This is a plain-text version of a dissertation. It should not be distributed or otherwise used without permission of the author. The author's current contact information is: // Gregory B. Newby, Assistant Professor in the School of Information // and Library Science, University of North Carolina at Chapel Hill // CB# 3360 Manning Hall, Chapel Hill, NC, 27599-3360 E: gbnewby@ils.unc.edu // V: 919-962-8064 F: 919-962-8071 W: http://www.ils.unc.edu/~gbnewby/ Towards Navigation for Information Retrieval by Gregory B. Newby B.A. State University of New York at Albany, 1987 M.A. State University of New York at Albany, 1988 Abstract of Dissertation Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Information Transfer in the Graduate School of Syracuse University May, 1993 Approved ____________________________ Professor Michael S. Nilan Date ________________________________ This work proposes navigation as a fundamental concept for information retrieval. A conceptual framework for navigation is developed, after Mead's (1960) notions of the importance of modelling the other for effective communi- cation. Navigation is defined as that behavior in which humans engage to make sense of an information space. Information space is defined as the set of concepts and relations among them stored by an information system. Unlike cognitive spaces, which humans possess, currently available information spaces are not generally subject to change in response to ongoing communica- tion. During human communication, people send and receive messages, which have the effect of changing their cognitive spaces. For effective communica- tion, each participant uses a model of the other to properly gauge the effects of her or his messages. For interaction with information systems, the necessity for a model is the same, but it is typically incumbent on the human user to conform to the generalized model which the system has of its users. Navigation is not a metaphor, it is human behavior which occurs when humans interact with information space. This work treats navigation through physical domains in the same way as navigation through information space, in that each requires the creation and maintenance of a cognitive model of what is navigated. Brookes' (1975) "exosomatic memory" is presented as a long-term goal of information systems. Providing more navigable systems is one step towards that goal, by facilitating human model building and moving towards human- computer interaction which is more similar to human-human interaction. An empirical investigation of navigation for information retrieval is completed. An information space is created which incorporates concept relations, intended as a step towards information spaces which match the cognitive spaces of users. A prototype information retrieval system is designed to navigate the information space, employing a visual interface and optional gesture-oriented input device. A user-based evaluation of the prototype environment is made. The outcome indicates that navigation, as conceptualized for this work, is a useful and fruitful outlook on information seeking behavior involving human-computer interaction. Towards Navigation for Information Retrieval by Gregory B. Newby B.A. State University of New York at Albany, 1987 M.A. State University of New York at Albany, 1988 Dissertation Submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Information Transfer in the Graduate School of Syracuse University May, 1993 Approved ____________________________ Professor Michael S. Nilan Date ________________________________ Coke is a trademark of the Coca-Cola Company. Dialog is a trademark of Dialog Systems, Inc. HyperCard is a trademark of Apple Computers. Iris, The Graphics Library (GL), The Geometry Engine, and Irix are copyrights of Silicon Graphics, Inc. PowerGlove is a trademark of Mattel, Inc. Unix is a trademark of AT&T. VM/XA is a copyright of IBM. Copyright 1993 Gregory B. Newby *** Insert Committee Approval Page Here *** T A B L E O F C O N T E N T S LIST OF TABLES AND FIGURES. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii ACKNOWLEDGEMENTS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii CHAPTER 1: INTRODUCTION AND CONCEPTUAL FRAMEWORK . . . . . . . . . . . . . . . . . . . . 1 1.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 A Conceptual Framework for Understanding Navigation. . . . . . . . . . . . . 6 1.1.1 Human Communication, Cognitive Movement, and Cognitive Space. . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.1.2 Information Space and Navigation. . . . . . . . . . . . . . . . . . 17 1.1.3 Models for Human Communication and Navigation. . . . . . . . . . . 25 1.1.4 Representation for Information Space. . . . . . . . . . . . . . . . 28 1.1.5 Facilitating Navigation . . . . . . . . . . . . . . . . . . . . . . 32 1.1.6 Communication and Navigation. . . . . . . . . . . . . . . . . . . . 33 1.2 Goals for Information Retrieval Systems. . . . . . . . . . . . . . . . . . . 37 1.2.1 Relevance-Based Information Retrieval . . . . . . . . . . . . . . . 38 1.2.2 Navigation-Based Information Retrieval. . . . . . . . . . . . . . . 41 1.2.3 Criteria for Evaluation of Information Retrieval Systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 1.3 Goals, Questions, and Definitions. . . . . . . . . . . . . . . . . . . . . . 50 1.4 Criteria for Evaluating this Work . . . . . . . . . . . . . . . . . . . . . 55 1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 CHAPTER 2: LITERATURE REVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 2.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59 2.1 Traditions in Information Retrieval. . . . . . . . . . . . . . . . . . . . . 60 2.1.1 Foundations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 2.1.2 Representation for IR . . . . . . . . . . . . . . . . . . . . . . . 62 2.1.3 Assigning Keyterms. . . . . . . . . . . . . . . . . . . . . . . . . 63 2.2 Non-Relevance-Based Approaches and the Retrieval Pro- cess . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 2.2.1 Browsing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 2.2.2 Cognitive Approaches. . . . . . . . . . . . . . . . . . . . . . . . 68 2.2.2.1 Dialog-Oriented Systems . . . . . . . . . . . . . . . . . 69 2.2.2.2 Anomalous States of Knowledge . . . . . . . . . . . . . . 70 2.2.2.3 Cognitive User Models . . . . . . . . . . . . . . . . . . 71 2.2.2.4 Sense Making. . . . . . . . . . . . . . . . . . . . . . . 72 2.3 Information Space. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 2.3.1 Spatial Representation for IR . . . . . . . . . . . . . . . . . . . 74 2.3.1.1 Spatial Representation with Orthogonal Vectors. . . . . . . . . . . . . . . . . . . . . . . . . . 75 2.3.1.2 Spatial Representation with Term Relations. . . . . . . . 76 2.3.2 Information Space in the IR Literature. . . . . . . . . . . . . . . 78 2.3.3 Information Space in the Psychological Literature . . . . . . . . . 80 2.4 Wayfinding and Visualization . . . . . . . . . . . . . . . . . . . . . . . . 82 2.5 Literature Central to the Current Work . . . . . . . . . . . . . . . . . . . 84 2.6 Outcome of the Literature Review . . . . . . . . . . . . . . . . . . . . . . 88 2.7 Information Space Revisited. . . . . . . . . . . . . . . . . . . . . . . . . 89 2.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 CHAPTER 3: METHODS OF INVESTIGATION. . . . . . . . . . . . . . . . . . . . . . . . . . . 91 3.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 3.1 Building an Information Space. . . . . . . . . . . . . . . . . . . . . . . . 93 3.1.1 Information Space via Multidimensional Scaling. . . . . . . . . . . 94 3.1.2 Building the Space. . . . . . . . . . . . . . . . . . . . . . . . . 98 3.2 Building the Interface . . . . . . . . . . . . . . . . . . . . . . . . . . 107 3.3 User-Based Evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . 112 3.3.1 Overview of Evaluation. . . . . . . . . . . . . . . . . . . . . . 113 3.3.2 Operationalization. . . . . . . . . . . . . . . . . . . . . . . . 114 3.4 Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 3.4.1 Respondents . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 3.4.2 Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 3.5 Error. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 CHAPTER 4: RESULTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 4.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 4.1 Analytic Techniques. . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 4.2 Task-related criteria. . . . . . . . . . . . . . . . . . . . . . . . . . . 132 4.2.1 Well-Defined Information Needs. . . . . . . . . . . . . . . . . . 134 4.2.2 Less Well-Defined Information Needs . . . . . . . . . . . . . . . 136 4.2.3 Learning and System Cues. . . . . . . . . . . . . . . . . . . . . 138 4.2.4 Training Time . . . . . . . . . . . . . . . . . . . . . . . . . . 141 4.2.6 Summary of Task-Related Criteria. . . . . . . . . . . . . . . . . 142 4.3 Model-related criteria . . . . . . . . . . . . . . . . . . . . . . . . . . 143 4.3.1 The Other . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 4.3.2 The Relationship to the Other . . . . . . . . . . . . . . . . . . 147 4.3.3 How to Change the Relationship to the Other . . . . . . . . . . . 151 4.3.4 Model of the Self . . . . . . . . . . . . . . . . . . . . . . . . 153 4.4 Analysis of Closed-Ended Data. . . . . . . . . . . . . . . . . . . . . . . 156 4.4.1 Searches which did not Result in Selection of a Document Surrogate . . . . . . . . . . . . . . . . . . . . . . . 160 4.4.2 Successful Searches: Prism . . . . . . . . . . . . . . . . . . . 162 4.4.4 Successful Searches: Space . . . . . . . . . . . . . . . . . . . 166 4.4.5 Respondent Overall System Evaluation. . . . . . . . . . . . . . . 171 4.4.6 Demographics. . . . . . . . . . . . . . . . . . . . . . . . . . . 176 4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178 CHAPTER 5: DISCUSSION AND CONCLUSION . . . . . . . . . . . . . . . . . . . . . . . . . 180 5.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 5.1 Navigation: Derived Understanding . . . . . . . . . . . . . . . . . . . . 182 5.2 Implications for Information Retrieval . . . . . . . . . . . . . . . . . . 184 5.3 Limitations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 5.3.1 The Conceptual Framework. . . . . . . . . . . . . . . . . . . . . 190 5.3.2 The Space System. . . . . . . . . . . . . . . . . . . . . . . . . 191 5.3.3 The Empirical Study . . . . . . . . . . . . . . . . . . . . . . . 195 5.4 Future Work on Navigation. . . . . . . . . . . . . . . . . . . . . . . . . 197 5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201 APPENDIX A: TRAINING PACKET. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202 APPENDIX B: QUESTIONNAIRE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221 APPENDIX C: INFORMATION NEED STATEMENTS. . . . . . . . . . . . . . . . . . . . . . . . 233 APPENDIX D: INSTRUCTIONS TO THE RESEARCH ASSISTANT . . . . . . . . . . . . . . . . . . 235 APPENDIX E: INFORMATION SPACE COORDINATES. . . . . . . . . . . . . . . . . . . . . . . 241 Section 1: Coordinates of 264 Eric "descriptors" and "major descriptors.". . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241 Section 2: The coordinates of 272 documents. . . . . . . . . . . . . . . . . . 246 APPENDIX F: SAMPLE ERIC DOCUMENTS. . . . . . . . . . . . . . . . . . . . . . . . . . . 252 APPENDIX G: CODEBOOK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255 APPENDIX H: OPEN-ENDED DATA BY ITEM. . . . . . . . . . . . . . . . . . . . . . . . . . 257 REFERENCES. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285 LIST OF TABLES AND FIGURES Figure 1.1: What do information retrieval systems match? . . . . . . . . . . . . . . . . 40 Table 1.1: Summary of Key Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Table 3.1: Overview of Procedures Taken to Generate a Spatial Repre- sentation of a Bibliographic Database. . . . . . . . . . . . . . . . . . . . . . 98 Table 3.2: The ERIC Database Subset. . . . . . . . . . . . . . . . . . . . . . . . . . . 99 Table 3.3: Statistics for Keyterms . . . . . . . . . . . . . . . . . . . . . . . . . . 102 Table 4.1: Success for Each System Across Tasks. . . . . . . . . . . . . . . . . . . . 135 Table 4.2: Time on Task. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 Table 4.3: Closed-ended scores for "satisfaction" item . . . . . . . . . . . . . . . . 157 Table 4.4: Relationship of Navigation and Satisfaction Scores. . . . . . . . . . . . . 158 Table 4.5: Demographic Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177 ACKNOWLEDGEMENTS Many people and organizations helped me during the completion of this work. My thanks goes (in alphabetical order) to: Abrams/Gentile Entertainment, Inc. and Chris Gentile for the PowerGlove converter box; The Advanced Graphics Research Laboratory at Syracuse University, with Dave Richers and Arnold Paul who helped set up early tests of my IR system; Roger Chen for chairing my defense; the database management group at Syracuse University for allowing me to develop a database subset of the ERIC database; Bruce Derr, Ronald Kalinowski, and the rest of the computer systems staff at Syracuse University for favors, advice, and indulgence; Judith Diamond for program- ming help; my fellow doctoral students at Syracuse University, who offered moral support and helpful advice; Wayne Fordyce, who gave me a startup account and advice for the Cornell supercomputer; the Graduate School of Library and Information Science at the University of Illinois at Urbana- Champaign, for their indulgence and faith in my work; David Micko for recruiting respondents and administering the user-based study; the National Center for Supercomputing Applications, which supported production of a videotape of my system; Joseph Woelfel for planting a seed that started all this, and serving as outside reader; Kent Yates for help with Iris system adminis- tration at the University of Illinois; Robert Zeh for programming assistance. Separate thanks go to my advisor and committee members: Michael S. Nilan (advisor), who offered constructive advice on how to present my work and made my career at Syracuse much easier; Jeffrey Katzer, who has the sharpest methodological mind on the block and offered keen criticism of my work; Sung Myaeng, who expressed strong interest in system design issues; Robert N. Oddy, who has knowledge of information retrieval that is astounding, and offered much practical advice on my space building and system design; and Michael B. Eisenberg, who brings exceptional writing talent and rigor to everything he does. Extra special thanks goes to my wonderful wife, Ilana, who gave me her support, encouragement, and inspiration. CHAPTER 1: INTRODUCTION AND CONCEPTUAL FRAMEWORK 1.0 Introduction The language of information retrieval is the language of navigation. Users look for information, they browse, they choose a new direction for a search, they find something that is close to what they are looking for. Navigation describes what information seekers do. This work introduces a conceptual framework with navigation as a fundamental concept about which information retrieval systems and other types of information systems might be built. Information retrieval systems are special types of information systems for two reasons. First, the size of the database is typically very large. Databases with hundreds of thousands of items are common, and