Shared Understanding: Implications for Computer Supported Cooperative Work

qualifying exam paper by

William T. Hunt

Advisor: Marilyn Mantei


This paper discusses shared understanding, a phenomenon involving both people and artifacts. The research on shared understanding is found under various names including: common ground, socially shared cognition, and distributed cognition. This paper will examine the meaning of shared understanding, review relevant research from the computer, cognitive, and social sciences, and suggest some research directions within the computer supported cooperative work area. Understanding in common and distributed understanding are prominently represented. Conversational approaches and the distributed cognition approach, (involving representations and trajectories of knowledge within people and artifacts) are emphasized. Possible situations for analysis, questions to address, and methodologies of choice are discussed.

    1. Abstract
    2. 1 Introduction
      1. 1.1 Cognition Between People
      2. 1.2 Cognition In the World
      3. 1.3 Shared Understanding in Computer Supported Cooperative Work
        1. 1.3.1 People to People
        2. 1.3.2 Computer is Medium
        3. 1.3.3 CSCW
      4. 1.4 Organization of Paper
    3. 2 Explication of Shared Understanding
      1. 2.1 Shared
      2. 2.2 Understanding
        1. 2.2.1 Meaning of Meaning
      3. 2.3 Shared Understanding
    4. 3 Review: Understanding in Common
      1. 3.1 Conversational Approaches
        1. 3.1.1 Grounding in Communication
          1. Collaborative Effort
          2. Grounding Changes with Purpose
            1. Grounding References
            2. Verbatim Content
          3. Grounding Changes with Medium
            1. Constraints on Grounding
            2. Costs of Grounding
        2. 3.1.2 Conversation Analysis
        3. 3.1.3 Computational Linguistics
          1. Heeman
          2. Traum
        4. 3.1.4 Critique of Conversational Approaches
      2. 3.2 Social Science Approaches
      3. 3.3 Computer Systems Approaches
        1. 3.3.1 Communication
          1. Tartar, Foster, et al.
          2. Whittaker, Brennan, et al.
          3. Winograd
          4. Scardamalia
        2. 3.3.2 Culture
          1. Eveland and Bikson
          2. Sproull and Kiesler
    5. 4 Review: Distributed Understanding
      1. 4.1 Distributed Cognition
        1. 4.1.1 Explication of Distributed Cognition
        2. 4.1.2 Studies
          1. Ship Navigation
          2. Perfective Software Maintenance
          3. Airline Crews
          4. Simulation
        3. 4.1.3 Critique of Distributed Cognition
      2. 4.2 Social Science Approaches
        1. 4.2.1 Star
        2. 4.2.2 Hantano & Inagaki
        3. 4.2.3 Rogoff
      3. 4.3 Distributed Artificial Intelligence
    6. 5 Research Directions
      1. 5.1 Situations and Questions
      2. 5.2 Methodologies and Unit of Analysis
    7. References

1 Introduction

Culture, constituted of tools, artifacts, and ways of thought, is what carries the past history of a society into the present, thereby both enabling and constraining current thinking. -- Lauren B. Resnick [1991, p. 18]

1.1 Cognition Between People

A prominent view of cognition and understanding is that they are processes occurring in someone's head or mind. This is especially true among American and British psychologists beginning in the 1950's when cognition studied by doing experiments on single unaided individuals in university laboratories [Norman, 1991, p. 17-19]. There was no emphasis on group activities, normal daily situations, or naturalistic observations.

However, this was not the case in the Soviet Union in the early 1900's where Lev Vygotsky's social-historical school, now known as activity theory, began [Rogoff & Wertsch, 1984, p. 1-6]. Vygotsky postulated that mental functioning occurs first between people in social interaction and later within the child's mind. Similar schools also arose in Scandinavia and Germany, under the banner of activity theory, action theory, and situated action.

Today, in many social and computer science disciplines, practitioners are examining the world from a situated activity perspective.

1.2 Cognition In the World

Besides being shared with others, cognition is also in the artifacts in the world. Gregory Bateson [1972] exemplified it well through a thought experiment:

Suppose I am a blind man, and I use a stick. I go tap, tap, tap. Where do I start? Is my mental system bounded at the handle of the stick? Is it bounded by my skin? Does it start halfway up the stick? Does it start at the tip of the stick? [p. 459]

Suppose the blind man eats his dinner. Then forks, knives, and spoons become relevant. As Michael Cole [1991] and Arthur Bentley [1954] argue, the mind cannot be bound by skin.

Don Norman [1991; 1992, p.19-25] talks of affordances. These are properties of objects. A table affords placing objects on it. A baseball affords throwing but does not afford sitting on. Spoons afford eating soup. So the blind man's understanding of the world is intricately tied to the affordance of his stick. We can talk of these properties of artifacts that extend our minds as affordances.

Cognition and understanding are also in the world as culture. Cole recounts experiments in which adults bounce infants dressed in blue diapers but treat infants in pink diapers gently. Cole also discusses the comments of parents upon first finding out about their baby's gender such as "I shall be worried to death when she's eighteen" and "It can't play rugby." (Interestingly, girls' rugby is Canada's fastest growing secondary school sport.) Here the cultural expectations of the parents are already putting constraints on the child's future experience of the world.

1.3 Shared Understanding in Computer Supported Cooperative Work

1.3.1 People to People

One definition of Computer Supported Cooperative Work (CSCW) is computer-assisted coordinated activity carried out by collaborating individuals [Baecker, 1992]. Typical technology includes electronic mail, conferencing systems, and electronic meeting rooms. CSCW arose as an interdisciplinary field, originally from telecommunications and computer science but CSCW now includes aspects from many disciplines of both the computer sciences (human-computer interaction, office systems, distributed systems, artificial intelligence), the social sciences (anthropology, group psychology, social psychology, organizational and management science, linguistics, cognitive science), and other studies (media, technology, and education).

In CSCW there is a focus on supporting person-to-person activities, a shift from the man-machine approach in other computer-related disciplines. This makes the study of shared understanding a natural endeavour within CSCW. Within the literature, however, issues of shared understanding are usually coupled with other issues , e.g., [Nunamaker, Dennis, et al., 1991; Tartar, Foster, et al., 1991].

1.3.2 Computer is Medium

Experience with technology teaches us that once a technology makes something possible, it gets applied, whether for good or bad. -- [Norman, 1992, p.14]

Knowledge in the head is subject to the limits posed by memory and attention; knowledge in the world reminds people. -- [ibid., p.165]

Don Norman [1988; 1991] studies the effects of artifacts, such things as lighting switch arrangements for rooms, stove knobs for burners, and floppy disk design. He talks of artifacts as changing the user's task [Norman, 1991]. Marshall and Eric McLuhan [1988] studied the effects of media, and, after 12 years of testing, provide four guiding questions to ask of any media or human artifact:

* What does it enhance or intensify?

* What does it render obsolete or displace?

* What does it retrieve that was previously obsolesced?

* What does it produce or become when pressed to an extreme?

Norman and the McLuhans come together on their understanding of the products of humans. Norman talks of artifacts as media and the McLuhans talk of media as artifacts. These two perspectives provide a common metaphor for computer systems: the computer is medium This knowledge in the world fits well with the multimedia aspect of CSCW.

1.3.3 CSCW

Two factors for examining CSCW work are time and space [Ellis, Gibbs, et al., 1991]. For instance, face-to-face meetings occur in the same space at the same time, whereas electronic mail involves people in different places sending and receiving messages asynchronously. Because small groups function differently from large ones, size becomes another factor [Nunamaker, Dennis, et al., 1991]. Other considerations for studying CSCW are the tasks performed and the technology used [McGrath, 1984; 1991]. For instance, a company may wish to support impromptu contacts between people in different offices [Mantei, Baecker, et al., 1991] or to provide a shared drawing surface for collaborators [Ishii & Kobayashi, 1992].

During actual work, the relevant factors involved in shared understanding can change. Consider the scenario of two people collaborating on a workshop The collaborators may brainstorm together on the same whiteboard but later work on separate parts. They may use a tool collaboratively to write the workshop notes, but not always at the same time.

1.4 Organization of Paper

The paper is divided into five major sections. This introduction described and motivated the study of shared understanding and its relevance to computer supported cooperative work. The next section provides an explication of shared understanding. The third and fourth sections describe research, the third on understanding in common and the fourth on distributed understanding. The last section provides some research directions.

2 Explication of Shared Understanding

What is the meaning of shared understanding? A common sense notion is that shared understanding involves two or more people being agreeing on ideas or appropriate actions. This could take place during a conversation or represent some cultural norms.

2.1 Shared

Using the Webster's Unabridged Dictionary, we find two major variations for the verb share, to possess in common and to distribute:

share v.t. 1 to divide in portions; to distribute in shares

2 to partake or enjoy with others; to seize and possess jointly or in common

i.t. 1 to have or take a share or part; to participate

If we look at the word share in Roget's Thesaurus [1987/1852], we find share in nine categories ("n" for noun, "v" for verb):

part (n), be one of (v), sharp edge (n), mete out (v), title deed (n), participation (n), give (n), apportion (v), feel (v).

2.2 Understanding

Similarly, the word understanding is in 13 Thesaurus categories:

agreement (relation), agreement (intellect), consensus knowledge, intelligence, imagination, concord, pacification, compact, feeling, friendliness, love, pity

For each of the words share and understanding, many of the distinct meanings are related. This phenomenon of a word having related meanings (which can be formed into a category) is called polysemy (not to be confused with homonymy, similarly spelt but different meaning words such as river bank and savings bank). George Lakoff [1987] demonstrated the relatedness of such meanings through case studies of the words: anger, over, and there.

