ELearning

Volume 6 Number 1 2009

www.wwwords.co.uk/ELEA

Visualization of Expert Chat Development

in a World of Warcraft Player Group

MARK CHEN
University of Washington, USA

ABSTRACT This article describes expertise development in a player group in the massively
multiplayer online game World of Warcraft using visualization of chat log data. Charts were created to
get a general sense of chat trends in a specific player group engaged in high-end raiding, a 40-person
collaborative activity. These charts helped identify patterns in the frequency of chat over time during
two specific gaming sessions. The sessions represented significant moments in the raid groups history:
the first time a particular monster, Ragnaros, was fought and one of the first times he was defeated.
Differences between the two cases show differences in the raiding practice and are evidence of learned
expertise. Based on this analysis, it is clear that this visualization process is a useful analysis tool within
ethnographic accounts of expertise development.

Imagine 40 people grouped together in a dark, hot, volcanic cavern deep beneath the earth. Some
of them appear to have been human at one point, but the flesh rotting off their frames clearly
points to some supernatural force. Others are muscular, green-skinned brutes or wiry, purple-
skinned figures sporting Mohawks and tusks. Finally, a few hefty, cow-like, bipedal forms stand
much taller than the others. Some in this exotic group are dancing, some are jumping up and
down, others are sitting and drinking water and various other liquids, but the majority of them are
just standing around, waiting or watching the large, spiky, snake-man creature in the middle of the
chamber. The humanoids are wearing an assortment of leather or metal armor and/or cloth or silk
robes, and they are equipped with glowing swords, maces, and staves. A few of them are discussing
the upcoming fight. One of them in particular is talking about the specific positions and roles for
the others during the fight. Many of the others are talking privately with each other at the same
time, sharing pleasantries or chatting about more mundane events, as if oblivious to their locale
and the upcoming fight.
The apparent leader, the one who was summarizing roles and positions, yells, Get in
positions! and everyone spreads out, running to various parts of the large cavern. A sizeable group
of them bunches up near a lava flow, directly across from the snake-man. Talk to Domo! yells the
raid leader, and one of the green orcs, decked out in full metal armor, rushes to the snake-man,
Majordomo Executus. Domo, seeing the orc approach, yells, Impudent whelps! Youve rushed
headlong to your own deaths! See now, the master stirs! He then summons his boss, the overlord
of this intricate cavern system known as Molten Core. His name is Ragnaros, and he emerges from
the center of the chamber, adding to the sweltering heat, his fiery, semi-liquid form towering and
massive like no other monster in this harsh land known as Azeroth.
Behold Ragnaros the Firelord! He who was ancient when this world was young! Bow
before him, mortals! Bow before your ending!
Surprisingly, Ragnaros bellows, TOO SOON! YOU HAVE AWAKENED ME TOO SOON,
EXECUTUS! WHAT IS THE MEANING OF THIS INTRUSION???
These mortal infidels, my lord! They have invaded your sanctum and seek to steal your
secrets!

54 http://dx.doi.org/10.2304/elea.2009.6.1.54
Visualization of Expert Chat Development

FOOL! YOU ALLOWED THESE INSECTS TO RUN RAMPANT THROUGH THE

HALLOWED CORE? AND NOW YOU LEAD THEM TO MY VERY LAIR? YOU HAVE FAILED
ME, EXECUTUS! JUSTICE SHALL BE MET, INDEED!
With that, Ragnaros slays Majordomo Executus with a flaming ball of fire.
NOW FOR YOU, INSECTS! BOLDLY, YOU SOUGHT THE POWER OF
RAGNAROS. NOW YOU SHALL SEE IT FIRSTHAND!
The raid leader, unfazed, yells, ATTACK! and a flurry of activity commences. Within
moments, the raiders are all dead.
This event was experienced repeatedly by a group of players in the massively multiplayer
online game (MMO) World of Warcraft (WoW) (Blizzard Entertainment, 2004) who delved into
Molten Core (MC) weekly for a period of about 10 months. The research in this article comes out
of an ethnography of this group and attempts to document the collaborative learning and expertise
development found within the group.

