Back to Vol. 2, No. 4 Table of Contents
[Conversation]
Collaborative U CMC Play E-Commerce Symposium Net Law InfoSpaces


Framing Flames:
The structure of argumentative messages on the net

Edward A. Mabry

Department of Communication
University of Wisconsin-Milwaukee

Table of Contents

Abstract

The purpose of this study was to assess the use, in computer-mediated communication, of the strategic message structuring tactic known as framing. Interlocutors in computer-mediated environments have software supported systemic resources facilitative of constructing messages using framing tactics in their argumentative discourse. It is hypothesized that framing strategies are related to the emotional tenor of a disputant's message and that a speaker's emotional involvement with an issue should be curvilinearly related to the appropriation of framing as an argumentative discourse strategy. Results from the analysis of 3000 messages, obtained from a diverse sampling of computer-mediated discussion groups and forums, provided support for the primary hypothesized relationship. A speaker's emotional involvement was significantly and curvilinearly related to two message framing devices (message dependency and coalition building) and a measure of conciliatory face-saving moves.

Participation in Computer-Mediated Groups

Growing use of computer-mediated communication (CMC) is being fostered by the increasing popularity of and expanding access to the global computer communications network known as the Internet (or, net). Internet use has increased as network members have created and maintained a wide variety of specialized discussion groups for exchanging messages using electronic mail. These electronic groupings vary in their formality and conversation-like continuity of message exchanges (Sudweeks, Collins and December, 1995). Many of these affinity groups represent relatively open communication environments that only loosel y conform to the topical identities originally project ed by their organizers, whose computer facilities often support the message distribution system (Rheingold, 1993). Permissive environments tacitly, if not actively, encourage the airing of controversies and usually evidence substantial tolerance in supporting ongoing arguments and disputes.

Argumentative exchanges among net group members are quite prevalent. The intensity and deviancy of such disembodied exchanges can become so heated and destructive (Hiltz et. al., 1989) that norms of network usage (referred to as netiquette), inculcated and reinforced among net users, specifically address the diligence users should display in how they write argumentative messages. These norms stress obligations for group and self-monitoring to insure that members maintain civility and communicative relevance (Kiesler et al., 1985; Lea et al., 1992). One frequently sanctioned breach of netiquette involves flames: messages that are precipitate, often personally derogatory, ad hominem attacks directed toward someone due to a position taken in a message distributed (posted) to the group (Siegel, Dubrovsky, Kiesler, and McGuire, 1986).

Argumentative Message Structure

The apparent acceptance, if not cultivation, of argumentative discourse in computer-mediated discussion groups stands in sharp contrast to the conventions of ordinary social conversation. Conversation theorists and researchers convincingly demonstrate that conversants typically display a "preference for agreement" in social interaction (Brown and Levinson, 1978; Holtgrave, 1986; McLaughlin, 1984). This paradox is compelling. Mediated groups are susceptible to more extensive, and often more intense, amounts of argumentative discussion than face-to-face groups (Kiesler and Sproull, 1992; Weisband, 1992). Conversely, routine social practices governing conversation are grounded in taken-for-granted assumptions of agreement and conciliatory behavior. And there is little reason to doubt that computer mediated messaging emulates primary linguistic characteristics of face-to-face interactions (Selfe and Meyer, 1991; Spitzer, 1986). The question arises as to how mediated groups manage this more adversarial communication context while retaining the discursive coherence and cohesiveness necessary for enacting socially appropriate rational discourse.

A strategy that has been proposed for maintaining the structural coherence of conversational arguments in face-to-face interaction is the practice known as recounting (Govier, 1985; Snoeck Hankemans, 1991). Recounting is the restating or summarizing of an opponent's position (including supporting or refutational stances) in the construction of one's own refutational or counter-refutational message. Snoeck Hankemans (1991) proposed that recounts may act to strengthen argument either indirectly or directly. Indirect advantage is gained by linking a refutational claim to the opponent's argument (should the claim be successful) with a reassertion of one's own stance. This combines a potentially successful attack with a potentially successful argumentative defense. Direct contribution to refutations of argumentative claims advanced by others is made in the juxtapositioning of counterclaims and supports. This structure amplifies the arguer's counter-refutation. It serves to punctuate an arguer's role-switching in moving between being the defender of a position and an attacker of the claims and counter-claims advanced by the other disputant.

