dama - final presentation

Motivation. Related works. Generating bill based Policy Networks. Results Conclusions Future work

Mining parliamentary data and news articlesto find patterns of collaboration between

politicians and third party actors.

Francisco Rodrıguez Drumond

DAMA & LARCA - UPC

July 7,2014

1 / 30


Social Networks: a natural tool for political analysis.

Nodes: Families of thepolitical landscape of XVcentury Florence.

Links: marriages betweenfamilies (alliances).

2 / 30


Analizing parliaments through SNs.

Why?

Main challenge: source of information (nodes andrelationships)

Co-sponsorship. [Fow06]Speeches. [TPL06]Strong and weak ties. [Kir11]

Can we discover relationships involving third-party actors?

Third party discoveryDefining meaningful relationships.

3 / 30


An overview of our task

4 / 30



5 / 30



6 / 30



7 / 30


SOPA: A motivating example.

Policy Networks (PN): Social networks for political analysis.

8 / 30


An overview of the literature.

Co-occurrence. [EESGGHAC14], [PSIO06].

Enriching links with the strength and semantics of relations.[Tan07],[PSB07],[ZAR03].

Beyond document co-occurrence. [NCSS06],[Bra06].

A (very) related paper. [MID+13]

9 / 30


A (very) related paper.

Moschopoulos (2013) Toward the automatic extraction of policynetworks using web links and documents

Two pre-computed PNs: Ireland and Greece.

Ground truth used for measuring correlations with similaritymeasures.

Web based.

Three types of similarity metrics:

Co-occurrence metrics (Set comparisons).Text-based metrics.Link-based metrics.

10 / 30


Generating bill based Policy Networks: the architecture.

11 / 30


Finding news articles that talk about a bill.

Topic modeling:

TF-IDF for keywordextraction.

One bill - one document.Whole set of bills as thecorpus.

1,2,3-ngrams.

Top 1000 keywords for eachbill.

Querying news articles:

Bills and news articlesmodeled as vectors

Cosine similarity forcomparison.

Rocchio’s rule for improvingqueries.

12 / 30


Selecting relevant news articles.

Threshold: point that maximizes:

threshold = argmaxp|p − (p.b′)b′|

Intuition: point at which there isno significant gain in score.

13 / 30


Entity extraction and preprocessing.

MITIE for entity extraction +

1 Entity Normalization‘The Univ. lumiere Lyon 2’ → ‘Univ Lumiere Lyon 2’

2 Mapping organization initials to the whole name‘The World Life Fund (WLF) has...’→ ‘World Life Fund’ = ‘WLF’

3 Mapping partial names with full names‘George Harrison preferred .... Harrison also...’→ ‘George Harrison’ = ‘Harrison’

4 Expanding names based on the news corpus‘Politecnica de Catalunya’→ ‘Universitat Politecnica de Catalunya’

14 / 30


Filtering relevant entities.

Problem: +3000 entities per bill

Noise.

Expensive comparisons.

Solution:

Document co-occurrence + Latent Semantic Indexing (LSI)for fast similarity computation.

Hierarchical Agglomerative Clustering (HAC) for groupingentities based on their similarity.

Politicians → seed entities.

Silhoutte for detecting the best cluster containing seedentities.

15 / 30


Computing and thresholding entity similarities.

Entities represented as vectors of 1...3-grams occurring inparagraphs they are mentioned in.

TF-IDF with sublinear TF scaling (tf = 1 + log(frequency))

Cosine similarity for comparing the vectors.

Elbows for detecting relevant entities for each entity.

Two entities e1 and e2 are related iff they are in each othersrelevant entities list.

16 / 30


Results.

Two bills:

BCN-World.Law of Popular Non-referendary Consults.

Look at:

Communities → colors.Influencers → node size.

17 / 30


BCN-World - Organizations.

Acencas

CUP

Tripartito

Canales Y Puertos De Tarragona

Puerto De Tarragona

Diputacion De Tarragona

Camara De Comercio De Tarragona Pimec

Govern

PSC

Parlament

Ciu

Veremonte

ERC

Icv-euia

PP

La Caixa

Melco

Hard Rock

Cepta

PPC

Sociedad Centre Medics Selva Maresme

Ciutadans

Grup Peralada

Camara Catalana

AECE

URV

Value RetailFerrari

Melia

Hard Rock Cafe

18 / 30


BCN-World - Persons-Organizations.