may include significant amounts of unstructured natural-language text. Second, the users and their purposes are heterogeneous. Designers of information retrieval (IR) systems have little power to specify the situations in which users will approach a system, or to predict the values, goals, experiences, or language they will bring with them (Nilan and Rosenbaum, 1991). These reasons make for large information spaces for IR (as defined and discussed in this chapter) which are difficult to structure for particular information use situations. Other types of information systems might also benefit from a focus on navigation, as defined and discussed here. Train schedules, book indexes, thesauri, training manuals, expert systems, and computer operating systems are other types of systems for which navigation as a fundamental concept might be applied but they typically focus on a more homogeneous range of users and uses than the bibliographic systems found in IR research. Defined generally as any mechanism for storing and retrieving information, information systems might include anything from office filing systems to artificial intelligences. The current work is concerned exclusively with computerized information retrieval systems, but perhaps has implications for non-computer- ized systems. A major redirection in the design of some computerized versions of the information systems listed above has been taking place in recent years, as graphical interfaces and window systems have become commonplace. These newer systems do focus on navigation, often explicitly, although without any sort of consistent conceptual framework such as that developed in this chapter. The X-Windows system, hypertext, computer browsing systems, and various other visual interfaces are all examples of information systems based on some aspects of navigation as discussed here. Historically, Bush's (1945) "As we may think" laid a goal which was easily within the reach of imagination, but has not yet been realized. I wonder if he could comprehend the vision of Gibson's (1984) NEUROMANCER, in which similarly futuristic incorporation of humans, machines, and information was predicted? Bush's "memex" was essentially an add-on to human memory -- it was able to identify information needed in response to ambiguous or incomplete queries, presumably based on what it knew about various users and processes. Input and output were via text (or perhaps voice). In Gibson's world, machines such as the "memex" have still not been realized. Instead, humans navigate through a worldwide multidimensional information space, called "Cyberspace." Input and output is accomplished via direct neural stimulation, and the entire scenario is augmented by computers. This work may take small steps towards making the visions of Bush, Gibson, and others a reality, through potential enhancements to IR systems and by expanding the notions of representation and cognitive matching within such systems. Several difficulties lie on the way to the achievement of such visions, however. One important one has to do with the nature of the interaction between people and computers. When a literature on information retrieval began to come into its own in the mid-1960's, computers did not provide highly interactive environments: they operated mostly in batch mode. Batch processing is when a program is introduced to the computer and run at a later time. Results are then returned to the programmer. The programmer then has to determine the quality of the output, make any changes indicated, and resubmit the job as needed. (The computer, meanwhile, has no memory of the previous time the job was submitted.) The "memex" would not exist in such an environment. Bush's vision pointed to ongoing interaction with the "memex," more akin to human communication than batch-mode retrieval. Unfortunately, the batch mode of processing is still the dominant model for computer interfaces today. Whether the user is creating a program, using an operating system, sending electronic mail, or using an IR system, the general process is one where the user submits and re-submits commands until he or she is satisfied with the result. Interaction in these cases is in real- time, with immediate system feedback -- however, the general process is one where the user submits and re-submits commands, and is solely responsible for quality control. The addition of a visual interface based on a desktop metaphor, and pointing device such as a mouse, might remove some of the onus from the user to remember specific command sequences, but the "hit or miss" nature of the interface is not removed by the simple addition of a different sort of target (a menu) and new device for interaction (a mouse, instead of a keyboard). These systems do not have models of either users or their tasks. This work postulates that users must navigate such systems, as navigation will be formally defined in this chapter, but the systems were not designed to facilitate navigation. This work presents navigation as a fundamental concept for information retrieval. This is contrasted with relevance, which is the concept at the heart of most of today's IR systems. Navigation is not proposed as an end goal for information retrieval, but rather as a redirection for research and design which will move away from the assumptions and practice of batch-mode IR and towards systems which are more like human communication systems. Eventual- ly, it is hoped, this path will lead towards realization of IR scenarios such as Bush dreamed of, or as proposed by Brookes (1975), where information systems operate as "exosomatic memory:" an external database which is accessed by humans as an extension of their individual memories. Users may be more empowered by navigable systems to select the information they want than by systems that do not focus on navigation. This is because a navigable system (as envisioned for the current work) would have two important qualities. First is the organization of a database so that it closely matches user perceptions of the database contents, in that the concepts, and relations among them, which the database has are similar to the concepts and relations as perceived by the system's users. Second, navigable systems provide cues which help the user to understand his or her status relative to the system -- the relationship between user and system -- and how to change that status. This chapter lays the framework for the rest of the work. The main component of the chapter consists of the development of a conceptual framework with which to consider navigation. Human information seeking behaviors are considered in the context of information seeking using computerized information retrieval systems. At the end of the chapter, specific criteria about which to build navigation-based IR systems are generated, and research questions are specified, both of which will drive the rest of the work. Chapter 2 reviews literatures related to this work which were not introduced as part of the conceptual framework. In Chapter 3, the three-part investigation of navigation is described. The first part was the construction of an information space which was intended to have qualities in common with human cognitive space, for the purpose of facilitating navigation. Then, a prototype IR environment was created using a visual interface to the information space. Finally, user-based methodologies were employed to investigate the applicability of the information space and prototype system for information retrieval. The results are analyzed in Chapter 4, and Chapter 5 summarizes the findings of this work. As will be discussed, the conceptual framework generated in this chapter was given some empirical support. Specifications for future system design are drawn in Chapter 5, and directions for future research. 1.1 A Conceptual Framework for Understanding Navigation This section introduces a conceptual framework from which to under- stand navigation for information systems. The section starts with some considerations for theory building, and then the framework building starts with a consideration of human communication. This work attempts to fit in several different areas of social science at once and play several of the traditional roles of information retrieval research found in the literature. This chapter is primarily concerned with theory building, in which concepts are defined and discussed and linked together, and laying the foundation for the rest of the work. Chapter 2 will give background, substantiation, and further direction for the conceptual, methodological, and system-building concerns of this work. Chapter 3 attempts to combine the conceptual framework from this chapter with practical methods for system design and evaluation in the literatures, so that a social scientific study of the framework can be carried out. This subsection addresses the current work in its role of laying the foundation for eventual theory. 1.1.1 Human Communication, Cognitive Movement, and Cognitive Space Human communication may be described as ongoing, dynamic interaction, in which two or more parties exchange messages and meanings (e.g., Cushman and Cahn, 1985). Human communication might be dyadic or involve more actors, it can take place synchronously or asynchronously, feedback may be immediate or only occasional. Scholars of human communication have divided the field into various domains: interpersonal communication, mass media, political communication, intrapersonal communication, computer-mediated communica- tion, etc. (Littlejohn, 1983). In all of the various types of human communica- tion, intentional agents use a variety of media to share meaning. The work of George Herbert Mead lies at the deep foundation of the current work (and is also central to several bodies of theory for human communication. See Littlejohn, 1983). Mead's work ranged across the entire realm of social existence, including functionalism and social behavior, language, ethics, time and space, and interpersonal and intrapersonal communication, from both a theoretical and philosophical standpoint (Aboulafia, 1991; Joas, 1980). He was more than a father figure to sociology and the symbolic interactionist perspective on human communication -- he was a seminal figure in 20th century Western thought (see, for example, Cronk, 1987). Mead's work did not receive all the attention it deserved from US scholars, but has seen the focused attention of such central figures in continental philosophy as Habermas, Tugendhat, and Joas, according to Aboulafia (1991). Aboulafia posits that Mead's work is at the core of much of 20th century social science, although he is often not credited. For the current work, Mead's notions of how humans communicate and the models that people have of each other and themselves are at the focus. Mead (1960) stressed the importance of models of the other for communication: in order for messages to be effectively sent and understood, each communicator must have a model of the other. The model might be at a basic stereotypical level, or it may be a result of intimate knowledge. In the case of an intimate model, messages can be transmitted more easily (Newby, 1988) -- take the example of long-term marriage partners who are able to send many messages effectively without many words. If the model is not accurate, "breakdown" may occur (in the sense of Maturana and Varela, 1987). For instance, a diner might order a ham-and-cheese sandwich in a vegetarian restaurant, or a librarian might deliver a book about friendly household pets to someone interested in taxidermy. The models of the other which humans possess make them capable of tailoring a message so that it might have the desired affect on another (although such processes are largely unconscious, and do not imply any necessary attempt at coercion by communicators). Mead did not go into detail about what the models might consist of, how they might be stored, or what processes integrate a given model with a desired meaning to form a particular message. He did propose that such models would include past experience with particular people and situations, and would involve integration of both general and specific knowledge (Mead, 1960). A model of the "generalized other," according to Mead, makes it possible to negotiate novel situations with reasonable success. The ability to model the self allows humans (and other beings) to relate a particular communication episode to their own existence. A uniquely human ability, according to Mead, is to model how another perceives the self. This enables people to place themselves "in the other person's shoes," and perceive a particular interaction or experience as another might. Taking someone else's point of view is known in common experience to help smooth interaction, and is at the basis of many types of communication instances (from, say, psycho- logical therapy to political solicitation). The model of how another perceives the self might not be accurate, of course, and is probably not consciously constructed. It is the model of the other, the model of the self (or self- concept), and the model of how the (possibly generalized) other perceives the self which enables people to send and receive the communication messages to which they are exposed. Communication scholars, psychologists, market researchers and others are interested in the effects of messages. In a simple model, the utterances and gestures made by someone in, say, a dyadic encounter, might be considered as messages sent to the other. In real-world situations, of course, it is difficult to identify unambiguously the nature of the messages being sent. The goal of mass media messages, such as advertisements, is cognitive movement. All communication results in some cognitive movement. The resultant cognitive movement may have been intended or not by the sender of the messages, just as the message may or may not have been sent intentionally. Cognitive movement is a change in meaning: when understandings of concepts, or the concepts themselves, are subjected to change. Cognitive movement might alternatively be called "learning" (but without implying active information seeking by the learner), "attitude change" (but referring to all that is known, not just attitudes -- whatever they are), or simply "change in knowledge" (but without implying that the change is great or significant, or even noticed). In the simple case, the effects of a single exposure to an advertisement may be gauged relative to pre-measured meanings (Woelfel, Holmes, Cody and Fink, 1988). In multi-party real-world interaction on an ongoing basis, messages are exchanged among actors so that meanings in the end are a combination of all messages sent with the pre-existing meanings (Kincaid, 1988. Note that while some of the mathematics in Kincaid's Conver- gence Theory have been found faulty, the basic premise has empirical support). A considerable body of empirical and theoretical work concerning cognitive movement has been generated by Dervin and Nilan (see Dervin, 1983; Nilan and Rosenbaum, 1991). Their "sense making" paradigm presents cognitive movement as a fundamental human condition, in which humans actively seek to make sense of their surroundings. Empirically, cognitive movement is measured by considering gaps in the knowledge of actors. Cognitive movement is accomplished when a gap is