A SEMANTIC APPROACH TO

VISUALIZING
Visualizing patterns in

ONLINE
CONVERSATIONS
online conversations
give participants a
better grasp of their
cyber environment
and their fellow
inhabitants.

Judith Donath

t any given moment, thousands upon thou- also representing the data accu-

A sands of conversations are taking place

online. They cover every imaginable topic,
from travel anxiety in a time of terrorism to tips
rately. This is a fine approach for
social scientists studying the
dynamics of online interactions. Yet
for our purpose it is also important
for teaching parrots to prattle. Some are public, that the visualization evoke an
others private. Some are synchronous chats, oth- appropriate intuitive response rep-
ers are archived discussions. Almost all resenting the feel of the conversa-
tion as well as depicting its
are text-based: easy to use, easy to dynamics. We call this
access, but lacking in visual approach “semantic
appeal and obscuring many of visualization,” for
the cues that aid social it introduces
interaction. Visualizing the assessments of
the meaning
patterns in these conversa- and relevance
tions can help the partici- of the data
pants gain a better sense of into the visual-
their social milieu: Who are ization process.
the other participants in the The Sociable
Media Group at
discussion and what are their the MIT Media Lab
roles? What are the social mores has been exploring ways of
that govern the interactions? visualizing online social interac-
tion for several years. Here, I discuss
This article discusses designing in the conversations—they are three such projects: Coterie, People-
visualizations of online conversa- interfaces rather than study tools. Garden, and the Loom Project.
tions in ways that help the viewer Traditional data visualization
perceive the online space as an focuses on making abstract num- Coterie
intuitive and legible social environ- bers and relationships into concrete, Coterie is a visualization of an
LISA HANEY

ment. These visualizations are spatialized images; the goal is to Internet Relay Chat (IRC) discus-
designed to be used by participants highlight important patterns while sion [9]. It highlights two key ele-

COMMUNICATIONS OF THE ACM April 2002/Vol. 45, No. 4 45

ments—the activity of the participants and the display fades and scrolls up with time; a channel with
structure of the conversation. In a standard IRC dis- a single closely debated topic will appear as a single,
play one can request a list of everyone logged into dense thread, a channel with multiple active topics
the channel, as individual IRC discussions are called, would have several dense threads, and a channel with
but in typical use only the participants who are desultory random discussion would have single state-
actively messaging are visible. ments appearing in various parts of the screen.
In Coterie, one can see at a glance how many peo- By visualizing presence and conversational activity,
ple are logged into a channel. The participants are rep- Coterie creates an environment that looks like the
resented as colored ovals that bounce and become conversation—coherent discussions form a solid, cen-
brighter when that person speaks. This highlights the tral core and scattered chats are, well, scattered all over
active participants and conveys the vitality of the dis- the place. One easily senses how populated the chan-
cussion. In IRC and other serial chats, multiple dis- nel is and who are the primary participants.
cussions often occur simultaneously, interwoven with To create an easily understood display of a conver-
each other—resulting often in a quite confusing sation, one must not only visualize the right data, but
stream of seeming non sequiturs that are in fact mul- depict it in an intuitive way. Let’s take a closer look
tiple interleaved conversations. at the design decisions made in creating Coterie.
Coterie analyzes the content of the conversation A key step is identifying what aspects of the con-

By VISUALIZING PRESENCE and conversational activity,

Coterie creates an environment that looks like
the conversation—coherent discussions form a
solid, central core and SCATTERED CHATS ARE, WELL,
SCATTERED ALL OVER THE PLACE. One easily senses
how populated the channel is and who are
the primary participants.

and attempts to sort the utterances into conversa- versation the visualization will highlight. In this case,
tional threads. This makes it easier to follow the dis- we focused on conversational cohesion. Conversa-
cussions. It also depicts the cohesiveness of the group tions are such a part of our everyday existence that we
and makes it apparent who are the initiators of new seldom think about their subtle and complex chore-
discussions. We use several heuristics for dividing the ography, yet upon close examination even the most
stream of messages into threads and assigning each mundane exchange resembles an intricate dance,
message to a thread. Coterie looks for repeated key improvised within the constraints of numerous com-
words and phrases, and places messages sharing them plex rules. One feature of this choreography is con-
in common conversations. It also looks for direct versational cohesion—how the participants maintain
addressing of a specific person. a common topic. During a conversation, people
Coterie’s algorithms include some biases for keep- introduce and sustain different topics. Strong conver-
ing a person in the conversation to which they are sational cohesion is a sign of cooperation, showing
assigned. However, if it is clear they are moving from that the members of the group share similar interests
one to another, it will display them as bouncing across and are motivated to sustain a common discussion,
the screen. Thus, participants focused on a single rather than individually attempting to redirect the
thread appear steady, while those whose attention is topic. The success of a new topic depends not only on
pulled in several directions are visibly more scattered. its inherent interest but also on its proponent: topics
Coterie also depicts the temporal patterns in the introduced by higher-status participants are more
conversation. Many traditional streamed chats add likely to catch on [2].
new lines as they arrive—looking at the screen, one Our focus on topic and cohesion led to the design
cannot tell if the 30 lines or so displayed have accu- that grouped messages by common content. The raw
mulated over the last minute or the last day. Coterie’s materials of Coterie’s visualizations—the colored

