Lecture Notes in Artificial Intelligence 7026

Subseries of Lecture Notes in Computer Science

LNAI Series Editors

Randy GoebelUniversity of Alberta, Edmonton, Canada

Yuzuru TanakaHokkaido University, Sapporo, Japan

Wolfgang WahlsterDFKI and Saarland University, Saarbrcken, Germany

LNAI Founding Series Editor

Joerg SiekmannDFKI and Saarland University, Saarbrcken, Germany

Luis Antunes H. Sofia Pinto (Eds.)

Progress inArtificial Intelligence15th Portuguese Conferenceon Artificial Intelligence, EPIA 2011Lisbon, Portugal, October 10-13, 2011Proceedings

13

Series Editors

Randy Goebel, University of Alberta, Edmonton, CanadaJrg Siekmann, University of Saarland, Saarbrcken, GermanyWolfgang Wahlster, DFKI and University of Saarland, Saarbrcken, Germany

Volume Editors

Luis AntunesGUESS/LabMAg/Universidade de LisboaFaculdade de Cincias, Departamento de InformticaCampo Grande, 749-016 Lisboa, PortugalE-mail: xarax@di.fc.ul.pt

H. Sofia PintoInstituto Superior Tcnico, ISTDepartment of Computer Science and Engineering, INESC-IDAvenida Rovisco Pais, 1049-001 Lisboa, PortugalE-mail: sofia@ontol.inesc-id.pt

ISSN 0302-9743 e-ISSN 1611-3349ISBN 978-3-642-24768-2 e-ISBN 978-3-642-24769-9DOI 10.1007/978-3-642-24769-9Springer Heidelberg Dordrecht London New York

Library of Congress Control Number: 2011938667

CR Subject Classification (1998): I.2, J.4, H.3, H.5.2, I.5, I.4, F.4.1, K.4

LNCS Sublibrary: SL 7 Artificial Intelligence

Springer-Verlag Berlin Heidelberg 2011This work is subject to copyright. All rights are reserved, whether the whole or part of the material isconcerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publicationor parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965,in its current version, and permission for use must always be obtained from Springer. Violations are liableto prosecution under the German Copyright Law.The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply,even in the absence of a specific statement, that such names are exempt from the relevant protective lawsand regulations and therefore free for general use.

Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Preface

The Portuguese Conference on Artificial Intelligence (EPIA) was first establishedin 1985, and since 1989 it has occurred biannually and is run as an internationalconference. It aims to promote research in artificial intelligence and has been alively scientific event for the exchange of ideas for more than 20 years. Research inartificial intelligence is deeply grounded in Portugal and its importance is visiblenot only by the sheer number of researchers holding a PhD (over 100), but also bythe success of numerous important conferences organized in Portugal. EPIA hasbeen organized all over the country, raising considerable success and interest inthe several local scientific communities, while never ceasing to attract researchersand practitioners from the whole country and a considerable participation offoreign scientists.

The 15th Portuguese Conference on Artificial Intelligence, EPIA 2011, tookplace in Lisbon at the Faculdade de Ciencias da Universidade de Lisboa (FCUL)and was co-organized by LabMAg/FCUL, INESC-ID, IST, and ISEL.

As in previous recent editions, EPIA 2011 was organized as a set of selectedand dedicated tracks. The tracks considered in this edition were:

Affective Computing Ambient Intelligence Environments Artificial Intelligence Methodologies for Games Artificial Intelligence in Transportation Systems Artificial Life and Evolutionary Algorithms Computational Logic with Applications General Artificial Intelligence Intelligent Robotics Knowledge Discovery and Business Intelligence Multi-Agent Systems: Theory and Applications Social Simulation and Modeling Text Mining and Applications The Doctoral Symposium on Artificial Intelligence

In this edition, a total of 203 contributions were received from 32 countries.All papers were reviewed in a double-blind process by at least three differentreviewers. In some cases, contributions were reviewed by up to five reviewers.This book groups together a set of 50 selected papers that were accepted forthe 15th Portuguese Conference on Artificial Intelligence, EPIA 2011, covering awide range of topics and perspectives. This corresponds to a 25% acceptance rate,ensuring the high quality of the event. Geographically, accepted papers camefrom Portugal, Spain, Brazil, France, Italy, Luxembourg, Mexico, and Poland(ordered by acceptance rate). However, the authors of accepted papers includedresearchers from institutions from Canada, Denmark, France, Iran, Sweden, andthe United Kingdom.

VI Preface

We wish to thank the members of all committees involved in the event, inparticular the Advisory Board, the Program Committee and the OrganizingCommittee. We would also like to thank all the Track Chairs, the reviewers andabove all the authors, without whom this event would not be possible. We wouldalso like to acknowledge and thank the use of EasyChair, which greatly easedthe management of the submissions, reviews and materials.

August 2011 Luis AntunesH. Sofia Pinto

Organization

The 15th Portuguese Conference on Artificial Inteligence (EPIA 2011) wasco-organized by the Laboratorio de Modelacao de Agentes (LabMAg)/Faculdadede Ciencias da Universidade de Lisboa (FC/UL); Instituto de Engenharia deSistemas e Computadores (INESC-ID)/Instituto Superior Tecnico da Universi-dade Tecnica de Lisboa (IST/UTL); Instituto Superior de Engenharia de Lisboa(ISEL).

General and Program Co-chairs

Luis Antunes (GUESS/LabMAg/UL)H. Sofia Pinto (INESC-ID/IST)

Organization Co-chairs

Rui Prada (INESC-ID/IST)Paulo Trigo (LabMAg/ISEL)

Advisory Board

Salvador Abreu Universidade de Evora/CENTRIAJose Julio Alferes Universidade Nova de LisboaPedro Barahona Universidade Nova de LisboaPavel Brazdil LIAAD/Universidade do PortoAmilcar Cardoso Universidade de CoimbraHelder Coelho Universidade de LisboaLuis Correia Universidade de LisboaErnesto Costa Universidade de CoimbraGael Dias Universidade da Beira InteriorPedro Rangel Henriques Universidade do MinhoJose Gabriel Lopes Universidade Nova de LisboaErnesto Morgado Siscog/Instituto Superior TecnicoJose Neves Universidade do MinhoEugenio Oliveira LIACC/Universidade do PortoArlindo Oliveira INESC-ID/Instituto Superior Tecnico/Cadence

Research Labs.Ana Paiva Instituto Superior TecnicoCarlos Ramos Instituto Superior de Engenharia do PortoLuis M. Rocha Indiana UniversityLuis Seabra Lopes Universidade de AveiroManuela Veloso Carnegie Mellon University

VIII Organization

Conference Track Chairs

Affective Computing Goreti Marreiros Andrew Ortony Ana Paiva

Ambient Intelligence Environments Paulo Novais Ana Almeida Sara Rodrguez Gonzalez

Artificial Intelligence Methodologies for Games Lus Paulo Reis Carlos Martinho Pedro Miguel Moreira Pedro Mariano

Artificial Intelligence in Transportation Systems Rosaldo Rossetti Elisabete Arsenio Jorge Lopes Matteo Vasirani

Artificial Life and Evolutionary Algorithms Sara Silva Francisco B. Pereira Leonardo Vannesch

Computational Logic with Applications Paulo Moura Vitor Beires Nogueira

General Artificial Intelligence H. Sofia Pinto Luis Antunes

Intelligent Robotics Lus Paulo Reis Lus Correia Nuno Lau

Knowledge Discovery and Business Intelligence Paulo Cortez Nuno Marques Lus Cavique Joao Gama Manuel Filipe Santos

Multi-Agent Systems: Theory and Applications Paulo Urbano Cesar Analide Fernando Lopes Henrique Lopes Cardoso

Organization IX

Social Simulation and Modeling Joao Balsa Antonio Carlos da Rocha Costa Armando Geller

Text Mining and Applications Joaquim Francisco Ferreira Da Silva Vitor Jorge Ramos Rocio Gael Dias Jose Gabriel Pereira Lopes

Doctoral Symposium on Artificial Intelligence Paulo Novais Cesar Analide Pedro Henriques

Program Committee

Affective Computing

Cesar Analide Universidade do MinhoAntonio Camurri University of GenoaAmilcar Cardoso University of CoimbraCristiano Castelfranchi Institute of Cognitive Sciences and TechnologiesHelder Coelho University of LisbonLaurence Devillers LIMSI-CNRS, Universite P11Anna Esposito Second University of NaplesHatice Gunes Imperial College LondonJennifer HealeyIan Horswill Northwestern UniversityEva HudlickaKostas Karpouzis National Technical University of AthensJose Machado Universidade MinhoJose Neves Universidade do MinhoPaulo Novais Universidade do MinhoJuan Pavon Universidad Complutense MadridFrank Pollick University of GlasgowBoon-Kiat QuekCarlos Ramos Instituto Superior de Engenharia do PortoRicardo Santos ESTGF/IPP

Ambient Intelligence Environments

Cesar Analide Universidade do MinhoCecilio Angulo Universitat Politecnica de CatalunyaJuan Augusto University of UlsterJavier Bajo Universidad Pontificia de SalamancaCarlos Bento Universidade de Coimbra

X Organization

Lourdes Borrajo University of VigoDavide Carneiro Universidade do MinhoDiane Cook Washington State UniversityJuan Corchado University of SalamancaRicardo Costa ESTGF.IPPJose Danado YDreams, YLabs ResearchEduardo Dias CITI - FCT/UNLAntonio Fernandez University of Castilla-La ManchaLino FigueiredoDiego Gachet Universidad Europea de MadridJunzhong Gu East China Normal UniversityHans Guesgen Massey UniversityJavier Jaen Polytechnic University of ValenciaRui Jose University of MinhoJoyca Lacroix Philips ResearchKristof Laerhoven TU DarmstadtGuillaume Lopez The University of TokyoJose Machado Universidade MinhoGoreti Marreiros ISEPRene MeierJose Molina Universidad Carlos III de MadridJose Neves Universidade do MinhoFrancisco Pereira Universidade de CoimbraDavy Preuveneers K.U. LeuvenCarlos Ramos Instituto Superior de Engenharia do PortoAndreas Riener Johannes Kepler University LinzFlorentino Riverola University of VigoTeresa Romao DI/FCT/UNLIchiro Satoh National Institute of InformaticsFrancisco Silva Universidade Federal do MaranhaoDante Tapia University of SalamancaMartijn Vastenburg TU DelftYu Zheng Microsoft Research Aisa