bridged. Methodologically, gaps are defined as questions or uncertainties actors experience. Cognitive movement through gap-bridging might result in increased understanding or new knowledge, but can also result in decreased understanding or identifica- tion of new gaps. Even if gaps are not successfully "bridged," cognitive movement might take place. -------------------------------------------------------------------------- Definition 1.1 Cognitive Movement: A fundamental human condition in which new meanings are sought or obtained. A change in what a human agent knows. -------------------------------------------------------------------------- Cognitive movement refers to changes in what is known and is limited in that it does not deal in how things come to be known, the mechanics of knowing, or thought processes (cognition). Human beings are considered to be active agents in the collection of information and maintenance of what they know. It is based on the premise that any information received (or created) by an individual will by definition result in some cognitive movement. This is not so much in the tradition of Shannon and Weaver's (1949) information processing theory in which information is that which reduces uncertainty and can be measured in "bits." It is more akin to Woelfel and Fink's (1980) description of how what is known changes as information is received, but things which are already well-known are less subject to change. Cognitive movement is not a metaphor, it is a way of describing human behavior. (Although there is nothing which physically moves, there is a change in what is known. By analogy, "counting" can be thought of as "moving along a number line." The generation of a spatio-temporal continuum for the discussion of counting is not inconsistent with the behavior under study, and does not imply an overlap with any physical domain. For the current purpose, "movement" refers to change, but does not imply that some object needs to move through space in order for "cognitive movement" to take place. The term "cognitive movement" was chosen because it may have less a priori meaning than terms such as "learning," because it does not imply a massive or significant change in what is known as might a term such as "cognitive change," and because the term "cognitive movement" has the benefit, as will be seen later, of fitting well with previous theory and research related to this work, such as Salton, in Salton and McGill, 1983, Woelfel, in Woelfel and Fink, 1980, and Koll, 1979). The fit between cognitive movement as defined above with Dervin's and Nilan's sense making approach to cognitive movement is fair, but not perfect. They study the information sought and received, the situation surrounding the information need situation, and the related affect. The current work is less concerned with the perceived changes in what is known, which are accessed by interviews with respondents, than with what is known. (This certainly does not imply that what is known by an individual can be readily ascertained, if at all. Indeed, both "cognitive movement" and "cognitive space," as introduced later, will be only partially explicated for the current purpose, and partially left in a black box). The activities concerning the active search for informa- tion are external to the conceptual framework presented here, as are, to a large extent, human perceptions of changes in knowledge. Thus, the methodological concern of Dervin and Nilan with gathering data on perceptions of information need situations is not a component to the framework being built here, but there is definite overlap of interest in cognitive movement, as it results from exposure to information. Notice that the "gaps" discussed by Dervin and Nilan are not meant to signify physical phenomena. They are metaphors for mental phenomena, the mechanics of which are outside of the sense making paradigm. This use of physical terminology to refer to cognitive phenomena will be employed later in a discussion of navigation and information space, in which the terms used are often applied to physical domains, but are appropriate at a higher level of conceptualization, where the higher level of conceptualization remains powerful, perhaps more so, for the physical domain. It might be inferred from Dervin (1983) and other works in the paradigm that the distinction between a physical gap or barrier and a mental one is practically non-existent, from the standpoint of theoretical consideration of the cognitive processes involved in bridging that gap. This is not to say that mental gaps are like physical phenomena. To the contrary: it is to say that a physical gap, and the act and necessity of crossing it, are cognitive phenomena. Therefore, physical phenomena may be considered a subset of cognitive phenomena when considering the cognitive processes involved in interacting with them. It is suggested here that cognitive space is the necessary medium in which cognitive movement takes place. In the current context, cognitive space is presented as a stepping stone to concepts more appropriate for consider- ation of information systems. For this purpose, cognitive space is made up of two things: concepts and relations among concepts. Such a cognitive space is sufficient, perhaps, to describe the substance of human communication and sense making behavior. Concepts might be simple referents to physical entities (boars, hogs, swine), more complicated entities (the constituency, females, the ERIC database), or vague ideas (democracy, freedom, philosophy). Relations might be temporal (as in the case of most of Dervin's and Nilan's works), based on perceived similarity (as in multidimensional scaling, a psychometric method adapted by Woelfel, described in Woelfel and Fink, 1980), typeless or simply directional (for most hypertext applications), or a variety of other types (hierarchical, causal, relational...). Changes to concepts or relations among concepts are the outcomes of cognitive movement. -------------------------------------------------------------------------- Definition 1.2 Cognitive space: The necessary medium in which cognitive movement takes place. Consists of concepts and relations among concepts. -------------------------------------------------------------------------- Cognitive space is sufficient, as explicated here, for this discussion of cognitive movement for human communication and sense making behavior, but does not provide understanding of human thought processes any more than the scholars cited here have, although it does provide a consistent base for consideration of various types of cognitive movement. Cognitive space is a construct employed to aid in consistency and understanding for cognitive movement. Cognitive space is, at best, a pale shadow or metaphor for the substrate of human thought. This means that while we known that "cognitive movement" takes place and can to some extent measure the causes and outcomes of the movement, we are left with only a high-level understanding of what has changed. This is the same situation as educators might face when discussing "learning:" they can measure outcomes of new knowledge, and postulate that some change in what is known has occurred (and would presumably agree with the statement that what is known consists of concepts and relations among them), but have little concrete understanding of how the concepts are "stored" in human memory, what the relations are, or how these things have changed. An additional component of cognitive space can be identified, but it is not a necessary part of the definition. "Dimensions," or "types" of relation- ships are a part of a cognitive space. For the current work, only a very low- level type of relationship, similarity, will be employed. Many other types of relationship can be identified. Hierarchical relationships, causal relationships, similarity on a particular quality (height, color, ease of use, etc.), and nominal relationships might exist, depending on the nature of the concepts. Identifica- tion of the different types of relationships, their associated situations, and their relative importance in both cognitive and information space are left for a future study. "Similarity," as used in this work, is proposed by Woelfel and Fink (1980) as a fundamental measure for cognitive research, in that all other measures, even nominal ones, incorporate notions of similarity. Their empirical measure for similarity, however, is "dissimilarity." This is because a ratio- level score of zero on a dissimilarity scale corresponds to identity (that is, a score of zero on a dissimilarity scale means the items measured are identical). For a corresponding scale measuring similarity, even an infinitely high number would not indicate identity. The term "similarity" will be used in this discussion in favor of "dissimilarity," because it is more familiar and because the distinction is more useful at a methodological level than a conceptual one. Methodologically (in Chapter 3), this work will make use of a scale which provides for an absolute (bounded) score for both similarity and dissimilarity. Similarity is a low-level type of relation among concepts in that it does not specify the different relationships which might be perceived among concepts by humans. However, similarity has been shown empirically to have the benefit of allowing for more specific types of relationships to emerge from similarity data. Multidimensional scaling (MDS), a psychometric measure of similarity (or dissimilarity), has been used for such purposes. For example, Woelfel and Fink investigated whether Osgood's Semantic Differential scale, which postulated an orthogonal relationship between the dimensions of good- bad, strong-weak, and active-passive, could be validated using MDS (Woelfel and Fink, 1980). They found that "goodness," "strength," and "activeness" did emerge from dissimilarity data as bidirectional components, but the relations were neither orthogonal nor on a unit circle, as predicted by Osgood. This and other studies indicate that similarity-based measures were shown to have the utility of semantic differential or Likert-type measures (see Babbie, 1990) for identifying types of relationships. This discussion of human communication, cognitive movement, and cognitive space is not meant as an answer to the questions that communication scholars and others have asked over the ages. It is meant as a conceptual framework for understanding human communication and sense making behavior, derived from the theory and empirical evidence of several para- digms. More importantly, it is the basis for considering navigation as a fundamental concept for information systems. 1.1.2 Information Space and Navigation In this subsection, the conceptual framework proposed above for human communication systems will be applied in consideration of information systems. In human communication, active agents exchange messages and meanings. When human agents interact with information systems, however, the exchange of messages may be one-sided (for instance, when someone reads a train schedule). In the case of two-way exchange of messages, as is more common when computers are involved, there is still little potential in existing systems for interaction between cognitive spaces as described for human communica- tion systems above. For instance, a user querying a database will get a response from the system (two-way communication), but the contents of the database will not be changed in response to ongoing user input. In a cognitive space, cognitive movement may take place. In almost all information systems, computerized or not, the concepts and relations among them are not subject to change. (That is, the concepts and relations among them are not subject to change in response to communication with users. The system designer or database manager may, of course, make changes.) In this case, no analog to cognitive movement of the system database may take place, and the space of the system is (by definition) not subject to change. The term introduced to refer to such a space is "information space." -------------------------------------------------------------------------- Definition 1.3 Information Space: The concepts and relations among them stored by an information system; typically not subject to change through interaction with the system's users. -------------------------------------------------------------------------- Information spaces may be book indices, databases, bibliographic collections, computer interfaces, and so forth. Information space, unlike cognitive space as described in the previous section, is more than a metaphor. For cognitive space, there is no currently accepted method for knowing all of the different concepts, relations among concepts, and whatever else might make up human knowledge. However, we do have full access to the contents of an information space, inasmuch as those contents were explicitly made a part of the space. The words used as labels for concepts are what are known explicitly, not the concepts themselves -- there may still be ambiguity in the meaning of the concepts. This is not to say that the space is somehow "objective" and entirely knowable: indeed, it is a premise of this entire work that all information spaces are subject to interpretation and individual perspective. For an information space, we are able to know both the concepts which are part of the space (or at least the words which serve as labels for the concepts) and any relationships they have to each other, such as relational or hierarchical relationships. Consider a relational database: each field is known, as are the contents of each field. Further, any relations among fields are also known, as they were formally and explicitly specified. As defined here, "information space" is not a metaphor, but it is not quite so "real" as cognitive movement. It is a concept which exists in various literatures (which will be reviewed in Chapter 2) without formal definition. Information space cannot be a metaphor, since there is nothing we can point to as its referent. It is a way of talking about the contents of an information system. (The argument for information space as non-metaphorical follows the same lines as for a number line as non-metaphorical: numbers do not exist on a line, nor do numbers have anything you can point to as directly indicating their existence at all. Likewise, concepts and relations among them are not found in an information space and are similarly non-corporal. Both number lines and information spaces can be used to organize and talk about their respective attributes, but they are not similes, analogies, or metaphors, they are simply helpful ways of talking about and delimiting things). There must be a resolution to the dual quality of information space as something which is known explicitly and at the same time only knowable through interpretation. The resolution is based on ontology: without an observer, the information space cannot be said to have any meaning at all. However, since there is an artifact which was generated by a human, we can use the formal specifications which generated the artifact as a basis for discussing the information space. This means that the information space may be referred to as "objective," but there is no way to interpret the contents of the space without an observer, making any perception or description