Climate change sentiment on Twitter: An unsolicited public opinion poll

Emily M. Cody,1, ∗ Andrew J. Reagan,1, † Lewis Mitchell,2, ‡

Peter Sheridan Dodds,1, § and Christopher M. Danforth1, ¶

1Department of Mathematics & Statistics, Vermont Complex Systems Center,Computational Story Lab, & the Vermont Advanced Computing Core,

The University of Vermont, Burlington, VT 05401.2School of Mathematical Sciences, The University of Adelaide, SA 5005, Australia

(Dated: July 7, 2019)

The consequences of anthropogenic climate change are extensively debated through scientificpapers, newspaper articles, and blogs. Newspaper articles may lack accuracy, while the severity offindings in scientific papers may be too opaque for the public to understand. Social media, however,is a forum where individuals of diverse backgrounds can share their thoughts and opinions. Asconsumption shifts from old media to new, Twitter has become a valuable resource for analyzingcurrent events and headline news. In this research, we analyze tweets containing the word “climate”collected between September 2008 and July 2014. We determine how collective sentiment varies inresponse to climate change news, events, and natural disasters. Words uncovered by our analysissuggest that responses to climate change news are predominately from climate change activistsrather than climate change deniers, indicating that Twitter is a valuable resource for the spread ofclimate change awareness.

INTRODUCTION

After decades receiving little attention from non-scientists, the impacts of climate change are now widelydiscussed through a variety of mediums. Originatingfrom scientific papers, newspaper articles, and blog posts,a broad spectrum of climate change opinions, subjects,and sentiments exist. Newspaper articles often dismissor sensationalize the effects of climate change due tojournalistic biases including personalization, dramatiza-tion and a need for novelty [6]. Scientific papers por-tray a much more realistic and consensus view of climatechange. These views, however, do not receive widespreadmedia attention due to several factors including journalpaywalls, formal scientific language, and technical resultsthat are not easy for the general public to understand [6].

According to the IPCC Fifth Assessment report, hu-mans are “very likely” (90-100% probability) to be re-sponsible for the increased warming of our planet [13],and this anthropogenic global warming is responsible forcertain weather extremes [15]. In April 2013, 63% ofAmericans reported that they believe climate change ishappening. This number, however, drops to 49% whenasked if climate change is being caused by humans. Thepercentage drops again to 38% when asked if peoplearound the world are currently being harmed by the con-sequences of climate change [24]. These beliefs and riskperceptions can vary by state or by county [18]. By con-trast, 97% of active, publishing, climate change scientistsagree that “human activity is a significant contributing

∗ emily.cody@uvm.edu† andy@andyreagan.com‡ lewis.mitchell@adelaide.edu.au§ peter.dodds@uvm.edu¶ chris.danforth@uvm.edu

factor in changing mean global temperatures” [2, 12].The general public learns most of what it knows aboutscience from the mass-media [42]. Coordination amongjournalists, policy actors, and scientists will help to im-prove reporting on climate change, by engaging the gen-eral public and creating a more informed decision-makingprocess [5].

One popular source of climate information that hasnot been heavily analyzed is social media. The Pew Re-search Center’s Project for Excellence in Journalism inJanuary of 2009 determined that topics involving globalwarming are much more prominent in the new, social me-dia [5]. In the last decade, there has been a shift fromthe consumption of traditional mass media (newspapersand broadcast television) to the consumption of socialmedia (blog posts, Twitter, etc.). This shift represents aswitch in communications from “one-to-many” to “many-to-many” [5]. Rather than a single journalist or scientisttelling the public exactly what to think, social media of-fers a mechanism for many people of diverse backgroundsto communicate and form their own opinions. Exposureis a key aspect in transforming a social problem into apublic issue [9], and social media is a potential avenuewhere climate change issues can be initially exposed.