Artificial Intelligence Methodologies for Games

Pedro G. Calero Universidad Complutense de MadridMarc Cavazza University of TeessideAntonio Coelho University of OportoFrank Dignum Utrecht UniversityStefan Edelkamp University of BremenKostas Karpouzis National Technical University of AthensPeter Kissmann TZI Universitat BremenNuno Lau University of AveiroCarlos Linares Lopez Universidad Carlos IIILuiz Moniz University of Lisbon

Organization XI

Alexander Nareyek National University of SingaporeJoao Pedro Neto University of LisbonJeff Orkin MITAna Paiva Instituto Superior TecnicoFilipe Pina Seed StudiosGuilherme Raimundo Instituto Superior TecnicoRui Rodrigues University of OportoLicinio Roque Universidade de CoimbraLuis Seabra Lopes University of AveiroA. Augusto de Sousa University of OportoMichael Thielscher University of New South WalesJulian Togelius IT University of CopenhagenMarco Vala Instituto Superior TecnicoGiorgios Yannakakis IT University of CopenhagenNelson Zagalo University of Minho

Artificial Intelligence in Transportation Systems

Ana Almeida Instituto Superior de Engenharia do PortoConstantinos Antoniou National Technical University of AthensRamachandran Balakrishna Caliper CorporationFederico Barber DSIC / Technical University of ValenciaAna Bazzan Universidade Federal do Rio Grande do SulCarlos Bento Universidade de CoimbraVicent Botti DSICEduardo Camponogara Federal University of Santa CatarinaAntonio Castro University of OportoHilmi Berk Celikoglu Technical University of IstanbulHussein Dia AecomJuergen Dunkel Hannover University for Applied Sciences and

ArtsAlberto Fernandez CETINIA, Rey Juan Carlos UniversityAdriana Giret Technical University of ValenciaFranziska Kluegl Orebro UniversityMaite Lopez-Sanchez Universitat de BarcelonaHelen Ma TU/eJose Manuel Menendez Universidad Politecnica de Madrid - UPMLus Nunes ISCTECristina Olaverri University of OportoEugenio Oliveira University of Oporto LIACCLus Osorio ISELSascha Ossowski Rey Juan Carlos UniversityFrancisco Pereira Universidade de CoimbraFrancisco Reinaldo UnilesteMGLuis Paulo Reis FEUPMiguel A. Salido DSIC / Technical University of Valencia

XII Organization

Majid Sarvi Monash UniversityJuergen Sauer University of OldenburgThomas Strang DLR German Aerospace CenterMan-Chun TanJose Telhada University of MinhoHarry Timmermans Eindhoven University of TechnologyGiuseppe Vizzari University of Milano-BicoccaFei Wang Xian Jiaotong UniversityDanny Weyns Katholieke Universiteit Leuven

Artificial Life and Evolutionary Algorithms

Wolfgang Banzhaf Memorial University of NewfoundlandHelio J. C. Barbosa Laboratorio Nacional de Computacao CientficaDaniela Besozzi University of MilanChristian Blum Universitat Politecnica de CatalunyaStefano Cagnoni University of ParmaPhilippe Caillou LRI, University of Paris Sud 11Luis Correia University of LisbonErnesto Costa University of CoimbraCarlos Cotta Universidad de MalagaIvanoe De Falco ICAR - CNRKalyanmoy Deb Indian Institute of Technology KanpurAntonio Della Cioppa University of SalernoAniko Ekart Aston UniversityAnna I. Esparcia-Alcazar S2 Grupo & Universidad Politecnica de ValenciaCarlos M. Fernandes University of GranadaJames A. Foster University of IdahoAntonio Gaspar-Cunha University of MinhoCarlos Gershenson Universidad Nacional Autonoma de MexicoMario Giacobini University of TurinJin-Kao Hao Universite AngersInman Harvey University of SussexWilliam B. Langdon University of EssexArnaud Liefooghe Universite Lille 1Fernando Lobo University of AlgarvePenousal Machado University of CoimbraAna Madureira Instituto Superior de Engenharia do PortoPedro Mariano University of LisbonRui Mendes CCTC - University of MinhoTelmo Menezes CNRS, ParisJuan Julian Merelo Guervos University of GranadaZbigniew Michalewicz University of AdelaideAlberto Moraglio University of KentShin Morishita University of YokohamaUna-May OReilly MIT

Organization XIII

Lus Paquete University of CoimbraAgostinho Rosa Technical University of LisbonMarc Schoenauer INRIARoberto Serra University of Modena e Reggio EmiliaAnabela Simoes Polytechnic Institute of CoimbraThomas Stuetzle Universite Libre de BruxellesRicardo H. C. Takahashi Universidade Federal de Minas GeraisJorge Tavares University of CoimbraLeonardo Trujillo Reyes Instituto Tecnologico de Tijuana

Computational Logic with Applications

Salvador Abreu Universidade de Evora and CENTRIAJose Julio Alferes Universidade Nova de LisboaRoberto Bagnara University of ParmaVtor Costa Universidade do PortoBart Demoen K.U. LeuvenDaniel Diaz Universite de Paris IPaulo Gomes CMS - CISUC, University of CoimbraGopal Gupta University of Texas at DallasAngelika Kimmig K.U. LeuvenJoao Leite Universidade Nova de LisboaAxel Pollers DERI, National University of Ireland, GalwayEnrico Pontelli New Mexico State UniversityPeter Robinson The University of QueenslandJoachim Schimpf Monash UniversityDavid Warren University of Stony Brook

General Artificial Intelligence

Salvador Abreu Universidade de Evora and CENTRIAJose Julio Alferes Universidade Nova de LisboaPedro Barahona Universidade Nova de LisboaPavel Brazdil LIAAD, University of OportoAmilcar Cardoso University of CoimbraHelder Coelho University of LisbonLuis Correia University of LisbonErnesto Costa University of CoimbraGael Dias University of Beira InteriorPedro Rangel Henriques Universidade do MinhoJose Gabriel Lopes Universidade Nova de LisboaErnesto Morgado Siscog and ISTJose Neves Universidade do MinhoEugenio Oliveira University of Oporto LIACCArlindo Oliveira IST/INESC-ID and Cadence Research

Laboratories

XIV Organization

Ana Paiva Instituto Superior TecnicoCarlos Ramos Instituto Superior de Engenharia do PortoLuis M. Rocha Indiana UniversityLuis Seabra Lopes University of AveiroManuela Veloso Carnegie Mellon University

Intelligent Robotics

Cesar Analide Universidade do MinhoKai Arras University of FreiburgStephen Balakirsky NISTJacky Baltes University of ManitobaReinaldo Bianchi Centro Universitario da FEIRodrigo Braga FEUPCarlos Carreto Instituto Politecnico da GuardaXiaoping Chen University of Science and Technology of ChinaAnna Helena Costa University of Sao PauloAugusto Loureiro Da Costa Universidade Federal da BahiaJorge Dias University of CoimbraMarco Dorigo Universite Libre de BruxellesPaulo Goncalves Polytechnic Institute of Castelo BrancoJohn Hallam University of Southern DenmarkKasper Hallenborg Maersk Institute, University of Southern

DenmarkHuosheng Hu University of EssexFumiya IidaLuca Iocchi Sapienza University of RomePedro Lima Instituto Superior TecnicoAntonio Paulo Moreira FEUPAntonio J. R. Neves IEETA, University of AveiroUrbano Nunes University of CoimbraFernando Osorio USP - ICMC - University of Sao PauloEnrico Pagello University of PaduaArmando J. Pinho University of AveiroMikhail Prokopenko CSIROIsabel Ribeiro Instituto Superior Tecnico / ISRMartin RiedmillerLuis Seabra Lopes University of AveiroSaeed Shiry Ghidary Amirkabir University of TechnologyArmando Sousa ISR-P, DEEC, FEUPGuy Theraulaz CNRS CRCAFlavio Tonidandel Centro Universitario da FEIPaulo Urbano University of LisbonManuela Veloso Carnegie Mellon University

Organization XV

Knowledge Discovery and Business Intelligence

Carlos Alzate K.U. LeuvenOrlando Belo University of MinhoAlbert Bifet University of WaikatoAgnes BraudRui Camacho LIACC/FEUP University of OportoLogbing Cao University of Technology SydneyAndre Carvalho Universidade de Sao PauloNing Chen GECAD, Instituto Superior de Engenharia do

PortoJose Costa Universidade Federal do Rio Grande do NorteCarlos Ferreira LIAAD INESC PortoPeter Geczy AISTPaulo Gomes CMS - CISUC, University of CoimbraBeatriz Iglesia University of East AngliaElena Ikonomovska IJS - Department of Knowledge TechnologiesAlpio Jorge FCUP / LIAAD, INESC PortoStephane Lallich University of Lyon 2Phillipe Lenca Telecom BretagneStefan LessmannHongbo Liu Dalian Maritime UniversityVtor Lobo CINAV - Escola NavalJose Machado Universidade do MinhoPatrick Meyer Institut Telecom BretagneSusana Nascimento Universidade Nova de LisboaFatima Rodrigues Institute of Engineering of PortoJoaquim Silva FCT/UNLCarlos Soares University of OportoMurate Testik Hacettepe UniversityTheodore TrafalisArmando VieiraAline Villavicencio UFRGS and University of Bath

Multi-Agent Systems: Theory and Applications

Huib Aldewereld University of UtrechtReyhan AydoganJavier Bajo Universidad Pontificia de SalamancaJoao Balsa GUESS/LabMAg/University of LisbonOlivier Boissier ENS Mines Saint-EtienneLus Botelho ISCTE-IULJuan Burgillo University of VigoJavier Carbo UC3MAmilcar Cardoso University of CoimbraCristiano Castelfranchi Institute of Cognitive Sciences and TechnologiesYun-Gyung Cheong Samsung Advanced Institute of TechnologyHelder Coelho University of Lisbon

