Prototypical Emotion Developments in Literary Genres EvgenyKimevgeny.kim@ims.uni-stuttgart.deUniversitätStuttgart,GermanySebastianPadósebastian.pado@ims.uni-stuttgart.deUniversitätStuttgart,GermanyRomanKlingerroman.klinger@ims.uni-stuttgart.deUniversitätStuttgart,Germany

Introduction Storytelling is a central form of human artisticexpression. An ingredient of the appeal of stories istheir emotional content. Readers of literature formexplicitmentalrepresentationsoffictionalcharacters’emotional states (Gernsbacher, Goldsmith, andRobertson,1992;Vega,1996).Evenmore,agripping“[…] literary work produces a complex emotionalexperience in the reader. This experience isinseparable from the depictive content of thenarrative”(Hogan,2011).Thisraisesthequestionoftherelationshipbetweenthenarrativeandemotionallevels in literature. We explore how computationalemotion analysis can contribute to thecharacterizationofstorygenres,whicharedifficulttodefine and for which various criteria have beenproposed,includingstylisticones(BiberandConrad,2009)andnarratologicalones(Chatman,1978). Ourhypothesisisthatgenrescanbelinkedtothedevelopment of predominant emotions over thecourse of the text. To test this, we present acomputationalmodelofmulti-labelemotionanalysisof literary genres and apply it to a set of Englishliterary works from the Project Gutenberg for fivegenres,namelyadventure,romance,mystery,sciencefiction,andhumorousfiction.Weidentifyprototypicalshapes for each genre and show that this analysisproduces results, which can find a place in thecomputationalanalysisof literarygenresandextendexistingstylometricapproaches.

Cuddon(2012)definesadventureas“aformoffiction[...] in which the hero conventionally undergoes aseriesoftestingandepisodicadventures”andmysteryasanarrativeinvolvingthe“detectionofcrime,withthe motives, actions, arraignment, judgement andpunishment of a criminal”. Baldick (2015) definedRomanceasnarrativeswith"improbableadventuresof idealized characters”. Today, however, the termcoversmany forms of fiction, including love stories.Weuse the term romance as a literary genre in thisbroadersense.Regardingscience fictionstories, it isgenerallyagreedthattheyare“[...]aboutanamazingvariety of things, topics, ideas. They include trips toother worlds, quests, the exploration of space...”(Cuddon, 2012). Humorous fiction is comicalliterature “written chiefly to amuse its audience"(Cuddon,2012).

Methods We calculate emotion scores for eight basicemotions, namely joy, sadness, trust, disgust, fear,anger,surprise,andanticipation(Plutchik,2001).Weuse the NRC Emotion Lexicon (Mohammad andTurney,2013).SincethedatainProjectGutenbergisdiachronic,thismethodofemotionrecognitionmightnotbeappropriateforoldertextsand,ingeneral,maysufferfromlowrecall.However,itcanbeconsideredahigh-precisionapproachsuitableforourpurpose.Toobtainanemotionanalysisforastory,westartbycomputingemotionscores forspansof text(Klinger,Suliya, andReiter, 2016). Formally,we compute thescorees(e, S) foranemotioneanda spanof tokensS=<tn,…,tm>as

whereDeisadictionarycontainingwordsassociatedwithemotioneand1t∈Dis1ift∈Dand0otherwise.Wedothisforeachofoureightemotions,obtaininganeightdimensional“emotionvector”foreachspan.Weanalyzed the stability of our results across differentsettingsandfoundthatdifferentdictionariesaffecttheactual values but not the relation between differenttime steps. These scores are not probabilities, butcouldbetransformedifneeded. Toobservedevelopmentover time,wecoulduseslidingwindows;however,continuoustimeseriesarenotoriouslydifficulttointerpret.Therefore,weadopta simpler scheme inspired by the five-act theory ofdramatic acts (Freytag, 1863), according to whichdramas are divided into five acts: exposition, risingaction, climax, falling action, and denouement. We

consequently divide each text into five successive,equal-sizedspans(sincedifferenttextshavedifferentlength, thesizeofactsvariesbetweentexts)thatweassume to correspond roughly to dramatic acts inFreytag’s theory, with exposition at position 1 anddenouement at position 5, and compute an eight-elementemotionvectorforeachAct.

Experimental Evaluation We now demonstrate how this emotionaggregation into five acts can contribute to theunderstandingofdifferentliterarygenres.

