AI & Society (2000) 14: 71-97

Collaborative Information Environments to Support Knowledge Construction by Communities

Gerry Stahl

Center for LifeLong Learning and Design

& Institute of Cognitive Science

University of Colorado at Boulder, USA

Abstract: In the information age, lifelong learning and collaboration are essential aspects of most innovative work. Fortunately, the computer technology which drives the information explosion also has the potential to help individuals and groups to learn much of what they need to know on demand. In particular, applications on the Internet can be designed to capture knowledge as it is generated within a community of practice and to deliver relevant knowledge when it is useful.

Computer-based design environments for skilled domain workers have recently graduated from research prototypes to commercial products, supporting the learning of individual designers. Such systems do not, however, adequately support the collaborative nature of work or the evolution of knowledge within communities of practice. If innovation is to be supported within collaborative efforts, these domain-oriented design environments (DODEs) must be extended to become collaborative information environments (CIEs), capable of providing effective community memories for managing information and learning within constantly evolving collaborative contexts. In particular, CIEs must provide functionality that facilitates the construction of new knowledge and the shared understanding necessary to use this knowledge effectively within communities of practice.

This paper reviews three stages of work on artificial (computer-based and Web-based) systems that augment the intelligence of people and organizations. NetSuite illustrates the DODE approach to supporting the work of individual designers with learning-on-demand. WebNet extends this model to CIEs that support collaborative learning by groups of designers. Finally, WebGuide shows how a computational perspectives mechanism for CIEs can support the construction of knowledge and of shared understanding within groups. According to recent theories of cognition, human intelligence is the product of tool use and of social mediations as well as of biological development; CIEs are designed to enhance this intelligence by providing computationally powerful tools that are supportive of social relations.

Keywords: Collaborative learning; Communities of practice; Construction of knowledge; Perspectives mechanism

1. Introduction:  The Need for Computer Support of LifeLong Collaborative Learning

The creation of innovative artifacts and helpful knowledge in our complex world – with its refined division of labor and its flood of information – requires continual learning and collaboration. Learning can no longer be conceived of as an activity confined to the classroom and to an individual’s early years. Learning must continue while one is engaged with other people as a worker, a citizen, and an adult learner for many reasons:

·       Innovative tasks are ill-defined; their solution involves continual learning and the creative construction of knowledge whose need could not have been foreseen (Rittel & Webber, 1984) .

·       There is too much knowledge, even within specific subject areas, for anyone to master it all in advance or on one’s own (Zuboff, 1988) .

·       The knowledge in many domains evolves rapidly and often depends upon the context of one’s task situation, including one’s support community (Senge, 1990) .

·       Frequently, the most important information has to do with a work group’s own structure and history, its standard practices and roles, the details and design rationale of its local accomplishments (Orr, 1990) .

·       People’s careers and self-directed interests require various new forms of learning at different stages as their roles in communities change (Argyris & Schön, 1978) .

·       Learning – especially collaborative learning – has become a new form of labor, an integral component of work and organizations (Lave & Wenger, 1991) .

·       Individual memory, attention, understanding are too limited for today’s complex tasks; divisions of labor are constantly shifting and learning is required to coordinate and respond to the changing demands on community members (Brown & Duguid, 1991) .

·       Learning necessarily includes organizational learning: social processes that involve shared understandings across groups. These fragile understandings are both reliant upon and in tension with individual learning, although they can also function as the cultural origin of individual comprehension (Vygotsky, 1930/1978) .

The pressure on individuals and groups to continually construct new knowledge out of massive sources of information strains the abilities of unaided human cognition. Carefully designed computer software promises to enhance the ability of communities to construct, organize, and share knowledge by supporting these processes. However, the design of such software remains an open research area (Stahl, 1999) .

The contemporary need to extend the learning process from schooling into organizational and community realms is known as lifelong learning. Our past research at the University of Colorado’s Center for LifeLong Learning and Design explored the computer support of lifelong learning with what we call domain-oriented design environments (DODEs). This paper argues for extending that approach to support work within communities of practice with what it will term collaborative information environments (CIEs) applied both to design tasks and to the construction of shared knowledge. The paper illustrates three stages our efforts have gone through in this direction during the current decade with illustrative software systems.

