Transcript Introduction to Mechanized Labor Marketplaces: Mechanical Turk

Introduction to Mechanized Labor
Marketplaces: Mechanical Turk
Uichin Lee
KAIST KSE
Mechanical Turk
•
Begin with a project
–
•
Break it into tasks and design your HIT
–
–
•
Define the goals and key components of your project. For example, your goal might be to clean your
business listing database so that you have accurate information for consumers.
Break the project into individual tasks; e.g., if you have 1,000 listings to verify, each listing would be
an individual task.
Next, design your Human Intelligence Tasks (HITs) by writing crisp and clear instructions, identifying
the specific outputs/inputs desired and how much you will pay to have work completed.
Publish HITs to the marketplace
–
You can load millions of HITs into the marketplace. Each HIT can have multiple assignments so that
different Workers can provide answers to the same set of questions and you can compare the results
to form an agreed-upon answer.
https://requester.mturk.com/tour/how_it_works
Mechanical Turk
•
Workers accept assignments
– If Workers need special skills to complete your tasks, you can require that they pass a
Qualification test before they are allowed to work on your HITs.
– You can also require other Qualifications such as the location of a Worker or that they have
completed a minimum number of HITs.
•
Workers submit assignments for review
– When a Worker completes your HIT, he or she submits an assignment for you to review.
•
Approve or reject assignments
– When your work items have been completed, you can review the results and approve or reject
them. You pay only for approved work.
•
Complete your project
– Congratulations! Your project has been completed and your Workers have been paid.
https://requester.mturk.com/tour/how_it_works
Screenshot
Screenshot
AMT Questions
• Who are the workers that complete these
tasks?
• What type of tasks can be completed in the
marketplace?
• How much does it cost?
• How fast can I get results back?
• How big is the AMT marketplace?
Analyzing the Amazon Mechanical Turk marketplace, P. G. Ipeirotis, Journal XRDS: Crossroads, 2010
Demographics of Mechanical Turk, P.G. Ipeirotis, NYU TR, 2010
Gender and Age
• Countries: 46.80% US, India: 34%, Misc: 19.2%
(from 66 different countries)
http://behind-the-enemy-lines.blogspot.com/2010/03/new-demographics-of-mechanical-turk.html
Educational Level
• Many of the workers are younger than overall
population, and this leads to higher
educational levels
Income Level
• Indian Turkers relatively have low income level
Marital Status and Household Size
• Lots of single workers
• Indian workers tend to belong to larger households
Level of Engagement on M-Turk
• Most workers spent less than a day per week, completing 20100 HITs, and earning less than $20 per week.
Motivation
• Why do you complete tasks in Mechanical Turk? Please check any of the
following that applies (multiple items possible):
– [1] Fruitful way to spend free time and get some cash (e.g., instead of
watching TV)
– [2] I find the tasks to be fun
– [3] To kill time
– [4] For "primary" income purposes (e.g., gas, bills, groceries, credit cards)
– [5] For "secondary" income purposes, pocket change (for hobbies, gadgets,
going out)
– [6] I am currently unemployed, or have only a part time job
[1]
[2]
[3]
Motivation
• Why do you complete tasks in Mechanical Turk? Please check any of the
following that applies:
– [1] Fruitful way to spend free time and get some cash (e.g., instead of watching
TV)
– [2] I find the tasks to be fun
– [3] To kill time
– [4] For "primary" income purposes (e.g., gas, bills, groceries, credit cards)
– [5] For "secondary" income purposes, pocket change (for hobbies, gadgets,
going out)
– [6] I am currently unemployed, or have only a part time job
[4]
[5]
[6]
Summary
• Significant fraction of Turkers are from US and
India (> 80% vs. 20% from misc)
• Turkers are younger (more than 50% from 21-35)
• More females (US) and more males (India)
• Turkers relatively have lower income
• Turkers relatively have smaller families
• Geographic distribution of Turkers and Internet
users is similar
Type of Tasks in M-Turk
Requester vs. Total Rewards
• Long tail nature of participation
Keywords vs. Ranks
Price Distribution
• HITgroups with a large number of HITs tend to have a low
price (e.g., $0.10)
• 90% of HITs pay less than $0.10
Posting and Serving Process
• Cumulative distribution plots:
– For each day, the value of tasks being posted by the AMT requesters,
and the value of the tasks that got completed in each day
• Median posting/completion rate: $1,040 vs. $1,155 per day
• M/M/1 queueing assumption: a task worth $1, average completion time
of 12.5 minutes
Value of posted/completed HITs in USD ($)
Posting and Serving Process
• Less posting over the weekends (requesters)
• Less work done on Monday due to less posting over the
weekends
Posting
Completion
Completion Time Distribution
Power law
distribution:
w/ alpha = -1.48
10 days
10 days
Running Experiments with
Amazon Mechanical-Turk
Gabriele Paolacci, Jesse Chandler, Jesse Chandler
Judgment and Decision Making, Vol. 5, No. 5,
August 2010
KSE 801: Human Computation and Crowdsourcing
Practical Advantages of M-Turk
• Supportive infrastructure:
– Fast recruiting
– Convenient to run experiments
– External site could be used (e.g., validation code)
• Subject identifiability and prescreening:
– M-Turk workers can be required to earn “qualifications” (or prescreening
questions) prior to completing a HIT
• Subject identifiability and longitudinal studies:
– Worker IDs can be used to explicitly re-contact former subjects or code can be
written that restricts the availability of a HIT to a predetermined list of workers
• Cultural diversity:
– Cross-cultural comparisons feasible (e.g., country, language, currency)
• Subject anonymity (not easy though)
– Ensuring worker’s anonymity (if external site is used)
– M-Turk studies can be exempted for the review of IRBs (Institutional Review
Boards) if anonymity is guaranteed
Tradeoffs of Different Recruiting Methods
A Comparative Study
• Tested various Judgment and Decision Making
(JDM) findings
– M-Turk, a traditional subject pool at a large
Midwestern US university, and visitors of online
discussion boards
– During April to May 2010
• Survey:
– Asian disease problem
– Linda problem
– Physician problem
Survey (Asian Disease Problem)
• Asian disease problem (called framing, Tversky and Kahnerman,
1981)
• Subjects read one of two hypothetical scenarios
– Imagine that the United States is preparing for the outbreak of an unusual
Asian disease, which is expected to kill 600 people. Two alternative programs
to combat the disease have been proposed. Assume that the exact scientific
estimates of the consequences of the programs are as follows:
– Problem 1: If Program A is adopted, 200 people will be saved. If Program B is
adopted, there is 1/3 probability that 600 people will be saved and 2/3
probability that no people will be saved. Which of the two programs would
you favor?
