Toth, 'EFFECTS OF COMBINING INTERACTIVE GRAPHICS AND TEXT IN COMPUTER-MEDIATED SMALL GROUP DECISION-MAKING', Arachnet Electronic Journal on Virtual Culture v2n02 (May 16, 1994) URL = http://hegel.lib.ncsu.edu/stacks/serials/aejvc/aejvc-v2n02-toth-effects The Arachnet Electronic Journal on Virtual Culture __________________________________________________________________ ISSN 1068-5723 May 16, 1994 Volume 2 Issue 2 TOTH V2N2 ======================================================== THE EFFECTS OF COMBINING INTERACTIVE GRAPHICS AND TEXT IN COMPUTER-MEDIATED SMALL GROUP DECISION-MAKING Jozsef A. Toth University of Pittsburgh jtoth+@pitt.edu Abstract Research in computer-mediated small group research has focused primarily on the medium of sentential messages. This paper details exploratory research which instead combines a synchronous sentential messaging medium with two-dimensional interactive graphics. Eleven three-person groups participated in a risk-taking, choice- dilemma task involving the collection of their prediscussion opinions, a discussion of the choice-dilemma, consensus attainment and collection of postdiscussion opinions. Two conditions, one where groups selected and received feedback of their graphics-based prediscussion opinions {g}, and a second, which also included a graphical representation of the prediscussion average {g+avg}, were coupled with a sentential communication medium. A third sentential-only {s} condition served as a control. In the condition with the graphical prediscussion opinions and average {g+avg}, groups sent proportionately more messages making persuasive arguments and proportionately fewer messages proposing values and unsubstantiated opinion about the group decision. In the graphical condition without the average {g}, the reverse effect was found to occur. In the control {s}, the same discussion parameters fell proportionately between the two graphics conditions. In both graphics conditions, the first advocate had a stronger influence on the group decision than in the control. The data suggest that the inclusion of two-dimensional graphics can either facilitate or inhibit normative and informational forms of social influence during the group decision-making process. 1.0 Introduction Investigation into how the use of computers can affect group and organizational behavior has burgeoned in the past decade. By comparing factors between face-to-face (FF) and computer-mediated (CM) small groups, significant computer-mediation effects have been discovered. These effects typically result from synchronous discussion programs or the transmission of asynchronous electronic mail messages. In such cases, the members of CM small groups (comprising 2, 3 or 4 people) have been isolated from each other verbally and nonverbally, so that their only means of communication has been through computer mediation. Significant CM effects include: o the temporal duration of discussion is increased, o the effects on the group decision due to members' high and low status are equalized, o the influence on the group decision of the group member who first advocates a decision proposal is attenuated, o individual participation in the group discussion is equalized, and o the shift between the aggregate opinions held by group members before, during and after discussions are exaggerated. (see Kiesler & Sproull, 1992, for a review) Throughout most of this work, it has been argued that the removal of vital _social context cues_ under CM conditions has either exacerbated or attenuated various social psychological phenomena in these groups. The tasks used in such research have primarily involved decision-making; the broader implications of the findings mentioned above bear ultimately on the everyday and widespread organizational use of computers. For instance, if a geographically disparate group conducts a good part of its communication and decision-making activity through electronic mail, what are the gross implications for the overall efficacy or quality of such a group's decisions -- as compared to an organization whose activities transpire primarily in face-to-face settings? The _modality_ in this computer-mediated research to date has involved only _sentential_ text. The primary focus of this paper will be to extend and refine the notion of computer-mediation so that it also includes two-dimensional interactive computer graphics. There are three primary reasons motivating such an inquiry. First, the wholesale use of low-cost, interactive, graphics- based, networked computer workstations is on the rise. Thus, it is not uncommon for networked group activity to include the integration of graphics and text. Second, and more to the point, if the use of computer-mediated sentential text has been found to have such a wide variety of effects on small-group behavior, what added effects might the combination of interactive text _and_ graphics have on groups? By integrating discourse-based sentential text with two- dimensional interactive computer graphics, this paper will demonstrate that there are indeed significant effects when the two are combined. Finally, if significant differences are to be found among communication modalities that extend beyond sentential text, what are the implications for the designers _and_ users of such systems in the coming years? This exposition will focus primarily on how the differences in communication modalities affect group members' attempts to _influence_ each other in the course of decision-making discussion. 2.0 Theoretical Background It would be highly desirable for the cognitive sciences to have developed comprehensive and integrated theories of attention, perception and cognition, plus associated theories concerning the effective design and use of information displays. At present, however, only fragments exist. In this section, two sub- disciplines will be reviewed in order to frame the ensuing empirical analysis and discussion. These sub-disciplines are (1) cognitive-perceptual efficiency and salience, and (2) small group processes involving influence and decision-making. 2.1 Cognitive-Perceptual Efficiency and Salience One theoretical assumption latent in the research described later acknowledges that there are fundamental differences in the way humans perceive and understand spoken and written language, on the one hand, and how they apprehend non-language visual information, on the other. More specifically, this bears on the differences between the modalities of sentential text and two- dimensional (2D) interactive computer graphics -- and how these differences might be utilized effectively in the design of human- computer interfaces for computer-mediated small groups. The intrinsic properties of these two modalities -- and how they bear on human attention, perception and cognition -- still are widely contested topics. For instance, written (e.g., sentential) and spoken language are presumed to share the same phonological component (Van Orden et al, 1990). However, no one has been able to determined whether sentential information is manipulated internally and represented as a canonical semantic structure -- a structure which also represents information from other non- language modalities (Kaplan & Simon, 1989). This paper does not seek to answer such questions. Rather, it assumes that the semantics of various modalities -- visual, verbal, written, etc. -- must somehow _interact_ within and between various internal modes of processing. For instance, reading the written word _DOG_ not only evokes a phonological activation of the spoken word "DOG", but also elicits dog-related mental images and sounds (e.g., barking, whining) (Paivio, 1986). Moreover, in problems that are _isomorphic_ (Kotovsky et al, 1985; Zhang & Norman, 1990), the manner in which each type of problem is presented to the problem solver yields widely disparate response times, levels of performance and comprehension (Lewis & Toth, 1992; Tabachnek & Simon 1992). A common theme permeating the work in problem isomorphs suggests that language-based representations consume more cognitive resources and are more prone to the effects of working memory (Schneider & Detweiler, 1987). By contrast, when consistently