Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Björn HartmannUniversity of California, Berkeley
EECS, Computer Science DivisionJuly 2012
Two Experiments in Crowd Education
Feedback and Free(?) Snacks
Björn HartmannUniversity of California, Berkeley
EECS, Computer Science DivisionJuly 2012
One Experiment in using Crowds for Education
Free (?) Snacks
UW MSR Summer Institute 2012 3
Crowds & Collective Intelligence
Expertise Sharing & (Informal)Learning
Crowd Platforms(Turkomatic, MobileWorks)
Communitysourcing(Local Crowds Grading Exams)
Expertise Sharing for End-User Programmers(HelpMeOut,
HyperSource)
Improved Design Tutorials
(ShowMeHow, Delta, Sifter, MixT)
Crowd Applications(BribeCaster, VeriFun)
Empirical Analyses
(Stack Overflow)
Teaching Crowd Workers(Shepherd)
Crowds & Collective Intelligence
Expertise Sharing & (Informal)Learning
Crowd Platforms(Turkomatic, MobileWorks)
Communitysourcing(Local Crowds Grading Exams)
Expertise Sharing for End-User Programmers(HelpMeOut,
HyperSource)
Improved Design Tutorials
(ShowMeHow, Delta, Sifter, MixT)
Crowd Applications(BribeCaster, VeriFun)
Empirical Analyses
(Stack Overflow)
Teaching Crowd Workers(Shepherd)
UW MSR Summer Institute 2012 6
UW MSR Summer Institute 2012 7
Crowds & Collective Intelligence
Expertise Sharing & (Informal)Learning
Crowd Platforms(Turkomatic, MobileWorks)
Communitysourcing(Local Crowds Grading Exams)
Expertise Sharing for End-User Programmers(HelpMeOut,
HyperSource)
Improved Design Tutorials
(ShowMeHow, Delta, Sifter, MixT)
Crowd Applications(BribeCaster, VeriFun)
Empirical Analyses
(Stack Overflow)
Teaching Crowd Workers(Shepherd)
UW MSR Summer Institute 2012 9
Word 2011
Word 2008
UW MSR Summer Institute 2012
Similar App, Different Command
?10
UW MSR Summer Institute 2012
ShowMeHow Interface Model Interface Model
Translation Map
Interaction Techniques
11
Crowds & Collective Intelligence
Expertise Sharing & (Informal)Learning
Crowd Platforms(Turkomatic, MobileWorks)
Communitysourcing(Local Crowds Grading Exams)
Expertise Sharing for End-User Programmers(HelpMeOut,
HyperSource)
Improved Design Tutorials
(ShowMeHow, Delta, Sifter, MixT)
Crowd Applications(BribeCaster, VeriFun)
Empirical Analyses
(Stack Overflow)
Teaching Crowd Workers(Shepherd)
expert
creative, multi-faceted
skill level
work complexity
novice
simple, transactional
Shepherd: Training Crowd Workers
w/ Steven Dow, Anand Kulkarni, Scott Klemmer
expert
creative, multi-faceted
skill levelnovice
simple, transactional
work complexity
Shepherd: Training Crowd Workers
Worker
Shepherd: Training Crowd Workers
Task submission
Assessor
Shepherd feedback
Assessment helped workers improve over time
95% HPD interval = [0.09, 0.44], p=0.001
0
0.1
0.2
0.3
0.4
0.250.10Slop
e Co
effici
ent
of R
atin
gs O
ver T
ime
external assessment
self assessment
no assessment
95% HPD interval = [-0.03, 0.23], p=0.08
95% HPD interval = [-0.13, 0.06], p=0.38
External assessment encourages more revisions
0
20
40
60
80
100
24.856.5
External Self
Percent of reviews edited
χ2=47.3, p<0.05
0
32
64
96
128
160
84137
Length change (chars)
F(1,81)=3.26, p=0.07
External Self
%
Crowds & Collective Intelligence
Expertise Sharing & (Informal)Learning
Crowd Platforms(Turkomatic, MobileWorks)
Communitysourcing(Local Crowds Grading Exams)
Expertise Sharing for End-User Programmers(HelpMeOut,
HyperSource)
Improved Design Tutorials
(ShowMeHow, Delta, Sifter, MixT)
Crowd Applications(BribeCaster, VeriFun)
Empirical Analyses
(Stack Overflow)
Teaching Crowd Workers(Shepherd)
UW MSR Summer Institute 2012
CommunitySourcing
19
Engaging Local Crowds to Perform Expert Work via Physical Kiosks
with Kurtis Heimerl, Brian Gawalt, Kuang Chen, Tapan Parikh
UW MSR Summer Institute 2012
Goal
Crowdsource the grading of (non-autogradable) examsRequires expert knowledgeVery time intensive - high volume
20
UW MSR Summer Institute 2012
Microtask Markets
Examples: Mechanical Turk, Crowdflower…Recruit from large pool of unskilled workersSafety in numbers: issue multiple assignments,aggregate the responses
2 2
2
1
2 3= 21+1 = ?