– Problem 2: If Program A is adopted, 400 people will die. If Program B is
adopted, there is 1/3 probability that nobody will die, and 2/3 probability that
600 people will die.
• Two scenarios are numerically identical, but the subjects responded very
differently
• In the scenario framed in terms of gains, subjects were risk-averse (72%
chose Program A); in the scenario framed in terms of losses, 78% of
subjects preferred Program B (Tversky and Kahnerman, 1981)
Survey (Linda Problem)
• Example: “Linda is 31 years old, single, outspoken, and very
bright. She majored in philosophy. As a student, she was
deeply concerned with issues of discrimination and social
justice, and also participated in anti-nuclear
demonstrations.”
• Which is more probable?
– Linda is a bank teller
– Linda is a bank teller and is active in the feminist movement
• Linda problem (Tversky & Kahneman, 1983)
– Demonstrates the conjunction fallacy
– People often fail to regard a combination of events as less
probable than a single event in the combination
• Probability of two events occurring together (in “conjunction”) is
always less than or equal to the probability of either one occurring
alone
Survey (Physician Problem)
• Physician problem demonstrates the outcome bias: a
surgeon deciding whether or not to do a risky surgery on a
patient.
– The surgery had a known probability of success (e.g., 92%)
– Subjects were presented with either a good or bad outcome (in
this case living or dying), and asked to rate the quality of the
surgeon's pre-operation decision.
• Judgment of quality of a decision is often dependent on the
valence of the outcome (Baron and Hershey, 1988)
• Subjects rated the quality of a physician’s decision to
perform an operation on a patient (on a 7-point scale)
– 1: incorrect and inexcusable, 7: clearly correct, and the opposite
decision would be inexcusable
– Those presented with bad outcomes rated the decision worse
than those who had good outcomes.
After Survey
• After survey, subjects completed the subjective
numeracy scale (SNS, 2007) called SNS score
– An eight-item self-report measure of perceived ability to
perform various mathematical tasks and preference for the
use of numerical vs. prose information
– Used as a parsimonious measurement of an individual’s
quantitative abilities
• Additional “catch trial” question: to test whether
subjects were attending to the questions (by requiring
precise and obvious answers)
– E.g., “while watching the television, have you ever had a
fatal heart attack?” (w/ six-point scale anchored on
“Never” and “Often”)
Configuration
• M-Turk:
– Pay: $0.10 (N=318 participated)
– Title: “Answer a short decision survey”
– Description: “Make some choices and judgments in this 5minute survey”
• Estimated completion time is included to provide workers with a
rough assessment of the reward/effort ratio (e.g., $1.71/hour)
• Lab subject pool:
– N=141 students from an introductory subject pool at a large
university
• Internet discussion board:
– Posted a link to the survey to several online discussion boards
that host online experiments in psychology
– Online for 2 weeks; and N=137 visitors took part in the survey
Subject Pools: Characteristics
• Subjects recruited from online discussion forums were
significantly less likely to complete the survey than the
subjects on M-Turk (69.3% vs. 91.6%, X2=20.915, p<.001)
• # of respondents who failed the catch trial is low, and not
significantly different across subject pools (X2(2,301)=0.187,
p=0.91)
• Subjects in the three subject pools did not differ
significantly in the SNS score: F(2, 299) = 1.193, p=0.30
Results on Experimental Tasks
• M-Turk is a reliable source of experimental data in JDM
Labor Supply
• Economic theory predicts that increasing the price paid for
labor will increase the supply of labor in most cases
• M-Turk experiment: after completing the demographic
survey and the first task (transcription), subjects were
randomly assigned to one of the four treatment groups and
offered the chance to perform another transcription for p
cents: 1, 5, 15, or 25
• Workers receiving high offers were more likely to accept