�ݺ�ߣ

REVEAL Project: Co-funded by the EU FP7 Programme Nr.: 610928 www.revealproject.eu © 2015 REVEAL consortium
Your Name – Your Company
your@email
www.yourwebsite.com
Stuart E. Middleton
University of Southampton IT Innovation Centre
sem@it-innovation.soton.ac.uk
www.it-innovation.soton.ac.uk
REVEAL Project - Trust and Credibility Analysis

RDSM 2015 Invited Talk Overview
1
 REVEAL Project
 Modality Analysis of Social Media Streams
 Geosemantics and Spatio-Temporal Grounding of Rumours
 Knowledge-based Approach to Trust and Credibility Modelling
 Future Work
Overview

Overview
2
 Objectives
 Enable users to reveal hidden ‘modalities’ such as reputation, influence or credibility of information
 Approach - Modality Extraction and Analysis
 Real-time modality extraction
 On-demand analytics capabilities
 Event-driven architecture using RabbitMQ to communicate
 Processing based on a scalable STORM cluster (real-time) & standalone HTTP services (on-demand)
 Journalism Use Case
 Newsgathering - Find newsworthy content and evidence to help verify this content
 Enterprise Use Case
 Forums - Identify and help newbies, track product feedback & sentiment & emerging trends
REVEAL Project

Modality Extraction
3
 Social Media Streams
 Twitter, You Tube, Instagram, Facebook, Four Square
 Search (historical) and Stream (real-time)
 Social Network Analysis
 Community Detection, Community Graph Extraction, Community Classification (e.g. topics), Role
Analysis (e.g. popular participant), Influence Models ...
 Content Analysis
 Image Feature Extraction (e.g. sky, city), Image Similarity Clustering, Multimedia Indexing, Image
Manipulation Detection, Topic Models, Original Content Detection, Text Stylometry ...
 Geospatial Analysis
 Geoparsing, Geosemantic Classification, Image-based Geolocation, Geospatial Topic Model ...
 Semantic Analysis
 Directed Linked-Data Crawler, Semantic Context (e.g. context summaries based on linked data) ...
Modality Analysis of Social Media Streams

Real-Time Annotation of Social Media
4
JSON Twitter @stuart_e_middle <tweet>
Author
Timestamp
URI's
Hashtags
...
Geoparsed
Location
Influence
Score
Topic
Model
JSON Facebook BBCNews <post>
JSON YouTube CNN <video>
JSON Instagram bbcnews <image>
11:30
11:35
11:40
11:45
11:50
Content Created
Text
Stylometry

Incremental Aggregation of Annotated Social Media
5
JSON Facebook BBCNews <post>
JSON YouTube CNN <video>
JSON Instagram bbcnews <image>
JSON Twitter @stuart.e.middle <tweet>
JSON Facebook Profile @gadgetshow
JSON Instagram gadgetshow <image>
Cross Check
- Timestamps
- Locations
- Authors
...
Trust Analysis
- Trusted sources
- Reputations
- Correlation to known facts
...
Trustworthy
Evidence
Credible
Evidence
User Feedback
User Feedback

Geosemantic Classification of Evidence
6
 What is ‘geosemantics’
 Study of context of spatial data - in our case contextual text relating to mentioned locations
 Our Approach – Geosemantic feature classification
 Geoparsing [Planet OpenStreetMap database] » LOC (high precision, native language)
 Text + POS + LOC + training set » classifier » context of how is location is talked about
 Classes » timeliness past | future | present, situatedness insitu | remote, confirmation confirm | deny
 Eyewitness reports » insitu
 Breaking content » present
 Denial of rumours » deny
 State of the art – Geosemantic text analysis
 Text + POS + training set » classifier » event type
 Location text » NLP Grammar » direction & distance
 e.g. trouble spotted 5 miles north of London
 Location text » sentiment analysis » good / bad opinion of text
 Resilience of approaches across event types and languages an issue
Geosemantics and Spatio-Temporal Grounding of Rumours