of the space necessarily "subjective" (an epistemological assumption which is revisited in the discussion of relevance in a following section). The key to this work lies in considering the outcome when a human agent's cognitive space interacts with the information space of an information system. If the information space is not subject to change, but the cognitive space (necessarily) is, then any convergence of meanings must be one way: the user must adjust to the meanings found in the information system. This is because the system, unlike a human communication partner, will not and cannot adjust for its user. Part of the long term goal of this work is to point out the necessity for making system information spaces more changeable, and more similar to human cognitive spaces. However, such information spaces are not currently the norm. This work will make use of non-dynamic information spaces as they currently exist, but with some enhancements which will be discussed later. The active formation of models so that communication may take place between humans could be described as negotiation: agents make changes in their models of the other and to knowledge about their relationship to the other during the communication process. Negotiation, as cognitive movement behavior, involves both knowing the other and making the self known to the other. While overt personal management skills sometimes come into play, negotiation is for the most part ongoing and without much conscious supervision. For any given communication instance or relationship, there are only some things which must be known for effective communication to take place. For instance, you might not need to know someone's religious affiliation to sell them a car, but such knowledge could be useful to sell a cemetery plot. -------------------------------------------------------------------------- Definition 1.4 Negotiation: The process by which human agents form and revise models of each other during communication. -------------------------------------------------------------------------- Negotiation refers to the often unconscious give-and-take while people get to know each other in some context -- it is not proposed as a bargaining process, but more of a "definition of the situation" (in the sense of Goffman, 1974, in Littlejohn, 1983). A sense of the ongoing, interactive nature of negotiation, as the term is used here, can be had from a reading of Cushman (1977), who talks of how the "rules" of a communication interaction are not obtainable explicitly, even from the participants, and yet they both govern and limit the interaction, and are in a constant state of flux. Negotiation is a necessary component for effective communication to take place: it is the part of a communication interaction which lays down the groundwork (alternatively stated, it defines the situation or produces the rules). Rules and definition of the situation, at least syntactically, imply that their associated activities happen up front -- at the very start of the interaction. The term "negotiation" allows for such up-front activities, but also implies that changes in the rules of an interaction or redefinitions of the situation are constant. There is no negotiation when the information space is not subject to an equivalent of the cognitive movement behavior of humans -- when the "other" can neither form nor revise a model. Instead, the user must navigate through the information space. Navigation means that the user must come to under- stand and act within the model of him or herself that the system has formed of him or her a priori. Navigation behavior occurs in humans when they must make sense of an information space, especially when they are just coming to know the space or a portion of it. -------------------------------------------------------------------------- Definition 1.5 Navigation: Human behavior to form and revise a model of an information space. Involves coming to understand both the information space and the model that an information system has of its (generalized) users. -------------------------------------------------------------------------- The difference between negotiation and navigation is that in negotiation all communicators actively form and reform models of each other. For navigation, model formation is one-sided: only the human component actively forms a model of the information space. The information space does not actively change in response to ongoing interaction with its human communica- tion partners (see Table 1.1). Table 1.1: Summary of Key Concepts -------------------------------------------------------------------------- Nature of Concept Concept Description negotiation human behavior occurs between human communicators navigation human behavior between human(s) [which possess cognitive space(s)] and information systems [which possess information space(s)] cognitive possessed by the dynamic knowledge possessed space humans by an individual information stored by non-dynamic knowledge space information systems -------------------------------------------------------------------------- * "knowledge" refers to known concepts and relations among them. The term "navigation" allows a distinction between the interactive modes of communication for human agents and the one-sided communication typical of current information systems. It also fits well with the concept of "informa- tion space" as discussed above. For instance, navigation behavior might lead to only a partial understanding of an information space -- further navigation would be needed to understand the rest of the space. This is similar to when people know each other and can communicate well in a business setting, but find gaps in their models of the other in social settings. An analogy for an information system might be someone who can use electronic mail within a computer operating system (a type of information system), but cannot perform other functions. Navigation is not strictly limited to the early stages of a relationship with an information system, however. The model one has of a system, and the understanding of its contents, might be less subject to change as time goes on. This is the same thing that happens when people form models of each other: at the beginning of a relationship, what is known about the other changes rapidly. After some time though, there is less need for change in the model. For understanding the information space, consider a word processor: a user could be adept at using a particular program, but then have problems adapting to a new version of the program if some menus or options were changed. Navigation behavior does not imply a physical domain in which to navigate, although some navigation behavior involves physical domains. The important part of navigation is not movement through a physical space, but the cognitive movement and model building on the part of the human navigator. Although the literatures on wayfinding and geographic systems (as will be discussed in Chapter 2) are mostly devoted to physical spaces such as cities or buildings, navigation behavior as described here and in those literatures are remarkably similar. People form a model of an information system, of another human agent or agents, or of a city or building in essentially the same way and for the same purpose. The purpose is to be able to understand one's surroundings and to be able to find what you want to find. The model is formed through exploration, application of models from similar situations and testing and refinement. "Wayfinding," as a field of study, seems more concerned with human cognitive behavior than locomotion in a physical environment. Navigation, as defined here, is not a metaphor, but is human behavior which occurs when people interact with information spaces. A philosophical assumption of this work is that a physical domain -- say, a building or city landscape -- cannot exist without a human to possess a model of the domain. Thus, a physical domain, from the point of view of a human agent, is simply another space in which to navigate. This assumption does not deny the existence of an arrangement of bricks or other building materials, but posits that everything that makes a building a building, rather than, say, a work of art, a tree, or a spaceship, exists only in the mind of a human agent. Navigation, as the term is defined here, is raised to a level which goes beyond physical relocation in a physical domain. Navigation accounts for the cognitive behavior and model building which necessarily precedes physical movement. Physical movement, seen in the light of the definition for navigation offered here, is secondary. Current efforts in the field of virtual reality (e.g. Rhein- gold, 1991), point out the need for de-emphasis on physical movement for navigation behavior. 1.1.3 Models for Human Communication and Navigation Four types of knowledge, or four components of the models necessary for human communication, can be identified. Three of these are especially important for a navigation-based approach to information retrieval. The models are: a model of the other, a model of the relationship to the other, a model of how to change that relationship, and a model of the self. Each will be described in additional detail in the following paragraphs. These models apply for any type of communication situation, including the situation under study here where one communicator, the information retrieval system, is not a human and has, at best, limited modelling capabilities. The model of the other tells us something about how the messages we send will be received. Any intentional communication has some goal, even if the goal is not always explicit. In order to be able to achieve some goal through communicating, we must know something about the other. According to Mead, we might have specific knowledge about a particular other, and specific shared experiences with it. In many cases, however, we do not. Therefore, we need to apply a model of the "generalized other." Mead stated that we do not generate our messages for what the other "is" (even if such a existential statement about the other were reasonable in Mead's theory), but at what we perceive the other to be. Mead believed the model of the other includes something unique to humans: a model of how the other perceives the self. This component of the model of the other can be used to fine-tune the interaction -- to not only attempt to predict the reaction of the other to our messages, but to guess the likely response that the other might make (based on their model of us). The model of the other would include an idea about what the other knows and who they are (i.e., their values, goals, and experiences). The model of the relationship to the other really contains two compo- nents, but they are hopelessly entangled. First is how you perceive the relationship. Second is how the other perceives the relationship. This model would include identification of common goals, any purpose to the relationship (or to a particular interaction), and history of the relationship. The model of how to change the relationship goes furthest beyond a particular relationship. That is, all sorts of general background knowledge from other interactions might be applied to make changes to a relationship. In some cases, changes to the relationship overall might be desired, for example when a co-worker is promoted and wishes to be treated as a superior. Much more frequently, the change to the relationship is momentary -- a change in the topic of discussion, or the introduction of some new component to the relationship. A model of the self is a central factor for intelligent behavior and a fundamental characteristic of human existence (Newby, 1988). This model perhaps plays a back seat in human communication in that it is deeply ingrained in the background of all our thoughts. It is also less likely to change during a particular interaction than the other models (the self-concept, once it emerges during childhood, is, according to Mead, quite stable and usually subject only to incremental change over time). This subsection has briefly introduced the four types of models which come into play during human communication. All four are important for interaction with information retrieval systems. Although the systems (as they currently exist) are not able to perform all of the modelling functions which humans do (or at least not nearly so dynamically), there is nonetheless an implicit model of each type possessed by the systems as implemented by their designers. A later subsection will generate model-based criteria which be applied to the study of IR systems which is a component of this work. 1.1.4 Representation for Information Space Information space, like cognitive space, consists of concepts and relations among them. The relations may be of various types. A representation of an information space consists simply of specifications of the types of relations present in that space and identification of the relations among some or all of the concepts in the space. Representation is of central concern for information retrieval systems. Whatever retrieval mechanisms are in place (say, Boolean set operators), they must perform their functions on a representation of the database. The internal machine representation of data is not of concern here (for instance, whether a particular machine treats a string of bits '01000010' as the character 'A' or not). Instead, the level and number of relations among items in the information space is of interest for information systems, for it is these relations which limit the type of searching mechanisms which may be employed. What follows is my brief analysis and comparison of some common representation schemes. Representation by keyword is the method used by most traditional information retrieval systems. In such a scheme, an index is created by which all document representations which contain a particular keyword may be quickly accessed. The type of relationship in such a scheme is purely logical: either a document representation "is" or "is not" associated with a particular keyword. Similarly, the coincidence of keywords in documents provides for only an existential relationship among documents: either they are related (that is, they have keywords in common) or the relationship is unknown (they do not have coincident keywords). Search mechanisms for such representation schemes are limited to combinations of sets of document representations which do or do not contain a particular keyword. Enhancements to keyword representation schemes include the keyword in context. This provides for relations among terms which occur in the same context in a document. In this scheme, there are some relations among keywords not found in the simpler method described in the previous paragraph. As for the basic keyword representation, though, document or keyword relations are limited to coincidence in a particular index entry, which indicates a relationship exists. Otherwise, the relationship is unknown (or may less precisely be said not to exist). More advanced mechanisms using keyword representation schemes allow for specification of adjacency or proximity. The relational database is another way of representing information. In such a scheme, categories of data are specified. Searching mechanisms can combine different data categories in one search -- for