46 April 2002/Vol. 45, No. 4 COMMUNICATIONS OF THE ACM

 Figure 1. A Coterie display showing three simultaneous
conversational threads: one related to aircraft, one about
screens, and one with a comment about a previous
statement’s usage. Six users are currently active, but
many more are listening.

ovals, the columns of text—are simple and seemingly

neutral. Yet they are used in a way that creates an intu-
itive feeling for the channel’s atmosphere. Cohesive
discussions have a single main column, while diver-
gent ones have entries scattered across the screen.
Active users are bright and mobile and those who are
participating in multiple threads are bounce around
from one to another.

PeopleGarden
PeopleGarden uses a flower and garden metaphor to
visualize participation on a message board [12]. Par-
ticipants are each represented by a flower. The longer
they have been involved, the higher the stem; the thriving and varied garden, replete with tall plants and
more they have posted, the more petals. Initial post- short plants, giant lush flowers and tiny new buds.
ings are shown in red, replies in blue. One can easily The viewer can quickly grasp the underlying situation
get a sense of an individual’s role as an active partic- that created the visualization.
ipant, long-time lurker, and so on. The flower imagery is highly evocative, which is
Each message board is a garden full of flowers. both its appeal and its limitation. Gardens are organic,
There is an inherent legibility to the flower metaphor. we associate them with life, beauty, freshness. This
A PeopleGarden visualization of a fading group where imagery is fine when the visualization is of, say, a wel-
a few participants occasionally still check in will look
like a neglected garden, with a scattering of tall, scrag- Figure 2. Two PeopleGarden visualizations. Left: a group
gly plants that are mostly stem and little flower. Its with a single dominant member. Right: a group with many
visualization of a lively, active group will resemble a members at different levels of participation.

COMMUNICATIONS OF THE ACM April 2002/Vol. 45, No. 4 47

coming support group. It
seems jarringly inappropri-
ate for depicting dry techni-
cal announcements or
vitriolic flame wars.
The problem is the
semantic overloading of the
symbolic representation
given the sparse information
to be visualized. If a visual-
ization evokes meaning
beyond the direct mapping
of the data, there needs to be
information in that data that
guides the choice and shape
of the visualization. If there
is no such guiding informa-
tion, then the evocative
quality of the visualization is
likely to be misleading and
inappropriate. With People-
Garden, the algorithms that
map posting patterns to
flower growth and shape are
direct mappings of the data,
but the flower metaphor also
has its own richly
metaphoric meaning that
should be taken into consid-
eration when applying this
visualization.

The Loom Project

The goal of the Loom Proj-
ect is to create evocative,
semantically based visual-
izations of Usenet news-
groups [3]. To accomplish
this, we have been explor-
ing the fundamental ques-
tions of what to visualize and how to visualize it Figure 3. Loom visualization. Here we look at the
through a series of analytic and design studies. threading patterns in two groups. Related posts are
What to visualize? There is a nearly infinite set of placed in a circle. Top: a group with almost no
statistics that can be derived from a database of threading—posts are sent but seldom replied to. The
Usenet postings. Resolving these statistics into socially single dots scattered across the screen are a sign of a
meaningful classifications is the essence of the analytic nonsocial group. Bottom: a group with dense
side of semantic visualization. threads. Numerous and dense circles suggest a vibrant
The initial process is to identify the categories of conversational arena. This is a primarily quantitative
interest and define them in terms of recognizable fea- visualization, one that was initially designed to help our
tures. For example, we may wish to depict the leaders study of what data to depict. Yet the design is not
of a group. A reasonable working definition might be entirely neutral. The circular form is suggestive of a con-
people who post frequently and are often replied to in versational gathering and the scattered dots lend
a positive way. This distinguishes them from other themselves to an interpretation of disconnection [6].
frequent posters such as trolls (deliberate troublemak-

