Information Extraction meets the Semantic Web

83 Pages·2017·0.99 MB·English

Undefined0(2016)1 1
IOSPress
Information Extraction meets the
Semantic Web: A Survey
Editor(s):AndreasHotho
Solicitedreview(s):DatBaNguyen,SimonScerri,Anonymous
Openreview(s):MichelleCheatham
JoseL.Martinez-Rodrigueza,AidanHoganb andIvanLopez-Arevaloa
aCinvestavTamaulipas,CiudadVictoria,Mexico
E-mail:{lmartinez,ilopez}@tamps.cinvestav.mx
bIMFDChile;DepartmentofComputerScience,UniversityofChile,Chile
E-mail:[email protected]
Abstract.WeprovideacomprehensivesurveyoftheresearchliteraturethatappliesInformationExtractiontechniquesinaSe-
manticWebsetting.Worksintheintersectionofthesetwoareascanbeseenfromtwooverlappingperspectives:usingSeman-
ticWebresources(languages/ontologies/knowledge-bases/tools)toimproveInformationExtraction,and/orusingInformation
Extraction to populate the Semantic Web. In more detail, we focus on the extraction and linking of three elements: entities,
concepts and relations. Extraction involves identifying (textual) mentions referring to such elements in a given unstructured
orsemi-structuredinputsource.Linkinginvolvesassociatingeachsuchmentionwithanappropriatedisambiguatedidentifier
referringtothesameelementinaSemanticWebknowledge-base(orontology),insomecasescreatinganewidentifierwhere
necessary.Withrespecttoentities,worksinvolving(Named)EntityRecognition,EntityDisambiguation,EntityLinking,etc.in
thecontextoftheSemanticWebareconsidered.Withrespecttoconcepts,worksinvolvingTerminologyExtraction,Keyword
Extraction,TopicModeling,TopicLabeling,etc.,inthecontextoftheSemanticWebareconsidered.Finally,withrespectto
relations,worksinvolvingRelationExtractioninthecontextoftheSemanticWebareconsidered.Thefocusofthemajorityof
thesurveyisonworksappliedtounstructuredsources(textinnaturallanguage);however,wealsoprovideanoverviewofworks
thatdevelopcustomtechniquesadaptedforsemi-structuredinputs,namelymarkupdocumentsandwebtables.
Keywords:InformationExtraction,EntityLinking,KeywordExtraction,TopicModeling,RelationExtraction,SemanticWeb
1. Introduction ing–forittobefeasibletoapplymanualannotationto
evenasignificantsubsetofwhatmightbeofrelevance.
The Semantic Web pursues a vision of the Web While the amount of structured data available on
where increased availability of structured content en- the Web has grown significantly in the past years,
ables higher levels of automation. Berners-Lee [20] there is still a significant gap between the coverage
described this goal as being to “enrich human read- of structured and unstructured data available on the
ablewebdatawithmachinereadableannotations,al- Web [248]. Mika referred to this as the semantic
lowingtheWeb’sevolutionasthebiggestdatabasein gap[205],wherebythedemandforstructureddataon
theworld”.However,makingannotationsoninforma- theWeboutstripsitssupply. Forexample,inananal-
tionfromtheWebisanon-trivialtaskforhumanusers, ysis of the 2013 Common Crawl dataset, Meusel et
particularly if some formal agreement is required to al. [201] found that of the 2.2 billion webpages con-
ensure that annotations are consistent across sources. sidered, 26.3% contained some structured metadata.
Likewise,thereissimplytoomuchinformationavail- Thus,despiteinitiativeslikeLinkingOpenData[274],
ableontheWeb–informationthatisconstantlychang- Schema.org [200,204] (promoted by Google, Mi-
0000-0000/16/$00.00©2016–IOSPressandtheauthors.Allrightsreserved
2 J.L.Martinez-Rodriguezetal./InformationExtractionmeetstheSemanticWeb
crosoft,Yahoo,andYandex)andtheOpenGraphPro- languagethatisfoundedononeofthecoreSemantic
tocol[127](promotedbyFacebook),thissemanticgap Webstandards:RDF/RDFS/OWL/SKOS/SPARQL.2
isstillobservableontheWebtoday[205,201]. ByInformationExtractionmethods,wefocusonthe
Asaresult,methodstoautomaticallyextractoren- extractionand/orlinkingofthreemainelementsfrom
hancethestructureofvariouscorporahavebeenacore an(unstructuredorsemi-structured)inputsource.
topic in the context of the Semantic Web. Such pro-
1. Entities: anything with named identity, typically
cessesareoftenbasedonInformationExtractionmeth-
anindividual(e.g.,Barack Obama,1961).
ods,whichinturnarerootedintechniquesfromareas
2. Concepts:aconceptualgroupingofelements.We
suchasNaturalLanguageProcessing,MachineLearn-
considertwotypesofconcepts:
ing and Information Retrieval. The combination of
– Classes: a named set of individuals (e.g.,
techniquesfromtheSemanticWebandfromInforma-
U.S. President(s));
tionExtractioncanbeseenfromtwoperspectives:on
– Topics: categories to which individuals or
the one hand, Information Extraction techniques can
documentsrelate(e.g,U.S. Politics).
beappliedtopopulatetheSemanticWeb,whileonthe
3. Relations:ann-arytupleofentities(n≥2)witha
other hand, Semantic Web techniques can be applied
predicatetermdenotingthetypeofrelation(e.g.,
to guide the Information Extraction process. In some
marry(BarackObama,MicheleObama,Chicago).
cases, both aspects are considered together, where an
existingSemanticWebontologyorknowledge-baseis More formally, we can consider entities as atomic el-
used to guide the extraction, which further populates ements from the domain, concepts as unary predi-
thegivenontologyand/orknowledge-base(KB).1 cates, and relations as n-ary (n ≥ 2) predicates. We
Inthepastyears,wehaveseenawealthofresearch take a rather liberal interpretation of concepts to in-
dedicatedtoInformationExtractioninaSemanticWeb cludebothclassesbasedonset-theoreticsubsumption
setting. While many such papers come from within ofinstances(e.g.,OWLclasses[135]),aswellastop-
the Semantic Web community, many recent works ics that form categories over which broader/narrower
have come from other communities, where, in partic- relations can be defined (e.g., SKOS concepts [206]).
