collective emotions in cyberspace...classification based on standard supervised, machine-learning....
TRANSCRIPT
1
Collective Emotions in Cyberspace
Projective objectives
and summary of main results
in the second project period1 Feb.2010 -31. Jan. 2011
In the name of CYBEREMOTIONS Consortium
Janusz Hołyst, Project Coordinator, Warsaw University of Technology, [email protected]
www.cyberemotions.eu
2
Collective Emotions in CyberspaceEuropean Union Research Project (FP7 FET)
Participant organisation name Leaders Country SpecializationWarsaw University of Technology Janusz Hołyst Poland Physics of complex systems Ecole Polytechnique Fédérale de Lausanne
Daniel Thalmann Stephane Gobron
Switzerland Virtual reality
University of Wolverhampton Michael Thelwall United Kingdom WebometricsÖsterreichische Studiengesellschaft für Kybernetik
Robert Trappl Marcin Skowron
Austria Human‐computer interactions
ETH Zürich Frank Schweitzer David Garcia
Switzerland Chair of systems design
Jozef Stefan Institute, Ljubljana Bosiljka Tadic Slovenia Physics of complex networksJacobs University, Bremen Arvid Kappas Germany PsychophysiologyTechnical University Berlin Matthias Trier Germany Dynamic network analysisGemius SA Anna Borowiec Poland Online research agency
Large-scale integrating project, ICT Call 3 Science of Complex Systems for Socially Intelligent ICT. Duration: 1 Feb. 2009 - 31. Jan. 2013. EC funding 3.6 M€
3
Egyptian Revolution: Egyptian protest leader Wael Ghonim’s Twitter message:“congratulations Egypt the criminal has left the palace.”
Egypt, Twitter and the Straw Man Revolution
….Twitter is not the root cause of these uprisings. Twitter was not repressed. Twitter did not get inspired by events in other countries. ….Twitter can help organize. Facebook can help get the word out. ….
Twitter revolution
4
Main project tasks
• automatic collection and classification of sentiment data in various e‐communities as well as cross‐validation of such classifiers using psycho‐physiological methodologies,
• qualitative and quantitative sentiment data analysis and data‐driven modeling of collective emotions by ABM, complex networks and fluctuation scaling paradigms,
• development of emotionally intelligent ICT tools such as affective dialog systems and graphically animated virtual agents that communicate by emotional interactions.
5
1. Developing and performing EmoChatting experimentswhere users had discussions in real time via their avatars with other avatars or simulated agents. The experiments integrated a sentence-based emotion generator and applied graphics engine using models of valence, arousal and dominance for emotional coordinates as well as asymmetric facial expressions. They were carried out by EPFL in close collaboration with OFAI, UW, ETH, WUT, and JUB groups. 2011_wp2_twoYearsSummary.mp4
During the second year of its project life CYBEREMOTIONS proved its ambition to be the leading world enterprise in the domain of affective interactions observed in e-communities. CyberEmotions datasets were mentioned in a list of 'The 70 Online Databases that Define our Planet', posted by the Physics arXiv Blog, published by MIT.
Main results from the second project period
6
Purpose of WP2 Emotionally reacting active agents (EPFL)
• Computer graphical metaphor of emotion applied to Virtual Reality
WP specification• Create a virtual society composed of VH, capable of reactions, emotions, and social behavior
• Develop interpersonal relationships and nonverbal communication in a virtual society
Graph
ics &
Emotion
7
A first CG emotional model
“Hello,nice day!”
Data miningClassifier
WP2
WP2
+
{v,a}selection
WP2
WP32010_wp2_firstYearSummary_v2.mov
8
VR‐server/clientsCG animations
‐ Facial and body emotional real‐time animated interpretations
A second CG emotional model
e.g. of dialog
A: “Hello”
A: “What? ”
B: “Hi chick!”
Multi classifiers- “SuperClassifier”- “ANEW”
B: “sorry…”
A: “It’s ok ”
Emotional model- {v,a,d}- Target- Polarity
Dynamic event manager- Multi user- Free interaction
VH emotional mind, eg.:- 3D emo. model {v,a,d}- Memory based on
history of dialogs
WP2
WP2
WP2WP5
WP3
9
Result: towards a virtual social environmentincluding verbal and non‐verbal communication
10
EmoChattingDialog system - Virtual Bartender
DS <-> WOZ
WP4: acquired data‐setsAu
stria
n Re
sear
ch I
nstit
ute
for
Art
ifici
al I
ntel
ligen
ce –
OFA
I
EmoDialogOnline Virtual Bartender
Role of system’s affective profile
21 IRC channelsSame time-frame
variety of channels & discussed topics
2009-2010Recent “hot topics” of discussion:politics, economy, social issues
2200 days of cooperationLong-term online communications
between members of Ubuntu communities
11
Dialog System vs. Wizard of Oz setting- emotional connection, dialog realism
Affect Listener ‐ Evaluation of Affective Dialog Systems
Aust
rian
Rese
arch
Ins
titut
e fo
r A
rtifi
cial
Int
ellig
ence
–O
FAI
Effect of affective profile in the interaction with a dialog system
12
Virtual Reality settingsAu
stria
n Re
sear
ch I
nstit
ute
for
Art
ifici
al I
ntel
ligen
ce –
OFA
I
• Integration of Affect Listener dialog system (OFAI), virtual reality event engine (EPFL) and sentiment classifier (UW)
• Comparison of dialog system with WOZ setting
• Dialog system results in pair with WOZ ratings• correlation coefficient for DS and WOZ:- chatting enjoyment 0.95- emotional connection 0.96- dialog realism 0.97• no statistically significant differencesin the participants ratings
13
14
2. Confirmation of the emergence of collective emotions in cyber‐communities by four project teams applying different
methods and using independent datasets
(i) avalanches distribution observed in BBC blogs and Digg data by JSI;(ii) non‐random clusters distribution observed in Blogs06, BBC Forum,
Digg and IRC channels by WUT; (iii) persistent character of sentiment dynamics observed by ETHZ for
IRC channels using the Hurst exponent analysis; (iv) causal sentiment triad distribution found in Network Motif Analysis
by TUB.