mapped, non-language visual representations result in resource- free and highly automatic modes of processing. Such modes are much more impervious to the effects of working memory, and possess a high degree of cognitive impenetrability. By "non- language", I mean ecologically available visual information that appeals to nomic visual processing and which has enjoyed a genetic advantage over spoken and written language by at least a few hundred million years. By "consistently mapped", I mean the work first detailed by Shiffrin & Schneider (1977) in which the choice and layout of stimuli had an extremely profound effect on whether controlled or automatic attentive and perceptual processing occurred. This paradigm was later extended to involve certain forms of cognitive processing, as well, in a chess context (Fisk & Lloyd, 1988). In the realm of text processing and comprehension, similar findings have been noted. For instance, the comprehension of text has been known to conflict with the reader's inferences and background knowledge, leading to inferences that were not part of the original text (Bransford, Barclay, & Franks, 1972). In other cases, it has been demonstrated that subjects erroneously "remember" aspects or details of the story that were not present in the original text (Loftus, Miller, & Burns, 1978). Regarding the understanding of non-linguistic information, similar results have been obtained in the understanding of novel and permuted stimuli (Mandler & Ritchey, 1977; Nickerson & Adams, 1979). One ramification of this consistently mapped phenomenon is an implication that the stimuli (or salient aspects of the stimuli) inherently possess the potential to be available for effortless, internal processing -- even in _absence_ of the original stimulus (just as is possible in certain forms of imagery). For example, keeping one's eyes closed, sustaining a mental image of the Towers of Hanoi problem, and then solving it, is quite within the realm of such nomic processing. It is argued that such "envisionment" is much more difficult to accomplish for spoken and written language with respect to the _original_ stimuli. In other words, solving a sentential version of the Towers of Hanoi (e.g., "the large disk and the small disk are on the middle peg and the medium disk is on the left peg") would be next to impossible. Unfortunately, what continues to evade cognitive scientists is a lawfully articulated set of principles and guidelines for the organization of such theoretical building-blocks into an ontology that then can be utilized by the applied sciences. For example, it couldn't have been predicted by most theories of the time that the radial-spatial aspects of the analogue fuel, temperature and velocity gauges (rather than the digital-numerical incarnations of the same) would present less of a hinderance to the motorist who must divide attention between the road ahead and the instrument display below. Sometimes, only trial-and-error can be trusted to establish what is effective and what isn't. Relating to the work reported in this paper, the _semantic_ impact of the use of one modality rather than another in terms of working memory limitations is of particular interest, as is how both modalities interact in the course of perception and cognition. If one modality is more resource-bounded (i.e., language) and the other is more resource-free (i.e., non- language), and if the information conveyed is isomorphic (e.g., the written number _9_ versus the _x_ or _y_ position of _9_ on a 2D graph), it will be hypothesized that the resource-free representation will have more of an influence on cognitive processing which takes such information into account. It also will be argued that this effect can persist even after removal of the original stimulus, when such a stimulus entails resource-free processing. 2.2 Influence and Decision-Making in Small Groups When FF and CM small groups engage in a discourse-based decision-making exercise, the typical type of task they perform is of the risk-taking _judgmental_ variety, for which no absolute or correct answer exists. An example of such a problem, known as a _choice-dilemma_, follows. Ms. H, a college senior, has studied the piano since childhood. She has won amateur prizes and given small recitals, suggesting that Ms. H has considerable musical talent. As graduation approaches, Ms. H has the choice of going to medical school to become a physician, a profession which would bring certain prestige and financial rewards; or entering a conservatory of music for advanced training with a well-known pianist. Ms. H realizes that even upon completion of her piano studies, which would take many more years and a lot of money, success as a concert pianist would not be assured. Keep in mind that the riskier alternative, entering a conservatory of music, is always assumed to be more desirable than the safer course, going to medical school. When an experimental group is presented such a problem, the members are asked to discuss the choice dilemma and to arrive at a consensus that reflects their aggregate advice to the central character, Ms. H. This advice is articulated as a single risk value ranging from 1-in-10 to 9-in-10, where 1-in-10 is riskiest and 9-in-10 is safest. This value reflects the lowest probability of success that the group would consider acceptable in order for Ms. H to give the risky alternative a try. Both before the discussion begins and after the discussion ends, each group member is asked to state, in isolation from the rest of the group, what his/her advice would be. Therefore, when the discussion begins, each group member has a preconceived notion of what s/he thinks the _consensus_ value should be. Since the individuals' values seldom are in agreement at the onset of the discussion, participants must adjust their initial positions as the discussion proceeds and the group consensus evolves. Many theories have been advanced explaining the phenomena which transpire in the course of decision-making discussions, and how group members compel each other to change their positions. Two of the most popular theories entail influence and decision rules. There are two varieties of influence: normative and informational (Ridgeway, 1984). _Normative influence_ centers on social norms, and how group members adhere to such norms in order to influence each others' behavior and thoughts. To sway a differing opinion, this type of influence can be as "primal" as brow-beating and shouting, or it can be more implicitly and culturally embedded in social class, gender or accepted social protocol (such as the influence of the majority). _Informational influence_ by contrast, is more "cerebral" in nature, in that it requires discussants to actually employ persuasive arguments that are grounded in fact, logic and inference in order to argue for or against a particular position. Finally, _decision rules_, which may include aspects of normative and informational influence, are analytical heuristics developed by social scientists. They sometimes can predict reliably what the group decision will be, given the group members' prediscussion opinions (Miller, 1989). One such rule is _majority wins_, where if a prediscussion majority exists (e.g., 1-in-10, 2-in-10, 9-in-10), the group decision (e.g., 3-in-10) will tend towards the majority. In this particular decision rule, the coercion of the majority implies a normative style of influence. Another such rule is the _average rule_, which typically occurs when there is an even distribution of opinion (e.g., 1-in-10, 5- in-10, 9-in-10). The group either quickly or as the result of an impasse decides to take the average of their opinions (e.g., 5- in-10). This decision rule is more grounded in the notion of informational influence. 