ular, general-knowledge Semantic Web KBs – such Thisisratherapracticaldecisionthatwillallowusto
asDBpedia[170],Freebase[26]andYAGO2[138]– drawtogetheracollectivesummaryofworksinthein-
have been broadly adopted as references for enhanc- terrelated areas of Terminology Extraction, Keyword
ing Information Extraction tasks. Given the wide va- Extraction,TopicModeling,etc.,underoneheading.
riety of works emerging in this particular intersection Returning to “extracting and/or linking”, we con-
fromvariouscommunities(sometimesunderdifferent sidertheextractionprocessasidentifyingmentionsre-
nomenclatures), we see that a comprehensive survey ferring to such entities/concepts/relations in the un-
isneededtodrawtogetherthetechniquesproposedin structuredorsemi-structuredinput,whileweconsider
suchworks.Ourgoalisthentoprovidesuchasurvey. the linking process as associating a disambiguated
identifier in a Semantic Web ontology/KB for a men-
Survey Scope: This survey provides an overview of tion, possibly creating one if not already present and
publishedworksthatdirectlyinvolvebothInformation usingittodisambiguateandlinkfurthermentions.
Extraction methods and Semantic Web technologies.
InformationExtractionTasks: Thesurveydealswith
Giventhatbothareverybroadareas,wemustberather
variousInformationExtractiontasks.Wenowgivean
explicitinourinclusioncriteria.
introductory summary of the main tasks considered
With respect to Semantic Web technologies, to be
(though we note that the survey will delve into each
included in the scope of a survey, a work must make
taskinmuchmoredepthlater):
non-trivialuseofanontology,knowledge-base,toolor
NamedEntityRecognition: demarcate the locations
ofmentionsofentitiesinaninputtext:
1Hereinweadopttheconventionthattheterm“ontology”refers – aka.EntityRecognition,EntityExtraction;
primarily to terminological knowledge, meaning that it describes
classesandpropertiesofthedomain,suchasperson,knows,coun-
try,etc.Ontheotherhand,weusetheterm“KB”torefertoprimar- 2Worksthatsimplymentiongeneraltermssuchas“semantic”or
ily“assertionalknowledge”,whichdescribesspecificentities(aka. “ontology”maybeexcludedbythiscriteriaiftheydonotalsodi-
individuals)ofthedomain,suchasBarackObama,China,etc. rectlyuseordependuponaSemanticWebstandard.
J.L.Martinez-Rodriguezetal./InformationExtractionmeetstheSemanticWeb 3
– e.g., in the sentence “Barack Obama was TopicLabeling: For clusters of words identified as
born in Hawaii”, mark the underlined abstracttopics,extractasingletermorphrasethat
phrasesasentitymentions. bestcharacterizesthetopic;
EntityLinking: associate mentions of entities with – aka. Topic Identification, esp. when linked
anappropriatedisambiguatedKBidentifier: with an ontology/KB identifier; often used
– involves,orissometimessynonymouswith, forthepurposesofTextClassification;
Entity Disambiguation;3 often used for the – e.g., identify that the topic { “cancer”,
purposesofSemanticAnnotation; “breast”,“doctor”,“chemotherapy”}is
– e.g., associate “Hawaii” with the DBpedia best characterized with the term “cancer”
identifier dbr:Hawaii for the U.S. state
(potentially linked to dbr:Cancer for the
(rather than the identifier for various songs
diseaseandnot,e.g.,theastrologicalsign).
orbooksbythesamename).4
RelationExtraction: Extract potentially n-ary rela-
TerminologyExtraction: extract the main phrases
tions (for n≥2) from an unstructured (i.e., text)
that denote concepts relevant to a given domain
orsemi-structured(e.g.,HTMLtable)source;
described by a corpus, sometimes inducing hier-
– a goal of the area of Open Information Ex-
archicalrelationsbetweenconcepts;
traction;
– aka.TermExtraction,oftenusedforthepur-
– e.g., in the sentence “Barack Obama was
posesofOntologyLearning;
born in Hawaii”, extract the binary rela-
– e.g., identify from a text on Oncology that
“breast cancer”and“melanoma”areim- tionwasBornIn(Barack Obama,Hawaii);
– binaryrelationsmayberepresentedasRDF
portantconceptsinthedomain;
– optionally identify that both of the above triples after linking entities and linking the
conceptsarespecializationsof“cancer”; predicate to an appropriate property (e.g.,
– termsmaybelinkedtoaKB/ontology. mapping wasBornIn to the DBpedia prop-
KeyphraseExtraction: extractthemainphrasesthat ertydbo:birthPlace);
categorize the subject/domain of a text (unlike – n-ary(n≥3)relationsareoftenrepresented
TerminologyExtraction,thefocusisoftenonde- withavariantofreification[133,271].
scribingthedocument,notthedomain);
Notethatwewilluseamoresimplifiednomenclature
– aka. Keyword Extraction, which is often {Entity,Concept,Relation} × {Extraction,Linking}
generically applied to cover extraction of
aspreviouslydescribedtostructureoursurveywiththe
multi-word phrases; often usedfor the pur-
goal of grouping related works together; thus, works
posesofSemanticAnnotation;
on Terminology Extraction, Keyphrase Extraction,
– e.g., identify that the keyphrases “breast
Topic Modeling and Topic Labeling will be grouped
cancer”and“mammogram”helptosumma-
undertheheadingofConceptExtractionandLinking.
rizethesubjectofaparticulardocument;
Again we are only interested in such tasks in the
– keyphrasesmaybelinkedtoaKB/ontology.
contextoftheSemanticWeb.Ourfocusisonunstruc-
TopicModeling: Cluster words/phrases frequently
tured (text) inputs, but we will also give an overview
co-occurring together in the same context; these
of methods for semi-structured inputs (markup docu-
clusters are then interpreted as being associated
mentsandtables)towardstheendofthesurvey.
toabstracttopicstowhichatextrelates;
– aka.TopicExtraction,TopicClassification; RelatedAreas,SurveysandNovelty: Thereareava-
– e.g., identify that words such as “cancer”,
rietyofareasthatrelateandoverlapwiththescopeof
“breast”,“doctor”,“chemotherapy”tend
thissurvey,andlikewisetherehavebeenanumberof
to co-occur frequently and thus conclude
previoussurveysintheseareas.Wenowdiscusssuch
that a document containing many such oc-
areasandsurveys,howtheyrelatetothecurrentcon-
currencesisaboutaparticularabstracttopic.
tribution,andoutlinethenoveltyofthecurrentsurvey.
