With sentiment, arousal, and emotion classification there is an opportunity for better triangulation of the overall conversational mood and tone of voice. With the emotionVis detection tool, we can dissect the conversation at large and in detail across time, actors and topical spaces,
1. EMOTION TEXT INFERENCE TOOL
FOR VISUAL ANALYTICS
ICIS Pre-workshop on Text
Mining as a Strategy of Inquiry
in Information Systems Research
Dublin 11/12/2016
2. 1. Problem Space
2. Purpose
3. Emotions Research (ICIS 2015)
4. Tool Development (DESRIST 2016)
5. Live Demo : Chipotle
6. Examples of Application : Denmark.dk
7. Text Mining Methodology & Feedback
3. ICIS 2015 / HARNESSING THE SEMANTIC SPACE
DESRIST 2016 / BUILDING A FEELINGS METER
1. To understand feelings that users choose to explicitly tag and publicly share.
2. To map the semantic space of ‘Facebook feelings’.
3. To explore how (if at all) do the user-categorized ‘Facebook feelings’ differ,
on the valence and arousal dimensions, from previously theorized mappings
of feelings (Russell, 1983; Scherer 2005)
4. To inform organizational practices related to social media analytics
(Holsapple et al. 2014), particularly sentiment analysis (cf. Stieglitz and
Dang-Xuan 2013).
5. To build an analytics tool capable of processing emotions on a more granular
level and reveal more about crowd sentiment; a tool that can easily be
incorporated into researcher and practitioner workflows.
19. FEELINGS
FOLKSONOMY
Facebook Feelings Tags, as generated
by the crowd.
1. feelings of excitement are the most
widely shared
2. positive-aroused feelings hold the
most 'gravitational pull’ in general
3. there are few motivations to
express neutrally-valenced
feelings with moderate levels of
arousal
4. on the valence spectrum, the most
negative feeling is that of sadness,
greater than disappointment,
anger or even disgust
27. CHALLENGES
q collecting large training sets
q cleaning social data
q training with many leakages
q forming a data-driven typology
q balanced v unbalanced classifier
q non-emotional class (having an emotionality threshold)
28. EMOTION V NON-EMOTION
How to impose a threshold of emotionality?
q Detect non-emotion - find a non-emotional dataset
q Detect ‘emotionality’ - use an existing dictionary (LIWC, etc),
q OR use feature list from our very large dataset pf emotion tags (with a large
empirical foundation)
1. Sort - arrange all the words (features) by how discriminative they are of a certain
class (emotion categories)….
2. Order - sorted by the productiveness of each feature.
3. Define - threshold across to impose across the board
29. TOP 20 - CORE EMOTIONS
Word Coefficient
miss 6.986016894
rip 5.141740386
rest 4.066502484
sad 3.721818005
missing 3.69276163
heart 3.176140494
devastated 3.112709379
pain 3.018786778
heaven 2.822618092
peace 2.629398536
help 2.600327012
accident 2.570705668
missed 2.554370967
ashamed 2.481876802
gone 2.469852326
vibestreet 2.428147187
grandma 2.425137259
lost 2.376184404
hurt 2.374763895
anymore 2.31673515
Word Coefficient
shame 4.817227383
fucking 4.624291437
hate 4.400117991
angry 4.244185619
trick 3.904443896
police 3.478886064
stupid 3.357348258
fuck 3.205525077
pissed 3.176277029
hell 3.109560947
ki 3.046973127
boycott 2.961132313
seized 2.905431273
government 2.895995141
snuggle 2.860197067
israel 2.799715298
ct 2.797375719
stolen 2.714096689
killed 2.658487516
georghiou 2.635029566
Word Coefficient
lol 7.11841604
lmao 5.232171744
funny 4.704103552
haha 3.520833606
lmfao 3.250737286
silly 3.209428659
pubic 3.075086933
hahaha 3.053358807
laughing 3.042177409
hilarious 2.771143801
brilliant 2.621092263
humor 2.556675867
boa 2.441085167
ce 2.40026206
claudia 2.395111804
kkkkk 2.367743641
weber 2.355344547
hahahaha 2.330192825
saw 2.278566134
oh 2.274346493
Word Coefficient
challenge 5.057432269
finished 4.105620733
proud 3.859225257
finally 3.821250998
starring 2.986186739
workout 2.900317902
congrats 2.745343277
productive 2.68759694
confidence 2.682003006
glory 2.547624651
barely 2.505130637
com 2.488742175
accomplished 2.477173108
completed 2.461110115
working 2.42237729
sweat 2.366149229
ready 2.352519136
work 2.256495611
niggaz 2.148209683
officially 2.12516032
ANGER EMPOWERED EXCITED SADNESS
Word Coefficient
confused4.434192281
omg4.269047991
scary3.607214015
worried3.468052141
seen 3.12186108
safe3.116267323
ebola3.008196773
animal2.984916385
pray2.831980246
scared 2.83058258
continued2.825664009
shocked2.806177611
ghost2.606510039
separated 2.60014044
alert2.551586858
thoughts2.485203644
comments 2.41972063
otha2.394228131
praying2.349163576
a4383222.334662301
FEAR
Word Coefficient
awesome 4.361401439
happy 3.194259797
pm 3.187567352
club 2.81594721
great 2.699492084
button 2.597035291
hhhmmm 2.559250826
enjoy 2.47965485
weekend 2.409352193
mall 2.364866299
best 2.328404284
sir 2.324124295
available 2.297224631
coach 2.268088716
india 2.22553258
apple 2.225355098
team 2.163325357
2013 2.137368371
mr 2.110431081
amazing 2.110370313
JOYFUL
32. The nations of Denmark
and Sweden had a
Twitter fight involving
moose and sperm
banks
Updated by Zack Beauchamp on July 7, 2016, 12:10 p.m. ET
@zackbeauchamp zack@vox.com
33. GOING TO WAR WITH SWEDEN
Conversation networks can be traced to see the spread of dialogue and identify influencers or groups (cliques)
35. FACEBOOK DATASETS
16,233 total contributions
2,283 published pieces (admin posts)
13,950 public reactions (comments)
Isolating the page discourse from the
community discussion allows us to contrast
the topics discussed, emotional tone, and
general behaviour between your actions and
that of the reactions from the crowd.