3.0 Hypotheses Tested The principal empirical parameter targeted in this exploratory work involves the opinions that group members hold individually about the central character's dilemma before the discussion begins (e.g., their "advice" to Ms. H). Historically, these prediscussion opinions, as well as the opinions established early in the discussion (realized as decision proposals) are known to have significant effects on the outcome of the discussion, or on group consensus. Varying the modality of the prediscussion opinions should affect subsequent reasoning markedly, in the ensuing discussion phase. The modality being explored may be more resource-bounded (sentential text) or more resource-free (2D graphics). Since a resource-free presentation of the prediscussion opinions involves visual information that is more salient and persistent, it is argued that such information can sustain the course of a ten minute discussion --even in the _absence_ of the original stimulus -- and significantly affect group members' reasoning. Bearing in mind what has been discussed regarding the way group members seek to influence each others' opinions, it can be expected that the salient information captured in the 2D graphical representation invariably will direct the group's reasoning and discussion. Regarding group behavior, this effect is expected to be rendered either in terms of normative or information influence. Two different graphical representations were created in order to test these postulated effects on group discussion and decision- making. The first type was a mapping of the three individual choice-dilemma prediscussion opinions onto a standard two- dimensional graph, whereby risk value corresponded to the spatial _y_ aspect of the graph {g}. The second type was identical in composition to the first, but also included the graphical representation of the prediscussion average {g+avg}. Three hypotheses were established regarding {g} and {g+avg}, with a third sentential-only representation (no graphics) serving as control {s}. The experimental design is summarized in Table 1. Table 1 SUMMARY OF ANTICIPATED FINDINGS ----------------------------------------------------------------- I. Group Consensus Hypotheses (H1, H3) Treatment Conditions (IVs) ---------------------------------------------- No First Advocacy || First Advocacy ---------------------- -------------------- g | g+avg | s || g | g+avg | s ------ --------- ----- ---- --------- ----- Level of high | x | || x | x | Accord ----------------------------- -------------------- (averaging medium x | | || | | effect, ----------------------------- -------------------- DVs) low | | || | | x ----------------------------- -------------------- none | | x || | | II. GROUP DISCUSSION HYPOTHESIS (H2) Treatment Conditions (IVs) ---------------------- g | g+avg | s ------ --------- ----- | | High x | | | ---------------- --------- ----- | Normative | Medium | | x | (opinion) ---------------- --------- ----- | | Low | x | | Level of ----------------------------------------------- Influence | g | g+avg | s (DV) | ------ --------- ----- | | High | x | | Informa- ---------------- --------- ----- | tional | Medium | | x | (persua- ---------------- --------- ----- | sion) | Low x | | ============================================================ g | g+avg | s ------ --------- ----- High | | x ---------------- --------- ----- Length of Discussion (DV) Medium x | | ---------------- --------- ----- Low | x | ============================================================ g | g+avg | s ------ --------- ----- High | x | ---------------- --------- ----- Conciseness of Content (DV) Medium x | | ---------------- --------- ----- Low | | x ----------------------------------------------------------------- Note: "x" indicates value of anticipated finding for treatment indicated H1. Individual and Group Opinion. It was predicted that the value of the group consensus would be most in accord with {g+avg}. The average, as such, would imply an informational, mathematically derived "consensus" which the group would be compelled to utilize. {g} also was expected to have an averaging effect on group consensus, but this effect was not expected to be as great as that for {g+avg}, since the average would be implicit. It would not necessarily be derived during the group discussions. In sum, the two graphics-based displays {g} and {g+avg} were expected to attenuate the _choice shift_ phenomenon (measured as the average of the absolute values of the differences between individual prediscussion opinions and the consensus). It was predicted that {s} would have no average effect; that instead it would induce a choice shift that has been noted in earlier work (Siegel et al, 1986). H2. Group Discussion. Since {g+avg} contained encapsulated, cohesive information about the group (i.e., prediscussion opinions summarized as the average), it was predicted that discussants would voice more informational influence content and less normative influence content than in {g} or {s}. It also was projected that discussion duration would be briefer and discussion content more concise compared to {g} or {s}. This effect was expected because subjects would arrive quickly at the consensus implied by the prediscussion graphical average. Regarding {g}, discussants were expected to use less informational influence and more normative influence, compared to {g+avg} or {s}. This effect was expected because the absence of a group prediscussion average would promote a lack of cohesion in aggregate opinion, instead reinforcing individual opinion exhibited by a more normative influence. Discussion in {g} was predicted to be briefer and more concise than in {s}. Faced with {s}, it was predicted that subjects would spend more time in discussion. This effect was expected because group members would have to articulate their prediscussion opinions sententially, at the onset of the discussion. H3. First Advocacy. Two forms of first advocacy were considered: explicit and implicit. The interpretation of _explicit first advocate_ follows the canonical definition used in computer-mediated and face-to- face research (Weisband, 1992; McGuire et al, 1987). This interpretation specifies that the first person make a specific, numerical decision proposal during the discussion phase (e.g., "I think it should be 1-in-10"). With inclusion of the graphical modality, a second type (hereafter referred to as the _implicit first advocate_), describes an individual who posts his/her opinion graphically but does not explicitly articulate the value of that opinion during the discussion. For example, the individual does not say "I think he should go for it" during the discussion phase, after having graphically selected a value of 1 in 10 in the prediscussion phase). In the case of {s}, the implicit first advocate would be referring to the opinion that s/he selected on the prediscussion questionnaire. It was predicted that in {g} and {g+avg}, the first explicit or implicit advocate would have a stronger influence on the group consensus than in {s}. The source of this influence was postulated to stem from the coupling of the cognitively and perceptually salient visual information with the subsequent reference to that information in the discussion. With {s}, it was predicted that the first explicit or implicit advocate would have the least influence on group consensus, thereby replicating results found in earlier CM research (Weisband, 1992). 4.0 Method 4.1 Design To test these hypotheses, a pilot study was conducted in April 1993. The within-group, repeated-measures experimental design comprised one independent variable (representational modality of group prediscussion opinions), involving three conditions: {s}: Prediscussion opinions were articulated in the sentential modality by the group members as their discussions began. This served as a control condition and an attempt to replicate results from earlier CM research {g}: Computer 2D graphical feedback of the group's prediscussion opinions to group members directly before the discussions began {g+avg}: Computer 2D graphical feedback of the group's prediscussion opinions to group members, including the prediscussion average, directly before the discussions began (see Figure 1, available as PostScript file [filename extension] from the ftp server where this paper was obtained). 4.2 Subjects The subjects were 33 University of Pittsburgh Students -- 18 men, 15 women, of varying ages and academic accomplishment. They were assigned to 11 groups of 3 members each. Five groups comprised two females and one male; five groups consisted of two males and one female; the final group contained three males. Gender was not expected to confound the selection criteria, based on findings from Straus (1991), in which effects of gender in CM groups working on judgmental tasks were not found to be significant. 