REVEAL Project: Co-funded by the EU FP7 Programme Nr.: 610928 www.revealproject.eu © 2015 REVEAL consortium 7
Open Street Map
Planet Database
Twitter Search & Streaming API
You Tube Search API
Instagram Search API
locations &
geometry
Keywords
Focus
Area(s)
RabbitMQ
HTTP, SQL,
SPARQL
JSON tweets
JSON tweets
+ Locations
JSON tweets
+ Locations
+ Class labels
Social
Media
Crawler
Geoparse
Geoparse
Geoparse
Geoparse
Geoparse
Geoclassify
Situation
Assessment
Visualization
Storm
Topology
Service
Key
Geospatial
Pre-processing
SQL content +
loc + class
CITATION geosemantics
Middleton, S.E. Krivcovs, V.
"Geoparsing and Geosemantic Analysis of Social Media for
Spatio-temporal Grounding of Rumours during Breaking News"
submitted TOIS 2015
CITATION geoparsing
Middleton, S.E. Middleton, L. Modafferi, S.
"Real-time Crisis Mapping of Natural Disasters using Social Media"
Intelligent Systems, IEEE , vol.29, no.2, pp.9,17, Mar.-Apr. 2014
CITATION tech blog
Middleton, S.E.
"From Twitter-based Crisis Mapping to Large-scale Real-Time
Situation Assessment with Trust and Credibility Analysis", 2014
http://revealproject.eu/

Case Study - NYSE flooding in 2012 (false rumour)
8
 Geosemantic filtering of evidence
 7000+ tweets in 5 minute analysis window
 114 ground truth tweets - WeatherChannel & CNN
 Geosemantic filter reduced content volume by 95%
 100% ground truth recall for CONFIRM class
 85% ground truth recall for DENY class
New York Stock Exchange
(building)
Battery Park
(park)

Case Study - Donetsk Airport 2015 (conflicting claims)
9
 Spatio-temporal mapping of evidence
 300,000 Tweets, YouTube & Instagram posts over 24 hours of Ukraine Crisis for 20th Jan 2015
 4 ground truth YouTube videos used by LifeNews reports on 20th Jan 2015
 Focus: Dontesk Airport cluster
 Ground truth URI's ranked 10,14 & 28 out of 30
Веселе
(suburb)
Donetsk Airport
(building)
Київський проспект
(road)

Definitions
10
 What is ‘credibility’ and ‘trust’ ?
 Trust and credibility are not well defined – below is our interpretation
 Credibility - consistency with other content (e.g. similar reports) and contextual information (e.g. local
geography)
 Trust - subjective assessment of likelihood of content being false
 A credible news report might still be false!
Knowledge-based Approach to Trust and Credibility Modelling

Approach
11
 Our Approach – Knowledge-based Trust Modelling
 Journalist already have personal sets of trusted sources they have come to rely upon
 Knowledge-based approach
 Journalist assert a-priori known facts (e.g. trusted sources, known locations)
 Evidence from streams asserted incrementally into a triple store (i.e. useekm + owlim)
 Simple inference » classify evidence » interactive exploration of evidence with journalist
 OWL classes & individuals, owl:Restriction, owl:intersectionOf, SPARQL, GeoSPARQL ...
 Not a black box - End users explore the evidence and we help them make a verification decision
 Scalable approach able to represent different viewpoints of Journalists
 State of the Art – Trust and Credibility Modelling
 Unsupervised learning (e.g. Bayesian Network, Damper Shafer) » trust prediction without explanation
 Supervised reputation models » trust prediction with explanation
 Heuristics & activity metrics » trust prediction with explanation

Early Results - Work in Progress
12

13

14

Roadmap for REVEAL Trust and Credibility Analysis
15
 REVEAL project runs until Sept 2016
 Ethnographic studies with Journalists
 Crawl content in parallel to journalists searching User Generated Content (UGC)
 Record ground truth by observing Journalists verifying news for real & explaining decisions
 Empirical analysis - compare automated decisions with ground truth
 MediaEval 2015 verification challenge
 Competition verifying images using a common Twitter dataset (10 different news events)
 Users trials 2015 - 2016
Future Work

REVEAL Project: Co-funded by the EU FP7 Programme Nr.: 610928 www.revealproject.eu © 2015 REVEAL consortium 16
Any questions?
Stuart E. Middleton
University of Southampton IT Innovation Centre
email: sem@it-innovation.soton.ac.uk
web: www.it-innovation.soton.ac.uk
twitter:@stuart_e_middle, @IT_Innov, @RevealEU
Many thanks for your attention!

�ݺ�ߣ

REVEAL Project - Trust and Credibility Analysis

More Related Content

REVEAL Project - Trust and Credibility Analysis