The entire Facebook wall was collected to
zoom out and listen to the community’s
discourse and reactions for almost 9 years.
2008-16 full history of Facebook wall
> The most active day ever was on April 26, 2016
36. ACTION & REACTION
• Admin posts have been
steadily declining over
the past 5 years.
• The current rate has
been between 10 and
20 posts per month in
the past two years.
• The community however
have been commenting
more and more in the
past three years
especially. Some months
recently have reached
up to 500 user
comments. PublishedPosts(monthly)CommunityCommentary
2,283 published pieces (admin posts)
13,950 public reactions (comments)
37. NEGATIVITY
Sentiment levels show a few days with positive and negative swings, which seem to be
happening at a more frequent rate in the past two years.
April 9th 2014 saw the lowest level of sentiment thus far.
38. AROUSAL
• Arousal has been climbing from the community. The number of days that say a spike in arousal levels
seems to be increasing in the last three years.
• April 26th 2016 saw the highest levels of arousal that the Facebook community has experienced yet.
39. AROUSAL
When we zoom in to just the event days (July 7-9), arousal peaked on July 7th, before the surge in volume.
40. EMOTION ANALYSIS
Joy is the single greatest emotion detected from the Denmark.dk community on Facebook (37.8%),
followed closely by excitement. Emotionality itself peaked on April 16, 2015, with birthday greetings
from the crowd on Queen Magrethe’s 75th birthday, consisting of mostly joy and excitement.
41. JOY
• Feeling ‘happy’, ‘fantastic’ and super were detected most, while ‘hopeful’ joy
much less.
• Within joyous posts, commonly used terms offer clues as to why Joy is detected
so much. These in include family, people, life, Copenhagen and visit.
42. MOST ACTIVE FACEBOOK USERS
1,717 unique actors have taken part in the
conversation over the past 8 years.
After removing spam posts (comments over 500
characters), a handful of people have contributed
the most (right) and are consider the most vocal
members of the community.
43. WHO EXHIBITED WHICH EMOTIONS THE MOST
One can also see whose comments have been the most Angry or Sad over time, for
example. Certain individuals may be of importance to be aware of.
44. EMOTIONAL ALIGNMENT
Crowd Reaction - Comments
Page Dialogue - Posts
Emotional signatures from the publication content and the community are similar, but slightly different.
Admin posts taken on an 44% excited tone, which is more muted in the community who are just as happy
(38%) as the page content, but have a more significant amount of sadness and anger.
Comparing with the Crowd
45. EVOLUTION OVER TIME
emotionality trenddominating emotions
PublishedPostsCommentary
Emotionality in general has been rising in admin published
posts. April 16, 2016 was an outlier in terms of high degrees
of emotionality by the community.
The admin published stories have consistently been
excited (orange) and happy (green) over time whereas
the community have had a greater degree of sadness
(blue) consistently.
46. OPPORTUNITY
q With three dimensions (valence, arousal, and distinct emotion) there
is a far better triangulation of the conversational mood overall.
q Coarse and fine-grain emotion categorization offer greater
contextual depth than valence.
q By visualizing these classifications in detail we can map emotional
signatures of conversations.
q By combining the classifications with other dimensions (time, actors and
topical spaces) we can empower practitioners to make meaning and
take action.
47. Chris Zimmerman
twitter : @socialbeit
chzi10ab@student.cbs.dk
Ravi Vatrapu Mari-Klara Stein Daniel Hardt
Copenhagen Business School - Department of IT Management