Here we study the social media site Twitter, which al-lows its users 140 characters to communicate whateverthey like within a “tweet”. Such expressions may includewhat individuals are thinking, doing, feeling, etc. Twit-ter has been used to explore a variety of social and lin-guistic phenomena [7, 26, 27], and used as a data sourceto create an earthquake reporting system in Japan [39],detect influenza outbreaks [3], and analyze overall publichealth [36]. An analysis of geo-tagged Twitter activity(tweets including a latitude and longitude) before, dur-ing, and after Hurricane Sandy using keywords relatedto the storm is given in [22]. They discover that Twitteractivity positively correlates with proximity to the storm

Typeset by REVTEX

arX

iv:1

505.

0380

4v1

[ph

ysic

s.so

c-ph

] 1

4 M

ay 2

015

2

and physical damage. It has also been shown that in-dividuals affected by a natural disaster are more likelyto strengthen interactions and form close-knit groups onTwitter immediately following the event [37]. Twitterhas also been used to examine human sentiment throughanalysis of variations in the specific words used by indi-viduals. In [11], Dodds et al. develop the “hedonometer”,a tool for measuring expressed happiness – positive andnegative sentiment – in large-scale text corpora. Since itsdevelopment, the hedonometer has been implemented instudies involving the happiness of cities and states [30],the happiness of the English language as a whole [21],and the relationship between individuals’ happiness andthat of those they connect with [4].

The majority of the topics trending on Twitter areheadlines or persistent news [23], making Twitter a valu-able source for studying climate change opinions. For ex-ample, in [1], subjective vs objective and positive vs neg-ative tweets mentioning climate change are coded manu-ally and analyzed over a one year time period. In thepresent study, we apply the hedonometer to a collec-tion of tweets containing the word “climate”. We col-lected roughly 1.5 million such tweets from Twitter’sgardenhose API (a random 10% of all messages) dur-ing the roughly 6 year period spanning September 14,2008 through July 14, 2014. Each collected tweet con-tains the word “climate” at least once, and we includeretweets in the collection. We apply the hedonometer tothe climate tweets during different time periods and com-pare them to a reference set of roughly 100 billion tweetsfrom which the climate-related tweets were filtered. Weanalyze highest and lowest happiness time periods us-ing word shift graphs developed in [11], and we discussspecific words contributing to each happiness score.

METHODS

The hedonometer is designed to calculate a happinessscore for a large collection of text, based on the happinessof the individual words used in the text. The instrumentuses sentiment scores collected by Kloumann et al. andDodds et al. [11, 21], where 10,222 of the most frequencyused English words in four disparate corpora were givenhappiness ratings using Amazon’s Mechanical Turk on-line marketplace. Fifty participants rated each word, andthe average rating becomes the word’s score. Each wordwas rated on a scale from 1 (least happy) to 9 (mosthappy) based on how the word made the participant feel.We omit clearly neutral or ambiguous words (scores be-tween 4 and 6) from the analysis. In the present study,we use the instrument to measure the average happinessof all tweets containing the word “climate” from Septem-ber 2008 to July 2014 on the timescales of day, week, andmonth. The word “climate” has a score of 5.8 and wasthus not included when calculating average happiness.

We recognize that not every tweet containing theword “climate” is about climate change. Some of these

tweets are about the economic, political, or socialclimate and some are ads for climate controlled cars.Through manual coding of a random sample of 1,500climate tweets, we determined that 93.5% of tweetscontaining the word “climate” are about the earth’sclimate or climate change. We calculated the happinessscore for both the entire sample and the sample withthe non-earth related climate tweets removed. Thescores were 5.905 and 5.899 respectively, a differenceof 0.1%. This difference is small enough to concludethat the non-earth related climate change tweets do notsubstantially alter the overall happiness score.

Based on the happiness patterns given by the hedo-nometer analysis, we select specific days for analysis us-ing word shift graphs. We use word shift graphs to com-pare the average happiness of two pieces of text, by rankordering the words that contribute the most to the in-crease or decrease in happiness. In this research, thecomparison text is all tweets containing the word “cli-mate”, and the reference text is a random 10% of alltweets. Hereafter, we refer to the full reference collectionas the “unfiltered tweets”.