XVI Organization

Juan M. Corchado University of SalamancaLuis Correia University of LisbonYves Demazeau CNRS - Laboratoire LIGFrank Dignum Utrecht UniversityVirginia Dignum TU DelftAmal Elfallah LIP6 - University of Pierre and Marie CurieMarc Esteva IIIA-CSICPaulo Ferreira Jr. Universidade Federal de PelotasMichael Fisher University of LiverpoolNicoletta Fornara University of LuganoYaAkov Gal Ben-Gurion University of the NegevAlejandro Guerra Universidad VeracruzanaJomi Hubner Federal University of Santa CatarinaWojtek Jamroga University of LuxembourgF. Jordan SrourNuno Lau University of AveiroJoao Leite Universidade Nova de LisboaChristian Lematre Universidad Autonoma Metropolitana, UAMLuis Macedo University of CoimbraPedro Mariano University of LisbonJohn-Jules Meyer Utrecht UniversityLuis Moniz University of LisbonPavlos Moraitis Paris Descartes UniversityJorg Muller Clausthal University of TechnologyPablo Noriega IIIAPaulo Novais Universidade do MinhoEugenio Oliveira University of Oporto LIACCAndrea Omicini Alma Mater Studiorum Universita di BolognaSantiago Ontanon Villar IIIA-CSICAntonio Pereira LIACC - DEI - FEUPAlexander Pokahr University of HamburgLuis Paulo Reis FEUPAna Paula Rocha FEUPAntonio Carlos da Rocha

Costa Universidade Federal do Rio GrandeJordi Sabater-Mir IIIA-CSICMurat Sensoy University of AberdeenOnn Shehory IBM Haifa Research LabJaime Sichman University of Sao PauloPaolo Torroni University of BolognaWamberto Vasconcelos University of AberdeenLaurent Vercouter ISCOD Team, Ecole des Mines de Saint-EtienneRosa Vicari Federal University of Rio Grande do SulNeil Yorke-Smith American University of Beirut and SRI

International

Organization XVII

Social Simulation and Modeling

Diana Adamatti Universidade Federal do Rio GrandeFrederic Amblard Universite des Sciences Sociales Toulouse 1Pedro AndradeLuis Antunes GUESS/LabMAg/University of LisbonPedro CamposAmilcar Cardoso University of CoimbraCristiano Castelfranchi Institute of Cognitive Sciences and TechnologiesHelder Coelho University of LisbonRosaria Conte Institute of Cognitive Sciences and TechnologiesNuno David ISCTEPaul Davidsson Blekinge Institute of TechnologyGracaliz Dimuro Universidade Federal do Rio GrandeBruce Edmonds Manchester Metropolitan University Business

SchoolNigel Gilbert University of SurreyLaszlo Gulyas Aitia International, Inc.Samer Hassan Universidad Complutense de MadridRainer Hegselmann Bayreuth UniversityWander Jager University of GroningenMarco Janssen Arizona State UniversityMaciej Latek George Mason UniversityJorge Louca ISCTEGustavo Lugo Federal University of Technology-ParanaLus Macedo University of CoimbraJean-Pierre Muller CIRADFernando Neto University of PernambucoFabio Okuyama IFRS - Campus Porto AlegreJuan Pavon Universidad Complutense MadridJuliette Rouchier CNRS-GREQAMDavid Sallach Argonne National Laboratory and the University

of ChicagoJaime Sichman University of Sao PauloPatricia Tedesco Center for Informatics / UFPEOswaldo Teran Universidad de Los AndesTakao TeranoKlaus Troitzsch University of Koblenz-LandauNatalie Van Der WalHarko Verhagen Stockholm University/KTH

Text Mining and Applications

Helena Ahonen-MykaSophia Ananiadou University of ManchesterJoao Balsa GUESS/LabMAg/University of Lisbon

XVIII Organization

Antonio Branco University of LisbonPavel Brazdil LIAAD, University of OportoLuisa CoheurBruno Cremilleux University of CaenWalter Daelemans University of AntwerpEric De La Clergerie INRIAAntoine Doucet University of CaenTomaz Erjavec Jozef Stefan InstituteMarcelo Finger Universidade de Sao PauloPablo Gamallo University of Santiago de CompostelaJoao Graca ISTBrigitte Grau LIMSI (CNRS)Gregory Grefenstette ExaleadDiana Inkpen University of OttawaMark Lee University of BirminghamNadine Lucas GREYC CNRS Caen UniversityBelinda Maia University of OportoNuno Marques Universidade Nova de LisboaAdeline Nazarenko LIPN UMR CNRS - Universite Paris NordManuel Palomar University of AlicantePaulo Quaresma Universidade de EvoraIrene Rodrigues Universidade de EvoraAntonio Sanfilippo Pacific Northwest National LaboratoryFrederique Segond XeroxIsabelle Tellier LIFO - University of OrleansRenata Vieira PUC-RSManuel Vilares Ferro University of VigoAline Villavicencio UFRGS and University of BathChristel Vrain LIFO - University of OrleansPierre Zweigenbaum LIMSI-CNRS

Doctoral Symposium on Artificial Intelligence

Salvador Abreu Universidade de Evora and CENTRIAJose Julio Alferes Universidade Nova de LisboaVictor Alves Universidade do MinhoPedro Barahona Universidade Nova de LisboaAmilcar Cardoso University of CoimbraLuis Correia University of LisbonPaulo Cortez University of MinhoGael Dias University of Beira InteriorEduardo Ferme Universidade da MadeiraJoaquim Filipe EST-Setubal/IPSJoao Gama University of OportoPaulo Gomes CMS - CISUC, University of CoimbraJoao Leite Universidade Nova de Lisboa

Organization XIX

Jose Gabriel Lopes Universidade Nova de LisboaJose Machado Universidade do MinhoSara Madeira ISTPedro Mariano University of LisbonGoreti Marreiros ISEPPaulo Oliveira UTADPaulo Quaresma Universidade de EvoraCarlos Ramos Instituto Superior de Engenharia do PortoLuis Paulo Reis FEUPAna Paula Rocha FEUPManuel Filipe Santos University of MinhoLus Torgo Universidade do PortoPaulo Urbano University of Lisbon

Additional Reviewers

Aiguzhinov, ArturAlmajano, PabloCalejo, MiguelCazzaniga, PaoloCosta, FranciscoCraveirinha, RuiDarriba Bilbao, Victor ManuelDay, JarethDodds, RicardoEitan, OranFaria, Brigida MonicaFerreira, LigiaFiser, DarjaGarcia, MarcosGebhardt, BirtheGoncalo Oliveira, HugoGoncalves, PatriciaGunura, KeithKurz, MarcLeach, DerekLjubesic, NikolaLloret, Elena

Lobo, JorgeMagalhaes, JoaoMaheshwari, NandanMelo, DoraMigliore, DavideNeves Ferreira Da Silva, Ana PaulaNojoumian, PeymanOliveira, MarciaPimentel, CesarPretto, AlbertoRibadas Pena, FranciscoSaldanha, RicardoSilveira, SaraSlota, MartinSpanoudakis, NikolaosVale, AlbertoVentura, RodrigoWagner, MarkusWerneck, NicolauWichert, AndreasWilkens, RodrigoXue, Feng

Table of Contents

Affective Computing

Sentiment Analysis of News Titles: The Role of Entities and a NewAffective Lexicon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Daniel Loureiro, Goreti Marreiros, and Jose Neves

Ambient Intelligence Environments

Providing Location Everywhere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15Ricardo Anacleto, Lino Figueiredo, Paulo Novais, and Ana Almeida

Modeling Context-Awareness in Agents for Ambient Intelligence:An Aspect-Oriented Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

Inmaculada Ayala, Mercedes Amor Pinilla, and Lidia Fuentes

Developing Dynamic Conflict Resolution Models Based on theInterpretation of Personal Conflict Styles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

Davide Carneiro, Marco Gomes, Paulo Novais, and Jose Neves

Organizations of Agents in Information Fusion Environments . . . . . . . . . . 59Dante I. Tapia, Fernando de la Prieta, Sara Rodrguez Gonzalez,Javier Bajo, and Juan M. Corchado

Artificial Intelligence Methodologies for Games

Wasp-Like Agents for Scheduling Production in Real-Time StrategyGames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

Marco Santos and Carlos Martinho

Artificial Intelligence in Transportation Systems

Operational Problems Recovery in Airlines A SpecializedMethodologies Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

Bruno Aguiar, Jose Torres, and Antonio J.M. Castro

Solving Heterogeneous Fleet Multiple Depot Vehicle SchedulingProblem as an Asymmetric Traveling Salesman Problem . . . . . . . . . . . . . . 98

Jorge Alpedrinha Ramos, Lus Paulo Reis, and Dulce Pedrosa

XXII Table of Contents

Artificial Life and Evolutionary Algorithms

Evolving Numerical Constants in Grammatical Evolution with theEphemeral Constant Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

Douglas A. Augusto, Helio J.C. Barbosa, Andre M.S. Barreto, andHeder S. Bernardino

The Evolution of Foraging in an Open-Ended SimulationEnvironment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

Tiago Baptista and Ernesto Costa

A Method to Reuse Old Populations in Genetic Algorithms . . . . . . . . . . . 138Mauro Castelli, Luca Manzoni, and Leonardo Vanneschi

Towards Artificial Evolution of Complex Behaviors Observed in InsectColonies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

Miguel Duarte, Anders Lyhne Christensen, and Sancho Oliveira

Network Regularity and the Influence of Asynchronism on theEvolution of Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

Carlos Grilo and Lus Correia

The Squares Problem and a Neutrality Analysis with ReNCoDe . . . . . . . 182Rui L. Lopes and Ernesto Costa

Particle Swarm Optimization for Gantry Control: A TeachingExperiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196

Paulo B. de Moura Oliveira, Eduardo J. Solteiro Pires, andJose Boaventura Cunha

Evolving Reaction-Diffusion Systems on GPU . . . . . . . . . . . . . . . . . . . . . . . 208Lidia Yamamoto, Wolfgang Banzhaf, and Pierre Collet