instance, an author/title search. Within each category, however, searching mechanisms are usually limited to those of a traditional keyword index. Note that the 'relations' among categories are mostly up to the user to judge (there is some matching of the information space to his or her cognitive space, created through the specifications for the information space), such as knowing that the name, address, and city fields are grouped together, and occupation, education, and professional affiliations are in a different group. Relationships within a category (say, the 'address' or 'name' field in a mailing list database) are the same as for keyword representation schemes. Relationships across categories (that 'name,' 'address,' 'city,' and 'state' are used for an envelope address, while 'occupation' is not) are left to an application program or the human user. In terms of retrieving desired information, relational databases provide an additional set of sets for Boolean queries, and perhaps some keyword in context operations, but few additional features. Similarity-based representation schemes have been employed to provide for searching mechanisms which allow continuous navigation of an information space instead of the discontinuous navigation found in previously mentioned schemes. Salton and his colleagues (e.g., Salton and McGill, 1983) have created vector spaces in which the keyterms are mutually unrelated, exactly as for keyword representation schemes. However, documents are spatially located at the centroid of their associated keywords in a multidimensional vector space. This is an important advance over other schemes in that it permits searching mechanisms which allow incremental changes to the information need -- changes of degree, not of an entire category. In such a scheme, there are no categories or types of relationships, but there are explicit and known relationships between every document in the database. A note on terminology: the word "keyword" does not appear in my dictionary. Neither does the word "keyterm." Since "keyword" would seem to refer to a single word, and "keyterm" would indicate either a single or multiple-word term or phrase, I have chosen to use the latter word throughout this work. I will use "keyword" when referring to methods or workers which have defined themselves to be working with "key words," as opposed to "key terms." For my purposes, keyterms are more open to the notion that a term, as used, might be ambiguous or subject to personal interpretation, whereas keywords are simply referents to objects which might be looked up in dictionaries or listings of subject headings. All the representation schemes considered in this subsection, and the others not mentioned here, result in information spaces. But the nature of the spaces, and the purposes to which they might be put, and the degree to which they facilitate navigation, are different. The discontinuous spaces, in which all relationships involve sets, are not highly navigable because so few relations exist. In a medium or large database, the majority of keywords and documents will have an unknown relationship. The trouble with using such a representa- tion scheme for searching for information might be seen by considering a different domain: shooting a gun. If a blindfolded person were trying to hit a target, and another person providing feedback, a keyword-only representa- tion scheme would be limited to feedback of the form: "you missed," and "you hit the target." Salton's vector representation provides an important advance over discontinuous schemes, in that (to continue our analogy), instructions to the would-be marksman may take the form, "you hit to the left of the target," and, "you hit to the right." Even a metric is available: "try aiming ten units to the right." Salton's information space is more navigable than a traditional keyword space because it incorporates relations among all documents in the space. This provides for an information space which more closely matches, at least in the qualities it possesses, the cognitive spaces of users, provided the assumption is accepted (after Woelfel and Fink, 1980, among others) that any concepts known to a person have some relation to each other. Salton's scheme is still deficient in that keywords are mutually unrelated, which is inconsistent with both theory (e.g., Woelfel and Fink, 1980), and with empirical evidence (e.g., Harper and van Rijsbergen, 1978). 1.1.5 Facilitating Navigation As Norman (1988) has pointed out, people do not need a user manual for a cup. How might information systems be built which do not require extensive learning time, or waste users' time by producing wrong, misleading, or incomplete results? There are two general approaches which can be taken to information system design: system-based or user-based. The system-based approach is typical of most systems of all sorts. System-based design means that the designer generates specifications for the system's functionality and means for accessing the system functions. In contrast, user-based design starts with the processes the user goes through and language he or she employs to accomplish the system task, and builds the system specifications around them. Combinations of system-based and user-based design are possible, such as employing system-based design for the functional descriptions but user-based design for the interface (Nilan et al., 1989). The advantage of user-based design is that a properly designed system will have minimal startup time: the functions and access methods will be organized as the user expects them to be. With system-based design, the user must first learn the secrets of the system and then translate his or her intentions into the language of the system. The philosophy of user-based system design can be restated using the terminology introduced in the current work. The goal is to minimize startup time and empower users to accomplish desired tasks without translating their needs into language suitable for a particular system. To meet this goal, the information space should match the cognitive space of the user, in the sense that the concepts and relations among them and the language used to reflect these are similar. I say again: The information space should match the cognitive space of the user. This means the system database would be as the user perceives the concepts and relations among them, not as the system designer perceives them. An information system designed in this way would be as a long-known friend, although a changeless and unresponsive friend: the user would be able to predict system responses in various situations. Such a system would be as a well-thumbed book, where desired passages could be found easily (even if they had never been seen before). The next step, to make information retrieval more like human communication, will necessitate the provision of a cognitive space for IR systems (instead of an information space), so that it may be subject to negotiation through ongoing contact with its environment. The final step, as currently envisioned, will be to move from human communication-like systems to exosomatic memory systems. 1.1.6 Communication and Navigation A long-term agenda for the type of information system envisioned in this chapter is to make the science fiction of today a reality -- to have systems which are able to respond to humans as other humans do. In short, to bring the (barely) navigable systems in use today up to the level of human communication systems. A purpose of this work is to consider similarities between human communication systems and information systems and to identify possible fruitful directions for information system design, which might lead to the ultimate goal of information systems as exosomatic memory. The discussion of models earlier in this section is equally appropriate for human communication and information systems. If people did not have a model of, say, a telephone directory (what it is, what it contains, how to use it, etc.), they would be hard-pressed to use it. Although the information space of an information system and the model it has of its users is not subject to change, there still is a definite model of the user which was instituted by the system designer. User-based design involves building a user model consistent with the model which the user expects the system to have of her- or himself. There is no one definition of "success" for information systems and success is not a criterion which can be applied to many human communication encounters. We can, however, think about what makes interaction with either humans or information systems easier or more efficient (in terms of the intended understanding of messages sent or received). Given the existence of an acceptable model of the other, only two components of a communication system can be possessed by actors to facilitate the process of either communication or navigation: 1. A model of one's relationship to the other (in the sense of Mead, 1960) 2. A model of how to change that relationship These components might be restated in Goffman's (1974, in Littlejohn, 1983) terms: people need a definition of the situation and a way of changing the situation. Stated again in Cushman's (1977) terms: people need to know the rules for an interaction and need to be empowered to change the rules. The specific types of knowledge which make up each model depend on the type of relationship. A model of a relationship might include history, similarities to other relationships, communication norms, and so forth. A model of how to change a relationship might include a variety of perceived cause- and-effect occurrences (such as a set of known commands), the predictability of the other, and knowledge of available communication channels. One example from human communication is when someone regularly purchases pizza for lunch at the same spot -- his or her model of the relationship is based on the roles of customer-salesperson, and includes a "script" (again borrowing Goffman's terminology) which the participants follow regularly. His or her model of how to change that relationship involves knowledge about other options for lunch (that is, non-pizza), and how to achieve these options. In order to break out of the "pizza script," the customer needs to change the model of the relationship which both parties have. In order to have a successful non-pizza interaction, in this trivial example, both participants need to adjust their understanding of what's going on by negotiating a new model. The extent to which this might be traumatic or difficult depends on how deeply ingrained the model is. Another example might be based on human-computer interaction. A regular user of word-processing software might receive a new version of the software. If for example, the old version ran under DOS and the new version runs under a windowing system, it might be that some of the command keys are different. In this case, the user, who was proficient in the old version, needs to learn to use the new version. This process, of changing both the (low-level) manner in which the user interacts with the word processor and of rendering the expert user to a less proficient status, is one of changing the relationship between the user and the system. However, the system's view of the system and the model it has of its generalized users (as created by the system designers) does not change in response to the user's changed model: In order to change the relationship, the user has to change, because the system will not. As Mead (1960) pointed out, a uniquely human ability is to model how another models the self -- to see oneself through another's eyes. As the second example demonstrates, the importance of such a model for information system use is paramount: it is incumbent on system users to know how the system expects the user to act. Few current information systems participate at all in helping the user to form a model of the system (at least through use - - training manuals, user guides, and help screens are all external to the system functionality. Such model-building activities might be analogous to the pizza-seller saying, "I am a pizza vendor. You may give me money, and I will sell you a pizza. In the event that you want a different sort of pizza than I have available, I will cook one for you"). The model of how to change the relationship has to do with predicting the impact of messages on the other. Again, this is a general but utterly central component of both human communication and information systems. Without a clear model of the other, you just muck about trying to form a message which generates the desired result -- like pushing buttons and turning knobs randomly in a control room, hoping to turn off a nuclear reactor. Communication and navigation involve the same sorts of processes on behalf of the actor -- both necessitate a model of the other and involve the exchange of messages. The main difference is the extent to which an information system is able to give the type of feedback typical of human communication systems: to dynamically change the model of the user, to understand the cognitive space of the user, and to incorporate the user's model of the system (and the data it contains) when responding to user input. 1.2 Goals for Information Retrieval Systems Notions of navigation for IR systems are informed by consideration of what ideal IR systems might be. An historical image of ideal IR was introduced by Bush in 1945. His "memex" was a machine that provided personal informa- tion on demand. With the advent of interactive computerized information systems in the 1960's, Bush's vision was pursued from two related angles. The first angle is what is here referred to as the "relevance-based" approach. The second has to do with "exosomatic memory." The following two subsections examine each of these angles in turn. A third subsection combines and extrapolates from the discussion thus far some specific criteria to apply to an evaluation of IR systems created in the relevance-based tradition, or in navigation, human communication, or exosomatic memory traditions. 