As we will see throughout this survey, Information
3InsomecasesEntityLinkingisconsideredtoincludebothrecog- Extraction (IE) from unstructured sources – i.e., tex-
nitionanddisambiguation;inothercases,itisconsideredsynony- tualcorporaexpressedprimarilyinnaturallanguage–
mouswithdisambiguationappliedafterrecognition.
reliesheavilyonNaturalLanguageProcessing(NLP).
4Weusewell-knownIRIprefixesasconsistentwiththelookup
A number of resources have been published within
servicehostedat:http://prefix.cc.AllURLsinthispaperwere
lastaccessedon2018/05/30. the intersection of NLP and the Semantic Web (SW),
4 J.L.Martinez-Rodriguezetal./InformationExtractionmeetstheSemanticWeb
wherewecanpoint,forexample,toarecentbookpub- Ontology-BasedInformationExtraction: refers to
lishedbyMaynardetal.[190]in2016,whichlikewise leveraging the formal knowledge of ontologies
covers topics relating to IE. However, while IE tools toguideatraditionalInformationExtractionpro-
mayoftendependonNLPprocessingtechniques,this cess over unstructured corpora. Such works fall
isnotalwaysthecase,wheremanymodernapproaches withinthescopeofthissurvey.Apriorsurveyof
totaskssuchasEntityLinkingdonotuseatraditional Ontology-BasedInformationExtractionwaspub-
NLP processing pipeline. Furthermore, unlike the in- lishedbyWimalasuriyaandDou[313]in2010.
troductorytextbookbyMaynardetal.[190],ourgoal OntologyLearning: helpsautomatethe(costly)pro-
here is to provide a comprehensive survey of the re- cessofontologybuildingbyinducingan(initial)
search works in the area. Note that we also provide a ontology from a domain-specific corpus. Ontol-
briefprimeronthemostimportantNLPtechniquesin ogyLearningalsooftenincludesOntologyPopu-
asupplementaryappendix,discussedlater. lation,meaningthatinstanceofconceptsandre-
On the other hand, Data Mining involves extract- lationsarealsoextracted.Suchworksfallwithin
ingpatternsinherentinadataset.ExampleDataMin- our scope. A survey of Ontology Learning was
ing tasks include classification, clustering, rule min- providedbyWongetal.[316]in2012.
ing, predictive analysis, outlier detection, recommen- KnowledgeExtraction: aims to lift an unstructured
dation, etc. Knowledge Discovery refers to a higher- or semi-structured corpus into an output de-
level process to help users extract knowledge from scribed using a knowledge representation for-
rawdata,whereatypicalpipelineinvolvesselectionof malism (such as OWL). Thus Knowledge Ex-
data,pre-processingandtransformationofdata,aData traction can be seen as Information Extraction
Miningphasetoextractpatterns,andfinallyevaluation but with a stronger focus on using knowledge
andvisualizationtoaidusersgainknowledgefromthe representation techniques to model outputs. In
raw data and provide feedback. Some IE techniques 2013, Gangemi [110] provided an introduction
may rely on extracting patterns from data, which can andcomparisonoffourteentoolsforKnowledge
beseenasaDataMiningstep5;however,Information Extractionoverunstructuredcorpora.
Extraction need not use Data Mining techniques, and Other related terms such as “Semantic Informa-
many Data Mining tasks – such as outlier detection tion Extraction” [108], “Knowledge-Based Informa-
– have only a tenuous relation to Information Extrac- tion Extraction” [139], “Knowledge-Graph Comple-
tion.AsurveyofapproachesthatcombineDataMin- tion” [178], and so forth, have also appeared in the
ing/KnowledgeDiscoverywiththeSemanticWebwas literature. However, many such titles are used specif-
publishedbyRistoskiandPaulheim[261]in2016. ically within a given community, whereas works in
With respect to our survey, both Natural Language theintersectionofIEandSWhaveappearedinmany
Processing and Data Mining form part of the back- communities. For example, “Knowledge Extraction”
groundofourscope,butasdiscussed,InformationEx- isusedpredominantlybytheSWcommunityandnot
tractionhasaratherdifferentfocustobothareas,nei- others.6 Hence our survey can be seen as drawing to-
thercoveringnorbeingcoveredbyeither. getherworksinsuch(sub-)areasunderamoregeneral
Ontheotherhand,relatingmorespecificallytothe scope:worksinvolvingIEtechniquesinaSWsetting.
intersectionofInformationExtractionandtheSeman-
Intended Audience: This survey is written for re-
ticWeb,wecanidentifythefollowing(sub-)areas:
searchers and practitioners who are already quite fa-
SemanticAnnotation: aims to annotate documents
miliarwiththemainSWstandardsandconcepts–such
with entities, classes, topics or facts, typically
astheRDF,RDFS,OWLandSPARQLstandards,etc.
based on an existing ontology/KB. Some works
– but are not necessarily familiar with IE techniques.
on Semantic Annotation fall within the scope of
Hence we will not introduce SW concepts (such as
oursurveyastheyincludeextractionandlinking
RDF, OWL, etc.) herein. Otherwise, our goal is to
of entities and/or concepts (though not typically
make the survey as accessible as possible. For exam-
relations). A survey focused on Semantic Anno- ple,inordertomakethesurveyself-contained,inAp-
tationwaspublishedbyUrenetal.[300]in2006.