Enactment of such strategies is interpretively similar to Bateson's (1972) conception of the perceptual "framing" that individuals engage in as they construe social situations. Bateson analogized this phenomenon to the principle of figure-ground separation in perception. A frame-of-reference is the perceptual differentiation called into use by drawing the social landscape (background) against which particular people, events, behaviors (figures) are compared. According to Bateson, frames provide important and unique bits of social information and contribute to metacommunication.

Aside from Snoeck Henkemans' theoretical explication, there has been little direct examination of argumentative framing, although indirect support for it is evident in studies of interlocution and conversational arguments (Antaki and Leudar, 1992). Holtgraves' (1986) examination of speech acts has shown that conversational rejoinders interact with speakers' face needs. Confrontational assertive rejoinders were more likely to occur in non-face-threatening episodes, whereas accepting nonconfrontational rejoinders was more likely in face-threatening situations. Predictably, accepting rejoinders were always perceived as more polite than assertive rejoinders.

Studies of small group influence recognize the importance of discussants' reactions to decision proposals, task opinions, and similar issues that often lead to arguments and group conflict. Alderton and Frey (1986) proposed using argumentative responses as an index of the impact of group arguments. Mabry and his colleagues (Mabry, Jackson, McPhee, and VanLear, 1990; Mabry, VanLear, Jackson, and McPhee, 1991) created and tested a method for analyzing group arguments that included the extent of responsiveness to a preceding message present in a succeeding speaker's argumentative message.

Research comparing computer-mediated groups to face-to-face groups also provides indirect clues to message strategies that may be in play when members attempt to influence group outcomes. Weisband (1992) has shown that advocacy sequences produce different outcomes in mediated versus face-to-face groups. Advocates in face-to-face groups, and early advocates in mediated groups, were significantly more likely to converge toward a group opinion. Third advocates (the final person in three-person groups) in computer-mediated groups were more likely to retain and express their pregroup opinions regardless of how discrepant that stance had become by the time it was made public.

Kiesler and Sproull (1992) noted similar processes in their assessment of mediated decision-making groups. They interpreted such patterns to be a concomitant of the more complex social organization required of mediated groups. Members often are simultaneously inputting messages. Some advocacy effects ultimately appear to be the resulof asynchrony. However, mediated groups also strive to accommodate diversity of participation. Thus, members may not initially respond to primacy effects while they are enacting turn-taking sequences. Therefore, time or other response latencies may dilute the primacy effects of first advocates and significantly increase the likelihood that later advocates will influence group thinking.

The Human-Technology Interface in CMC

One aspect of participating in CMC systems could be unique with respect to its impact on the structure of argumentative discourse. Virtually all systems software supporting the electronic mail (email) function on which network communication is based includes so-called "cut-and-paste" message editing features. The combination of electronic filing, making it possible to save exchanged messages, and efficient keyboard editing utilities affords computer-mediated interactants access to very powerful message construction tools.

An arguer can easily create a new message, say a counterargument to a position advanced in a previously distributed message, by inserting segments of the disputed message along with counterclaims the arguer wishes to advance. In this way, edited insertions (when they are used) function as strategic recounts and frame the arguer's discursive moves.

Following is an example of a brief, modestly confrontational exchange of email messages, employing cut-and-paste editing:

#: 312513 S14/Chatter
16-Mar-93 13:54:29
Sb: #312252-#Babylon 5
Fm: <Pseudonym A>
To: <Pseudonym B>
> > > The "logging arc" had a rather curious creative problem, being that at its heart is was a parody of something which the general public had not seen yet. < < <

Do you think that arc was an out and out parody of TWIN PEAKS? It had some of that flavor, I agree, but I didn't see it that way.
> > > But, it did have David Straithairn, which is a positive point in anything. < < <

Which player was he?


