Foltz, P.W. (1996) Comprehension, Coherence and Strategies in Hpyertext and Linear text. In Rouet, J.-F., Levonen, J.J., Dillon, A.P. & Spiro, R.J. (Eds.) Hypertext and Cognition. Hillsdale, NJ: Lawrence Erlbaum Associates.
Hypertext presents a new way to read on-line text that differs from reading standard linear text. Text is typically presented in a linear form, in which there is a single way to progress through the text, starting at the beginning and reading to the end. However, in hypertext, information can be represented in a semantic network in which multiple related sections of the text are connected to each other. A user may then browse through the sections of the text, jumping from one text section to another. This permits a reader to choose a path through the text that will be most relevant to his or her interests.
The features in hypertext supply flexibility to the reader when compared to reading linear text such as books. Clearly some of this flexibility does exist in books (e.g. table of contents and indexes), but it is not as widely used or exploited. Hypertext permits readers to use these features automatically rather than requiring readers to manually refer to them as needed. This provides additional control to the reader in determining the order that the text is to be read, and allows the reader to read the text as if it were specifically tailored to the reader's background and interests. This flexibility does promise an advantage of personalization and eases the burden of finding information, However, is this flexibility actually good or useful to the reader?
The concept of using the associative paths in hypertext to retrieve and read information has caused great excitement. The promise of universally available hypertexts has been touted as "a seamless and reunited computer world" (advertisement for the Xanadu Hypertext ), having "the potential to become a significant application area; equaling or perhaps exceeding that of word processing, spreadsheets and general database applications" (Begoray, 1990, p. 121), and "because hypertext has the power to change the way we understand and experience texts, it offers radical promises and challenges to students, teachers and theorists of literature" (Landow, 1989, p. 174). Using associative retrieval paths is similar to the way retrieval is performed from human memory, and this may be part of the appeal of hypertext to researchers and developers when they state that hypertext systems will improve a user's ability to find and use information.
While a variety of hypertexts have been developed over the past 20 years, it is often not clear whether there are strong advantages for hypertext. Research in hypertext has often failed to show any significant advantages for reading a hypertext compared to the equivalent text in linear form. In addition, the effectiveness of various features that can be used in hypertexts can vary greatly depending on the domain and content of the text and the goals of the reader. Up to this point, no standards or definitive rules exist on how to develop an effective hypertext. However, because hypertext encompasses such domains as user-interface design, psychology, education, and information retrieval, theory from these domains can be applied to hypertext in order to aid in the understanding and development of effective hypertexts.
Over the last 40 years, a large body of research has developed which focuses on linear text comprehension, both from a theoretical and an evaluative standpoint. Text research has permitted predictions of comprehension based on such factors as the structure of the text, the background knowledge of the reader, and the reader's abilities. This chapter examines the evaluation of hypertext from the perspective of text comprehension.
The development of hypertext systems has created research into how to design and use hypertexts. Much of this research has focused on computer science and interface issues rather than the cognitive aspects of hypertexts. Some of the major areas of hypertext research include: the development of the underlying representations of information and connections (e.g., Botafogo & Shneiderman, 1991; Furuta & Stotts, 1989), methods of connecting, structuring and retrieving the information (e.g., Croft & Turtle, 1989; Crouch, Crouch, & Andreas, 1989), designing hypertexts for supporting argumentation (e.g., Conklin & Begeman, 1989; Fischer, McCall, & Morch, 1989), studying information seeking in hypertexts (e.g., Marchionini & Shneiderman, 1987; Weyer, 1982), and the role of rhetoric and writing in hypertexts (e.g., Bolter, 1991; Britton & Glynn, 1989).
While there have been a variety of areas of research on hypertext, one of the primary goals of the research has been to evaluate hypertexts and hypertext features in order to bring about improvements. From this research, we can then get an idea as to how well hypertexts have succeeded and what hypertext features are the most effective.
Research comparing hypertext to linear text provides some measure as to what situations will be best suited for the presentation of textual information using hypertext. While there have been claims of the superiority of hypertext over linear text (e.g., Martin, 1990; Nelson, 1967), research into this area has not been uniformly successful in showing an advantage for hypertext. The chapter by Rouet and Levonen in this volume discuss the details of these studies. Overall, the results of the studies vary greatly. A large part of the variability may be due to the different tasks, hypertext systems and the types of texts used. Therefore, the results indicate that there are several factors that must be considered in order to determine whether a hypertext or linear text would work best.
Because hypertexts have incorporated many new features that are not found in linear text, such as linking, the ability to search for information, guided tours and overview maps, it is often not clear which of the features may be effective in improving the users' ability to read and comprehend the text. Without information about what features work best, it is difficult to know what features to incorporate into the design of a new hypertext. For this reason, there have been studies comparing the various hypertext features.
One factor that may affect reading hypertext is the resolution of the screen. Work by Gould (Gould & Grischkowsky, 1984; Gould, Alfaro, Fonn, Haupt, Minuto, & Salaun, 1987) found that reading time is significantly slowed when reading a text on a computer screen as compared to reading the same text on paper. However, when a high resolution screen with anti-aliased fonts are used, the reading speeds are equivalent. Thus, studies comparing linear text on paper to hypertexts on screens may find some reading differences merely due to the resolution of the text rather than any factors of the actual structure of the text. For this reason, research that makes comparisons of hypertext and linear texts should ensure that they both have equivalent resolution of the text.
With the multiple paths that are possible when reading a hypertext, there is a greater navigation load on the reader than with linear text. In order to simplify the reader's task then, additional navigation features must be provided. In a linear text, a table of contents and an index are the two typical navigational features. Since linear text is often organized hierarchically, the table of contents serves as an outline with pointers to pages for each entry. However, in a hypertext, the structure may be much more complicated. For that reason, a common feature in a hypertext is a map. A map permits a reader to see a representation of the text's structure along with the relationships between the different text sections. Monk, Walsh & Dix (1988) found that with a map, subjects were much faster at being able to answer questions in a hypertext than with no map. However, in a similar study, Hammond & Allinson (1989) found no differences in answer accuracy or task time whether or not a map was provided. Nevertheless, they found that with a map, subjects did tend to visit more of the text nodes than the subjects who did not have a map. There were large differences in the style of navigation based on the subject's task. Subjects who were instructed to just explore the information space tended to use a guided tour facility, while subjects with the more directed tasks of answering particular questions used mostly the map and index to help navigate.
Thus, we see that there are varying results as to whether certain features aid the user of a hypertext. In addition, factors such as the user's goals and background knowledge can influence whether a particular feature is useful. Nielsen (1989) performed a meta-analysis of 92 benchmark measures that had been taken from usability issues that had been tested in hypertext research. In all of the measures he examined, he found that there were actually very few results which had large effects. From this analysis, he concluded that there are actually very few studies that have shown the real world impact of hypertext systems. In addition, the lack of strong effects in this area of research could mostly be explained by large individual differences among users, tasks and texts. Thus, while there has been a lot of research in hypertext, few concrete results exist to provide strong evidence on how well hypertexts will work or guidelines on how to create effective hypertexts.
Another source of how to develop hypertexts comes from development guidelines. Several books and articles have been published containing guidelines and rules for hypertext development (Nielsen, 1990; Martin, 1990). These guidelines concentrate on issues such as how much text should be contained in a node, what hypertext features to use, and how the information should be structured. However, most of these guidelines do not provide concrete design rules based on theory, but instead present abstract rules based more on common sense. The fact that the rules are fairly abstract is not surprising given that writing a hypertext is as difficult, if not more difficult, than writing linear text. With linear text, we do have some guidelines, but they still are somewhat abstract, and effective writing depends more on large amounts of practice than just being able to follow a set of writing rules. In addition, the style of writing will depend greatly on the points the author wants to express and the domain in which the author is writing.
