Developing Smart Cities Services through Semantic Analysis of Social Streams

Developing Smart Cities Services through Semantic Analysis

of Social StreamsCataldo Musto, Giovanni Semeraro, Marco de Gemmis, Pasquale Lops

(Università degli Studi di Bari ‘Aldo Moro’, Italy - SWAP Research Group)

WDS4SC 2015 WWW 2015 Workshop on

Web Data Science and Smart Cities Florence (Italy) - May 19, 2015

Outline• Background

• Information Overload • Social Content Analytics

• CrowdPulse • Social Data Extraction • Semantic Tagging • Sentiment Analysis • Processing & Visualization

• Use Cases • L’Aquila Social Urban Network • The Italian Hate Map

• Conclusions

2Cataldo Musto, Giovanni Semeraro, Marco de Gemmis, Pasquale Lops Developing Smart Cities Services through Semantic Analysis of Social Streams. WDS4SC 2015 Workshop, Florence (Italy) 19.05.2015

Background


Background

4

Information Overload

Cataldo Musto, Giovanni Semeraro, Marco de Gemmis, Pasquale Lops Developing Smart Cities Services through Semantic Analysis of Social Streams. WDS4SC 2015 Workshop, Florence (Italy) 19.05.2015


5

… in digital life



6… in real life


7

Obstacleor

Opportunity?



Background (again)


9

Social Networks

can be considered as novel data silosCataldo Musto, Giovanni Semeraro, Marco de Gemmis, Pasquale Lops Developing Smart Cities Services through Semantic Analysis of Social Streams. WDS4SC 2015 Workshop, Florence (Italy) 19.05.2015

10

Social Networks

information about preferencesCataldo Musto, Giovanni Semeraro, Marco de Gemmis, Pasquale Lops Developing Smart Cities Services through Semantic Analysis of Social Streams. WDS4SC 2015 Workshop, Florence (Italy) 19.05.2015

11

Social Networks

information about connectionsCataldo Musto, Giovanni Semeraro, Marco de Gemmis, Pasquale Lops Developing Smart Cities Services through Semantic Analysis of Social Streams. WDS4SC 2015 Workshop, Florence (Italy) 19.05.2015

12

Social Networks

information about people feelingsCataldo Musto, Giovanni Semeraro, Marco de Gemmis, Pasquale Lops Developing Smart Cities Services through Semantic Analysis of Social Streams. WDS4SC 2015 Workshop, Florence (Italy) 19.05.2015

13

Social Networks

changed the rule for content analytics


14

Social Content AnalyticsSuccessful Use Cases

- Online brand monitoring

- Social CRM- Real-time polls

All these applications share a common insight


15

Social Content AnalyticsResearch Question

Is it possible to aggregate rough human-generated data to get complex people-based findings?


16

Our contribution: CrowdPulse

A framework for real-time Semantic Analysis of Social Streams


17

CrowdPulse

Social Data Extraction


features

Semantic Tagging

Sentiment Analysis Processing & Visualization


workflowCrowdPulse


Step 1: Social Data ExtractionCrowdPulse


Step 1: Social Data Extraction

Extraction

Source

Heuristics

CrowdPulse



Extraction

Source

Heuristics

CrowdPulse



Extraction

Source

Heuristics

ContentUserGeo

Content+Geo

#www2015#democrats

#traffic

@barack_obama@comunefi

#earthquake

Page

Group

CrowdPulse



Extraction

Source

Heuristics

ContentUserGeo

Content+Geo

#www2015#democrats

#traffic

@barack_obama@comunefi

#earthquake

Page

GroupWe only extract public content

CrowdPulse


Step 2: Semantic TaggingCrowdPulse

25

aquila

??

(eagle)

(italian city)

(italian)

Semantic TaggingMotivations

Poor Semantics


Keyword-based representation introduces a lot of noise in the analysis

26

aquila

??

(eagle)

(italian city)

(italian)



(Please, do something: l’Aquila is going to die!)(Please, do something: the eagle is going to die!)

