innovation' and the second states that `hard work leads to innovation.'" Statements like "necessity leads to invention" are exactly what we aim to capture in a representation of points. Why not ground a language used to support linking in the sorts of explanations that expert indexers use?
Another source of motivation for point linking comes from observing how good readers behave. When good readers read a text, they boil it down, abstracting its main points. These points help readers to determine what questions they should ask themselves about a text and to integrate what they read with what they already know. If people use summarized points internally to attach new knowledge to pre-existing knowledge, why not use them in an external automated system which does the same?
Slot Name Legal Fillers Example Concept-1 (Any concept) Simulations Mode Do, Should, Can Do Sense Indeed, Not, Indeed15 Anti Relation (Predefined Enable relation) Concept-2 (Any concept) Learning-By-Doin gTable 3. The point frame
The two concept slots indicate what topics an author addresses in a point. As in the other linking methods, indexers are expected to expand the set of legal concepts to reflect the universe of discourse in their particular ASK system. To provide guidance, the representation language for points provides a predefined concept hierarchy (identical to that provided for elaborated topic linking).
The mode, sense, and relation fields indicate what an author says about the topics in a point (i.e., how the concepts in a point relate to each other). We claim that there are only a small number of important ways in which speakers (and authors) relate concepts to each other. So, the language provides comprehensive, fixed vocabularies for the three fields which deal with how concepts interrelate. The example point demonstrates the strength of this approach. It seems reasonable that a language should contain predefined terms such as Do, Indeed, and Enable, but unreasonable to expect it to include Simulations and Learning-By-Doing.
The mode and sense fields allow an indexer to twist the meaning of a point, while the relation field provides the pivot around which the meaning may be twisted. The point language provides a predefined hierarchy of relations (Table 4). Unlike the concept hierarchy, this one is not expected to be routinely extended by indexers. The relations it contains are intended to capture (sometimes at a high level) the important conceptual relationships which hold between concepts in the points people make. We have encoded the stories in Engines using the point language and have found the current hierarchy sufficient to capture the relationships required for the approximately 750 points which resulted.
Have-Causal-RelationTable 4: The predefined relation hierarchy
Apply-Theory Have-Subclass Contain-Step
Have-Economic-Relatio Consume n Create
Have-Storytelling-Re Know lation
Have-Teaser Have-Value-Judgment Have-Time-Relation
To determine which points raise and answer questions, the method employs a library of "point association questions" (PAQs). Like the linking rules used in narrative linking, PAQs operate on structured representations of stories to
Join the source point: <Concept-1 Does Indeed Contain <Concept-2>] To the target point: <Concept-3> Does Indeed Contain <Concept-2> Through the question: "What else also contains a <Concept-2>?" Using the link type: AlternativesFigure 3: An example point association question
infer links which correspond to a specialized sense of one of the CACs. More specifically each PAQ captures one type of question used in ASK system browsing links. Figure 3 shows an example.
To create links, the linker considers each point in a corpus of material as a potential source point, creating questions the point raises and then locating answers to them. For example, the example PAQ enables the program to create the link shown in Figure 2. One of the points the source story in this link makes is Dustin Does Indeed Contain Social Simulation. Since this point matches the PAQ's pattern for source points , the program instantiates the pattern for questions, thereby raising the alternatives question "What else also contains a social simulation?" To find target points, the program also instantiates the pattern for target points, creating the more specific pattern <Concept-3> Does Indeed Contain Social Simulation. The program then searches for target points which match the pattern, locating the point George Does Indeed Contain Social Simulation, which is one of the points made by the "George" story. Since this point answers the question, the linker draws a link between the "Dustin" and "George" stories, labeling it with he alternatives link type and the appropriate question.
We do not yet have empirical results from point linking. However, we expect that it will achieve a recall rate which approaches that of elaborated concept linking, though it will be slightly lower. The reason for the difference is that point linking as described here lacks one type of information available to elaborated concept linking: the set of concepts which a story mentions but does not elaborate (generally, only those concepts which are elaborated will be included in points). As a practical matter, this means that point linking will likely be less likely to infer context links which give general descriptions of peripheral concepts.