As a first polysemic look, we could consider the good (concord, pacification, compact), the true (relation, agreement , consensus knowledge, intelligence, imagination), and the beautiful (feeling, friendliness, love, pity) meanings of understanding. This paper uses a number of these meanings of understanding.

2.2.1 Meaning of Meaning

Many people in many disciplines have something to say about meaning, e.g., Frege [1960] and Putnam [1975] from philosophy, de Saussure [Selden, 1989] and Kristeva [1986] from literary theory. There is a truth-conditional model [Barwise, 1988; Halpern, 1986; Halpern & Moses, 1990; Larson, 1990], a meaning triangle model [Ogden & Richards, 1946/23; Regoczei & Hirst, 1990], and a representational system [Newell, 1981; Norman, 1991]. Katherine Nelson [1985, p. 249], a developmental psychologist, formulates the meaning of meaning as meaning for:

* one's communicative partners (reference)

* oneself (denotation)

* the community at large (sense).

Along with Putnam's insistence that meaning encompass both people and the world, Nelson's work is particularly appropriate for shared understanding research in three ways:

* the literature often discusses communicative, individual, and cultural qualities of shared meaning

* Nelson discusses how each aspect of shared meaning is acquired

* Nelson discusses how each is studied and by which disciplines

2.3 Shared Understanding

In looking at shared cognition, Cole [ibid.] has taken a considered look at the common and distributed meanings of the word share. He finds that often both meanings are found in the same situation, e.g., "they shared the taxi" and "they shared the batch of cookies." He further raises the questions, "How can shared cognition be both within and between individuals?" and "Where is distributed cognition located?"

The next two sections of this paper will review work related to these two notions of to possess in common and to distribute with further discussions along communicative or cultural lines as appropriate.

3 Review: Understanding in Common

In the study of shared understanding as having or putting in common, this paper will look at it from three general prospectives, a conversational approach, a social science approach, and a computer systems approach. The conversational approach will focus on issues of grounding communication, the social science approach will focus on cultural issues, and the computer systems approach will examine understanding within multiuser systems.

3.1 Conversational Approaches

3.1.1 Grounding in Communication

Conversation requires much shared information to be successful, i.e., mutual knowledge, beliefs, and assumptions. This information is called common ground and conversational partners are constantly coordinating with each other to ground the content of their conversation [Clark & Brennan, 1991]. Grounding is the process of seeking and providing evidence of understanding in conversation.

The message model of conversation [Reddy, 1979] says that communication is a conduit, where the medium is neutral and the hearer is passive. This message view claims that people listen to words, decode them, and interpret the words against a common ground. From this viewpoint, one who overhears a conversation should understand it as well as the participants.

But Michael Schober and Herbert Clark [1989] conducted experiments showing that active participants have significantly better recall of the content of a conversation than do passive observers. They put forth the collaborative view claiming that conversational partners understand one another through moment by moment collaboration.

Since understanding cannot be perfect, partners must try to meet some grounding criterion to establish that they mutually understand one another for current purposes [Clark & Wilkes-Gibbs, 1986]. Grounding is the process of trying to reach this mutual belief.

Herbert Clark and E. F. Schaefer [1989] proposed the contribution model. In this model, a contribution to a conversation must have a presentation phase and an acceptance phase. Consider the partners Alan and Barbara. In the presentation phase, Alan presents utterance u to Barbara. He does so on the assumption that, if Barbara gives evidence e or stronger, he can believe that she understands what he means by u. In the acceptance phase, Barbara accepts utterance u by giving evidence e that she believes she understands what Alan means by u. She does so on the assumption that, once Alan registers that evidence, he will also believe that she understands. E.g.:

Presentation phase> Alan: Now, - um do you and your husband have a j-car

Acceptance phase > Barbara: - have a car?

Alan: Yeah

Barbara: No

Much of the contribution model rests on the methodology of conversational analysis (CA). This originated in a school of anthropology called ethnomethodology that studies everyday person-to-person interactions [Sacks, Schegloff, et al., 1974]. Conversation analysis seeks to explain patterns in conversational exchanges at the level of negotiations.

The contribution model says that people seek positive evidence for understanding. Three common forms of seeking positive evidence are acknowledgments, relevant next turn, and continued attention. Acknowledgments are frequently sounds or gestures rather that written words, e.g., uh huh, yeah, (head nod). These are called back-channel responses.

A second form of evidence is relevant next turn. Consider the exchange:

Alice: Did you know that mother had been drinking?

Bob: I don't think mother has been drinking at all.

Since the answer is relevant to the question, it is evidence that Alice has been understood.

A third form of positive evidence is continued attention. Undivided attention by a partner is evidence for understanding. Puzzled looks or shifting gaze suggest loss of common ground.

Collaborative Effort
In 1975 Paul Grice [1989] proposed some conversational maxims based on minimizing individual effort. Grice's four conversational efficiency rules are:

Quantity: Be as informative as is required but not more so

Quality: Say what is true, not that for which you lack sufficient evidence

Relation: Be relevant

Manner: Be easy to understand, not obscure, ambiguous, verbose, disconnected

But there are problems with these maxims. They do not allow for grounding and seem to assume flawless presentations and trouble-free acceptances. Some major problems with this principle arise from time pressures, errors, and ignorance [Clark & Wilkes-Gibbs, 1986].

Speakers appear to limit the time and effort they are willing to spend in planning and producing an utterance, for instance:

A: That tree has, uh, uh

B: Tentworms.

A: Yeah.

B: Yeah.

Speakers often make errors or improper utterances when they could easily prevent them by taking more time and effort. For instance Alan fumbled in the earlier conversation:

Alan: Now, - um do you and your husband have a j-car

Sometimes the speaker doesn't know enough to formulate a proper utterance. For instance, Herbert Clark and Deana Wilkes-Gibbs [1986] conducted experiments where partners were to put cards with novel patterns in a specific order but could only communicate verbally to one another. A typical utterance was: Um, the next one's the person ice skating that has two arms?, a question, inviting participation with the partner.

Clark & Wilkes-Gibbs proposed the

principle of least collaborative effort: in conversation, the participants try to minimize their collaborative effort-- the work that both do from the initiation of each contribution to its mutual acceptance.

This principle helps account for some phenomena in CA analysis such as speakers preferring both to make and to initiate their own repairs, i.e., interlocutors create extra turns when involved in speech repairs. Susan Brennan [1990] conducted several experiments that further elaborated this principle. She suggests five refinements [pp. 106-109]:

* Adjust the grounding criterion to meet the current purpose, taking into account the strength of evidence

* Fleasibly divide the labor

* Use the information available in common ground both for presenting and accepting utterances

* Provide evidence as often as you can afford to

* If your partner does not accept your presentation, either revise your hypothesis about what she understands or seek evidence for your hypothesis

Grounding Changes with Purpose
Conversationalists generally try to establish a collective purpose [Grice, 1989]. This can be to plan an event, to solve a problem, to gossip, to get acquainted. The method of grounding changes as purposes change. Different classes of purposes require different techniques for grounding, for instance, reference (e.g., identifying a landmark) and verbatim content (e.g., giving a telephone number).
Grounding References
It is sometimes necessary to establish the mutual belief that all conversationalists have identified the correct referent, i.e., establish referential identity. Four techniques are commonly used:

Alternative Descriptions: Referring expressions are used to refer to objects, e.g.,
(note: `*' are speech overlaps, `( )' are comments, `(( ))' are inserts, `-' are pauses)

A: Well, that young gentleman from - ((the park)).

B: Joe Joe Wright you mean? - - *(- - laughs)*

A: *yes, (laughs) yes*

B: ((God)), I thought it was old Joe Wright who((`d)) walked in at first

Indicative Gestures: Pointing, looking, or touching often ground referents, e.g.,

B: How would you describe the color of this flower?

S: You mean this one (pointing)?

B: Yes.

S: It's off yellow.

Referential Installments: A referent may need to be identified first before an entire utterance is understood, e.g., providing assembly instructions. Consider:

S: Take the spout--the little one that looks like the end of an oil can -

A: Okay.

S: And put that on the opening in the other large tube. With the round top.

If S said, "Put the spout, on the opening...," it could not have been understood without grounding the spout.

Trial References: Grounding process are sometimes initiated in mid-utterance, e.g.,

A: So I wrote off to . Bill, . uh who ((had)) presumably disappeared by this time, certainly, a man called Annegra?

B: Yeah. Allegra.

A: Allegra, uh replied, . uh and I . put . two other people

Verbatim Content
Sometimes the exact content must be confirmed, such as a telephone number. Three common grounding techniques for verbatim content are: verbatim displays, installments, and spelling. Often the second speaker just repeats the content in verbal displays or uses spelling. When there is much information, speakers often make displays in installments.
Grounding Changes with Medium
The effort and tasks required to ground a conversation change with different media. Giving directions is quite different when talking over the telephone than it is when talking face-to-face. The principle of least collaborative effort predicts that the medium that affords the least collaborative effort will be the one used for the purpose at hand.

There is a great variety in media choice as well: answering machines, fax machines, email, keyboard teleconferencing, video teleconferencing. The matter becomes further complicated when one treats cognitive artifacts (that change our tasks [Norman, 1991]) as media. This is particularly apparent with collaborative systems and tools such as electronic meeting rooms, shared whiteboards, shared writing or drawing tools, etc.

In approaching the this issue, Herbert Clark and Susan Brennan [ibid.] discuss 8 media constraints and 11 media costs important for grounding and the relationships between the two factors.