“Fate qualcosa per favore, l’Aquila sta morendo!”

?

27



• Entity Linking Algorithms• Input: textual content • Output: identification and

disambiguation of the entities mentioned in the text.

(1) http://tagme.di.unipi.it

(2) http://spotlight.dbpedia.org

28

Step 2: Semantic Tagging

Solution: semantic processing of extracted content

Algorithms

CrowdPulse


http://tagme.di.unipi.it

http://spotlight.dbpedia.org

29


Entity Linking: identification and disambiguation of the entities mentioned in the text.


30


Non-trivial NLP tasks (stopwords removal, n-grams identification, named entities recognition and disambiguation) are automatically performed



CrowdPulse

31

Step 2: Semantic Tagging


Each entity is a reference to a Wikipedia page http://it.wikipedia.org/wiki/Massimo_Cialente

IMPORTANT!


http://en.wikipedia.org/wiki/The_Wachowskis

We enriched the entity-based representation by exploiting the Wikipedia categories’ tree


CrowdPulseStep 2: Semantic Tagging

We enriched the entity-based representation by exploiting the Wikipedia categories’ tree



Many interesting (new) features come into play!(e.g. italian politics, L’Aquila mayors, Democrats politics)

The final representation of each content is obtained by merging the entities identified in the text with the most relevant Wikipedia categories each entity is linked to.

Features = Entities + Wikipedia Categories



35

CrowdPulse


Step 3: Sentiment Analysis

36

Sentiment AnalysisMotivations


Is this content conveying any opinion?

37

Sentiment AnalysisMotivations


Is this content conveying any opinion?

This is a crucial issue if people-based findings have to be generated

38

Sentiment AnalysisDefinition

“It is the field of study that analyzes people’s

opinions, sentiments, evaluations, appraisals, attitudes, and emotions towards entities such as

products, services, organizations, individuals, issues, events, topics, and

their attributes “ (*)

(Pang, Bo, and Lillian Lee. "Opinion mining and sentiment analysis." Foundations and trends in information retrieval, 2008)


We concentrated on the polarity detection task

39

Sentiment Analysis


How to develop an (unsupervised) sentiment analysis algorithm?

40

External lexical resourcesassociate a polarity score to each term.

joy +++

frustration - -


Sentiment AnalysisLexicon

41

SenticNet(*)

(*) Cambria, Erik, Daniel Olsher, and Dheeraj Rajagopal. "SenticNet 3: a common and common-sense knowledge base for cognition-driven sentiment analysis." Twenty-eighth AAAI conference on artificial intelligence. 2014.

Inspired by the Hourglass of Emotions model

Each term is represented on the ground of the intensity of four basic emotional dimensions (sensitivity, aptitude, attention, pleasantness)

The activation level of each dimension defines 16 basic emotions


Sentiment Analysis

42

Sentiment AnalysisSenticNet

According to the triggered emotions, each term is provided with an aggregated polarity score


43

SenticNet

SenticNet models a sentiment score for some bigrams and trigrams as well!


Sentiment Analysis

44

Insight:The polarity of a textual content (e.g. a

microblog posts) depends on the polarity of the microphrases which compose it.


Sentiment AnalysisMethodology

45



A microphrase is built whenever a splitting cue

is found in the text



46





Conjunctions, adverbs and punctuations are used as

splitting cues



47






splitting cues

example: “I don’t like this food, it’s terrible”



48






splitting cues

example: “I don’t like this food, it’s terrible”{ { m1 m2

splittingcue



49

Insight:

pol(C) = ∑ pol(mi)

The polarity of a textual content (e.g. a microblog posts) depends on the polarity of the microphrases which compose it.

i=1

k

Content microphrase

T={m1…mk}



50

Insight:

pol(T) = ∑ pol(mi)i=1

k

The polarity of a content depends on the polarity of the micro-phrases which

compose it.

pol(mi) = ∑ score(tj)j=1

term

n

T={m1…mk}

Mi={t1…tn}

Content microphrase



51

Insight:

pol(T) = ∑ pol(mi)i=1

k

The polarity of a microphrase depends on the polarity of the terms which compose it.