The studies described above shed some light on what features may be used to develop a good hypertext system, however, they are very insular, only examining particular features with particular texts. Without an ability to generalize outside of the texts and tasks, with every new feature or text developed, another evaluation study would need to be performed to determine its effectiveness. An alternative approach is to use a theoretical background to drive the design of the hypertext. This then would permit comparisons of features through using the theory to make the performance predictions. Theory-based approaches have been used in the study of human-computer interactions (e.g. Card, Moran & Newell, 1984; Kieras & Polson, 1985; Lewis & Polson, 1990). These approaches use a theoretical model of the user and system in order to derive predictions of usability. They typically involve modeling such factors as the users' knowledge, possible states of the computer, and possible actions a user can take, and have had some success at predicting usability of systems and of particular features. While theory-based design has not been fully incorporated into the majority of software designers, its success for designing interfaces suggests that it can similarly be used for the design of hypertext systems.
There have been a few theoretical approaches to studying and designing hypertext. While not based on some of the strict modeling criteria used in the modeling mentioned above, these approaches do use theoretical bases or guidelines for their development and testing. One approach has been to use user-interface guidelines, such as dialog rules, to determine how a hypertext interface should be designed (Hardman & Sharrat, 1989; Shneiderman, 1987). This approach permits pieces of the interface of a hypertext to be developed based on proven guidelines. A second approach has been to use "formative design evaluation" (Egan, Remde, Gomez, Landauer, Eberhardt & Lochbaum, 1989, Landauer, Egan, Remde, Lesk, Lochbaum, & Ketchum, 1992). In formative design evaluation, development and changes of features are based on psychological guidelines. Behavioral studies are then performed on the system in order to determine what features can be improved and in what ways. This iterative design method allows a comparison of the features and quantitative measures of the improvement from previous versions. A third approach has been to develop hypertexts from the point of view of information retrieval theory. In this approach, models from information retrieval have been applied in order to determine how to structure the information. These methods include, probabilistic models of retrieval (Croft & Turtle, 1989), hierarchical clustering (Crouch, Crouch & Andreas, 1989), and petri nets (Furuta & Stotts, 1989) in order to connect the nodes of information. A more cognitively-based approach is Cognitive Flexibility Theory (Spiro & Jehng, 1990). Cognitive Flexibility Theory represents the textual information in multiple perspectives that permit the reader to learn and use the information flexibly.
While these approaches have helped improve the design of hypertexts, one area that has been neglected is to consider the text in terms of its discourse cues. Charney (1987, 1994) has suggested this area as a possible area for improving hypertexts. However, thus far, there has been little theoretical research in this area. This may be partially due to the fact that up to this point, developers of hypertexts tend to be primarily computer scientists and not psychologists with skills and background in issues of text comprehension. Thus, a different theoretical approach to hypertext is to examine it in terms of text comprehension and discourse cues.
Over the past 30 years, there has been a large amount of research in text comprehension, primarily in the fields of psychology and education. The goals of text comprehension research are to understand what factors of the reader and the text influence the ease of comprehending a text and to make some predictions as to how easy a text will be to comprehend. Through modeling both the text and the reader's knowledge and abilities, researchers have been able to both develop better texts and a better understanding of the human comprehension processes.
Research in text comprehension has examined a variety of factors that influence comprehension. These factors include: the role of coherence in a text, the role of the readers' background knowledge, the role of the narrative schema of the text, and the role of the reader's cognitive abilities. In addition, studies have examined the strategies readers use when going through a text and the role these strategies can play for comprehending a text. Not only do these factors play a major role in linear text comprehension, they also play an equally important role in the comprehending of hypertexts.
Hypertext differs from linear text in some fundamental ways. A hypertext provides more flexibility to the reader in choosing where to go in the text. A hypertext also provides the reader with more methods to employ in order to find the relevant information in the text and to move through the different sections of the text. However, aside from these differences, the primary goal of both hypertexts and linear texts is to convey information in a coherent form to a reader. In this manner, the reader should be able to extract the relevant information from the text that should reflect the intended goals of both the author and the reader. For this reason, we can examine how the results from research in text comprehension can have implications to the design and evaluation of hypertexts.
Researchers in the field of text comprehension have used user models to predict what information will be learned from a linear text. One primary approach has been to examine comprehension using the Kintsch model of text comprehension (Kintsch, 1988; van Dijk & Kintsch, 1983). The Kintsch model has been used for predicting the comprehension of text based on such factors as what features will be remembered from the text (van Dijk & Kintsch, 1983), the role of background knowledge (Britton & Gulgoz, 1991; van Dijk & Kintsch, 1983), the role of coherence and readability (Kintsch & Vipond, 1979), and also goal planning in such domains as computer mail systems (Mannes & Kintsch, 1991) and the UNIX operating system (Doane, Kintsch & Polson, 1990).
When reading a linear text, processing occurs at many levels. These levels range from the low level processes of recognizing individual words up to high level processes of deriving the gist of the information in the text. These processes work together simultaneously to extract meaning from the text. Meaning, however, is represented at different levels.
In the Kintsch model, the reader's memory for text is represented at 3 levels, a surface representation of the words and sentences, the meaning of the text (textbase) and a general representation of what is described by the text incorporating outside background knowledge (situation model). As text is read, the text is incorporated into the readers' representation of the information. Information from the surface representation of the text is quickly lost (e.g., Bransford & Franks, 1971). However, some of the abstracted information from the surface structure of the text is incorporated into the textbase, represented as propositions. Propositions serve as semantic primitives representing the information acquired (e.g. Kintsch, 1974; Fodor, Fodor & Garrett, 1975).
The propositions are connected to each other in the textbase through semantic coherence relations. In semantic coherence, constituents of the text will be coherent if they share some form of semantic relatedness in the discourse. The amount of coherence is therefore represented by the amount of shared meaning and referential relations. These coherence relations are based on standard rhetorical devices in the text, such as causality, use of pronouns, and word repetition. The semantic coherence can be represented both at the local level of individual propositions in the textbase and at the global level of the macrostructure of the text.
The mental representation of these interconnected propositions in the textbase (coherence graph) takes the form of a hierarchical structure with higher level concepts represented as super-ordinate propositions which are connected to lower level concepts represented as subordinate propositions (Kintsch & van Dijk, 1978). Recall of propositions follows this hierarchy, with propositions from the upper part of the tree being more likely to be recalled than lower level propositions (Britton, Meyer, Hodge & Glynn, 1980; Kintsch & Keenan, 1973; Meyer, 1973). At a higher level, readers generate a macrostructure or gist of the text. The macrostructure is the result of the readers' inferential processes, with readers forming hypotheses of the overall meaning of paragraphs, chapters and whole books. In this manner, the macrostructure incorporates the reader's background knowledge with the text in forming these hypotheses. The resulting macrostructure of the text is similar to the textbase in that it is also represented as a hierarchical structure with higher level concepts represented at the top.
The process of incorporating propositions into the textbase is a process of maintaining coherence. Propositions that have overlapping arguments, and thus are semantically related, create coherence. However, if the current proposition being processed does not share arguments with propositions in short-term memory, then a bridging inference must be made by the reader in order to maintain coherence (e.g., Kintsch & van Dijk, 1978). When a reader makes a bridging inference, the reader must use knowledge from his or her situation model in order to fill in the missing information. Writers often assume readers will have the appropriate background knowledge to make the proper bridging inferences, and knowledgeable readers can make these inferences automatically. However, if a reader does not have the proper background knowledge, these inferences consume additional resources of the reader, typically resulting in lower comprehension. Thus, the amount of coherence in the text, can be used to make predictions of comprehension (e.g., Kintsch & Vipond, 1979; Miller & Kintsch, 1980). Coherence of a text can be assessed by converting a text into its propositions and then calculating the number of arguments that overlap from one proposition to the next. The greater amount of propositional overlap corresponds to a greater amount of coherence.
