Emotion Mining Promotionsvortrag English

download Emotion Mining Promotionsvortrag English

of 5

Transcript of Emotion Mining Promotionsvortrag English

  • 7/29/2019 Emotion Mining Promotionsvortrag English

    1/5

    1MultimediaConcepts

    andApplications

    AlexanderOsherenko

    OpinionMiningand

    LexicalAffectSensing

    PromotionsvortragAlexanderOsherenko

    Betreuer:Prof.Dr.ElisabethAndre,Prof.Dr.Dr.WolfgangMinker

    30.06.2010

    2MultimediaConcepts

    andApplications

    AlexanderOsherenko

    Outline

    Introduction

    Challenges Researchquestions

    Previous approaches

    Studiedapproaches Statistical

    Semantic

    Hybrid

    Viafusion

    Summary Contributions

    Outlook

    30.06.2010

    3MultimediaConceptsandApplications

    AlexanderOsherenko

    OpinionMining

    Movie Review (long text) www.reelviews.net

    Grammatically correct text Definitely expressed opinion,but emotionally different

    words

    30.06.2010 4MultimediaConceptsandApplications

    AlexanderOsherenko

    AffectRecognition

    Naturallanguage utterances (short text)

    Notalways grammatically correct text

    Repetitions,repairs,fill words,incorrect wordings

    Textis important,but not everything

    - We have, Prudence.

    - I m okay.

    - Er m, wel l , i t s beenr easonabl e day so f ar.Er m, bi t bor i ng, but ,er, hopef ul l y t he day

    wi l l pi ck up.

    30.06.2010

    5MultimediaConcepts

    andApplications

    AlexanderOsherenko

    Challenges

    Bigvariability inexpression ofemotions Speaker andautorspecific

    Situationspecific

    Genrespecific:movie reviews,chats,emals etc.

    Emotions are expressed not always clearly Irony

    UnterdrckteEmotions

    MixedEmotions

    Corporaare difficult toobtain Many texts andtalks dont contain emotions that are interesting for us

    It is not always easy toEsistnichtimmereinfach,eineGrundwahrheitzufinden.

    530.06.2010 6MultimediaConceptsandApplications

    AlexanderOsherenko

    Challenges(Software)

    Recognition

    Simulation

    Modelling

    According to taxonomyof applications usingemotional awareness (Batliner et al., 2006):

    30.06.2010

  • 7/29/2019 Emotion Mining Promotionsvortrag English

    2/5

    7MultimediaConcepts

    andApplications

    AlexanderOsherenko

    Challenges(Applications)

    Opinion Mining

    Sort documents not according tothe topic,but ratheraccording tothe opinion

    Emotionrecognition incall centers Forchoosing the appropriate dialogue strategy Should the caller speak with ahumanoperator?

    Emotionrecognition inacar Entertainmentsoftware considers the emotionalstate of

    the driver andherdriving style

    730.06.2010 8MultimediaConceptsandApplications

    AlexanderOsherenko

    Emotionmodels

    Discrete categories

    Forinstance,Ekmancategories(1999):Wut,Abscheu,Furcht,Freude,Trauer,berraschung

    Continous emotions

    Representation through thedimensions (for instance,Erregungundvalence,orEvaluationandactivation)

    joy

    Higherarousal

    Positive valence

    surprise

    sadness

    Lowerarousal

    disgust

    fear

    Negative valence

    anger

    affection

    bored

    30.06.2010

    9MultimediaConceptsandApplications

    AlexanderOsherenko

    Emotionsinthethesis

    9

    negative positive

    Mapping of continiousemotions onto discretecategories

    30.06.2010 10MultimediaConceptsandApplications

    AlexanderOsherenko

    Existingapproaches Informationclassification

    Statistical approach:

    Movie reviews:[Pang etal.,2002],[Pang,B.,Lee,L.2004]

    Product reviews:[Daveetal.,2003]

    Weblogs:[Riloff etal.,2006]

    Articles from newspapers:[Diederichetal.,2000]

    Conversation abstracts:[MairesseF.etal.,2007]

    Semantic approach:

    Sentences from weblogs:[Neviarouskaya etal.,2007]

    Acoustic approaches:

    Berlindatabase,Danish corpus,SmartKom corpus:[Vogtetal.,2008]

    Lexical,stylometric,acoustic features

    Emotionwords,negations, intensifers

    Nosystematic combination ofinformation

    Nostudy ofmultipletextgenres

    1030.06.2010

    11MultimediaConcepts

    andApplications

    AlexanderOsherenko

    Researchquestionsaccordingto

    emotionrecognitionfromspeech1. What linguistic features should be extracted for

    automatic opinion mining andhow toevaluatethem?

    2. Datadriven or knowledgebased emotionrecognition?

    3. How could other modalities,for instance,acousticinformation contribute toimprovement ofrecognition rates?

    1130.06.2010 12MultimediaConceptsandApplications

    AlexanderOsherenko

    Studiedcorpora

    12

    215movie reviews04stars inincrement 0.5stars

    Movie reviewsBMRC

    759sentencesPositive,negative,unclassifiable

    Englishsentences

    Fifty WordFiction

    (FWF)

    Genre Emotionclasses Data amount

    Pang MovieReview

    Movie reviews Positive,negative 2000moviereviews

    SensitiveArtificial

    Listener (SAL)

