This is the talk I gave at the Academica Sinica Inst. for Information Science in Taiwan. It focuses on our Wikipedia and Amazon Mechanical Turk research.
2010-02-22 Wikipedia MTurk Research talk given in Taiwan's Academica Sinica
1. Ed H. Chi
Area Manager and Principal Scientist
Augmented Social Cognition Area
Palo Alto Research Center
2. Cognition:
the
ability
to
remember,
think,
and
reason;
the
faculty
of
knowing.
Social
Cognition:
the
ability
of
a
group
to
remember,
think,
and
reason;
the
construction
of
knowledge
structures
by
a
group.
– (not
quite
the
same
as
in
the
branch
of
psychology
that
studies
the
cognitive
processes
involved
in
social
interaction,
though
included)
Augmented
Social
Cognition:
Supported
by
systems,
the
enhancement
of
the
ability
of
a
group
to
remember,
think,
and
reason;
the
system-‐supported
construction
of
knowledge
structures
by
a
group.
Citation:
Chi,
IEEE
Computer,
Sept
2008
2010-02-22 Ed H. Chi ASC Overview 2
2
3. Characteriza*on
Models
Evalua*ons
Prototypes
Characterize activity on social systems with analytics
Model interaction social and community dynamics and variables
Prototype tools to increase benefits or reduce cost
Evaluate prototypes via Living Laboratories with real users
3
2010-02-22 Ed H. Chi ASC Overview 3
4. Characterization and Modeling:
– Community Analytics and Wikipedia Dynamics
Prototyping:
– Social Transparency thru WikiDashboard
Evaluation:
– Evaluations using Amazon Mechanical Turk
4
2010-02-22 Ed H. Chi ASC Overview 4
7. Mediator
Pattern
-‐
Terri
Schiavo
Anonymous (vandals/
spammers)
Sympathetic to
husband
Mediators
Sympathetic to parents
2010-02-22 Ed H. Chi ASC Overview 7
8. Measure
of
controversy
• Controversial”
tag
• Use
#
revisions
tagged
controversial
2010-02-22 Ed H. Chi ASC Overview 8
9. Page
metrics
• Possible
metrics
for
identifying
conflict
in
articles
Metric type Page Type
Revisions (#) Article, talk, article/talk
Page length Article, talk, article/talk
Unique editors Article, talk, article/talk
Unique editors / revisions Article, talk
Links from other articles Article, talk
Links to other articles Article, talk
Anonymous edits (#, %) Article, talk
Administrator edits (#, %) Article, talk
Minor edits (#, %) Article, talk
Reverts (#, by unique
Article
editors)
2010-02-22 Ed H. Chi ASC Overview 9
19. Edits beget edits
– more number of previous edits, more number of new edits
Growth rate depends on current population N
r = growth rate of the population
N(t) = N 0 ⋅ e rt
dN
= r⋅ N
dt
Growth rate Current
of population €
population
€ 2010-02-22 Ed H. Chi ASC Overview 19
19
20. Ecological population growth model
– r, growth rate of the population
– K, carrying capacity (due to resource limitation)
