Orion: An Integrated Multimedia Content
Moderation System for Web Services
Yusuke Fujisaka
Akihabara Lab., CyberAgent, Inc.
fujisaka_yusuke@cyberagent.co.jp
Our business
Media Internet AD
Game Startup
Our media services
AbemaTV (AbemaTV, Inc.)
● Free-to-view internet TV with TVCFs
● 30M+ downloads
Ameba
● “Ameblo”: Japan’s largest blog service
● 20,000+ official blogs
Tapple (MatchingAgent, Inc.)
● Japan’s largest dating app
● 3.5M+ users, 100M+ matches
AWA (AWA, Inc.)
● Music subscription service
● 16M+ downloads, 45M+ musics
Agenda
1. Motivation
2. System overview
3. Orion’s effect
4. Conclusion
Motivation
● Social Networking Services (SNS) rely on User Generated Content (UGC)
● Some UGC are viewed as spam
● Platform needs aims to eliminate spam from SNS
Motivation
● Social Networking Services (SNS) rely on User Generated Content (UGC)
● Some UGC are viewed as spam
● Platform needs aims to eliminate spam from SNS
Spam characteristics
● Only a small fraction of content and users are involved with spam
All post
Spam post
〜 1/1000
〜 1/200
Spam characteristics
● Types of spam include:
○ Adult content
○ Grotesque content
○ Duplicate posts originated by certain bot
○ Abusive posts
○ Criminal posts
○ etc.
● Spam affects users not only psychologically, but also physically
● Spam may reduce the reliability of SNS
● Spam trends changes
Filtering vs. Operator
Case 1: Deploy filter systems to moderate UGC
Pros:
● Cost efficient
● Ability to handle huge amount of data
Cons:
● Models must upgrade to follow spam trends
● False-(positive, negative) happens
○ Spam UGC remains on service
○ obviously safe UGCs mistakenly deleted
○ → Service satisfaction may decrease
Filtering vs. Operator
Case 2: Operators control spam messages
Pros:
● Humans always follow trend
○ Operators classify UGCs as same view as users
● Reduce incorrect tagging
○ If operators can effectively moderate contents
Cons:
● Cost inefficient
● Resource limited
Filtering with Operator
● We need to manage a large amount of data, cost efficiently and avoid
incorrect labelling
● Two steps to process
○ Step 1: Deploy automatic filters to extract contents
including suspicious words or behavior
○ Step 2: Perform manual operation to detect actual
spam contents and remove them
Safe data:
Not caught by filter
Step
2
Step
1
Suspicious
contents
Spam
System overview
● Orion: integrated content moderation system
○ Combination of “automatic filtering” and “manual moderation”
Service
log
Service
Streaming
Metadata
DB
Filter Moderation API
Admin API
Web Server
Operator
Feedback
Queue
Retrieval
Engine
Content
DB
Automatic modules Manual modules
Streaming module
● Collects user posts from services
● Filters suspicious content as defined by each service
○ 300+ filters to mark content for moderation
○ Maximum coverage, low latency required
○ Determine whether operator check is required
Correction check
User level check
Filtering / moderation mark
Save to DB
Gather UGCs from service
Word filter Repeat post filter
ML-based filter Image filter
User level
● “Well-behaved” users are considered to not require content checking.
● What is “well-behaved” user?
○ Those who post frequently without spam
● User level
○ “Problem users’” posts must be checked
regardless of filtering
○ “Safe users” need not be checked as often
Problem user
General user
New user
Safe user
Total post #
Deleted post #
Moderation service
● Operators can moderate in service-dedicated window
● Dummy posts & quality checks are included
Analyze / Reporting
● We collect information from a variety of sources
○ Spam category, service, operator...