6Herewemean“KnowledgeExtraction”inanIE-relatedcontext.
5Infact,thetitle“InformationExtraction”pre-datesthatofthe Otherworksongeneratingexplanationsfromneuralnetworksuse
title“DataMining”initsmoderninterpretation. thesameterminanunrelatedmanner.
J.L.Martinez-Rodriguezetal./InformationExtractionmeetstheSemanticWeb 5
pendixAweprovideadetailedprimeronsometradi- Scholarforrelatedpapers,merginganddeduplicating
tionalNLPandIEprocesses;thetechniquesdiscussed lists of candidate papers (numbering in the thousands
inthisappendixare,ingeneral,notinthescopeofthe in total); (II) we initially apply a rough filter for rele-
survey,sincetheydonotinvolveSWresources,butare vance based on the title and type of publication; (III)
heavily used by works that fall in scope. We recom- wefilterforrelevancebyabstract;and(IV)finallywe
mend readers unfamiliar with the IE area to read the filterforrelevancebythebodyofthepaper.
appendix as a primer prior to proceeding to the main To collect further literature, while reading relevant
bodyofthesurvey.KnowledgeofsomecoreInforma- papers,wealsotakenoteofotherworksreferencedin
tion Retrieval concepts – such as TF–IDF, PageRank, relatedworks,worksthatcitemoreprominentrelevant
cosinesimilarity,etc.–andsomecoreMachineLearn- papers, and also check the bibliography of prominent
ingconcepts–suchaslogisticregression,SVM,neural authorsintheareaforotherpapersthattheyhavewrit-
networks,etc.–maybenecessarytounderstandfiner ten;suchworkswereaddedinphaseIIItobelaterfil-
details,butnottounderstandthemainconcepts. teredinphaseIV.Table2presentsthenumbersofpa-
persconsideredbyeachphaseofthemethodology.7
Nomenclature: TheareaofInformationExtractionis
associatedwithadiversenomenclaturethatmayvary Table1
in use and connotation from author to author. Such Keywordsusedtosearchforcandidatepapers
variationsmayattimesbesubtleandatothertimesbe E/C/RlistkeywordsrelatingtoEntities,ConceptsandRelations;
entirely incompatible. Part of this relates to the vari- SWlistskeywordsrelatingtotheSemanticWeb
ousareasinwhichInformationExtractionhasbeenap-
Type Keywordset
pliedandthevarietyofareasfromwhichitdrawsin-
fluence. We will attempt to use generalized terminol- E "coreference resolution", "entity disambiguation",
"entity linking", "entity recognition",
ogyandindicatewhenterminologyvaries. "entity resolution", "named entity",
"semantic annotation"
SurveyMethodology: Basedonthepreviousdiscus-
C "concept models", "glossary extraction",
sion,thissurveyincludespapersthat:
"group detection", "keyphrase assignment",
"keyphrase extraction", "keyphrase recognition",
– deal with extraction and/or linking of entities,
"keyword assignment", "keyword extraction",
conceptsand/orrelations, "keyword recognition", "latent variable models",
"LDA" "LSA", "pLSA", "term extraction",
– dealwithsomeSemanticWebstandard–namely
"term recognition", "terminology mining",
RDF,RDFSorOWL–oraresourcepublishedor "topic extraction", "topic identification",
"topic modeling"
otherwiseusingthosestandards,
– have details published, in English, in a relevant R "OpenIE", "open information extraction",
"open knowledge extraction", "relation detection",
workshop,conferenceorjournalsince1999,
"relation extraction", "semantic relation"
– considerextractionfromunstructuredsources.
SW "linked data", "ontology", "OWL", "RDF", "RDFS",
For finding in-scope papers, our methodology be- "semantic web", "SPARQL", "web of data"
gins with a definition of keyphrases appropriate to
thesectionathand.Thesekeyphrasesaredividedinto
Weprovidefurtherdetailsofoursurveyonline,in-
listsofIE-relatedterms(e.g.,“entity extraction”, cludingthelistsofpapersconsideredbyeachphase.8
“entity linking”) and SW-related terms (e.g.,
We may include out-of-scope papers to the extent
“ontology”,“linked data”),whereweapplyacon-
that they serve as important background for the in-
junctionoftheirproductstocreatesearchphrases(e.g.,
scopepapers:forexample,itisimportantforanunini-
“entity extraction ontology”).Giventhediverse
tiated reader to understand some of the core tech-
terminology used in different communities, often we
niques considered in the traditional Information Ex-
need to try many variants of keyphrases to capture
tractionareaandtounderstandsomeofthecorestan-
as many papers as possible. Table 1 lists the base
dards and resources considered in the core Semantic
keyphrasesusedtosearchforpapers;thefinalkeyword
Web area. Furthermore, though not part of the main
searchesaregivenbytheset(E∪C∪R)(cid:107)SW,where
“(cid:107)”denotesconcatenation(withadelimitingspace).
7Table2referstopapersconsideringtextasinput;afurther20
Our survey methodology consists of four initial
papersconsideringsemi-structuredinputsarepresentedlaterinthe
phasestosearch,extractandfilterpapers.Foreachde-
survey,whichwillbringthetotalto109selectedpapers.
fined keyphrase, we (I) perform a search on Google 8http://www.tamps.cinvestav.mx/~lmartinez/survey/
6 J.L.Martinez-Rodriguezetal./InformationExtractionmeetstheSemanticWeb
Table2 ontheotherhand,thereferenceKBmaycontainenti-
Numberofpapersincludedinthesurvey(byPhase)
tiesfromhundredsoftypes.Hence,whilesomeEntity
E/C/Rdenotecountsofhighlightedpapersinthissurveyrelatingto Extraction & Linking tools rely on off-the-shelf NER
Entities,Concepts,andRelations,resp.;
tools, others define bespoke methods for identifying
ΣdenotesthesumofE+C+Rbyphase.
entitymentionsintext,typicallyusingentities’labels
Phase E C R Σ intheKBasadictionarytoguidetheextraction.