Expertise and Change in Expert Practice

Many previous studies on expertise have treated expertise development as one of individual
cognition by focusing on differences between novices and experts in school or professional settings
(e.g. Chi et al, 1981; Norman et al, 2006). Yet, expertise development occurs in all the domains of
activity in which people participate. Looking at expertise development in these various settings is
important for understanding consequential learning across settings. Recently, one domain that has
seen a rise in expertise research is that of games studies (e.g. Squire, 2005). Instead of treating
expertise development as within the confines of the individual, other scholars have been rethinking
expertise as inextricably tied to local settings and social situations (e.g. Hutchins, 1995a; Bell, et al,
2006; Collins & Evans, 2007). Expertise development is a sociocultural process, and acquiring
expertise is, as Collins & Evans (2007) note, a matter of socialization into the practices of an expert
group (p. 3).
The practices of an expert group, however, also change over time, especially with a newly
formed group that has to learn effective methods to succeed or with an established group
encountering new challenges. Groups of players of World of Warcraft who join together to tackle
shared tasks are one such type of group that has to learn and develop new forms of expert practice.
Studying these groups and other informal learning groups may help us understand the elements of
group dynamics necessary for success in all the various ways expertise may be defined.
Participating in the group (also known as a raid) gives a clearer sense of the kinds of in-game talk
and practice happening, how they changed over time, and how this change related to expertise
development at the level of the group. Additionally, being in the trenches led to questioning
whether all players of the raid could and did participate equally.
This article documents an attempt to confirm initial thoughts about the ways in which a
particular raid groups practice evolved and whether certain categories of players participated (that
is, used in-game chat) less than others did. Through visualizations (Tufte, 1983, 1990, 1997), a
comparison of chat logs from two different nights in the in-game dungeon MC is made to examine
patterns in the main form of communication that could not readily be seen before. This allowed
further theorizing about the nature of chat during high-end game play. Two key examples are the
focus of this articles analysis: (1) the first time the group encountered Ragnaros, the last monster in
MC, but was unable to defeat him, and (2) an encounter three months later, which depicts only the
second time the group was finally able to defeat Ragnaros, having done so once the week before.
The difference in how chat was used gives evidence to how the groups actions and practices as
raiders changed to enable them to succeed.

Methods
Larger Study
This visualization study is part of a larger research project that documents the practice of two
particular expert player groups in World of Warcraft from fall 2005 to fall 2006. The core of the
research methods was ethnographic in nature, and, in fact, the author gained access to and was

55
Mark Chen

playing with the two player groups before researching WoW. Similar to the work of other games
scholars who write about their gaming groups (Taylor, 2006; Steinkuehler, 2007), this research
brings a disciplined knowledge of gaming to orient to the kinds of activities and meaning-making
taking place in and around the game. It attempts to document the practices of the two raid groups
through a distributed cognition (Hutchins, 1995b) lens, mapping the learning pathways of the
groups as systemic wholes, which includes their use of technological/material resources.
The bulk of this research data is text and voice chat, supplemented with the discourse these
player groups engaged in through online forums and other out-of-game arenas. Visualizing the chat
data is complementary and supplemental to a deeper read of the discourse during play.

Setting
World of Warcraft follows a tradition of role-playing games loosely based on Dungeons & Dragons
(Gygax & Arneson, 1974; Wizards of the Coast, 2008) set in a Tolkienesque fantasy world full of
exotic locales, aggressive monsters, and glory to be had. Players create a character to play by
choosing its class (warrior, rogue, etc.), race (human, orc, etc.), and sex (see Figure 1). Character
class and race determine ones initial attribute values (strength, agility, etc.), which, in turn, shape
the available abilities or actions one can perform. The abilities from one class complement those
from a different class, encouraging players to team up and cooperate. As a player journeys through
the land with his or her character, completing quests and defeating monsters, the character gains
experience points (XP). After a certain amount of XP, the character advances an experience level
and becomes more powerful through a rise in his or her attribute values and access to new abilities.
Additionally, the corpses of defeated monsters can be searched for valuable items (known as loot)
that help characters outfit themselves and be better prepared for future encounters. Some loot, for
example, is enchanted and gives additional bonuses to a characters attributes.

Figure 1. World of Warcraft character creation screen showing the different

races and classes players could choose to play. Featured is a male orc rogue.

56
Visualization of Expert Chat Development

During the time of data collection for this project, WoW had a level cap of 60, which means that
characters started out at level one and could only advance to level 60, at which point no more XP
could be gained. (The level cap at the time of this writing is now 80.) Once reaching level 60, the
only way to improve ones character was to join a raid that went to endgame dungeons (such as
the example which introduced this article) to kill the monsters within for the loot they dropped.
Once formed, these expert groups must learn how to work collectively and to coordinate with each
other on team-based activities. As part of this study, the author joined a 40-person raid group for
participant observation that met each week within the game to delve into the dungeon known as
Molten Core (MC) for a period of roughly 10 months (October 2005-July 2006).

Figure 2. World of Warcraft in-game location known as Molten Core and some of the monsters found within.