Section 1 illustrates how computer support for lifelong learning has already been developed for individuals such as designers. It argues, however, that DODEs – such as the commercial product NetSuite – that deliver domain knowledge to individuals when it is relevant to their task are not sufficient for supporting innovative work within collaborative communities. Section 2 sketches a theory of how software productivity environments for design work by individuals can be extended to support organizational learning in collaborative work settings known as communities of practice; a scenario of a prototype system called WebNet illustrates this. Section 3 discusses the need for mechanisms within CIEs to help community members construct knowledge in their own personal perspectives while also negotiating shared understanding about evolving community knowledge; this is illustrated by the perspectives mechanism in WebGuide, discussed in terms of three applications. A concluding section locates this discussion within the context of AI and society.

2. Augmenting the Work of Individual Designers

In this first Section we discuss how our DODE approach – which has now emerged in commercial products – provides support for individual designers. However, because design (such as the layout, configuration, and maintenance of computer networks) now typically takes place within communities of practice, it is desirable to provide computer support at the level of these communities as well as at the individual designer’s level and to include local community knowledge as well as domain knowledge. Note that much of what is described in this section about our DODE systems applies to a broad family of design critiquing systems developed by others for domains such as medicine (Miller, 1986) , civil engineering (Fu et al., 1997) , and software development (Robbins & Redmiles, 1998) .

2.1 Domain-Oriented Design Environments

Many innovative work tasks can be conceived of as design processes: elaborating a new idea, planning a presentation, balancing conflicting proposals or writing a visionary report, for example. While designing can proceed on an intuitive level based on tacit expertise, it periodically encounters breakdowns in understanding where explicit reflection on new knowledge may be needed (Schön, 1983) . Thereby, designing entails learning.

For the past decade, we have explored the creation of DODEs to support workers as designers. These systems are domain-oriented: they incorporate knowledge specific to the work domain. They are able to recognize when certain breakdowns in understanding have occurred and can respond to them with appropriate information (Stahl et al., 1993) . They support learning-on-demand.

To go beyond the power of pencil-and-paper representations, software systems for lifelong learning must “understand” something of the tasks they are supporting. This is accomplished by building into the system knowledge of the domain, including design objects and design rationale. A DODE typically provides a computational workspace within which a designer can construct an artifact and represent components of the artifact being constructed. Unlike a CAD system, in which the software only stores positions of lines, a DODE maintains a representation of objects that are meaningful in the domain. For instance, an environment for local-area network (LAN) design (a primary example in this paper) allows a designer to construct a network design by arranging items from a palette representing workstations, servers, routers, cables, and other devices from the LAN domain. Information about each device is represented in the system.

A DODE can contain domain knowledge about constraints, rules of thumb, and design rationale. It uses this information to respond to a current design state with active advice. Our systems used a mechanism we call critiquing (Fischer et al., 1993/1998) . The system maintains a representation of the semantics of the design situation: usually the two-dimensional location of palette items representing design components. Critic rules are applied to the design representation. When a rule “fires,” it posts a message alerting the designer that a problem might exist. The message includes links to information such as design rationale associated with the critic rule.

For instance, a LAN DODE might notice that the length of a cable in a design exceeds the specifications for that type of cable, that a router is needed to connect two subnets, or that two connected devices are incompatible. At this point, the system could signal a possible design breakdown and provide domain knowledge relevant to the cited problem. The evaluation of the situation and the choice of action is up to the human designer, but now the designer has been given access to information relevant to making a decision (Fischer et al., 1996) .

2.2 NetSuite: A Commercial Product

Many of the ideas in our DODEs are now appearing in commercial products, independently of our efforts. In particular, there are environments for designing LANs. As an example, consider NetSuite, a highly rated system that illustrates current best practices in LAN design support. This is a high-functionality system for skilled domain professionals who are willing to learn to use its rich set of capabilities (see Figure 1). NetSuite contains a wealth of domain knowledge. Its palette of devices that can be placed in the construction area numbers over 5,000, with more downloadable from the vendor every month. Each device has associated parameters defining its characteristics, limitations, and compatibilities – domain knowledge used by the critics that validate designs.