Seedcollection(I) 2,418 8,666 8,008 19,092 Once entity mentions are extracted from the text,
Filterbytitle(II) 114 167 148 429 thenextphaseinvolveslinking–ordisambiguating–
Filterbyabstract(III) 100 115 102 317 these mentions by assigning them to KB identifiers;
typicallyeachmentioninthetextisassociatedwitha
Finallist(IV) 25 36 28 89
singleKBidentifierchosenbytheprocessasthemost
likelymatch,orisassociatedwithmultipleKBidenti-
survey, in Section 5, we provide a brief overview of
fiersandanassociatedweight(aka.support)indicating
otherwiserelatedpapersthatconsidersemi-structured
confidenceinthematchesthatallowtheapplicationto
inputsources,suchasmarkupdocuments,tables,etc.
choosewhichentitylinkstotrust.
Survey Structure: The structure of the remainder of
Example: In Listing 1, we provide an excerpt of
thissurveyisasfollows:
an EEL response given by the online DBpedia Spot-
Section2 discusses extraction and linking of entities light demo10 in JSON format. Within the result, the
forunstructuredsources. “@URI” attribute is the selected identifier obtained
Section3 discussesextractionandlinkingofconcepts from DBpedia, the “@support” is a degree of confi-
forunstructuredsources. denceinthematch,the“@types”listmatchesclasses
Section4 discussesextractionandlinkingofrelations from the KB, the “@surfaceForm” represents the
forunstructuredsources. text of the entity mention, the “@offset” indicates
Section5 discusses techniques adapted specifically
the character position of the mention in the text,
for extracting entities, concepts and/or relations the “@similarityScore” indicates the strength of
fromsemi-structuredsources.
a match with the entity label in the KB, and the
Section6 concludesthesurveywithadiscussion. “@percentageOfSecondRank” indicates the ratio of
Additionally,AppendixAprovidesaprimeronclas- thesupportcomputedforthefirst-andsecond-ranked
sical Information Extraction techniques for readers documentsthusindicatingthelevelofambiguity.
previouslyunfamiliarwiththeIEarea;werecommend Ofcourse,theexactdetailsoftheoutputofanEEL
suchareadertoreviewthismaterialbeforecontinuing. processwillvaryfromtooltotool,butsuchatoolwill
minimallyreturnaKBidentifierandthelocationofthe
entitymention;asupportwillalsooftenbereturned.
2. EntityExtraction&Linking
Applications: EEL is used in a variety of applica-
EntityExtraction&Linking(EEL)9referstoidenti- tions,suchassemanticannotation[41],whereentities
fyingmentionsofentitiesinatextandlinkingthemto mentioned in text can be further detailed with refer-
areferenceKBprovidedasinput. encedatafromtheKB;semanticsearch[295],where
EntityExtractioncanbeperformedusinganoff-the- search over textual collections can be enhanced – for
shelfNamedEntityRecognition(NER)toolasusedin example,todisambiguateentitiesortofindcategories
traditional IE scenarios (see Appendix A.1); however of relevant entities – through the structure provided
suchtoolstypicallyextractentitiesforlimitednumbers bytheKB;questionanswering[299],wheretheinput
of types, such as persons, organizations, places, etc.; text is a user question and the EEL process can iden-
tifywhichentitiesintheKBthequestionrefersto;fo-
cused archival [79], where the goal is to collect and
9Wenotethatnamingconventionscanvarywidely:sometimes
preserve documents relating to particular entities; de-
Named Entity Linking (NEL) is used; sometimes the acronym
(N)ERDisusedfor(Named)EntityRecognition&Disambiguation; tectingemergingentities[136],whereentitiesthatdo
sometimesEELisusedasasynonymforNED;otherphrasescan not yet appear in the KB, but may be candidates for
alsobeused,suchasNamedEntityExtraction(NEE),orNamedEn-
tityResolution,orvariationsontheideaofsemanticannotationor
semantictagging(whichweconsiderapplicationsofEEL). 10http://dbpedia-spotlight.github.io/demo/
J.L.Martinez-Rodriguezetal./InformationExtractionmeetstheSemanticWeb 7
addingtotheKB,areextracted.11 EELcanalsoserve
Listing1:DBpediaSpotlightEELexample asthebasisforlaterIEprocesses,suchasTopicMod-
Input: Bryan Cranston is an American actor. He is eling,RelationExtraction,etc.,asdiscussedlater.
(cid:44)→ known for portraying "Walter White" in the
(cid:44)→ drama series Breaking Bad. Process: As stated by various authors [62,153,243,
Output: 246], the EEL process is typically composed of two
{ main steps: recognition, where relevant entity men-
"@text": "Bryan Cranston is an American actor. He
(cid:44)→ is known for portraying \"Walter White\" tionsinthetextarefound;anddisambiguation,where
(cid:44)→ in the drama series Breaking Bad.", entity mentions are mapped to candidate identifiers
"@confidence": "0.35",
"@support": "0", withafinalweightedconfidence.Sincethesestepsare
"@types": "", (often) loosely coupled, this section surveys the vari-
"@sparql": "",
"@policy": "whitelist", ous techniques proposed for the recognition task and
"Resources": [ thereafterdiscussesdisambiguation.
{ "@URI": "http://dbpedia.org/resource/
(cid:44)→ Bryan_Cranston",
"@support": "199", 2.1. Recognition
"@types": "DBpedia:Agent,Schema:Person,Http://
(cid:44)→ xmlns.com/foaf/0.1/Person,DBpedia:Person
(cid:44)→ ", The goal of EEL is to extract and link entity men-
"@surfaceForm": "Bryan Cranston",
"@offset": "0", tions in a text with entity identifiers in a KB; some
"@similarityScore": "1.0",
tools may additionally detect and propose identifiers
"@percentageOfSecondRank": "0.0" },
{ "@URI": "http://dbpedia.org/resource/ for emerging entities that are not yet found in the
(cid:44)→ United_States", KB [238,247,231]. In both cases, the first step is to
"@support": "560750",
"@types": "Schema:Place,DBpedia:Place,DBpedia: markentitymentionsinthetextthatcanbelinked(or
(cid:44)→ PopulatedPlace,Schema:Country,DBpedia:
proposed as an addition) to the KB. Thus traditional
(cid:44)→ Country",
"@surfaceForm": "American", NERtools–discussedinAppendixA.1–canbeused.