One application of this predictive power has been to improve the comprehensibility of texts. Britton & Gulgoz (1991) used the Kintsch model to identify locations in a text where a reader would have to make these bridging inferences in order to maintain coherence of the text. The chosen text was on the topic of the air war in Vietnam and so undergraduates had very little background knowledge to make the correct bridging inferences at these incoherent locations. They created a revised version of the text in which they inserted into the text the inferences that would be needed to be made at the various locations. They found that undergraduates reading the revised text had significantly better recall and a better mental representation of the text than those who had read the original version.
In a linear text, a writer typically maintains a set of coherent arguments through the text. At the local level, a writer makes words and sentences flow together through common referents. At a global level, a writer similarly makes paragraphs and sections flow from one to the other in a coherent manner. This aids the reader in structuring the information in the text to fit into the knowledge structures of what has been read previously. If there is little global coherence between sections, then the user must make bridging inferences in order to maintain coherence (e.g., Kintsch & Vipond, 1979). For readers without appropriate background knowledge, these inferences can consume the resources of the reader, typically resulting in lower comprehension.
In a hypertext, at any text section there are usually a variety of other sections to which a reader can jump. However, it may not be possible for a writer to anticipate all the possible places to which a reader may jump and therefore, it may also not be possible to maintain good macrocoherence for all possible links. Although the writer of the hypertext may code all the links, it would still be difficult to write each section so that it would cohere well with every possible section to which a reader could jump. Jumps which are not coherent could result in additional processing load for the reader, as the reader generates the necessary inferences to incorporate the textual information from the new node into what has been previously read. Thus, a many of the possible links in a hypertext may cause difficulty in the reader's comprehension.
One exception to this problem of hypertext coherence may be found in argument-based hypertext systems (e.g., Conklin & Begeman, 1989; Fischer, McCall, & Morch, 1989). These systems take into account the role of coherence through only allowing jumps between nodes in which a coherent argument has been previously set between the two. These coherent links are created through careful hand-coding of all possible links. While argument-based hypertext involves a lot more hand crafting in order to create only these coherent links, it avoids the problems of readers jumping to nodes using links which may cause an incoherent transition.
As a text is read, there is a large cognitive load on the reader as the reader is decoding the text and incorporating the textual information into his or her knowledge base. The key to incorporating the information into the reader's knowledge base is partly dependent on the amount of background knowledge of the reader. A reader's background knowledge permits the information to be incorporated into pre-existing knowledge structures. So readers who do not have an adequate amount of background knowledge on the subject of a text will have lower comprehension of the text (Spilich, Vesonder, Chiesi & Voss, 1979; Voss, Vesonder & Spilich, 1980). The background knowledge permits the reader to provide a coherence to the text, permitting better bridging inferences between non-coherent sections and also permitting additional elaborative inferences. In addition, there is evidence that the background knowledge influences the processing of information for developing at the situational model level, but not at the propositional level (Fincher-Kiefer, Post, Greene & Voss, 1988). Thus, additional background knowledge of the text makes it easier for developing relevant macropropositions, resulting in a better representation of the text for the reader.
The amount of background knowledge can also differentially affect the readers of a hypertext. As in linear text, readers with background knowledge on the domain of the text will be better at encoding information from the text than those without the background knowledge. The high knowledge readers will have the correct conceptual structures in which to integrate the new information and therefore the reader will tend to have better recall of the text.
However, due to the flexibility of hypertexts, readers with little background knowledge may have additional difficulties when compared to readers of equivalent linear texts. Readers with background knowledge will already have the correct conceptual structures for the domain. The hypertext structure may therefore be very familiar for them. However, a reader with little knowledge of the domain of the text will not be familiar with the structure of the hypertext. Since one of the concepts of hypertext is to permit more flexibility for the reader in choosing where to go, a low knowledge reader may not be able to accurately choose the relevant text sections. Thus, low knowledge readers may have additional problems of navigating through the hypertext structure. This problem may not be as evident in linear texts, since the linear text provides a single (default) path to read through the text. Low knowledge readers can always take this path, even if they are having trouble comprehending the text.
Similar to the background knowledge, the narrative schema of a text can aid in the comprehension of that text. A knowledgeable reader can use the narrative schema to provide a structure used in which to organize the text (e.g., Black & Bower, 1979; van Dijk & Kintsch, 1983; Dillon, 1991). Placing a text in a well known narrative schema can improve the comprehension of that text (e.g., Poulsen, Kintsch, Kintsch, & Premack, 1979).
In linear texts, there are a variety of common narrative schemata employed. Most readers are familiar with them and can use their knowledge of a schema in order to help structure and integrate the textual information. However, a narrative schema is primarily dependent on a familiar organization of the presentation of the information. With the novel structures found in hypertexts, much of any familiar narrative schema will not be evident. This can cause difficulties for the readers of a hypertext because they can not effectively organize the textual information into their situation model. While readers of hypertexts may not currently be able to rely on a familiar narrative schema, this may change in the future. As hypertexts become more accepted and widespread, writers of hypertext may develop standard rhetorical styles. Readers who are then familiar with those rhetorical styles can use that knowledge to help in their structuring of the information in an effective manner.
With the large amount of both high-level and low-level processing that must take place in order to comprehend a text, one can expect differences in comprehension based on the reader's abilities. Skilled readers tend to have better skills at exploiting context cues and other textual constraints. They are able to make better hypotheses about the meaning of words (Perfetti & Roth, 1981), and are more responsive to the rhetorical structure of the text (Meyer, Brandt & Bluth, 1980). On the other hand, poor readers' decoding skills are not as effective and instead they compensate through using context-dependent hypothesis testing. Thus, skilled readers are able to use parallel automatic processing to form better hypotheses about the meaning of the text as they read through it and are not as dependent on the contextual cues of the text. In this manner, if contextual cues are missing or are confusing, then the performance of poor readers will be degraded to an even greater extent.
Skilled readers of linear text are more responsive to the rhetorical structure of the text and possess better decoding skills. In contrast, poor readers must rely more on the context to help in decoding. Hypertext may provide fewer context cues than equivalent linear texts. This could be the case in a hypertext in which a reader is not given a lot of information about the structure of the text or about where to go in the text. In a linear text, a poor reader could rely on information from the linear structure for this context, but the same context may not be as evident in the hypertext. On the other hand, a well-structured hypertext that provides a map of its structure may provide additional context cues that would not be present in a linear text. In such a case, a poor reader may find the hypertext to be more of an advantage than the equivalent linear text.
Hypertexts also cause additional processing load by making the reader responsible for navigating the text. Skilled readers, who can process the text automatically will not have as much interference from the controlled processing task of having to make choices of where to go as they read. Readers with poor reading skills are using a lot more controlled processing for their reading processing and thus will likely have a greater amount of interference from the additional task of navigating the text. This will not let them generate as many hypotheses about the text as they read it, making it harder to integrate the information.
Thus, the abilities of the reader may interact with the type of text format and how that text format is implemented. A well structured hypertext which does not impose a large navigational load on the reader may be an advantage for poor readers. On the other hand, poor readers may have great difficulty with a less well structured hypertext when compared to the equivalent linear text.
While much of the results described above are based on tasks in which a reader reads through a text in a single linear order, readers also use various strategies for reading through a linear text. Not only do the reader's abilities and the static characteristics of the text determine the comprehension, but also the structural, syntactic and semantic signals in the text can affect comprehension. These signals provide evidence for the macro-relevance of the individual sections of text . A variety of studies have examined the role of different types of text signals on comprehension, including studies of titles (Bransford & Johnson, 1972), frequency of mention (Perfetti & Goldman, 1974), enumeration (Lorch & Chen, 1986), and initial sentences (Kieras, 1981).