4000000
K
3500000
3000000
dN N Population
2500000
= r ⋅ N ⋅ (1− ) 2000000
dt K 1500000
1000000
500000
0
2000 2002 2004 2006 2008 2010
Year
2010-02-22 Ed H. Chi ASC Overview 20
20
21. Follows a logistic growth curve
New Article
http://en.wikipedia.org/wiki/Wikipedia:Modelling_Wikipedia’s_growth
2010-02-22 Ed H. Chi ASC Overview 21
21
22. Carrying Capacity as a function of time.
K(t)
Population
2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
Year
2010-02-22 Ed H. Chi ASC Overview 22
22
23. Biological system
– Competition increases as
population hit the limits of the
ecology
– Advantage go to members of the
population that have competitive
dominance over others
Analogy
– Limited opportunities to make
novel contributions
– Increased patterns of conflict and
dominance
2010-02-22 Ed H. Chi ASC Overview 23
23
25. Highly skewed contribution pattern
– Top 3% users contribute 50%+ edits
– A lot of single-edit users
Five Editor Classes
– Monthly edit count
– No bot, vandalism included in the analysis
– 1000+: editors who made more than 1000 edits in that month
– 100-999
– 10-99
– 2-9
– 1
2010-02-22 Ed H. Chi ASC Overview 25
25
26. Monthly Edits by Editor Class (in thousands)
2010-02-22 Ed H. Chi ASC Overview 26
26
28. Monthly Ratio of Reverted Edits
2010-02-22 Ed H. Chi ASC Overview 28
28
29. Two interpretations:
– Overall increased resistance
from the Wikipedia community
to changing content
– Disparity of treatment of edits
» Occasional editors have been
reverted in a higher rate
Example of increased
patterns of conflict and
dominance
Photo: http://www.flickr.com/photos/efan78/3619921561/
2010-02-22 Ed H. Chi ASC Overview 29
29
33. “Wikipedia is the best thing ever. Anyone in the world can write
anything they want about any subject, so you know you’re getting the
best possible information.”
– Steve Carell, The Office
2010-02-22 Ed H. Chi ASC Overview 33
33
34. Content in Wikipedia can be added or
changed by anyone
Because of this, WP has become one of the
most important resources on the web
– Hundreds of thousands of contributors
– Over 2 million articles
– 5th most used websites (Alexa.com)
Also because of this, is viewed with
skepticism by readers, press, researchers
2010-02-22 Ed H. Chi ASC Overview 34
34
37. “Wikipedia, just by its nature, is
impossible to trust completely. I don't
think this can necessarily be
changed.”
2010-02-22 Ed H. Chi ASC Overview 37
37
38. Risks with using Wikipedia
– Accuracy of content
– Motives of editors
– Expertise of editors
– Stability of article
– Coverage of topics
– Quality of cited information
Insufficient information to evaluate
trustworthiness
2010-02-22 Ed H. Chi ASC Overview 38
38
39. Transparency of social dynamics can reduce conflict and coordination
issues
Attribution encourages contribution
– WikiDashboard: Social dashboard for wikis
– Prototype system: http://wikidashboard.parc.com
Visualization for every wiki page
showing edit history timeline and
top individual editors
Can drill down into activity history
for specific editors and view edits
to see changes side-by-side
Citation: Suh et al.
CHI 2008 Proceedings
39
2010-02-22 Ed H. Chi ASC Overview 39
43. Surfacing information
• Numerous studies mining Wikipedia revision
history to surface trust-relevant information
– Adler & Alfaro, 2007; Dondio et al., 2006; Kittur et al., 2007;
Viegas et al., 2004; Zeng et al., 2006
Suh, Chi, Kittur, & Pendleton, CHI2008
• But how much impact can this have on user
perceptions in a system which is inherently
mutable?
43
44. Hypotheses
1. Visualization will impact perceptions of trust
2. Compared to baseline, visualization will
impact trust both positively and negatively
3. Visualization should have most impact when
high uncertainty about article
• Low quality
• High controversy
44
45. Design
• 3 x 2 x 2 design
Controversial Uncontroversial
Visualization Abortion Volcano
High quality
• High stability George Bush Shark
• Low stability
• Baseline (none) Pro-life feminism Disk
defragmenter Low quality
Scientology and
celebrities Beeswax
45
53. Method
• Users recruited via Amazon’s Mechanical Turk
– 253 participants
– 673 ratings
– 7 cents per rating
– Kittur, Chi, & Suh, CHI 2008: Crowdsourcing user studies
• To ensure salience and valid answers, participants
answered:
– In what time period was this article the least stable?
– How stable has this article been for the last month?
– Who was the last editor?
– How trustworthy do you consider the above editor?