Finally, we analyze four events including three naturaldisasters and one climate rally using happiness time seriesand word shift graphs. These events include HurricaneIrene (August 2011), Hurricane Sandy (October 2012), amidwest tornado outbreak (May 2013), and the Forwardon Climate Rally (February 2013).

RESULTS

Fig. 1 gives the raw and relative frequencies of theword “climate” over the study period. We calculate therelative frequencies by dividing the daily count of “cli-mate” by the daily sum of the 50,000 most frequentlyused words in the gardenhose sample. From this figure,we can see that while the raw count increases over time,the relative frequency decreases over time. This decreasecan either be attributed to reduced engagement on theissue since the maximum relative frequency in December2009, during Copenhagen Climate Change Conference, oran increase in overall topic diversity of tweets as Twit-ter grows in popularity. The observed increase in rawcount can largely be attributed to the growth of Twit-ter during the study period from approximately 1 milliontweets per day in 2008 to approximately 500 million in2014. In addition, demographic changes in the user pop-ulation clearly led to a decrease in the relative usage ofthe word “climate”.

Fig. 2 shows the average happiness of the climatetweets by day, by week, and by month during the 6 yeartime span. The average happiness of unfiltered tweetsis shown by a dotted red line. Several high and lowdates are indicated in the figure. The average happinessof tweets containing the word “climate” is consistentlylower than the happiness of the entire set of tweets.

3

2009 2010 2011 2012 2013 20140

2000

4000

6000

8000

10000

Cou

nt

Daily count of "climate"

2009 2010 2011 2012 2013 20140

0.5

1

1.5

x 10−4

Rel

ativ

e fre

quen

cy

Daily relative frequency of "climate"

100

102

104

10−6

10−5

10−4

10−3

Figure 1. The daily raw frequencies (top) and relative frequencies (bottom) of the word “climate” on Twitter from September2008 to July 2014. The insets (in red) show the same quantity with a logarithmically spaced y-axis.

2009 2010 2011 2012 2013 20145

5.5

6

6.5

7

Aver

age

Happ

ines

s

Happiness of Tweets Containing "climate" by Day

12/28

03/16

04/0908/26 04/30

10/09 04/04 01/08 08/06

2009 2010 2011 2012 2013 20145.5

6

6.5

Aver

age

Happ

ines

s

By Week

12/2803/15

05/17 04/29

10/12 05/09 07/18 10/28 11/1010/20

2009 2010 2011 2012 2013 20145.5

6

6.5

Aver

age

Happ

ines

s

By Month

Mar Apr

Oct Oct

Figure 2. Average happiness of tweets containing the word “climate” from September 2008 to July 2014 by day (top), by week(middle), and by month (bottom). The average happiness of all tweets during the same time period is shown with a dottedred line. Several of the happiest and saddest dates are indicated on each plot, and are explored in subsequent figures.

Several outlier days, indicated on the figure, do havean average happiness higher than the unfiltered tweets.Upon recovering the actual tweets, we discover that onMarch 16, 2009, for example, the word “progress” was

used 408 times in 479 overall climate tweets. “Progress”has a happiness score of 7.26, which increases the av-erage happiness for that particular day. Increasing thetime period for which the average happiness is measured

4

−20 −10 0 10 20

1

5

10

15

20

25

30

35

40

45

50

+↓loveno −↓

+↓meshit −↓

don’t −↓+↓haha−↑fight

+↓lol−↑crisis

+↓youhate −↓science +↑

−↑bill+↓like

+↓happynever −↓energy +↑

−↑threat−↑not

−↑pollution+↓hahaha

new +↑−↑against

−↑denialcan’t −↓bitch −↓dont −↓miss −↓ass −↓

−↑hell−↑tax+↓my−↑risk

−↑stopdamn −↓

−↑deny+↓good

−↑disasterheaven +↑

+↓lifedie −↓

+↓birthday−↑combat−↑denying

we +↑−↑war

−↑poor−↑hurricane

world +↑mad −↓

Per word average happiness shif t δh av g, r (%)