In NetSuite, one designs a LAN from scratch, placing devices and cables from the palette. As the design progresses, the system validates it, critiquing it according to rules and parameters stored in its domain knowledge. The designer is informed about relevant issues in a number of ways: lists of devices to substitute into a design are restricted by the system to compatible choices, limited design rationale is displayed with the option of linking to further details, and technical terms are defined with hypertext links. In addition to the construction area, there are LAN tools, such as an automated IP address generator, and utilities for reporting on physically existing LAN configurations. When a design is completed, a bill-of-materials can be printed out and an HTML page can be produced for display on the Internet. NetSuite is a knowledgeable, well constructed system to support an individual LAN designer.

2.3 The Need to Go Further

Based on our understanding of organizational learning and our investigation of LAN design communities, we believe that in a domain like LAN management no closed system will suffice. The domain knowledge required to go beyond the functionality of NetSuite is too open-ended, too constantly changing, and too dependent upon local circumstances. The next generation of commercial DODEs will have to support extensibility by end-users and collaboration within communities of practice. While a system like NetSuite has its place in helping to design complex networks from scratch, most work of LAN managers involves extending existing networks, debugging breakdowns in service, and planning for future technologies.

Many LAN management organizations rely on home-grown information systems because they believe that critical parts of their local information are unique. A community of practice has its own ways of doing things. Generally, these local practices are understood tacitly and are propagated through apprenticeship (Lave & Wenger, 1991) . This causes problems when the old-timer who set things up is gone and when a newcomer does not know who to ask or even what to ask. A community memory is needed that captures local knowledge when it is generated (e.g., when a device is configured) and delivers knowledge when needed (when there is a problem with that device) without being explicitly queried.

The burden of entering all this information in the system must be distributed among the people doing the work and must be supported computationally to minimize the effort required. This means:

1.     The DODE knowledge base should be integrated with work practices in ways that capture knowledge as it is created.

2.     The benefits of maintaining the knowledge base have to be clearly experienced by participants.

3.     There may need to be an accepted distribution of roles related to the functioning of the organizational memory.

4.     The software environment must be thoroughly interactive so that users can easily enter data and comments.

5.     The information base should be seeded with basic domain knowledge so that users do not have to enter everything and so that the system is useful from the start.

6.     As the information space grows, there should be ways for people to restructure it so that its organization and functionality keep pace with its evolving contents and uses (Fischer et al., 1999) .

DODEs must be extended in these ways to support communities of practice, not just isolated designers. This reflects a shift of emphasis from technical domain knowledge to local socially-based community knowledge.

3. Supporting Communities of Practice

In this Section, we briefly define “community of practice” – a level of analysis increasingly important within discussions of computer-supported cooperative work (CSCW) – and suggest that these communities need group memories to carry on their work. The notion of DODEs must be extended to support the collaborative learning that needs to take place within these communities. A scenario demonstrates how a CIE prototype named WebNet can do this.

3.1 Community Memories

3.1.1 Communities of Practice

All work within a division of labor is social (Marx, 1867/1976) . The job that one person performs is also performed similarly by others and relies upon vast social networks. That is, work is defined by social practices that are propagated through socialization, apprenticeship, training, schooling, and culture (Bourdieu, 1972/1995; Giddens, 1984; Lave & Wenger, 1991) , as well as by explicit standards. Often, work is performed by collaborating teams that form communities of practice within or across organizations (Brown & Duguid, 1991) . These communities evolve their own styles of communication and expression, or genres (Bakhtin, 1986; Yates & Orlikowski, 1992) .

For instance, interviews we conducted showed that computer network managers at our university work in concert. They need to share information about what they have done and how it is done with other team members and with other LAN managers elsewhere. For such a community, information about their own situation and local terminology may be even more important than generic domain knowledge (Orr, 1990) . Support for LAN managers must provide memory about how individual local devices have been configured as well as offer domain knowledge about standards, protocols, compatibilities, and naming conventions.

Communities of practice can be co-located within an organization (e.g., at our university) or across a discipline (e.g., all managers of university networks). Before the World Wide Web existed, most computer support for communities of practice targeted individuals with desktop applications. The knowledge in the systems was mostly static domain knowledge. With intranets and dynamic Web sites, it is now possible to support distributed communities and also to maintain interactive and evolving information about local circumstances and group history. Communities of practice need to be able to maintain their own memories. (The problem of adoption of organizational memory technologies by specific communities involves complex social issues beyond the scope of this paper. For a review of common issues and positive and negative examples of responses, see (Grudin, 1990; Orlikowski, 1992; Orlikowski et al., 1995) .)