Molten Core is a volcanic cave located in a fiery, barren landscape of WoW. The sounds of lava
flows and rushing hot air provide steady background noise as raid groups fight the monsters inside.
These monsters include several big boss monsters with names like Majordomo Executus and
Ragnaros and many more generic monsters like rocky Molten Giants and two-headed Core
Hounds (see Figure 2). Like all WoW monsters, each monster in MC has a set of abilities it uses
when fighting. For example, Molten Giants have a Stomp ability that damages everyone around
them. Part of successfully raiding a dungeon, then, means learning effective approaches to each
encounter.
The raid group examined for this analysis met twice a week for roughly seven months and
then just once a week for three months as they became more efficient in their monster killing. Each
session lasted about five hours, and each week the group would attempt to kill as many of the boss
monsters as possible before the dungeon reset every Tuesday. That is to say, every week the group
would start anew because MC would be set back to its initial state and all of the bosses would
reappear. This mechanic was deliberately designed into the game to allow groups to achieve
progress through iterative attempts to clear the dungeon. Some of the regular monsters, however,
reappeared (known as repopping or respawning) after a few hours, making backtracking during a
session difficult. Only after seven or eight months of attempts was the group able to clear the
dungeon completely before it reset the following week.
The last three months of this 7-8 month period were spent achieving the ultimate goal of
raiding MC, defeating the last boss monster, Ragnaros (see Figure 3) and collecting the epic loot
he drops. In sum, then, it took the group five months of regular practice in MC to learn how to kill
efficiently the monsters before Ragnaros. Then it took another three months to learn how to

57
Mark Chen

execute successfully the Ragnaros fight. Like most WoW boss monsters, when he dies, Ragnaros
only drops three or four items. This meant that raid groups typically continue to visit the dungeon
to defeat Ragnaros multiple times even after they have solved the complex problem of his defeat
so that every raid member, in theory at least, can receive a loot reward. Various incentives are used
by different groups in order to insure that raid members do not just leave the group once they
received the loot they were personally after. One common method is to award points for
participation that can then be spent in in-group loot auctions. The most famous of this system is
called dragon kill points or DKP (Malone, 2007). Another method, a laid back one used by the
group under study, is by randomly awarding raid members with loot, giving veteran raiders
weighted chances of winning, and emphasizing the groups values of friendship over loot (Chen,
2009).

Figure 3. Ragnaros, the last boss monster in Molten Core. The skeletal remains
of the raiders littering the floor around him are good indicators of his size.

Analysis
Multiple changes to both practice and talk needed to occur for the group to defeat Ragnaros. What
follows are four pairs of stacked bar charts (Figures 4, and 7-9) showing patterns in the text chat
that was happening on two different nights in MC, focusing on the Ragnaros fight. The two nights
represent the raid groups first serious attempt at fighting Ragnaros on February 24, 2006, and the
second time the raid was able to defeat him on May 19, 2006. Sadly, no data exists for the first time
the group defeated Ragnaros the week before.
Each session has been cropped to include only the time spent preparing for, fighting, and then
debriefing the encounters with Ragnaros. On the first night, the raid group attempted to defeat him
four times before it gave up. After each defeat, the raid had to spend time to resurrect itself,
regroup, and try again. On the second night examined, the raid finally killed Ragnaros on its third
attempt. In the first case, preparations for the fight with Ragnaros began at 9:15 p.m. and the night
ended at 10:51 p.m. In the second case, the Ragnaros battle was longer, starting at 7:37 p.m. and
ending at 9:32 p.m. These two times approximately one and a half hours and two hours

58
Visualization of Expert Chat Development

respectively make up the last part of these two game sessions. About two more hours were spent
clearing regular monsters before getting to the activity represented in the charts.
For most of the sessions, voice chat chatter is relatively constant but the gray bands show
specifically moments when the majority of voice chat focused on strategy or on-task talk about the
Ragnaros fight. For both gaming sessions, the voice chat was not audio recorded but field notes
were taken.
Comparing the two cases across Figures 4 to 9, we can clearly see more participation turns in
the second, successful case. Participation turns, Iacono & Weisband (1997) suggest, can show
evidence of trust within the group. The amount of activity needed from each specialized role and
the way in which many of these activities depended on coordinated simultaneous action was such
that it was impossible for any single player to keep track of what others were doing in the moment
without their own responsibilities suffering. Therefore, trust among group members was needed
for the raid to function efficiently. This trust only existed, however, when players were confident
that other players were playing their roles well.
It is clear that the second night did indeed contain more turns of talk, suggesting that it is
possible that trust was increased, reflecting the behavior of a more confident group. If one looks at
the kind of talk happening in the two nights, however, it becomes evident that more is happening
than simply a rise in participation turns. In fact, when the talk is broken down into on-task or
general chatter (Figure 7), it becomes clear that the level of strategy talk stays the same or lowers.
The need for strategy talk lessens with experience, allowing for more banter among the group
members and reflecting the general comfort group members have gained with their shared activity.
Also, an anomaly appears in the second case where the time between the second and third
attempts to defeat Ragnaros is longer than the time between other battle sections. Here, the group
was unable to revive itself after the second fight, a task normally done by someone known as a safe
rezzer usually a character who can self-resurrect (rez) through an item called a soulstone and
then resurrect others. Soulstones can be made only once every 30 minutes, meaning that the group
had a limited rate of reviving itself.
After the first attempt to kill Ragnaros and while the raid was being revived, someone
accidentally attracted Ragnaross attention. He then proceeded to kill everyone again, (see the
green band in the second chart across Figures 4-9), forcing another safe rezzer to use his or her self-
rezzing ability (i.e. soulstone). By the time the group failed to kill Ragnaros a second time, it had
run out of safe rezzers, meaning that the members would all have to release their in-game
characters from their corpses, appearing as ghosts at the nearest graveyard to laboriously run back
to the beginning of the dungeon to reclaim their bodies. Not having a safe rezzer was a huge
setback since some of the regular monsters in MC had reappeared, requiring the group to spend
additional valuable time killing them yet again to get back to Ragnaross location.
Eventually, the group decided to have a couple of members who were rogues a character
class that can sneak around undetected by monsters provided they dont stray too close attempt
to steal their way into Ragnaross chamber and use an alternative item to bring the safe rezzer back
to life (see the blue band in the second case in Figures 4-9). One of the rogues died on the way but
the second succeeded, enabling the raid to revive itself in Ragnaross chamber and attempt
(successfully this time) to kill him one last time. It is possible that more chatter occurred in this
second case than would have normally happened since members of the raid gave remarks of
appreciation to the rogue who was able to sneak to Ragnaros.
What follows are analyses of the chat based on more specific categories and coding schemes
to highlight additional patterns in the talk that are indicative of the development of expertise more
specifically.

