Improving responsiveness of public services in housing by monitoring social media impact

Preview:

DESCRIPTION

Copyright Bojan Cestnik at CeDEM14

Citation preview

1/26

INTRODUCTION:

• Motivation

• Sentiment analysis

• Desire: Include reason and emotion in social media and PR

analysis

METHOD AND RESULTS:

• Method workflow

• Used formulas

• Results on a case study

CONCLUSION AND FURTHER WORK

TALK OVERVIEW

2/26

PROBLEM:

• Competent managing of public relations requires substantial

amount of resources and skill

• Expert intuition might not always be accurate

• D. Kahneman: Thinking, Fast and Slow

• Every inappropriate response might cause a „tsunami“ effect of

media inquiries and/or public activities

• Cyprus crisis in March 2013: the Dutch Finance Minister

Dijsselbloem said: „The Cyprus deal will be used as a template for

the future solutions of similar Eurozone banking problems“

MOTIVATION I

3/26

SOCIAL NETWORKS (APHORISMS BY NOSHIR CONTRACTOR):

• Social networks:

• It‘s not what you know, it‘s who you know.

• Cognitive social networks:

• It‘s not who you know, it‘s who they think you know.

• Knowledge networks:

• It‘s not who you know, it‘s what they think you know.

• Cognitive knowledge networks:

• It‘s not who you know, it‘s what who you know knows.

MOTIVATION II

4/26

GOAL:

• Support a process of managing public relation within an e-gov

organization with a sentiment analysis technology

RELATED EXAMPLES:

• Presidential election in 2012 in Slovenia (emotions from Twitter)

• Monitoring the influence of emotions in the press to financial

markets (EU project First)

MOTIVATION III

5/26

EMOTIONS FROM TWITTER

6/26

FACEBOOK (MARCH 2014):

• 1,28 billion monthly active users

• Average user has 130 friends

• 802 million users log in every day

TWITTER (APRIL 2014):

• About a billion members

• 255 million monthly active users (77% outside US)

• 100 million daily active users

• 500 million tweets sent every day

• Average user has 208 followers

SOCIAL MEDIA STATISTICS

7/26

SENTIMENT ANALYSIS:

• Rational arguments constitute foundations of science, economics

and law

• Emotions put flavor to our everyday lives in politics and business

• Explanatory models based on reason alone often fail to account

for the complexity of reality

• An attempt to overcome such limitations by combining rational

models and emotional explanatory approach resulted in a new

method called sentiment analysis

• Sentiment analysis aims to automatically elicit emotions like

happy-sad or positive-neutral-negative from fragments of text

INTRODUCTION I

8/26

INTENTION

9/26

SENTIMENT ANALYSIS IMPLEMENTED:

• Simple sentiments: positive and negative

• More complex sentiments:

• joy, surprise, anger, disgust, fear, sadness

• Difficulty: language used in social media

• Relatively low accuracy of sentiment classification

• Sentiment analysis still useful on a large scale

INTRODUCTION II

10/26

GOAL:

• Support a process of managing public relation within an e-gov

organization with a sentiment analysis technology

METHOD OVERVIEW:

• Analysis of user posts to a forum

• Workflow that includes receiving questions from media,

generating answers, storing and analyzing textual data

CASE STUDY:

• Sentiment of user posts to forum

• Archive of journalists‘ questions and answers in the period

between October 2007 and November 2012

METHOD

11/26

METHOD WORKFLOW

12/26

METHOD 1 WORKFLOW

Monitor

news from press

and broadcasting media

Monitor

social media posts

Articles Posts

Tagged data

News articles,

broadcasts

Posts to forums,

Twitter, Facebook

ResponsesAnalyse media and

prepare responses

Data

storage

13/26

THE HOUSING FUND OF THE REPUBLIC OF SLOVENIA:

• Founded in 1991

• Offer loans under favorable terms to citizens

• Encourage savings in housing

• Build, sell and rent apartments

• Past project: offer housing subventions to young families

IMPORTANT SLOVENIAN PUBLIC INSTITUTION:

• Considerable media attention

FINANCIAL FIGURES:

• 429 M€ assets

• 125 M€ in long term loans to citizens

THE CLIENT

14/26

USED DATASETS:

• Training dataset: 345 preselected short questions in Slovene

language containing negative, neutral and positive wording

• Testing dataset:

• 298 journalists’ questions and answers in the period between

October 2007 and November 2012

• 103 press releases, 41 explanations, and 8 press conferences

• 296 posts to the forum from March 2010 till October 2013

FORMULAS:

• Word frequencies and conditional probabilities of emotion states

AVERAGE SENTIMENT:

• Workflow that includes receiving questions from media,

generating answers, storing and analyzing textual data

RESULTS

15/26

TRAINING:

TESTING:

COMBINATION:

S = ROUND ( P(☺) * 7 + P(�) * 4 + P(�) * 1) – 4

FORMULAS

16/26

SENTIMENT PROBABILITIES FOR GIVEN

WORDSword w p(☺☺☺☺ | w) p(���� | w) p(���� | w)

advantrage 0,50 0,31 0,19

efficient 0,54 0,01 0,45

kind 0,55 0,30 0,15

...

blame 0,20 0,01 0,79

angry 0,19 0,00 0,80

reject 0,11 0,07 0,82

...

saving 0,45 0,54 0,01

good 0,30 0,56 0,14

return 0,28 0,64 0,08

17/26

PRESS AND FORUM SENTIMENT

COMPARISON

18/26

AVERAGE SENTIMENT

19/26

SENTIMENT IN FINANCE

20/26

SENTIMENT IN DNEVNIK

21/26

SENTIMENT IN DELO

22/26

SENTIMENT IN RTV SLO

23/26

SENTIMENT BY THE NUMBER OF SUB-

QUESTIONS

24/26

PICTURE TAKEN BY A CONCERNED

CITIZEN

25/26

RESULTS HIGHLIGHTS

• Approach to agile sentiment analysis used at the Housing Fund

• Following the sentiment in social networks and user forums

• Officers can validate their intuitive ideas with the analysis‘ results

• Consequence of the analysis: More frequent and regular press

conferences

FURTHER WORK

• Improve the user interface to speed-up the decision process

• Extend the analysis to other social media sources like Twitter and

Facebook

CONCLUSION & FURTHER WORK

26/26

SENTIMENT AND DAYS TO ANSWER

alenka.kern@ssrs.si, bojan.cestnik@temida.si

Recommended