The arguer (A), in the context of an exchange of messages regarding impressions of television programs, disagrees with an interpretation of a plot-line advanced by another discussant (B). The left (<) and right (>) angle brackets are used as graphical markers by A to display the segment of B's message that A wishes to counterargue. A's counterclaim, in ordinary sentence formatting that visually sets it apart as a separate- yet concatenated-message act, follows immediately below B's bracketed text. Also note A's use of cut-and-paste editing to accomplish a different goal. A second bracketed text segment of B's message is followed by a question from A seeking clarifying information from B. Both of A's reactions function as inducements for B to send a responding message to A on this topic. Thus, A's message signals both an interest and a willingness by A to continue the mediated dialogue with B-even if there was a point of contention on the floor.

Research Objectives and Hypotheses

Little is known about communicators' uses of software supported message framing devices in the enactment of online arguing. Particularly important are the implications of this form of message construction in the management of interaction involvement and maintenance of rationally versus emotionally driven argumentative stances. Message editing is a rationalistic activity requiring concentration and attention to the content of arguers' language and reasoning choices. Conversely, emotionally intense messages, especially negative ones that are often perceived as flames, seem ill-suited to the cognitive demands of message editing. Enacting highly rationalized argumentative strategies, juxtaposing claims and counterclaims in a complex message, requires a deliberateness that would escape most arguers in the heat of battle.

It is assumed a message's relational implications, particularly argumentative convergence (or divergence), is related to message structure. Positive or negative emotional overtones should evoke less internally structured messages. Likewise, rational and emotional message characteristics ought to be linearly related to the confrontational intensity. These assumptions lead to the following hypothesis:

H1 Message framing and message continuity evoking strategies should be curvilinearly related to message intent.


It follows that message structure and coherence should be related to other message characteristics like compositional style and purpose. Message structure is only one facet of message content. Different message content strategies could be used in conjunction with varying levels of emotionality. In general, message content like appeasement, conciliation, or aggressiveness, should be linearly related to message emotionality.Thus, the second hypothesis of the study asserts:

H2 There is a positive (negative) linear relationship between message connotativeness and message emotionality.


Methods

Overview

This study was conducted as part of an international project in computer-mediated collaborative research. The principal goal of the project was to facilitate the investigation of actual computer-mediated messaging activities in Internet computer-mediated discussion groups. A thorough explanation of the project is in Sudweeks and Rafaeli (1996).

Message Sampling and Analysis

A complete discussion of sampling is available in Rafaeli et al. (1994) and will be abbreviated here. Various computer bulletin boards, lists, and newsgroups were canvassed for a period of approximately one month. Messages were randomly sampled across days and times. Sampling of a given group terminated when a target of 100 messages was collected. Messages were sampled beginning on a randomly chosen Monday (March 15, 1993), and took, in some cases, up to six months before the target number of messages was reached. Only lists in English were retained for the message pool. Approximately 3000 messages, from 30 different online discussion groups, have been coded and comprise the project's database.

A standardized message content analysis coding protocol was collaboratively developed by participating project members in the first year of their work together. The content analysis scheme measured 46 message variables of which 40 were hand-coded and 6 were machine coded. The coding protocol required trained coders to read the literal text of a message and apply all code variables to each message (wherever possible). A message was evaluated on whether it contains content descriptive of facts, opinions, humour, challenges, metacommunications, presence of graphic art, formality of composition, quoted material, emotional tone (or flames), and sender characteristics (e.g., gender, status) (see Rafaeli et. al., 1994, for a more complete explanation of message analysis categories).