Comparing Chat Patterns: chat channel used

The first pair of charts (Figure 4) shows the timeline view of the number of utterances per chat
channel (differentiated by color) during the raid. These charts clearly show that most of the talk
occurred in the [Raid] channel (orange), a common channel that all raid members could see. Areas
where there were high frequencies of [Yell] (light gray) utterances coincide with actual fights with
Ragnaros because most of the yells that occurred during our sessions were the raid leader yelling

59
Mark Chen

commands such as ATTACK! during the fights. Ragnaros also yelled during the fights (such as
DIE, INSECT! any time one of the raiders died) but yells from non-human actors were not
included in the charts.
A second pattern that stands out in the first pair of charts (Figure 4) is the high level of talk in
the [madrogues] channel (yellow). The [madrogues] channel is a specialized, private channel that
all members of the rogue class were able to subscribe to enabling them to talk about rogue-specific
strategies only amongst themselves without disrupting the entire groups process with their
chatter. Each role within this raid group established a corresponding channel specialized to their
specific class (and therefore function). In these data cases, only the rogue channel was collected;
however, one can imagine similar, parallel levels of chat occurring in other, similarly specialized
channels (see Chen, 2009).

Figure 4. Chat visualization from two contrasting cases in the

Ragnaros fight in World of Warcraft, broken down by chat channel.

Third, the brown areas in Figure 4 generally show moments of participants emoting with their
characters exchanging typed actions (e.g. hugging another character, flexing ones muscles) that
appear in-game as non-verbal lines of chat (e.g. as [character name] hugs you or [character name]
flexes his muscles). Such emotes also include players rolling (using an in-game mechanism to
generate a random number) for loot that dropped from Ragnaros after the final, successful fight.
The corresponding flurry of emoting activity is represented by the large brown areas toward the
latter end of the second timeline.
Finally, the unusual pattern in the [Whisper] channel (pink) requires additional explanation.
Whispers are private chat between two players. Only the two players exchanging whispers (sender
and recipient) can see the chat; therefore, all the whispers that were captured are only between
another player and the author. On the first night at 22:00 (10 p.m.), whispers were exchanged with
the raid leader prior to the first fight with Ragnaros to clarify where the rogues were supposed to
be positioned during the Ragnaros fight. Examination of the content from the chat logs of the
[madrogues] channel reveals a similar confusion as to the appropriate location of the rogue
characters:
21:42:43.156 [madrogues] Thoguht: why do we start here again instead of on the other side?
21:42:44.500 [Raid] Maxwell: we go in 15 seconds
21:42:45.781 [Raid] Maxwell: <CTRaid> Maxwell has performed a ready check.
21:42:55.062 [madrogues] Rover: tis a mystery
21:42:56.531 [madrogues] Rebecca: no idea
21:43:08.750 [Raid] Maxwell: SUMMON RAGS
21:43:13.968 [madrogues] Thoguht: are we supposed to melee on the west side?

60
Visualization of Expert Chat Development

21:43:21.875 Majordomo Executus yells: Impudent whelps! Youve rushed headlong to your own
deaths! See now, the master stirs!
Unfortunately, this confusion was happening right as the raid group was attacking Maxwell, the
raid leader, had signaled for the start of the event (SUMMON RAGS) so there was no time to
clarify positions, and the conversation on [madrogues] did not continue since the rogues were
focused on more pressing matters. After the first failure, it became clear that the rogues were not in
optimal places (see Figure 5); by the time the raid took down Ragnaros three months later, the
rogues had learned where to position themselves without the earlier need for discussion and
clarification (see Figure 6).

Figure 5. Screen shots from the first evening that failed, showing the rogues starting positions and their movement to
final positions for the Ragnaros fight. They are spread out and not in an optimal location, which led to their early deaths.

