[PDF] Comprehensive Supersense Disambiguation of English

View - Download Comprehensive Supersense Disambiguation of English

Previous PDF

Next PDF

Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Long Papers), pages 185-196

Melbourne, Australia, July 15 - 20, 2018.

c

2018 Association for Computational Linguistics185ComprehensiveSupersenseDisambiguationof

EnglishPr epositionsandPossessives

NathanSchneider

GeorgetownUniversity

JenaD.Hwang

IHMC

VivekSrikumar

UniversityofUtah

JakobPrange

AustinBlodgett

GeorgetownUniversity

SarahR.Moeller

UniversityofColoradoBoulder

AviramStern

AdiBitan

OmriAbend

HebrewUniversityof Jerusalem

Abstract

Semanticrelationsare oftensignaled with

prepositionalorpossessi ve marking - but extremepolysemybedevilstheir analysis andautomatic interpretation.Weintroduce anew annotationscheme,corpus,andtask forthedisambiguation ofprepositionsand possessivesinEnglish.Unlikepre vious approaches,our annotationsarecompre- hensivewithrespecttotypesandtok ens ofthesemark ers;usebroadly applicable supersenseclassesrather thanfine-grained dictionarydefinitions; uniteprepositions andpossessiv esunderthesameclassinven- tory;anddis tinguishbetweena marker's lexicalcontributionandtheroleitmarksin thecontext ofapredicateorscene.Strong interannotatoragreementrates, aswellas encouragingdisambiguationresults with establishedsupervisedmethods, speakto theviabilityof theschemeand task.

1Introduction

Grammar,aspera commonmetaphor, gives speak-

ersofa languageashared toolboxtoconstruct and deconstructmeaningfula ndfluentutterances. Be- inghighlyana lytic,Englishrelies heavilyonword orderandclosed-class functionwords likeprepo- sitions,determiners,and conjunctions.Though functionw ordsbearlittlesemanticcontent,the y arenev erthelesscrucialtothemeaning.Consider prepositions:they serve,forexample,to convey placeandtime (Wemetat/in/outsidetherestaur ant for/afteranhour),toe xpressconfigurationalrela- tionshipslik equantity,possession,part/whole,and membership( thecoats ofdozensofchildreninthe class),andto indicatesemantic rolesinar gument structure( Grandmacookeddinner forthechildr en nathan.schneider@georgetown.edu (1)I wasbook edfor/DURATION2nightsat/LOCUSthis hotelin/TIMEOct2007. (2)I wentto/GOALohmafter/EXPLANATION;TIME readingsomeof/QUANTITY;WHOLEthere views. (3)Itw asvery upsettingtoseethiskindof/SPECIES behaviorespeciallyin_front_of/LOCUS my/SOCIALREL;GESTALTfouryear_old.

Figure1:Annotatedsentencesfrom ourcorpus.

vs.Grandmacookedthec hildrenfordinner).Fre- quentprepositionslik eforaremaddeninglypoly- semous,theirinterpretation dependingespecially ontheobject ofthepreposition - Irode thebus for5dollars/minutes - andthego vernorof the prepositionalphrase (PP):IUbered/ask edfor$5. requiressomeform ofsensedisambiguation, but arrivingatalinguisticrepresentation thatisfle xible enoughtogeneralize acrossusagesand types,yet simpleenoughto supportreliableannotation, has beenadaunting challenge(§2).

Thisw orkrepresentsanewattempt tostrike that

balance.Buildingon priorwork, weargue foran approachtodescribing Englishprepositionand pos- sessivesemanticswithbroadcov erage.Giv enthe semanticov erlapbetweenprepositionsandposses- hood),weanalyze themusingthe samein ventory ofsemanticlabels. 1

Ourcontributions include:

•anew hierarchicalinventory("SNACS") of50supersense classes,extensi velydocu- mentedinguidelines forEnglish(§ 3); annotations:alltypes andtokens ofpreposi- tionsandpossessi vesare disambiguated(§4; examplesentencesappearinfigure 1); •aninterannotatoragreement studythat 1

Someusesof certainother closed-classmarkers -

intransitiveparticles,subordinators,infinitive to - arealso included(§3.1).

186showstheschemeisreliable andgeneralizes

acrossgenres - andfor thefirsttime demon- stratingempiricallythat thelexical semantics ofapreposition cansometimesbe detached fromthePP' ssemanticrole (§5); •disambiguationexperimentswithtwo super- visedclassificationarchitectures toestablish thedifficulty ofthetask(§6).

2Background:Disambiguationof

PrepositionsandPossessiv es

Studiesof prepositionsemanticsin linguistics

andcognitiv esciencehavegenerallyfocusedon thedomainsof spaceandtime (e.g.,Herskovits,

1986;BowermanandChoi,2001;Regier,1996;

Khetarpaletal. ,2009;XuandK emp,2010;Zwarts

andWinter ,2000)oron motivated polysemystruc- turesthatco veradditional meaningsbeyondcore spatialsenses( Brugman,1981;Lakoff,1987;Tyler andEvans ,2003;Lindstromberg,2010).Posses- siveconstructionscanlikewisedenote anumberof semanticrelations,and various factors - including semantics - influencewhetherattrib utiv eposses- sioninEnglish willbee xpressedwithof,orwith 's andpossessiv epronouns(the'genitivealternation';

Heine,2006;Wolketal.,2013;Shihetal. ,2015).