Constraints on Grounding
     Constraint                          Description                     
copresence             A and B share the same physical environment       
visibility             A and B are visible to one another                
audibility             A and B communicate by speaking                   
cotemporality          B receives at roughly the same time as A          
simultaneity           A and B can send and receive at once and          
sequentiality          A's and B's turns cannot get out of sequence      
reviewability          B can review A's messages                         
revisability           A can revise messages from B                      

Table 1: Constraints on Grounding

The 8 constraints along with a brief definition are in Table 1. Each medium has a different collection of constraints associated with it. For instance, telephone has audibility, cotemporality, simultaneity, and sequentiality constraints while electronic mail has reviewability and revisability constraints.

Costs of Grounding
      Cost                     Description              
formulation        formulate and reformulate            
production         producing the utterance              
reception          receiving a message                  
understanding      understanding a message              
start-up           starting a new discourse             
delay              planning and revising before         
asynchrony         timing of discourse exchanges        
speaker change     changing speakers                    
display            presenting an object of the          
fault              producing a mistake                  
repair             repairing a mistake                  

Table 2: Costs of Grounding

The 11 costs with brief descriptions are in Table 2. These are not independent of one another. The speaker pays formulation and production costs while the addressee pays reception and understanding costs. Both pay for the others. For each cost, certain constraints are more important than others. For instance, display costs are usually higher in media without copresence. If the medium has a high degree of telepresence (having the sense that someone is present) then the cost is reduced.

Costs have tradeoffs. For instance, Philip Cohen [1984] reported the following same-purpose conversations, one over the telephone and the other by keyboard:


S: Uh, now there's a little plastic blue cap

J: Yep.

S: Put that on the top hole in the cylinder you just worked with.


K: Next, take the blue car that has the pink thing on it and screw it to the blue piece you just screwed on.

On the telephone, referential installments were used, but not by keyboard. Cohen claims that speakers attempt to achieve more detailed goals than do the keyboard users. This can be explained using the principle of least effort and grounding costs. Using the telephone, production and repair costs are cheap, so having several conversational turns is acceptable. Using the keyboard, the display cost is cheap compared to the repair cost, so displaying more information in one turn is cheaper.

Different media are more desirable for different purposes. Face-to-face is preferable to email for negotiations, while email is preferred for coordinating schedules [Finholt, Sproull, et al., 1990].

The analysis of grounding in CSCW systems will provide insight into its workings and benefit its design.

3.1.2 Conversation Analysis

Conversation analysis (CA) is an empirical, descriptive method used by sociologists and anthropologists to study human conversation [Schegloff, 1991, p.152]. Several distinctions from CA that are generally useful for shared understanding research are: turn-taking, adjacency pairs, restarts, repairs, silences, and closings.

Of particular interest is third person repair. Typically, the first speaker recognizes a misunderstanding by the second speaker in the second turn and hence corrects it in the third turn. An example with memory trouble is:

M: What' Stuart have to say.

N: Didn' I tell you?

M: No:

N: He's coming

M: Oh that's righ*t*

N: *Y*eah he's coming he's coming.

M: Oh that's right, *he's* going to Berkeley

N: *Yes.*

M says no; she's forgotten the shared knowledge of Stuart coming to Berkeley. She repairs it two positions later.

Repair with unit trouble is more complex and involves problems with the discourse unit itself. Schegloff [1991] discusses both repairs in detail.

3.1.3 Computational Linguistics

Two computational models of shared understanding are discussed, one for referring expressions, the other using acknowledgments and repairs to ground the intentions behind utterances. Importantly, each model combines a plan-based approach (computational) with insight from conversation analysis (providing analysis from everyday communicative settings).
Peter Heeman [1991] develops a computational model of how conversational participants make referring actions successful. The model uses the planning paradigm, viewing language as a goal-directed behaviour. What is interesting in this model is that it builds on the Clark and Wilkes-Gibbs proposal of referring action. This model states that after a referring expression (or refashioned referring expression) is presented, the hearer either accepts it, rejects it, or postpones the judgment. If the expression is not accepted, one participant replaces or expands it. The process continues until the final referring expression is mutually accepted.

Clark and Wilkes-Gibbs do not account for why the expression is accepted, rejected, or postponed, but Heeman proposes a model:

* if exactly one object matches the expression, the expression is accepted

* if no objects match the expression, the expression is rejected

* if many objects match the expression, the judgment is postponed

* if rejected, the expression (or part of it) needs to be replaced

* if postponed, the expression needs to be expanded or part of it needs to be replaced

* if by plan recognition the hearer can choose a plausible candidate, then the hearer can make the change, otherwise the speaker must make the choice

Furthermore, in refashioning the expression, Heeman draws on Clark and Wilkes-Gibbs principle of least collaborative effort. Through his goal-oriented approach, Heeman makes the current purposes of collaboration computational. Here is an example dialogue with the system (S):

I: See the weird creature.

S: In the corner?

A: No, on the television.

S: Okay.

David Traum [1991] develops a computational model for how conversants reach a state of shared understanding of the speaker's intention. Traum draws on CA to create discourse units and invoke both acknowledgments and repairs as part of a plan-based system called TRAINS-90. Applying John Searle's [1969] speech act theory, Traum implements acts, including turn-taking acts and grounding acts to establish shared understanding. Here is an example dialogue with the system (S):

M: system, why don't we uhh take uhh engine E-two

and go get tanker T-one

and bring it back to city D

S: okay

M: and why don't we . use engine E-three .. to uhh

go to city I to get..get boxcar B-eight,

go tm city B to get tanker T-two

go to city B to get tanker B-seven

S: sorry, those are boxcars, you mean

M: aaah I'm sorry, yes

I wanna get boxcar seven and eight and tanker T-two

S: okay

and tanker T-two at B

M: yes

S: yes

3.1.4 Critique of Conversational Approaches

In general, the grounding approach has a lot to offer in the analysis of shared understanding. However, it is not clear how local references might build into some integrated or more abstract understanding. Nor do cultural factors seem to find a natural place in the approach. At a finer grain, constraints such as visibility may be too coarse, i.e., significant differences in grounding ability may occur in being able to see someone's: room, body, hands, face, eyes. Furthermore, there is a need to examine how constraints and simple costs combine, perhaps in complex ways.

The descriptive approach of CA analysts makes it subject to investigator bias with little theory building beyond a single conversation. Graeme Hirst [1991] addresses the questions: "Is CA too rich?" (requiring more flexibility than possible from systems) and "Is CA too poor?" (ignoring users' intents). He shows that ascribing intent is necessary for such notions as repair, closing, or rules as resources. The use of CA as a framework for studying the communication of meaning and intention between agents is appropriate for CSCW.

Computational linguistic approaches may have too narrow an application for direct use in CSCW but may be appropriate in the future.

3.2 Social Science Approaches

Several researchers in various social science disciplines are looking at socially shared cognition. Several of the people and ideas are briefly discussed below.

Lauren Resnick [1991] discusses thinking as social practice in her introduction to the definitive book [Resnick, Levine, et al., 1991] from the social sciences on shared understanding. This book is a collection of papers from investigators with diverse disciplinary backgrounds: anthropology, sociology, linguistics, and various schools of psychology. She sets the tone by looking at cognition as being pervaded by the social.

John Levine and Richard Moreland [1991] have written an excellent introduction to culturally shared understanding in work groups. Briefly, culture as shared thoughts has roots in the group as a whole, in its individual members, and in the work. Culture as shared customs has several aspects: its jargon, its special symbols, its accounts (stories), its everyday routines, and its rituals. Membership into a group goes through phases with associated roles and transitions, as shown in Table 3.

  Membership Phases      Role Transitions      Roles      
                        socialization       newcomers     
                        maintenance         oldtimers     

Table 3: Individual Phases, Transitions, and Roles within a Workgroup

Robert Krauss and Susan Fussell [1991] carry out several carefully designed experiments to study how speakers craft their communications based on hypotheses about the expected listener. There are two general sources of these hypotheses. The first is prior beliefs and expectations about others. The other is the dynamics of interaction. The experiments on expectations confirmed the hypothesis that the way people describe things and formulate messages depends upon the social category of the recipient. The experiments on interaction dynamics looked at the evolution of referring expressions, and listener response effects. The value of this work is in providing experimental evidence for various social effects on shared understanding.

Reid Hastie and Nancy Pennington [1991] looked at jury decision-making, in mock criminal trial juries where consensus was required. They found that communication was a mix of facts, values, and efforts to persuade. Factions formed within the juries making it necessary to think in terms of interacting social subsystems made up of cognitively interacting individuals. The individuals were evidence-driven but the groups were sometimes verdict-driven, leading to decisions without complete convergence of interpretation. The shared understanding was on the good but not always on the true or the beautiful. Several of the polysemic meanings of understanding are present, and the situation of understanding in common also contained distributed understanding.

Perret-Clermont, Perret, et al. [1991] provide an interesting shift in attitude and approach between a first and second set of experiments. When they observed that social circumstances affected answers and strategies which in turn affected personal cognitive development, they shifted their unit of analysis from the individual's cognitive behaviour to the social interaction itself.

Jean Lave [1991] argues not to think of the process of socially shared cognition as ending in the internalization of knowledge by individuals, but as a process of becoming a member of a community of sustained practice.