Word

rankr

T r e f : unfi l tered (havg=5.99)Tcom p: cl imate (havg=5.84)

Tex t s i z e :T r e f T c omp

+↓ +↑

−↑ −↓

Balanc e :

−387 : +287

−100 0

100

101

102

103

104

!ri=1δ hav g, i

Words  on  the  right  contribute  to  an  increase  in  happiness  in  the  climate  tweets    

Words  on  the  le5  contribute  to  a  decrease  in  happiness  in  the  climate  tweets    

A  yellow  bar  with  a  down  arrow  indicates  that  a  happy  word  was  used  less  

A  purple  bar  with  an  up  arrow  indicates  that  a  sad  word  was  used  more  

A  yellow  bar  with  an  up  arrow  indicates  that  a  happy  word  was  used  more    

A  purple  bar  with  a  down  arrow  indicates  that  a  sad  word  was  used  less  

Number  of  words  contribu:ng  to  the  total  shi5  in  happiness  

Comparison  of  number  of  words  in  each  of  the  4  happiness  shi5  categories  

The  size  of  the  text  including  all  tweets  versus  the  size  of  the  text  including  climate  tweets  

Using  all  tweets  as  the  reference  text,  the  words  are  ordered  based  on  their  net  contribu6on  toward  making  climate  tweets  sadder.  

Figure 3. A word shift graph comparing the happiness of tweets containing the word “climate” to all unfiltered tweets. Thereference text is roughly 100 billion tweets from September 2008 to July 2014. The comparison text is tweets containing theword “climate” from September 2008 to July 2014. A yellow bar indicates a word with an above average happiness score. Apurple bar indicates a word with below average happiness score. A down arrow indicates that this word is used less withintweets containing the word “climate”. An up arrow indicates that this word is used more within tweets containing the word“climate”. Words on the left side of the graph are contributing to making the comparison text (climate tweets) less happy.Words on the right side of the graph are contributing to making the comparison text more happy. The small plot in the lowerleft corner shows how the individual words contribute to the total shift in happiness. The gray squares in the lower right cornercompare the sizes of the two texts, roughly 107 vs 1012 words. The circles in the lower right corner indicate how many happywords were used more or less and how many sad words were used more or less in the comparison text.

(moving down the panels in Fig. 2), the outlier days be-come less significant, and there are fewer time periodswhen the climate tweets are happier than the referencetweets. After averaging weekly and monthly happinessscores, we see other significant dates appearing as peaksor troughs in Fig. 2. For example, the week of October28, 2012 appears as one of the saddest weeks for climatediscussion on Twitter. This is the week when HurricaneSandy made landfall on the east coast of the U.S. Forthe same reason, October 2012 also appears as one of thesaddest months for climate discussion.

The word shift graph in Fig. 3 shows which words con-tributed most to the shift in happiness between climatetweets and unfiltered tweets. The total average happi-ness of the reference text (unfiltered tweets) is 5.99 whilethe total average happiness of the comparison text (cli-

mate tweets) is 5.84. This change in happiness is due tothe fact that many positively rated words are used lessand many negatively rated words are used more whendiscussing the climate.

The word “love” contributes most to the change in hap-piness. Climate change is not typically a positive subjectof discussion, and tweets do not typically profess love forit. Rather, people discuss how climate change is a “fight”,“crisis”, or a “threat”. All of these words contribute tothe drop in happiness. Words such as “pollution”, “de-nial”, “tax”, and “war” are all negative, and are usedrelatively more frequently in climate tweets, contribut-ing to the drop in happiness. The words “disaster” and“hurricane” are used more frequently in climate tweets,suggesting that the subject of climate change co-occurswith mention of natural disasters, and strong evidence

5

exists proving Twitter is a valid indicator of real timeattention to natural disasters [38].

