atVariables
ASTAncestors
Method,Class,SuperclassandInterfaceSubtokensDeclarationModi ersVariableTypeAll
MethodDeclarationsASTAncestors
CyclomaticComplexityFieldsSubtokens
Class,SuperclassandInterfacesSubtokensMethodImplementationSubtokensDeclarationModi ersReturnTypeSiblingMethods
NumberofArgumentsThrownExceptionsAll
TypeDeclarationsFieldSubtokens
SuperclassandInterfaceSubtokensContainedMethodsSubtokensAll
@5%@20%@5%@20%-0.3-1.2-0.1-0.21.23.64.09.111.77.91.310.08.87.76.417.0-----1.00.1-0.1-0.34.81.30.2-0.17.84.20.521.74.90.420.85.1----0.90.00.70.6-0.81.93.7-24.87.03.4-1.64.0-0.44.3-8.85.7-5.15.3-1.513.6
2.61.01.93.85.8-2.32.3-7.63.14.70.93.22.36.1-2.17.53.18.78.510.3
@5%@20%@5%@20%2.01.12.22.12.14.110.99.813.47.46.57.010.56.45.519.3----0.71.80.71.55.01.50.50.68.84.60.0217.55.20.716.85.3----0.80.10.20.3-0.91.91.8-23.07.30.9-1.83.6-1.23.2-11.45.35.214.2-4.42.6
2.50.81.43.35.7-0.91.4-6.92.83.80.62.61.24.9-1.16.50.42.52.33.6
Figure6:Evaluationofsinglepointsuggestionsforvariablesatrankk=1averagedacrossallprojects.The“features”and“no-features”modelslacksuf cientcon dencetomakesuggestionsatthehighersuggestionfrequencies.
slightlyasthepredictioncon dencedecreases.Interestingly,thetoken-levelmodelsareunabletomakeanysuggestionsbeyondasuggestionfrequencyof15%.Forallothertokens,thetoken-levelmethodsreturnthespecialUNKtoken,indicatingthatthemodelsexpectaneologismwhichtheycannotpredict.Incontrast,thesubtokenmodelssustainagoodF1score,evenforlargesugges-tionfrequencies.Thisisduetothefactthatthesubtokenmodelslearnnamingconventionsatthesubtokenlevel,capturinglinguis-ticpatterns[5]suchasthatspeci cfunctionsmaycontainvarioussubtokense.g.get,set,has,is.
Table4showsafulllistoftheeffectthateachfeaturehasontheperformanceoftheneuralmodelsatrankk=5.Asexpected,thereturntype,thesubtokensoftheclasswherethemethodisdeclaredinandthesubtokensofthevariablesandmethodinvocationsinsidethatmethodprovidethemostsubstantialperformanceincreases.Basedontheseresults,weconcludethatweareabletosuggestaccuratemethodnamesandthatoursuggestionsarebetterthanpreviousapproaches.WethereforeanswerRQ3intheaf rmative.ClassDe nitionsAccuracyIntheprevioussection,theperfor-manceoftheneuralmodelonsuggestingnamesformethoddecla-rationswasshown.Inthissection,weevaluatetheneuralmodel
whenmakingsuggestionsforclassde nitions.Figure7bshowstheperformanceofthen-gramlanguagemodelandtheneuralmodelsforclassnamede nitions.Incontrasttothepreviousmodels,thetoken-levelmodelscannotmakeanysuggestions,alwayssuggest-ingtheUNKtoken.However,thesubtokenmodelisabletomakesuggestionsevenathighsuggestionfrequenciesmaintaininganF1scoreofmorethan40%outperformingthen-grammodel.
Thankstotheabilityofthesubtokenmodeltosuggestneologismsthesubtoken-levelmodelisabletosuggestclassde nitionnamesthatithasneverseenbefore,withagoodF1score.Table4showsthatthesubtokensofthesuperclassandinterfacesthatatypeisimplementingareinformativeaboutthenameoftheclass.Addition-ally,whencombiningalltheavailablefeatures,wegetasigni cantincreaseinF1score.Thus,weansweryestoRQ4aswell;weareabletosuggestaccuratetype(class)names.
6.RELATEDWORK
NamingInSoftwareEngineeringNamingincodehasachievedafairamountofresearchattention.Therehasbeenpriorresearchintoidentifyingpoorlynamedartifacts.HøstandØstvold[26]de-velopedatechniqueforautomaticallyinferringnamingrulesformethodsbasedonthereturntype,control ow,http://www.77cn.com.cningtheserulestheyfoundandreported“namingbugs”byidentifyingmethodswhosenamescontainedruleviolations.Ar-naoudovaetal.presentedacatalogof“linguisticanti-patterns”incodethatleadtodevelopersmisunderstandingcodeandbuiltadetectorofsuchanti-patterns[5].Binkleyusedpartofspeechtag-gingto nd eldnamesthatviolateacceptedpatterns,e.g.the eldcreate_mp4beginswithaverbandimpliesanactionwhichisacommonpatternforamethodratherthana eld[11].Ourworkiscomplementary,aswemakesuggestionsfornameswhennamingbugsarefound,anti-patternsoccur,ornamingrulesareviolated.DeLucioetal.attemptedtoautomaticallynamesourcecodeartifactsusingLSIandLDAandfoundthatthisapproachdoesn’tworkaswellassimplermethodssuchasusingwordsfromclassand