○ Unique IDs sent from each service are used to identify the information
● Reporting assures quality of moderation
○ If an operator failed to identify dummy spam data, it will be indicated on the report
○ Reports are displayed on a Tableau server
Effect > Spam removal efficiency
● 35+ services in use
● Orion filters and moderates millions of pages of content
New service User level applies New service
All post
Suspicious post
Deleted post
Effect > Spam removal efficiency
● Ratio comparing 2014-2015 vs. 2017-2018
(%) Check/All Delete/Check Delete/All
Min Max Ave Min Max Ave Min Max Ave
‘14-’15 1.17 26.44 7.62 0.10 2.86 0.43 0.004 0.756 0.034
Change 0.61x 5.04x 2.97x
‘17-’18 3.09 6.32 4.66 1.51 3.64 2.17 0.063 0.165 0.101
Effect > Moderation effect
● Orion has been effective since deployment
○ Criminal activity among our company’s services has greatly declined
○ No criminal case has observed in late 2017
→ Time period
→Criminalcase#
→ Orion operational
Conclusion
● Content moderation should not rely solely on automatic classification nor
manual moderation
● We introduced Orion, which integrates automatic filtering and manual
moderation
○ UGCs are screened by various filters and suspicious UGCs are send for manual moderation
○ Operators are monitored to ensure a high moderation quality
● On deploying Orion, the amount of UGC requiring manual moderation
decreased, and the number of criminal posts sharply declined
Bibliography
[1] Roberts, Sarah T. "Commercial content moderation: Digital laborers' dirty work." (2016).
[2] Sawyer, Michael S. "Filters, Fair Use & Feedback: User-Generated Content Principles and the DMCA." Berkeley Tech. LJ
24 (2009): 363.
[3] Ghosh, Arpita, Satyen Kale, and Preston McAfee. "Who moderates the moderators?: crowdsourcing abuse detection in
user-generated content." Proceedings of the 12th ACM conference on Electronic commerce. ACM, 2011.
[4] Wang, Gang, et al. "Social turing tests: Crowdsourcing sybil detection." arXiv preprint arXiv:1205.3856 (2012).
[5] Aoe, Jun‐Ichi, Katsushi Morimoto, and Takashi Sato. "An efficient implementation of trie structures." Software: Practice
and Experience 22.9 (1992): 695-721.
Thank you.

Orion an integrated multimedia content moderation system for web services

  • 1.
    Orion: An IntegratedMultimedia Content Moderation System for Web Services Yusuke Fujisaka Akihabara Lab., CyberAgent, Inc. fujisaka_yusuke@cyberagent.co.jp
  • 2.
  • 3.
    Our media services AbemaTV(AbemaTV, Inc.) ● Free-to-view internet TV with TVCFs ● 30M+ downloads Ameba ● “Ameblo”: Japan’s largest blog service ● 20,000+ official blogs Tapple (MatchingAgent, Inc.) ● Japan’s largest dating app ● 3.5M+ users, 100M+ matches AWA (AWA, Inc.) ● Music subscription service ● 16M+ downloads, 45M+ musics
  • 4.
    Agenda 1. Motivation 2. Systemoverview 3. Orion’s effect 4. Conclusion
  • 5.
    Motivation ● Social NetworkingServices (SNS) rely on User Generated Content (UGC) ● Some UGC are viewed as spam ● Platform needs aims to eliminate spam from SNS
  • 6.
    Motivation ● Social NetworkingServices (SNS) rely on User Generated Content (UGC) ● Some UGC are viewed as spam ● Platform needs aims to eliminate spam from SNS
  • 7.
    Spam characteristics ● Onlya small fraction of content and users are involved with spam All post Spam post 〜 1/1000 〜 1/200
  • 8.
    Spam characteristics ● Typesof spam include: ○ Adult content ○ Grotesque content ○ Duplicate posts originated by certain bot ○ Abusive posts ○ Criminal posts ○ etc. ● Spam affects users not only psychologically, but also physically ● Spam may reduce the reliability of SNS ● Spam trends changes
  • 9.
    Filtering vs. Operator Case1: Deploy filter systems to moderate UGC Pros: ● Cost efficient ● Ability to handle huge amount of data Cons: ● Models must upgrade to follow spam trends ● False-(positive, negative) happens ○ Spam UGC remains on service ○ obviously safe UGCs mistakenly deleted ○ → Service satisfaction may decrease
  • 10.