4.3 Procedure At the beginning of the experimental session, each group met face-to-face with the experimenter. The group members then were seated individually at Sun Microsystems Sparcstations with chromatic monitors. Two subjects were located in the same room, but were separated physically by two office partitions and approximately 20 feet, ensuring verbal and non-verbal isolation. The third subject was seated in an adjacent room and was separated from the other two subjects by a concrete wall, again guaranteeing isolation. All software described below, with the exception of the Internet Relay Chat (IRC) program, was implemented using an X Windows- based graphics package, and was developed in the Unix-based Sun Openwindows graphical display environment. The workstations were connected on a local area network. Each workstation was equipped with a keyboard, three-button mouse and a pencil to fill out prediscussion and postdiscussion questionnaires. Further details of the hardware and software configuration are available from the author. The experimental session consisted of four trials, one practice trial and three trials -- one for each experimental condition. The practice trial was always first. The three experimental conditions were ordered in 3! combinations, totaling six possible experimental sessions. Five of the combinations were performed by two groups each. Since there were 11 groups, the only ordering that was performed by one group was {g+avg} -- {g} -- {s}. The duration of the sessions lasted anywhere from 60-to-100 minutes. The experimental trial, following the classical small group experimental paradigm (McGuire et al, 1987), was partitioned into four sequence-invariant phases: (i) collection of individual prediscussion opinions, (ii) group discussion, (iii) group consensus attainment, and (iv) collection of individual postdiscussion opinions. The computer workstation monitor display was partitioned by software into four like-sized windows, and followed the ``What You See is What I See'' (WYSIWIS) interface design philosophy. An _Instruction_ Window at the upper left provided instructions informing the group what they should do next. A graphics-based _Feedback_ Window at the upper right (for {g} and {g+avg} only), allowed subjects to select prediscussion, consensus and postdiscussion opinions with the three-button mouse. It also displayed aggregate opinions. Any of the three buttons on the mouse performed the same function when selecting a graphical item. A _Problem_ Window at the lower left displayed the choice-dilemma problem. For {s}, the problem was "Ms. H". For {g}, the problem was a college graduate facing a choice of a difficult or an easy Ph.D. program. For {g+avg}, the central character had to choose between staying with a lifelong career at a large company, or instead moving to a risky, yet potentially lucrative start-up company. A _Discussion_ Window at the lower right, running the Internet Relay Chat (IRC) program, allowed subjects to type synchronous messages to each other during the discussion phase (ii). The discussion scrolled from bottom to top the discussion proceeded, such that the discussants could see prior statements from the conversation. The experimenter was seated at a fourth workstation, which ran IRC and networked the other three workstations, such that individual opinions could be collected electronically and stored to disk during {g} and {g+avg}. For instance, when collecting prediscussion opinions via mouse through the feedback window, the server collected opinions from all three subjects. Then, it would broadcast the three opinions back to each subject's workstation. When each workstation received the broadcast, it would update the feedback window and display the group's opinions in the format consistent with either experimental condition {g} or {g+avg}. This collect/broadcast loop was the same as well for the consensus and postdiscussion phases in {g} and {g+avg}. All mouse selections in the feedback window and keystrokes in the discussion window were recorded locally on each workstation for later analysis. Each message that the subject typed in IRC was prepended with a unique time-stamp denoting minutes and seconds. The subjects were aware that the experimenter was viewing the discussion, and that their responses were being recorded. They were allowed to ask any procedural questions during the practice trial. Occasionally, procedural questions were asked during the experimental trials, but typically occurred at the end of the discussion (e.g, "what do we do next?"). For the control condition {s} prediscussion phase (i), a questionnaire was distributed to each group member. Then, they read the dilemma involving the central character and the risky/safe alternatives. Next, they provided their individual written opinion. After the completed questionnaires were collected by the experimenter, in phase (ii) the group was instructed to discuss the problem using the IRC program. The instructions they read were as follows. Imagine that as a group, the three of you are advising Ms. H. Your task is to discuss Ms. H's dilemma and decide on the lowest probability that you would consider acceptable for Ms. H to continue with her musical training. Your group decision should consist of a single value ranging from 1-in-10 (riskiest) to 9-in-10 (safest). The group was permitted to discuss the problem for as long as it took to attain consensus. A timer in the feedback window displayed to each group member the number of minutes elapsed in the conversation. Once consensus was attained in (iii), the experimenter again distributed questionnaires to the group members (after which the trial was completed). The {g} and {g+avg} trials varied from the {s} trial in two important ways. (a) Instead of written questionnaires, prediscussion (i), consensus (iii) and postdiscussion (iv) subjects selected opinions interactively from the feedback window, via the three-button mouse. The nine _y_ axis labels, "1-in-10" through "9-in-10", could be selected by positioning the mouse pointer over the item and clicking a mouse button. (b) Directly before the discussion began, the group viewed their prediscussion opinions in the Feedback Window. Each subject's name was enclosed in a graphical box and shown along the _y_ position in the display to the right of the _y_ axis that corresponded to his/her prediscussion opinion. The _x_ and _y_ axes met at a common "origin". The _y_ axis contained 9 hash marks to denote each of the 9 discrete risk values. If two (or all three) subjects had the same opinion, the boxes were located at the same _y_ position and concatenated in alphabetical order. The boxes always were the width of the longest subject-name so that the boxes would be the same size. This was done so that potentially different box sizes would not confound results by implying the "importance" of one opinion over another. For {g}, only the opinions were displayed. {g+avg} was identical to {g}, except that the average of subjects' prediscussion opinions also was displayed, along with the numerical value of that average. If that value was non- discrete (e.g., 2.3), it was displayed as a _y_ axis label. For both displays, the _x_ position of the boxes in the graph was held constant. The displays in {g} and {g+avg} were presented to the subjects for 90 seconds, then removed before the discussion phase (ii) began. In this fashion, the displays during the discussion phase were by-and-large the same for all three conditions. This sequencing was considered carefully, since it was hypothesized that any effects of prediscussion modality would appear during the discussion. Introducing additional interactive graphics during the discussion phase would have only confounded these effects, and potentially could have resulted in an undesired interaction effect between the prediscussion and discussion phases. In {g} and {g+avg}, after the consensus value was selected (iii), the value was displayed to the subjects again in the Feedback Window, for 45 seconds. The value was displayed as before, except that the subjects' graphical boxes were concatenated and ordered alphabetically at the _y_ position, reflecting the group decision. In one case, the group agreed on a non-discrete value (e.g., 2.33), therefore they were not able to select the appropriate consensus value. The practice trial was identical to {g} except that some of the instructions were more verbose. 