Data Wecollect2113booksfromProjectGutenbergthatbelong to five genres found in the Brown corpus(Francis and Kucera, 1979), namely adventure (585books), romance (383 books),mystery (380 books),sciencefiction(562books),andhumorousfiction(203books).Thecorpusisavailablefromtheauthorsuponrequest. Theselection isbasedon theLibraryofCongressSubjectHeadingsinthemetadata.Weselectallbooksthat contain theword “Fiction” in combinationwithoneofthefollowinglabels:“Adventurestories”,“Lovestories”, “Romantic fiction”, “Detective and mysterystories”, “Science fiction” or “Humor”. Furthermore,we reject books with the following labels: “Shortstories”, “Complete works”, “Volume”, “Chapter”,“Collection”, “Part”. This constraint is aimed atexcluding files which contain partial or multipleworks.

Qualitative analysis Each plot in Figure 1 depicts the act-by-actdevelopment for one emotion with their emotionscore es(e,S). Sincewe interpret shapes rather thanvalues,weomitthelegend.Theaverageoverallbooksisshowninadark-coloredline.Theareaaroundthatline corresponds to confidence intervals at a 95%confidencelevel. Forsadness,anger,fear,anddisgust,allfivegenresshow a close correspondence, namely a consistentincrease of the emotion fromAct 1 through Act 5 –corresponding to gripping narratives. Mystery andsciencefictionlackthedropinangerandtendtoendwith higher levels of this emotion. Joy is inverse totheseemotions,showingadecreasingtendencyfromAct 1 to Act 5 for all genres with exception ofhumorousfictionthatshowsaplateaubetweenActs1and4.

Inadventures,allemotionsincreaseinthefirsthalfofthebooks,butdropsharplybetweenAct4andAct5.ThisisconsistentwithWhetter(2008),accordingtowhom adventures are marked by a late drop inemotionaltensionwhenthehero’smisfortunescometo an end. The only exception is trust that showsincrease towards the end for all genres, which isespeciallynoticeableinadventures.Apotentialreasonis that prototypical adventure novels are ‘upbeat’ inthattheycultivatevirtuessuchascourageandloyalty(Baldick,2015,p.5)anddepictheroesthatdonotlosetrustevenamidunexpecteddangers. Theresultsforanticipationandsurpriseshowlesscoherent tendencies which we find difficult tointerpret.Thesetwoemotionsappearlessconstitutivetothenarrativestructureofgenres,atleastthosethatwecurrently consider: anticipationandsurprise canoccur under (almost) any circumstances. Mysteryfiction has a slightly different pattern, whereanticipationexhibitssteadyincreasefromActs1to4anditspeakcoincideswiththepeakforsurpriseatAct4.

Quantitative analysis We analyze the genre-specific time course ofemotions quantitatively by computing associationsbetween genres and the Act in which an emotion“peaks”inastory.Wedefinearandomvariablevieforan emotion peak as vie=1 iff the highest value ofemotion e is in Act i. The association between eachgenre and emotion peaks vie follows point-wisemutualinformation(ChurchandHanks,1990),

where probabilities are computed as relativefrequenciesoverthedataset. Figure 2 gives insight into the genre-specificemotion patterns. For instance, disgust ischaracteristic of Act 4 for all genres. The onlyexceptionissciencefictionthatdoesnotlistdisgustorsurpriseamongthetop10features.Trustisimportantat the beginning and in the end of adventures andscience fiction, but is missing in mystery. Similarly,romance fiction is not characterized by anticipationamong top-ranked features, corresponding to its“anticipation” curve that decreases monotone frombeginning to end. Interestingly, humor is the onlygenre that does not contain joy among the top 10features.

Figure 1: Genre-specific emotion curves

Figure 2: Top 10 PMI features for each genre

Related work Sentiment and emotion in fiction have beenpreviouslyaddressedcomputationallybyMohammad(2012),Nalisnick andBaird (2013),Heuser,Moretti,and Steiner (2016), among others. Samothrakis andFasli(2015) istheonlyworkweareawareofwhichdiscussesemotionsincontextofgenres. The study most related to ours is Reagan et al.(2016).Theyproposeapipelinethattracksemotionsintext.Theirmainclaimisthatstoriestypicallyfollowoneoutofsix“emotionalarcs”regardinghappiness.