More than 40 people participated as coders. Coders were furnished with a codebook (distributed electronically) and provided with on-line training activities and guidance. Training involved coding a set of sample messages chosen to cover the range of code variables. Coders rated the messages and returned their results via electronic mail. Low concordance on the training sample of messages led to additional coaching and more training messages being sent for test coding. High agreement with the preferred responses to training messages qualified the person as coder for the purpose of receiving messages to be analyzed in the main study. Completed sets of formatted message analysis codes were returned via electronic file transfer to a host computer system. Work was automatically screened using custom software to debug technical errors (e.g., off line formats, typographical errors) and any rejected codes were sent back to the coder for correction.

Reliability

Two conventional methods for assessing reliability were used: Brennan and Prediger's (1981) modification to Cohen's kappa, k(n), coefficient and Cronbach's (1951) alpha coefficient. The k(n) coefficient permits an assessment of interratreliability under conditions where marginal values of an n x n coding matrix are free to vary. Kappa was computed for all variables used in this study. Alpha was computed for items that were treated as scaleable. Reliability and descriptive information for variables needed in testing study hypotheses is contained in Table 1.

Table 1. Reliability and descriptive statistics for variables analyzed.
Independent variables
Dependent variables
Hypothesis 1
Emotional tone
Message dependency
Quotations (lines)

n

%

n

%

n

%
Neutral

1765

58.8

 None

907

30.2

 None

2115

70.5
Friendly

767

25.6

 1 message

1600

53.3

 1-10

666

22.2
Diverging

205

6.8

 2+ messages

178

5.9

 11-25

163

5.4
Disagreeing

125

4.2

 Series

302

10.1

 26+

49

1.6
Tension

65

2.2
k(n) = 0.592
k(n) = 0.882
Antagonism

46

1.5
alpha(8) = 0.609
alpha (7) = 0.833
Hostility

20

0.7
k(n) = .525
Hypothesis 2
alpha (6) = .670
+/- Coalescence
Apology

n

%

n

%
Strong agreement

156

5.6

 None

2802

93.4
Mild agreement

229

7.0

 Mild

139

4.6
None

2087

69.6

 Full

52

1.7
Agree/Disagree

141

4.7
k(n) = 0.940
Mild disagreement

254

8.5
alpha (6) = 0.778
Strong disagreement

126

4.2
k(n) = 0.525
alpha (7) = 0.749
Conciliation
Challenges

n

%

n

%
None

2865

95.5

 No

2877

95.9
Avoids tension

60

2.0

 Yes

116

3.9
Reduces tension

68

2.3
k(n) = 0.950
k(n) = 0.948

Note. Percentages may not sum to 100 percent due to missing data. Numbers in [ ] are number of online groups for which Cronbach's alpha could be computed. Emotional Tone is the only Independent Variable tested in Hypotheses 1 and 2.

A sample of 1,000 messages comprising 100 messages from each of 10 computer-mediated discussion groups/lists, were double-coded and thus constituted the data for computing reliability analyses. Because coders were not fixed across lists, reliability calculations were performed on a listwise basis (for each 100 cross-coded message sample) and averaged. Listwise alpha coefficients could not be computed for Conciliation and Challenges variables due to attentuated variances caused by high percentages of agreement among coders (in excess of 90 percent). Bracketed entries in Table 1 show the number of lists used in deriving mean alpha coefficients. Two variables, Opinion and Fact, did not attain acceptable k(n) coefficients and were dropped from the analyses. Both items were related to the test of Hypothesis 2. However, neither item was deemed critical to an adequate test of the hypothesis.

Statistical Analyses

The primary analytic strategy was to assess the influence of a message's Emotional Tone (its level of argumentativeness) on message structure and attribute variables. Consistent with Hypothesis 1, it was expected that Emotional Tone would not be linearly related to message variables measuring argumentative framing tactics. The framing tactics are reflected in Message Dependency (referencing of previously distributed messages) and length of Quotations (the amount of previous message material inserted into a coded message). The hypothesis was tested using univariate trend analyses with Emotional Tone as the independent variable (Kerlinger and Pedhazur, 1973).