53
54. Results
main effects of quality and controversy:
• high-quality articles > low-quality articles (F(1, 425) = 25.37, p < .001)
• uncontroversial articles > controversial articles (F(1, 425) = 4.69, p = .
031)
54
55. Results
interaction effects of quality and controversy:
• high quality articles were rated equally trustworthy whether controversial
or not, while
• low quality articles were rated lower when they were controversial than
when they were uncontroversial service.
55
56. Results
1. Significant effect of
visualization
– High > low, p < .001
2. Viz has both positive and
negative effects
– High > baseline, p < .001
– Low > baseline, p < .01
3. No interaction of
visualization with either
quality or controversy
– Robust across conditions
56
57. Results
1. Significant effect of
visualization
– High > low, p < .001
2. Viz has both positive and
negative effects
– High > baseline, p < .001
– Low > baseline, p < .01
3. No interaction of
visualization with either
quality or controversy
– Robust across conditions
57
58. Results
1. Significant effect of
visualization
– High > low, p < .001
2. Viz has both positive and
negative effects
– High > baseline, p < .001
– Low > baseline, p < .01
3. No interaction effect of
visualization with either
quality or controversy
– Robust across conditions
58
60. User studies
• Getting input from users is important in HCI
– surveys
– rapid prototyping
– usability tests
– cognitive walkthroughs
– performance measures
– quantitative ratings
61. User studies
• Getting input from users is expensive
– Time costs
– Monetary costs
• Often have to trade off costs with sample size
62. Online solutions
• Online user surveys
• Remote usability testing
• Online experiments
• But still have difficulties
– Rely on practitioner for recruiting participants
– Limited pool of participants
63. Crowdsourcing
• Make tasks available for anyone online to complete
• Quickly access a large user pool, collect data, and
compensate users
• Experiences at PARC:
– CSL UbiComp group
– ISL’s NLTT group
64. Crowdsourcing
• Make tasks available for anyone online to complete
• Quickly access a large user pool, collect data, and
compensate users
• Example: NASA Clickworkers
– 100k+ volunteers identified Mars craters from
space photographs
– Aggregate results “virtually indistinguishable” from
expert geologists
experts
crowds
http://clickworkers.arc.nasa.gov
65. Amazon’s Mechanical turk
• Market for “human intelligence tasks”
• Typically short, objective tasks
– Tag an image
– Find a webpage
– Evaluate relevance of search results
• Users complete for a few pennies each
67. Using Mechanical Turk for user studies
Traditional user Mechanical Turk
studies
Task complexity Complex Simple
Long Short
Task subjectivity Subjective Objective
Opinions Verifiable
User information Targeted demographics Unknown demographics
High interactivity Limited interactivity
Can Mechanical Turk be usefully used for user studies?
68. Task
• Assess quality of Wikipedia articles
• Started with ratings from expert Wikipedians
– 14 articles (e.g., “Germany”, “Noam Chomsky”)
– 7-point scale
• Can we get matching ratings with mechanical turk?
69. Experiment 1
• Rate articles on 7-point scales:
– Well written
– Factually accurate
– Overall quality
• Free-text input:
– What improvements does the article need?
• Paid $0.05 each
70. Experiment 1: Good news
• 58 users made 210 ratings (15 per article)
– $10.50 total
• Fast results
– 44% within a day, 100% within two days
– Many completed within minutes
71. Experiment 1: Bad news
• Correlation between turkers and Wikipedians
only marginally significant (r=.50, p=.07)
• Worse, 59% potentially invalid responses
Experiment 1
Invalid 49%
comments
<1 min 31%
responses
• Nearly 75% of these done by only 8 users
72. Not a good start
• Summary so far:
– Only marginal correlation with experts.
– Heavy gaming of the system by a minority
• Possible Response:
– Can make sure these gamers are not rewarded
– Ban them from doing your hits in the future
– Create a reputation system [Delores Lab]
• Can we change how we collect user input ?
73. Design changes
• Use verifiable questions to signal monitoring
– “How many sections does the article have?”