pol(mi) = ∑ score(tj)j=1

term

n

T={m1…mk}

Mi={t1…tn}

Tweet microphrase



52

CrowdPulse



53

CrowdPulse



Overall sentiment: :-(

54

CrowdPulse



Overall sentiment: :-(The process can be iterated over a larger set of content, to get findings about the feeling of the

population regards a certain topic

55

CrowdPulse



Overall sentiment: :-(

56

CrowdPulse


Step 4: Processing & Visualization

57

CrowdPulse


Step 4: Domain-specific processing

Supervised learning

Unsupervised learning

Linguistic Analysis

classification, regression tasks

clustering

building word spaces, similarity between concepts, analysis of terms usage, etc.

58

CrowdPulse



Supervised learning


Linguistic Analysis


clustering


CrowdPulse natively supports all these methodologies

59

CrowdPulse



Supervised learning


Linguistic Analysis


clustering


The choice is typically scenario-dependent

60

CrowdPulse


Step 4: Data Visualization

An interactive analytics console is made available for

each project

Descriptive statistics can be built in real-time and can be

immediately shown

61

CrowdPulse


Recap

62

Use Cases



L’Aquila Social Urban Network The Italian Hate Map1. 2.

Use Cases


Use CasesL’Aquila Social Urban Network

April 6, 20095.8 magnitude earthquake20 billions damages70,000 people displaced309 people died



2015: six years later7 billions fundings still needed22,000 people still displacedDiaspora



19 ‘new towns’ around l’Aquila 15,200 people today live there



What about the consequences?

Loss of trust, sense of belonging, relationships



Loss of social capital



Research Question:Is it possible to extract and process social

media to monitor in real time people feelings, opinions and sentiments about the current

state of the social capital of L’Aquila?



Research Question:

We can use CrowdPulse

Is it possible to extract and process social media to monitor in real time people feelings, opinions and sentiments about the current

state of the social capital of L’Aquila?



Heuristics: - Twitter users (local newspapers, mention to politicians) - Twitter content+geo (50km around l’Aquila with specific hashtags as #laquila #earthquake, etc)

CROWDPULSE SETTINGS



CROWDPULSE SETTINGS

Heuristics: - Facebook groups (identified after a thorough analysis) - Facebook pages (identified after a thorough analysis)



Extracted content (example)

Tweets about the fear of new earthquakes.

Facebook posts about citizens’ proposals.

Tweets about people worried of the situation.Tweets about new buildings in the city.

CROWDPULSE SETTINGS



Sentiment Analysis and Semantic Tagging of the content

CROWDPULSE SETTINGS



How to map each content with the social indicator it refers to?

CROWDPULSE SETTINGS



Given a fixed set of social capital indicators, we built a classification model to associate each content (along with

its sentiment) to the social indicator it refers to.

CROWDPULSE SETTINGS



Tweet about new buildings in the city.

Input: Social indicators + semantic representation of the content




Domain-specific processing: Classification model




Output: (multi-class) classification + sentiment





The score of a social indicator is the average sentiment of all the content referring to it.



CROWDPULSE OUTPUT

Overall score of the social indicators between March and August 2014



CROWDPULSE OUTPUT

COMMUNITY PROMOTER

DEFINES SOME INITIATIVES TO EMPOWER THE SOCIAL CAPITAL

MONITORS THE STATE OF THE SOCIAL INDICATORS

Real-world applicationof the output

Conclusions


L’Aquila Social Urban Network

Crowdsourcing-based approach

Social content about L’Aquila are extracted and processed in real-time

Machine Learning exploited to build a classification

model

Sentiment Analysis used to provide each social

indicator with a score

1. 2.

3. 4. Analytics Console used to monitor the state of the social

capital in real-time

Almost 500,000 social content extracted and analyzed.