Hypothesis 2 asserted that message attributes consistent with seeking positional convergence or divergence should be linearly related to Emotional Tone. This hypothesis also could be tested using univariate trend analyses with Emotional Tone as the independent variable and Coalescence, Apology and Conciliation as dependent variables. Additionally, a measure of verbalized Challenges in messages was judged as too range restricted for entry into the trend analysis. It was addressed in a chi square analysis that compared it with the Emotional Tone variable.

Results

Hypothesis 1: Message Framing

Message Dependency

The trend analysis confirmed the hypothesis concerning a curvilinear relationship between Emotional Tone and Message Dependency. The weighted quartic ANOVA was significant (F [1,2980] = 30.62, p < .0001). Multiple range tests using the conservative Scheffe method (Winer, 1971) revealed a significant (p < .05) increment in the amount of references to prior messages between affectively neutral messages compared to messages observed to be diverging, friendly, disagreeing, evoking tension; message referencing also was significantly higher in antagonistic messages compared to neutral, hostile, diverging, friendly, and disagreeing messages. As the pattern of means in Table 2 indicates, dependency increased as a message's Emotional Tone became increasingly negative but tended to flatten out between disagreement and antagonism, ultimately dropping off at the point where a message was overtly hostile.

Table 2. Means and standard deviations for variables tested in Hypothesis 1.
Emotional tone
Message dependency
Quotation
Mean
SD
Mean
SD
Neutrality
1.71
0.78
1.26
0.54
Friendliness
2.23
0.87
1.40
0.64
Diverging
2.27
0.78
1.78
0.87
Disagreement
2.37
0.85
1.74
0.83
Tension
2.39
0.93
1.89
1.00
Antagonism
2.89
1.04
1.97
0.95
Hostility
2.05
1.05
2.05
1.32

Quotation

A similar relationship between Emotional Tone and Quotation linage could not be established. The weighted linear ANOVA trend model was highly signficant (F (1,2986) = 266.95, p < .0001). The distribution of means corroborates the ANOVA results (Table 2). The upward trend for increasing length of quoted material insertions into messages is broken only by a plateauing between diverging and disagreement after which the linear trend re-emerges and continues unbroken.

Hypothesis 2: Message Content Characteristics

Positive/negative coalescence

The ANOVA trend analysis results for Coalescence were curvilinear, rather than linear as predicted, thereby failing to confirm the hypothesis. The value of the weighted quintic term was F (1,899) = 18.32, p < .0001). The shape of the trend, as reflected in the patterns of significant Scheffe comparisons between means showed a double-bend, or wave, pattern across levels of Emotional Tone (Table 3). Messages rated as neutral, friendly, or hostile were also more likely to contain bids for agreement or convergence in positions compared to diverging, disagreeing, tension-evoking, or antagonistic messages. The mean for Coalescence returned to a point of mild agreement from nearly a plateau point of strong disagreement. As might be expected, Coalescence means for emotional neutrality and hostility explained most of the variance in significant Scheffe comparisons of mean range distributions. Hostile messages containing agreements, capitulations, or compromises are relatively complex messages. Thus, these results indicate greater message complexity is being signaled in the Coalescence scale than was initially predicted.

Table 3. Means and standard deviations for variables tested in Hypothesis 2.
Emotional tone
Apology
Coalescence
Conciliation
Mean
SD
Mean
SD
Mean
SD
Neutrality
1.06
0.27
2.17
1.01
1.01
0.15
Friendliness
1.06
0.28
2.37
1.15
1.05
0.29
Diverging
1.10
0.37
3.57
0.81
1.18
0.53
Disagreement
1.12
0.41
4.24
0.98
1.16
0.50
Tension
1.34
0.67
3.83
1.40
1.43
0.72
Antagonism
1.35
0.71
4.10
1.22
1.61
0.88
Hostility
1.80
0.83
2.16
1.27
1.75
0.79

Conciliatory Statements

Results for the trend analysis of the relationship of Emotional Tone'with communicators' use of conciliatory statements also were not in the expected direction.

