Emotional Profiling of Locations based on Social Media
1. Emotional profiling of locations
based on social media
Juan Bernabé Moreno
Álvaro Tejeda Lorente
Carlos Porcel Gallego
Hamido Fujita
Enrique Herrera Viedma
The Third International Conference
on Information Technology and Quantitative Management (ITQM 2015)
Presented in:
3. Introduction: three questions we address
1) Can you describe the emotional profile of the people
in a particular place?
2) Can you compare two different places from the
emotional perspective?
3) Can you trace how the emotional profile of a location
changes over time?
In this paper, we want to give answers to following questions:
For that, we suggest a method to extract emotional information from
the stream of geo-localized Social Media interactions in a location.
4. Social Media in a nutshell
• Social Media has positioned as a channel for getting
unfiltered feedback from customers.
• Social Media works like a sensor: 24x7 near real time
feedback
• Platforms like 4square created new business models
just based on geo-positioning
• Main stream players like FB, Twitter, etc implemented
geo-location to enrich user experience.
SM icons by: http://colaja.deviantart.com/
• Mobile Internet emerged and transformed the way
people interacted with the SM platforms
• Interactions became more pervasive (from everywhere)
and geo-location capabilities were added on top
5. Our Model: definitions
The User in the location under analysis
The Social Network for the user
… and for all Users
The User Interaction
The Entity and the interactions related to the Brand
6. Extracting emotions from
Natural Language (1/3)
Osgood et al. (1957) * identified a basis of three psychological variables as semantic
differentials:
§ Valence
–the
pleasantness
of
the
emo.on–, (bad <-> good)
§ Arousal
–the
intensity
of
emo.on
provoked
by
the
s.mulus– (passive <-> active)
§ Dominance
–the
degree
of
control
exerted
by
the
emo.on–
(weak <-> strong)
ANEW “Affective Norms for English Words”
Different methods have been created to extract emotions from user generated content:
§ machine learning methods
§ support vector machines
§ maximum entropy approaches
§ concept-level analysis of natural language -
e.g.: affective ontologies , etc-.
§ Dictionary based approaches, e.g.:
*) C. Osgood, G. Suci, and P. Tannenbaum. The Measurement of Meaning. University of Illinois, Urbana,
IL, 1957.
Not suitable for Social Media user generated content, because:
§ Is it too short
§ The interactions are often not connected to each other
The solutions so far to extract emotions from text
The problem with SM
The semantic modeling of emotions
The solution
7. § Study: participants shown lists of isolated words
§ Asked to grade each word’s valence, arousal, and
dominance level
§ Integer scale of 1–9
§ N =1034 words—previously identified as bearing
emotional weight
§ Participants = College students
§ Results published by Bradley and Lang (1999) *
Extracting emotions from
Natural Language (2/3)
In this paper we use the eANEW extended to
over 13K lexemes **)
**) Warriner, A.B., Kuperman, V., Brysbaert, M.. Norms of valence, arousal, and dominance for
13,915 English lemmas. Behavior research methods 2013;45(4):1191–1207
*) Bradley, M., Lang, P.. Affective norms for English words (anew): Technical manual and affective
ratings. Gainsville, FL: Center for Research in Psychophysiology, University of Florida 1999
ANEW “Affective Norms for English Words”
8. Extracting emotions from
Natural Language (3/3)
§ J. Russell in 1980 * suggested a model to map,
the valence, arousal pairs to named moods on a
2 circumplex plane
§ In our paper we use the evolution of the
Russell’s circumplex plane enhanced to more
named moods as of Paltoglou et all 2013 **
Russell’s circumplex plane
**) G. Paltoglou, M. Thelwall, Seeing stars of valence and arousal in blog posts, Aective Computing,
IEEE Transactions on 4 (1) (2013) 116–123
*) Russell, J.A.. A circumplex model of affect. Journal of personality and social psychology
1980;39(6):1161.
9. Applying emotional extraction
to Social Media
Defining the Emotional Rating for a SM interaction it
valence, arousal or
dominance of the
term tj in eANEX
Assuming a normal distributionDefining the Emotional Profile of a Location
Kernel Density Multivariate Function
21. What is it good for?
• Psychology domain:
– Measuring difference between places, understanding what
motivated these differences and tracing changes over time.
– At individual level, understanding how a particular person fits
into a community by comparing their emotional profiles and
unveiling certain adaptability issues.
• Marketing domain:
– Tailoring a campaign message according to the emotional
baseline of a location might multiply its effectiveness
– Understanding which locations are suitable for which products
or services
• Political domain:
– Understanding which message is suitable for which audience.
– Analysing emotional differences between affine locations and
hostile locations
22. Conclusions
• Geo-located Social Media interactions open a door to new insights
about locations
• Applying dictionary approaches it‘s possible to extract emotions
from user generated SM content
• With the Valence-Arousal-Model it‘s possible to semantically
represent emotions
• The Russell‘s circumplex plane allows for mapping VAD values to
named moods
• Combining all of them, it‘s possible to:
– Ceate emotional profiles of locations
– Compare two locations from the emotional point of view
– Trace how an emotional profile changes over time for a location
23. Taking it forward, we clearly see 2 areas for future research:
• Adopting a user centric approach
– for example, creating emotional profiles of users over a longer period of
time, which then are mapped to locations for better consistency or
considering the segmentation by gender and educational class already
present in the extended ANEW –
• Bias removal to make the insights representative for the entire population of a
location, not just the geo-located Social Media users.
Future Research