Computational Logic with Applications

Optimal Divide and Query . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224David Insa and Josep Silva

A Subterm-Based Global Trie for Tabled Evaluation of LogicPrograms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239

Joao Raimundo and Ricardo Rocha

General Artificial Intelligence

Intention-Based Decision Making with Evolution Prospection . . . . . . . . . . 254The Anh Han and Lus Moniz Pereira

Table of Contents XXIII

Unsupervised Music Genre Classification with a Model-BasedApproach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268

Lus Barreira, Sofia Cavaco, and Joaquim Ferreira da Silva

Constrained Sequential Pattern Knowledge in Multi-relationalLearning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282

Carlos Abreu Ferreira, Joao Gama, and Vtor Santos Costa

Summarizing Frequent Itemsets via Pignistic Transformation . . . . . . . . . . 297Francisco Guil-Reyes and Mara-Teresa Daza-Gonzalez

A Simulated Annealing Algorithm for the Problem of Minimal AdditionChains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311

Adan Jose-Garcia, Hillel Romero-Monsivais,Cindy G. Hernandez-Morales, Arturo Rodriguez-Cristerna,Ivan Rivera-Islas, and Jose Torres-Jimenez

Novelty Detection Using Graphical Models for Semantic RoomClassification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326

Andre Susano Pinto, Andrzej Pronobis, and Luis Paulo Reis

Intelligent Robotics

A Reinforcement Learning Based Method for Optimizing the Processof Decision Making in Fire Brigade Agents . . . . . . . . . . . . . . . . . . . . . . . . . . 340

Abbas Abdolmaleki, Mostafa Movahedi, Sajjad Salehi,Nuno Lau, and Lus Paulo Reis

Humanoid Behaviors: From Simulation to a Real Robot . . . . . . . . . . . . . . . 352Edgar Domingues, Nuno Lau, Bruno Pimentel, Nima Shafii,Lus Paulo Reis, and Antonio J.R. Neves

Market-Based Dynamic Task Allocation Using Heuristically AcceleratedReinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365

Jose Angelo Gurzoni Jr., Flavio Tonidandel, andReinaldo A.C. Bianchi

Shop Floor Scheduling in a Mobile Robotic Environment . . . . . . . . . . . . . . 377Andry Maykol Pinto, Lus F. Rocha, Antonio Paulo Moreira, andPaulo G. Costa

Humanized Robot Dancing: Humanoid Motion Retargeting Based in aMetrical Representation of Human Dance Styles . . . . . . . . . . . . . . . . . . . . . 392

Paulo Sousa, Joao L. Oliveira, Luis Paulo Reis, and Fabien Gouyon

XXIV Table of Contents

Knowledge Discovery and Business Intelligence

Bankruptcy Trajectory Analysis on French Companies UsingSelf-Organizing Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407

Ning Chen, Bernardete Ribeiro, and Armando S. Vieira

Network Node Label Acquisition and Tracking . . . . . . . . . . . . . . . . . . . . . . . 418Sarvenaz Choobdar, Fernando Silva, and Pedro Ribeiro

Learning to Rank for Expert Search in Digital Libraries of AcademicPublications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431

Catarina Moreira, Pavel Calado, and Bruno Martins

Thematic Fuzzy Clusters with an Additive Spectral Approach . . . . . . . . . 446Susana Nascimento, Rui Felizardo, and Boris Mirkin

Automatically Enriching a Thesaurus with Information fromDictionaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462

Hugo Goncalo Oliveira and Paulo Gomes

Visualizing the Evolution of Social Networks . . . . . . . . . . . . . . . . . . . . . . . . 476Marcia Oliveira and Joao Gama

Using Data Mining Techniques to Predict Deformability Properties ofJet Grouting Laboratory Formulations over Time . . . . . . . . . . . . . . . . . . . . 491

Joaquim Tinoco, Antonio Gomes Correia, and Paulo Cortez

Multi-Agent Systems: Theory and Applications

Doubtful Deviations and Farsighted Play . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506Wojciech Jamroga and Matthijs Melissen

Uncertainty and Novelty-Based Selective Attention in the CollaborativeExploration of Unknown Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521

Luis Macedo, Miguel Tavares, Pedro Gaspar, and Amlcar Cardoso

A Dynamic Agents Behavior Model for Computational Trust . . . . . . . . . . 536Joana Urbano, Ana Paula Rocha, and Eugenio Oliveira

The BMC Method for the Existential Part of RTCTLK and InterleavedInterpreted Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 551

Bozena Wozna-Szczesniak, Agnieszka Zbrzezny, andAndrzej Zbrzezny

Social Simulation and Modeling

Building Spatiotemporal Emotional Maps for Social Systems . . . . . . . . . . 566Pedro Catre, Luis Cardoso, Luis Macedo, and Amlcar Cardoso

Table of Contents XXV

Text Mining and Applications

An Exploratory Study on the Impact of Temporal Features on theClassification and Clustering of Future-Related Web Documents . . . . . . . 581

Ricardo Campos, Gael Dias, and Alpio Jorge

Using the Web to Validate Lexico-Semantic Relations . . . . . . . . . . . . . . . . 597Hernani Pereira Costa, Hugo Goncalo Oliveira, and Paulo Gomes

A Resource-Based Method for Named Entity Extraction andClassification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 610

Pablo Gamallo and Marcos Garcia

Measuring Spelling Similarity for Cognate Identification . . . . . . . . . . . . . . 624Lus Gomes and Jose Gabriel Pereira Lopes

Identifying Automatic Posting Systems in Microblogs . . . . . . . . . . . . . . . . . 634Gustavo Laboreiro, Lus Sarmento, and Eugenio Oliveira

Determining the Polarity of Words through a Common OnlineDictionary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649

Antonio Paulo-Santos, Carlos Ramos, and Nuno C. Marques

A Bootstrapping Approach for Training a NER with ConditionalRandom Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 664

Jorge Teixeira, Lus Sarmento, and Eugenio Oliveira

Doctoral Symposium on Artificial Intelligence

Domain-Splitting Generalized Nogoods from Restarts . . . . . . . . . . . . . . . . . 679Lus Baptista and Francisco Azevedo

A Proposal for Transactions in the Semantic Web . . . . . . . . . . . . . . . . . . . . 690Ana Sofia Gomes and Jose Julio Alferes

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705

Sentiment Analysis of News Titles

The Role of Entities and a New Affective Lexicon

Daniel Loureiro1, Goreti Marreiros1, and Jose Neves2

1 ISEP, GECAD - Knowledge Engineering and Decision Support Group, Portugalloureiro97@gmail.com, mgt@isep.ipp.pt

2 Minho University, Portugaljneves@di.uminho.pt

Abstract. The growth of content on the web has been followed byincreasing interest in opinion mining. This field of research relies onaccurate recognition of emotion from textual data. Theres been muchresearch in sentiment analysis lately, but it always focuses on the sameelements. Sentiment analysis traditionally depends on linguistic corpora,or common sense knowledge bases, to provide extra dimensions of infor-mation to the text being analyzed. Previous research hasnt yet exploreda fully automatic method to evaluate how events associated to certainentities may impact each individuals sentiment perception. This projectpresents a method to assign valence ratings to entities, using informationfrom their Wikipedia page, and considering user preferences gatheredfrom the users Facebook profile. Furthermore, a new affective lexicon iscompiled entirely from existing corpora, without any intervention fromthe coders.

1 Introduction

The ability to recognize and classify emotion has been the topic of much researcheffort over the history of computer science. Pioneers in artificial intelligence, suchas Minsky[16] and Picard[22], have described it as one of the hallmarks of theirfield. The growth of interest in opinion mining for content on the web (e.g., prod-uct reviews, news articles, etc.)[20] combined with the permanence of textuallybased computer interfaces reinforces the need for a computer system capable ofclassifying emotion accurately[9]. The intricacies of emotion expression stand asa challenge both for the computer systems that try to decode it and the psycho-logical theories that try to explain it.

Much research into sentiment analysis uses psychological models of emotion.The OCC emotion model[18] has found widespread usage, as it provides a cog-nitive theory for the classification of discrete emotions (e.g. joy, relief) that canbe characterized by a computationally tractable set of rules. Sometimes, othermodels such as Ekmans six basic categories of emotions are used instead[4],though that one isnt associated to a cognitive theory and lacks a well-definedrule-based structure. The preferred psychological theory for the classification

L. Antunes and H.S. Pinto (Eds.): EPIA 2011, LNAI 7026, pp. 114, 2011.c Springer-Verlag Berlin Heidelberg 2011

2 D. Loureiro, G. Marreiros, and J. Neves

of emotion is usually accompanied by natural language processing (NLP) tech-niques for a detailed understanding of the text.

This project uses the OCC emotion model and several techniques for lin-guistic parsing, word stemming, word sense disambiguation and named entityrecognition. This linguistic data is enriched using large databases in an attemptto extract enough information for a substantiated classification. The two mainnovelties introduced in this paper are the following: (1) the automatic genera-tion of an objective affective lexicon, and (2) a user centric emotion evaluationtechnique that considers personal preferences towards entities.

The rest of the paper is organized as follows. Section 2 starts by exposingthe motivation behind this project. Section 3 briefly explains some similar ap-proaches for sentiment analysis. Sections 5 through 8 explore the implemen-tation, discussing our methods for parsing the news title, word-level valenceassignment, recognizing entities and classifying emotion. A description of theresults has been detailed in section 9.

2 Background

This project is a follow-up to the work done by Vinagre, on user-centric emotionclassification of news headlines[17]. Vinagre proposed a system for gatheringnews items of a specific emotion based on a users preferences and personality.It classified the emotion of a news headline according to the OCC emotionmodel and assessed a users personality based on its Big-Five[8] measurement.Theoretically, this provided the system with the ability to predict the usersreaction to the news item. However, the sentiment analysis procedure was hand-made for each news headline. This project is an effort to implement a method forautomatic emotion classification that was lacking in the aforementioned work.