James Wertsch [1991] presents the sociocultural approach in which the unit of analysis is the social language. The approach calls on the researcher to:

* identify and characterize particular social language

* specify how social languages reflect as well as create particular sociocultural settings; corrective to strictly psychological

* examine processes whereby appropriating various social languages affects intrapsychological functioning; mastering a social language is a process whereby psychological processes are formed

William Damon [1991] warns against losing sight of the individual, asking, "What do individuals bring to a setting?" and "What do they take away, in a lasting sense?"

3.3 Computer Systems Approaches

Within the CSCW literature, issues of shared understanding are often coupled with issues of communication or issues of culture. Below are some examples of research in which shared understanding has been addressed.

3.3.1 Communication

Tartar, Foster, et al.
Deborah Tartar, Gregg Foster, et al. [1991] examined the communication shortcomings in Cognoter software used in the Colab. The Colab project was an attempt to provide computer support for small group meetings [Stefik, 1986; Stefik, Bobrow, et al., 1987; Stefik, Foster, et al., 1987]. The CoLab room consisted of a large video screen (called Liveboard) and three computer workstations whose components were a screen, a keyboard, a mouse, and a video switch. The three workstations and Liveboard were connected in an Ethernet network. Any screen could be viewed on the Liveboard by flipping the video switch. Workstation screens could be projected onto the Liveboard or any other workstation through the network.

Cognoter was the most developed software for the lab. It was designed to aid small groups in forming plans and outlines. A three part process of brainstorming, idea organization, and idea evaluation was explicitly implemented.

The Cognoter software had two types of windows: item organization windows and edit windows. The item organization windows were shared spaces, i.e., an identical window appeared on every screen and each user could change its contents. This what-you-see-is-what-I-see (WYSIWIS) window is discussed at length in Stefik, Bobrow, et al. To create an item, a user would open an edit window, name it, type in it, and close it. The software then placed the item as an icon in an item organization window with the name showing. Once there, anyone could move the item, open it, or edit it.

The users had unexpected communication breakdowns which led to redesigning the software [Tartar, Foster, et al., 1991]. A major reason for the problems with understanding one another was the parcel post communication model used to build Cognoter. Based on this model, people send, receive, and open parcels in any order.

Experiments were done with two outside groups performing two tasks of two hours in length. The two groups were quite different in there use of Cognoter. Members of the first group basically worked alone, avoiding the group software. The second group, after much frustration, eventually used the software as a video (not computational) workspace, i.e., they watched one person type for the group. The worst frustrations were both connected with shared understanding: lack of information visibility and mistaken references. Some of the comments about lack of visibility were:

* "Why can't I see that?"

* "I don't see what use it is to have a big screen if we can't all contribute to it."

* "Click DONE so I can see it."

* P1: P2, do you have anything you want to say?

P2: I won't be able to see it up there, right?"

One account of mistaken reference is [Tartar, Foster, et al., 1991, p. 191]:

* (P2 was using the video network to look at P1's machine. There was a general discussion about creating a new item)

* (P1 hit a mouse button, thereby creating an edit window for entering a new item)

* (P3 suggested that P2 should type instead of P1)

* (P2 went to switch her display back to her own machine)

* (P1 looked at her screen, appeared surprised, typed four characters (the title of the item) and moused the DONE button, thus sending it to the others)

* (P2 found that her display had gone black (into idle) and was confused)

P2: What! Who did that?

(At virtually the same time, the item that P1 had created appeared on P2's and P3's screens)

* P3: P1

* P2: (hitting a space bar which caused the machine to come out of idle)

Oh, it was my fault.

* P1: I did not!

* P3: P1, Let P2 type!

* P1: I am!

* P2: I've forgotten what I was going to type.

These problems led to rethinking about communication and abandoning the parcel post model in favour of an interactive model as discussed in section 3.1.1 Grounding in Communication. The analysis resulted in a new version called Cnoter with shared editing and moving facilities, faster communication between machines, and consistent positioning of windows across machines. The lack of information visibility resulted in shared editing where the edit window automatically opened on all screens and anyone could edit (using a mouse button). Mistaken references were greatly reduced by the faster communications and the consistency of movements and positioning across all screens.

Whittaker, Brennan, et al.
Steve Whittaker, Susan Brennan, et al. [1991] conducted experiments in which separated participants shared an electronic Whiteboard. The researchers videotaped six groups of three collaborators who could type and draw in the shared space. For analysis, input was divided into three categories: ARTIFACTS, complex textual objects such as lists and tables, PROSE; text unconnected to ARTIFACTS; DEIXIS, drawings or objects used to point . Input units (marks or text by the same person made in continuous duration) were classified into temporal (duration and overlap), spatial (located near previous input, near last self input, or elsewhere), and deletions. Furthermore, users created both personal spaces and public spaces.

An overall result was a 41% overlap in information presentation, much greater than ordinary conversation (5%). This was accounted for by the permanence of information on the Whiteboard. More particularly, ARTIFACTS remained on the Whiteboard five times longer than PROSE. A proposed explanation was that ARTIFACTS represent the content of communication, whereas PROSE represents the process, such as clarification or negotiation. This explanation was further supported by a higher percentage of consensus decisions used for deleting ARTIFACTS (versus PROSE). However, given that the two tasks required ARTIFACTS, one, creating a list, and the other, coordinating calendars (i.e., a table), weakens the explanation. Some version of this shared knowledge (one of Star's boundary objects) needs to be around for the duration of the task.

Another result was that there was more ARTIFACT overlap than PROSE overlap and ARTIFACT overlap was more likely to be in personal spaces. The explanation proposed is as ARTIFACT construction does not need close coordination between collaborators whereas PROSE does. PROSE, on the other hand, was more likely to be next to the previous input, supporting the coordination explanation.

Videotape analysis may settle two issues. First, more time was spent in self regions than either previous or elsewhere regions for both PROSE and ARTIFACTS. The explanation could be that the screen is used as a scratch pad rather than for communication. Secondly, ARTIFACTS are bigger than PROSE. How much of the temporal effect (duration and overlap) are just creation time?

In a further experiment a speech channel was added and PROSE all but disappeared. This supports the idea of PROSE as process with the speech channel being used for process. DEIXIS also reduced greatly, exemplified by oral voting.

The authors make the following suggestions:

* provide ARTIFACTS for shared data structures

* provide support for turn-taking depending upon whether sequential communication tasks are performed

* provide switching between serial and parallel modes if necessary

* notice that permanence frees participants from the need for incremental discussion

This study has several implications for shared understanding. It emphasizes the value of boundary objects in group support systems. The relatively low counts for previous spatial placement (lowest for all unit types) and the high counts for near self placement (PROSE and ARTIFACTS) suggests that, for the given tasks and media, the building of shared understanding in one's own space is preferred to using part of a shared space. With 56% of all input being PROSE (which becomes speech in the second experiment), a lot of supportive activity is needed to build joint ARTIFACTS. Speech is a preferred over text for communicative activity. This last point is further supported by experiments by [McCarthy, Miles, et al., 1991].

Terry Winograd provides a very explicit model of conversation as action in his software, the Coordinator [Winograd, 1988]. To initiate a conversation, one must identify the type of action: request, offer, promise, what if, inform, question, note. Each has very specific possible further actions for continuing the conversation. For instance, a request can either be declined, countered, or lead to a promise. This conversation can ultimately be cancelled or declared complete. The point of it all, is to make the intentions of the conversationalists explicit and thus improve interactions.

Bowres and Churcher [1988] argue that this forcing of explicitness gets in the way of conversation rather than improving it. From a shared understanding point of view, this is an attempt at grounding conversation at a higher level. The question is whether it forces attention to an extra layer of understanding on the conversationalists and detracts from the content originally intended for grounding. It is one thing to analyze the conversation from such a perspective; it is quite another to potentially change its course by forcing a structure. Studies could be done with and without Coordinator to determine effects on understanding. Other structured communication approaches are: [Fischer, Lemke, et al., 1991; Lee, 1990; Malone, Grant, et al., 1989; Shepherd, Mayer, et al., 1990].

Marlene Scardamalia and Carl Bereiter [1991] are studying the construction of knowledge by children who use the Computer Supported Intentional Learning Environment known as CSILE. Children take some topic of interest and research it, through books, people, films, or whatever. They then enter information into CSILE in the form of labeled notes and drawings. They can look at everything in CSILE by searching through the labels. Any item can be commented upon, thus creating another item for the knowledge base. By these activities the class builds its own understanding of many topics and the students carry on discussions about each other's notions. The children's understanding changes through this interaction.

3.3.2 Culture

Eveland and Bikson
J. Eveland and Tora Bikson [1988] conducted a year-long field study of two task forces from a large corporation. Each had a mandate to develop a set of recommendations about pre-retirement planning. Both groups consisted of older employees, half in each group were retired and half were pre-retirement age. One group had standard mechanisms such as meetings, phone calls, and surface mail, while the second group had an electronic network as well as the standard mechanisms. This study showed significant changes over time (evaluated every three months) and significant differences among the four subgroups.

Communication increased over time as measured by recognition and contact densities. Significant were increases in recognition (of names) and contact (with group members) over time, greater recognition by the electronic group over the standard group, and greater contact by retirees over employees.

The leadership structure and meeting participation patterns were quite different for the two task forces. Using the five most central people in the four evaluation times as a measure of leadership, 1 of 13 leaders in the standard group was retired. For the electronic group, 9 of 16 leaders were retired. Looking at the last time period, most of the face-to-face meetings of the standard task force were unscheduled (usually at the corporation) with only 12% of the retirees present. The electronic task force had over twice as many total meetings with the vast majority being scheduled. Seventy-five percent of the retirees attended these scheduled meetings.

