twitter cancer awareness campaigns as … · regression (evaluation on combined(2) test set) pes p...
TRANSCRIPT
![Page 1: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/1.jpg)
TWITTER CANCER AWARENESS CAMPAIGNS AS #IDENTITYPROJECTS FOR PREVENTION AND EARLY DETECTION BEHAVIOR
Lunch Lecture Inter-Actief – April, 4th 2017
@AnnaPriante – [email protected]
07/04/2017 Footer text: to modify choose
'View' or 'Insert' (for Office 2007 or
later) then 'Header & Footer'
1
![Page 2: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/2.jpg)
2
Agenda
Something about me
PhD project
Social Identity Classifier and applications
Q&A
![Page 3: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/3.jpg)
#WhoIAm in 160 characters and some more
![Page 4: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/4.jpg)
TWITTER CANCER
AWARENESS CAMPAIGNS
AS #IDENTITYPROJECTS
A multi-level study of the
influence of online identities
on offline behaviors
![Page 5: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/5.jpg)
![Page 6: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/6.jpg)
![Page 7: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/7.jpg)
![Page 8: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/8.jpg)
![Page 9: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/9.jpg)
![Page 10: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/10.jpg)
![Page 11: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/11.jpg)
TWITTER CANCER
AWARENESS CAMPAIGNS
AS #IDENTITYPROJECTS
A multi-level study of the
influence of online identities
on offline behaviors
![Page 12: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/12.jpg)
Effectiveness of online campaigns
Online (low-
threshold) behavior
Meaningful offline
behavior
![Page 13: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/13.jpg)
How can we study
campaigns effectiveness?
Conceptualize, explore and
test how online identities
explain the effectiveness of
online campaigns
ONLINE
IDENTITY
offline
online
EFFECTIVENESS OF
ONLINE CAMPAIGNS
![Page 14: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/14.jpg)
Identities on Twitter
Profile & Descriptions
Tweets & retweets
Hashtags
#movember #modatagranttwente
#moustache #menshealth
![Page 15: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/15.jpg)
SOCIAL SCIENCES
MEET
BIG DATA
![Page 16: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/16.jpg)
University of Utrecht - 11 January 2017 16
‘Big social data’: all tweets from 9 campaigns and
6 cancer types since 2008
> 300 million
Tweets
![Page 17: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/17.jpg)
#WhoAmI IN 160 CHARACTERS?
CLASSIFYING SOCIAL IDENTITIES BASED ON
TWITTER PROFILE DESCRIPTIONS
A. Priante, D. Hiemstra, T. van den Broek, A. Saeed, M. Ehrenhard, A. Need
Proceedings of the first EMNLP Workshop on Natural Language Processing and
Computational Social Science, Association for Computational Linguistics, ISBN 978-
1-945626-26-5, pages 55-65, 2016
17
Main contributions:
1. Social theory can be used to guide NLP methods, and NLP methods provide
input to revisit social theory
2. Three social identity classifiers of Twitter profile descriptions (English) for
scaling-up online identity research to massive datasets
![Page 18: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/18.jpg)
From language to identity and behavior
18
LANGUAGE
SOCIAL IDENTITY
BEHAVIORS
![Page 19: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/19.jpg)
Social Theory
SOCIAL IDENTITY: individual’s self-concept derived from social
roles or memberships to social groups
19
Relational Occupational Political
Ethnic/Religious Stigmatized
5-Category Online Social Identity Classification
![Page 20: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/20.jpg)
Data & Methods
20
• Movember campaign
• Random tweets
• From UK, US & AUS (country classifier)
English Tweets
N = 2,633
Profile
description
• Human Annotation (2 coders)
• Codebook
• Inter-rater reliability (KALPHA) > 0.850
Qualitative Content Analysis
• Binary text classification
• TF-IDF weighting & ChiSquare for FS
• 4 classification algorithms (SVM, BNB, LR, RF)
• Stratified 10-Fold Cross Validation
• Evaluation measures: precision (P), recall (R) and F-score.
Automatic Text
Classification
![Page 21: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/21.jpg)
Experiment 1: PARTIALLY WORKS!
21
• 5-category online social identity classification
• Stratified 10F Cross Validation on 3 training sets
• Optimal classification algorithm: Logistic Regression (evaluation on 3 test sets)
![Page 22: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/22.jpg)
Experiment 1: PARTIALLY WORKS!
22
• 5-category online social identity classification
• Stratified 10F Cross Validation on 3 training sets
• Optimal classification algorithm: Logistic Regression (evaluation on 3 test sets)
RELATIONAL OCCUPATIONAL
P R F P R F
0.757 0.648 0.699 0.743 0.791 0.766
POLITICAL ETHNIC/RELIGIOUS STIGMATIZED
P R F P R F P R F
0.600 0.200 0.300 0.661 0.460 0.543 0.958 0.273 0.425
![Page 23: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/23.jpg)
REVISITING SOCIAL THEORY
23
• Political, Ethnic/Religious and Stigmatized identities are more engaged in collective action
• They have a collective, action-oriented nature and may often overlap
• We merge Political, Ethnic/Religious and Stigmatized identities (PES identity)
Distributions of social identities on Twitter (based on the
annotated sample).
28%
36%
4%
13%
7%
45%
0%
5%
10%
15%
20%
25%
30%
35%
40%
45%
50%
Relational Occupational Political Ethnic/religious Stigmatized N/a
PES
![Page 24: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/24.jpg)
Experiment 2: WORKS!
24
• 3-identity category merger (PES identity)
• Stratified 10F Cross Validation on Combined(1)
training set
• Optimal classification algorithm: Logistic
Regression (evaluation on Combined(2) test set)
PES
P R F
0.857 0.466 0.604
We succeeded in developing 3 social identity classifiers for relational, occupational and PES
identities!
Social theory can be used to guide NLP methods, and NLP methods
provide input to revisit social theory
![Page 25: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/25.jpg)
WHAT’S NEXT?
THE ROLE OF OCCUPATIONAL IDENTITY AND
COMMUNICATION NETWORKS
IN CHARITABLE GIVING DURING ONLINE CAMPAIGNS
![Page 26: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/26.jpg)
26
Occupational
Identity
(Classifier)
Charitable
Giving
Communication
Network
+
Frequency of
Tweets
Structural position
in the network
![Page 27: TWITTER CANCER AWARENESS CAMPAIGNS AS … · Regression (evaluation on Combined(2) test set) PES P R F 0.857 0.466 0.604 We succeeded in developing 3 social identity classifiers for](https://reader036.vdocuments.site/reader036/viewer/2022081617/60422211fac898167e7487d2/html5/thumbnails/27.jpg)
Practical contributions for Movember
27
Campaign organizers should consider to target people with
occupational identities because they are likely to engage in charitable
giving
Limit behavioral choices to those relevant to charitable giving by
lightening the intensity of Twitter activity in favor of the creation of
communicative networks, social identification, interconnectedness and
interdependence