61
Mark Chen

Figure 6. Screen shots from the second evening that succeeded, showing that, months later, the rogues have learned
exactly where to stand during the fight, evident through their bunching up and standing on top of each other.

62
Visualization of Expert Chat Development

Comparing Chat Patterns: on-task talk

Coding the different chat utterances based on on-task and off-task talk uncovers that more on-task
talk (sub-categorized as strategy, management, buffing, and CTRA [CT Raid Assist: see below]
talk) happened during the successful case than during the first case. Specifically, in Figure 7, pink
areas indicate strategy-related talk (e.g. argumentation on different approaches to an upcoming
fight or notifications of events during a battle), and purple areas indicate buff actions or talk about
buffs. A buff is a temporary bonus given by one character to another that improves a characters
abilities in some way. Many character classes have the ability to buff others but they often have to
choose one buff to activate from a selection of buffs they have available.

Figure 7. Chat visualization from two contrasting cases, broken down by type of on-task talk.

As Figure 7 shows, buffing in both cases tended to happen between fights or after all raid members
had to be revived after a failed attempt (because death removes buffs). It is clear, however, that
while strategy talk remained at fairly consistent or lower levels, much more buffing was happening
in the second case. It is expected, as stated earlier, that, as a group develops expertise, they will
need to talk about coordination strategy less and instead will simply do whatever actions are
needed to complete the task. In this second case, raid members have learned exactly which buffs to
cast or apply on each other, and they do this buffing more or less automatically.
Also coded are lines on the raids management or maintenance, such as inviting new raid
members, as well as lines that were generated by an add-on called CT Raid Assist (CTRA) for its
internal maintenance. Add-ons are third-party modifications to the games user interface that
players can install to streamline or otherwise enhance their play experience. Most add-ons reveal
previously hidden aspects of the game, making clear, for example, the numbers behind certain
abilities. CTRA keeps track of raid-specific issues, such as notifying players when a boss event is
about to happen, but for it to work, all players must have it installed. At the time of this data
collection, all raid groups that the author had participated with or had heard of from other players
used CTRA.
A review of the content of the chat logs reveals that much more strategy talk occurred during
the fights themselves on the successful night because of the time-dependent nature of the fight.
Below is an excerpt of the chat from 20:13, the minute in the second case with the most utterances
of strategy talk:
21:13:03.000 Maxwell yells: ATTACK!
21:13:03.140 [Raid] Maxwell: ATTACK!
21:13:07.421 Ragnaros yells: TASTE THE FLAMES OF SULFURON!
21:13:07.671 [Raid] Maxwell: *** AE Knockback ***

63
Mark Chen

21:13:10.250 [Raid] Maxwell: *** Ragnaros has Emerged. 3 minutes until submerge. ***
21:13:12.484 Ragnaros yells: DIE, INSECT!
21:13:12.734 [Raid] Roger: 33%
21:13:20.937 [madrogues] Roger: dump on him, just dont draw aggro.
21:13:24.328 Rand has died.
21:13:38.343 Ragnaros yells: BY FIRE BE PURGED!
21:13:38.890 [Raid] Pall: out of mana
21:13:41.671 Ragnaros yells: TASTE THE FLAMES OF SULFURON!
21:13:42.312 [Raid] Maxwell: *** AE Knockback ***
21:13:56.921 [Raid] Paula: down
21:13:57.671 [Raid] Mandy: 1m 42
The battle with Ragnaros occurred in timed phases. By the time of the successful battle (the second
case), the group had learned the timing of the fight, and new practices had emerged. Through the
use of an add-on, Maxwell was able to call out specific events (e.g. *** AE Knockback ***) in
addition to his previous practice of calling out directions (e.g. ATTACK!) The raid also adopted
the use of one raid member as a timekeeper who called out in the [Raid] channel how much time
was left before each next phase (Mandy: 1m 42). Upon entering each new phase and at other
moments of significance, our raid leader also called out a reminder so that the rest of the raiders
could prepare to do whatever new actions were required in the new time phase. These new
practices were only taken up by the raid through experience with the fight and its synthetic use of
non-human resources such as add-ons, and these changes are reflected as patterns in the visualized
chat data.

Comparing Chat Patterns: raid leader talk

In both examples examined, the raid group had one main leader as well as three other players who
were sub-leaders. Coding each of their chat lines (Figure 8) shows how much talk was made by raid
leaders (dark blue) versus regular raid members (light blue).

Figure 8. Chat visualization from two contrasting cases with raid leader talk differentiated from raid member talk.

These charts were needed to explore whether an unexpected pattern in raid leader talk occurred.
They confirm, however, that raid leader talk rose during fighting sections, reflecting the calling out
occurring during encounters. As discussed in the previous section, most of the raid leader talk
consisted of calling out specific events. In the first case, however, these calling out moments were
restricted to common events found in all fights in World of Warcraft, such as when the group would

64
Visualization of Expert Chat Development

start the encounter (ATTACK!) The second case clearly shows greater raid leader activity. Again,
going back to the chat log itself reveals that this is explained by the leaders new practice of calling
out events that were specific to the Ragnaros fight.

