do you feel what i feel? - github pages · • sympathy, apology, and complaining influence...

1
Do you feel what I feel? Social Aspects of Emotions in Twitter Conversations Suin Kim, JinYeong Bak, Yohan Jo and Alice Oh Department of Computer Science KAIST, Daejeon, South Korea [email protected], {jy.bak, yohan.jo}@kaist.ac.kr, [email protected] Research Questions Do conversational partners influence each other’s sentiment and emotion? What topics can affect conversational partner’s sentiment? How can we find sentiments and emotions from text? Our Work 1. Emotion discovery using a probabilistic topic model 2. Analysis of sentiment and emotion transition patterns 3. Identification of major topics that cause emotion influence #Users #Dyads #Chains #Tweets Tweets per chain 62,952 77,850 153,054 871,544 5.69 Conversation Data We define consecutive replies of tweets using the reply button as a chain. Only chains of length >= 4 are considered. Contributions Interlocutors converge toward a common sentiment/emotion as conversation proceeds. Twitter users express lots of love and joy even when the partner expresses a negative emotion. Sympathy, apology, and complaining influence emotion. A: Sorry to hear about your bags. If you would like us to get someone to contact you DM us your reference and contact number. B: it’s on it’s way to manch. If the woman on the check in desk in Miami hadn’t been trying to be all smart! Been no problem. A: Sorry about that. Pleased to hear they located it quickly for you though. B: mistakes happen. Apology Complaint Sentiment Patterns in Conversations Interlocutors tend to converge toward a common sentiment as conversation proceeds. What if two interlocutors have opposing sentiments? Overall sentiment of u in conversation v: Emotions in sentiment-opposing conversations Agony - Sympathy Complaint - Apology Senti(u, v )= p u,v - n u,v p u,v + n u,v Example of sentiment-opposing conversation ASUM A B 1 2 3 4 A B 1 2 3 4 Conversation corpus Sentiment seed words :) :D XD (: :-) =) :)) :] =D =] :-D 8) :p :( :/ ): :’( :S =/ :- ( =( :-/ : (( ]: :\ /: Positive Negative + Topic 18 song album music new her she amazing listening awesome great Topic 8 thank :D much very welcome Topic 21 morning day hope hello Topic 59 she :( car were off home her phone house night Topic 95 eat want food chicken hungry Topic 73 she hate being never Thank you very much. :D Topic 8, Positive Topics per each sentiment Sentiment-Topic classification Sentiment-Topic distribution 0 1 Positive Negative 0 0.6 Topics A B 1 2 3 4 Finding Conversational Topics Using ASUM (LDA-based joint model of topics and sentiments) ASUM automatically discovers topics for each sentiment. Twitter Conversation Tweet http://uilab.kaist.ac.kr Social Aspects of Sentiments Q: How does the sentiment in a tweet lead to the sentiment in the reply? : Sentiment transition 0.24 0.36 0.35 0.13 0.27 0.16 Neutral 19.4% 0.63 0.38 0.48 xx :) thank nice very <3 haha :-) okay :( lmao shit didn’t man don’t bad :/ sorry people phone day twitter song want team money cool morning great never work hey off hope sure :P way say tomorrow only Negative 32.1% Positive 48.5% User A User B Having a tough day today. RIP Harrison. I’ll miss you a ton :/ Just pray about it. God will help you. Not really religious, but thanks man. :) If you need talk you know I’m here. Time Sliding Window (Negative) (Positive) Analyzing sentiment transition by sliding window method Positive Negative Topic 17 lmao ur ass lol shit Topic 48 weather rain :( hot cold Topic 38 :( school exam tomorrow year Topic 54 sleep work still awake tired Positive Positive Topic 16 follow thank please followed welcome Topic 8 thank :D much very welcome Topic 19 heart xD big kisses much Topic 21 morning day hope hello happy Negative Positive Topic 44 hope better sorry feel soon Topic 59 home want house sleep bed Topic 47 today week still feel work Topic 37 money buy want had pay Negative Negative Topic 72 shit bro da chillin nigga Topic 49 game lmao shit team win Topic 42 off still bad feel down Topic 32 ur loool lool looool man Sympathy Swear Mocking Teasing Complaint Greeting Topics that influence the conversational partner’s sentiment Q: How does the sentiment and the topic of a tweet influence the partner’s sentiment? : Sentiment influence A: Self-transitions make up the largest proportion. Transitions to positive sentiments are high. Primary Secondary & Tertiary Love Affection, Lust, Longing, Adoration, Fondness, Liking, Arousal, Desire, Infatuation, Longing, .. Joy Cheerfulness, Zest, Contentment, Pride, Optimism, Enthrallment, Relief Amusement, ... Surprise Surprise Amazement, Astonishment Anger Irritability, Exasperation, Rage, Disgust, Envy, Torment Aggravation, Agitation, Annoyance, ... Sadness Suffering, Sadness, Disappointment, Shame, Neglect, Sympathy Agony, Anguish, ... Fear Horror, Nervousness Alarm, Shock, Fear, Horror, Terror, Panic, Hysteria, Anxiety, ... Parrott’s tree-structured list of emotion Social Aspects of Emotions Finding emotions from tweets 0.19 0.11 0.11 0.16 0.11 0.11 0.1 0.12 0.14 0.1 0.11 Love Fear Sadness Surprise Joy Anger 0.68 0.55 0.65 0.64 0.4 0.46 Corr measures correlation between emotion c and topic t. Spec measures specialization of certain emotion for topic t. Topics with high Spec score will be classified as emotion topics. Spec(t)= max c Corr (c, t) P c Corr (c, t) - max c Corr (c, t) Corr (c, t)= γ c n c n c X i=1 P (w i |φ t ) Sadness Joy Anger Joy Surprise Topic 81 better feel hope pain sick :( Topic 99 hope sorry better feel soon hugs Topic 23 day night party home fun come Topic 17 eat food cream make chocolate drink Love Topic 18 song album amazing listening great Anger Topic 79 song shit new lmao never Emotion topics 0 0.6 Low Spec 0 0.8 High Spec Q: How does the emotion and the topic of a tweet influence the partner’s emotion? : Emotion influence Q: How does the emotion in a tweet lead to the emotion in the reply? : Emotion transition Sadness Topic 99 hope sorry better feel soon aww okay hugs feeling happened Topic 86 :’( miss him sorry want sad wish wanna much cry Joy Topic 31 day hope morning great weekend Topic 23 party home fun tomorrow Topic 40 happy birthday th hahaha year Topic 46 thank much very hope Love Topic 1 her smiles laughs eyes want Topic 47 school luck exam tomorrow Topic 3 twitter justin follow selena fan Topic 9 pic hair look nice Fear Topic 60 :/ school tomorrow exam year doing much day next maths Topic 87 dont lmao tell want him her please face who omg Anger Topic 95 eat want food chicken hungry Topic 73 she hate being never Topic 59 she :( car were off home her phone house night Surprise Topic 18 song album music new her she amazing listening awesome great Topic 30 where live here same awesome Topic 68 she old than look