5.0 Dependent Measures and Results 5.1 Content Coding For the 33 trials, duration of discussion, number of utterances, prediscussion, consensus and postdiscussion opinions were evaluated. Unique codes were edited into the data files using the Gnu Emacs text editor. A parser written in the Lucid Lisp programming language then extracted and tabulated the coded information. In contrast with earlier CM work, the conversational unit of analysis in this study was defined as each subject's _message_ in the IRC medium -- even if one message comprised two or three sentences. Most messages, no matter how long, typically carried one basic semantic theme or idea. [1] The discussions in this experiment took on three salient characteristics that appeared consistently in all 33 discussions, reflecting a common discourse _structure_. Subjects spent their discussion time in one of three registers. (a) Arguing or weighing the dilemma (the risky alternative against the safe alternative), and in turn, providing persuasive arguments for or against either alternative -- usually in order to support or refute a proposed consensus value. At times, such persuasive arguments were accompanied by a consensus proposal. Hereafter, this discourse type will be referred to as _persuasive arguments_. (b) Proposing various values for the group consensus value, sometimes accompanied by social pressure and/or unsubstantiated claims. Hereafter, this discourse type will be referred to as _normative arguments_. (c) Settling on and affirming the consensus value, an activity which ended all 33 discussions. Hereafter, this discourse type will be referred to as _consensus attainment_. With few exceptions, all three types of registers occurred in contiguous sequences, involving only messages in the context of one of the three types. For instance, in a few messages subjects would weigh the risky and safe alternatives, then switch for several messages to proposing various consensus values, then for a few more messages make substantive arguments for or against their proposals, then finally come to agreement. For the purposes of this discussion, these contiguous sequences are recognized as distinct discourse _types_. More detailed examples of these types follow: Example 1: This sample is from a {g+avg} trial and is a superlative example of _persuasive arguments_: if it doesn't work out he is qualified enough [then] it should be relatively easy for him to find employment elsewhere i disagree that jobs would be easy for him to find...it's a rough job market but then again is this scenario set during a recession or during a good economic period?? what's the current economy? who can tell? i am assuming that the scenario is time-invariant With a degree in EE and 5 years experience he can find a job. i've got friends with the same experience who are unemployed! exactly it may take a few months or a year (to find a new job) I would say again that one has to take risks in life. anyway assuming that the scenario is time-invariant the current conditions of the economy should not matter much Example 2: This next example, from a {g} trial, demonstrates the interplay between: (a) _normative arguments_, in which a majority-style normative pressure is applied, and (b) _persuasive arguments_" (author's comments enclosed in parentheses) Let's go with 5 in 10. i've changed my mind- 2in 10 what do you think? I say 2 in 10 come on [s1] I think [s1] is being difficult. tell me one reason why I should agree with you guys. just as you said before - "if she puts her mind to it"! (majority-style normative pressure doesn't work, so switch to persuasive arguments) Why did you come to school here. You had a very good chance of success at a community college but chose the tougher school why? why? (continue on to consensus attainment) ok 2 in 10 (s1 acquiesces and the group reaches a decision.) Example 3: This example is from an {s} trial. It exemplifies a typical exchange, as the discussion begins. The second utterance, by s1, is an instance of an _explicit decision proposal_. what is your choice on Ms. H 3 in 10 Remember what are the lowest odds you would accept. I went with 8 in 10. 3 in 10 is really risky. I would still take a chance Example 4: These are opening utterances from a {g+avg} trial. They are a splendid example of an _implicit decision proposal_. s2 has selected 6-in-10, s1 7-in-10, and s0 2-in-10. Note how s1 and s2 discuss their opinions, as well as s0's opinion, without ever explicitly mentioning the value of these opinions. They instead refer to the graphical display which they had seen initially and that is no longer present. The way the economy is today I wouldn't take too many risks. I'm waiting to hear from you two. I don't like living in fear for the future of my kid I think you are right [s2]. [s0] that is some kind of risk there why? Later in the discussion, as the group consensus begins to coalesce, the first explicit advocacy appears. I will move to 5 in 10 and that's it! [s0] what say you? well [s2] i'll do the same Example 5: Finally, this example illustrates what happens typically during _consensus attainment_. I COULD GO WITH 4 IN 10 Have we decided? o.k. then it is in agreement with 4-in-10 yes Yes 4 in 10 5.2 Statistical Results From hypotheses H1, H2 and H3, seven dependent measures were analyzed for statistical significance using the one-way analysis of variance method. The measures and results are summarized in Table 2. For the three experimental conditions, there was no significant difference in the discussion durations. The choice shift between the average of the prediscussion opinions and the group consensus was not found to be significant. Attitude polarization, or the shift between the prediscussion average and postdiscussion average, also was not found to be significant. The lack of choice-shift is not consistent with earlier sentential CM research (i.e., Siegel et al, 1986). [2] The first hypothesis (H1) was not supported. H1 had supposed that the group consensus and postdiscussion opinions would be affected by the information conveyed in the graphics conditions {g} and {g+avg} to a greater extent than in the sentential condition {s}. Although it had been postulated that modality might have an effect on discussion duration (H2), the data did not support this conjecture, either. Likewise, the total messages for each condition did not vary significantly (H2). The number of messages occurring during the persuasive arguments phases as a percentage of the total messages in each discussion was found to be significant (p=.006). Likewise, the number of messages involving normative arguments as a percentage of the total messages in the discussion also was found to vary significantly (p=.0001). The number of messages involving consensus attainment as a percentage of the total messages in the discussion was not found to vary significantly. These data confirm parts of hypothesis H2, establishing the fact that the presence of the graphical modality in both {g} and {g+avg} had an effect on discussion structure and content. Regarding H3, there were significant effects regarding the explicit first advocate (p=.006), and the implicit first advocate (p=.013). These data supports the hypothesis that the explicit and implicit first advocate would have a greater influence on the group consensus than in {s}. More detailed pairwise comparisons of the significant results are summarized in Table 3. [3] There were proportionately more normative arguments in {g} (m=.38) than in {s} (m=.21, p<.01) or {g+avg} (m=.08, p<.05). Likewise, for {s} there were more normative arguments than in {g+avg} (p<.01). In sum, for normative arguments, proportionately: {g} > {s} > {g+avg}, which supports portions of H2. There were proportionately more persuasive arguments in {g+avg} (m=.58) than in {s} (m=.45, p<.1) or {g} (m=.32, p<.01). There were more persuasive arguments in {s} than in {g} (p<.1). In sum, for persuasive arguments: {g+avg} > {s} > {g}, which also supports H2. Regarding explicit and implicit first advocacy (H3), {g} and {g+avg} did not vary significantly from each other, as predicted. The differences between {s} and {g}, as predicted, were significant (explicit p=.