– “How many images does the article have?”
– “How many references does the article have?”
74. Design changes
• Use verifiable questions to signal monitoring
• Make malicious answers as high cost as
good-faith answers
– “Provide 4-6 keywords that would give someone a
good summary of the contents of the article”
75. Design changes
• Use verifiable questions to signal monitoring
• Make malicious answers as high cost as
good-faith answers
• Make verifiable answers useful for completing
task
– Used tasks similar to how Wikipedians described
evaluating quality (organization, presentation,
references)
76. Design changes
• Use verifiable questions to signal monitoring
• Make malicious answers as high cost as
good-faith answers
• Make verifiable answers useful for completing
task
• Put verifiable tasks before subjective
responses
– First do objective tasks and summarization
– Only then evaluate subjective quality
– Ecological validity?
77. Experiment 2: Results
• 124 users provided 277 ratings (~20 per article)
• Significant positive correlation with Wikipedians (r=.
66, p=.01)
• Smaller proportion malicious responses
• Increased time on task
Experiment 1 Experiment 2
Invalid 49% 3%
comments
<1 min 31% 7%
responses
Median time 1:30 4:06
78. Generalizing to other user studies
• Combine objective and subjective questions
– Rapid prototyping: ask verifiable questions about
content/design of prototype before subjective
evaluation
– User surveys: ask common-knowledge questions
before asking for opinions
79. Limitations of mechanical turk
• No control of users’ environment
– Potential for different browsers, physical
distractions
– General problem with online experimentation
• Not designed for user studies
– Difficult to do between-subjects design
– Involves some programming
• Users
– Uncertainty about user demographics, expertise
80. Conclusion
• Mechanical Turk offers the practitioner a way to
access a large user pool and quickly collect data at
low cost
• Good results require careful task design
1. Use verifiable questions to signal monitoring
2. Make malicious answers as high cost as good-faith
answers
3. Make verifiable answers useful for completing task
4. Put verifiable tasks before subjective responses
81. Ed
H.
Chi
(manager,
PS)
Peter
Pirolli
(RF)
Lichan
Hong
Bongwon
Suh
Les
Nelson
Rowan
Nairn
Gregorio
Convertino
Interns/Collaborators:
Sanjay
Kairam,
Jilin
Chen
(UMinn),
Michael
Bernstein
(MIT)
http://asc-‐parc.blogspot.com
Ed H. Chi ASC Overview 81
2010-02-22
83. r, growth rate dN N
= rN(1− )
K, carrying capacity dt K
4000000
3500000
3000000
€ K dominates
2500000 r dominates when N K
2000000
when N is small N
1500000
N (1− ) ≈ 0
1000000
(1− ) ≈ 1 K
500000 K
0
2000 2002 2004 2006 2008 2010
Year
2010-02-22 Ed H. Chi ASC Overview 83
€ €
84. r-Strategist
– Growth or exploitation
– Less-crowded niches / produce many offspring
K-Strategist
– Conservation
– Strong competitors in crowded niches / invest more heavily in
fewer offspring
Evolution cycle
– Resilience of an ecological system
– Gunderson & Holling 2001
2010-02-22 Ed H. Chi ASC Overview 84
85. Exponential growth model dN
– Growth rate depends on the current N
= r*N
dt
Ecological population growth model
– r, growth rate of the population
– K, carrying capacity (due to resource limitation)
€
dN N
= rN(1− )
dt K
2010-02-22
€ Ed H. Chi ASC Overview 85
86. People-ware
– Growing resistance to changing content
– Coordination cost and bureaucracy
Knowledge-ware: Availability of easy topics to write about
Tool-ware: Quality of tools used by editors and admins
http://www.aerostich.com/
http://www.mikestreetmedia.co.uk/blog/wp-content/uploads/2009/01/knowledge.jpg
2010-02-22 Ed H. Chi ASC Overview 86
http://youropenbook.agitprop.co.uk/growing.php?p=2