21
UW MSR Summer Institute 2012
Microtask Market Benefits
Pros:High Task VolumeQuick ResponsesSupport for Small Tasks
Cons…
22
UW MSR Summer Institute 2012
Microtask Market Drawbacks
Lack of expert workers(experts have full-time jobs)Hard ceiling on knowledge tasks?
4 22
x
-1
9 -2 = ?
23
eπ=
UW MSR Summer Institute 2012
Freelancing
Examples: Elance,oDesk…Hire the best applicant (as judged by you)
1 yr 0
2 mo
3 mo
3 yr 6 mo
How muchgraphic design experience do you have?
24
UW MSR Summer Institute 2012
Innovation Contests
Examples: 99Designs, InnoCentive, TopCoder…Ask a set of experts a questionTake the best response (as judged by you)
a o
q
c
d p
Design a new logo.
25
UW MSR Summer Institute 2012
Innovation Contests & Freelancing
Pros:Expert, high quality workSupports large, complex tasks
Cons:Short tasks don’t workLow volume
26
UW MSR Summer Institute 2012
How might we crowdsource short, complex work at high volume?
27
UW MSR Summer Institute 2012
CommunitySourcing
1 Bring the crowdsourcing to the crowdPlace a work kiosk where communities of these experts congregate and have free time
28
UW MSR Summer Institute 2012
CommunitySourcing
1 Bring the crowdsourcing to the crowd2 Pay them in what they valueProvide rewards that fit the users and their context, not just small change
29
UW MSR Summer Institute 2012
CommunitySourcing
1 Bring the crowdsourcing to the crowd2 Pay them in what they value3 Not one big market - many small onesTarget workers with tasks that are situated in the context and location
30
UW MSR Summer Institute 2012
Task: Grade CS Exams
Community: Computer Science students
31
UW MSR Summer Institute 2012
Task: Grade CS Exams
Community: Computer Science studentsLocation:Outside Main CS lecture hall
32
UW MSR Summer Institute 2012
Task: Grade CS Exams
Community: Computer Science studentsLocation:Outside Main CS lecture hallReward: (Unhealthy) Food.
33
UW MSR Summer Institute 2012 34
Touchscreen
Cardreader
Micro-controller+ Laptop
UW MSR Summer Institute 2012 35
UW MSR Summer Institute 2012 36
UW MSR Summer Institute 2012 37
UW MSR Summer Institute 2012 38
UW MSR Summer Institute 2012
Exams
Generated a 13 question exam out of prior CS2 exams available online8 Participants + 1 “Ringer”
39
UW MSR Summer Institute 2012
Exams
105 questions answered in totalAnswers were scanned and chunked by captricity.comEach question graded on a 0-4 point scale
40
UW MSR Summer Institute 2012
Example Questions
Q: Briefly explain the difference between a (Java) instance variable and a (Java) class variable.
A: Each object has its own distinct copy of an instance variable. But all the objects in a class share just one 'copy' of a class variable.
41
UW MSR Summer Institute 2012 42
Answer Key
GradingSlider
Question
StudentAnswer
UW MSR Summer Institute 2012 43
UW MSR Summer Institute 2012
Rewards (in Umati credits)
44
10 – Small granola bar
20 – Normal candy
30 – Energy bar
UW MSR Summer Institute 2012 45
1 Week300+ Unique Users7500+ Answers Graded
$200 of Candy
UW MSR Summer Institute 2012
Comparisons
Mechanical TurkRebuilt interface using web
46
UW MSR Summer Institute 2012
Comparisons
Expert graders (n=10), from our university
47
UW MSR Summer Institute 2012
Comparison: Mechanism
Gold standard: median grade of all experts
We sampled (with replacement) from each distribution of raters (experts, umatiers, turks) and compared to the gold standard for each exam question.
48
UW MSR Summer Institute 2012
Accuracy
49
UW MSR Summer Institute 2012
Accuracy
50
O1 Expert: 78.3%
UW MSR Summer Institute 2012
Accuracy
51
O3 Umatiers: 71.2%
UW MSR Summer Institute 2012
Accuracy
52
O 5 Turkers:25.6%
UW MSR Summer Institute 2012
Accuracy
53
O O 10 Umatiers:78.9%
1 Expert: 78.3%
UW MSR Summer Institute 2012
Results - Cost
54
Umati: 2.25¢/answer; Turk: 2.3¢/answer; Expert: 34¢/answer
UW MSR Summer Institute 2012
Results – Lower Cost
55
OO
11¢/grade
{
UW MSR Summer Institute 2012
Results – Higher Accuracy
56
O
O2%
{
UW MSR Summer Institute 2012
Potential Biases
Novelty effect likely330 users in 4 days!However, usage was growing at study end…
Computer Science Bias: Students more likely to crowdsource?
57
UW MSR Summer Institute 2012
Conclusion
CommunitySourcing works!Accuracy++/Cost-- if you…Place a crowdsourcing kiosk within an expert communityProviding rewards relevant to that communityTargeting expert skills available in that community
58
UW MSR Summer Institute 2012
Open Questions
Scaling: How do we support concurrent usage without losing “community” feel?Location: How important is it that the tasks are local (vs. just the rewards?)Generalizability: What other tasks/communities/rewards fit with communitysourcing?
59