3.1.2 Digital Memories for Communities of Practice

Human and social evolution can be viewed as the successive development of increasingly effective forms of memory for learning, storing, and sharing knowledge. Biological evolution gave us episodic, mimetic, and mythical memory; then cultural evolution provided oral and written — external and shared ¾ memory; finally modern technological evolution generates digital (computer-based) and global (Internet-based) memories (Donald, 1991; Norman, 1993) .

At each stage, the development of hardware capabilities must be followed by the definition and adoption of appropriate skills and practices before the potential of the new information technology can begin to be realized. External memories, incorporating symbolic representations, facilitated the growth of complex societies and sophisticated scientific understandings. Their effectiveness relied upon the spread of literacy and industrialization. Similarly, while the proliferation of networked computers ushers in the possibility of capturing new knowledge as it is produced within work groups and delivering relevant information on demand, the achievement of this potential requires the careful design of information systems, software interfaces, and work practices. New computer-based organizational memories must be matched with new social structures that produce and reproduce patterns of organizational learning (Giddens, 1984; Lave & Wenger, 1991) .

Community memories are to communities of practice what human memories are to individuals. They make use of explicit, external, symbolic representations that allow for shared understanding within a community. They make organizational learning possible within the group (Ackerman & McDonald, 1996; Argyris & Schön, 1978; Borghoff & Parechi, 1998; Buckingham Shum & Hammond, 1994; Senge, 1990) .

3.1.3 Integrative Systems for Community Memory

Effective community memory relies on integration. Tools for representing design artifacts and other work tasks must be related to rich repositories of information that can be brought to bear when needed. Communication about artifacts under development should be tied to the artifact so they retain their context of significance and their association with each other. Also, members of the community of practice must be integrated with each other in ways that allow something one member learned in the past to be delivered to other members when they need it in the future. One model for such integration – on an individual level – is the human brain, which stores a wealth of memories over a lifetime of experience, thought, and learning in a highly inter-related associative network that permits effective recall based on subjective relevance. This – and not the traditional model of computer memory as an array of independent bits of objective information – is the model that must be extended to community memories.

Of course, we want to implement community memories using computer memory. Perhaps the most important goal is integration in order to allow the definition of associations and other inter-relationships. For instance, in a system like those to be discussed in Section 3 using perspectives, it is necessary for all information to be uniformly structured with indications of perspective and linking relationships. A traditional way to integrate information in a computer system is with a relational database. This allows associations to be established among arbitrary data. It also provides mechanisms like SQL queries to retrieve information based on specifications in a rather comprehensive language. Integrating all the information of a design environment in a unified database makes it possible to build bridges from the current task representation to any other information. Certainly, object-oriented or hybrid databases and distributed systems that integrate data on multiple computers can provide the same advantages. Nor does an underlying query language like SQL have to be exposed to users; front-end interfaces can be much more graphical and domain-oriented (Buckingham Shum, 1998) .

Communities themselves must also be integrated. The Web provides a convenient technology for integrating the members of a community of practice, even if they are physically dispersed or do not share a homogeneous computer platform. In particular, intranets are Web sites designed for communication within a specific community rather than world-wide. WebNet, for instance, is intranet-based software that we prototyped for LAN management communities. It includes a variety of communication media as well as community memory repositories and collaborative productivity tools. It will be discussed later in this Section.

Dynamic Web pages can be interactive in the sense that they accept user inputs through selection buttons and text entry forms. Unlike most forms on the Web that only provide information (like product orders, customer preferences, or user demographics) to the webmaster, intranet feedback may be made immediately available to the user community that generated it. For instance, the WebNet scenario below includes an interactive glossary. When someone modifies a glossary definition the new definition is displayed to anyone looking at the glossary. Community members can readily comment on the definitions or change them. The history of the changes and comments made by the community is shared by the group. In this way, intranet technology can be used to build systems that are CIEs in which community members deposit knowledge as they acquire it so that other members can learn when they need to or want to, and can communicate about it. This illustrates computer support for collaborative learning with digital memories belonging to communities of practice.