1.2.1 Relevance-Based Information Retrieval Relevance is not a simple concept and it is not at the heart of the current work. It is a topic of considerable history in the literature, and will not be fully treated here. Of current importance is the role of relevance in the design and evaluation of typical functional or experimental information retrieval systems. Despite personal views or philosophies of their designers, most IR systems are designed in the "rationalist" tradition. "At its simplest," according to Winograd and Flores (1988, cited in Schamber, Eisenberg, and Nilan, 1990), "the rationalistic view accepts the existence of an objective reality, made up of things bearing properties and entering into relations." If this "simple" view of the rationalist tradition (as discussed further by Schamber, Eisenberg, and Nilan) is used to understand the design and evaluation of most IR systems in the literature, it is clear that these systems are well placed within this tradition. These qualities of IR systems within the rational tradition are demonstrative of the assumptions which I believe underlie them: - Systems are designed which assume the language a user employs to represent the user's information need (that is, queries are equated to information needs). - Systems are typically evaluated with sets of pre-con- structed queries and relevance judgements which are independent of each other and divorced from a user context. - Relevance judgements are typically binary (yes/no). - Documents are represented in only one way (that is, the keyterms chosen to represent a document are intended to be context free and static). - Queries are independent of each other -- there is no allowance for ongoing or dynamic interaction be- tween system and user. The most common vision of a perfect IR system is one which provides all documents which are relevant to a particular information need and none of the non-relevant documents. As Figure 1.1 illustrates, even if the assumptions of relevance-based IR were accepted, there is a substantial problem having to do with the fact that these systems are not, in practice, matching information needs to information. Instead, they match document surrogates with queries. For the current work, general notions of relevance for IR are not rejected. It is the rationalist tradition within which most relevance-based systems are designed and evaluated which is called into question. Relevance seems a fair evaluation measure for retrieval situations in which users have a very firm notion of what they seek and a sound basis for making judgements about the appropriateness of documents (or surrogates). Perhaps, as researchers move towards the "dynamic, situational" approach to relevance advocated in Schamber, Eisenberg, and Nilan (1990), the role of the rationalist tradition in the design and evaluation of IR systems will decrease. Figure 1.1: What do information retrieval systems match? 1.2.2 Navigation-Based Information Retrieval The more common vision of ideal IR, as mentioned above, has to do with retrieving all relevant documents and no non-relevant documents. A less common vision of ideal IR, but one which is perhaps more closely associated with Bush's "memex," has to do with what Brookes (1975) called "exosomatic memory." Exosomatic memory, according to Brookes, is an IR system operating as an extension of human memory. Such a system might possess intimate knowledge of particular users or user groups and models of the types of processes the users would go through. Multiple placement of items in the database, multiple access methods, and serendipitous retrieval of items are assumed as parts of such a system because introspection reveals that they are in evidence in human memory processes. Exosomatic memory is an ostensible long term goal of relevance-based approaches to IR (discussed further in Chapter 2), but the qualities of exosomatic memory systems seem inconsistent with the assumptions usually inherent in relevance-based IR systems. An intermediate goal I see as being on the way to exosomatic memory is to develop information systems which are closer to human communication systems. A criterion about which to design IR systems to eventually meet that goal is that they facilitate navigation. Navigation is proposed as the most suitable alternative to relevance, and the only alternative I can think of, until there is a way to have information systems which are able to dynamically refine models of their users, systems which communicate as humans do. This work is not intended as a large step towards ideal notions of "exosomatic memory," but it is introduced as a change in direction for IR which facilitates the type of cognitive movement for model formation and change described in the preceding section, more so than systems which do not focus on navigation. Making IR systems more navigable is proposed as a step towards human communication systems for information retrieval, followed eventually by exosomatic memory. Navigation is not the only path which might be taken towards exosomatic memory. Researchers and theoreticians operating in the rationalistic model, in the pursuit of relevance, share the goal of exosomatic memory (at least as demonstrated by the ongoing citation of Bush's 1945 article). As stated earlier, I do not see a path from relevance-based IR to exosomatic memory -- the reliance on matching for relevance-based IR (see Figure 1) is inconsistent with the qualities described in the above paragraph. The practice of evaluating systems in terms of precision and recall, without consideration of the ability to, for example, serendipitously discover new links among items, does not indicate steps being taken towards exosomatic memory. Other paths towards exosomatic memory might include a focus on the neurobiological aspects of information -- to study how human memory works, and develop ways of encoding thought patterns using external devices. Or, study of how people, such as reference librarians, can retrieve information on demand for others -- trying to systemize the processes which people go through, possibly for incorporation in an expert system. These and other approaches have been bypassed for the current work in favor of navigation, because navigation has the dual benefits of being based (as described in this chapter) in a large body of theory of human communication and of being comparable (as described in Chapter 3) to relevance-based approaches to IR. Four phases might be identified along the path towards the ultimate goal of exosomatic memory for IR. First is where we are now: systems which do not explicitly facilitate navigation, and do not attempt to match their information spaces to user cognitive spaces. These systems do possess information spaces, but the spaces are created largely without theoretically coherent modelling of the user. Second is the phase which this work strives towards: making systems more navigable, by providing information spaces which better match users' cognitive spaces and through various other techniques which will be discussed in this chapter and Chapter 3. Navigation is provided in place of negotiation as found in human communication, and navigable information spaces are provided as alternatives to true cognitive spaces. As will be discussed in Chapters 2 and 3, navigation may be enhanced both by providing a more navigable information space (that is, one in which the information space is well matched to a user or user group's cognitive space, in that appropriate concepts have appropriate relations to one other), and by providing navigation cues. The third proposed phase towards exosomatic memory is to make IR systems which perform as human communication systems. Such systems do not imply any sort of human-like intelligence on the part of IR systems, but do necessitate the ability of the systems to negotiate models of the other, as humans do. During ongoing communication, the systems would dynamically modify their database contents, organization, and presentation according to the situation at hand, making their databases, by my definition, cognitive spaces and no longer mere information spaces. The final phase in this research program is to provide for exosomatic memory or a memex-like concept which might be more akin to intrapersonal communication than interpersonal communication. The path to the final phase is not clear at this time. For the present effort, we return to consideration of what might constitute reasonable steps towards the goal of making navigable information systems, which will clear the way for providing systems which are more like human communication systems. Navigation, as described here, can include a full range of information seeking activities, from goal-directed cognitive movement with a specific end-point, or moving towards a vaguely defined goal, or browsing. (These are three types of cognitive movement typically found in the IR literature. Dervin, 1983, lists other types of cognitive movement which are equally suitable, such as resolving uncertainty, getting clarification, solving a problem, overcoming an obstacle, etc.) Navigable information systems are to be built by creating environments where a model of one's relationship to a system, and a model of how to change that relationship, is readily obtainable. Some possible characteristics of IR systems which are built with navigation as a fundamental concept might include: explicit cues as to one's status relative to the system and how to change that status, the use of document representation schemes which closely match how users perceive the documents, keyterms, and relations among them (that is, a correspondence between the user's cognitive space and the system's information space); an explicit model of the information space, in order for users to understand the system information space as it relates to their cognitive space; cues indicating where various items or concepts of importance to the user are located in the information space; the ability to take different views, build maps, retrace steps, and so forth -- to generate the type of interactive and dynamic environment for the exchange of messages typical of human communication. Empowerment of users through navigable systems could result from any of the specific qualities in this paragraph, which fall into one of two general categories. First is the provision of a system information space which more closely matches users' cognitive spaces. Second, navigable systems provide cues which help the user to understand his or her status relative to the system -- the relationship between user and system -- and how to change that status. The assumptions about information retrieval and navigation which underlie the current work are: - We do not yet know enough about user needs, user situa- tions, user goals, etc. to achieve the goal of presenting only documents that would be judged "relevant." Even in a human communication situation, the actors cannot predict fully the results of their messages. - Information retrieval is an information seeking activity, the outcome of which is a user's cognitive movement. - User cognitive space is not static, but is in a constant state of change. Therefore, information retrieval is a dynamic process in which a user's cognitive space inter- acts with a system's information space. - Documents (and the information they contain) are not static: Their meaning, content, and importance exist only according to what they are being used for and who is using them. The notion of a document having some rele- vance value independent of a user is not accepted. - Information spaces might have different meanings for different users, and can have no independent or "objec- tive" meaning. - Users may have very specific goals when approaching an IR system, or very undefined goals. Much use of IR systems is exploratory in nature. The concept of cognitive movement applies to all types of information seeking behavior. - The facilitation of navigation is one step towards making IR systems more like human communication systems and eventually achieving the ultimate goal of IR systems as exosomatic memory. Based on these assumptions, an immediate goal for IR systems is to capitalize on the dynamic, changing features of users and their information needs, and provide systems which facilitate navigation. The intermediate goal is to provide IR systems which approximate human communication systems. The long-term goal is exosomatic memory. This work starts towards providing more navigable systems by first building on assumptions which are consistent with navigation as described here (and therefore not part of traditional design approaches or the rationalist tradition). An effort is made to provide ways of making information spaces which are more consistent with system user's cognitive spaces. Methods of providing knowledge about one's status relative to the system and of how to change that status are developed. 1.2.3 Criteria for Evaluation of Information Retrieval Systems At this point, most of the conceptual framework for this consideration of navigation has been created. An important task remaining is to consider how the definitions from the discussion thus far lead to criteria by which to evaluate information retrieval systems, including those which attempt to take steps towards eventual goals of human communication systems and exosomatic memory. This subsection presents some criteria for consideration in evaluating IR systems. The specific criteria which follow can be divided according to the phase in IR system development at which they apply. For each phase, all the criteria from the previous phases would still be appropriate. 1. Relevance-based retrieval Does the system allow for identification of a particu- lar document from a large collection? Can specific information needs be met ("specific" signifying that a relevance judgement can be made by a user with reasonable knowledge of the subject domain)? Is the nature of the system and its limitations clear? Are cues present to help the user to know how to proceed during a search? 2. Navigable retrieval systems Does the information space share characteristics of the cognitive space (of a user or user group)? Can a fruitful subset of the information space be identified? (where fruitful means that the contents or organization can be used for a variety of purpos- es, not just the retrieval of particular documents to meet specific information needs)? Is the language of the system flexible (e.g., can you use different terms or follow different procedures and end up with a similar set of retrieved docu- ments, or in a similar location in the information space)? Can different perspectives or points of view be applied to the information space? Can incremental changes be made in the system response set? 3. Human communication systems Can the system remember past interactions? Can "world knowledge" be employed to help meet information needs, such as by drawing on experi- ence with other users or purposes, or applying knowledge from the information space? Is there an ability to spot differences in seemingly similar situations, or spot similarities in seemingly different situations? Does the system change somehow in response to messages from its users, on a continuing basis? 