WP6/JSI: Quantitative Analysis of User's Collective Behaviors
M. Mitrović, G. Paltoglou and B. Tadić, JSTAT (2011) P02005
Mapping the data on bipartite networks:Communities on the networks (Eigenvalue spectral analysis).
Time-series analysis and avalanches of positive/negative comments.
Distribution of sizes of emotional avalanches Indicates Self-Organized Criticality in the dynamics.
Accurate mapping of high-resolution data onto bipartite graphs: Users (bulits) and
Post&Comments (squares), direction of links indicate user's actions; Color: Emotional content
WP6/JSI: Agent‐Based Models on Networks
Emotional agents – 2D emotional states (arousal, valence);
Emotion dynamics on networks; 2 types of AB model on
Growing bipartite networkFixed social network
Rules of actions and parameters extracted from the Data of Blogs and Diggs and MySpace network.
Emotional state of each agent depends his/hers connections on the network.
M. Mitrović and B. Tadić, DRAFT (2011)
Simulated time-series of charge expressed in comments of 5 communities found on the
emergent network: g2 and g3 (with negative charge) continue to grow.
17
Example discussion +1 0 -1 -1 -1 +1 0 0 +1 -1
We define an emotional cluster of size n as a chain of n consecutive messages with similar sentiment orientations (i.e. negative, positive or neutral).
Classification based on standard supervised, machine-learning.Emotions e { +1, 0 , -1}
hierarchical extension - a document is initially classified by the algorithm as objective or subjective and in the latter case a second-stage classification determines its polarity, either positive or negative.
WP1, Warsaw
18
Growth probability for cluster of size n
neepneep )|()|(
The presence of a longer cluster of coherent emotional expressions increases a possibility to follow the cluster by a comment with the same emotion.
Conditional probability for cluster growth increases as a power-law with cluster length.
19
20
Demonstrator CMXViewer (D 8.2)
Workpackage 8 (TUB)
Development of Demonstrator Software CMXViewer (D 8.2):• Animated Graph Visualization
representing sentiment propagation processes
• Dynamic visualization of longitudinal network data in 2D and 3D showing sentiment dissemination processes
• Sentiment-based dynamic link coloring
21
Sentiment network visualisation (TU Berlin)
22DIGG discussion network (Discussion ID 11223766) rendered with the CMXViewer (TUB)
23
Central component
24
Negative subnetwork
25
Positive subnetwork
26
Sentiment Triad Census Analysis (D 8.3)
Workpackage 8 (TUB)
27
Collective emotions of cybercommunities detected by various methods
t
Sentiment Triad Census AnalysisEmotional persitence of IRC chatts
Emotional clustersEmotional avalanches
Hurst eponents
28
Sentiment Networks EvolutionIRC channel interaction network
(IRC channel "edubuntu_2007_11")
Movies\movie1.wmvMovies\movie2.wmv
29
WP3 Summary (Wolverhampton) 1. SentiStrength algorithm evaluation and
improvement‐ detects positive and negative sentiment strength in short informal text
2. Six human classified sentiment strength data sets with >1000 classifications (MySpace, Twitter, BBC, Digg,YouTube, RunnersWorld)
3. Ternary lexicon‐based classifier for social media: Twitter, MySpace, Digg
4. Ordinal prediction of valence and arousal on forum posts: LiveJournal
5. Real‐valued prediction on [1,9] scale of valence/arousal: BBC forum discussions
30
#oscars% m
atch
ing
post
sSe
ntim
ent
stre
ngth
Subj
.
Increase in negative sentiment strength
Date and time
Date and time
9 Feb 2010
9 Feb 2010
9 Mar 2010
9 Mar 2010
An analysis of Twitter sentiment around the top 30 mediaevents showed that increases in interest were typicallyassociated with increases in negative sentiment, even
for positive events – such as the oscars
Thelwall, M., Buckley, K., & Paltoglou, G. (2011). Sentiment in Twitter events. Journal of the American Society for Information Science and Technology, 61(12), 2544–2558.
31
EMG(smiling, frowning)
EKG(heart rate)
EDA(sweating)
Continuous recording of psychophysiology during participation in a forum discussion
32
EKG (heart rate)
Sample of recording output with a selection of channels
EMG(smiling, frowning)
Digital online event identifier (“marker”)
33
34
SummaryNew data sets:• corpora of posts at Twitter and Newsgroups; YouTube• physiological responses after reading and writing emotional posts
Confirmation of collective character of emotions in various e‐communities:• BBC blogs and Digg (avalanches) • Blogs06, BBC Forum, Digg and IRC channels by WUT (clustering)• IRC channels (Hurst exponents)• Digg (Motif Analysis)
Models of emotions dynamic in cyberecommunities • Agent based models• Stochastic models
EmoChatting and EmoDialog experiments• Based on multidimenional emotion models • Applications of avatars or agents • Role of assymetrical facial expressions
35
Collaboration to other Projects and Programmes
• Coordination Action ASSYST• Flagship initiative FuturIcT• Flagships Midterm Conference, Warsaw,
November 2011• Proposal Automatic Detection of Affective
Mails (ADAM) submitted to Swiss-Polish Research Programme.
36
More results will be presented at Partners presentations