3.2 Extending the DODE Approach to CIEs for Design

To provide computer support for collaborative learning with CIEs, we first have to understand the process of collaborative learning. Based on this analysis, we can see how to extend the basic characteristics of a DODE to create a CIE.

3.2.1 The Process of Collaborative Learning

The ability of designers to proceed based on their tacit existing expertise (Polanyi, 1962) periodically breaks down and they have to rebuild their understanding of the situation through explicit reflection (Schön, 1983) . This reflective stage can be helped if they have good community support and effective computer support to bring relevant new information to bear on their problem. When they have comprehended the problem and incorporated the new understanding in their personal memories, we say they have learned. The process of design typically follows this cycle of breakdown and reinterpretation in learning (see Figure 2, cycle on left) (Stahl, 1993a) .

When design tasks take place in a collaborative context, the reflection results in articulation of solutions in language or in other symbolic representations. The articulated new knowledge can be shared within the community of practice. Such knowledge, created by the community, can be used in future situations to help a member overcome a breakdown in understanding. This cycle of collaboration is called organizational learning (see Figure 2, upper cycle). The personal reflection and collaborative articulation of shared perspectives makes innovation possible (Boland & Tenkasi, 1995; Tomasello et al., 1993) .

Organizational learning can be supported by computer-based systems of organizational memory if the articulated knowledge is captured in a digital symbolic representation. The information must be stored and organized in a format that facilitates its subsequent identification and retrieval. In order to provide computer support, the software must be able to recognize breakdown situations when particular items of stored information might be useful to human reflection (see Figure 2, lower cycle) (Stahl, 1993b) . DODEs provide computer support for design by individuals. They need to be extended to collaborative information environments (CIEs) to support organizational learning in communities of practice.

3.2.2 Extending the DODE Approach to CIEs for Design

The key to active computer support that goes significantly beyond printed external memories is to have the system deliver the right information at the right time in the right way (Fischer et al., 1993/1998) . To do this, the software must be able to analyze the state of the work being undertaken, identify likely breakdowns, locate relevant information, and deliver that information in a timely manner.

Systems like NetSuite and our older prototypes used critics based on domain knowledge to deliver information relevant to the current state of a design artifact being constructed in the design environment work space (see Figure 3, left).

One can generalize from the critiquing approach of these DODEs to arrive at an overall architecture for organizational memories. The core difference between a DODE and a CIE is that a DODE focuses on delivering domain knowledge, conceived of as relatively static and universal, while a CIE is built around forms of community memory, treated as constantly evolving and largely specific to a particular community of practice. Where DODEs relied heavily on a set of critic rules predefined as part of the domain knowledge, CIEs generalize the function of the critiquing mechanisms.

In a CIE, it is still necessary to maintain some representation of the task as a basis for the software to take action. This task representation plays the role of the design artifact in a DODE, triggering critics and generally defining the work context in order to decide what is relevant. This is most naturally accomplished if work is done within the software environment. For instance, if communication about designs takes place within the system where the design is constructed, then annotations and email messages can be linked directly to the design elements they discuss. This reduces problems of deixis (comments referring to “that” object “over there”). It also allows related items to be linked together automatically. In a rich information space there may be many relationships of interest between new work artifacts and items in the organizational memory. For instance, when a LAN manager debugs a network, links between network diagrams, topology designs, LAN diary entries, device tables, and an interactive glossary of local terminology can be browsed to discover relevant information.

The general problem for a CIE is to define analysis mechanisms that can bridge from the task representation to relevant community memory information items to support learning on demand (see Figure 3, right).

To take a very different example, suppose you are writing a paper within a software environment that includes a digital library of papers written by you and your colleagues. Then an analysis mechanism to support your learning might compare sentences or paragraphs in your draft (which functions as a task representation) to text from other papers and from email discussions (the community memory) to find excerpts of potential interest to deliver for your learning. We use latent semantic analysis (Landauer & Dumais, 1997) to mine our email repository (Lindstaedt & Schneider, 1997) and are exploring similar uses of this mechanism to link task representations to textual information to support organizational learning. Other retrieval mechanisms might be appropriate for mining catalogs of software agents or components, design elements, and other sorts of organizational memories.

Using our example of LAN design, we next show how a CIE might function in this domain. We present a scenario of use of WebNet, a prototype we developed to extend our DODE concept to explicitly support communities of LAN designers.