The researchers also addressed the question of what contact means for electronic messages. They distinguished between receiving any messages, completed message loops (message and reply), and completed point-to-point message loops (i.e., no group mailings).

Although the retirees had low participation without the electronic network, they became the most involved subgroup with email, continuing beyond the project. The interplay of initial retired culture and electronic media resulted in a new culture and a new shared understanding.

Sproull and Kiesler
Les Sproull and Sarah Kiesler [1991] talk about experiences with email within corporations. They concluded that distribution lists, bulletin boards, and conferences were windows on the corporation and they provided a better understanding of the culture. A second effect was that email provided a voice for the voiceless by allowing peripheral people to speak and flattened the existing hierarchical flow of information.

4 Review: Distributed Understanding

In the study of shared understanding in its distributed sense, there are three chief sources of research coming from the overlapping areas of distributed cognition, socially shared cognition, and distributed artificial intelligence. Distributed cognition focusses on the interaction of people and artifacts, socially shared cognition focusses on social and cultural relationships including developmental issues, and distributed artificial intelligence focusses on the interactions of (usually) artificial agents with each other.

4.1 Distributed Cognition

4.1.1 Explication of Distributed Cognition

Distributed Cognition is a new endeavour that seeks to understand intelligence at a systems level [Flor & Hutchins, 1991]. It purports to do this by studying:

* the representation of knowledge both inside the heads of individuals and in the world

* the propagation of knowledge among different individuals and artifacts

* the transformations which external structures undergo when operated on by individuals and artifacts

The external representations, which are most closely examined, are goal required external structures created by agents in the system and those structures used to support these goal relevant structures. Such structures can include drawings, typed or written statements, utterances, gestures, and physical models.

An assumption of this method is that the system studied is a goal-driven complex cognitive system. This implies that a goal is attempted by transforming external structures. If a goal is achieved, the structures are each in a goal state.

Typically, the method of study involves charting the goal required external structures and examining the other supporting external structures. The resulting descriptions can then be used to:

* infer what task-relevant thoughts people had (or internal states of agents)

* help improve the system (artifacts and actors)

* characterize system properties.

New devices are usually intended to enhance the accomplishment of a task. When viewing the whole system (user and artifact), the artifact appears to expand some functional capacity of the task performer. However, from the user's point of view, the task is often changed [Norman, 1991]. The artifact has replaced the original task with a different task, one that may have radically different cognitive capacities than those for the original task. Consider the problem of determining the speed of a ship given two fix positions taken 3 minutes apart and the distance travelled in the interim being 1500 yards. Edwin Hutchins [1990] suggests four different conditions, each requiring a different task:

* pencil, paper, algebra, conversion knowledge, formula: D = RT

* pocket calculator, algebra, conversion knowledge, formula: D = RT

* nautical slide rule, knowledge to operate it

* knowledge of the three-minute rule

The basic issue in developing artifacts is the mapping between the representing world (artificial) and represented world (actual). These artificial devices are called cognitive artifacts and are "designed to maintain, display, or operate upon information to serve a representational function." Two qualities of artifacts usually get confounded in most physical devices, but often get separated in software. These are control and representation.

One type of artifact that combines both of these qualities is the object symbol. A scroll bar is a good example because it is controlled by changing its representation. Object symbols may facilitate understanding.

4.1.2 Studies

Ship Navigation
Edwin Hutchins [1990] studied the process of navigating a ship in restricted waters. He looked at "...the members of the navigation team together with the team's social organization and the tools of the navigation trade as a system of socially distributed computations." [p. 194] The six team members included two bearing takers (BT), a bearing time recorder (BTR), a plotter (PLOT), a keeper of the deck log (DL), and a fathometer operator (F).

The goal required was to ascertain the ship's position following the fix cycle, which is:

* the F takes a reading of the depth beneath the ship

* the BTs on either side of the ship take bearings (directions) using landmarks viewed through a telescopic sighting device called an alidade

* the BTR talks to the BTs to get the bearings and writes them in the bearing log

* the PLOT plots the bearings reported on the chart and projects the ship's positions for the next few fixes

* the DL records all events of consequence including commands

* the PLOT and BTR decide on the next set of landmarks

* the BTR relays landmarks to the BTs

From the distribution cognition methodology, the goal required external structures are the bearings (and ship's position). The trajectory of the bearing representation is as follows:

* an image of the bearing in degree marks is seen by the BT through the alidade

* an utterance of the bearing is made by the BT over the telephone circuit to the BTR

* a written recording is make by the BTR into the bearing log

* a plotted line is made by the PLOT onto the chart

Plans are often circumvented in specific situations [Suchman, 1987]. The PLOT, an oldtimer in the group, may go help a new BT find a landmark. Furthermore, the open layout of the ship, may allow the BTR, PLOT, and DL to help and to learn from each other.

For the team, goal supporting external structures are conversations, gestures, recordings, and extra sightings. Internal structures (thoughts, images) have to be determined by analysis of the external structures (or direct knowledge as a participant/observer). Thus we see how shared understanding across a navigation team may be studied.

Perfective Software Maintenance
Nick Flor and Edwin Hutchins [1991] analyzed team programming during a perfective software maintenance experiment by using videotapes, transcripts, and keystroke recordings. Two professional programmers were given the task of adding functionality to an existing program. The goal of this study was to characterize some of the system properties important for the successful performance of the task.

In the experiment, several system level properties were characterized:

* The Reuse of System Knowledge * The Sharing of Goals and Plans

* Searching through Larger Spaces of Alternatives * Efficient Communication

* Joint Productions of Ambiguous Plan Segments * Shared memory for old plans

* Divisions of labor and collaborative interaction systems

Prominent are aspects of shared understanding. Several are exemplified in Table 4. LM (who does all the typing) engages RC in finishing her plan segment and he complies. This shows that he understands her plan (lines 8-10). After LM has finished typing, RC takes a turn at getting his plan accepted (11), and this copy-and-modify plan becomes shared (12-16).

The culturally shared knowledge of yank, escape, j, k, p, and y (vi editor jargon) is common ground for the two programmers. This knowledge is needed for efficient communication in lines 11-15. Because LM and RC share the knowledge of where the new procedure whisper belongs, RC can underspecify his command in line 16. However, assuming too much can lead to miscommunication as in lines 17-18 where LM doesn't understand RC's plan to have the player line followed by a get_response line.

line  subject                    activity /key entry                   
 #      /key                                                           
 6    LM        (laughs) Why don't I just change it and you just       
                tell me                                                
      key       case `w':,RET> display_prompt("                        
 7    RC        I got a better idea LM why don't you do a yank 3 and   
 8    LM        enter player?                                          
 9    LM        (laughs)                                               
10    RC        yes                                                    
      key       Enter player>");                                       
11    RC        ok go up, go up. yank. yank yank                       
12    LM        I'm looking for escape.  A yank?                       
      key       <esc>                                                  
      LM        A yank?                                                
13    RC        yeah don't type it all.  Do a three yank yank.         
      key       kkkkk                                                  
14    LM        right there?                                           
15    RC        yeah, three y y                                        
      key       3yy                                                    
16    RC        then go down to where you whisper and do a P           
      key       jjjjjp                                                 
17    RC        and yank player get response                           
18    LM        yank what?                                             
19    RC        get response.                                          
20    LM        oh                                                     

Table 4: Partial Transcript of Perfective Maintenance Session
Airline Crews
Some of Don Norman's [1992] studies of distributed cognition have focused on commercial airplane crews. He found that pilots created some of their own artifacts to better understand the state of affairs in the cockpit. The most common (and effective) cognitive aids were:

* speed bugs: little metal or plastic tabs that pilots move around the outside of the airspeed indicator to help remember critical settings

* crew-provided devices: written notes, coffee cups, and tape

* checklists

Speed bugs change the task of memorizing critical speeds, such as takeoff and landing, to a perceptual task of observing the pointer being above or below the appropriate speed bug. Speed bugs started with grease pencil or tape marks and now are in danger of becoming a burden because instrument designers are starting to clutter the dials with too many bugs.

Airplane crews provide themselves with many devices to remember the myriad of regular and special bits of information and procedures needed for each flight. Notes are often taped at various locations in the cockpit. An empty coffee cup may be placed over a flap handle. These put the knowledge into the world.

Checklists serve multiple functions. The normal checklists serve as checks on action already completed. Abnormal checklists serve more as triggers to remind crews about unfinished activities. They are often associated with problems and emergencies. Checklists also provide a mechanism for shared understanding within the crew. By having one crew member read the list and others doing the appropriate operations or checks, the entire crew knows the state of the aircraft. A final function of checklists is to provide a record of the actions and state of affairs, providing group memory as well as satisfying legal requirements.

Edwin Hutchins [1991] studied how the social organization of distributed cognition affects the cognitive properties of groups. He looked at confirmation bias, i.e., the propensity to affirm prior interpretations and to discount, ignore, or reinterpret counter evidence. Using a connectionist model [Rumelhart, Smolensky, et al., 1986], Hutchins manipulated environmental input and four parameters: who talks to whom, what they talk about, how persuasive they are, and when they communicate. He studied the process of communities arriving at shared versus differing understandings. The general conclusions drawn were:

* there is a tradeoff between allowing for a diversity of interpretation and the ability to reach a decision

* voting does not always produce the same results as would be achieved by further communication

* confirmation bias for group cognition depends strongly upon how the group distributes the tasks of cognition among its members and not on the initial individual biases.

