�ݺ�ߣ

REVEAL Project: Co-funded by the EU FP7 Programme Nr.: 610928 www.revealproject.eu © 2016 REVEAL consortium
Stefanie Wiegand & Stuart E. Middleton
University of Southampton IT Innovation Centre
{sw,sem}@it-innovation.soton.ac.uk
Veracity & Velocity of Social Media Content
during Breaking News:
Analysis of November 2015 Paris Shootings

REVEAL Project: Co-funded by the EU FP7 Programme Nr.: 610928 www.revealproject.eu © 2016 REVEAL consortium 1
 Introduction
 Experiment
 Results
 Discussion
 Future work
Overview

What's this all about?
2
 Problems:
 Journalists doing breaking UGC verification – speed vs. accuracy
 Echo chamber can make false rumours go viral
 Automate information gathering – Journalists make the final decision
 Ideas:
 First 60 mins of a UGC post filter by attribution to trusted sources
 Visualise traffic patterns for posts attributed to trusted and untrusted sources
 Can traffic analysis help to verify / debunk content?
 First 5 mins rank UGC not seen before by mention count
 Provide a ranked list of likely eyewitness UGC every 5 mins
 Can we produce a high quality eyewitness UGC feed?
Introduction

Experiment setup
3
 Data
 5 viral UGC posts (3 eyewitness, 2 debunked) - manually identified
 38GB of serialised data covering the first 6h after the first attack
 5.9M posts, ~40k attributed sources, ~418k unique URLs
 ~160k - 1.8M posts in the first hour per UGC test case
 Technology
 Target UGC Image/Video → TinEye → Duplicate Images/Videos
 Posts → Text extraction → Sources → PostgreSQL
 PostgreSQL → Triple store → Trust knowledge model → Trusted posts
Experiment

Experiment method
4
 Verification (Experiment 1)
 Filter (un-)trusted content in first 60 mins of 5 target UGC posts
 Examine velocity of trusted and untrusted sources mentioning target UGC
 When is target UGC attributed to trusted sources?
 Identification (Experiment 2)
 Temporally segment first 5 mins of posts for 5 target event times
 Filter old URLs (including alternative URLs)
 Rank by mention frequency
 Does target UGC appear highly in ranked list?
Experiment

Experiment 1 - Case P1
5
Results
0
50
100
150
200
250
300
350
400
10 20 30 40 50 60
contentitems[#]
time [min]
P1
trusted unknown untrusted total

6
Results
0
200
400
600
800
1000
1200
10 20 30 40 50 60
contentitems[#]
time [min]
P2

7
Results
0
50
100
150
200
250
10 20 30 40 50 60
contentitems[#]
time [min]
P3

7
Results
0
1
2
3
4
5
10 20 30 40 50 60
contentitems[#]
time [min]
trusted/untrusted P3
trusted untrusted

Experiment 1 - Case D1
9
Results
0
500
1000
1500
2000
2500
3000
3500
10 20 30 40 50 60
contentitems[#]
time [min]
D1

Experiment 1 - Case D2
9
Results
0
500
1000
1500
2000
2500
3000
3500
10 20 30 40 50 60
contentitems[#]
time [min]
D2

Experiment 2
11
Results
Target Image ID P1 P2 P3 D1 D2
number of followers of author 335 1.4k 218 2.8k 151k
content likes 11 408 35 17k 29k
content retweets 83 3.3k 194 22k 30k
total # of tweets
in 60 minute window
483918 162111 811079 1501000 1837173
total # of unique mentioned URLs in
60 minute window
785 4331 535 7907 13252
ranking of target image set in total for
5 minute segment
(top x percent)
9 / 653
(2%)
1 / 603
(1%)
61 / 1097
(6%)
427 / 11605
(4%)
1 / 11337
(1%)
total number of eyewitness content in
5 minute segment
25 2 12 29 30
unique number of eyewitness content
in 5 minute segment
4 1 4 13 14

How is this useful to journalists?
12
 Posts by trusted matter for verification
 Wisdom of the crowds is not always wisdom at all
 Twitter "echo chamber" is less useful than a post by a trusted source
 Easier/faster to spot new eyewitness UGC
 Filter feeds to 10s of posts not 1000s of posts
 Reduce information overload for journalists in first 5 mins
 Additional analysis can improve eyewitness UGC further
 Eyewitness classification
 Image analysis (e.g. Exif metadata)
 Author profile pages
Discussion

Where to go from here
13
 Cross check known facts
 Extend knowledge model to support this
 e.g. image classification of weather/lighting ↔ time & location of event
 e.g. mentions of known event actors
 Use linked open data to visualise source bias
 this can include political, religious or other bias
 Observational study of journalists verifying UGC
 Journalist experts show best practice verification on specific examples
 We train our algorithms on observed best practice
 We check our algorithms results against journalists ground truth
Future work

REVEAL Project: Co-funded by the EU FP7 Programme Nr.: 610928 www.revealproject.eu © 2016 REVEAL consortium 14
Any questions?
Stefanie Wiegand & Stuart E. Middleton
University of Southampton IT Innovation Centre
email: {sw|sem}@it-innovation.soton.ac.uk
web: www.it-innovation.soton.ac.uk
twitter: @RevealEU, @IT_Innov, @stuart_e_middle
Many thanks for your attention!

�ݺ�ߣ

Veracity & Velocity of Social Media Content during Breaking News

More Related Content

Veracity & Velocity of Social Media Content during Breaking News