While these signals can affect comprehension, they also determine what a reader looks for in a text and what the reader ignores. A reader may look through a text and find relevant sections based on these signals and skip over sections that do not seem to be as relevant. In this way, a linear text may not always be read linearly. There has been some research on the strategies used in reading a text. However, reading strategies have not been investigated as much as some of the comprehension processes in reading, since there is a lot less control of the conditions and it is harder to determine what information a subject has read and the subjects' motivations for choosing the particular information from the text.
Goldman and Saul (1990) identified a number of strategies used by subjects reading text passages. These strategies were identified both at the global level of the text and at the more local levels. Subjects read individual sentences and could go backward and forward through the sentences. At the global level, they identified three approaches: Once through, in which subjects read straight through a text, Review, in which subjects went to the passage end and then reviewed sentences, and the Regress approach, in which they went back to previous sentences throughout the text. They found that subjects almost always used more than one approach to reading a passage. Thus, readers employed a great deal of flexibility in reading the passages. However, their approaches for getting through the text were done in such a manner as to establish both local and global discourse coherence. From these results, they proposed a model of strategy competition for reading strategies. In their model, the choices of where to go in the text follow a set of procedural rules. These procedural rules work at both the global and local level to maintain coherence of the text, to react to the textual features that serve as cues to the coherence relations and to make strategy choices when coherence can not be established.
Thus, reading is not always a static process that proceeds in a single order. Readers can employ a variety of strategies for proceeding through a text. These strategies may be determined by a variety of factors, including, the reader's knowledge for the domain, the reader's goals, the reader's skills, and the characteristics of the text. However, the reader must still work to maintain the overriding goal of proceeding through the text in a coherent manner.
The examination of strategies in hypertext have primarily concentrated on the search strategies of readers when hypertexts have been used as information retrieval systems. Since hypertexts provide additional navigational flexibility to readers, one could expect that readers would employ a variety of strategies. As evident from the Goldman and Saul (1990) study, readers use a variety of strategies in order to maintain coherence through a linear text. It is expected that readers of hypertexts would do the same, since they must still maintain coherence in order to develop a coherent mental representation of the text.
Part of the goal of developing hypertexts is to have mechanisms to simplify the navigational strategies of readers. Hypertexts were developed so that readers could get to the relevant sections in a more efficient manner than linear texts. For this reason, it is expected that the strategies of the reader can depend greatly on what mechanisms are implemented in the hypertext. A hypertext based primarily on search will tend to structure a reader's strategies to using search. However, these mechanism must also match the goals of the reader and these goals can influence the choice of strategies. Hypertext typically permits a reader to see more of the structure of the text and therefore can make it somewhat easier to find the relevant places in the text. Thus, the reading strategies in hypertext could be expanded from those found in reading a linear text and could depend on the mechanisms provided in the hypertext and the goals of the reader.
Overall, there are a variety of factors that can cause differences in readers' comprehension and strategies of a text written in hypertext and linear form. Foltz (1993) ran two experiments that focused on two of these factors, the coherence of the text and how the goals of the reader affected reading strategies. The first experiment examined the comprehension and strategies of readers using either a linear text or one of two hypertext environments. The second experiment used verbal reports to investigate readers' strategies in hypertext.
In the first experiment, readers' comprehension and strategies were measured when using either a linear text or one of two hypertext versions of a chapter from an undergraduate level economics textbook. The goals of the readers were manipulated so that half of the readers read the chapter for general knowledge, while the other half read the chapter in order to find certain specific pieces of information.
This design permitted several comparisons. The first comparison was whether there were differences in readers' comprehension and strategies between the hypertexts and the linear texts. It was hypothesized that since readers would have little background knowledge on the subject of the text, the lack of coherence in the hypertext may adversely affect the readers' comprehension when compared to the comprehension of the readers of the linear text. The second comparison was whether the goals of the reader interacted with the format of the text. More specifically, did readers use different strategies based on their goals, and did one format of the text afford better comprehension than another? It was hypothesized that hypertext may be more suited for tasks involving information search, so readers with specific goals, in which they must find the relevant information, may perform better on the hypertext, while readers with general reading goals would find an advantage for the linear text, since it presented a single coherent set of information on the text.
An undergraduate level economics textbook chapter was converted into a hypertext using guidelines from several works on hypertext (e.g., Martin, 1990; Nielsen, 1990; Shneiderman, 1987). This included using the authors' section titles to designate nodes and the authors' outline for generating a hierarchy. The resulting hierarchical hypertext contained 6018 words consisting of 26 nodes. In addition to the hierarchical structure, there were also 17 cross- links where the authors made references to other sections. The hypertext permitted readers to navigate through the text using buttons for following both the hierarchical structure and the cross links. In addition, there was a map of the hypertext structure, which permitted jumping to any point in the text.
As described above, one of the hypothesized problems with hypertext may be that the jumps from node to node may not present as coherent a chapter as that of the linear text, resulting in lower comprehension. In order to investigate this idea, a second, more coherent hypertext was also developed. This hypertext was exactly the same as the first one, but also provided additional coherence in the text by automatically modifying the text in order to make it more coherent. In order to determine when a non-coherent jump was made in the hypertext, a macro-propropositional analysis of the text was first performed. In this analysis, the text was reduced to a set of macro-propositions, describing the gist of each sentence. The analysis indicated where possible non-coherent jumps would occur while reading the hypertext and what missing pieces of information would be needed for any jump in order to make the transition more coherent. Then, whenever non-coherent jumps were made, the computer added in an additional first paragraph of the necessary missing macropropositions, in sentence form, in order to make the transition more coherent. An example of one of the hypertext nodes with the inserted coherence paragraph, labeled "General Background" is shown in Figure 1.
Undergraduate subjects with no background in economics read one of the three text formats. Half of the readers were told to find specific information in the text that was distributed across several sections, while the other half of the readers read the text for general knowledge. After reading the text, the subjects answered comprehension questions and wrote an essay on the chapter. Some of the questions required subjects to recall information from the text, while other questions required subjects to use what they had learned from the text and apply it to new examples. This permitted an evaluation of both what they remembered from the text and what they learned from the text.
The surprising results of the study were that very few performance differences existed between the readers of the hypertexts and the linear text. There were no differences between text formats for the amount of time it took subjects to read the text. The subjects' comprehension also was equivalent for the different text formats. There were no differences in the number of macropropositions subjects generated in their essay and no differences in the subjects' scores on the short answer essay and multiple choice test. However, subjects who used the hypertexts did score better on questions that assessed their knowledge of the structure of the text. There were also no significant performance differences between the two types of hypertexts.
Manipulating the goals of the reader did have an effect. Subjects searching for the specific information in the text were faster but showed lower comprehension on the general knowledge questions. However, these measures did not interact with type of text, indicating that different types of text format did not differentially affect the comprehension of readers with different reading goals. Thus, the hypertexts did not help the readers search for the information any more than the linear text.
These results are similar to other studies in finding few comprehension differences between hypertext and linear text. However, in order to understand the lack of differences, it is important to look at readers' strategies. The computer kept a log of each subjects' progress and section choices as they read through the text. This permitted an analysis of the strategies that the subjects used for getting through the text and a characterization of the order that they read the text. Any node or page on which a subjects spent more than five seconds was counted as being read by the subject. Thus, in the linear text, subjects could still make non-coherent transitions by quickly paging forward or backward. The order that the subjects read the text was then compared to a macro-propositional analysis of the Text in order to count the number of coherent transitions made by the subjects. A coherent transition was one that did not violate the macrostructure of the text (i.e., the context of the node they were reading was similar to the context of the node to which they jumped).
For all groups in the study, 80 to 90 percent of the transitions were made in a manner that was coherent with respect to the macrostructure of the text. There were no differences due to text format or readers goals. Although this is not a surprising result for readers of the linear text, it suggests that readers of the hypertext primarily made transitions within the same context and seldom used any features to move them into very different parts of the hypertext. This coherence measure also correlated with the number of macropropositions generated in the subjects' essays indicating that the more coherently the text was read, the better the essay. Thus, the readers' comprehension of the hypertext was better if the reader chose to go through it in a coherent manner. The fact that there were no group differences in the coherent transition measure suggests that the lack of comprehension differences between the three text formats could be due to ceiling effects. All subjects were seeing approximately the same text in about an equivalently coherent manner, and thus their comprehension would not differ greatly. These coherence results therefore indicate that subjects tended to use a very rational approach of getting through the hypertexts, trying to maintain coherence.