Corpus-basedcomputationalw orkonsemantic

disambiguationspecifically ofprepositionsand possessives 2 fallsintotwocate gories:the lexi-

Dahlmeieretal. ,2009;TratzandHovy,2009;

Hovyetal.,2010,2011;TratzandHovy,2013),

andthesemanticclassapproach( Moldovanetal.,

2004;BadulescuandMoldo van,2009;O'Hara

andWiebe ,2009;SrikumarandRoth,2011,2013;

Schneideretal. ,2015,2016;Hwangetal.,2017,

seealsoMülleret al.,2012forGerman).The lexicographicapproachcancapturefiner -grained meaningdistinctions,at ariskof relyinguponid- iosyncraticandpotentially incompletedictionary definitions.Thesemantic classapproach,which we followhere,focuseson commonalitiesinmeaning acrossmultiplele xicalitems,and aimstogeneral- 2 Ofcourse,meanings marked byprepositions/possessiv es areto someextent capturedinpredicate-ar gumentorgraph- basedmeaningrepresentations (e.g.,Palmeretal.,2005;Fill- moreand Baker,2009;Oepenetal. ,2016;Banarescuetal. ,

2013)anddomain-centric representationslike TimeMLand

ISO-Space(Pustejovskyetal.,2003,2012).

izemoreeasily tone wtypesand usages.

Themostrecent class-basedapproach topreposi-

tionswas ourinitialframework of75preposition supersensesarrangedin amultipleinheritance tax- onomy(Schneideretal. ,2015,2016).Itw asbased largelyonrelation/roleinv entoriesofSrikumar andRoth(2013)andV erbNet(Bonialet al.,2011;

Palmeretal.,2017).Thefr ameworkwasrealizedin

version3.0ofour comprehensively annotatedcor- pus,STREUSLE 3 (Schneideretal. ,2016).How- cleartous over time.

First,aspointed outby Hwangetal.(2017),the

flawedbecauseitin somecasesputs intoconflict thesemanticrole ofthePP withrespectto apred- icate,andthe lexicalsemantics ofthe preposition itself.Hwangetal.(2017)suggesteda solution, discussedin§ 3.3,but didnotconductanannotation studyorrelease acorpusto establishitsfeasibility empirically.Weaddressthat gaphere.

Second,75cate goriesisan unwieldynumber

forbothannotators anddisambiguation systems.

Somearequite specializedande xtremelyrarein

STREUSLE3.0,which causesdatasparseness is-

suesforsupervised learning.Inf act,the onlypub- lisheddisambiguationsystem forprepositionsuper - sensescollapsedthe distinctionstoj ust12labels (GonenandGoldber g,2016).Hwangetal.(2017) remarkedthatsolvingtheaforementioned problem couldremov etheneedformanyofthespecialized categoriesandmakethe taxonomymoretractable forannotatorsand systems.W esubstantiatethis here,defininga newhierarch ywithjust 50cate- gories(SNA CS,§3)and providingdisambiguation resultsforthe fullsetof distinctions.

Finally,giventhe semanticoverlapofposses-

sivecaseandtheprepositionof,wesa wanop- portunitytobroaden theapplication ofthescheme toincludepossessi ves. Ourreannotatedcorpus,

STREUSLE4.0, thushassupersense annotations

forov er1000possessivetokensthat werenotse- manticallyannotatedin version3.0. Weinclude theseinour annotationanddisambiguation experi- mentsalongsidereannotated prepositiontokens.

3AnnotationScheme

3.1LexicalCategories ofInterest

Apartfromcanonical prepositionsandpossessi ves,

thereareman ylexically andsemanticallyoverlap- 3 https://github.com/nert-gu/streusle/

187pingclosed-classitems whicharesometimes clas-

sifiedasother partsofspeech, suchasadv erbs,par- ticles,andsubordinating conjunctions.TheCam- bridgeGrammarof theEnglishLanguage(Huddle- stonandPullum ,2002)argues foranexpansive definitionof 'preposition'thatw ouldencompass theseothercate gories.As apracticalmeasure,we decidedto encourageannotatorsto focusonthe se- manticsofthese functionalitemsrather thantheir syntax,sowe takean inclusive stance.

Anotherconsiderationis developing annotation

guidelinesthatcan beadaptedfor otherlanguages.

Thisincludeslanguages whichhav epostpositions,

circumpositions,orinpositionsratherthan prepo- sitions;thegeneral termforsuch itemsis adpositions. 4

Englishpossessiv emarking(via'sor

possessivepronounslikemy)ismore generallyan exampleofcasemarking.Notethat prepositions (4a-4c)differinwordorder frompossessiv es(4d), thoughsemantically theobjectof thepreposition andthepossessi venominal patterntogether: (4)a.eat inarestaurant b.themaninablueshirt c.thewife oftheambassador d.theambassador 'swife

Cross-linguistically,adpositionsand casemark-

ingareclosely related,andin generalboth gram- maticalstrategies canexpresssimilarkinds ofse- manticrelations. Thismotiv atesacommon seman- ticinv entoryforadpositionsandcase.

Wealsocover multiwordprepositions (e.g.,

out_of,in_front_of),intransiti veparticles(Heflew away),purposeinfiniti ve clauses(Openthedoor toletinsome air 5 ),prepositionswith clausalcom- plements(Itrained beforethepartystarted ),andquotesdbs_dbs12.pdfusesText_18

[PDF] english short stories for beginners

[PDF] english tenses french equivalent

[PDF] english to french translation exercises a level

[PDF] english to french translation exercises with answers

[PDF] english turkish

[PDF] english vocabulary arabic pdf

[PDF] english vocabulary in use advanced pdf

[PDF] english vocabulary in use beginner pdf

[PDF] english vocabulary in use pre intermediate & intermediate pdf

[PDF] english vocabulary in use upper intermediate pdf

[PDF] english vocabulary words with meaning pdf

[PDF] english french vocabulary pdf

[PDF] engraissement bovin charolais

[PDF] engraissement bovin définition

[PDF] engraissement bovin france