The ANOVA analysis resulted in a significant weighted quadratic trend: F (1,2987) = 37.90, p < .0001. The pattern of significant Scheffe pairwise mean comparisons emulated a classic arching configuration showing greater intensity of conciliation as Emotional Tone turned more negative. Emotional neutrality and friendliness were associated with significantly fewer expressions of conciliation than divergence, disagreement or tension containing messages. However, the intensity of conciliatory behavior again significantly increased as message content also included tense, antagonistic, or hostile argumentative statements.

Apology

ANOVA trend analysis results for the use of Aporevealed a strong quadratic trend (F (1,2986) = 56.55, p < .0001). Inspection of means for level of Apology (Table 3) included in the composition of a message clearly indicates that apologizing intensity escalates with message confrontiveness. The quadratic form of the results points to a clear interaction between Emotional Tone and Apology that is supported by the pattern of Scheffe mean comparison results. Mean rates of apologies did not register significant changes until the emotional tenor of a message reached at least a perceived state of tension display. Thus, the perceived affect level of a message probably must shift towards a more negative bias before apologizing seems warranted.

Challenges

The chi square analysis of Emotional Tone and Challenges did reach statistical significance: X2(6) = 250.89, p <0001. However, contrary to the implicit rationale of the hypothesis, Challenges decreased as Emotional Tone shifted from positive to negative poles. Thus, for instance, there were three times as many Challenges in messages rated "friendly" (f = 31; 26.7%) compared to messages rated "hostile" (f = 10; 8.6%).

Discussion

The results provided partial support for the hypothesized relationships. A communicator's emotional involvement and use of message framing devices (making pointed references to prior messages and quoting from those messages) are systematically related.

The study also tested hypothesized relationships between a communicator's emotional involvement in their communication goals and use of communication practices that facilitate or impede the development and strengthening of personal relationships. Although not linearly related to a message's emotional tone as hypothesized, being conciliatory and apologetic increased as message affect increased.

Thus, there is some reason to believe that communicators try to neutralize the effects of negative emotional spirals when they arise. However, the study also found evide nce that communicators are just as likely to move towards polarizing sentime nts in a dispute.

Messages seeking positive or negative coalescence on an issue were significantly related to a message's perceived emotional tone. And, in a somewhat surprising finding, communicators used more confrontive challenging messages when a message was low in affective tenor.

The implications of these findings are intriguing. First, they clearly indicate that conversational and argumentative structuring is apparent in mediated groups. Technologies supporting mediated groups offer resources that facilitate the use of complex and adaptive message framing strategies for enacting social argumentation.

Second, as Antaki and Leudar (1992) reason, there is an apparent duality in discourse that is found in both face-to-face conversations and computer-mediated interactions. Dialogues of all sorts often turn from platforms for agreement to the exchanging of claims (contentions) and counterclaims. These moves are accomplished as cooperative, but argumentative, dialogue games.

Third, results support the efficacy of applying conceptual models of communication developed for explaining face-to-face interaction to mediated environments. This is another instance of corroboration for the utility of mediated communication providing communication resource opportunities similar to those expected from face-to-face interaction (Kiesler and Sproull, 1992).

References

About the Author

Edward A. Mabry (BA, Sociology; MA, Speech Communication, CSC-Long Beach; PhD, Speech and Communication, Bowling Green State University) is an Associate Professor of Communication, Department of Communication, University of Wisconsin-Milwaukee, where he has held a variety of positions including department chair and director of graduate studies. His interests are in group and interpersonal communication, and the impact of mediated communication on instrumental performances and personal relationships. He is the co-author (with R. E. Barnes) of The Dynamics of Small Group Communication and has contributed papers to a number of national and international journals including Small Group Research, Speech (Communication) Monographs, and Human Communication Research (where he has served as a member of the editorial board).
Address: Department of Communication, University of Wisconsin - Milwaukee, PO Box 413, Milwaukee WI 53201, USA.