The results from Experiment 1 showed that, although there were few differences in comprehension due to the three text formats, the reading strategies used by the subjects were all very similar. This indicates that subjects used heuristics to employ information from the text to guide them through the text in a coherent manner. While Experiment 1 showed evidence of these strategies, the experiment did not provide data that showed the reasons why the subjects used these strategies and what type of information from the text helped them to use these strategies. In the second experiment, verbal protocols (e.g. Ericsson & Simon, 1984) were collected about readers' strategies from six subjects as they read through the hypertext. In this experiment, two of the subjects read through the hypertext with instructions to find specific information, while the other four subjects read for general knowledge. Subjects were instructed to read the texts silently, but to describe out loud everything they were thinking as they read through the texts. If at any point the subjects were silent and were not reading, and also when they made any transition in the hypertext, the experimenter prompted them to describe what they were thinking. The results therefore provide a qualitative measure of subjects reasoning about their navigation strategies.
The verbal protocols indicated that subjects seldom wanted to stray from the hierarchical structure and they expressed interest in reading all the text in one area of the hypertext before moving on to other areas.
I think I'm done in government. I've covered all, covered that. Uhuh, I'm going to go to... how markets solve the basic economic problems. (Subject hg3001)
This strategy has somewhat of the same effect as the depth-first strategy. By reading one part of the hierarchy before moving on to other parts, the subject avoids moving out of the current context at any point. Indeed, if this strategy is applied recursively, the two strategies are equivalent.
Subjects with the general reading goals used reading strategies for covering the whole text in a coherent manner. However, when they made non-coherent transitions, they often expressed confusion and then tried to get back to where they had been previously. On the other hand, the subjects with specific reading goals tried to maintain the context around the specific information they needed to find.
Um, I was mostly looking at the general topics and trying not to skip, go all the way down to the, that in case there might be other topics to the left that were on the same kind of hierarchy, or whatever. I tried to look at, get an overview of those other three before I went over to there, in case they had something to do with that [....] I just think I needed an overview of the little things around it in order to get at the answers to that question. (subject hs3004)
Thus, it was not sufficient to find the relevant information in the text. In order for that information to make sense, the subject needed background context that could be provided by reading the text from the nodes that were around the relevant node. In addition, based on their time-stamped record, when subjects were instructed to read with goal of finding particular pieces of information, rather than using jumps in the hypertext to go down to the specific information, they tended to start at the upper level nodes above the desired information and then worked their way down to the specific information they had to read. They expressed that this would give them the appropriate background context.
The verbal protocols show that subjects used strategies for maintaining the coherence of the text. It is also important to know what information they used to help guide them to use these reading strategies. From both the verbal protocols and the time-stamped record of actions, subjects seemed to rely primarily on the map to guide their strategies. Subjects consulted the map an average of 8 times throughout the session. This indicates that they used the map often to help orient themselves. The subjects' protocols also showed that they relied on the map for guidance.
I'm thinking umm, I'm trying to decide whether to go back to the map of the chapter and work my over actually to keep working my way down the map. (Subject hs3004 )
In addition to the verbalizations about the map, subjects also appeared to use the titles of nodes in order to guide them. One subject identified the three nodes that were under the node titled, "The three functions of government", and decided he had better cover all three functions. Thus, while the map of the hierarchical structure appeared to be the primary piece of information used to guide the subjects, the titles could identify relationships between items in the hierarchy and allow the subject to navigate based on following the labels.
These results corroborate the results found in Experiment 1. The results show that reading a hypertext is not just a reading process, but also a process of problem solving. In order to understand the text, subjects developed heuristics for maintaining coherence with an unfamiliar text domain and text format The verbal protocols showed that the subjects were aware of their strategies to read the information around one area of the hierarchy before moving on to other areas. In this manner, their strategies allowed them to always stay within the same textual context.
Despite the differences in the text formats, in the above experiments, subjects used strategies to navigate the texts in a coherent manner. The similar approaches to reading the text resulted in similar amounts of text comprehension for the different text formats. For navigating the linear text, subjects merely relied on the linear order of the text for following the coherent path. For the hypertexts, subjects relied on additional constraints in the text such as the map and node titles which signaled the structure of the text. These signals served as guides for locating additional nodes that would flow coherently with the previously read nodes.
Manipulating the goals of the reader changed the readers' strategies. Readers with the specific goals saw fewer nodes and pages of the text and read for a shorter amount of time. However, although they were supposed to find information that was in separate places in the hierarchy, there were no differences in the number of coherent transitions made based on the readers' goals. This indicates that even when readers need to find specific information in an unfamiliar text, they still may need to have the appropriate background context. The verbal protocols from Experiment 2 confirmed that although subjects were aware of where the relevant nodes were, they read down to those nodes in order to put the relevant nodes in the appropriate context. In addition, the goal manipulation did not interact with the text format. Thus, even though subjects had different reading goals, they employed similar reading strategies on the different text formats.
The modified version of the hypertext that provided additional coherence showed no real differences from the original hypertext. Subjects may have ignored the added information that was provided to them by the coherent hypertext. However, based on the strategies used by the subjects in Experiment 1 and on the protocols in Experiment 2, the general background context provided by the coherent hypertext appears to be the type of information that the subjects needed to maintain coherence and to provide a context for what they were currently reading. Thus, subjects may have not noticed this information sufficiently or ignored when it appeared. It is also possible that not enough information was provided in these "background context" paragraphs in order to be efficiently used. However, because subjects tended to navigate the coherent hypertext in a coherent manner, this information was also seldom presented to them. Thus, there may not have been enough opportunities for them to use the provided background context to produce significant differences between the hypertexts. Although this research failed to show that the automatic coherence aided the readers, this type of approach may be of use to hypertexts. Given the right type of information in these background context paragraphs and proper training of the readers, readers would not have had to rely as much on following coherent paths. Instead they could used the provided background context to give them the necessary coherence when making jumps out of a particular context.
The strategies used by the subjects show evidence for the dominance of global comprehension in reading. The subjects used a rational approach to reading, maintaining an order of reading that was consistent with the macrostructure of the text. Even when the structure of the text had been modified from its linear form, they chose paths through the text that would flow coherently. This is consistent with the readers' goals to form a coherent macrostructure. Thus, readers of a hypertext are opportunistic. They look for the cues that will lead them to the most coherent path through the text. This behavior is also consistent with rules for coherent reading strategies set out in Goldman and Saul (1990).
With readers using rules to maintain coherence, this means that readers must develop a set of rules and then choose what rules to use for different situations. In this way, reading becomes an application of problem solving. The problem of navigating a hypertext coherently is similar to the problem of search in a problem space in an unfamiliar domain (e.g., Newell, 1980; Newell & Simon, 1972). The reader has a set of possible actions for getting through the text (problem space) and a set of search methods for finding the most effective paths through the text. Since the subjects in this experiment were all expert readers, they had fairly powerful search methods for discovering coherent relationships. However, the subjects were all in an unfamiliar text domain (economics), which put some limits on the ability of the search methods from guaranteeing that they would find the most coherent path.
This approach to navigating through the hypertext is also consistent with investigations of exploratory learning in computer systems, such as the CE+ model of Lewis and Polson (1990). Their model of behavior in unfamiliar computer domains uses a label-following, hill climbing search method in order to choose the relevant actions. In label-following (e.g., Engelbeck, 1986), the user's goals are matched with the possible actions, and an action is chosen through the degree of overlap of terms between the two. The label following heuristic is similar to the methods the hypertext readers used for determining where to make the next coherent transition. For finding specific information, subjects could match their reading goals to the possible nodes they could read to identify relevant nodes of the text and then work their way down to the appropriate nodes. For subjects with general reading goals, they could use the titles as guides for determining the relationships between nodes in order to determine what transitions would be the most coherent.