3.3 WebNet: Scenario of a CIE for Design

3.3.1 Critiquing and Information Delivery

Kay is a graduate student who works part-time to maintain her department’s LAN. The department has a budget to extend its network and has asked Kay to come up with a design. Kay brings up WebNet in her Web browser. She opens up the design of her department’s current LAN in the LAN Design Environment, an Agentsheets (Repenning, 1994) simulation applet. Kay starts to add a new subnet. Noticing that there is no icon for an Iris graphics workstation in her palette, Kay selects the WebNet menu item for the Simulations Repository Web page (see Figure 4, left frame). This opens a Web site that contains simulation agents that other Agentsheets users have programmed. WebNet opens the repository to display agents that are appropriate for WebNet simulations. Kay locates a simulation agent that someone else has created with the behavior of an Iris workstation. She adds this to her palette and to her design.

When Kay runs the LAN simulation, WebNet proactively inserts a router (see Figure 4, upper right), and informs Kay that a router is needed at the intersection of the two subnets. WebNet displays some basic information about routers and suggests several Web sites with details about different routers from commercial vendors (see Figure 4, lower right). Here, WebNet has signaled a breakdown in Kay’s designing and provided easy access to sources of information for her to learn what she needs to know on demand. This information includes generic domain knowledge like definitions of technical terms, current equipment details like costs, and community memory from related historical emails.

WebNet points to several email messages from Kay’s colleagues that discuss router issues and how they have been handled locally. The Email Archive includes all emails sent to Kay’s LAN management workgroup in the past. Relevant emails are retrieved and ordered by the Email Archive software (Lindstaedt, 1996) based on their semantic relatedness to a query. In Kay’s situation, WebNet automatically generates a query describing the simulation context, particularly the need for a router. The repository can also be browsed, using a hierarchy of categories developed by the user community.

Kay reviews the email to find out which routers are preferred by her colleagues. Then she looks up the latest specs, options, and costs on the Web pages of router suppliers. Kay adds the router she wants to the simulation and re-runs the simulation to check it. She saves her new design in a catalog of local LAN layouts. Then she sends an email message to her co-workers telling them to take a look at the new design in WebNet’s catalog. She also asks Jay, her mentor at Network Services, to check her work.

3.3.2 Interactive and Evolving Knowledge

Jay studies Kay’s design in his Web browser. He realizes that the Iris computer that Kay has added is powerful enough to perform the routing function itself. He knows that this knowledge has to be added to the simulation in order to make this option obvious to novices like Kay when they work in the simulation. Agentsheets includes an end-user programming language that allows Jay to reprogram the Iris workstation agent (Repenning, 1994) . To see how other people have programmed similar functionality, Jay finds a server agent in the Simulations Repository and looks at its program. He adapts it to modify the behavior of the Iris agent and stores this agent back in the repository. Then he redefines the router critic rule in the simulation. He also sends Kay an email describing the advantages of doing the routing in software on the Iris; WebNet may make this email available to people in situations like Kay’s in the future.

When he is finished, Jay tests his changes by going through the process that Kay followed. This time, the definition of router supplied by WebNet catches his eye. He realizes that this definition could also include knowledge about the option of performing routing in workstation software. The definitions that WebNet provides are stored in an interactive glossary. Jay goes to the WebNet glossary entry for “router” and clicks on the “Edit Definition” button. He adds a sentence to the existing definition, noting that routing can sometimes be performed by server software. He saves this definition and then clicks on “Make Annotations”. This lets him add a comment suggesting that readers look at the simulation he has just modified for an example of software routing. Other community members may add their own comments, expressing their views of the pros and cons of this approach. Any glossary user can quickly review the history of definitions and comments – as well as contribute their own thoughts.

3.3.3 Community Memory

It is now two years later. Kay has graduated and been replaced by Bea. The subnet that Kay had added crashed last night due to print queue problems. Bea uses the LAN Management Info component of WebNet to trace back through a series of email trouble reports and entries in LAN diaries. The LAN Management Information component of WebNet consists of four integrated information sources: a Trouble Queue of reported problems, a Host Table listing device configurations, a LAN Diary detailing chronological modifications to the LAN and a Technical Glossary defining local hardware names and aliases. These four sources are accessed through a common interface that provides for interactivity and linking of related items.