4.1.3 Critique of Distributed Cognition

The major criticism of distributed cognition is its glossing over the internal structures of agents within a system. These are also part of the system's mental state. For instance, long periods of inaction need to be interpreted. The role of participant/observer would be helpful is some situations to determine these states. The goal-oriented interpretation goes some of the way toward this, but their interpretation could incorporate more of the beautiful (affective) aspects of understanding.

4.2 Social Science Approaches

4.2.1 Star

Susan Leigh Star [Gerson & Star, 1986; 1989; 1992] addresses what she calls the due process questions of distributed understanding, i.e., in combining or collecting evidence from different viewpoints, how do you decide that sufficient, reliable, and fair amounts of evidence have been collected? Who, or what, does the reconciling, according to what set of rules? She made three observations based on two studies, one with clinical and basic physiologists and the other with professional and amateur biologists:

* Different groups can cooperate without having good models of each other. They can successfully work together while employing different units of analysis, methods of aggregating data, and different abstractions of data.

* They can cooperate while having different goals, time horizons, and audiences to satisfy.

* These activities are facilitated by creating boundary objects which can be adapted locally to needs and constraints while maintaining a global identity.

The boundary objects observed by Star are: repositories, ideal types, terrain with coincident boundaries, and forms & labels. Table 5 lists the conditions, results, and advantages of the four boundary objects.

    object        conditions           results             advantages        example     
repository      unit of          repositories such     modularity          libraries     
                analysis         as libraries or                           and museum    
                heterogeneous    museums                                                 
ideal type or   differences in   deletion of local     adaptability        early         
platonic        abstraction      contingencies and                         atlases of    
object                           global rules                              the brain     
terrain with    different        common boundaries,    resolution of       state of      
coincident      aggregations     different contents    different goals     California    
boundaries      of data                                                                  
forms and       dispersed        immutable mobiles,    local               standardized  
labels          workers          standard indexes      uncertainties       forms         

Table 5: Boundary Object Characteristics

Repositories are ordered piles of objects indexed in a standardized fashion. Ideal Types (or Platonic Objects) are objects that are abstracted from all relevant domains and thus do not have local details. Two or more Terrain with Coincident Boundaries have the same boundaries but different internal contents. Forms and Labels are standardized forms that contain only information that remains immutable across distances.

4.2.2 Hantano & Inagaki

Giyoo Hatano and Kayoko Inagaki [1991] studied adolescents in a learning environment to examine collective attempts at acquiring knowledge and to determine how much knowledge is shared. The setting was a classroom using the Japanese science education method called Hypothesis-Experiment-Instruction. For each student the procedure involves hearing a question with alternatives, choosing an answer, seeing everyone's answer, discussing answers with others, choosing again, and testing (or reading) answer.

Hatano and Inagaki observed that partisanship formed around positions with the students distributing themselves among the roles of supporters, criticizers, and assessors. For each position, arguments were developed by collaboration. Those assessing arguments would use cognitive factors as well as social factors, such as gaining supporters or silencing the opposition.

The conclusions drawn were:

* collective comprehension often takes place when cognitive and social motivations work in concert. Partisan motivation provides for an effective division of labour.

* many ways of comprehending can only be achieved by constructive interaction but those involved may have different ways of comprehension.

4.2.3 Rogoff

Barbara Rogoff [1991] studied teachers and apprentices in a shared problem solving setting. She concluded that the model of participation changes understanding should replace the model of packaged information gets internalized and changes understanding. This is empirical evidence for preferring the contribution model over the message model in section 3.1 Conversational Approaches.

4.3 Distributed Artificial Intelligence

Distributed artificial intelligence (DAI) researchers are interested in understanding and modeling both action and knowledge in collaborative enterprises. Natural and artificial agents interact without central control or knowledge of the overall system. The understanding within the system is distributed but agents can build common understanding among themselves. Some of the key researchers and some issues they address are described below.

Carl Hewitt [1985; 1986] discusses an open systems perspective that is characterized by:

* Concurrency: need to handle simultaneous influx from many sources

* Asynchrony: in receiving and responding to input

* Decentralized Control: local decisions are made

* Inconsistent Information: coming from different sources

* Arm's length Relationships: incomplete knowledge of other agents

* Continuous Operation: local failures accommodated

Because of the indeterminacy of open systems, reasoning is by due process including commitments, negotiations, and trials of strength.

Les Gasser [1991; 1992; Gasser, Rouquette, et al., 1989] develops the foundations of DAI further in the directions of open systems. He emphasizes multiple actors, multiple contexts, multiple representations, resource limitations, compatibility-based problems, and robustness. Another issue Gasser addresses is action-at-a-distance. This is the communication grounding problem when the two agents are separated in space, time, or semantics.

Eric Werner [1988; 1989; 1991] addresses issues of social goals and cooperative behaviour. He develops a theory of agents incorporating the information and intention of other agents. Werner further provides algorithms for generating cooperation styles with the core procedure attempting to satisfy as many of the goals of the cooperating agents as possible. The two dimensions of cooperation (from cooperation to antagonism) and self-interest (from self-interest to self-destruction) provide a space in which a spectrum of styles is modeled.

Bernardo Huberman [1988; 1990; Huberman & Hogg, 1988] is primarily interested in the overall system behaviour of multi-agent computational ecologies. He examines biological, human, and computational societies, making mathematical models based on only knowledge of individual agents. The emergence or lack of emergence of stable features is of particular importance. Two features for which stability was not found were incomplete knowledge and delayed information.

Phil Cohen & Hector Levesque [1991a; 1991b; Cohen, Levesque, et al., 1990] develop a formal model of joint activity and teamwork. Using empirical evidence [Cohen, 1984], they argue for joint commitments and intentions in task-oriented dialogues. Using a modal language of belief, goal, action, and time, Cohen and Levesque define joint intention out of joint persistent goal and, among other things, explain the confirmation of referential understanding with their model.

Leigh Star [Gerson & Star, 1986; 1989; 1992] discusses distributed understanding through boundary objects (see section 4.2 Social Science Approaches) and provides a dialogue among the activity theory (see section 1.1 Cognition Between People), American symbolic interactionism (a school of sociology) and DAI. She elaborates on the insights each of these communities of practice (including CSCW work) provide as part of a new pluralism and argues for a continuing dialogue.

5 Research Directions

5.1 Situations and Questions

Paraphrasing Liam Bannon and Kjeld Schmidt [1989], CSCW should endeavor to understand the way people and artifacts interact together, with the objective of designing adequate computer-based technologies. Below are research questions framed within existing situations.

One situation is the ongoing preparation for the November 1992 CSCW conference in Toronto. The audio/visual committee, with two co-chairs, is a good candidate for study. Knowledge could be tracked through the various communication media they use. Some questions could be about misunderstanding, losing, or disregarding information in various media, and during specific transitions. Questions could be raised about the conditions in which retrieval of knowledge is easy or difficult. Joint document preparation could open new questions about what facilitates understanding. Cultural factors could be examined through life styles and changes.

Examining the communication between the tutorial chairman and the presenters at the 1993 CHI conference in Amsterdam could be done with similar techniques to those just described.

.A study could be made of children building shared knowledge using CSCILE. Two groups could be studied, one as the system currently stands, one with some new features. Understanding could be tracked, including some joint writing endeavours. Various questions about factors affecting the building of understanding are possible. These include properties of the knowledge such as reviewability and connectability as well as social and cultural factors. One such question could be about the effect of being notified when someone else had reviewed your own work.

The Telepresence Project will be providing computer-mediated multimedia connections to centres in Toronto, Ottawa, Germany, Spain, France, and Italy. A study of the development of shared understanding across diverse cultures could be made. Cultural and time lag questions could be addressed along with the media issues discussed above. As the project develops, specific applications of the Telepresence technology may present themselves for study.

The CAVECAT (Computer Audio Video Enhanced Collaboration And Telepresence) system is a system of enhanced workstations connected by a digital+audio+video network [Mantei, Baecker, et al., 1991]. It is readily available to dozens of people within the Dynamic Graphic Project at the University of Toronto. At least two scenarios for study are possible. The daily activities of selected frequent users could be monitored, both for CAVECAT usage and other media. Analysis of chosen means and success at shared understanding could be made. Second, a number of short-term projects or activities could be similarly analyzed. Additional questions could be about the choosing or lack of choosing CAVECAT for communications.

5.2 Methodologies and Unit of Analysis

The general methodologies would draw primarily on the distributed cognition and communication approaches reviewed above. Collections of communications would be made, including videotapes. Transcripts would be necessary. Tracking knowledge across people and artifacts would be done. A look at acknowledgments, misunderstandings, and repairs in conversation would be common. Inferring plans and goals would also be part of the methodology. Cultural and social considerations would be examined for relevance and effect. The use of artifacts as boundary objects would be noted, especially their structure and connectivity. Shared document creation would receive special attention.

The preferred unit of analysis would be the situated activity, where people are doing tasks meaningful to their own work.


Baecker, Ronald (1992). Readings in Groupware and Computer Supported Cooperative Work: Assisting Human-Human Collaboration. San Mateo, CA: Morgan Kaufmann Publishers, Inc. Forthcoming.

Bannon, Liam J. and Schmidt, Kjeld (1989). CSCW: Four Characters in Search of a Context. Proceedings of the First European Conference on Computer-Supported Cooperative Work (EC-CSCW'89). Slough, UK: Computer Sciences House. pp. 358-372. September. Also in John M. Bowers and Steven D. Benford (eds.) Studies in Computer Supported Cooperative Work: Theory, Practice and Design. 1991. Amsterdam, The Netherlands: North Holland. pp.3-16.