The Lewis and Polson research suggests that one of the keys to aiding in the problem solving is to design for successful guessing (see also, Norman 1986, 1988). This is a similar necessity for hypertexts. A reader of a hypertext must be able to guess whether following a particular link will both lead toward the relevant information and also be coherent with the current context. In order for a reader to make these successful guesses, this information needs to be evident through the structure of the text and the titles of the nodes.
While one of the original goals of this research was to determine the effect of the text's coherence on readers of hypertexts and linear texts, the results suggest that the similar coherence strategies overshadowed the effects of the text formats. In linear text, coherence is a good predictor of comprehension (e.g., van Dijk & Kintsch, 1983; Kintsch & Vipond, 1979). In this research, comprehension was equivalent because the resulting coherence of the text formats were equivalent. Even when modifying the goals of the reader, the text was still read coherently, resulting in few differences in comprehension. Thus, due to a lack of variability in coherence, this research can not draw conclusions about how much the coherence can break down in a hypertext and thereby affect the comprehension.
Readers' strategies indicate the strong role coherence plays in reading a hypertext. In the development of hypertext, the coherence between linked nodes is not often considered. Hypertext guidelines tell creators to link related items, however, there are few guidelines that tell them to make the text of linked nodes cohere. It is typically assumed that because two nodes are linked by some common piece of information, the reader can then generate the correct inferences about the link and incorporate the new information into his or her representation of the text. The results of this study show that subjects avoided the cases of "loosely" linked information (e.g., cross-hierarchical links), and instead primarily made transitions to highly related nodes.
It should be noted that this research examined the strategies for readers who were not familiar with the domain of the text. The results could change greatly with readers who are knowledgeable of the domain. With more background knowledge, readers could perform better problem solving to choose better coherent paths through the text. These readers would have the correct knowledge to make informed guesses on where to go in the text. With the proper knowledge structures, they could also make non-coherent jumps without adverse affects because they would possess the appropriate background knowledge to make the necessary bridging inferences. Thus, although readers with little background knowledge of the text domain, may have to rely on coherent paths in order to read through the text, a knowledgeable reader may be better able to exploit some of the less coherent links in the hypertext.
The results of these studies indicate that text comprehension theory can provide predictions for the design of hypertexts. Based on the theory, the primary design issues are to consider the type of text, how it will be used, and who will be using it. These issues can affect both what the reader can learn from the text and the type of strategies the reader will employ. In the case of these studies, the text was designed primarily for students to study concepts in economics. Students reading the text would have little background knowledge on the topic and would have the goals of either reading through the text or for searching for information in the text. Thus, these results can not be generalized to all hypertexts in all situations.
One of the primary considerations in the design of a hypertext should be to focus on the readers' strategies. These strategies can vary, depending both on the goals of the reader and the background knowledge of the reader. A reader who wants to find specific information in a hypertext needs to have simple methods to locate that information. However having the specific information without the appropriate background context can be useless. Thus, for hypertexts it is not just a matter of getting a reader to the correct place in the text, but also ensuring that the relevant information is represented in a meaningful context. The results from Experiment 2 indicated that subjects finding specific information needed the appropriate background context before they could go ahead and read the specific information.
For readers with extensive background knowledge, text comprehension theory predicts that this meaningful context may not be needed. However, for readers without the proper background knowledge, the additional information will be more valuable. For this reason, it would make sense to insert the appropriate context, or permit the novice reader to read through all the background context in order to get the relevant information. This concept would be somewhat similar to the idea of "guided tours" in hypertexts in which an author provides a single path through the text with the appropriate context (e.g., Marshall & Irish, 1989). It is also similar to the "training wheels" methodology for interfaces which kept novices on the correct path, only allowing them the use of a subset of the available functions (e.g., Carroll & Carrithers, 1984).
In addition to providing an appropriate context to readers, it is important to support familiar reading strategies. Readers will navigate through the text trying to find coherent paths. Thus, readers should be able to identify these paths. This can be done is several ways. Providing a clear structure of the text will aid the reader in finding these paths. Although a large text may have an enormous structure that can not be visualized as a whole, providing local information around the current node will provide both additional context for the reader and additional information as to what nodes will produce a coherent representation.
Along with having a clear structure, providing good labels for the nodes will also aid readers. Since novice readers tend to use a label-following strategy, labels that clearly indicate the role of a particular node, will help the reader in successful guessing of the appropriate coherent path (e.g., Lewis & Polson, 1990). Support for successful guessing in a hypertext should be similar to that of using a standard rhetorical structure in linear text. Readers who are familiar with the structure will be able to exploit it to their advantage. Those who are unfamiliar with it will have to expend cognitive effort in problem solving at the expense of the primary task of comprehending the information from the text.
One of the common tasks in developing hypertext is converting a linear text into a hypertext. While changing a text to a hypertext may provide a lot more flexibility to the readers, these conversions typically concentrate on restructuring the information rather than making large scale changes to the text. A converted linear text may still contain many of the ordering signals that were in the original text. These signals could include such text factors as enumeration and the mention of information that occurred previously in the linear text. These factors may cause the text to be read in the same linear fashion as the original text. The text may also have to be modified in order to improve its coherence. Restructuring a text will cause more possible ways to get to any particular node. The text in that node will have to be written in such a way that a jump into that node will be coherent. Thus, the conversion of a linear text to a well-written hypertext can involve an extensive amount of modification of the original text in order for the text to be more appropriate to being in a hypertext format.
One way around the problem of structuring when converting a text to hypertext is to keep the text in the linear form. SuperBook (Egan, et al., 1989, Landauer, et al., 1992) uses such an approach. The text is kept in its original form, but also enhanced through a variety of features to facilitate navigation. These features include a dynamic table of contents, a fisheye view of the current context, and search which posts the hits against the table of contents. This permits readers to get to the relevant places in the text efficiently but also permits them to read the text in the linear order.
One of the current difficulties in developing a hypertext is that hypertext writing environments are still in their infancy. Improved environments will help a writer see the constraints of the text as the text is developed. Such features will permit the writer to be better able to consider the coherence of the text from node and how the text should best be structured.
Thus, developing a hypertext is not just a matter of developing a text with all relevant interconnections. Instead, an effective hypertext must be based on a coherent structure. This structure must present an interface to the reader that will make the reader's strategies simple to be carried out.
The studies described above show the importance that readers place on maintaining the coherence of the text as they navigate a hypertext, however there remain many research issues that still must be explored in hypertexts. These results, and many other studies indicate that hypertexts, as they are currently implemented, are not suitable for all reading applications.
While this research sought to identify the effect of differences in the amount of coherence in the hypertext and the linear text, the strategies of the readers created equivalent amounts of coherent transitions. Not all hypertexts would permit subjects to perform as many coherent transitions as the hypertext used in this research. Thus, it would be instrumental to evaluate hypertexts which may cause a greater amount of non-coherent transitions in order to determine how much the coherence affects the comprehension. An alternative is to examine hypertexts that have large amounts of coherence. Hypertexts based on rhetorical models, such as systems for argumentation, enforce the coherence to a large extent. An analysis of reading comprehension on such systems may find that they produce larger amounts of comprehension and also provide a more appropriate representation of arguments than the equivalent linear version.
