278
Table2
PennTreebankPart-Of-Speech(POS)tags.CCCDDTEXFWINJJJJRJJSLSMDNNNNPNNPSNNSPDTPOSPRP
J.M.Chenlo,D.E.Losada/InformationSciences280(2014)275–288
CoordinatingconjunctionCardinalnumberDeterminer
ExistentialthereForeignword
PrepositionorsubordinatingconjunctionAdjective
Adjective,comparativeAdjective,superlativeListitemmarkerModal
Noun,singularormassPropernoun,singularPropernoun,pluralNoun,pluralPredeterminerPossessiveendingPronoun,personalPRP$RBRBRRBSRPSYMTOUHVBVBDVBGVBNVBPVBZWDTWPWP$WRBPronoun,possessiveAdverb
Adverb,comparativeAdverb,superlativeParticleSymbolTo
Interjection
Verb,baseformVerb,pasttense
Verb,gerundorpresentparticipleVerb,pastparticiple
Verb,non-3rdpersonsingularpresentVerb,3rdpersonsingularpresentWh-determinerWh-pronoum
Possessivewh-pronoumWh-adverb
inavectorrepresentation.Thesevectorsandthecorrespondinggroundtruthlabels–subjectivity/polarityclassassignments–feedtheclassi er.Inourexperiments,weconsideredthefollowingsetsoffeatures:
Vocabularyfeatures.Thesearebinaryfeaturesbasedontheoccurrenceofunigramsandbigramsinthesentence.Weonlyrepresentedunigramsandbigramsthatoccuratleastfourtimesinthecorpus.Unigramsandbigramsarevaluabletodetectspeci cdomain-dependent(opinionated)expressions.ThediscriminativepowerofthistypeofcontentfeatureshasbeendemonstratedbyseveralOpinionMiningstudies[31,16].Ourrepresentationisbinarybecauseweworkatsen-tencelevel.Non-binaryweightingschemes[27,28]aremoresuitedtolargerchunksoftext.
Positionalfeatures.Positionalevidencecouldbeagoodguidanceforsubjectivityorpolarityclassi cation.RecentstudiesindicatedthatthepositionofsentencesinadocumentisanimportantcueinSentimentAnalysis[31,29,10].Weincludedfeaturesinourstudythatencodetheabsolutepositionofthesentencewithinthedocument(e.g.,2forthesecondsen-tenceinthedocument),anditsrelativeposition(i.e.,absolutepositionnormalisedbythenumberofsentencesinthedocument).
Part-of-speechfeatures.Thepart-of-speech(POS)ofeachwordcanbevaluableforanalysingsentiments[41].ForeachPOStag–e.g.,JJforadjectives–wede nedonesentencefeature:thecountofoccurrencesofthetaginthesentence.TextwasprocessedbytheStanfordLog-linearPart-Of-SpeechTagger,1whichassignsTreebankPOStags[26](seeTable2). Syntacticpatterns.Inadditiontosentimentwords,manyotherlanguagecompositionsexpressorimplysentimentandopinions[23].Forinstance,Turney[41]de ned vesyntacticpatternstoextractopinionsfromreviews.Thesepatternshavebecomereferencerulesforautomaticallydeterminingopinions[23].Turney’spatternsaresequencesofPOStags(seeTable3)andweencodedtheseasbinaryfeatures(representingtheappearanceofeverypatterninthesentence). Sentimentlexiconfeatures.Thesefeaturesaccountforthenumberofopinionatedtermsthatoccurinthesentence.Sen-timentwordsarerecognizablyadominantfactorinOpinionMining[30,23],wheremanyresearchershavesuccessfullyemployedlexiconstorevealopinions,e.g.,[22,20].Weincludedlexicon-basedfeaturesasfollows.Forsubjectivityclas-si cation,wecomputedthenumberofsubjectivetermsinthesentence.Forpolarityclassi cation,werepresentedthenumberofpositiveandthenumberofnegativetermsinthesentence.Wealsoincludedtheproportionofsubjective,posi-tiveandnegativeterms,andthenumberandproportionofexclamationsandinterrogations.Interrogationsandexclama-tionshavebeensuccessfullyappliedinother ne-grainedOpinionMiningscenarios,suchassentimentdetectionintweets[1].TheopinionatedtermswereobtainedfromtheOpinionFinder(OF)[43]sentimentlexicon.
RSTfeatures.RhetoricalStructureTheory(RST)[25]isoneoftheleadingdiscoursetheories.Thistheoryexplainshowtextscanbesplitintosegmentsthatarerhetoricallyrelatedtooneanother.Withinthisstructure,textsegmentscanbeeithernucleiorsatellites,withnucleibeingassumedtobemoresigni cantthansatelliteswithrespecttounderstand-ingandinterpretingatext.Manytypesofrelationsbetweentextsegmentsexist;themainpaperonRSTde nes23typesofrelations[25].Asatellitemayforinstancebeanelaboration,anexplanationoranevaluationonwhatisexplainedinanucleus.Evidenceonspeci ctypesofsatellitescanserveasaguidancefortheopiniondetectionprocess.Forexample,anattributionrelationshipcouldbeindicativeofsubjectivity.WeusedSPADE(Sentence-levelPArsingofDiscoursE)[36],whichcreatesRSTtreesforindividualsentencesandweincludedbinaryfeaturesassociatedtotheappearanceofeverytypeofRSTrelation(seeTable4)inagivensentence.Everysentencehasonlyoneofthesefeaturessetto1(determinedbythesentence’stop-levelnucleus–satelliterelationship).
1
doc.guandang.net.
OpinionFinder: A system for subjectivity analysis ...2.3.1 Subjective Sentence Classi?cation The ?...features, including those obtained from the ...
4.Study on Features and Translation of English ...Subjectivity in Translation 浅论翻译 中的译者主体性... An Empirical Research Based on CLEC 二语习得...
Essential design features of the BSC Strategy/ ...(2003) Subjectivity and the weighting of ... an empirical study of the effectiveness of the ...
sentence to enhance the are (McNeill, 2000, ...Tendency to subjectivity 3. Lacks firm theoretical...empirical studies need to be carried out 3. ...
An Empirical Study on First-year English Majors’...I-R.Transfer of Sentence Processing Strategies : ...s Subjectivity and the Imagination of New ...
a book on empirical studies in interpreting, ... along literal sentence under the guidance of ...subjectivity of interpretation aspects, for example,...
through nothing more than a single sentence. ...inherent integrity or subjectivity of their ...empirical studies in education that can help ...
Empirical Methods in Natural Language Processing (...topic Table 3: Features for neutral-polar classi...found useful for sentence-level subjectivity classi...
sentence simple-present tense passive vioce non-...loose adherence to style ◆more empirical vs. ...description subjectivity vs. objectivty 8 2. ...
empirical evidence in the creation of pronunciation... sentence or verifications, answersto comprehension ...Thus, such ratingshave some degree of subjectivity...
热门文档