Comparing Chat Patterns: gender

Figure 9 shows talk by women players (green areas) versus talk by men (purple). Note that, here,
gender differentiation was based on voice samples only as no off-screen demographic information
was collected from players (miscategorization here is very possible although it should be noted that
there were no border cases where categorization appeared troubling). One of the main reasons for
the data visualizations presented here was an informal hypothesis of disproportionate talk based on
gender during raids. Specifically, while the data was being collected, there was a suspicion that
women were participating with lower frequency than men were. Figure 9 confirms this suspicion.
In the first case, 11 women and 31 men were in the raid (including substitutions during the raid
process), meaning that women constituted roughly 26% of the raid participants, yet, women
account for only 16.7% of the talk. The second night shows an even bigger imbalance, with women
constituting roughly 22% of the raid participants (9 women and 32 men total) yet accounting for
only 7.5% of the participation. Full details can be seen in Table I.

Figure 9. Chat visualization from two contrasting cases, broken down by gender.

Women Men % women

Feb 24 Players 11 31 26.2

(Failure)
Chat utterances 125 623 16.7
Chat utterances w/o leader data 91 455 16.7
May 19 Players 9 32 22.0
(Success) Chat utterances 108 1323 7.5
Chat utterances w/o leader data 88 851 9.4

Table I. Percentage of participants who were women

(based on voice sample) versus percentage of chat made by women.

The raid leader, who accounted for more of the talk than anyone else (see Figure 8), was male; two
of the raid sub-leaders were women. Excluding the raid leader as well as the sub-leaders, the
percentage of talk by women does not change for the first example raid examined but does change

65
Mark Chen

for the second evening slightly (7.5% versus 9.4% respectively). For the first example, in other
words, talk from raid leaders does not confound the ratio of talk in the whole raid in general. For
the second example, however, because the main raid leader had adopted the practice of yelling out
specific events, his portion of the chat was greater. In both cases, it is clear that the initial suspicions
were correct and that women were participating in the raid chat disproportionately so. It is possible
that women in these data used private channels more than men did, although we have no data as
yet to support this claim.

Discussion
Such visualization charts have limitations. They only show general patterns without going into the
intricate details of raiding practice. For example, we can see talk in the various channels but we
have no clear sense of how each line of talk relates to previous lines, whether a conversation or
argumentation is happening, etc. Without such additional analyses, it is difficult to see the
distributed yet collective nature of the raid group or the overall joviality and tone of the raid
sessions. Additional developments within the group between the first unsuccessful raid and the
second successful example are also not accounted for. For example, during the interim the group
adopted the use of a specific third-party add-on to lessen cognitive load (Sweller, 1988), keeping
track of relational data during fights (see Figure 6). Additionally, the group developed expert
practices along the way that are simply not represented by analyses of chat transcripts alone. These
include assigning specialized roles to certain players (e.g. timekeeper), using outside material
resources (such as web strategy guides for boss fights), or collaborating in online forum discussion
of strategies for maximizing efficiency (a common practice within MMOs like WoW known as
theorycrafting).
Still, such visualizations do offer us particular insight into important dynamics of developing
expertise within groups such as raiding guilds, and they offer a quick way to see patterns happening
in the data. Such visualizations should be seen as complementary to a careful analysis of the chat
itself.
Several insights came from the charts in this article. First, it is obvious that more talk occurred
during the second case, yet the amount of strategy-specific talk did not rise. Instead, we see a rise in
buffing and general chatter or banter, indicating that the group had more experience and expertise
with the shared activity. The raid members no longer had to spend as much time focused on the
task; parts of the task had become automated.
Second, it is surprising that the time spent fighting and preparing for fights in the successful
raiding case actually took longer than in the first case because often experts are seen as being able
to complete tasks faster than novices can. This time extension, however, does make complete sense
during the fights; the raid members were not dying within moments and therefore the fights lasted
longer. The longer preparation time is interesting, and much of it was spent by the raid leader
describing the encounter and each of the class roles during the fight.
Third, during the second case, more strategy talk was spent on fight-specific events and less
was spent on clarification and negotiation of roles. In other words, there was less speculation and
more automated, declarative responses to specific situations. This is evident in a general pattern of
an increase in strategy talk during fights and a slight decrease between fights. Much of this increase
during fights can be attributed to the new practice of the raid leader calling out specific events
during the activity that elicited reaction by the other raid members. Part of expertise is recognizing
patterns accurately and responding appropriately. This group clearly became able to react to
familiar situations. The fight did not play out the same way each time; it differed slightly based on
the raid groups composition and random luck, as the damage done by the groups attacks and
Ragnaross strike backs varied. Thus being able to recognize the pattern of the fight and adapt to
micro-differences was important for the group to succeed.
Finally, initial notions about the disproportionate talk by women have been confirmed.
Unfortunately, it is unclear why this was happening. This might be a normal, benign phenomenon,
or it could be evidence of an unequal situation. When considering studies on group entities, it is
sometimes possible to lose track of the individuals making up the group and to lose a critical focus

66
Visualization of Expert Chat Development

on the inner workings of the group. Armed with this knowledge, we can be better prepared for
future studies.