4. Exosomatic memory systems Can the system anticipate information needs? Can it search for information independently? Can the system learn from past experience? Implied in the criteria for relevance-based systems are some of the components of Mead's models. A model of the other, a model of the relationship to the other, and a model of how to change that relationship, are all implied. In navigation-based systems, the nature of the desired model of the other changes, so that not just a "learnable" system is presented, but one which somehow corresponds to the user's expectations. That is, the user should be able to rapidly form a model of the other, the relationship to the other, and how to change that relationship, which is largely correct (if not fully detailed), but not explicitly communicated by the system. In other words, the system functionality and organization, not just the information space, should match the expectations of its users. Models implied for human-communication systems include some basic model of the self possessed by the system, and an ability for the system to reconfigure it's functionality and presentation to better match the perceptions of the users. With exosomatic memory systems, the model would necessarily incorporate a large portion of the user's views on the world -- something beyond our current understanding of the sorts of knowledge people use to get through the day, but not unimaginable. Note that some of the criteria above, especially those dealing with model issues, are more appropriately considered from either the theoretician's or system designer's point of view. Others are appropriate for an empirical investigation. The set of criteria in this subsection will be revisited and refined in Chapters 2 and 3, and will form the basis of the analysis and comparison to be described later in this work. 1.3 Goals, Questions, and Definitions The research goal of the current work is to explicate and investigate navigation as a fundamental concept for information retrieval and for information systems in general. "Fundamental concept" means that navigation should be at the heart of current IR systems and at the heart of information seeking behavior employing such systems. As a fundamental concept, navigation (as laid out in this chapter) describes what users of IR systems and other information systems do, even if navigation was not an explicit design specification for the systems. Navigation and its role in the literatures relating to IR will be explored further in Chapter 2. As the literature review will show, navigation is consistent with the goals of various information systems, including existing traditional approaches. The conceptual framework provided in this chapter indicates that navigation is not a true end goal for information systems, but is a reasonable step on the way to systems which are akin to human communication and eventually, human memory. The approach this work takes is to first create a conceptual framework from which to understand navigation and related concepts, and then to create an environment in which navigation can be examined and evaluate the environ- ment for information retrieval. This chapter has started the necessary framework. Chapter 2 examines a wide range of literatures with implications for navigation in information retrieval and ends with suggestions for information system design for navigable systems. The rest of the work describes the creation and evaluation of the IR environment. As navigation has not been formally addressed in the literature of information science and an IR system that focuses on navigation as a fundamental concept has not previously been built, this study is exploratory in nature. The research questions addressed by the current study are: 1. Is navigation a useful approach for operationalizing information retrieval? 2. What perceptions by users engaged in information seeking help to understand the use of navigation for conceptual- ization of information retrieval and information system design? Let us examine the key concepts from the research questions. "Navigation" is as defined earlier in this chapter: human behavior to form and revise a model of an information space. Navigation as an approach for IR involves creating an information space and an information retrieval environ- ment such that the two types of model building are facilitated: the ability of a human user to form and revise a model of his or her relationship with an IR system, and the ability of the user to form and revise a model of how to change that relationship. "Useful" refers to the capability of a navigation-based IR system for accomplishing information seeking tasks as perceived by users of the system. It is a user-based criterion which will be evaluated in the context of users attempting to accomplish basic information seeking tasks with a navigation- based system. Minimally, if a navigation-based system can be employed to meet information needs, it can be said to be useful. However, usefulness might be better considered relative to some other criterion. As Chapter 3 will describe, the usefulness of a navigation-based system will be evaluated relative to a system which was not based on navigation. Criteria will be employed to assess the relative usefulness of the naviga- tion-based system. One set of criteria will include the time taken to accomplish similar tasks and the self-report of satisfaction on a Likert scale. A more important set of criteria are the written comments of users engaged in search tasks (again, Chapter 3 will provide details on the specific comments which will be solicited). Usefulness is not an absolute criterion -- the answer to the first research question cannot be a "yes" or "no." User comments, basic utility, and traditional time-on-task measures will all indicate the ways in which navigation is useful for operationalizing information retrieval. Chapter 2 will introduce some particular desirable qualities for navigable systems which are operation- alized in Chapter 3. Thus, navigation will be evaluated for usefulness as a concept about which to build IR systems and the usefulness of particular aspects of a navigable system will assessed separately. "Operationalizing information retrieval," in the first research question, is a way of saying that navigation will not be addressed from the armchair. An operational IR system, based on the philosophical and theoretical concepts from this chapter and including desired system qualities as will be described in Chapter 2, will be constructed and evaluated for its usefulness. Despite the arguable value of this empirical approach to validation of the claims for navigation introduced in this chapter, problems are introduced in the pursuit of knowledge through the creation and evaluation of a single system. Foremost is the difficulty of having a single system to evaluate: if negative or uncertain answers emerge to the research questions, then it will be uncertain whether the results are due to the inappropriateness of navigation or simply to the particular implementation of navigation. Another problem is that the creation of an evaluation system takes away from possible resources for a more detailed empirical study of navigation. Both of these problems must be endured due to the lack of a suitable navigation-based environment for the comparison. Let us turn to the second research question. "Perceptions by users" indicates that this is a user-based study, employing user-based methods. As such, I am interested in the perceptions of users engaged in information seeking activities with a navigation-based system. This is not simply a system-design experiment: as an effort towards making IR systems which are similar to human communication systems, such an outcome would produce a philosophical conflict. The perceptions to be elicited will be described fully in Chapter 3 and will include unstructured commentary on navigation concepts as found in the evaluation system, open-ended responses to items concerned with particular characteristics of IR systems and navigation environments, and responses to closed-ended items concerning satisfaction and the navigability of the information space. Both the evaluation system and its associated information space will be the subject of user commentary. The term "information seeking" is used less broadly here than in the works of Dervin and Nilan. It is here limited to the types of behaviors typically studied in information science research. Conceptually, information seeking might be limited in the current context to the use of IR systems to retrieve document surrogates in response to an information need. The information need might be for a known item, for information suitable for a particular query, or for general access (browsing). As Chapter 3 will show, information seeking will be operationalized in the narrow context of looking for document surrogates which meet information needs stated in sets of keyterms or sentences. This narrow application of "information seeking" may serve to limit the interpretation of the "usefulness" of navigation. In spite of this limitation, this application of information seeking is very similar to the types of tasks for which IR systems are evaluated in the literatures on information science, but with the benefit of retaining philosophical consistency with the assumptions laid out in this chapter. The understanding of the "use of navigation for conceptualization of information retrieval" refers to the insight gained throughout this work for the conceptual framework for navigation described in this chapter. Naviga- tion, and the long-term research program in which it is a step, is proposed as a fundamental concept for information retrieval. Do Chapters 3, 4, and 5 bear this proposal out? What insight into navigation and navigation for information retrieval will be revealed during the process? The understanding of the use of navigation for "information system design" points not to the theory of navigation, but to the practice. Do the steps derived for the construction of a navigable IR environment work? Which steps seem to hold the most promise, and what perceptions by users help to indicate directions for future research? 1.4 Criteria for Evaluating this Work As an exploratory user-based study, it was difficult to anticipate whither the work would lead. Research questions were a starting point, but neither the evaluation environment nor the user perceptions could be predicted at the outset. This work on navigation was intended as a first step on a path towards exosomatic memory for information retrieval with a proposed route via human communication systems. Two uncertainties about these steps needed resolution. First was the path itself: the conceptual framework for navigation was laid here, but only part of the human communication phase and almost none of the exosomatic memory phase was given as much careful treatment. One criterion for this work was that it yielded insight into whether there is a path beyond navigation and whether navigation, as conceived here, was really its own path or simply a diversion from existing approaches to information retrieval. The second uncertainty was about whether the steps that are taken here were well-taken: were the methodological considerations derived from the conceptual framework described here well-suited for movement along the path to the ultimate goals of IR? This work did not end with the creation of a perfect navigation-based IR system and it did not wait for user perceptions to dictate all of the next steps for reconsideration of navigation for conceptual and methodological purposes. Instead, the work closed when it was clear that there were no further steps to be taken with the few stepping stones laid here. When there were sufficient data to indicate that, first, navigation as a fundamental concept for information retrieval could be looked at in a new light, and second, that alternatives for the design of functional IR systems could be considered with improved knowledge of the role of navigation, this work was drawn to an end. The internal criteria about which this work was organized seem fairly clear. From an outside point of view, the role of this work in the literature on IR and criteria which other scholars might apply to consider the relative merits of the work -- not only as a part of the literature, but as a document meant to serve the purpose of a demonstration of competence for a doctoral degree -- might not be clear. The lack of clarity is because this work does not fit the traditional mold and strives to do several things at one time. Not entirely conceptual, I attempted to lay a framework with immediate uses for this study and possible long-term applicability to examining the domain of IR. Not an empirical study, an empirical study was carried out in the context of evaluating a system designed for navigable IR. And not a system design exercise, this work did involve considerable effort in the area of system design. For an outside reader, the only criterion which made sense to apply to this work was that of contribution: how might this work contribute to the literature on IR and related literatures? Again, the contributions could come from several sources, including the conceptual framework, the empirical design and analysis, and the system-building. Inasmuch as the contribution of this work might not have been evident at the outset, and could be difficult to predict, the other criteria to be applied had to do with the thoughtfulness of the work, the skill with which the tools of system design and analysis were used, and the understanding expressed here of the strengths and weaknesses of the work. 1.5 Conclusion Navigation was a key skill for information seekers before the advent of computer technology (as described in Chapter 2). It was necessary for information seeking with existing information systems. The assumptions of relevance-based IR have resulted in systems which do not provide knowledge about the user-system relationship nor sufficient cues as to how to change that status. More importantly, there has not been a focus on creating information systems which are well-matched to users' cognitive spaces. This work investigated the usefulness of navigation for operationalizing information retrieval. It assessed both the conceptual framework as described in this chapter and the practical criteria derived from it as applied in an empirical information seeking setting, both within the context of a system designed for navigability. As an exploratory study, there was a limited extent to which firm answers concerning the applicability of navigation to information retrieval would be found. Instead, insight into the research questions, and development of grounds for future research were be sought. The work did not represent a deviation from the essence of information seeking behavior as was found in the literatures on human behavior and information retrieval, but was simply a redirection for research and develop- ment based on a the start of a conceptual framework presented here. More navigable systems were proposed as a step towards IR which is more like human communication, towards the eventual goal of IR as exosomatic memory. This chapter has laid the groundwork for the rest of the study. The next chapter examines the role of navigation in various literatures related to IR, and lead towards refinement of the criteria and system-building steps which were only partially developed in this chapter. Chapter 3 introduces the system which was built, how the information space was constructed, and the methods employed to evaluate it. Chapter 4 presents the results of the system design and evaluation, and Chapter 5 draws the work to a close with a revisitation of all the qualities of this work, including the strengths, weaknesses, and implications for the future. CHAPTER 2: LITERATURE REVIEW 2.0 Introduction This chapter discusses several bodies of literature that inform the role of navigation for information systems. While most are not central to the conceptual framework from Chapter 1, they are important either for the framework's support, or because they are pertinent for methodology or design for the theory and practice of IR. The chapter starts with traditional issues for IR system design, including the general batch-oriented outlook of the relevance-based approach and the representation and indexing schemes associated with that approach. The following section considers some non- relevance-based approaches to IR, including a focus on browsing and several viewpoints on the cognitive aspects of users as they may help to understand the information seeking and use process. The next section deals with information space and introduces some of the groundwork which helped define the path for the current work (as discussed further in Chapter 3). The creation of information spaces which are more navigable than those produced by relevance-based approaches to IR has a long tradition, but such spaces have not been at the heart of common systems to date. The following section introduces wayfinding and visualization, which were drawn upon for the system design process described in Chapter 3. Finally, three sections summarize the literature, gather implications for information space and navigation, and give special treatment to researchers whose ideas were particularly influential on this work. 2.1 Traditions in Information Retrieval This section examines some of the more traditional approaches taken to information retrieval. Most of the qualities of current IR systems (including the system used for comparison in Chapter 3) may be found in the literature described here. The section starts with a discussion of some of the founda- tions of IR, including a brief (and not necessarily unbiased) history of the field. Two following sections introduce traditions in document representation and the assignment of keyterms for documents. 