Felip Puig

Francesc Xavier Mena

PimecDiputacion De Tarragona

Govern

Veremonte

Josep Gonzalez

Josep PobletPere Granados

Xavier Adsera

Salvador Guillermo

Isidre Faine

Xavier Pallares

Francesc Homs

Antoni Belmonte

Cepta

Dolors Llobet

Caixabank

Macia Alavedra

Javier De La Rosa

Daniel De Alfonso

Hortensia GrauJoan Herrera

Santi Vila

Jordi VilajoanaLluis Rullan

Melco

Xavier Sabate

Josep Felix Ballesteros

Puerto De Tarragona

Enrique Bañuelos

Josep Andreu

URV

Isabel Vallet

Joan Pons

Icv-euia

Pere Aragones

Parlament

Hard Rock

Grup Peralada

PSC

PP

Jordi Turull

Marta Rovira

Miquel Salazar

Jordi PonsJaume Amat

Sociedad Centre Medics Selva Maresme

Pere Navarro

Andreu Mas-colellCaixa

Damia Calvet

Oriol Junqueras

Artur Mas

Camara Catalana

Ciu

ERC

PPC

Albert Batet

Alicia Romero

Melia

Value Retail

Alejandro Fernandez

Enric MilloAlicia Sanchez-camacho

CUPEnric Genesca

Tripartito

Ciutadans

Agusti Colom

Camara De Comercio De Tarragona

Ferrari

La Roca

Ernest Maragall

Josep Mayoral

Carles Pellicer

Acencas

Francesc Perendreu

AECE

Jordi Sierra

19 / 30


Law of Popular Non-referendary Consults. - Organizations.

Manresadecideix

UDC

Comunidad Valenciana

Senado

Upyd

Omnium

Tribunal Constitucional

Moviment Arenyenc

PP

Juzgado De Lo Contencioso

Parlament

GreenpeaceSolidaritat

ERC PSC

Podemos

PSOE

Icv-euia

Ciudadanos

Congreso Ciu

Compromis

CDCCUP

BNG

Barcelona DecideixReagrupament

Bildu

ANC

Els Verds

20 / 30


Law of Popular Non-referendary Consults. - Persons.