Acknowledgement
This work is funded by the National Science Foundation through the Science of Learning Center
program under grant SBE-0354453.

References
Adobe Systems ([1987] 2008) Photoshop [computer software]. San Jose: Adobe Systems.
Bell, P., Bricker, L.A., Lee, T., Reeve, S. & Zimmerman, H.T. (2006) Understanding the Cultural Foundations
of Childrens Biological Knowledge: insights from everyday cognition research, in S.A. Barab, K.E. Hay
& D.T. Hickey (Eds) Proceedings of the Seventh International Conference of the Learning Sciences, 1029-1035.
Mahwah: Erlbaum.
Blizzard Entertainment (2004) World of Warcraft. Irvine, CA: Blizzard Entertainment.
Chen, M.G. (2009) Communication, Coordination, and Camaraderie in World of Warcraft, Games and Culture,
4, 47-73. http://dx.doi.org/10.1177/1555412008325478
Chi, M.T.H., Feltovich, P. & Glaser, R. (1981) Categorization and Representation of Physics Problems by
Experts and Novices, Cognitive Science, 5, 121-152. http://dx.doi.org/10.1207/s15516709cog0502_2
Collins, H. & Evans, R. (2007) Rethinking Expertise. Chicago: Chicago University Press.
Goodwin, C. (1994) Professional Vision, American Anthropologist, 96, 606-633.
http://dx.doi.org/10.1525/aa.1994.96.3.02a00100
Gygax, G. & Arneson, D. (1974) Dungeons & Dragons [table-top game]. Published under Tactical Studies
Rules, Inc. (TSR).
Horney, M. (1994) Interactive Data Visualization in Qualitative Research, Computer Graphics, 28(1), 38-40.
http://dx.doi.org/10.1145/181505.181513
Hutchins, E. (1995a) Cognition in the Wild. Cambridge, MA: MIT Press.
Hutchins, E. (1995b). How a Cockpit Remembers Its Speeds, Cognitive Science, 19, 265-288.
Iacono, C.S. & Weisband, S. (1997) Developing Trust in Virtual Teams, in Proceedings of the Thirtieth Hawaii
International Conference on System Sciences, vol. 2, 412-420. Los Alamitos, CA: IEEE Computer Society
Press.
Malone, K.M. (2007) Dragon Kill Points: the economics of power gamers. http://ssrn.com/abstract=1008035
(accessed January 7, 2008).
Marcelionis, A. (2008) amCharts [computer software]. http://www.amcharts.com
Microsoft Corporation ([1985] 2007) Excel [computer software]. Redmond, WA: Microsoft Corporation.
Norman, G., Eva, K., Brooks, L. & Hamstra, S. (2006) Expertise in Medicine and Surgery, in K.A. Ericsson,
N. Charness, P.J. Feltovich & R.R. Hoffman (Eds) The Cambridge Handbook of Expertise and Expert
Performance, 339-354. Cambridge: Cambridge University Press.
Ruberg, L.F. & Moore, D.M. (1995) Visualizing Qualitative Data in a Study of Student Interactions within a
Computer Mediated Environment, in Imagery and Visual Literacy: selected readings from the annual
conference of the international visual literacy association 1994, 112-119.
http://eric.ed.gov/ERICDocs/data/ericdocs2sql/content_storage_01/0000019b/80/13/b4/3b.pdf
Squire, K.D. (2005) Educating the Fighter: buttonmashing, seeing, being. On The Horizon The Strategic
Planning Resource for Education Professionals, 13(2), 75-88.
Steinkuehler, C. (2007) Massively Multiplayer Online Gaming as a Constellation of Literacy Practices, E-
Learning, 4(3), 297-318. http://dx.doi.org/10.2304/elea.2007.4.3.297
Stevens, R. (2000) Who Counts What as Math? Emergent and Assigned Mathematics Problems in a Project-
Based Classroom, in J. Boaler (Ed.) Multiple Perspectives on Mathematics Teaching and Learning, 105-144.
Westport, CT: Ablex.
Sweller, J. (1988) Cognitive Load during Problem Solving: effects on learning, Cognitive Science, 12, 257-285.
Taylor, T.L. (2006) Play between Worlds: exploring online game culture. Cambridge, MA: MIT Press.
Tufte, E.R. (1983) The Visual Display of Quantitative Information. Cheshire, CT: Graphics Press.
Tufte, E.R. (1990) Envisioning Enformation. Cheshire, CT: Graphics Press.

67
Mark Chen

Tufte, E.R. (1997) Visual Explanations: images and quantities, evidence and narrative. Cheshire, CT: Graphics
Press.
Wizards of the Coast (2008) Dungeons & Dragons, 4th edn. [table-top game].