"@offset": "21",
However, in the context of EEL where a target KB is
"@similarityScore": "0.9940788480408",
"@percentageOfSecondRank": "0.003612999020" }, givenasinput,therecanbekeydifferencesbetweena
{ "@URI": "http://dbpedia.org/resource/Actor",
typicalEELrecognitionphaseandtraditionalNER:
"@support": "35596",
"@types": "",
"@surfaceForm": "actor",
– Incaseswhereemergingentitiesarenotdetected,
"@offset": "30",
"@similarityScore": "0.9999710345342", the KB can provide a full list of target entity la-
"@percentageOfSecondRank": "2.433621943E−5" }, bels, which can be stored in a dictionary that is
{ "@URI": "http://dbpedia.org/resource/
(cid:44)→ Walter_White_(Breaking_Bad)", usedtofindmentionsofthoseentities.Whiledic-
"@support": "856",
tionaries can be found in traditional NER sce-
"@types": "DBpedia:Agent,Schema:Person,Http://
(cid:44)→ xmlns.com/foaf/0.1/Person,DBpedia:Person narios,theseoftenrefertoindividualtokensthat
(cid:44)→ ,DBpedia:FictionalCharacter", stronglyindicateanentityofagiventype,suchas
"@surfaceForm": "Walter White",
"@offset": "66", commonfirstorfamilynames,listsofplacesand
"@similarityScore": "0.9999999999753",
companies, etc. On the other hand, in EEL sce-
"@percentageOfSecondRank": "2.471061675E−11" },
{ "@URI": "http://dbpedia.org/resource/Drama", narios,thedictionarycanbepopulatedwithcom-
"@support": "6217",
pleteentitylabelsfromtheKBforawiderrange
"@types": "",
"@surfaceForm": "drama", oftypes;inscenariosnotinvolvingemergingen-
"@offset": "87",
tities,thisdictionarywillbecompletefortheen-
"@similarityScore": "0.8446404328140",
"@percentageOfSecondRank": "0.1565036704039" }, tities to recognize. Of course, this can lead to a
{ "@URI": "http://dbpedia.org/resource/
(cid:44)→ Breaking_Bad", verylargedictionary,dependingontheKBused.
"@support": "638", – Relating to the previous point, (particularly) in
"@types": "Schema:CreativeWork,DBpedia:Work,
(cid:44)→ DBpedia:TelevisionShow", scenarios where a complete dictionary is avail-
"@surfaceForm": "Breaking Bad", able,thelinebetweenextractionandlinkingcan
"@offset": "100",
"@similarityScore": "1.0", become blurred since labels in the dictionary
"@percentageOfSecondRank": "4.6189529850E−23" } from the KB will often be associated with KB
]
}
11Emerging entities are also sometimes known as Out-Of
Knowledge-Base(OOKB)entitiesorNotInLexicon(NIL)entities.
8 J.L.Martinez-Rodriguezetal./InformationExtractionmeetstheSemanticWeb
identifiers; hence, dictionary-based detection of Selectionofentities: InthecontextofEEL,anobvi-
entities will also provide initial links to the KB. oussourcefromwhichtoformthedictionaryisthela-
Such approaches are sometimes known as End- belsoftargetentitiesintheKB.InmanyInformation
to-End(E2E)approaches[247],whereextraction Extractionscenarios,KBspertainingtogeneralknowl-
andlinkingphasesbecomemoretightlycoupled. edgeareemployed;themostcommonlyusedare:
– In traditional NER scenarios, extracted entity
DBpedia[170] A KB extracted from Wikipedia and
mentions are typically associated with a type,
used by ADEL [247], DBpedia Spotlight [199],
usuallywithrespecttoanumberoftrainedtypes
ExPoSe [238], Kan-Dis [144], NERSO [124],
suchasperson,organization,andlocation.How-
Seznam[94],SDA[42]andTHD[82],aswellas
ever, in many EEL scenarios, the types are al-
works by Exner and Nugues [95], Nebhi [226],
readygivenbytheKBandareinfactoftenmuch
Gianninietal.[116],amongstothers;
richerthanwhattraditionalNERmodelssupport.
Freebase[26] A collaboratively-edited KB – previ-
ously hosted by Google but now discontinued in
In this section, we thus begin by discussing the favor of Wikidata [293] – used by JERL [183],
preparation of a dictionary and methods used for rec- Kan-Dis [144], NEMO [68], Neofonie [157],
ognizingentitiesinthecontextofEEL. NereL[280],Seznam[94],Tulip[180],aswellas
worksbyZhengetal.[330],amongstothers;
2.1.1. Dictionary
Wikidata[309] A collaboratively-edited KB hosted
The predominant method for performing EEL re-
by the Wikimedia Foundation that, although re-
liesonusingadictionary–alsoknownasalexiconor
leased more recently than other KBs, has been
gazetteer –whichmapslabelsoftargetentitiesinthe
usedbyHERD[283];
KBtotheiridentifiers;forexample,adictionarymight
YAGO(2)[138] AnotherKBextractedfromWikipedia
mapthelabel“Bryan Cranston”totheDBpediaIRI with richer meta-data, used by AIDA [139],
dbr:Bryan_Cranston. In fact, a single KB entity AIDA-Light[230],CohELL[121],J-NERD[231],
mayhavemultiplelabels(aka.aliases)thatmaptoone KORE [137] and LINDEN [277], as well as
identifier, such as “Bryan Cranston”, “Bryan Lee worksbyAbedinietal.[1],amongstothers.