We also expect that point linking will have a precision rate which approaches that of narrative linking because both methods use structured descriptions and explicit inference rules to privilege the most important links between stories. Elaborated concept linking has low precision because its representation provides scant information about the roles that concepts play in the stories that discuss them (and the method uses even less than is provided, since it does not currently consider how concepts are elaborated when creating links, only that they are). Narrative linking has high precision because its structured representation makes clear the roles concepts play in stories. We found, however, that most narrative linking rules tapped into only a few slots of the narrative frame. Hence, simpler structures should be able to achieve comparable results. Points are such simpler structures, designed to capture just those local fragments of representation which are important to make good links (including fragments that narrative frames could not capture because of their reliance on the intentional chain).
* Indexers have difficulty constructing global, structured representations. Hence we have devised linking methods which require simpler "snippets" of representation. We have also added predefined taxonomies which provide indexers with representational guidance.
* Different audiences of indexers require different linking methods. Specifically, novice indexers require methods with less complex representational demands than expert indexers.
* The central challenge a linking method faces is to filter out poor links without eliminating good ones. Hence, the primary reason to add knowledge to a linking system is to address this challenge.
* A straightforward linking method, simple concept linking, requires a minimum of representation and can suggest a large proportion of the hypermedia links that should be created. The critical shortcoming of this method is that it also suggests many poor links.
* Elaborated concept linking, a method which adds a minimum of representation to redress the shortcomings of concept linking, appears to suit the needs of novice indexers who require an easy-to-learn and easy-to-use method, but at the cost of poor precision.
* Point linking should suit the needs of professional indexers who can invest the effort to learn a more complex representation to achieve more thorough and accurate linking.
2. J. Conklin, M. L. Begeman, gIBIS: A hypertext tool for exploratory policy discussion, The ACM Conference on Computer-Supported Cooperative Work (CSCW'88) 1988), pp. 140-152.
3. D. C. Edelson, D. K. O'Neill, The CoVis Collaboratory Notebook: Supporting collaborative scientific inquiry, Proceedings of The 1994 National Educational Computing Conference 1994),
4. R. Bareiss, R. Osgood, Applying AI models to the design of exploratory hypermedia systems, The Proceedings of the Fifth ACM Conference on Hypertext (ACM Press, Seattle, WA, 1993), pp. 94-105.
5. R. C. Schank, Cognitive Science I, 421-441 (1977).
6. C. Cleary, R. Bareiss, A descriptive model of question asking during story acquisition interviews, Proceedings of the Sixteenth Annual Conference of the Cognitive Science Society Atlanta, GA, 1994),
7. R. Osgood, Ph.D., Northwestern University (1994).
8. R. C. Schank, C. Cleary, . (Lawrence Erlbaum Associates, Hillsdale, NJ, 1995),
9. M. Bernstein, An apprentice that discovers hypertext links, Proceedings of the European Conference on Hypertext Versailles, France, 1990), pp. 212-223.
10. G. Salton, J. Allan, Selective text utilization and text traversal, The Proceedings of the Fifth ACM Conference on Hypertext (ACM Press, Seattle, WA, 1993), pp. 131-144.
11. D. T. Chang, HieNet: A user-centered approach for automatic link generation, The Proceedings of the Fifth ACM Conference on Hypertext (ACM Press, Seattle, WA, 1993), pp. 145-158.
12. J. G. Carbonell, in Machine learning: An artificial intelligence approach(Morgan Kaufmann, Los Altos, CA, 1986) pp. 371-392.
13. J. Nanard, M. Nanard, Should anchors be typed too? An experiment with MacWeb, The Proceedings of the Fifth ACM Conference on Hypertext (ACM Press, Seattle, WA, 1993), pp. 51-62.
14. R. Osgood, R. Bareiss, Automating index generation for constructing large-scale conversational hypermedia systems, Proceedings of the Eleventh National Conference on Artificial Intelligence (AAAI Press, Menlo Park, 1993), pp. 309-314.
15. R. C. Schank, R. P. Abelson, Scripts, Plans, Goals, and Understanding (Lawrence Erlbaum Associates, Hillsdale, NJ, 1977).
16. R. Spiro, J. Jehng, in Cognition, Education, and Multimedia: Exploring Ideas in High Technology D. Nix, R. Spiro, Eds. (Lawrence Erlbaum, Hillsdale, NJ, 1990).
17. A. Collins, J. S. Brown, S. E. Newman, in Knowing, learning, and instruction: Essays in honor of Robert Glaser L. B. Resnick, Eds. (Lawrence Erlbaum Associates, Hillsdale, NJ, 1983) pp. 453-494.
18. R. C. Schank, et al., Cognitive Science 6, 255-275 (1982).