Upload: others

Post on 08-Jan-2020

9 views

Category:

Documents


0 download

TRANSCRIPT

Do you feel what I feel?Social Aspects of Emotions in Twitter Conversations

Suin Kim, JinYeong Bak, Yohan Jo and Alice OhDepartment of Computer Science

KAIST, Daejeon, South [email protected], {jy.bak, yohan.jo}@kaist.ac.kr, [email protected]

Research Questions• Do conversational partners influence each other’s sentiment

and emotion?• What topics can affect conversational partner’s sentiment?• How can we find sentiments and emotions from text?

Our Work1. Emotion discovery using a probabilistic topic model2. Analysis of sentiment and emotion transition patterns3. Identification of major topics that cause emotion influence #Users #Dyads #Chains #Tweets Tweets per chain

62,952 77,850 153,054 871,544 5.69

Conversation DataWe define consecutive replies of tweets using the reply button as a chain.Only chains of length >= 4 are considered.

Contributions• Interlocutors converge toward a common sentiment/emotion as

conversation proceeds.• Twitter users express lots of love and joy even when the partner

expresses a negative emotion.• Sympathy, apology, and complaining influence emotion.

A: Sorry to hear about your bags. If you would like us to get someone to contact you DM us your reference and contact number.

B: it’s on it’s way to manch. If the woman on the check in desk in Miami hadn’t been trying to be all smart! Been no problem.

A: Sorry about that. Pleased to hear they located it quickly for you though.

B: mistakes happen.■Apology ■Complaint

Sentiment Patterns in Conversations• Interlocutors tend to converge toward a common sentiment as conversation proceeds.

• What if two interlocutors have opposing sentiments?Overall sentiment of u in conversation v: Emotions in sentiment-opposing conversations

• Agony - Sympathy• Complaint - Apology

Senti(u, v) =pu,v � nu,v

pu,v + nu,v

Example of sentiment-opposing conversation

ASUM

A B1

2

3

4

A B1

2

3

4

Conv

ersa

tion

corp

us

Sentiment seed words:) :D XD (: :-) =) :)) :] =D =] :-D 8) :p

:( :/ ): :’( :S =/ :-( =( :-/ :(( ]: :\ /:

Positive Negative

+

Topic 18songalbummusicnewhersheamazinglisteningawesomegreat

Topic 8thank:DmuchverywelcomeTopic 21morningdayhopehello

Topic 59she:(carwereoffhomeherphonehousenight

Topic 95eatwantfoodchickenhungryTopic 73shehatebeingnever

Thank you very much. :D

Topic 8, Positive

Topics pereach sentiment

Sentiment-Topic classification

Sentiment-Topicdistribution

0

1

Positive Negative

0

0.6

Topics

A B1

2

3

4

Finding Conversational Topics Using ASUM (LDA-based joint model of topics and sentiments)ASUM automatically discovers topics for each sentiment.