Cranston”,“Bryan L. Cranston”,etc.Furthermore,
some labels may be ambiguous, where a single label These KBs are tightly coupled with owl:sameAs
maymaptoasetofidentifiers;forexample,“Boston” links establishing KB-level coreference and are also
maymaptodbr:Boston,dbr:Boston_(band),and tightlycoupledwithWikipedia;thisimpliesthatonce
entities are linked to one such KB, they can be tran-
so forth. Hence a dictionary may map KB labels to
sitively linked to the other KBs mentioned, and vice
identifiers in a many-to-many fashion. Finally, for
versa.KBsthataretightlycoupledwithWikipediain
each KB identifier, a dictionary may contain contex-
thismannerarepopularchoicesforEELsincetheyde-
tual features to help disambiguate entities in a later
scribe a comprehensive set of entities that cover nu-
stage;forexample,contextinformationmaytellusthat
merous domains of general interest; furthermore, the
dbr:Bostonistypedasdbo:CityintheKB,orthat
text of Wikipedia articles on such entities can form a
known mentions of dbr:Boston in a text often have
usefulsourceofcontextualinformation.
wordslike“population”or“metropolitan”nearby.
On the other hand, many of the entities in these
Thus, with respect to dictionaries, the first impor-
general-interestKBsmaybeirrelevantforcertainap-
tant aspect is the selection of entities to consider (or,
plicationscenarios.Somesystemssupportselectinga
indeed, the source from which to extract a selection
subset of entities from the KB to form the dictionary,
ofentities).Thesecondimportantaspect–particularly
potentially pertaining to a given domain or a selec-
given large dictionaries and/or large corpora of text – tion of types. For example, DBpedia Spotlight [199]
istheuseofoptimizedindexesthatallowforefficient can build a dictionary from the DBpedia entities re-
matchingofmentionswithdictionarylabels.Thethird turned as results for a given SPARQL query. Such a
aspect to consider is the enrichment of each entity in pre-selectionofrelevantentitiescanhelpreduceambi-
thedictionarywithcontextualinformationtoimprove guityandtailorEELforagivenapplication.
the precision of matches. We now discuss these three Conversely, in EEL scenarios targeting niche do-
aspectsofdictionariesinturn. mains not covered by Wikipedia and its related KBs,
J.L.Martinez-Rodriguezetal./InformationExtractionmeetstheSemanticWeb 9
customKBsmayberequired.Forexample,forthepur- dictionary preloaded into the index, the text can then
poses of supporting multilingual EEL, Babelfy [215] bestreamedthroughtheindextofind(prefix)matches.
constructsitsownKBfromaunificationofWikipedia, Phrasesaretypicallyindexedseparatelytoallowboth
WordNet, and BabelNet. In the context of Microsoft word-levelandphrase-levelmatching.Thisalgorithm
Research, JERL [183] uses a proprietary KB (Mi- isimplementedbyGATE[69]andLingPipe[38],with
crosoft’sSatori)alongsideFreebase.Otherapproaches thelatterbeingusedbyDBpediaSpotlight[199].
makeminimalassumptionsabouttheKBused,where The main drawback of tries is that, for the match-
earlier EEL approaches such as SemTag [80] and ing process to be performed efficiently, the dictio-
KIM[250]onlyassumethatKBentitiesareassociated nary index must fit in memory, which may be pro-
with labels (in experiments, SemTag [80] uses Stan- hibitive for very large dictionaries. For these reasons,
ford’s TAP KB, while KIM [250] uses a custom KB theLucene/SolrTaggerimplementsamoregeneralfi-
calledKIMO). nitestatetransducerthatalsoreusessuffixesandbyte-
encodings to reduce space [71]; this index is used by
Dictionarymatchingandindexing: Inordertomatch
HERD[283]andTulip[180]tostoreKBlabels.
mentions with the dictionary in an efficient manner –
Inothercases,ratherthanusingtraditionalInforma-
with O(1) or O(log(n)) lookup performance – opti-
tion Extraction frameworks, some authors have pro-
mized data structures are required, which depend on
posedtoimplementcustomindexingmethods.Togive
the form of matching employed. The need for effi-
some examples, KIM [250] uses a hash-based index
ciency is particularly important for some of the KBs
overtokensinanentitymention12;AIDA-Light[230]
previouslymentioned,wherethenumberoftargeten-
usesaLocalitySensitiveHashing(LSH)indextofind
titiesinvolvedcangointothemillions.Thesizeofthe
approximate matches in cases where an initial exact-
inputcorporaisalsoanimportantconsideration:while
matchlookupfails;andsoforth.
slower(butpotentiallymoreaccurate)matchingalgo-
Of course, the problem of indexing the dictionary
rithms can be tolerated for smaller inputs, such algo-
is closely related to the problem of inverted indexing
rithmsareimpracticalforlargerinputtexts.
inInformationRetrieval,wherekeywordsareindexed
Amajorchallengeisthatdesirablematchesmaynot
againstthedocumentsthatcontainthem.Suchinverted
be an exact match, but may rather only be captured
indexeshaveproventheirscalabilityandefficiencyin
by an approximate string-matching algorithm. While
Web search engines such as Google, Bing, etc., and
one could consider, for example, approximate match-
likewisesupportsimpleformsofapproximatematch-
ingbasedonregularexpressionsoreditdistances,such
ingbasedon,forexample,stemmingorlemmatization,
measures do not lend themselves naturally to index-
which pre-normalize document and query keywords.
based approaches. Instead, for large dictionaries, or
Exploiting this natural link to Information Retrieval,
large input corpora, it may be necessary to trade re-
the ADEL [247], AGDISTIS [301], Kan-Dis [144],
call(i.e.,thepercentageofcorrectspotscaptured)for
TagMe [99] and WAT [243] systems use inverted-
efficiency by using coarser matching methods. Like-
indexingschemessuchasLucene13andElastic14.
wise, it is important to note that KBs such as DBpe-
To manage the structured data associated with en-
dia enumerate multiple “alias” labels for entities (ex-
tities, such as identifiers or contextual features, some
tracted from the redirect entries in Wikipedia), which
tools use more relational-style data management sys-
ifincludedinthedictionary,canhelptoimproverecall
whileusingcoarsermatchingmethods. tems.Forexample,AIDA[139]usesthePostgreSQL
A popular approach to index the dictionary is to relationaldatabasetoretrieveentitycandidates,while
usesomevariationonaprefixtree(aka.trie),suchas ADEL[247]andNeofonie[157]usetheCouchbase15
usedbytheAho–Corasickstring-searchingalgorithm, andRedis16NoSQLstores,respectively,tomanagethe
which can find mentions of an input list of strings labelsandmeta-dataoftheirdictionaries.