Use CasesThe Italian Hate Map

http://users.humboldt.edu/mstephens/hate/hate_map.html

Inspired by the Hate Map built by

the Humboldt University

joint research with a psychologists team of Rome University and a

no-profit agency focused on human

rights





Insight:To aggregate rough people-based data in order to analyze

complex phenomena.



Use Cases

Research Question:Is it possible to extract and process social media

to detect intolerant content posted on social networks and identify the most at-risk areas of the

Italian country?

The Italian Hate Map


Use Cases

Research Question:Is it possible to extract and process social media

to detect intolerant content posted on social networks and identify the most at-risk areas of the

Italian country?

We can use CrowdPulse



Use Cases

Heuristics: Twitter content- 76 intolerant seed terms, defined by the psychologists teams - 5 intolerance dimensions: violence (against women), racism,

homophobia, disability, anti-semitism

CROWDPULSE SETTINGS



Use Cases

Extracted content (seed term: nano/midget)

Tweet about an Italian ministry

CROWDPULSE SETTINGS

Tweet about iPod nano

Tweet about an Italian football player



Use Cases

Many non-intolerant Tweets are extracted!

Tweet about an Italian ministry

CROWDPULSE SETTINGS

Tweet about iPod nano

Tweet about an Italian football playerX

X



Use Cases

Non-intolerant Tweets are detected and filtered out.

CROWDPULSE SETTINGS



Use Cases

Ironic Tweets are detected and filtered out.

CROWDPULSE SETTINGS



Use CasesCROWDPULSE SETTINGS



Use Cases

We have to build a map, so we only need geotagged content

CROWDPULSE SETTINGS



Use Cases

We have to build a map, so we only need geotagged content

CROWDPULSE SETTINGS



Use CasesCROWDPULSE SETTINGS


Definition of heuristics to increase the number of geotagged Tweets



Dimension #Tweets #Geo %Geo

Homophobia 110,774 8,501 7,66%

Racism 154,170 1,940 1,24%

Violence 1,102,494 28,886 2,62%

Disability 479,654 3,410 0,75%

Anti-Semitism 6,000 1,150 18,03%


Use CasesCROWDPULSE OUTPUT


Violence against women Disability

Racism Homophobia


Use CasesCROWDPULSE OUTPUT


Given the maps and given the output of the linguistic analysis of intolerant Tweets (co-occurrences between terms,

timelapse, etc.), the psychologists team defined some guidelines to tackle and prevent intolerant behaviors.

These guidelines have been freely distributed to public administration on early 2015.

Conclusions


Crowdsourcing-based approach

Social content containing the seed terms is extracted and processed in

real-time

Semantic Processing exploited to delete non-intolerant

Tweets

Sentiment Analysis

used to filter out Tweet with irony

1. 2.

3. 4. Analytics Console used to build real-time hate

maps

Almost 2,000,000 social content extracted and analyzed.


Lessons Learned


Lessons Learned

102

Pipeline of state of the art techniquesEntity Linking, Sentiment Analysis, Machine Learning, Data Visualization

Use Cases.L’Aquila Social Urban Network The Italian Hate Map

DEFINITION OF A FRAMEWORK FOR REAL-TIME SEMANTIC CONTENT ANALYSIS

1.2.


Lessons Learned

103

Pipeline of state of the art techniquesEntity Linking, Sentiment Analysis, Machine Learning, Data Visualization

Use Cases.L’Aquila Social Urban Network The Italian Hate Map

DEFINITION OF A FRAMEWORK FOR REAL-TIME SEMANTIC CONTENT ANALYSIS

1.2.


The outcomes of both use cases showed that very complex phenomena can be analyzed in a totally new

way, thanks to the huge availability of textual data

Future Research

104

Integration of further machine learning techniques, and further data visualization formalisms

Evaluation of the real impact of the framework in real-world dynamics (e.g., do intolerant behaviors decrease thanks to the Hate Map?)


Improvement of the algorithms for semantic tagging, text classification and sentiment analysis

questions?Cataldo Musto, Ph.D [email protected]

@cataldomusto

mailto:[email protected]

Developing Smart Cities Services through Semantic Analysis of Social Streams

Technology