<.05; implicit p<.05), and {s} and {g+avg}, were highly significant (explicit p<.01; implicit p<.01). The non-significant results in the {s} condition replicates results found in earlier face-to-face and sentential CM research (Weisband, 1992). Table 4 summarizes actual findings versus those anticipated. Table 2 SUMMARY OF RESULTS FROM PILOT EXPERIMENT ------------------------------------------------------------ Measure Mean SD F p ------------------------------------------------------------ Discussion Duration (min.) s 7.62 3.29 ns g 8.20 4.49 g+avg 8.64 5.81 Total Messages s 25.91 12.56 ns g 26.09 17.59 g+avg 27.73 19.04 Choice Shift s 4.27 1.21 ns g 4.67 2.02 g+avg 4.18 1.75 Attitude Polarization s 1.30 .58 ns g 1.10 .47 g+avg 1.03 .48 % Msgs. Normative Arguments s .21 .17 12.412 .0001 g .38 .13 g+avg .08 .12 % Msgs. Persuasive Arguments s .45 .15 6.066 .006 g .32 .18 g+avg .58 .19 % Msgs. Consensus Attainment s .34 .13 ns g .30 .13 g+avg .34 .17 Explicit First Advocate s 1.82 1.33 6.111 .006 g .79 1.17 g+avg .27 .47 Implicit First Advocate s 2.27 1.42 5.091 .013 g .94 1.20 g+avg .82 .87 ------------------------------------------------------------ Table 3 PAIR-WISE COMPARISON OF SIGNIFICANT CONDITIONS --------------------------------------------------- Measure Mean SD p < --------------------------------------------------- % Msgs. Normative Arguments s .21 .16 .01 g .38 .13 s .21 .16 .01 g+avg .08 .11 g .38 .13 .05 g+avg .08 .11 % Msgs. Persuasive Arguments s .45 .14 .1 g .32 .17 s .45 .14 .1 g+avg .58 .18 g .32 .17 .01 g+avg .58 .18 Explicit First Advocate s 1.82 1.33 .05 g .79 1.17 s 1.82 1.33 .01 g+avg .27 .47 g .79 1.17 ns g+avg .27 .47 Implicit First Advocate s 2.27 1.42 .05 g .94 1.20 s 2.27 1.42 .01 g+avg .82 .87 g .94 1.20 ns g+avg .82 .87 ------------------------------------------------------------ Table 4 SUMMARY OF ANTICIPATED VS. ACTUAL FINDINGS ----------------------------------------------------------------- I. Group Consensus Hypotheses (H1, H3) Treatment Conditions (IVs) ---------------------------------------------- No First Advocacy || First Advocacy ---------------------- -------------------- g | g+avg | s || g | g+avg | s ------ --------- ----- ---- --------- ----- Level of high | x# | || x* | x* | Accord ----------------------------- -------------------- (averaging medium x# | | || | | effect, ----------------------------- -------------------- DVs) low | | || | | x* ----------------------------- -------------------- none | | x# || | | II. GROUP DISCUSSION HYPOTHESIS (H2) Treatment Conditions (IVs) ---------------------- g | g+avg | s ------ --------- ----- | | High x* | | | ---------------- --------- ----- | Normative | Medium | | x* | (opinion) ---------------- --------- ----- | | Low | x* | | Level of ----------------------------------------------- Influence | g | g+avg | s (DV) | ------ --------- ----- | | High | x* | | Informa- ---------------- --------- ----- | tional | Medium | | x* | (persua- ---------------- --------- ----- | sion) | Low x* | | ============================================================ g | g+avg | s ------ --------- ----- High | | x# ---------------- --------- ----- Length of Discussion (DV) Medium x# | | ---------------- --------- ----- Low | x# | ============================================================ g | g+avg | s ------ --------- ----- High | x# | ---------------- --------- ----- Conciseness of Content (DV) Medium x# | | ---------------- --------- ----- Low | | x# ----------------------------------------------------------------- Note: "x" indicates value of anticipated finding for treatment indicated; "*" indicates anticipated finding was confirmed at a significant level; "#" indicates anticipated finding was not confirmed at a significant level 6.0 Discussion To summarize, the two discourse types _persuasive arguments_ and _normative arguments_ varied significantly with respect to the three experimental conditions {g}, {g+avg} and {s}. The third type, _consensus attainment_, occurred in equal proportions for all three conditions. These data suggest that the bulk of the modality effects rested exclusively in the phases of discussion where group members attempted to hash out substantiated and unsubstantiated opinions, normative influence, decision proposals, inference and logical arguments (i.e., informational influence) in order to attain a consensus. In addition, both the _explicit_ and _implicit_ first advocates in the two graphics conditions exerted significant effects on the outcomes of the group discussions, whereas the control condition {s} did not. Three parameters are considered central to the ensuing discussion: (1) individual prediscussion opinions, (2) individual opinions which are voiced during the group discussion, and (2) the group consensus which invariably emerges during the group discussion. What will be considered of great theoretical interest -- keeping cognitive and perceptual salience in mind -- are the relationships between these opinions and the communication modalities of the three experimental conditions. First considering the effect of the graphical modality {g+avg} on the group discourse, why would groups spend proportionately _less_ time discussing opinions and asserting normative influence and _more_ time substantiating their opinions with persuasive arguments, thus asserting more informational influence? Owing to the theorized effects of perceptual and cognitive salience discussed earlier in the background section, it is suggested that when group members see each others' opinions as part of the group average {g+avg}, those opinions present an irrefutable implied informational "decision proposal" or "group consensus", even before any discussion begins. The graphical display frames subjects' individual opinions with a very salient notion of consensus, denoted by the informational average. Thus, during the discussion (with the explicit knowledge that each member will eventually have to adjust his/her opinion in order to move towards a consensus), s/he will be more motivated to argue - - using persuasive arguments that are grounded in fact, logic and inference -- to justify and protect his/her position. Since individual members are more committed to their positions (further reinforced by their persuasive arguments), it certainly would not serve them well to submit alternate positions. This tack also would explain the small proportion of normative arguments. Regarding the {g} condition, why do groups engage in discussion proportions that diametrically oppose those of {g+avg}? It is proposed that when the discussants encounter the {g} graphical display, without explicitly seeing the average (as in {g+avg}), it can be assumed that _no_ decision proposal or group consensus has yet begun to take shape (before the onset of discussion). The discussants have no external reference against which to measure their individual opinions. They are not compelled to defend their positions, nor do they have a salient notion of opinion _and_ consensus, as in {g+avg}. On the contrary, the discussants have a very strong notion of their opinions, and these opinions overshadow any sense of consensus. Naturally, this contributes to a lack of cohesion as the discussion begins, and as individual opinions change and the group consensus takes shape. This explains the greater proportions of normative arguments. To put it another way, one could consider the effect of {g} almost "anti-consensual" in nature. This also explains the attenuation of persuasive arguments, since the discussants are not nearly as compelled to justify their opinions as in {g+avg}. With the {s} condition serving as a control for the experiment, the "neutral" modality of sentential text (with respect to {g} and {g+avg}) favored