22 1965 International Conference on Computational

221965InternationalConferenceonComputationalLinguisticsSENTENCEGENERATIONBYSEMANTICCONCORDANCEToshiyukiSakaiandMakotoNagaoDepartmentofElectricalEngineering,KyotoUniversityKyoto,Japan~~'~a~l:~..:,~,P.~o-~o'lSakai&NagaoIABSTRACTGenerationofEnglishsentenceisrealizedinthefollowingthreesteps.First,thegenerationofkernelsentencebyphrasestructurerules;second,theapplicationoftransformationalrulestothekernelsentence;andfina-llythecompletionofasentencebythemorphophonemicmodifications.Atthefirststageofgeneratingkernelsentence,thesemanticsofwordsarefullyutilized.Themethodissuchthatapairofwordsinthegenerat-ionprocess(subjectnounandpredicateverb,verbandobjectorcomplement,sdJectiveandmodifiednounetc.)isselectedinaccordancewiththeseman-ticcategorieswhichareattachedtoeachwordintheworddictionary.Thesemanticcategoriesaredeterminedbyconsideringboththemeaningofwordsthemselvesandalsothefunctioningofwordsinsentences.Atthestageoftransformationalrules,sentenceisconsiderednotasasimplestringbutastheonehavingtheinternaltreestructure,andthetransformationalrulesareappliedtothistreestructure.Forthesetwostagesthegenerationprocessisformalizedstrictlyandisrealizedinacomputerprogramming.Wehavepresentedinrelationtothetransformationalrulesamethodofsentencegenerationnotfromtheaxiom(fromth@topGfthetree)butfromanypoint,fromwhichthewholetreeisconstructed.Wehavealsoproposedthatthemorphophonemicrulescanbepresentedasakindofoperatorsoperatingonwordsintheneighbourhoodofageneratedstring.I.INTRODUCTIONAtpresentonbehalfoftheexperienceobtainedfromtheresearchalre-adydone,wehavefairlyexactknowledgeabouttherealdifficultiesintheMTresearchwhicharetobesolvedinthenearfuture.Amongtheseproblemsthemostimportantonesmightbehowtoconstructthesyntaxofalanguageandhowtograspthesemanticsofsentencesofthelanguage.Therehavebeenmanyexcellentcontributionstotheproblemsofsyntax,buttherearestillfewtotheproblemsofsemanticsandtheinterrelation-shipbetweensyntaxandsemantics.WehavetriedaninvestigationinthisareabythemethodofgenerationofEnglishsentence.ThefirstpapereverpublishedconcerningthegenerationofsentencesmighthavebeenthatbyProf.V.H.YngveofMITin1961.Wehaveadoptedthemethodonceagainfromthefollowingpointsofview.(i)itactsasapowerfultesttothestudyofsentencestructure,(ii)tothestudyofsemantics,(iii)tothestudyoftherelationshipbetweensyntaxandsemanticsinnaturallanguage.Generallyspeakingthesentencegenerationmethod,contrarytotheanalysisofagivensentence(whichisguaranteedtobeacorrectone),tendstodemandthesevereconstructionofsyntacticrulesandwordselec-tionrules.Itmayseematthepresentlevelofmechanicaltranslationthatthetreatmentofthesentencepropertyinitsentiretyistoodiffi-culttorealize.ButifwehopetohavethetranslationasperfectasSakai~Nagao2possible,wearenecessarilytoconfrontwiththisproblem.Thequalityofalinguistictheoryofalanguagemaybebestexaminedbythegenerationofsentencesaccordingtothespecifiedlinguistictheory.Especiallytheeffectoftheinterrelationshipbetweensyntaxandsemanticsseemstobeclearlyexemplifiedbythissotospeakcrudetest.Thusthesentencegenerationmethodsurelyrespondstothisoveralltreatmentofthesentenceproperty.Thismaybeconsideredasasteptowardsthegeneraltheoryofnaturallanguage.2.EVALUATIONOFTHEGENERATEDSENTENCE2.1Evaluationofthegeneratedsentence-syntax-Severalmethodsaredevelopedforthedescriptionofsentencestructure.WerepresentthesyntaxofEnglishbyaphrasestructuregrammar,transfor-mationalgrammar,andmophophonemicrules.Thekernelsentenceisgeneratedbythephrasestructuregrammar,thensomepropertransformationalrulesareappliedtoit,andthenthemodificationofthesentencebythemorpho-phonemicrulesproducesthefinaloutput.Theserulesshouldgenerateconceivablesentencestructures,althoughtheactuallyusedsentencestructureshaveseveralconstraints.Theseareforexample,(i)thedepthofthesentencestructure(ii)thecoordinationstructure(iii)theintrinsicunsymmetryofsentencestructure---progressivestruc-ture,topheavystructureetc.Ingeneraltheruleswhicharesuitableforanalysisofagivensentenceseemtodifferfromtheruleswhichgenerategoodsentences.Thedifferencebetweenthesetwoisthedifferencebetweentheactualspokensentencesandtheconceivablesentences.Herewecanseeman'stendencytothelanguagestructure.Thereforeitwillnotbeworthlesstoknowthefrequencyratiosofthephrasesusedintheactualsentences.2.2Evaluationofthegeneratedsentence-semantics-Thenextquestion,andthemoredifficultonethantheformer,isthedeterminationofwhatisthepropermeaningfulsentence.Thetestforthesemanticsentenceanomalyisfarmoredifficultthanthetestforthegrammaticalityofthesentence.Herewecanthinkofthefollowingthreelevelsofcriteriafortherightsentence.(i)Thegrammaticalsentencewhichisspokenorwrittenbytheaverageperson(andthesentencewhichconveysaconcreteconceptwithoutknowingthecircumstancesthesentenceisspoken).Heregrammaticalcoversthephonology,phonemics,morphology,syntaxetc.Thesesen-tencewhicharegrammaticalbutwhicharecontradictoryinmeaningandwhichwedonotspeakaretoberejected.(ii)Thesentenceswhichareincompleteinthewordusages,inflexionsandsoonbutwhichconveyclearunderstandableconcepts.Thesearetheso-calledcorrigibles