The particular problem that Bea is working on was submitted to her through the Trouble Queue. Bea starts her investigation with the Host Table, reviewing how the printer, routers, and servers have been configured. This information includes links to LAN Diary entries dating back to Kay’s work and providing the rationale for how decisions were made by the various people who managed the LAN. Bea also searches the Trouble Queue for incidents involving the print queue and related device configurations. Many of the relevant entries in the four sources are linked together, providing paths to guide Bea on an insightful path through the community history. After successfully debugging the problem using the community memory stored in WebNet, Bea documents the solution by making entries and new cross links in the LAN Management Information sources: the Trouble Queue, Host Table, LAN Diary, and Glossary.

In this scenario, Kay, Jay, and Bea have used WebNet as a design, communication, and memory system to support both their immediate tasks and the future work of their community. Knowledge has been constructed by people working on their own, but within a community context. Their knowledge has been integrated within a multi-component community memory, that provides support for further knowledge building. This scenario – in which simulations, various repositories, electronic diaries, communication media and other utilities are integrated with work processes – suggests how complexly integrated CIEs can support communities of practice.

4. Perspectives on Shared, Evolving Knowledge Construction

In this Section we propose a mechanism designed to make a CIE like WebNet more effective in supporting the interactions between individuals and groups in communities of practice. We call this mechanism “perspectives”. The perspectives mechanism allows a shared repository of knowledge to be structured in ways that allow for both individual work and the negotiation of shared results. To illustrate this approach to collaboration, we describe a CIE called WebGuide, which is an example of computer supported collaborative learning (CSCL) (Crook, 1994; Koschmann, 1996; O'Malley, 1995) .

4.1 Perspectives: A Collaboration Support Mechanism

The concept of perspectives comes from the hermeneutic philosophy of interpretation of Heidegger and Gadamer (Gadamer, 1960/1988; Heidegger, 1927/1996) . According to this philosophy, all understanding is situated within interpretive perspectives: knowledge is fundamentally perspectival. This is in accord with recent work in cognitive science that argues for theories of socially situated activity (Lave & Wenger, 1991; Winograd & Flores, 1986) . These theories extend the hermeneutic approach to take into account the role of social structures in contributing to molding the construction of knowledge (Vygotsky, 1930/1978) . Communities of practice play an important role in the social construction of knowledge (Brown & Duguid, 1991) .

Knowledge here is the interpretation of information as meaningful within the context of personal and/or group perspectives. Such interpretation by individuals is typically an automatic and tacit process that people are not aware of (Polanyi, 1962; Stahl, 1993b) . It is generally supported by cultural habits (Bourdieu, 1972/1995) and partakes of processes of social structuration (Giddens, 1984) . This tacit and subjective personal opinion evolves into shared knowledge primarily through communication and argumentation within groups (Habermas, 1981/1984) .

Collaborative work typically involves both individual and group activities. Individuals engage in personal perspective-making and also collaborate in perspective-taking (Boland & Tenkasi, 1995) . That is, individuals construct not only elements of domain knowledge, but also their own “take” on the domain, a way of understanding the network of knowledge that makes up the domain. An essential aspect of making one’s perspective on a domain of knowledge is to take on the perspectives of other people in the community. Learning to interpret the world through someone else’s eyes and then adopting this view as part of one's own intellectual repertoire is a fundamental mechanism of learning. Collaborative learning can be viewed as a dialectic between these two processes of perspective making and perspective taking. This interaction takes place at both the individual and group levels of analysis – and it is a primary mode of interchange between the two levels.

While the Web provides an obvious medium for collaborative work, it provides no support for the interplay of individual and group understanding that drives collaboration. First, we need ways to find and work with information that matches our personal needs, interests, and capabilities. Then we need means for bringing our individual knowledge together to build a shared understanding and collaborative products. Enhancing the Web with perspectives may be an effective way to accomplish this.

As a mechanism for computer-based information systems, the term perspective means that a particular, restricted segment of an information repository is being considered, stored, categorized, and annotated. This segment consists of the information that is relevant to a particular person or group, possibly personalized in its display or organization to the needs and interests of that individual or team (Stahl, 1995) . Computer support for perspectives allows people in a group to interact with a shared community memory; everyone views and maintains their own perspective on the information without interfering with content displayed in the perspectives of other group members.