48 April 2002/Vol. 45, No. 4 COMMUNICATIONS OF THE ACM

ers), automatic newsfeeds, and the excessively verbose. research, both in terms of what to depict and how to
And it distinguishes them from frequent posters who depict it. Categorizing social data is a highly subjective
are often replied to negatively—these we may initially and culturally determined process. The designer’s
label as “provocateurs.” In this example, the category background, values, and prejudices will affect how a
is about the individuals in the group, but categories category is labeled and how it is matched to an expres-
can also be classifications of the situation, that is, the sive image. Most work in visualization has shied away
conversational thread or the group as a whole. One from such issues, preferring to present itself as neutral,
might choose to look at conversational groupings (as quantitative renderings of neutral, quantitative data.
depicted in Figure 3), the affective tone of messages, Yet there is no such thing as a truly neutral depiction
the group’s circadian rhythms, insularity, and so on. and we argue it is better to be cognizant of the ren-
How to visualize it? The traditional approach is to dering’s subjective and emotional qualities than to
depict the data, whether the raw statistics or the simply let them happen as they may. Furthermore,
results of the classification process, in a neutral, quan- when the rendering is of social data for participatory
titative way. In the group leader example, one could use, the neutral stance may not even be appropriate. It
graph the number of postings, the percent replied to, is useful to think of the design of such participatory
and the percent of positive responses as a basic graph. visualizations as being a form of architecture rather
Or one could use a similar chart to show the number than a branch of statistical graphing. Architects strive
of participants classified as “provocateurs,” the num- to bring out the evocative and symbolic qualities of
ber classified as “leaders,” and so on. In either case, the their buildings while working within the constraints
interpretation is not inherent in the picture, only in of functionality and the need to fit within the larger
the viewers’ knowledge of the meaning of the data. environment.
The semantic approach, on the other hand, makes While rules can guide architecture, the design of a
use of our knowledge about the social meaning of the particular building comes from the aesthetic sense and
categories we have defined and creates a visualization intellect of its architect. Similarly, we are working
that reflects this knowledge. To use a simplistic exam- toward creating visualizations of online conversations
ple, a semantic visualization might use simple shapes that draw from both the quantitative depiction of
and saturated colors for the leaders, and spiky shapes their patterns and rhythm and a semantic approach to
and acidic colors to depict the provocateurs. We have utilizing meaningful representations. c
been drawing from a variety of cognitive science and
graphic design studies that look at how different ele- References
ments such as color, shape, position, motion, among 1. Arnheim, R. Visual Thinking. University of California Press, Berkeley,
CA, 1972.
others, create different impressions for the viewer [1, 2. Bonvillain, N. Language, Culture and Communication. Prentice Hall,
4, 5]. One of our future research goals is to develop Englewood Cliffs, NJ, 1993.
3. Boyd, D., Lee, H-Y, Ramage, D. and Donath, J. Developing legible
new approaches to mapping meaning and visual visualizations for online social spaces. In Proceedings of the Hawaii
design, such as a system that lets a population of users International Conference on System Sciences. (Jan.7–10, 2002, Big Island,
evolve graphic designs they feel suited for depicting Hawaii).
4. Dondis, D. A Primer of Visual Literacy. MIT Press, Cambridge, MA, 1973.
various social patterns [7]. 5. Lakoff, G. and Johnson, M. Metaphors We Live By. The University of
Chicago Press, Chicago, IL, 1980.
6. Milgram, S. The drawing power of crowds. The Individual in the Social
Conclusion World (2nd edition). McGraw-Hill, New York, 1992.
It is important to remember that all visualizations 7. Sims, K. Artificial evolution for computer graphics. In Proceedings of
will have some evocative quality. We do not think in SIGGRAPH’91. ACM, New York, NY, 319–328.
8. Spertus, E. Smokey: Automatic recognition of hostile messages. In Pro-
pure abstractions; rather, our thinking is metaphoric ceedings of the Ninth Innovative Applications of Artificial Intelligence
and grounded in the spatial world [5]. For instance, (IAAI-97).
9. Spiegel, D. Coterie: A visualization of the conversational dynamics
things higher seem positive—and thus how the axes within IRC. MIT Master’s Thesis, 2001.
of a chart are labeled can subtly but strongly influ- 10. Tannen, D. Repetition in conversation as spontaneous formulaity. Text
ence how it is interpreted [11]. Colors and shapes are 7, 3 (1987), 215–243.
11. Tufte, E.R. Envisioning Information. Graphics Press, Chesire, CT 1990.
highly evocative and once introduced we must be 12. Xiong, R. and Donath, J. PeopleGarden: Creating data portraits for
very sensitive to how they appear. Indeed, one argu- users. In Proceedings of UIST 1999.
ment for deliberately designing evocative visualiza-
tions for online social environments is the existing Judith Donath ([email protected]) is assistant professor of
default textual interfaces are themselves evocative, Media Arts and Sciences and the director of the Social Media Group at
MIT’s Media Lab in Cambridge, MA.
they simply evoke an aura of business-like monotony
rather than the lively social scene that actually exists.
There are many complex issues that arise in this

COMMUNICATIONS OF THE ACM April 2002/Vol. 45, No. 4 49