3 Related Work

Previous approaches in sentiment analysis have employed many techniques likekeyword spotting, lexical affinity, machine learning or knowledge basedapproaches[9]. Each technique has its strength and weaknesses as described in[9, 20]. Here well be looking at some examples most similar to our approach.

3.1 Lexicon-Based Approaches

The first works in sentiment analysis classified text based exclusively on thepresence of affective words. This method, called keyword spotting, requires aword list with affective word valences, known as an affective lexicon. Among themost used are the General Inquirer[27] and Ortonys Affective Lexicon[19], thatprovide sets of words from different emotions categories and can be divided intopositive or negative types. The keyword spotting method relies on the occurrenceof words recognized by the affective lexicon and can be easily tricked by negation.

Sentiment Analysis of News Titles 3

For instance, while the sentence We won the lottery. would be assigned apositive valence, its negation We didnt win the lottery. would likely havethe same valence assigned to it. Moreover, sentences like Were always late formeetings. would be assigned a neutral valence due to the lack of affective words.Wilson et al.[30] proposed an interesting solution to handle the issue of negationby identifying contextual polarity and polarity shifters.

The development of linguistic corpora with information on the semantics ofwords brought with it new methods for sentiment analysis, more specifically,content-based approaches.

3.2 Content-Based Approaches

The most used resources in content-based sentiment analysis are WordNet[6]and ConceptNet[10]. WordNet is a lexical database for the English languagewith over 150,000 nouns, verbs, adjectives and adverbs organized by a numberof semantic relations such as synonym sets (known as synsets), hierarchies, andothers. ConceptNet is a common sense knowledge base with over a million factsorganized as different concepts, interconnected through a set of relations, such asUsedFor, IsA, Desires, and more. Both databases are sufficiently complexto derive an affective lexicon from a group of affective keywords.

The authors of ConceptNet developed a popular approach for sentiment anal-ysis that used the knowledge base to classify sentences into one of Ekmans sixbasic categories of emotions[9]. A set of affective concepts were manually an-notated for the emotion categories, and then used in propagation models to inferemotion in other concepts related to them. The accuracy of this approach hasnot yet been investigated.

A few years later, SenseNet[25] was developed. This approach used WordNetto detect polarity values of words and sentence-level textual data. The systemworks from a set of assumptions and one of them describes some hand-craftedequations that count the positive and negative senses from the definitions inWordNet. SenseNet also turns to ConceptNet when processing a concept notfound on its database. It only produces valenced results, therefore it can onlydifferentiate between positive, negative or neutral emotions.

Theres a significant problem with approaches dependent on ConceptNet. Thecommon sense knowledge base has a very uneven distribution of specificity. Thisincreases the noise in affect sensing from related concepts. We estimate thatWordNet suffers from the same problem to some extent, but its impact is lessenedfrom a distribution across a small group of categories (hypernyms, hyponyms,meronyms, etc.).

3.3 News Sentiment Analysis

Theres already some research in sentiment analysis for news titles specifically.The UPAR7[3] is a content-based approach that classifies emotion using both va-lence ratings and a set of booster categories of words associated to one or more

4 D. Loureiro, G. Marreiros, and J. Neves

of Ekmans six basic emotions. The valence ratings are extracted from Word-Net Affect[29] and SentiWordNet[5], while the booster categories are taken froma selected group of hypernyms from WordNet. For instance, words that haveweapon system as an hypernym, boost anger, fear and sadness emotions. Thisis an interesting approach and it performed adequately at the SemEval-2007[28]affective text task. However, its unknown if UPAR7s reliance on hypernymsand selected categories performs as well outside the SemEval task.

The same researchers that developed SenseNet later developed the Emo-tion Sensitive News Agent (ESNA)[26]. ESNA uses the sentiment ratings fromSenseNet to determine the appropriate classification for news items accordingto the OCC model of emotion. It requires users to select news sources fromRSS feeds according to their interests, and rate a list of named entities accord-ing to their personal feelings towards them. This approach is very similar toVinagres[17], from which our project stems from. Were only aware of these twoapproaches to integrate user preferences and interests into the classification ofemotion with the OCC model. Notice that while Vinagre didnt provide a methodfor automatic emotion classification, Shaikh et al. didnt provide a method forautomatic gathering of user preferences.

4 Architecture Overview

Determining the emotional classification of a news title requires parsing the sen-tence and analyzing its components. A title must be put through a pipelinewhere many different linguistic aspects are scrutinized. To get an accurate as-sessment we need to look at the semantics in work and have knowledge aboutthe content and its listener. After analyzing these aspects we look at a modelof emotions and find out the appropriate classification. Our approach pursued apipelined architecture with the following stages: Linguistic Parsing, Word-levelValence Assignment, Entity Processing and Emotion Classification, as illustratedin Figure 1.

Fig. 1. Four stage pipeline architecture. Each step contains the methods and resourcesused.

Sentiment Analysis of News Titles 5

5 Linguistic Parsing of the News Title

Natural Language Processing (NLP) is used to get a handle on the semanticrelations of the phrases in the headline. In this stage we also look at the differentsenses that a word can assume and try to map it correctly to the appropriateword on the lexicon.

5.1 Grammatical Relations

Grammatical relations are explored using a natural language parser, the Stan-ford Parser[14], which is typically used to examine the grammatical structureof sentences. The Stanford Parser is a probabilistic parser that produces themost likely analysis based on knowledge gained from hand-parsed sentences.It has some problems with the capitalization of consecutive words, as some-times found in news titles, that arent addressed in this paper. We cant usea Part-of-Speech (POS) tagger to decapitalize words because it would interferewith the entity recognition method used later. The solution is to manually fixor avoid improperly capitalized news titles. After fixing that capitalization is-sue, the parser performs adequately and we use of a specific structure it createscalled typed-dependencies. This structure provides a simple representation of thegrammatical relations between pairs of words, with over 55 types of relations.The sentence Michael Phelps won eight gold medals at Beijing. generates thefollowing typed-dependencies:

nn(Phelps, Michael) nsubj(won, Phelps)num(medals, eight) amod(medals, gold)dobj(won, medals) prep(won, at)pobj(at, Beijing)

The dependencies are relations between a governor (also known as a head)and a dependent. The grammatical relations are defined in[13]. This project uses22 of the most frequent typed-dependencies relations to organize the sentenceunder a subject-verb-object (SVO) phrase typology with details on prepositions,adjectives, modifiers, possessives, and other relations the parser may find.

The subject phrase is extracted from the governors of the nominal subject(nsubj) and passive nominal subject(nsubjpas) relations. The object phrase isextracted from the governors of the direct object (dobj) and object of a prepo-sition (pobj) relations. Finally, the verb phrase is extracted from the dependentsof all those. When the parser cant find the expected relations necessary build theSVO, it still tries to build it from some unused relations, such as prepositionalrelations. The parser is robust enough to handle most syntactic structures.

The SVO is still compounded with information from other relations. Thenoun compound modifier (nn) extends nouns, the adjectival modifier (amod)

6 D. Loureiro, G. Marreiros, and J. Neves

modifies the meaning of nouns, the possessive modifier (poss) associaties nounsto their possessives, among others. These are encoded in the SVO using specialcharacters. The prepositional phrases use square brackets, the possessives usecurly brackets and the modifiers use parenthesis.

The sentence parsed before ends up looking like this when wrapped under theSVO structure:

Michael Phelps / won [ at Beijing ] / eight medals (gold)

5.2 Word Sense Disambiguation

A central issue in NLP is the process of assigning the correct sense to a word,according to the context in which it is found. This is a key obstacle to sur-pass in text-based sentiment analysis, since assigning the wrong sense to a wordwith multiple meanings can result in a vastly different classification of the wholesentence. This can be seen in published news titles such as Drunk Gets NineMonths in Violin Case or Complaints About NBA Referees Growing Ugly,taken from[23].

Much research in the field of word sense disambiguation (WSD) uses WordNetslarge database to assign senses to words. One such solution is SenseRelate[21],which finds the correct sense for each word by determining the similarity betweenthat word and its neighbors. The result is a set of senses with maximum semanticrelatedness. This method doesnt require any training and was shown to performcompetitively.

This project gives preference to WordNet driven WSD, because the disam-biguated senses can be affectively rated with the same method for word-levelvalence assignment used for all words in the news title.

6 Word-Level Valence Assignment

Most research on sentiment analysis uses a set of affectively annotated words.This normally involves a word list where each word has a numerical valencerating assigned to it. Sometimes the authors compile a set of affective wordsand rate them individually[25, 9], or group them into emotion categories. Thispractice may be suitable when theres some expectation about which words willbe analyzed, but a general purpose affective lexicon sensor requires a word listthat is as objective as possible.

The results of an early study by Pang et al.[20] on movie reviews suggest thatcoming up with the right set of keywords might be less trivial than one mightinitially think.

Sentiment Analysis of News Titles 7

6.1 Corpus of Choice

The Affective Norms for English Words (ANEW)[2] corpus provides an extensiveword list (1034 words) with numerical ratings (1-9) compiled by several experi-mental focus groups. It considers three dimensions of affect: valence, arousal anddominance. Ratings for each range semantically from pleasant to unpleasant,excited to calm and controlling to submissive, respectively. A rating below4 is low, above 6 is high, and neutral in between. Low valences are interpretedas negative and high valences as positive, According to ANEW, the word as-sault has a low valence (unpleasant), high arousal (excited) and low dominance(submissive) while the word comfort has high valence (pleasant), low arousal(calm) and neutral dominance.

This three dimensional rating allows for finer distinctions between words andcombining dimensions results in improved performances. Combining valence andarousal reinforces the difference between valences of similar ratings, for instance,the words kindness and excitement have very similar positive valence ratingsbut the higher arousal rating for excitement reinforces its positivity. In ourwork we only use valence and arousal dimensions, as dominance ratings are tooscarcely differentiated to have an impact in this implementation.

Furthermore, the ANEW corpus has each rating accompanied by a standard-deviation value. This opens up the possibility to integrate the Big Five model[8]into the affective assessment so that depending on the users profile, differentratings can be adopted within the standard-deviation. For example, a higheragreeableness profile could use ratings within the upper range of the standard-deviation.

