Modelling Valence and Arousal in Facebook Posts LyleUngar D.Preo¸tiuc-Pietro,H.A.Schwartz G.Park,J.Eichsteadt,M.Kern,E.Shulman PositivePsychologyCenter UniversityofPennsylvania 16June2016 Motivation DataSources Productreviews Affectivestates Opinionstowardsproducts, Feelingstowardsselfor restaurants,events,etc. others. Long,morestructured Short,lessstructured Modelsofproductsentimentandemotionshouldbedifferent Motivation Modelsofemotion DiscreteEmotions DimensionalModels MostpopularinNLPareEkman’s Eachaffectivestateisa sixemotions: anger,disgust,fear, combinationofreal-valued joysadness,surprise components Caucasian Facial Expressions of EmotionMatsumoto & Ekman - Japanese and MmTnvwaeoouleodsrnteoicnplped(ohR(epyoupusrseilsosaneelrdloni1egts9niimtc8tha0eel,nsPctyio)rssactnuenemmdr2sap:0rloe0xu5)s)al Someemotionsdrivenbysimilar words(hell,bad→sadness,fear, anger) Emotion Circumplex Source: Jonker & Van der Merwe - Emotion episodes of Afrikaans-speaking employees in the workplace Applications Goal: Automatedlarge-scalepsychologicalstudies • measuringtime-of-dayandday-of-weekmoodswings • andwhatcausesthem • mentalillnessdetection • bipolar,schizophrenicbreaks... • analysingmoviesandbooks • andhowtheyvaryinemotioncontent • correlatingwithexternaleffects • e.g. weather,sportsgameoutcomes,... Measuring Valence and Arousal • Valence(orsentimentorpolarity) • 1(verynegative)–5(neutral/objective)–9(verypositive) • Arousal(orintensity) • 1(neutral/objectivepost)–9(veryhighintensity) Examples Message V A Is the one whoz GOing to Light Up your 7 8 Day!!!!!!!!!!!! Blessedwithababyboytoday... 7.5 2 theboringlifeisback:(... 3 2.5 ISSUPERSTRESSEDANDITSJUSTTHESEC- 2.5 7 ONDMONTHOFSCHOOL..D: Exampleofpostsannotatedwithaveragevalence(V)andarousal(A) ratings. Data Source 3120Facebookposts Stratifiedby: • Age(13-35) • Gender(M/F) Eachmessagefromadistinctuser Allmessagesfromthesametimeinterval Annotation Twoannotators: • psychologystudents • receivedtraininginannotatingthesetraits,including anchoring • nodistractionsthatmayaffecttheymood(music,etc.) Messagesareun-ratableiftheyarenotinEnglishorcontainno cues • 235messages(∼7.5%) • CohensKappaκ=.93 Annotation Results 900 500 700 mber of posts500 mber of posts300 Nu300 Nu 100 100 1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9 Valence Arousal Valence Arousal Histogramsofaverageratingscores. Valence–Arousal→r = 0.222 Valence–Arousal→r = 0.085(ignoringneutralposts)