Conclusion, discussion and future work We investigated the relationship betweenemotional development in literature and genre andobserve differences among emotions. The genre ofadventurestandsout,especiallyconcerningtheendof

thestoryarc.Ourresultscanprovideanovelstartingpoint for characterizing similarities and differenceswithinandbetweenliterarygenres. Our observations require further investigationregardingtheunderlyingfactors.Forinstance,itmightbe argued that the pattern for mystery stories isdominated by the subgenre of crime fiction. Futureworkwillcombineourdistantreadingapproachwithclose and scalable reading. Furthermore, to improveemotion recognition, we plan to use distributionalmethods forexpanding theexisting lexical resourcesand adapting them to texts from different historicalperiods(cf.Buechel,Hellrich,andHahn,2016).

Acknowledgements This research has been conducted within theCenterforReflectedTextAnalytics(CRETA),whichisfunded by the German Ministry for Education andResearch(BMBF).

Bibliography Baldick, C. (2015). TheOxford dictionary of literary terms.

OUPOxford.Biber,D.andConradS.(2009).Register,Genre,andStyle.

CambridgeUniversityPress.Buechel,S.,HellrichJ.,andHahnU.(2016).Feelingsfrom

the Past – Adapting Affective Lexicons for HistoricalEmotionAnalysis.In:LT4DH2016,p.54.

Chatman,S.(1978).StoryandDiscourse:NarrativeStructure

andFictionandFilm.CornellUniversityPress.Church, K.W. and Hanks, P. (1990). Word association

norms,mutualinformation,andlexicography.Computa-tionalLinguistics,16(1),pp.22-29.

Cuddon,J.A.(2012).Dictionaryofliterarytermsandliterary

theory.JohnWiley&Sons.Francis,W.N.andKuceraH.(1979).Browncorpusmanual.

BrownUniversity,Rhodesisland.Freytag, G. (1863).Die Technik des Dramas. Leipzig, Ger-

many:Hirzel.Gernsbacher,M.A.,GoldsmithH.H.,andRobertsonR.R.

(1992).Doreadersmentallyrepresentcharacters’emo-tionalstates?Cognition&Emotion,6(2),pp.89-111.

Heuser,R.,MorettiF.,andSteinerE.(2016).TheEmo-

tionsofLondon.Stanford:StanfordLiteraryLab,Stan-fordUniversity.Availableat:https://litlab.stan-ford.edu/LiteraryLabPamphlet13.pdf[accessed5Mar.2017].

Hogan,P.C.(2011).Whatliteratureteachesusaboutemo-

tion.CambridgeUniversityPress.Klinger,R., SuliyaS.S., andReiterN. (2016).Automatic

EmotionDetection forQuantitative Literary Studies: Acase study based on Franz Kafka’s “Das Schloss” und“Amerika. In: Digital Humanities 2016: Conference Ab-stracts. JagiellonianUniversity&PedagogicalUniversity,Krakow,Poland,pp.826-828.

Mohammad,S.M.(2012).Fromonceuponatimetohappily

everafter:Trackingemotionsinmailandbooks.DecisionSupportSystems,53(4),pp.730-741.

Mohammad,S.M.andTurneyP.D.(2013).Crowdsourc-

ingaword-emotionassociation lexicon.ComputationalIntelligence,29(3),pp.436-465.

Nalisnick, E. T. andBairdH. S. (2013). Extracting senti-

mentnetworksfromShakespeare’splays.In:DocumentAnalysis and Recognition (ICDAR) 2013 12th Interna-tionalConferenceon.IEEE,pp.758-762.

Plutchik,R.(2001).TheNatureofEmotions.Humanemo-

tionshavedeepevolutionaryroots,a fact thatmayex-plaintheircomplexityandprovidetoolsforclinicalprac-tice.AmericanScientist,89(4),pp.344-350.

Reagan, A. J., Mitchell L., Kiley D., Danforth C.M., and

Dodds P. S. (2016). The emotional arcs of stories aredominatedbysixbasicshapes.EPJDataScience,5(1),p.31.

Samothrakis,S.andFasliM. (2015).Emotionalsentence

annotationhelpspredictfictiongenre.PloSOne,10(11).Vega,M.(1996).Therepresentationofchangingemotions

inreadingcomprehension.Cognition&Emotion,10(3),pp.303-322.

Whetter,K.S.(2008).Understandinggenreandmedievalro-

mance.AshgatePublishing,Ltd.