6.2 Beyond the Corpus

Despite ANEW being a fairly large corpus, it was found to be insufficient forreliable sensing. Word stemming enabled broader coverage of the ratings foundin the corpus, so that different words with the same word stem will return thesame rating in the word list. Mihalcea et. al[15] argued that lemmatization isntappropriate for words in the ANEW corpus because it contains more than oneword for each lemma and its different ratings would be lost. Word stemmingshares the same problem. Thus, it was decided to simply include both represen-tations in the lexicon,the word stem and the regular word.

Finally, we included some tokens that are part of English grammar and arentcovered by neither the ANEW corpus nor WordNet. These include pronouns,prepositions, determiners and numerals, totaling an addition of nearly 300 neu-trally rated tokens.The end result is a word list with more than 5000 words rated in two dimensionswhere, besides the nearly 300 words for grammar neutralization, it was fully de-rived from the ANEW corpus. Through this process a larger corpus was derivedfrom a scientifically validated one without adding any unnecessary subjectivity.

8 D. Loureiro, G. Marreiros, and J. Neves

6.3 Affinity Rating

The affective lexicon cant possibly cover all words in the English dictionary. Re-gardless of whatever techniques used to expand it, there will always be affectivewords left unrated. To cover these omissions we turn once again to WordNet,but now to infer valence ratings using the affective lexicon presented before.

There have been many approaches to rate the valences of unknown words froma preexisting lexicon using WordNet. Some assigned the unknown words ratingby averaging the valences of the words in its synonym sets[25]. Others consid-ered more relations, like hypernyms and hyponyms[3]. Our approach averagesthe synonym sets, and also the affective words contained in its definition. Weconsider the words definition because its content is often more affectively co-herent than the words on its hierarchies, which quickly get off topic. If we wishto make the system more reactive, we apply a reducing factor to the neutralratings. The calculations for valence assignment are detailed below.

PositiveWeight = AvgPosV alence AvgPosArousal (1)

NeutralWeight = AvgNtrV alence AvgNtrArousal ReducingFactor (2)

NegativeWeight = AvgNegV alence AvgNegArousal (3)

WordV alenceLabel=Max(PositiveWeight, NeutralWeight, NegativeWeight)(4)

WordV alenceIntensity = AvgXV alence where X is given in (4). (5)

7 Recognizing Entities

An often ignored aspect of sentiment analysis is the role that entities play inthe attribution of emotion. The desirability of events is tied to the subjects andobjects involved. For instance, the headline Saddam Hussein arrested in Iraq.evokes a positive emotion only when considering that the entity in case is asso-ciated to negative emotions. Other examples like, Netherlands loses World Cupto Spain can have opposite emotion classifications depending on the readerspreferences. This section explains how we solve this issue.

7.1 Extracting Entities

This project uses the Stanford Named Entity Recognizer[7] to extract enti-ties from the news titles. It can recognize the names of persons, organizationsand locations and label sequences of words accordingly. The perceived affec-tive valence of a recognized person or organization entity is determined by amethod specifically for sentiment analysis of entities. We determined that loca-tions should always have a neutral valence rating to avoid biased stances in casesof polemic locations.

Sentiment Analysis of News Titles 9

7.2 Sentiment Analysis of Entities

After extracting the entities found on news titles, theyre classified using in-formation on their Wikipedia page. Most entities that people have an opinionabout are likely to be relevant enough to have a page on Wikipedia. This in-cludes celebrities, politicians, scientists, writers, organizations, sports teams, andmany more. These pages have lots of information, and theres bound to be someoverlap between the words on the pages and the words in the affective lexicon.By averaging the valence ratings of these overlapping words, its possible to as-sign a valence rating to entities. This process is essentially keyword spottingfor wikipages, but here its issues with reliance on surface features are miti-gated by Wikipedias community effort on writing clear and descriptive text.This project uses DBPedia[1] to gather information from Wikipedia pages. DB-Pedia is an online knowledge base that extracts structured information fromWikipedia and provides its own formal ontology. Using DBPedia were also notabusing Wikipedias resources.

The classification of entities according to the users preferences relies onInterestMaps[11], a semantic network of interests collected from profile pages onsocial networks. We use a variation on InterestMaps that gathers user informa-tion from their Facebook profile through the websites download backup profilefeature, and then builds a network of interests using DBPedias ontology[12].The users likes and interests are explored in Wikipedia to provide the systeminformation on other entities that may also be of interest.

The inclusion of InterestMaps into sentiment analysis allows more person-alized emotion classifications. Whichever entities are mentioned on the usersprofile page can reinforce or reverse the intensity of the associated emotions. Forinstance, consider a fan of the Rolling Stones (as expressed on his user profile).A news title like Mick Jagger releases new single. would evoke a positive emo-tion with high intensity, because his InterestMap would have information on allthings related to the band, namely Mick Jagger. This applies to all contentthat appears on Facebook and has a page on Wikipedia.

8 Rule-Based Emotion Classification

This project uses the OCC emotion models rule based structure to determinewhat emotions are associated to a news title. There are a total of six emotioncategories in the model (e.g. well-being, attribution, attraction, etc.) and eachis related to a set of variables (e.g. desirability, appealingness, familiarity, etc.).These variables are resolved by analyzing the properties of the subject, verband object phrases. Afterwards, the model is applied and the sentiment analysisprocedure reaches its end.

8.1 Phrase Sentiment Analysis

The linguistic parsing stage built a SVO structure with information on the mostrelevant linguistic relations of the news title. Among these, we have the sub-ject, verb and object in addition to all modifiers, possessive determiners and

10 D. Loureiro, G. Marreiros, and J. Neves

prepositional phrases acting on them. Since every word already has a valencerating assigned to it from the word-level valence assignment and entity pro-cessing stages, the valence ratings for the phrases are calculated following theserules:

If the phrase doesnt contain modifiers, possessive determiners or preposi-tional phrases, then its rating is calculated from the average of all wordratings.

If there are prepositional phrases present, their ratings are scaled down bya factor of 0.8 and added to the average of the other words.

If there are modifiers or possessive determiners present, then the phraseinherits their valence ratings. The impact of prepositional phrases is reduced,because it often strays from the main topic of the sentence.

8.2 Cognitive Model of Emotion

After decoding a news titles linguistic structure and affective contents, we lookat underlying cognitive structure through the OCC emotion model. This rule-based model uses knowledge and intensity variables to characterize 22 differentemotion types. The values for each variable can be assigned using a conditionalset of rules. Shaikh et.al presented a method for variable attribution from a lin-guistic and affective standpoint[24]. This method considered 16 variables fromthe OCC emotion model and described how to assign their values using linguisticand affective data.

The following code demonstrates how we implement Shaikhs et. al interpre-tation of the OCC model.

if Valenced_reaction == True:if Self_reaction == Pleased":if Self_presumption == Desirable":

if Cognitive_strength == Self":Joy = True

if Prospect == Positive":if Status == Unconfirmed":if Object_familiarity == High":Hope = True

9 Evaluation

The results of this projects method for word-level valence assignment were mea-sured using affective words extracted from the General Inquirer as the goldstandard. The General Inquirer contains a positive and a negative word list withabout 3500 words combined. In order to make results comparable, we labelledour numerical ratings above 6 as positive, below 4 as negative and neutral in

Sentiment Analysis of News Titles 11

between. Table 1 shows the results. The variable recall measures the ratio of rec-ognized words, while accuracy and precision measure the binary classificationsof positive and negative valence ratings. As expected, the recall values for theANEW corpus were very low because of its significantly smaller size than theword lists its being measured against.

Table 1. Measure of overlap between our method for word-level valence assignment,and the positive and negative word lists from General Inquirer

Lexicon Accuracy Precision Recall

ANEW 90.44% 87.12% 33.9%

ANEW & WordNet 76.20% 69.45% 91.96%

For measuring the results of sentiment analysis for news titles, the trial testsfrom SemEval 2007 affective text task were used as the gold standard. These tri-als divide emotions into the regular joy, sadness, fear, surprise, angerand disgust categories. Since our method classifies different emotion types, wehad to reduce both ours and SemEvals classifications into positive and negativeemotion groups. From SemEval, the joy category joined the positive group andthe rest (except surprise) joined the negative group. Since the OCC model di-vides each emotion category into positive and negative types, each type joinsthe corresponding group. Table 2 shows the results. We noticed that many ofthe mistakes that occurred were not only involved with the affective aspects, butalso the linguistic analysis often failed to produce a valid SVO structure, hencethe somewhat low recall.

Table 2. Performance results for sentiment analysis of news titles

Comparison Accuracy Precision Recall

SemEval 2004 57.78% 51.92% 60.52%

Heres the output of an example from SemEval as processed by the system.

Headline: Nigeria hostage feared dead is freed."Word-level valence assignment:Nigeria is an entity of type Location: Neutral (5.00)hostage is a Noun included in the lexicon: Negative (2.2)feared is a Verb included in the lexicon: Negative (2.25)dead is a Adjective included in the lexicon: Negative (1.94)is is a Verb included in the lexicon: Neutral (5.00)freed is a Verb disambiguated(free.v.06): Positive (8.14)

(from 43 tokens)

12 D. Loureiro, G. Marreiros, and J. Neves

Phrase analysis:Subject: - Nigeria hostage ( feared dead ) v: negativeVerb: - freed v: positiveObject: - v:

Entities = [Nigeria], no preferenceObject familiarity: Low (0)Emotion_intensity: Neutral (0.018)Emotion Classified: Relief, Joy

Notice that the word/phrase/sentence analysis makes it possible to understandwhich elements contributed to the emotion classification. A purely probablisticclassifier couldnt provide this sort of insight into sentiment analysis.

10 Conclusion

In summary, this paper presented an approach for sentiment analysis that pro-vides an alternative affective lexicon and introduces the element of InterestMapsto consider user preferences. While the latter didnt have the opportunity to betested, the lexicon showed satisfactory results. The emotion classification processreproduced and combined techniques from other works in a unique configura-tion, showing some encouraging results. Our approach relied on the cognitivestructure of sentences and analyzed many linguistic and emotional aspects. Thepipeline nature of this process helps to understand what elements associate head-lines to particular emotions. In the future, well compare our approach to otheraffective resources and explore some applications.