Models of text comprehension (e.g., van Dijk & Kintsch, 1983) can be applied to hypertext as well as to linear text. In both cases the key is to be able to measure the appropriateness of the text's structures and the resulting human's representation of those structures. The primary difference between the current text comprehension models for linear text and those for the hypertext is that a model for hypertext must also take into account the reader's behavior as he or she navigates through the text. In this way, the model must take into account more than just the text, but also the reader's strategies for dealing with the computer interface (e.g., Polson & Lewis, 1990). A model of hypertext comprehension must consider both the information the reader gains from the text and how that information can affect the readers' choice of strategies for proceeding through the text. Thus, this involves both an understanding of the constraints of the text and also of the constraints of the user interface. Given a successful model of these factors, we can generate better guidelines on how to develop more readable hypertexts which support more successful reading strategies.
From many of the studies in the field of hypertext, hypertexts have often performed worse, or just as well as the equivalent linear text. This raises the question: What are the areas where a hypertext will be more successful than a linear text, and what sorts of features will the hypertext need for this success?
One of the remaining problems of paper texts is still the fact that their indexes are inadequate and unwieldy to use. A current success for hypertext is as a search engine. Although this blurs the distinction between information retrieval systems and hypertexts, the ability to search in through a text simplifies many tasks. Hypertexts, such as SuperBook, permit readers to find the relevant information in an efficient manner. Although current computer indexing methods still leave much room for improvement (e.g., Dumais, 1988), the ability to move from one piece of relevant information to another can not be matched in equivalent paper versions of large amounts of text.
A second area where there is a great deal of potential for hypertext systems is for representing textual information that is not as easily presented in linear form. Such domains as legal argumentation and design knowledge have such problems. Any piece of textual information can have many related arguments and counter-arguments. These can be represented graphically with links between the arguments. Because the arguments are designed to follow each other, the links are constrained to be coherent. Thus the reader of the hypertext may follow the links and still get a coherent representation of the information. While these systems may provide an advantage to the reader, they do however, put a large load on the writer. Unlike preparing a single coherent set of arguments, the writer must determine the relationship between all pieces of information and structure them in a much more complicated manner than the linear equivalent. In the future, better computerized methods based on models of the text may be able to do some of this structuring, simplifying the task of the writer.
A third potential area for success in hypertexts is suggested by the investigation in this research of providing dynamic coherence to the hypertext. The goal of dynamic coherence is to provide additional information to the reader so that the text automatically contains information that is more appropriate to the reader's representation of the text. Readers of a text will bring large differences in reading abilities and background knowledge to a text. Through modeling the reader's knowledge, abilities, and goals, hypertexts could be developed that perform dynamic structuring of the text. These hypertexts could include additional background knowledge to readers with little knowledge and make the text more coherent for readers with low reading abilities. For readers with general reading goals or trying to get an overview, the text could be presented as a single path through the text. For readers trying to find specific information, the text could provide search capabilities and appropriate background context of any relevant items.
While the concept of personalized text sounds like a wonderful idea, accurately gauging the readers background knowledge, abilities and goals can be difficult. A lot of this will depend on being able to develop appropriate user models. These models would need to couple information about what is contained in the text, what information a reader already knows, and what information the reader needs to know.
Success for hypertexts lies in exploiting the powers of both the computer and the writer to generate better personalized texts. However, in order to do this, we must first have accurate models of such factors as: what a reader knows, what a reader needs, what methods a reader can use, and what information is contained in the text. Thus, the future of hypertext will depend on improving both models of the user and models of the text.
The author thanks Walter Kintsch, Peter Polson, Anders Ericsson, Reid Hastie, and Gerhard Fischer for their advice and comments on this project. He also thanks Adrienne Lee, Thomas Landauer, Susan Dumais, and Dennis Egan for their help. Support was provided by grants from the Army Research Institute ARI project MDA 903-86-CO143, the National Institute of Health and Bell Communications Research.
Begoray, J. A. (1990). An introduction to hypermedia issues, systems and application areas. International Journal of Man-Machine Studies, 33, 121-147.
Black, J. B., & Bower, G. H. (1979). Episodes as chunks in narrative memory. Journal of Verbal Learning and Verbal Behavior, 18, 309-318.
Bolter, J. D. (1991). Writing Space: The computer, hypertext, and the history of writing. Hillsdale, NJ: Lawrence Erlbaum.
Botafogo, R. A., & Shneiderman, B. (1991). Identifying aggregates in hypertext structures. In Proceedings of Hypertext '91, (pp. 63-74). San Antonio, TX: ACM.
Bransford, J., & Johnson, M. K. (1972). Contextual prerequisites for understanding: Some investigations of comprehension and recall. Journal of Verbal Learning and Verbal Behavior, 11, 717-726.
Bransford, J. D., & Franks, J. J. (1971). The abstraction of linguistic ideas. Cognitive Psychology, 2, 331-350.
Britton, B. K., & Glynn, S. M. (1989). Computer Writing Environments: Theory, Research, and Design. Hillsdale, NJ: Lawrence Erlbaum.
Britton, B. K., & Gulgoz, S. (1991). Using Kintsch's computational model to improve instructional text: Effects of repairing inference calls on recall and cognitive structures. Journal of Educational Psychology, 83, 329-345.
Britton, B. K., Meyer, B. J., Hodge, M. H., & Glynn, S. M. (1980). Effects of the organization of text on memory: Tests of retrieval and response criterion hypotheses. Journal of Experimental Psychology: Human Learning and Memory, 6(5), 620-629.
Card, S., Moran, T. P., & Newell, A. (1983). The Psychology of Human-Computer Interaction. Hillsdale, NJ: Lawrence Erlbaum.
Carroll, J. M., & Carrithers, C. (1984). Training wheels in a user interface. Communications of the ACM, 27(8), 800-806.
Charney, D. (1987). Comprehending non-linear text: The role of discourse cues and reading strategies. In Proceedings of Hypertext '87, (pp. 109-120). Chapel Hill, NC: ACM.
Charney, D. (1994). The impact of hypertext on processes of reading and writing. In S. J. Hilligoss and C.L. Selfe (Eds.) Literacy and Computers.. New York: MLA.
Conklin, J., & Begeman, M. L. (1989). gIBIS: A tool for all reasons. Journal of the American Society for Information Science, 40(3), 200-213.
Croft, W. B., & Turtle, H. (1989). A retrieval model incorporating hypertext links. In Proceedings of Hypertext '89, (pp. 213-224). Pittsburgh: ACM.
Crouch, D. B., Crouch, C. J., & Andreas, G. (1989). The use of cluster hierarchies in hypertext information retrieval. In Proceedings of Hypertext '89, (pp. 225-238). Pittsburgh: ACM.
Dillon, A. (1991). Readers' models of text structures: The case of academic articles. International Journal of Man-Machine Studies, 35, 913-925.
Doane, S., Kintsch, W., & Polson, P. G. (1990). Action planning: Producing UNIX commands. In Proceedings of the 11th annual meeting of the Cognitive Science Society, 7 (pp. 30-57). Ann Arbor, Michigan: Erlbaum.
Dumais, S. T. (1988). Textual information retrieval. In M. Helander (Eds.), Handbook of Human-Computer Interaction. (pp. 673-700). New York: Elsevier.
Egan, D. E., Remde, J. R., Gomez, L. M., Landauer, T. K., Eberhardt, J., & Lochbaum, C. C. (1989). Formative design-evaluation of 'SuperBook'. Transactions on Information Systems, 7(1), 30-57.
Engelbeck, G. E. (1986) Exceptions to generalizations: Implications for formal models of human-computer interaction. Unpublished master's thesis, University of Colorado, Department of Psychology, Boulder, Co.
Ericsson, K. A., & Simon, H. A. (1984). Protocol analysis: Verbal reports as data. Cambridge, MA: MIT Press.
Fincher-Kiefer, R., Post, T. A., Greene, T. R., & Voss, J. F. (1988). On the role of prior knowledge and task demands in the processing of text. Journal of Memory and Language, 27, 416-428.
Fischer, G., McCall, R., & Morch, A. (1989). JANUS: Integrating hypertext with a knowledge-based design environment. In Proceedings of Hypertext '89, (pp. 105-118). Pittsburgh, PA: ACM.