withinaninputtextintimelineartothecombinedsize
of the inputs and output. The main idea is to repre- 12ThisimplementationwaslaterintegratedintoGATE:https:
sent the dictionary as a prefix tree where nodes refer //gate.ac.uk/sale/tao/splitch13.html
to letters, and transitions refer to sequences of letters 13http://lucene.apache.org/core/
14https://www.elastic.co;notethatElasticSearchisinfact
in a dictionary word; further transitions are put from
basedonLucene
failed matches (dead-ends) to the node representing 15http://www.couchbase.com
thelongestmatchingprefixinthedictionary.Withthe 16https://redis.io/
10 J.L.Martinez-Rodriguezetal./InformationExtractionmeetstheSemanticWeb
Contextualfeatures: Ratherthanbeingaflatmapof thatmaynotbeavailableinothertextualcorpora,
entity labels to (sets of) KB identifiers, dictionaries including disambiguation pages, redirections of
often include contextual features to later help disam- aliases, category information, info-boxes, article
biguatecandidatelinks.Suchcontextualfeaturesmay edithistory,andsoforth.17
becategorizedasbeingstructuredorunstructured.
We will further discuss how contextual features –
Structuredcontextualfeaturesarethosethatcanbe
stored as part of the dictionary – can be used for dis-
extracteddirectlyfromastructuredorsemi-structured
ambiguationlaterinthissection.
source.InthecontextofEEL,suchfeaturesareoften
2.1.2. Spotting
extracted from the reference KB itself. For example,
Wenowassumeadictionarythatmapslabels(e.g.,
eachentityinthedictionarycanbeassociatedwiththe
“Bryan Cranston”, “Bryan Lee Cranston”, etc.)
(labelsofthe)typesofthatentity,butalsoperhapsthe
to a (set of) KB identifier(s) for the entity ques-
labelsofthepropertiesthataredefinedforit,oracount
tion (e.g„ “dbr:Bryan_Cranston”) and potentially
ofthenumberoftriplesitisassociatedwith,ortheen-
some contextual information (e.g., often co-occurs
titiesitisrelatedto,oritscentrality(andthus“impor-
with“dbr:Breaking_Bad”,anchortextoftenusesthe
tance”)inthegraph-structureoftheKB,andsoforth.
term “Heisenberg”, etc.). In the next step, we iden-
Ontheotherhand,unstructuredcontextualfeatures
tify entity mentions in the input text. We refer to this
are those that must instead be extracted from textual
processasspotting,wherewesurveykeyapproaches.
corpora. In most cases, this will involving extracting
statistics and patterns from an external reference cor- Token-based: Given that entity mentions may con-
pusthatpotentiallyhasalreadyhaditsentitieslabeled sistofmultiplesequentialwords–aka.n-grams–the
(and linked with the KB). Such features may capture brute-forceoptionwouldbetosendalln-gramsinthe
patternsintextsurroundingthementionsofanentity, inputtexttothedictionary,fornupto,say,themaxi-
entities that are frequently mentioned close together, mumnumberofwordsfoundinadictionaryentry,ora
patternsintheanchor-textoflinkstoapageaboutthat fixedparameter.Werefergenericallytothesen-grams
entity, in how many documents a particular entity is astokensandtothesemethodsforextractingn-grams
mentioned, how many times it tends to be mentioned astokenization.Sometimesthesemethodsarereferred
inaparticulardocument,andsoforth;clearlysuchin- toaswindow-basedspottingorrecognitiontechniques.
formationwillnotbeavailablefromtheKBitself. A number of systems use such a form of tokeniza-
A very common choice of text corpora for extract- tion. SemTag uses the TAP ontology for seeking en-
ing both structured and unstructured contextual fea- titymentionsthatmatchtokensfromtheinputtext.In
tures is Wikipedia, whose use in this setting was – to AIDA-Light [230], AGDISTIS [301], Lupedia [203],
thebestofourknowledge–firstproposedbyBunescu and NERSO [124], recognition uses sliding windows
and Pasca [33], then later followed by many other overthetextforvarying-lengthn-grams.
subsequent works [67,99,42,255,39,40,243,246]. The Although relatively straightforward, a fundamental
widespread use of Wikipedia can be explained by the weaknesswithtoken-basedmethodsrelatestoperfor-
uniqueadvantagesithasforsuchtasks: mance:givenalargetext,thedictionary-lookupimple-
mentation will have to be very efficient to deal with
– The text in Wikipedia is primarily factual and thenumberoftokensatypicalsuchprocesswillgen-
availableinavarietyoflanguages. erate, many of which will be irrelevant. While some
– Wikipedia has broad coverage, with documents basicfeatures,suchascapitalization,canalsobetaken
aboutentitiesinavarietyofdomains. intoaccounttofilter(some)tokens,still,notallmen-
– ArticlesinWikipediacanbedirectlylinkedtothe tions may have capitalization, and many irrelevant or
entities they describe in various KBs, including incoherent entities can still be retrieved; for example,
DBpedia,Freebase,Wikidata,YAGO(2),etc. by decomposing the text “New York City”, the sec-
– MentionsofentitiesinWikipediaoftenprovidea ond bi-gram may produce York City in England as a
linktothearticleaboutthatentity,thusproviding candidate,though(probably)irrelevanttothemention.
labeled examples of entity mentions and associ- Suchentitiesareknownasoverlappingentities,where
atedexamplesofanchortextinvariouscontexts. post-processingmustbeapplied(discussedlater).
– Aside from the usual textual features such as
term frequencies and co-occurrences, a variety 17Informationfrominfo-boxes,disambiguation,redirectsandcat-
of richer features are available from Wikipedia egoriesarealsorepresentedinastructuredformatinDBpedia.