Acknowledgements. This work was funded by a research integration grantfrom Fundacao para a Ciencia e Tecnologia (FCT).

References

[1] Bizer, C., Lehmann, J., Kobilarov, G., Auer, S., Becker, C., Cyganiak, R.,Hellmann, S.: DBpedia A Crystallization Point for the Web of Data. Journalof Web Semantics: Science, Services and Agents on the World Wide Web (7),154165 (2009)

[2] Bradley, M.M., Lang, P.J., Cuthbert, B.N.: Affective Norms for English Wordsin Center for the Study of Emotion and Attention National Institute of MentalHealth, University of Florida (1997)

[3] Chaumartin, F.: UPAR7: A knowledge-based system for headline sentiment tag-ging. In: SemEval 2007, Prague, ACL, pp. 422425 (2007)

[4] Ekman, P.: Facial expression of emotion. American Psychologist 48, 384392(1993)

Sentiment Analysis of News Titles 13

[5] Esuli, A., Sebastiani, F.: SentiWordnet: A Publicly Available Lexical in Resourcefor Opinion Mining. In: LREC 2006 (2006)

[6] Fellbaum, C.: WordNet: An Electronic Lexical Databases. MIT Press, Cambridge(1999)

[7] Finkel, J., Grenager, T., Manning, C.: Incorporating Non-local Information intoInformation Extraction Systems by Gibbs Sampling. In: Proceedings of the 43ndAnnual Meeting of the Association for Computational Linguistics (ACL 2005),pp. 363370 (2005)

[8] Goldberg, R.: The structure of phenotypic personality traits. American Psychol-ogist 48(1), 2634 (1993)

[9] Liu, H., Lieberman, H., Selker, T.: A Model of Textual Affect Sensing using Real-World Knowledge. In: Proceedings of the 2003 International Conference on Intel-ligent User Interfaces, pp. 125132 (2003)

[10] Liu, H., Singh, P.: ConceptNet: A Practical Commonsense Reasoning Toolkit. BTTechnology Journal 22(4), 211226 (2004)

[11] Liu, H., Maes, P.: Interestmap: Harvesting social network profiles for recommen-dations in Beyond Personalization - IUI (2005)

[12] Loureiro, D.: Facebooks hidden feature: User Models and InterestMaps. In: 4thMeeting of Young Researchers, UP, IJUP (2011)

[13] Marneffe, M., Manning, C.: Stanford typed dependencies manual (2008)[14] Marneffe, M., MacCartney, B., Manning, C.D.: Generating Typed Dependency

Parses from Phrase Structure Parses in LREC (2006)[15] Mihalcea, R., Liu, H.: A corpus-based approach to finding happiness. In: The

AAAI Spring Symposium on Computational Approaches to Weblogs (2006)[16] Minsky, M.: The Emotion Machine. Simon & Schuster, New York (2006)[17] Vinagre, E., Marreiros, G., Ramos, C., Figueiredo, L.: An Emotional and Context-

Aware Model for Adapting RSS News to Users and Groups. In: Lopes, L.S., Lau,N., Mariano, P., Rocha, L.M. (eds.) EPIA 2009. LNCS, vol. 5816, pp. 187198.Springer, Heidelberg (2009)

[18] Ortony, A., Clore, G., Collins, A.: The Cognitive Structure of Emotions.Cambridge University Press, New York (1988)

[19] Ortony, A., Clore, L., Foss, A.: The referential structure of the affective lexicon.Cognitive Science 11, 341364 (1987)

[20] Pang, B., Lee, L.: Opinion mining and sentiment analysis. Foundations andTrends. Information Retrieval 1(1-2), 1135 (2008)

[21] Patwardhan, S., Banerjee, S., Pedersen, T.: SenseRelate: TargetWord A general-ized framework for word sense disambiguation. In: Proc. of AAAI 2005 (2005)

[22] Picard, R.: Affective Computing. The MIT Press, Massachusetts (1997)[23] Pinker, S.: The Language Instinct Perennial. HarperCollins (1994)[24] Shaikh, M.A.M., Prendinger, H., Mitsuru, I.: Rules of Emotions: A Linguistic

Interpretation of an Emotion Model for Affect Sensing from Texts. In: Paiva,A.C.R., Prada, R., Picard, R.W. (eds.) ACII 2007. LNCS, vol. 4738, pp. 737738.Springer, Heidelberg (2007)

[25] Shaikh, M., Ishizuka, M., Prendinger, H.: SenseNet: A Linguistic Tool to VisualizeNumerical Valence Based Sentiment of Textual Data. In: Proc. ICON 2007 5thIntl Conf. on Natural Language (2007)

[26] Shaikh, M., Prendinger, H., Ishizuka, M.: Emotion Sensitive News Agent: AnApproach Towards User Centric Emotion Sensing from the News. In: ACM Inter-national Conference on Web Intelligence, pp. 614620 (2007)

14 D. Loureiro, G. Marreiros, and J. Neves

[27] Stone, J., Dunphy, D., Smith, M., Ogilvie, D.: The General Inquirer: A ComputerApproach to Content Analysis. MIT Press, Cambridge (1966)

[28] Strapparava, C., Mihalcea, R.: SemEval-2007 Task 14: Affective Text. In: Pro-ceedings of the 45th Aunual Meeting of ACL (2007)

[29] Strapparava, C., Valitutti, A.: Wordnet-affect: an affective extension of wordnet.In: Proceedings of the 4th Internation Conference on Language Resources andEvaluation, Lisbon, pp. 10831086 (2004)

[30] Wilson, T., Wiebe, J., Hoffmann, P.: Recognizing contextual polarity in phrase-level sentiment analysis. In: Proceedings of the HLT 2005 (2005)

Providing Location Everywhere

Ricardo Anacleto, Lino Figueiredo, Paulo Novais, and Ana Almeida

GECAD - Knowledge Engineering and Decision Support,R. Dr. Antonio Bernardino de Almeida, 431,

4200-072 Porto, Portugal{rmao,lbf,amn}@isep.ipp.pt,

pjon@di.uminho.pt

Abstract. The ability to locate an individual is an essential part ofmany applications, specially the mobile ones. Obtaining this locationin an open environment is relatively simple through GPS (Global Posi-tioning System), but indoors or even in dense environments this type oflocation system doesnt provide a good accuracy. There are already sys-tems that try to suppress these limitations, but most of them need theexistence of a structured environment to work. Since Inertial NavigationSystems (INS) try to suppress the need of a structured environment wepropose an INS based on Micro Electrical Mechanical Systems (MEMS)that is capable of, in real time, compute the position of an individualeverywhere.

Keywords: Location Systems, Dead Reckoning, GPS, MEMS, MachineLearning, Optimization.

1 Introduction

Location information is an important source of context for ubiquitous computingsystems. Using this information, applications can provide richer, more produc-tive and more rewarding user experiences. Also, location-awareness brings manypossibilities that make mobile devices even more effective and convenient, atwork and in leisure.

The ability of mobile applications to locate an individual can be exploitedin order to provide information to help or to assist in decision-making.Someexamples include electronic systems to help people with visual impairments [10],support systems for a tourist guide at an exhibition [5] and navigation systemsfor armies [7] [19].

In an open environment we use GPS (Global Positioning System) to retrieveusers location with good accuracy. However, indoors or in a more dense envi-ronment (big cities with tall buildings, dense forests, etc) GPS doesnt work ordoesnt provide satisfactory accuracy. Consequently, location-aware applicationssometimes dont have access to the user location. The authors would like to acknowledge FCT, FEDER, POCTI, POSI, POCI

and POSC for their support to GECAD unit, and the project PSIS(PTDC/TRA/72152/2006).

L. Antunes and H.S. Pinto (Eds.): EPIA 2011, LNAI 7026, pp. 1528, 2011.c Springer-Verlag Berlin Heidelberg 2011

16 R. Anacleto et al.

The aim of this work is to design a system capable of, in real time, determiningthe position of an individual (preferably in a non-structured environment), whereGPS is not capable.

The motivation for this project comes from previous publications, where arecommendation system to support a tourist when he goes on vacations, waspresented [1] [2]. Mobile devices currently available on the market already havebuilt-in GPS, which provides the necessary location context to recommend placesof interest and to aid the tourist in planning trip visits [17]. However, this as-sistance, with current technology, can only be used in environments with GPSsignal. In closed or dense environments the system cant easily retrieve the userlocation context. For example, in an art gallery if the system knows the touristcurrent position inside the building, it can recommend artworks to view andlearn even more about the tourist tastes.

To remove this limitation, a system that provides precise indoor location ofpeople becomes necessary. Actually, there are already some proposed systemsthat retrieve indoor location with good precision. However most of these solutionsforce the existence of a structured environment. This can be a possible solutionwhen GPS isnt available, but only indoors. In a dense forest we dont have thiskind of systems. To retrieve the location on this type of terrain can be veryuseful for position knowledge of a firemans team. Also, the implementation ofa structured environment using this type of technologies is very expensive andit becomes unfeasible to incorporate this type of systems in all the buildings ofthe world.

Since location without using a structured environment remains an open re-search problem, the main goal of the proposed project is to minimize deploymentand infrastructure costs and provide location everywhere.

As users of these types of devices (and applications) are on foot, the INS isone of the most appropriate solutions to be used. This system consists of MicroElectrical Mechanical Systems (MEMS) devices, which can be accelerometers,gyroscopes and other types of sensors. MEMS are small in size, which allows easyintegration into clothing. Usually, they communicate with a central module usinga wireless network (e.g., Bluetooth). These devices obtain individual movementsinformation independently of the building infrastructure. All this sensory setrequires the implementation of a sensor fusion. This includes algorithms that caninterpret the sensors information and thereby determine the individual position.The collected information, in addition to the motion speed and direction, mustalso be able to determine the step width and the individual position (sitting,lying or standing).