persuasive arguments over normative arguments. When the discussants first announce their prediscussion opinions to one another -- in contrast with {g} and {g+avg}, where they are preannounced graphically -- it is postulated that greater emphasis is placed on the cohesiveness of group opinion. In other words, by stating one's prediscussion opinion in {s}, one already is suggesting explicitly a value for the group consensus. It also should be kept in mind that group members do not see each others' questionnaires before the discussion begins. As in {g+avg}, group members quickly develop a sense of what the consensus value _might_ be, thereby inducing a need to justify one's own position, through persuasive arguments. The influence of implicit and explicit first advocates on group consensus in {g} and {g+avg} also met with very interesting (and favorable) results. Both graphics conditions demonstrated a statistically greater influence on the consensus than did the control. In both graphics conditions, the cognitively and perceptually salient information from the prediscussion graphical feedback ostensibly anchored discussants closer to the value of the first advocate's position. As compared with the sentential condition, this resulted in a significantly smaller shift of opinion from the moment of the first advocacy to the consensus. Owing to the lack of salience, however, the first advocates in the {s} discussions did not have as much of an influence on consensus as in the other two conditions. The notions of "opinion" and "consensus" are more susceptible to working memory effects and the conflation of information gleaned from the discourse, resulting in the shift of opinion away from the first advocate. What is interesting generally about the results when compared with traditional face-to-face research is the fairly clear delineation between normative and informational influence that occurs in most of the discussions. In some cases, a decision proposal might have been appended to an otherwise persuasive argument, but in general the distinction between the two was quite clear. Also of great interest were the almost identical proportions (roughly one-third each) that were spent in consensus attainment for all three conditions. The non-significant variances of discussion duration, discussion times, number of messages, choice shift and attituded polarization also were notable. These aggregate data provided a superlative backdrop against which influence-related parameters could be examined. Further decomposition and classification of discourse types in future research could reveal even more interesting discussion dynamics. Moreover, detailed study of the decision rules that discussants resort to could shed additional light on the issues raised thus far. Regarding the three parameters (individual prediscussion opinion, individual opinion during the discussion, and group consensus), Table 5 organizes the relative saliencies, along with the two significant outcomes: persuasive arguments and normative arguments. The metrics "high", "med", and "low" are assigned loosely to these five entities, in order to denote the relative saliencies associated with (1) the experimental manipulations (i.e., the first four columns) and (2) the results from the content coding of the discussion (i.e., the last two columns). Beginning with the row describing the {g} condition, note that the saliency of the prediscussion opinions (as well as the discussion opinions) are "hi" owing to the graphical representation, and group consensus is "low", since it develops solely during the conversation in the sentential modality. Next, for {g+avg}, the prediscussion opinions are "hi", as in {g}, but the discussion opinions and group consensus are "low" and "high" respectively -- directly opposite the "high" and "low" of {g}. Moreover, the proportions of informational (i.e., persuasive) arguments and normative (i.e., normative arguments) reflect the same diametric property. To summarize, what can be said about the effects of salience on group discussion with respect to influence? First, it appears that informational influence can be associated better with the notion of aggregate group opinion -- consensus -- more than it can be associated with individual opinion. Likewise, normative influence can be associated better with individual opinion than it can be associated with aggregate opinion. Regarding first advocacy (Table 6), the same experimental conditions (i.e., first four columns) give rise to similar notions in terms of the correspondence between an individual's opinion and the _value_ (not the formation) of consensus. Opinions unfold temporally, through the course of discussion. The true linear notion of "conversation" should be contrasted with the "clustering of opinion" (the graphical case in both {g} and {g+avg}). Table 5 "SALIENCY" MATRIX FOR DISCUSSION PARAMETERS ----------------------------------------------------------------- Predisc. Disc. Group Persuasive Normative Condition Opinions Opinions Consensus Arguments Arguments --------------------------------------------------------------- g high high low low high g+avg high low high high low s low low low med med --------------------------------------------------------------- Table 6 "SALIENCY" MATRIX FOR FIRST ADVOCACY ----------------------------------------------------------------- Predisc. Disc. Group Explicit Implicit Condition Opinions Opinions Consensus 1st Adv. 1st Adv. --------------------------------------------------------------- g high high low med med g+avg high low high high med s low low low low low --------------------------------------------------------------- 7.0 Conclusions The empirical results discussed in the discussion section appear to support the idea that _communication modality_ has definite effects on how CM group members attempt to influence each other during decision-making discussion. When they view 2D graphical information about their prediscussion opinions (which includes "meta" information such as the average of their opinions), group members attempt to assert more _informational influence_ on one another. The purpose of this section is to discuss briefly the implications of these empirical data and how they bear on issues of Human-Computer Interaction (HCI) in the realm of computer- mediated (CM) small groups. One is compelled to ask the question, "What constitutes a 'good decision', as opposed to a 'bad decision', in group decision- making". Unfortunately, such issues are highly subjective and context-dependent. But more importantly, it is not the role of the social scientist to ascribe values such as "good", "bad" or "quality" to any group's decision-making activities. Likewise, for the designer of a graphics-based display, these data suggest that the composition of the display will exacerbate some small group tendencies while attenuating others. Suggesting design guidelines at this phase of research certainly is premature -- more work needs to be done. Although I do not advocate that such displays should carry a sign which reads: "WARNING! This display will cause your group to engage in more persuasive arguments!", self-awareness on the part of the designer and of the users is key. The designer, in particular, could run the risk of creating a display which statistically might increase the users' tendency to bully one another into a decision, engaging solely in some perhaps not-so-desirable normative discussion behaviors. Knowing that the members of electronic groups might be susceptible to the effects observed in this study, how might such effects bear on the loss of inhibition, status equalization, riskier decision-making, and variations between normative and informational influence -- as it might relate (for example) to the case of the Rodney King? Were such potential effects on the computer-mediated group envisioned by the original designers of the electronic mail system used in the Los Angeles Police patrol cars? The answer, in all likelihood, is: probably not. The design intent, as all designers know, typically is altruistic. Given these known and potential effects on small group behavior, how might the designers of tools seek to curtail or avoid such behavior (which by society's standards, in such extreme cases, is unacceptable)? The design position that I suggest the reader consider seriously is that of a _pro-active_ and empirical design methodology. In such a methodology, (a) the designer would seek to understand user responses observed in this study, in the context of the tool being developed, (b) the designer would design a tool which makes such user responses more salient, perhaps by abstracting salient aspects of user responses into other modalities like interactive 2D graphics -- so that the design intent is more effectively conveyed to the potential user, (c) the designer actually would test for the exacerbation or attenuation of such effects on the computer-mediated group in a controlled experimental setting, perhaps accompanied by follow-up in the field, and finally, (d) if effects exceed acceptable criteria (e.g., as detected by analysis of the group's conversations), the designer would redesign the tool and repeat this cycle until user responses pass acceptable criteria. Although this kind of methodology might raise objections, it is not unlike the "third eye" brake light mechanism that has become a regulated standard in American automobiles. Its integration with existing brake lights (typically two) was prompted by faulty user responses, which confused braking intent with other rear-end visual cues, such as tail lights. This sometimes resulted in rear-end collisions. With addition of the unambiguous "third eye", braking intent was more explicit, resulting in proportionately fewer rear-end collisions. In the work I have reported here, I am making a similar case for CM-HCI. Footnotes 1. In two cases, one {s} and one {g+avg}, single messages were partitioned into two separate messages, since they entailed semantic ideas which were irrefutably distinct. 2. The _degree_ of shift, however, is considerably greater compared with earlier research. For instance, in Siegel (1986) the mean choice-shift for synchronous computer-mediated discussion was .89. Since that study used a completely different discussion medium (e.g., terminal and mainframe-based as opposed to workstation-based), it is not safe at this point to infer or induce what additional factors might have led to these differences. 3. The Duncan comparison procedure was performed using the SPSS statistical software package. 4. This most certainly motivates the design of a follow-up experiment in which each group member would see each other's opinions (e.g., on the _y_ axis, as in the present experiment) over time (e.g., on the _x_ axis). Acknowledgement Ellice Forman, Stephen Hirtle, Jill Larkin, Alan Lesgold, John Levine, Mike Lewis, Dirk Mahling, Richard Moreland and Walter Schneider. Special thanks to Sara Kiesler for her help in cultivating many of the foundational ideas that have gone into this research. Thanks to the Industrial Networking Institute at Carnegie Mellon and the University of Pittsburgh Department of Information Science for funding the subjects who participated in this study. References Bransford, J. D, Barclay, J. R., & Franks, J. J. (1973). Sentence memory: A constructive versus interpretive approach. Cognitive Psychology 3 (1972), 193-209. Fisk, A. D., & Lloyd, S. J. (1988). The Role of Stimulus-to-Rule Consistency in Learning Rapid Application of Spatial Rules. Human Factors 30(1), 35-49. Kaplan, S. & Simon, H. A. (1989). Foundations of Cognitive Science. In M. I. Posner (Ed.), Foundations of Cognitive Science. Cambridge, MA: MIT Press. Kiesler, S., & Sproull, L. (1992). Group Decision Making and Communication Technology. Organizational Behavior and Human Decision Processes 52, 96-123. Kotovsky, K., Hayes, J. R. & Simon, H. A. (1985). Why are some problems hard? Evidence from tower of hanoi. Cognitive Psychology 17, 248-294. Lewis, C. M., & Toth, J. A. (1992). Situated Cognition in Diagrammatic Reasoning. In N. H. Narayanan (Ed.), Working Notes, 47-52, AAAI Spring Symposium on Reasoning with Diagrammatic Representations, March 25-27, 1992, Stanford University. Loftus, E. F., Miller, D. G. & Burns, H. J. (1978). Semantic integration of verbal information into a visual memory. Journal of Experimental Psychology: Human Learning and Memory 4, 19-31. Mandler, J. M. & Ritchey, G. H. (1977). Long-term memory for pictures. Journal of Experimental Psychology: Human Learning and Memory 3, 386-396. Miller, C. E. (1989). The social psychological effects of group decision rules. In P. B. Paulus (Ed.), Psychology of group influence (2nd ed.), 327-355. Hillsdale, NJ: Lawrence Erlbaum Associates. Nickerson, R. S., & Adams, M. J. (1979). Long-term memory for a common object. Cognitive Psychology 11, 287-307. Paivio, A. (1986). Mental representations: A dual coding approach. New York: Oxford University Press. Ridgeway, C. L. (1984). Dominance, performance, and status in groups: A theoretical analysis. In E. J. Lawler (Ed.), Advances in group processes 1, 59-93. Greenwich, CT: JAI Press. Schneider, W., & Detweiler, M. (1987). A connectionist/control architecture for working memory. In G. H. Bower (Ed.), The psychology of learning and motivation 21, 53-119. Orlando, FL: Academic Press. Straus, S. G. (1991). Does the Medium Matter: An Investigation of Process, Performance, and Affect in Computer-Mediated and Face-to-Face Groups. Working paper, Graduate School of Industrial Administration, Carnegie Mellon University, Pittsburgh, PA. Tabachnek, H. & Simon, H. A. (1992). Reasoning About Economic Markets, in N. H. Narayanan (Ed.), Working Notes, 59-64, AAAI Spring Symposium on Reasoning with Diagrammatic Representations, March 25-27, 1992, Stanford University. Van Orden, G. C., Pennington, B. F. & Stone, G. O. (1990). Word identification in reading and the promise of subsymbolic psycholinguistics. Psychological Review 97(4), 488-522. Weisband, S. P. (1992). Group discussion and first advocacy effects in computer-mediated and face-to-face decision making groups. Organizational Behavior and Human Decision Processes 53, 352-380. Zhang, J. & Norman, D. A. (1990). The Interaction of Internal and External Information in a Problem-Solving Task. Technical Report 9005, Dept of Cognitive Science UCSD, La Jolla, CA. _____ Articles and Sections of this issue of the _Electronic Journal on Virtual Culture_ may be retrieved via anonymous ftp to byrd.mu.wvnet.edu or via e-mail message addressed to LISTSERV@KENTVM or LISTSERV@KENTVM.KENT.EDU (instructions below) or GOPHER gopher.cic.net Papers may be submitted at anytime by email or send/file to: Ermel Stepp - Editor-in-Chief, _Electronic Journal on Virtual Culture_ M034050@MARSHALL.WVNET.EDU _________________________________ *Copyright Declaration* Copyright of articles published by Electronic Journal on Virtual Culture is held by the author of a given article. If an article is re-published elsewhere it must include a statement that it was originally published by Electronic Journal on Virtual Culture. The EJVC Editors reserve the right to maintain permanent archival copies of all submissions and to provide print copies to appropriate indexing services for for indexing and microforming. _________________________________ _________________________________ _THE ELECTRONIC JOURNAL ON VIRTUAL CULTURE_ ISSN 1068-5327 Ermel Stepp, Marshall University, Editor-in-Chief M034050@Marshall.wvnet.edu Diane (Di) Kovacs, Kent State University, Co-Editor DKOVACS@Kentvm.Kent.edu ____________________________ GOPHER Instructions ____________________________ GOPHER to gopher.cic.net 70 ____________________________ Anonymous FTP Instructions ____________________________ ftp byrd.mu.wvnet.edu login anonymous password: users' electronic address cd /pub/ejvc type EJVC.INDEX.FTP get filename (where filename = exact name of file in INDEX) quit LISTSERV Retrieval Instructions _______________________________ Send e-mail addressed to LISTSERV@KENTVM (Bitnet) or LISTSERV@KENTVM.KENT.EDU Leave the subject line empty. The message must read: GET EJVCV2N2 CONTENTS Use this file to identify particular articles or sections then send e-mail to LISTSERV@KENTVM or LISTSERV@KENTVM.KENT.EDU with the command: GET where is the name of the article or section (e.g., author name) and is the V#N# of that issue of EJVC