One problem that typically arises is that isolated perspectives of group members tend to diverge instead of converging as work proceeds. Structuring perspectives to encourage perspective-taking, sharing, and negotiation offers a solution to this by allowing members of a group to communicate about what information to include as mutually acceptable. The problem with negotiation is generally that it delays work on information while potentially lengthy negotiations are underway. Here, a careful structuring of perspectives provides a solution, allowing work to continue within personal perspectives while the contents of shared perspectives are being negotiated. We believe that perspectives structured for negotiation is an important approach that can provide powerful support for collaborative use of large information spaces on the Web.

The idea of perspectives traces its lineage to hypertext ideas like "trail blazing" (Bush, 1945) , "transclusion" (Nelson, 1981) , and "virtual copies" (Mittal et al., 1986) – techniques for defining and sharing alternative views on large hypermedia spaces. At the University of Colorado, we have been building desktop applications with perspectives for the past decade (McCall et al., 1990; Stahl, 1995; Stahl et al., 1995) and are now starting to use perspectives on the Web.

Earlier versions of the perspectives mechanism defined different contexts associated with items of information. For instance, in an architectural DODE information about electrical systems could be grouped in an "electrical context" or "electrician's perspective." In a CIE, this mechanism is used to support collaboration by defining personal and group perspectives in which collaborating individuals can develop their own ideas and negotiate shared positions. These informational contexts can come to represent perspectives on knowledge. While some collaboration support systems provide personal and/or group workspaces (e.g., (Scardamalia & Bereiter, 1996) ), the perspectives implementation described below is innovative in supporting hierarchies of perspective inheritance.

The most important characteristics of the perspective mechanism (Stahl, 1993a) that we have been exploring are:

·       Individual community members have access to what appears to be their own information source. This is called their personal perspective. It consists of items from a shared central information repository that are tagged as being visible within that particular perspective (or in any perspective inherited by that perspective).

·       Community member A can integrate an item from B’s perspective into A’s personal perspective by creating a link or virtual copy of the item. If B modifies the original item, then it changes in A’s perspective as well. However, if A modifies the item, a new item is actually created for A, so that B’s perspective is not changed. This arrangement generally makes sense because A wants to view (or inherit) B’s item, even if it evolves. However, B should not be affected by the actions of someone who copied one of B’s items.

·       Alternatively, A can physically copy the contents of an item from B’s perspective. In this case, the copies are not linked to each other in any way. Since A and B are viewing physically distinct items now, either can make changes without affecting the other’s perspective.

·       When A creates a virtual copy of an item from B’s perspective, A can decide if she will also get virtual copies of items related to that one, or if she will create her own sub-network for her copy of that item. Arbitrarily large sub-networks of information can be inherited with no overhead using the virtual copy mechanism.

·       Items of information can be created, edited, rearranged, linked together, or deleted by users within their personal perspective without affecting the work of others.

·       New perspectives can be created by users. Perspectives can inherit from existing perspectives. Thus, a team perspective can be created that includes virtual copies of all contents of the inherited perspectives of the team members.

·       There is an inheritance tree of perspectives; descendants inherit the contents of their ancestor perspectives. Changes (additions, edits, deletions) in the ancestor are seen in descendent perspectives, but not vice versa.

·       A hierarchy of team, sub-team, and individual perspectives can be built to match the needs of a particular community.

This model of perspectives has the important advantage of letting team members inherit the content of their team’s perspective and other information sources without having to generate it from scratch. They can then experiment with this content on their own without worrying about affecting what others see. This is advantageous as long as one only wants to use someone else’s information to develop one’s own perspective. It has frequently been noted in computer science literature (Boland & Tenkasi, 1995; Floyd, 1992) that different stakeholders engaged in the development and use of a system (e.g., designers, testers, marketing, management, end-users) always think about and judge issues from different perspectives and that these differences must be taken into account.

However, if one wants to influence the content of team members’ perspectives, then this approach is limited because one cannot change someone else’s content directly. It is of course important for supporting collaborative work that the perspectives maintain at least a partial overlap of their contents in order to reach successful mutual understanding and coordination. The underlying subjective opinions must be intertwined to establish intersubjective understanding (Habermas, 1981/1984; Tomasello et al., 1993)