This type of system uses the PDR (Pedestrian Dead Reckoning) technique [4].PDR normally is composed by three key technologies: tracking of the sensorsbehavior, walking locomotion detection and walking velocity estimation. Butunfortunately, large deviations of these sensors can affect performance, as wellas the various forms in which a human can move, so this is the projects biggestchallenge: to correct the sensor deviations.

Providing Location Everywhere 17

A module working only with PDR is not able to ensure that the geographicalpositions are accurate within a few meters. In fact, although these deviationsmay be small for every millisecond, the positioning error caused by a sustainableuse of the system can exceed one meter in 10 seconds [23].

In order to provide a better contextualization of the existing work on thefield, Section 2 gives a brief description of the most relevant INS and indoorlocation systems. In Section 3 we present our main objectives to implement anINS and how we intend to complete them, presenting the main challenges andthe methodology we will use to suppress them. Finally, Section 4 presents someconclusions regarding the presented problem.

2 State of the Art

Indoor localization technologies hold promise for many ambient intelligence ap-plications to facilitate, for example, decision-making. However, many of the ex-istent systems cant work indoors or in too dense environments. This happensbecause they work based on the location obtained by GPS.

Several localization techniques already exist to give indoor positioning, butsome of them doesnt provide good accuracy or are too difficult to implement.From the selected articles we have divided them into two main categories basedon their localization techniques: Radio Frequency Waves and Pedestrian DeadReckoning.

Radio frequency waves are solutions that estimate the location of a mobiletarget in the environment by measuring one or more properties of an electro-magnetic wave radiated by a transmitter and received by a mobile station. Theseproperties typically depend on the distance traveled by the signal and the sur-rounding environment characteristics. In radio frequency we include technologiessuch as IEEE 802.11, Infrared, Bluetooth and Radio-Frequency IDentification(RFID) tags.

As examples of these systems we have: RADAR (Wireless LAN) [3], ActiveBadge (Infrared) [25], a home localization system (Bluetooth) [12], a project forthe location of objects and people using RFID tags [22], among others [11]. Themain problem with these systems is that they need a structured environment inorder to work. This makes them dependent on the particular context and stillimpractical and expensive.

Pedestrian Dead Reckoning systems use sensors to provide location updates,calculated using information about a previously-estimated location. This posi-tion estimation is commonly based on inertial sensors. Since they yield relativepositioning information only, an absolute reference is required to specify thedisplacement reported by an inertial measurement in absolute coordinates.

We have divided the PDR systems into two groups: the pure ones and thehybrid solutions. The first use only inertial sensors to give position (after hav-ing a reference point), and the latter use WLAN, Bluetooth and other type ofstructured indoor location systems to correct the inertial sensors data.

18 R. Anacleto et al.

2.1 Dead Reckoning with Structured Environments

The solutions based on inertial sensing are subject to big measurement deviationscaused by thermal changes in the sensor circuit. This is because the longerthe time indoors, the greater will be the diversion from virtual trajectory tothe real trajectory. In other words, there is error accumulation and to make arecalibration of the sensors, some projects use a structured environment.

This structured environment can be constituted by surveillance cameras likeKourogi [15] proposes. In his paper two enhancements for PDR performanceare introduced: map matching and dynamic estimation of walking parameters.The users location and orientation are updated by fusing the measurementsfrom the PDR estimation and the maps. The surveillance cameras are used tomeasure walking velocity in order to recalibrate the sensor parameters. Also aparticle filter [18] is used for probabilistic data fusing (based on a Bayesian filter).Probability distribution of the users location is predicted from the estimatedposition, orientation and its uncertainties. The results are satisfactory, but theexistence of lots of people in the building will, certainly, confuse the system.Another problem is the cost of this system (over $1500).

Also Wi-Fi signal strength can be used to tackle the traditional drift problemsassociated with inertial tracking [26]. Woodman proposes a framework consti-tuted by a hip-mounted mobile PC which is used to log data obtained from afoot-mounted Inertial Measurement Unit (IMU). The IMU contains three or-thogonal gyroscopes and accelerometers, which report angular velocity and ac-celeration respectively. The logs are then post processed on a desktop machine(so it isnt a real-time location system). Like Kourogis system, this uses Bayesianfilters to probabilistically estimate the state of a dynamic system based on noisymeasurements. The particle filter update consists of three steps: Re-sampling,Propagation and Correction. To reduce the cubic-in-time drift problem of thefoot-mounted IMU, they apply a Zero Velocity Update (ZVU) [27], in which theknown direction of acceleration due to gravity is used to correct tilt errors whichare accumulated during the previous step. Besides Wi-Fi, it also uses a map thatacts like a collection of planar floor polygons. The system obtains the ReceivedSignal Strength Indication (RSSI) information by querying the Wi-Fi hardware.Each query returns a list of visible Wi-Fi access points and corresponding RSSImeasurements. Each floor polygon corresponds to a surface in the building onwhich a pedestrians foot may be grounded. Each edge of a floor polygon is eitheran impassable wall or a connection to the edge of another polygon. Then thesystem tries to make a correspondence between the Wi-Fi signals, the map andthe IMU. A system evaluation was performed in a three floor building, with atotal area of 8725m2, and the error proved to be 0.73m in 95% of the time.

The system proposed by Evennou and Marx [8], it is also based on an IEEE802.11 wireless network. The absolute positional information obtained fromthe wireless network was combined with the relative displacements and rota-tions reported by a gyroscope, a dual-axis accelerometer and a pressure sensor.

Providing Location Everywhere 19

The information obtained from all the sensors was combined through Kalman[21] and particle filtering. To calculate the user displacement, an accelerometerwas used to count the number of steps taken. Then, a constant estimate ofthe users step length was used to calculate the user total displacement withinthe environment. The authors showed that combining the information from thebank of sensors yielded improved localization when compared to the usage of eachsensor separately. The authors conducted an experiment using their multi-sensorlocalization system. The localization accuracies reported during such experimentranged from 1.5 to 3.3 meters.

Renaudin [20] proposes a solution based on RFID tags and inertial MEMS.This system was developed with the intuition to be used by firemens. MEMS andRFID are hybridized in a structure based on an Extended Kalman Filter and ageographical database. While progressing indoors, the fire fighters deploy RFIDtags that are used to correct the large errors affecting MEMS performances.The first team of fire fighters attaches a RFID tag each time it passes a doorand when moving from one floor to another, at the beginning and at the endof the stairway. Upon installation, the geographical coordinates of the tag areassociated with the tag ID. The RFID tag database is a collection of locationcoordinates of all the building doors and stairs. This information is then builtover the evacuation emergency maps. So, the Extended Kalman Filter uses the3D coordinates of each detected RFID tag to relocate the trajectory.

The IMU is composed by three sensor modules. The module attached on theshank has a gyroscope, measuring the angular rate of the shank in the sagittalplane, and an accelerometer oriented in the vertical plane that allows the gaitanalysis. The trunk module contains a triad of gyroscopes, magnetometers andaccelerometers providing the orientation information. The thigh module containsan accelerometer measuring the frontal acceleration, which permits posture anal-ysis. The algorithms that compute the route use all the information available ina database to extract the optimum walking path.

The proposed pedestrian navigation solution has been tested on the campusof the Ecole Polytechnique Federale de Lausanne, in Switzerland. As expected,the PDR position error grows with time, whereas this hybrid positioning solutionremains under a certain limit close to 5 meters.

2.2 Pure Dead Reckoning

Like we have already seen, there are already systems prepared for indoor posi-tioning. But if we want a system that can provide our location anywhere anddoesnt rely in a structured environment, another solution should be used.

One possibility is the use of hybrid solutions based on INS and GPS sig-nal. INS complements the GPS giving location where GPS cant (indoor or indense environment). The INS uses as start location point the GPS last knowncoordinate.

As example of an INS system of this type we have NavMote [9] that integrates,MEMS and GPS, in a wireless-device that is in the persons abdominal area.This device consists of an accelerometer and a magnetic compass integrated in

20 R. Anacleto et al.

a generic wireless controller board, with the radio, a processing unit, and powerstorage all integrated. This device records user movements and stores them inthe system. Sensor data are constantly being stored in a 4MB flash memorywhich allows an operation time of 1.7 hours with a sampling frequency of 30Hz. Subsequently, the compressed data is transferred to the main system whenthe wireless-device comes within range of a sensor network (called NetMote).These sensor networks can be spread along the building. On the main system,data are processed into an estimate of the pedestrian trajectory based on a PDRalgorithm. The main system is also responsible for trajectory displaying, mapmatching and other purposes.

This PDR approach uses the acceleration signal pattern to detect the step oc-currences and the magnetic compass to provide continuous azimuth information.Based on a simplified kinematic model of a persons gait, the walked distance isprovided by summing up the size of each step over the step count. The systemshowed an error of only 3%. However, this solution doesnt provide user loca-tion in real time. For a 3 minute walk, 45 seconds of download time from thewireless-device to the main system and 25 seconds of system filtration time aretypical. So the system is mainly used to track or monitor the users positionwith potential security applications rather than to provide high-level interactivetravel support.

Walder [24] proposes a system to be used on emergency situations, in largebuildings and underground structures. It combines inertial measurements (us-ing 3D accelerometer, gyroscope and 3D magnetometer) of moving persons withbuilding floor plans (CAD drawings) enriched with additional data (e.g., po-sitions of fire extinguishers, room allocation). Because the proposed INS is notaccurate enough, the system tries to improve positioning by a permanent interac-tion between the mobile positioning sensors and floor plans stored in the buildinginformation model. The system is composed by four components: position deter-mination, position verification and correction, user interface and communication.The position determination reads raw sensor data and then computes positioncoordinates. Once computed, a position is verified and corrected, based on thetagged floor plans, if necessary.

The correction of noise and drift is performed by a comparison of real integra-tion with an assumed step length after each step. A normalized step length canbe defined as a system parameter, taking into account the users body size andweight. This step length will be adapted during the movement, depending on themovement pattern, the moving direction (e.g., curve radius) or the recognizedenvironment (e.g., stairs). In the worst case scenario, when the automatic cor-rection doesnt work properly, the user can communicate his current position tothe system. Identifying his position by giving inform