    Naturallanguagedialogues

    Positiveaktive,negativeaktive,positivpassive,negativepassive,neutral

    574uerungen

    CwPR Productreviews

    1 5stars 300productreviews

    BMRCS Englishsentences

    Positiveaktive,negativeaktive,positivpassive,negativepassive,neutral

    1010sentences

    30.06.2010

  • 7/29/2019 Emotion Mining Promotionsvortrag English

    3/5

    13MultimediaConcepts

    andApplications

    AlexanderOsherenko

    Mainideaofthethesis Noexplicit rules for mapping texts onto emotions

    Statistical Approach Extract relevantfeatures from texts andtrain classifiers

    Emotionrecognition difficult without meaningconsideration Semantic Approach

    Search for emotionalpatterns inrelevantparts ofsentences andmap them onto emotions

    Combination ofthe semantic andthe statisticapproaches HybridApproach

    Classification improvement through consideration ofadditionalmodalities Fusion

    1330.06.2010 14MultimediaConceptsandApplications

    AlexanderOsherenko

    Statisticalapproach

    Learning phase

    14

    Testing phase

    Feature extraktion/Feature evalutation

    (Preprocessing)Learningdata

    Classifiertraining

    OpinionTestingdata

    Classification

    30.06.2010

    15MultimediaConceptsandApplications

    AlexanderOsherenko

    StatisticalApproach(Dissertation) Corpora(2,5,5and9classes)

    Features Lexikalical features:

    (Lemmatized)words inthe frequency list,Whissell,BNC

    Stylometricfeatures: Featuressuchasstatndard deviation ofword lengths,ofsentence lengths,

    digramsetc.

    Deictic features:

    Timeandlocation references,pronouns,stopwordsetc.

    Grammatical features: Interjections,repetitions etc.

    Klassifizierung(SVM)1530.06.2010 16MultimediaConcepts

    andApplications

    AlexanderOsherenko

    SALresults

    Bestresults:words,buttheirnumberisverybig Wordfeaturesarenotknownforeverycorpusincontrasttoother

    featuregroups16

    31.35%Grammatical features

    59.65%Deictic features

    58.97%Stylometrical features

    59.6%Lemmatized word lists

    60.21%Non-lemmatized word lists

    SALCorpus/Features

    30.06.2010

    17MultimediaConcepts

    andApplications

    AlexanderOsherenko

    SemanticApproach

    Recognition oftypical patterns inemotionalutterances

    Interjections:Oh!It is disgusting!

    Repetitions:It is very very expensive!

    Intensifiers:It is very unplesant!

    Negations:Nomovie is sogood asthis one!vs. Itis not agoodmovie.

    1730.06.2010 18MultimediaConceptsandApplications

    AlexanderOsherenko

    SemanticApproach(Dissertation)

    18

    Syntactic Analysis- Stanford Parser -

    Semantic Analysis- SPIN Parser-

    I am nothappy.

    Output of Stanford Parser: (ROOT (S (NP (PRP I)) (VP(VBP am) (RB not) (ADJ P (J J happy))) (. .)))

    Output of SPIN parser: Negation(not) EmotionalWord(happy) EmotionalPhrase(semCat: low_neg)

    30.06.2010

  • 7/29/2019 Emotion Mining Promotionsvortrag English

    4/5

    19MultimediaConcepts

    andApplications

    AlexanderOsherenko

    FWFresults

    Statisticalapproach:37.20%19

    44.22Average

    45.21Last phrase

    44.79First phrasePhrases

    46.04Average

    47.24Last phrase

    47.20First phraseSubsentences

    42.79Average

    47.45Last phrase

    45.41First phraseWhole text

    45.92Average

    47.64Last phrase

    47.20First phraseMajority

    RStrategyGranularity

    30.06.2010 20MultimediaConceptsandApplications

    AlexanderOsherenko

    HybridApproach

    Longtexts

    20

    Shorttexts

    Result:better than statistic approach but worse than semantic

    approach

    Statistical analysisSemantic analysisSentences Opinion

    Statistic analysisSemanticanalysisSentence Emotion

    Semanticanalysis

    Statistic analysisSentence Emotion

    Result doublechoice by chance

    30.06.2010

    21MultimediaConceptsandApplications

    AlexanderOsherenko

    Fusion

    Featurefusion:combines features from differentmodalities

    21

    Classifier

    Classifier

    Acoustic features

    Linguistic features

    Choice

    Deicision fusion:makes choice ofdecisions ofmultipleclassifiers

    Classifier

    Acoustic features

    Linguistic features

    30.06.2010 22MultimediaConceptsandApplications

    AlexanderOsherenko

    1. Corpus(additionally acoustic information)

    2. FeatureandDecision fusion

    3. Visualization astree

    Fusionis beneficial especially if nolanguage context isconsidered.

    Fusion(Dissertation)

    30.06.2010

    23MultimediaConcepts

    andApplications

    AlexanderOsherenko

    Contributions

    Comprehensive analysis ofapproaches toopinion miningandlexical affect sensing using differentcorporarealization inanew software

    Extraction andevaluation offeatures toopinion mining andlexical affect sensing

    Differentiated semantic approach

    Implementation ofintroduced approaches inEmoText

    Hybridapproach

    Multimodalfusion

    2330.06.2010 24MultimediaConceptsandApplications

    AlexanderOsherenko

    Outlook

    1. Newmodalities

    2. Application development

    3. Combinated emotion andpersonality modeling

    BigFive

    30.06.2010

  • 7/29/2019 Emotion Mining Promotionsvortrag English

    5/5

    25MultimediaConcepts

    andApplications

    AlexanderOsherenko

    Dissertationdefence

    Thank you!

    30.06.2010