Barwise, Jon (1988). Three Views of Common Knowledge. In Moshe Y. Vardi (ed.) Proceedings of the Second Conference on Theoretical Aspects of Reasoning about Knowledge. New York: Association for Computing Machinery Press. pp. 365-379.

Bateson, Gregory (1972). Form, Substance, and Difference. In Gregory Bateson (ed.) Steps to an Ecology of Mind. New York: Ballantine Books. pp. 448-466.

Bentley, Arther Fisher (1954). The Human Skin: Philosophy's Last Line of Defense. In Bentley Arthur Fisher (ed.) Inquiry into Inquiries: Essays in Social Theory. Boston: The Beacon Press. pp. 195-211.

Bowres, J. and Churcher, J. (1988). Local and Global Structuring of Computer Mediated Communication: Developing Linguistic Perspectives on CSCW in Cosmos. Proceedings of the Conference on Computer-Supported Cooperative Work (CSCW'88). New York: Association for Computing Machinery Press. pp. 125-139. September.

Brennan, Susan Elise (1990). Seeking and Providing Evidence for Mutual Understanding. PhD Thesis. Stanford University, Stanford, CA.

Clark, Herbert H. and Brennan, Susan E. (1991). Grounding in Communication. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 127-149.

Clark, Herbert H. and Schaefer, Edward F. (1989). Contributing to Discourse. In Cognitive Science. 13(2):259-294.

Clark, Herbert H. and Wilkes-Gibbs, Deana (1986). Referring as a Collaborative Process. In Cognition. 22(1):1-39.

Cohen, Philip R. (1984). The Pragmatics of Referring and the Modality of Communication. In Computational Linguistics. 10(2):97-146.

Cohen, Philip R. and Levesque, Hector J. (1991a). Confirmations and Joint Action. Proceedings of the 12th International Joint Conference on Artificial Intelligence. San Mateo, CA: Morgan Kaufmann Publishers, Inc. pp. 951-957. August.

Cohen, Philip R. and Levesque, Hector J. (1991b). Teamwork. In Nous. XXV(4):487-512. Also, Technical Note 504. Artificial Intelligence Center, SRI International Menlo Park, CA.

Cohen, Philip R., Levesque, Hector J., Nunes, Jose H. T., and Oviatt, Sharon L. (1990). Task-Oriented Dialogue as a Joint Activity. Proceedings of the Pacific Rim International Conference on Artificial Intelligence. pp. 203-207. November.

Cole, Michael (1991). Conclusion. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 398-417.

Damon, William (1991). Problems of Direction in Socially Shared Cognition. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 384-397.

Ellis, Clarence A., Gibbs, Simon J., and Rein, Gail L. (1991). Groupware: Some Issues and Experiences. In Communications of the ACM. 34(1):38-58. January.

Eveland, J. D. and Bikson, Tora K. (1988). Work Group Structures and Computer Support: A Field Experiment. In ACM Transactions on Office Information Systems. 6(4):354-379.

Finholt, Tom, Sproull, Lee, and Kiesler, Sara (1990). Communication and Performance in ad hoc Tasks Groups. In Jolene Galegher, Robert E. Kraut, and Carmen Egido (eds.) Intellectual Teamwork: Social and Technological Foundations of Cooperative Work. Hillsdale, NJ: Lawrence Erlbaum Associates, Publishers. pp. 291-325.

Fischer, Gerhard, Lemke, Andreas C., McCall, Raymond, and Morch, Anders I. (1991). Making Argumentation Serve Design. In Human Computer Interaction. 6(3,4):393-419.

Flor, Nick V. and Hutchins, Edwin L. (1991). Analyzing Distributed Cognition in Software Teams: A Case Study of Team Programming During Perfective Software Maintenance. In Jurgen Koenemann-Belliveau, Thomas G. Moher, and Scott P. Robertson (eds.) Empirical Studies of Programmers: Fourth Workshop. Norwood, NJ: Ablex Publishing Corporation. pp. 36-63.

Frege, Gottlob (1960). On Sense and Reference. In Peter Geach and Max Black (eds.) Translations from the Philosophical Writings of Gottlob Frege. 2nd edition Oxford, UK: Basil Blackwell. pp. 56-78.

Gasser, Les (1991). Social Conceptions of Knowledge and Action: DAI Foundations and Open Systems Semantics. In Artificial Intelligence. 47:107-138.

Gasser, Les (1992). Boundaries, Identity, and Aggregation: Plurality Issues in Multiagent Systems. In Y. Demazeau and Eric Werner (eds.) Decentralized Artificial Intelligence. Vol. III. Amsterdam, The Netherlands: Elsevier Science Publishers B. V. To appear.

Gasser, Les, Rouquette, Nicholas F., Hill, Randall W., and Lieb, John (1989). Representing and Using Organizational Knowledge in Distributed AI Systems. In Les Gasser and Michael N. Huhns (eds.) Distributed Artificial Intelligence. Vol. II. San Mateo, CA: Morgan Kaufmann Publishers, Inc. pp. 55-78.

Gerson, Elihu M. and Star, Susan Leigh (1986). Analyzing Due Process in the Workplace. In ACM Transactions on Office Information Systems. 4(3):257-270. July.

Grice, H. Paul (1989). Studies in the Way of Words. Cambridge, MA: Harvard University Press.

Halpern, Joseph Y. (1986). Reasoning About Knowledge: An Overview. Theoretical Aspects of Reasoning about Knowledge: Proceedings of the 1986 Conference. New York: Association for Computing Machinery Press. pp. 1-17.

Halpern, Joseph Y. and Moses, Yoram (1990). Knowledge and Common Knowledge in a Distributed Environment. In Journal of the Association for Computing Machinery. 37(3):549-587. July.

Hastie, Reid and Pennington, Nancy (1991). Cognitive and Social Processes in Decision Making. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 308-327.

Hatano, Giyoo and Inagaki, Kayoko (1991). Sharing Cognition Through Collective Comprehension Activity. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 331-348.

Heeman, Peter Anthony (1991). A Computational Model of Collaboration on Referring Expressions. Masters Thesis. University of Toronto, Toronto, Ontario. Also, Technical Report CSRI-251. Computer Systems Research Institute. University of Toronto, Toronto, Ontario.

Hewitt, Carl (1985). The Challenge of Open Systems. In Byte. 10(4):223-242. April.

Hewitt, Carl (1986). Offices are Open Systems. In ACM Transactions on Office Information Systems. 4(3):271-287. July. Also in Bernardo A. Huberman (ed.) The Ecology of Computation. 1988. Amsterdam, The Netherlands: Elsevier Science Publishers B. V. pp.5-24.

Hirst, Graeme (1991). Does Conversation Analysis have a Role in Computational Linguistics? In Computational Linguistics. 17(2):211-227. June.

Huberman, Bernardo A. (1988). The Ecology of Computations. In Bernardo A. Huberman (ed.) The Ecology of Computations. Amsterdam, The Netherlands: Elsevier Science Publishers B. V. pp. 1-4.

Huberman, Bernardo A. (1990). The Performance of Cooperative Processes. In Physica D. 42:38-47.

Huberman, Bernardo A. and Hogg, Tad (1988). The Behaviour of Computational Ecologies. In Bernardo A. Huberman (ed.) The Ecology of Computations. Amsterdam, The Netherlands: Elsevier Science Publishers B. V. pp. 77-116.

Hutchins, Edwin (1990). The Technology of Team Navigation. In Jolene Galegher, Robert E. Kraut, and Carmen Egido (eds.) Intellectual Teamwork: Social and Technological Foundations of Cooperative Work. Hillsdale, NJ: Lawrence Erlbaum Associates, Publishers. pp. 191-220.

Hutchins, Edwin (1991). The Social Organization of Distributed Cognition. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 283-307.

Ishii, Hiroshi and Kobayashi, Minoru (1992). ClearBoard: A Seamless Medium for Shared Drawing and Conversation with Eye Contact. In Penny Bauersfeld, John Bennett, and Gene Lynch (eds.) ACM Conference on Human Factors in Computing Systems: Reaching Through Technology (CHI'92). Reading, MA: Addison-Wesley Publishing Company, Inc. pp. 525-532. April-May.

Krauss, Robert M. and Fussell, Susan R. (1991). Constructing Shared Communicative Environments. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 173-200.

Kristeva, Julia (1986). The Kristeva Reader. Toril Moi (ed.) New York: Columbia University Press.

Lakoff, George (1987). Women, Fire, and Dangerous Things: What Categories Reveal about the Mind. Chicago: University of Chicago Press.

Larson, Richard K. (1990). Semantics. In Daniel N. Osherson and Howard Lasnik (eds.) An Invitation to Cognitive Science. Vol. 1: Language. Cambridge, MA: The MIT Press. pp. 23-42.

Lave, Jean (1991). Situating Learning in Communities of Practice. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 63-82.

Lee, Jintae (1990). SIBYL: A Tool for Managing Group Decision Rationale. Proceedings of the Conference on Computer-Supported Cooperative Work (CSCW'90). New York: Association for Computing Machinery Press. pp. 79-92. October.

Levine, John M. and Moreland, Richard L. (1991). Culture and Socialization in Work Groups. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 257-279.

Malone, Thomas W., Grant, Kenneth R., Lai, Kum-Yew, Rao, Ramana, and Rosenblitt, David A. (1989). The Information Lens: An Intelligent System for Information Sharing and Coordination. In Margrethe H. Olson (ed.) Technological Support for Work Group Collaboration. Hillsdale, NJ: Lawrence Erlbaum Associates, Publishers. pp. 65-88.