TwitterConversation

Tweet

http://uilab.kaist.ac.kr

Social Aspects of SentimentsQ: How does the sentiment in a tweet lead to the sentiment in the reply? : Sentiment transition

0.24

0.36

0.35

0.13

0.270.16

Neutral19.4%

0.63

0.38

0.48

xx:)

thanknicevery

<3

haha

:-)

okay

:( lmao

shit

didn’tman

don’t

bad :/

sorry

people phonedaytwitter song

wantteam

money

cool

morninggreat

never

workhey

off

hopesure :P

waysaytomorrow

only

Negative32.1%

Positive48.5%

User A

User B

Having a tough day today. RIP Harrison. I’ll

miss you a ton :/

Just pray about it. God will help you.

Not really religious, but thanks man. :)

If you need talk you know I’m here.

Time

Sliding Window

(Negative) (Positive)

Analyzing sentiment transition by sliding window method

Positive → Negative

Topic 17lmaourasslolshitTopic 48weatherrain:(hotcold

Topic 38:(schoolexamtomorrowyearTopic 54sleepwork stillawaketired

Positive → Positive

Topic 16followthankpleasefollowedwelcomeTopic 8thank:Dmuchverywelcome

Topic 19heartxDbigkissesmuchTopic 21morningdayhopehellohappy

Negative → Positive

Topic 44hopebettersorryfeelsoonTopic 59homewanthousesleepbed

Topic 47todayweekstillfeelworkTopic 37moneybuywanthadpay

Negative → Negative

Topic 72shitbrodachillinniggaTopic 49gamelmaoshitteamwin

Topic 42offstillbadfeeldownTopic 32urlooolloolloooolman

SympathySwear

MockingTeasing

ComplaintGreeting

Topics that influence the conversational partner’s sentiment

Q: How does the sentiment and the topic of a tweet influence the partner’s sentiment? : Sentiment influence

A: ① Self-transitions make up the largest proportion. ② Transitions to positive sentiments are high.

Primary Secondary & Tertiary

Love Affection, Lust, Longing, Adoration, Fondness, Liking, Arousal, Desire, Infatuation, Longing, ..

Joy Cheerfulness, Zest, Contentment, Pride, Optimism, Enthrallment, Relief Amusement, ...

Surprise Surprise Amazement, Astonishment

Anger Irritability, Exasperation, Rage, Disgust, Envy, Torment Aggravation, Agitation, Annoyance, ...

Sadness Suffering, Sadness, Disappointment, Shame, Neglect, Sympathy Agony, Anguish, ...

Fear Horror, Nervousness Alarm, Shock, Fear, Horror, Terror, Panic, Hysteria, Anxiety, ...

Parrott’s tree-structured list of emotion

Social Aspects of Emotions

Finding emotions from tweets0.19

0.110.11

0.16

0.11

0.11

0.1

0.120.14

0.1

0.11

Love Fear

Sadness

SurpriseJoy

Anger

0.68

0.55

0.65 0.64

0.4

0.46

• Corr measures correlation between emotion c and topic t.

• Spec measures specialization of certain emotion for topic t.

Topics with high Spec score will be classified as emotion topics.

Spec(t) =

maxc Corr(c, t)Pc Corr(c, t)�maxc Corr(c, t)

Corr(c, t) =�c

nc

ncX

i=1

P (wi|�t)Sadness → JoyAnger → Joy SurpriseTopic 81betterfeelhopepainsick:(

Topic 99hopesorrybetterfeelsoonhugs

Topic 23daynightpartyhomefuncome

Topic 17eatfoodcreammakechocolatedrink

→LoveTopic 18songalbumamazinglisteninggreat

→AngerTopic 79songshitnewlmaonever

Emotion topics

0

0.6

Low Spec 0

0.8

High Spec

Q: How does the emotion and the topic of a tweet influence the partner’s emotion? : Emotion influence

Q: How does the emotion in a tweet lead to the emotion in the reply? : Emotion transition

SadnessTopic 99hopesorrybetterfeelsoonawwokayhugsfeelinghappened

Topic 86:’(misshimsorrywantsadwishwannamuchcry

JoyTopic 31dayhopemorninggreatweekendTopic 23partyhomefuntomorrow

Topic 40happybirthdaythhahahayearTopic 46thankmuchveryhope

LoveTopic 1hersmileslaughseyeswantTopic 47schoolluckexamtomorrow

Topic 3twitterjustinfollowselenafanTopic 9pichairlooknice

FearTopic 60:/schooltomorrowexamyeardoingmuchdaynextmaths

Topic 87dontlmaotellwanthimherpleasefacewhoomg

AngerTopic 95eatwantfoodchickenhungryTopic 73shehatebeingnever

Topic 59she:(carwereoffhomeherphonehousenight

SurpriseTopic 18songalbummusicnewhersheamazinglisteningawesomegreat

Topic 30whereliveheresameawesomeTopic 68sheoldthanlook