    Filtering vs. Operator Case2: Operators control spam messages Pros: ● Humans always follow trend ○ Operators classify UGCs as same view as users ● Reduce incorrect tagging ○ If operators can effectively moderate contents Cons: ● Cost inefficient ● Resource limited
  • 11.
    Filtering with Operator ●We need to manage a large amount of data, cost efficiently and avoid incorrect labelling ● Two steps to process ○ Step 1: Deploy automatic filters to extract contents including suspicious words or behavior ○ Step 2: Perform manual operation to detect actual spam contents and remove them Safe data: Not caught by filter Step 2 Step 1 Suspicious contents Spam
  • 12.
    System overview ● Orion:integrated content moderation system ○ Combination of “automatic filtering” and “manual moderation” Service log Service Streaming Metadata DB Filter Moderation API Admin API Web Server Operator Feedback Queue Retrieval Engine Content DB Automatic modules Manual modules
  • 13.
    Streaming module ● Collectsuser posts from services ● Filters suspicious content as defined by each service ○ 300+ filters to mark content for moderation ○ Maximum coverage, low latency required ○ Determine whether operator check is required Correction check User level check Filtering / moderation mark Save to DB Gather UGCs from service Word filter Repeat post filter ML-based filter Image filter
  • 14.
    User level ● “Well-behaved”users are considered to not require content checking. ● What is “well-behaved” user? ○ Those who post frequently without spam ● User level ○ “Problem users’” posts must be checked regardless of filtering ○ “Safe users” need not be checked as often Problem user General user New user Safe user Total post # Deleted post #
  • 15.
    Moderation service ● Operatorscan moderate in service-dedicated window ● Dummy posts & quality checks are included
  • 16.
    Analyze / Reporting ●We collect information from a variety of sources ○ Spam category, service, operator... ○ Unique IDs sent from each service are used to identify the information ● Reporting assures quality of moderation ○ If an operator failed to identify dummy spam data, it will be indicated on the report ○ Reports are displayed on a Tableau server
  • 17.
    Effect > Spamremoval efficiency ● 35+ services in use ● Orion filters and moderates millions of pages of content New service User level applies New service All post Suspicious post Deleted post
  • 18.
    Effect > Spamremoval efficiency ● Ratio comparing 2014-2015 vs. 2017-2018 (%) Check/All Delete/Check Delete/All Min Max Ave Min Max Ave Min Max Ave ‘14-’15 1.17 26.44 7.62 0.10 2.86 0.43 0.004 0.756 0.034 Change 0.61x 5.04x 2.97x ‘17-’18 3.09 6.32 4.66 1.51 3.64 2.17 0.063 0.165 0.101
  • 19.
    Effect > Moderationeffect ● Orion has been effective since deployment ○ Criminal activity among our company’s services has greatly declined ○ No criminal case has observed in late 2017 → Time period →Criminalcase# → Orion operational
  • 20.
    Conclusion ● Content moderationshould not rely solely on automatic classification nor manual moderation ● We introduced Orion, which integrates automatic filtering and manual moderation ○ UGCs are screened by various filters and suspicious UGCs are send for manual moderation ○ Operators are monitored to ensure a high moderation quality ● On deploying Orion, the amount of UGC requiring manual moderation decreased, and the number of criminal posts sharply declined
  • 21.
    Bibliography [1] Roberts, SarahT. "Commercial content moderation: Digital laborers' dirty work." (2016). [2] Sawyer, Michael S. "Filters, Fair Use & Feedback: User-Generated Content Principles and the DMCA." Berkeley Tech. LJ 24 (2009): 363. [3] Ghosh, Arpita, Satyen Kale, and Preston McAfee. "Who moderates the moderators?: crowdsourcing abuse detection in user-generated content." Proceedings of the 12th ACM conference on Electronic commerce. ACM, 2011. [4] Wang, Gang, et al. "Social turing tests: Crowdsourcing sybil detection." arXiv preprint arXiv:1205.3856 (2012). [5] Aoe, Jun‐Ichi, Katsushi Morimoto, and Takashi Sato. "An efficient implementation of trie structures." Software: Practice and Experience 22.9 (1992): 695-721.
  • 22.