APPENDIX
Overview of Data Visualization Generation Methods
Literature regarding the specific use of visualization tools for qualitative studies in education, with
the exceptions of Horney (1994) and Ruberg & Moore (1995), is scarce. Yet, we can learn much
from methodological discussions around the use of visuals to help analyze and illustrate qualitative
data, such as those written by Goodwin (1994) and Stevens (2000). A first stab (the rogue action of
choice) at what could be discovered through visualization was initiated without knowing if
anything worthy of note would be discovered. In so doing, occlusion issues that stem from
technical data manipulation inherent in the process of making these charts became apparent. For
example, when the charts divided by gender were first created, neither leader talk nor the authors
talk was excluded. There was a bias initially towards men simply because half of the things in the
[Whisper] channel were from the author (a man). Therefore, perhaps it will be useful to take a
moment to explain how the charts in this article were created.
While playing World of Warcraft, an add-on or extension to the game was used that
downloaded text chat from within the game into external text files on a computer hard drive. The
data included: date, time, actor/speaker, and the content of the chat utterance. These log files were
then formatted and saved with a text editor that supported regular expressions a way of
matching patterns of text using various syntactically meaningful symbols. The new files included
explicit channel and actor information as separate columns (Figure A1):

Figure A1. Raw chat long on top; edited file below with information separated into columns.

For example, the following expression was used to change all lines with <character name>
whispers such that they appeared in the [Whisper] category within a channels field:
Search for: ([0-9])\t([A-Z][a-z]{1,}) whispers:
Replace with: \t[Whisper]\t\2\t
The result is that Bob whispers: how are you? is transformed into [Whisper] Bob How are you?
After the files were saved with information separated in columns, they were imported into
Microsoft Excel (Microsoft, [1985] 2007), which was used to create initial pivot tables and charts to
get a sense of what was possible (E. Speckman, personal communication, November 2008). It was
clear that visualizing the data might prove useful, yet there did not appear to be any easy way of
coding each line with participant information such as alias, gender, or character class. To resolve

68
Visualization of Expert Chat Development

the issue, the files were imported into a MySQL database and an additional table (aliases) was
created containing only participant metadata. By creating this relational database, linking
information from one table to another was possible using the SQL JOIN command. For example,
the following code was used to output a new file that included all of the data in the chat logs plus
information about the actor on each chat line (output seen in Figure A2):
SELECT 20060519.Time, 20060519.Channel, aliases.Alias, aliases.Class, aliases.SexReal,
aliases.Leader, 20060519.Chat
FROM 20060519 JOIN aliases ON (20060519.Name = aliases.Name)

Figure A2. SQL output of the joined tables (20060519 and aliases).

Queries were then run on the database counting the number of times something occurred per
minute and saved as comma separated values (CSV) text files. Each query counted something
different, such as a specific chat channel or the number of times a woman spoke. For example, the
query used to count the number of times a rogue spoke was the following:
SELECT 20060519.Time, Count(20060519.Time)
FROM 20060519 JOIN aliases ON (20060519.Name = aliases.Name)
WHERE 20060519.Channel NOT LIKE [Guild] AND aliases.Class LIKE Rogue
GROUP BY 20060519.Time
This produced rows formatted with just two columns: the minute and the number of times a rogue
spoke in that minute.
Files were aggregated and manipulated in Excel so that, for example, the file for character class
included columns for each class. The resulting table (see Figure A3) was saved as a flat CSV file and
then read by a Flash-based web application, amCharts (Marcelionis, 2008) to make the timeline
charts. Their production required editing and saving HTML and settings files with a text editor for
each chart pair. Screenshots of the charts were then taken and edited with Adobe Photoshop
(Adobe, [1987] 2008), adding the hanging color bands and notes.

Figure A3. The first few rows of the

class text file read by amCharts.

69
Mark Chen

After creating several charts in this fashion, it became clear that some of the talk patterns were
confounded by whether the talk was on-task, so the tab-delimited log files from the first regular
expression step were reformatted to include codes that flagged strategy talk, talk about buffs, etc.
These new files were then run through the database scrubber to create the CSV files needed for the
charting software.
Ironically, the process described above may be just as time consuming as a standard analysis
of chat discourse. Eventually, the hope is that some way of automating this process can create
multiple charts quickly to help the researcher identify patterns and anomalies to study as a way of
supplementing the necessary work of analyzing the raw chat.

MARK CHEN is a PhD Candidate in the College of Education at the University of Washington-
Seattle, USA. He writes about online game communities, everyday learning and expertise
development within them, and issues of socialization that arise out of the emergent social
structures and practices these communities adopt. Correspondence: Mark Chen, LIFE Center,
University of Washington, 1100 NE 45th Street, Suite 200, Seattle, WA 98105, USA
(markchen@u.washington.edu). Weblog: markdangerchen.net

70