Lehendakari Ibarretxe

Manuel Fraga

Carles Mora

Joan Ridao

Uriel Bertran

Felip Puig

Jordi Fabrega

Joan Saura

Miquel Calcada

Joan Puigcercos

Lluis Corominas

Jaume De Frontanya

Jose Blanco

Jose Manuel Durao BarrosoJose Luis Rodriguez Zapatero

Mariano Rajoy

Josep Maria Pelegri

Josep Lluis Carod-roviraJosep Antoni DuranJose Zaragoza

Maria Emilia Casas

Maria Mas Jose Montilla

Leon

Josep Pique

Josep Camprubi

Lauren Uria

Ramon Torramade

Jordi Pujol

Josep Rull

Pedro Sanchez

Rodrigo Rato

Javier Arenas

Jose Maria Aznar

Maria Dolores De Cospedal

Carles MartiJordi Hereu

Joan ClosXavier Trias

Jordi Portabella

Joan Ferran

Ricard Goma

Ferran Mascarell

Pasqual Maragall

Franco

Miquel Iceta

Soraya Saenz De Santamaria

Patxi Lopez

Angel Acebes

Pere Jover

Marc Carrillo

Eduardo Zaplana

Oriol Pujol

Dolors Camats

Jordi Molto

Laia Bonet

Joan Botella

Joana Ortega

Joan Tarda

Santiago Rodriguez

Joan Herrera

Ferran Requejo

Abogado Del Estado

Andreu Mumbru

Artur Mas

Assumpta Escarp

Antoni Castells

Jone Goyricelaya

Ernest Benach

Angel Ros

Francesc Homs

Jordi Ausas

Nuria De Gispert

Oriol Junqueras

Albert RiveraJordi Turull

Marta Rovira

Alicia Sanchez-camacho

Ramon Espadaler

Pere Navarro

David Fernandez

Maurici Lucena

Alfredo Perez RubalcabaCarme Forcadell

Anna Simo

Joan Ignasi ElenaJoan Rigol

Esperanza Aguirre

Jose Manuel Soria

Laia Ortiz

Carles Viver Pi-sunyer

Rosa Diez

Carme Garcia

Josep Duran Lleida

Alfred Morales

Josep Maria Alvarez

Muriel Casals

Alfred Bosch

Cayo Lara

Pedro Arriola

Jose Maria Mena

Josep Maria Terricabras

Dolors Batalla

Arnaldo Otegi

Joan Carles Gallego

Alicia Sanchez Camacho

Jose Domingo

Angel Colom

Carme Capdevila

Joan Carretero

Francesc Ribera

Ahora Marti

Ernest Maragall

Jorge Fernandez Diaz

Alfons Lopez Tena

Alfonso Alonso

Marina Llansana

Marc Sanglas

Alberto Ruiz Gallardon David Cameron

Lluis Companys

Carme Chacon

Santi Vila

Jose Antonio Perez Tapias

Jordi Guillot

Josep Felix Ballesteros

Isabel ValletAlberto Fernandez Diaz

Antonio Hernando

Cristobal Montoro

Jaume Collboni

Ban Ki-moon

Ada Colau

21 / 30


Conclusions.

1 An unbiased, low-cost, automated tool to aid the process ofPolicy Network generation and analysis.

2 The system automatically:

1 Detect entities related to a bill.2 Computes and thresholds similarity measures for SN

generation.

3 The method works better for finding relationships betweenorganizations than for persons, particularly politicians.

22 / 30


Contributions.

1 The use of bills as a cornerstone relating political actors,allowing to:

Understand better the discovered relations.Find fine-grained relationships which would otherwise bemissed.

2 A method for combining parliamentary open data and newspapers for PN generation.

3 An unsupervised method for automatically detecting relevantentities of a given topic from a corpus of documents given aset of seed entities.

23 / 30


Future work.

1 A more rigorous evaluation and problem definition.

2 Improving the PN generation phase.

3 Generative models.

4 Use-case driven PN generation.

5 Time component.

6 Signed Social Network Analysis

24 / 30


The end.

Merci beacoup!Gracies!Grazie!Multumesc!

Questions?

25 / 30


Understanding the representation of entities anddocuments.

26 / 30


References I

Roger B Bradford, Application of latent semantic indexing ingenerating graphs of terrorist networks, Intelligence andSecurity Informatics, Springer, 2006, pp. 674–675.

Jesus Espinal-Enrıquez, J Mario Siqueiros-Garcıa, RodrigoGarcıa-Herrera, and Sergio Antonio Alcala-Corona, Aliterature-based approach to a narco-network, SocialInformatics, Springer, 2014, pp. 97–101.

James H Fowler, Connecting the congress: A study ofcosponsorship networks, Political Analysis 14 (2006), no. 4,456–487.

Justin H Kirkland, The relational determinants of legislativeoutcomes: Strong and weak ties between legislators, TheJournal of Politics 73 (2011), no. 03, 887–898.

27 / 30


References II

Theodosis Moschopoulos, Elias Iosif, Leeda Demetropoulou,Alexandros Potamianos, and Shrikanth Shri Narayanan,Toward the automatic extraction of policy networks using weblinks and documents, Knowledge and Data Engineering, IEEETransactions on 25 (2013), no. 10, 2404–2417.

David Newman, Chaitanya Chemudugunta, Padhraic Smyth,and Mark Steyvers, Analyzing entities and topics in newsarticles using statistical topic models, Intelligence and SecurityInformatics, Springer, 2006, pp. 93–104.

Bruno Pouliquen, Ralf Steinberger, and Clive Best, Automaticdetection of quotations in multilingual news, Proceedings ofRecent Advances in Natural Language Processing, 2007,pp. 487–492.

28 / 30


References III

Bruno Pouliquen, Ralf Steinberger, Camelia Ignat, and TamaraOellinger, arXiv preprint cs/0609066 (2006).

Hristo Tanev, Unsupervised learning of social networks from amultiple-source news corpus, MuLTISOuRcE, MuLTILINguALINfORMATION ExTRAc-TION ANd SuMMARIzATION(2007), 33.

Matt Thomas, Bo Pang, and Lillian Lee, Get out the vote:Determining support or opposition from congressionalfloor-debate transcripts, Proceedings of the 2006 conferenceon empirical methods in natural language processing,Association for Computational Linguistics, 2006, pp. 327–335.

29 / 30


References IV

Dmitry Zelenko, Chinatsu Aone, and Anthony Richardella,Kernel methods for relation extraction, The Journal ofMachine Learning Research 3 (2003), 1083–1106.

30 / 30

dama - final presentation

Documents