2.1.1 Foundations As do many fields, "Library and Information Science" claims its roots in antiquity. The transfer and storage of information have been critical components of survival throughout history. The history of the library shows that great stores of information sources were coveted by powerful rulers throughout the ages (Boorstin, 1983). Modern library and information science is not directed at withholding information resources from all but the very powerful. To the contrary, it is the charge of the modern information scientist to provide access to information resources to all people. In his Libraries of the Future (1965), J.C.R. Licklider envisioned that computers would be used to enable all people to access the information they need from the huge corpus which modern society has created. Information retrieval came into being as an area of study about the time the future role of computers in society was being identified. At the heart of the early literature on information retrieval are rationalist notions of relevance (as presented in Chapter 1). There have been many calls in the literature for different approaches to IR, some of which are included in this chapter. IR textbooks such as Pao's Concepts of Information Retrieval (1989) and Gerrie's Online Information Systems (1983) still offer few views of the field that are not relevance-based. The topics discussed in this chapter include the topics found in texts on information retrieval, such as the evaluation of IR systems, the interaction between system and user, the nature of information needs, and the representation of information. Unlike the (seemingly) related field of artificial intelligence, there is relatively little philosophy associated with the field of information retrieval. IR is a practice-oriented field in which the large number of installed computerized card catalogs, CD-ROM systems, and online systems tend to drive notions of what IR should look like. Workers and theoreticians in artificial intelligence realized long ago that their assumptions about what constituted intelligent behavior needed to be questioned and investigated (e.g., Hof- stadter, 1979. See 'comp.ai,' a USENET newsgroup, for ongoing and current examples of such discussion). Questions about the "intelligent" role of the computer in IR and the necessary "understanding" that the computer must have of its users' needs have seldom been asked in the IR literature. While one might argue that it is the human intelligence that drives effective IR, not a computer intelligence, this potential man-machine symbiosis has seldom been the focus of IR work. 2.1.2 Representation for IR Document representation has a history and scope beyond the field of information retrieval. Representation has to do with the specification and maintenance of relations among some set of items. For our purposes, the relations are known by the formal specifications used to generate them, such as cataloging rules or database format statements. The items, for IR, are usually surrogates or labels for concepts. An important part of the history of document representation has to do with the practice of indexers and catalogers, who have the job of insuring that once a text is put somewhere it can be found again later (see Cleveland and Cleveland, 1983). Classification systems, such as the Dewey Decimal System and Library of Congress Subject Headings, were invented to help classify documents according to their subject area. This section considers the assignment of keyterms, or indexing, a foundation of representation for IR. A later section discusses variations on traditional indexing, specifically having to do with spatial representation. 2.1.3 Assigning Keyterms This subsection is concerned with the activities of assigning keyterms to documents for information retrieval. The goal of indexers and catalogers is to adequately describe the subject area of a given text (book, journal article, photograph, etc.) so that it may be retrieved later. (Catalogers also identify the publisher, author, title, and other information which will not be considered in the current work.) In practice, this involves assigning a set of keyterms to the text, usually chosen from a controlled vocabulary. In an ideal world, the vocabulary and criteria for selecting keyterms would be so well-defined that any trained indexer would choose the same terms for a given text. This ideal vision is not achieved in the real world (e.g., Sievert and Andrews, 1991). The whole practice of assigning keyterms to texts is firmly a part of the rationalist tradition. The job of indexers and catalogers is to assign keyterms which are context free -- independent of any particular user or use. The controlled vocabularies used for such purposes are carefully constructed so that there is minimal ambiguity in the meaning of any assigned term and a minimum number of terms may be assigned (Cleveland and Cleveland, 1983). (In human language, of course, there is considerable ambiguity in the meaning of terms. c.f. Roget, 1977.) For practical reasons, keyterms are assumed for the most part to be independent of one another. A thesaurus may be used to look up related terms, but in all cases the number of terms and the number of relations among them are minimized for indexing and cataloging purposes. The historical practical reason for assigning a small number of keyterms is that only so many terms could fit on a catalog card, or in a computer record, before their numbers made storage and retrieval difficult. The reasons for minimizing the relations among the terms have to do first with reliance on trained indexers and catalogers to assign terms unambiguously, second with the difficulties in assigning term relations which are free of context, and finally with the representation schemes which are incapable of handling large-scale term relationships. That third reason for independence of keyterms for indexing is of particular importance for this work. Document representation for com- puterized information retrieval started and has largely remained at the independent keyterm stage. A small set of keyterms, perhaps six to twelve, are assigned to a document. This set and its associated bibliographic information is the "document surrogate" which is retrieved. The user provides a set of keyterms which represent his or her information need and the IR system matches the user's keyterms to document surrogates. (See almost any IR system, such as those offered by InfoTrac and Dialog, for an example of this type of retrieval.) Note that even in systems which provide access to the full text of an abstract, the same keyterm-based access applies. The main difference between searching a set of keyterms and an abstract is the number of keyterms: one hopes that some of the potential ambiguity in the words chosen to represent a document will be avoided by increasing the odds for overlap between system and user language. Thus, traditional indexing and representation for IR lessen the chance of an isomorphism between user cognitive space and system information space in that the concepts and especially the relations among them are almost definitely not as perceived by any particular user with a particular information need. A later section deals with spatial representation, which is of primary importance for this work. Other forms of representation exist in the literature, but are not considered by the author to have the capability of providing a system information space which matches an individual user's or user groups' cognitive spaces. The argument for this statement is a philosophical one: spatial representation allows for all concepts in an information space to be related to each other. The other representation schemes in the literature do not allow for fully relational representation. The assumption is made that all concepts are related to one another (while admitting that not all relations will be "rich" or particularly meaningful. This argument is given empirical support in the work of Woelfel and his colleagues, e.g., Woelfel and Fink, 1980). For a review of document representation and retrieval methods not included here, such as probabilistic retrieval and various weighting schemes, see Vickery (1971), Richmond (1972), and Batten (1973). Perhaps it is indicative of the continued use of the same representation schemes that there has been no ARIST chapter on "document description and representation" since Batten's chapter. Nor has there been any real change in how things are represented, with the possible exception of Oddy's work (Oddy, 1977; Oddy and Balakrishnan, 1991). Note that Farradane, among others, has advocated what he calls a relational approach to document representation for traditional indexing, but his techniques have not been applied on a large scale (Farra- dane, 1952; Farradane and Thompson, 1980). Nor do Farradane's methods fit well within the framework for information space and cognitive movement developed for the current work. 2.2 Non-Relevance-Based Approaches and the Retrieval Process As mentioned in Chapter 1, the assumptions associated with relevance- based IR may be somewhat appropriate for informed users who have well- defined information needs. These users are familiar with their subject matter, know the terminology of the system, and may have information needs such that binary relevance judgements are reasonable and reasonably independent (for instance, they might be looking for articles suitable for a literature review). There is a whole spectrum of information users and information uses, however, that are not facilitated by the assumptions of the relevance-based approach. This section examines some of the work done on those types of users and uses. 2.2.1 Browsing Browsing has a long history in non-computerized approaches to information use (see Kent and Lancour, 1970). There have been a number of studies to identify the role of browsing and the range of browsing behavior, in information seeking and use (e.g., Morse, 1973; Apted, 1971). Researchers such as Hildreth (1982) and Palay and Fox (1981) realized that users were browsing with IR systems and started to identify the types of behavior that indicated a less directed search than the designers of relevance-based IR systems had intended, and to think about how to design systems which enable those types of behaviors. There has been a recent emergence of computerized browsing systems for IR, usually with the specific audience of users who are not in a strongly goal-directed search mode (Burgess and Swigger, 1986; Larson, 1986; Shneiderman, Shafer, Simon, and Weldon, 1986; Marchionini and Shneiderman, 1988; Pejtersen, 1989). Some researchers who produce browsing systems, such as Canter, Rivers, and Storrs (1985), specifically mention "navigation" as an activity that users perform. Whether they mention navigation or not, most of the browsing systems in the literature provide a visual interface such as HyperCard. Only a few researchers, such as Noerr and Noerr (1985) and Cove and Walsh (1988) consider various representations for data and other more theoretical topics for browsing. Hypertext has the potential to allow user-based links between items in the database, but the literature does not show any work on matching such information spaces to particular cognitive spaces. The literature on browsing and the implementation of IR systems for browsing provides a firm theoretical and practical basis for the implementation of navigation systems for IR and does not fit with the relevance-based assumptions introduced in Chapter 1. Navigation is able to encompass a range of information seeking behavior that goes from that implicitly assumed by relevance-based systems to the type of non-directed and informal behavior which browsing systems attempt to provide for. 2.2.2 Cognitive Approaches Human communication is an interactive, ongoing process. In librarian- ship, recognition of the nature of human communication for the expression and understanding of information needs has manifested in the literature on the reference librarian (Dervin and Dewdney, 1986; Taylor, 1968). For other areas of study, such as expert systems, the eventual resolution of an information problem comes about from an ongoing interaction between system and user (Parsaye and Chignell, 1988). For IR, there is a body of literature directed at interactive systems. Much of this literature points towards the eventual goal outlined in Chapter 1 of having information systems which are more like human communication systems. 2.2.2.1 Dialog-Oriented Systems A small number of IR systems adapt to ongoing user input, attempting to identify a suitable set of documents to retrieve. This is an adaptation of "relevance feedback," and such systems might be called relevance feedback systems. Most of the literature on relevance feedback systems fits within the rationalist tradition (e.g., Salton and McGill, 1983). In contrast, dialog-oriented systems, as Oddy's THOMAS (1977) and later PTHOMAS (Oddy and Balakrishnan, 1991) involve an ongoing communication process, in which the user helps direct the system towards the most useful set of documents in the database. In this process, the system eventually "learns" the user's information need (although systems found in the literature do not remember the interaction for future searches or searchers). The documents are represented in a richly connected network with keyterms. The dialog style of "querying" the system is closer to a human communication approach to IR than to the relevance-based approach. Except for PTHOMAS, there are few explicit links in the literature between relevance feedback of any type and connectionist models or neural networks (Rumelhart and McClelland, 1986). However, the document represen- tations and retrieval processes for relevance feedback systems are closely related to connectionist models which, after all, produce information spaces. In neural networks, input layers (which receive data from the external environment) and output layers (which provide data to the environment) are connected (usually by a middle layer). Such networks may use either local representation in which one concept is associated with one input or output "neuron," or d