Mantei, Marilyn M., Baecker, Ronald M., Sellen, Abigail J., Buxton, William A. S., Milligan, Thomas, and Wellman, Barry (1991). Experiences in the Use of a Media Space. In Scott P. Robertson, Gary M. Olson, and Judith S. Olson (eds.) ACM Conference on Human Factors in Computing Systems: Reaching Through Technology (CHI'91). Reading, MA: Addison-Wesley Publishing Company, Inc. pp. 203-208.

McCarthy, John C., Miles, Victoria C., and Monk, Andrew F. (1991). An Experimental Study of Common Ground in Text-Based Communication. In Scott P. Robertson, Gary M. Olson, and Judith S. Olson (eds.) ACM Conference on Human Factors in Computing Systems: Reaching Through Technology (CHI'91). Reading, MA: Addison-Wesley Publishing Company, Inc. pp. 209-215.

McGrath, Joe E. (1984). Groups: Interaction and Performance. Englewood Cliffs, NJ: Prentice-Hall, Inc.

McGrath, Joe E. (1991). Time, Interaction, and Performance. In Small Group Research. 22(2):147-174. May.

McLuhan, Marshall and McLuhan, Eric (1988). Laws of Media. Toronto, Ontario, Canada: University of Toronto Press.

Nelson, Katherine (1985). Making Sense: the Acquisition of Shared Meaning. New York: Academic Press.

Newell, Alan (1981). The Knowledge Level. In The AI Magazine. 2:1-20. Also, in Artificial Intelligence, 1982. 87:87-127.

Norman, Donald A. (1988). The Psychology of Everyday Things. New York: Basic Books. Also, The Design of Everyday Things. 1990. New York: Doubleday.

Norman, Donald A. (1991). Cognitive Artifacts. In John M. Carroll (ed.) Designing Interaction: Psychology at the Human-Computer Interface. Cambridge, UK: Cambridge University Press. pp. 17-38.

Norman, Donald A. (1992). Turn Signals are the Facial Expressions of Automobiles. Reading, MA: Addison-Wesley Publishing Company, Inc.

Nunamaker, J. F., Dennis, Alan R., Valacich, Joseph S., Vogel, Douglas R., and George, Joey F. (1991). Electronic Meeting Systems to Support Group Work. In Communications of the ACM. 34(7):40-61. July.

Ogden, Charles K. and Richards, Ivor A. (1946/23). The Meaning of Meaning. New York: Harcort, Brace, and World.

Perret-Clermont, Anne-Nelly, Perret, Jean-Francois, and Bell, Nancy (1991). The Social Construction of Meaning and Cognitive Activity in Elementary School Children. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 41-62.

Putnam, Hilary (1975). The Meaning of "Meaning". In Keith Gunderson (ed.) Language, Mind and Knowledge. Vol. VII. Minneapolis, MN: University of Minnesota Press. pp. 131-193.

Reddy, Michael J. (1979). The Conduit Metaphor- a Case of Frame Conflict in our Language about Language. In Andrew Ortony (ed.) Metaphor and Thought. Cambridge, UK: Cambridge University Press. pp. 284-324.

Regoczei, Stephen and Hirst, Graeme (1990). The Meaning Triangle as a Tool for the Acquisition of Abstract, Conceptual Knowledge. In International Journal of Man-Machine Studies. 33(5):505-520.

Resnick, Lauren B. (1991). Shared Cognition: Thinking as Social Practice. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 1-20.

Resnick, Lauren B., Levine, John M., and Teasley, Stephanie D. (1991). Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association.

Roget, Peter Mark (1987/1852). Roget's Thesaurus. Betty Kirkpatrick (ed.) Essex, UK: Longman Group UK Limited.

Rogoff, Barbara (1991). Social Interaction as Apprenticeship in Thinking: Guidance and Participation in Spatial Planning. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 349-364.

Rogoff, Barbara and Wertsch, V. James (1984). Children's Learning in the `Zone of Proximal Development'. San Francisco, CA: Jossey-Bass Inc., Publishers.

Rumelhart, David E., Smolensky, Paul, McClelland, James L., and Hinton, Geoffrey E. (1986). Schemata and Sequential Thought Processes in PDP Models. In James L. McClelland, David E. Rumelhart, and The PDP Group (eds.) Parallel Distributed Processing: Explorations in the Microstructure of Cognition:. Vol. 2 Psychological and Biological Models. Cambridge, MA: The MIT Press. pp. 7-57.

Sacks, Harvey, Schegloff, Emanuel A., and Jefferson, Gail (1974). A Simplest Systematics for the Organization of Turn-Taking in Conversation. In Language. 50(4):696-735.

Scardamalia, Marlene and Bereiter, Carl (1991). Higher Levels of Agency for Children in Knowledge Building: A Challenge for the Design of New Knowledge Media. In The Journal of the Learning Sciences. 1(1):37-68.

Schegloff, Emanuel A. (1991). Conversation Analysis and Socially Shared Cognition. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 150-171.

Schober, Michael F. and Clark, Herbert H. (1989). Understanding by Addressees and Overhearers. In Cognitive Psychology. 21(2):211-232.

Searle, John (1969). Speech Acts: An Essay in the Philosophy of Language. Cambridge, UK: Cambridge University Press.

Selden, Raman (1989). A Reader's Guide to Contemporary Literary Theory. Lexington, Kentucky: University Press of Kentucky.

Shepherd, Allan, Mayer, Neils, and Kuchinsky, Allan (1990). Strudel-An Extensible Electronic Conversation Toolkit. Proceedings of the Conference on Computer-Supported Cooperative Work (CSCW'90). New York: Association for Computing Machinery Press. pp. 93-104. October.

Sproull, Lee and Kiesler, Sara (1991). Increasing Personal Connections. In Lee Sproull and Sara Kiesler (eds.) Connections: New Ways of Working in the Networked Organization. Cambridge, MA: The MIT Press. pp. 79-101.

Star, Susan Leigh (1989). The Structure of Ill-Structured Solutions: Boundary Objects and Heterogeneous Distributed Problem Solving. In Les Gasser and Michael N. Huhns (eds.) Distributed Artificial Intelligence. Vol. II. San Mateo, CA: Morgan Kaufmann Publishers, Inc. pp. 37-54.

Star, Susan Leigh (1992). Working Together: Symbolic Interactionism, Activity Theory, and Distributed Artificial Intelligence. In Yrjo Engestrom and David Middleton (eds.) Communication and Cognition at Work. To Appear.

Stefik, Mark (1986). The Next Knowledge Medium. In The AI Magazine. VII(1):34-46. spring. Also in Bernardo A. Huberman (ed.) The Ecology of Computation. 1988. Amsterdam, The Netherlands: Elsevier Science Publishers B. V.

Stefik, Mark, Bobrow, Daniel G., Foster, Gregg, Lanning, S., and Tatar, Deborah (1987). WYSIWIS Revised: Early Experiences with Multiuser Interfaces. In ACM Transactions on Office Information Systems. 5(2):147-167. April.

Stefik, Mark, Foster, Gregg, Bobrow, Daniel G., Kahn, K., Lanning, S., and Suchman, Lucy (1987). Beyond the Chalkboard: Computer Support for Collaboration and Problem Solving in Meetings. In Communications of the ACM. 30(1):32-47. January. Also in Irene Greif (ed.) Computer-Supported Cooperative Work: A Book of Readings. 1988. San Mateo, CA: Morgan Kaufmann Publishers, Inc.

Suchman, Lucy (1987). Plans and Situated Action. Cambridge, UK: Cambridge University Press.

Tartar, Deborah G., Foster, Greg, and Bobrow, Daniel G. (1991). Design for Conversation: Lessons from Cognoter. In International Journal of Man-Machine Studies. 34(2):185-209. February. Also in Saul Greenberg (ed.) Computer-supported Cooperative Work and Groupware. 1991. New York: Harcourt Brace Jovanovich, Publishers. pp.55-79.

Traum, David R. (1991). Towards a Computational Theory of Grounding in Natural Language Conversation. Technical Report 401. Computer Science Department, University of Rochester, Rochester, NY. October.

Werner, Eric (1988). Toward a Theory of Communication and Cooperation for Multiagent Planning. In Moshe Y. Vardi (ed.) Proceedings of the Second Conference on Theoretical Aspects of Reasoning about Knowledge. New York: Association for Computing Machinery Press. pp. 129-143. November, 1987.

Werner, Eric (1989). Cooperating Agents: A Unified Theory of Communication and Social Structure. In Les Gasser and Michael N. Huhns (eds.) Distributed Artificial Intelligence. Vol. II. San Mateo, CA: Morgan Kaufmann Publishers, Inc. pp. 3-36.

Werner, Eric (1991). Generating Cooperative and Noncooperative Behavior. unpublished paper (International Symposium on Next Generation Human Interface (Friend21: 91), Institute for Personalized Information Environment, Tokyo, Japan). November.

Wertsch, James V. (1991). A Sociocultural Approach to Socially Shared Cognition. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley (eds.) Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association. pp. 85-100.

Whittaker, Steve, Brennan, Susan E., and Clark, Herbert H. (1991). Co-ordinating Activity: An Analysis of Interaction in Computer-Supported Co-operative Work. In Scott P. Robertson, Gary M. Olson, and Judith S. Olson (eds.) ACM Conference on Human Factors in Computing Systems: Reaching Through Technology (CHI'91). Reading, MA: Addison-Wesley Publishing Company, Inc. pp. 361-367.

Winograd, Terry (1988). Where the Action Is. In Byte. 13(13):256A-258. December.