Fodor, J. D., Fodor, J. A., & Garrett, M. F. (1975). The psychological unreality of semantic representations. Linguistic Inquiry, 6, 515-531.
Foltz, P. W. (1993). Readers' comprehension and strategies in linear text an hypertext. Unpublished Doctoral dissertation. University of Colorado, Boulder.
Furuta, R., & Stotts, P. D. (1989). Programming browsing semantics in Trellis. In Proceedings of Hypertext '89, (pp. 27-42). Pittsburgh, PA: ACM.
Goldman, S. R., & Saul, E. U. (1990). Flexibility in text processing: A strategy competition model. Learning and Individual Differences, 2(2), 181-219.
Gould, J. D., Alfaro, L., Fonn, R., Haupt, B., Minuto, A., & Salaun, J. (1987). Why reading was slower from CRT displays than from paper. In Proceedings of the ACM CHI + GI '87, (pp. 7-11). Toronto, Canada: ACM.
Gould, J. D., & Grischkowsky, N. (1984). Doing the same work with hard copy and with cathode ray tube (CRT) computer terminals. Human Factors, 26, 296-300.
Hammond, N., & Allinson, L. (1989). Extending hypertext for learning: An investigation of access and guidance tools. In People and Computers V, Nottingham, UK.
Hardman, L., & Sharrat, B. (1989). User-Centered hypertext design: The application of HCI design principles and guidelines. In Proceedings of Hypertext 2, York, UK
Kieras, D., & Polson, P. G. (1985). An approach to the formal analysis of user complexity. International Journal of Man-Machine Studies, 22, 365-394.
Kieras, D. E. (1981). The role of major referents and sentence topic in the construction of passage macrostructure. Discourse Processes, 4, 1-15.
Kintsch, W. (1974). The representation of meaning in memory. Hillsdale, NJ: Erlbaum.
Kintsch, W. (1988). The use of knowledge in discourse processing: A construction-integration model. Psychological Review, 95, 363-394.
Kintsch, W., & Keenan, J. M. (1973). Reading rate and retention as a function of the number of propositions in base structure of sentences. Cognitive Psychology, 5, 257-274.
Kintsch, W., & van Dijk, T. A. (1978). Toward a model of text comprehension and production. Psychological Review, 85, 363-394.
Kintsch, W., & Vipond, D. (1979). Reading comprehension and readability in educational practice and psychological theory. In L. G. Nilsson (Eds.), Perspectives on Memory Research. Hillsdale, NJ: Erlbaum.
Landauer, T. K., Egan, D., Remde, J., Lesk, M. J., Lochbaum, C. C., & Ketchum, D. (1993). Enhancing the usability of text through computer delivery and formative evaluation: The SuperBook project. In C. McKnight, A. Dillon & J. Richardson, (Eds.), Hypertext: A Psychological Perspective. New York: E. Horwood.
Landow, G. P. (1989). Hypertext in literary education, criticism, and scholarship. Computers and the Humanities, 23, 173-198.
Lewis, C. & Polson, P. G. (1990). Theory-based design for easily learned interfaces. HCI, 5, 191-220.
Lorch, R. F., & Chen, A. H. (1986). Effects of number signals on reading and recall. Journal of Educational Psychology, 78, 263-279.
Mannes, S., & Kintsch, W. (1991). Routine computing tasks: Planning as understanding. Cognitive Science, 15, 305-342.
Marchionini, G., & Shneiderman, B. (1987). Finding facts vs. browsing knowledge in hypertext systems. IEEE Computer, 21(1), 70-80.
Marshal, C. C., & Irish, P. M. (1989). Guided tours and on-line presentations: How authors make existing hypertext intelligible for readers. In Proceedings of Hypertext '89, (pp. 15-26). Pittsburgh, PA: ACM.
Martin, J. (1990). Hyperdocuments and How to Create Them. Englewood Cliffs, NJ: Prentice-Hall.
Meyer, B. J. F. (1973). The organization of prose and its effects on memory. Amsterdam: North Holland.
Meyer, B. J. F., Brandt, D. M., & Bluth, G. J. (1980). Use of top-level structure in text: Key for reading comprehension in ninth-grade students. Reading Research Quarterly, 16, 72-103.
Miller, J. R., & Kintsch, W. (1980). Readability and recall of short prose passages: A theoretical analysis. Journal of Experimental Psychology: Human Learning and Memory, 6(4), 335-354.
Monk, A. F., Walsh, P., & Dix, A. J. (1988). A comparison of hypertext, scrolling, and folding as mechanisms for program browsing. In D. M. J. &. R. Winder (Eds.), People and Computers IV. (pp. 421-435). Cambridge: Cambridge University Press.
Nelson, T. (1967). Getting it out of our system. In G. Schechter (Eds.), Information Retrieval: A critical review. Wash., D.C.: Thompson Books.
Newell, A. (1980). Reasoning, problem solving and decision processes: The problem space as a fundamental category. In R. Nickerson (Eds.), Attention and performance VIII. (pp. 693-718). Hillsdale, NJ: Lawrence Erlbaum.
Newell, A., & Simon, H. A. (1972). Human problem solving. Englewood Cliffs, NJ: Prentice-Hall.
Nielsen, J. (1989). The matters that really matter for hypertext usability. In Proceedings of Hypertext '89, (pp. 239-248). Pittsburgh, PA: ACM.
Nielsen, J. (1990). Hypertext and Hypermedia. San Diego, CA: Academic Press.
Norman, D. A. (1986). Cognitive engineering. In D. A. Norman & S. W. Draper (Eds.), User centered system design: New perspectives in human-computer interaction. (pp. 31-61). Hillsdale, NJ: Lawrence Erlbaum.
Norman, D. A. (1988). The psychology of everyday things. New York: Basic Books.
Perfetti, C. A., & Goldman, S. R. (1974). Thematization of sentence retrieval. Journal of Verbal Learning and Verbal Behavior, 13, 70-79.
Perfetti, C. A., & Roth, S. (1981). Some of the interactive processes in reading and their role in reading skill. In A. Lesgold & C. Perfetti (Eds.), Interactive processes in reading. Hillsdale, NJ: Lawrence Erlbaum.
Poulsen, D., Kintsch, E., Kintsch, W., & Premack, D. (1979). Children's comprehension and memory for stories. Journal of Experimental Child Psychology, 28, 379-403.
Shneiderman, B. (1987). User interface design and evaluation for an electronic encyclopedia. In G. Salvendy (Eds.), Cognitive Engineering in the Design of Human-Computer Interaction and Expert Systems. (pp. 207-223). New York: Elsevier.
Spilich, G. J., Vesonder, G. T., Chiesi, H. L., & Voss, J. F. (1979). Text processing of domain-related information for individuals with high and low domain knowledge. Journal of Verbal Learning and Verbal Behavior, 18, 275-290.
Spiro, R. J., & Jehng, J. C.. (1990). Cognitive Flexibility and hypertext: Theory and technology for the nonlinear and multidimensional traversal of complex subject matter. In D. Nix and R. J. Spiro (Eds.) Cognition, Education and Multimedia: Exploring Ideas in High Technology. Hillsdale, NJ: Lawrence Erlbaum
van Dijk, T. A., & Kintsch, W. (1983). Strategies of Discourse Comprehension. New York: Academic Press.
Voss, J. F., Vesonder, G. T., & Spilich, G. J. (1980). Text generation and recall by high-knowledge and low-knowledge subjects. Journal of Verbal Learning and Verbal Behavior, 19, 651-667.
Weyer, S. A. (1982). The design of a dynamic book for information search. International Journal of Man-Machine Studies, 17, 87-107.
 "For the short term, the best recommendation probably is to pay close attention to the authoring principles implicit in other writers' hypertext and try to emulate the principles you like." (Nielsen, 1990, p. 164)