Post on 02-Nov-2014
description
Yahoo! AnswersCommunity Moderation
Micah Alpern, Director of User ExperienceYahoo! Search
•Yahoo! AnswersAsk, Answer, Discover
• 1.2 million Q&A per day
• Experiential Knowledge
•Yahoo! AnswersAsk, Answer, Discover
• 1.2 million Q&A per day
• Experiential Knowledge
Activity
Activity
Answers is a marketplace
•People want lots of good answers to their questions quickly.
This means wide distribution of your freshest content.
Higher visibility for abuse in Questions than Answers.
“If you don't fix it right away, it's not worth fixing.”
by getxsickphotography
Traditional Moderation Failed
•As Answers grew traditional moderation methods failed to scale with the community.
•Manual customer care systems: Had slow response times
Treated all abuse reports the same
Had high false positive and false negative rates
Were high cost and scaled up with traffic
Machine learning?
A team of PHDs worked on a black box abuse classifier.
Pre-filter of bad content to be reviewed by Customer Care.
Could only handle questions, not the answers
After a while that broke as well
60% of reports were fake abuse by users trying to report others.
Community Moderation
• Deployed a newCommunity Moderation system Empowered trusted Answers users to help moderate content by allowing their report abuse actions to automatically delete content.
System didn’t reveal reputation scores to users,
Encouraged them to report accurately so they could gain more community influence.
Illustration by Bryce Glass
To address these challenges:
Reputation System
Answers Moderation Model
Hide Content
Item?
ContentReporter
Author
Answers Moderation ModelREPORTER
Abuse Suspicion Reputation
REPORTER Community InvestmentReputation
User’s Abusive
Reputation
Confirmed Abuse
Reporter Reputation
Abuse Reporter
Bootstrap Reputation
AbuseReporter
Reputation
CONTENT ITEM AbuseReputation
AUTHOROverall Asker
Reputation
AUTHOROverall Answerer
Reputation
AUTHOROverall Asker
Reputation
Question Quality
Reputation
AUTHOROverall
Answerer Reputation
Answer Quality
Reputation
Hide Content
Item?
the story of a report
an abuse report comes in
We need to decided: Who do we believe? Do we hide the content?
A question is asked..
Author Reporter
The content
A question is asked..
What do we know about the content? Have others reported it
Machined learned “Junk detector” score
Language specific
Requires tons of training data
!= a quality detector
CONTENT ITEM
AbuseReputation
the reporter
What do we know about the reporter?
Reporter
08.17.08 4:41pm
08.20.08 6:23 pm
10.04.08 1:08 pm
11.06.08 7:03 am
Past reporting behavior
hiddennot
hiddenhide overturned
Outcome
not hidden
Confirmed Abuse
Reporter Reputation CONTENT
ITEM AbuseReputation
AbuseReporter
Reputation Past reporting performance
and it’s outcome
For many reporters we don’t have enough feedback.
How do we solve the cold start problem?
Cold Start:Reporter Reputation
Reporter
Suspicion: Implied Karma
Hard evidence
Examine signals across synonymous identities
Confirmed Abuse
Reporter Reputation Abuse
Reporter Reputation
CONTENT ITEM AbuseReputation
Take this result and update the running reputation of the content reported.
Reporter’s Abuse
Suspicion Reputation
Reporter’s Community InvestmentReputation
User’s Abusive
Reputation
Abuse Reporter
Bootstrap Reputation
Do we trust the author?How good are they at Asking questions?How good are they at Answering?Do they have a history of abuse?
AUTHOROverall Asker
Reputation
AUTHOROverall Answerer
Reputation
AUTHOROverall Asker
Reputation
Question Quality
Reputation
AUTHOROverall
Answerer Reputation
Answer Quality
Reputation
Users Abusive
Reputation
we make a decision
an abuse report comes in
A question is asked..
Author Reporter
decisions can be revised
an abuse report comes in
A question is asked..
Author Reporter
Trust level is re-instated andtemporary protection assigned.
Trust level for Reporting is decreased.
CraigsListAfter 3 flags the post is deletedOther systems: Ignore new user flags
In Answers:• Its harder to remove content from good contributors.
• Easier for a even a new person to remove a spammer.
Naive moderation
So how did we do?
Speed
Accuracy
Concussion
• Reputation system helped Answers:
Scale & empower the community to mange abuse
Provide a better experience to Askers and Answerers
Questions?
Bryce Glass
Randy Farmer
Ori Zaltzman
http://buildingreputation.com/
Micah Alpern
malpern@alpern.org
twitter:malpern
SPAM
• “We have other methods for identifying and dealing with spam, but as a very popular site we are a target for spammers so are continually upping our game to find new ways to identify them quicker and remove their content faster (though we shouldn’t share the methods we take). We recognize spam damages the experience for our users and so are committed to reducing spam on Answers as much as possible.”
Chat and Low quality content
• “You mention this a bit in the presentation, but generally we acknowledge there are different types of questions being asked and answered on the site and accept that currently we’re limited in how we are identifying and highlighting knowledge based Q&A vs Social Q&A. However, we are looking at ways to easier promote the type of content a user wants to see.”
Accuracy
• “focus is on experiential knowledge, practical help based on someone’s opinion or experience. Different type of knowledge than covered by wikipedia. We don’t want to publish content which is misleading, incorrect or harmful but use the “report abuse” system and our trusted users to balance any dubious advice.