On the positive side, we see that relatively less profan-ity is used when discussing the climate, with the excep-tion of the word “hell”. We also see that “heaven” is usedmore often. From our inspection of the tweets, it is likelythat these two words appear because of a famous quote byMark Twain: “Go to heaven for the climate and hell forthe company” [41]. Of the 97 non-earth related climatetweets from our 1,500 tweet sample, 8 of them referencedthis quote. The word “energy” is also used more dur-ing climate discussions. This indicates that there may bea connection between energy related topics and climaterelated topics. As energy consumption and types of en-ergy sources can contribute to climate change, it is notsurprising to see the two topics discussed together.

Using the first half of our dataset, Dodds et. al. [11]calculated the average happiness of tweets containingseveral individual keywords including “climate”. Theyfound that tweets containing the word “climate” were,on average, similar in ambient happiness to those con-taining the words “no”, “rain”, “oil”, and “cold” (seeTable 2 [11]).

Analysis of Specific Dates

While Fig. 3 shows a shift in happiness for all climatetweets collected in the 6 year period, we now move to an-alyzing specific climate change-related time periods andevents that correspond to spikes or dips in happiness.Fig. 4 gives word shift graphs for three of the happiestdays according to the hedonometer analysis. These datesare indicated in the top plot in Fig. 2. The word shiftgraphs use unfiltered tweets as the reference text and cli-mate tweets as the comparison text for the date given ineach title.

Fig. 4(a) shows that climate tweets were happier thanunfiltered tweets on December 28, 2008. This is due inpart to a decrease in the word “no”, and an increase inthe words “united”, “play”, and “hopes”. On this day,there were “high hopes” for the U.S. response to climatechange. An example tweet by OneWorld News is givenin Fig. 5(a) [33].

Fig. 4(b) shows that climate tweets were happier thanunfiltered tweets on April 9, 2009, largely due to the in-crease in positive words “book”, “energy”, and “prize”.Twitter users were discussing the release of a new bookcalled Sustainable Energy Without the Hot Air by DavidJC MacKay [28]. Also on this date, users were postingabout a Climate Prize given to a solar-powered cookerin a contest for green ideas. Example tweets includeFig. 5(b) and (c) [10, 40]. Finally, Fig. 4(c) shows that cli-mate tweets were happier than unfiltered tweets on April30, 2012. This is due to the increased usage of the words“dear”, “new”, “protect”, “forest”, “save”, and “please”.On this date, Twitter users were reaching out to Brazil-ian president Dilma to save the Amazon rainforest, e.g.,

Fig. 5(d) [43].Similarly, Fig. 6 gives word shift graphs for three of

the saddest days according to the hedonometer analy-sis. These dates are indicated in the top panel in Fig. 2.Fig. 6(a) shows an increase in many negative words onOctober 9, 2008. Topics of conversation in tweets con-taining “climate” include the threat posed by climatechange to a tropical species, a British climate bill, and theU.S. economic crisis. Example tweets include Fig. 5(e-g)[8, 31, 34].

Fig. 6(b) shows an increase in negative words“poor”,“assault”, “battle”, and “bill” on April 4, 2010.Popular topics of conversation on this date included aCalifornia climate law and President Obama’s oil-drillingplan. Example tweets include Fig. 5(h) and (i) [32, 35].Finally, Fig. 6(c) shows that the words “don’t” and“stop” contributed most to the decrease in happiness onAugust 6, 2011. A topic of conversation on this date wasthe Keystone XL pipeline, a proposed extension to thecurrent Keystone Pipeline. An example tweet is given inFig. 5(j) [16].

This per day analysis of tweets containing “climate”shows that many of the important issues pertaining toclimate change appear on Twitter, and demonstrate dif-ferent levels of happiness based on the events that areunfolding. In the following section, we investigate spe-cific climate change events that may exhibit a peak or adip in happiness. First, we analyze the climate changediscussion during several natural disasters that may haveraised awareness of some of the consequences of climatechange. Then, we analyze a non-weather related eventpertaining to climate change.

Natural Disasters

Natural disasters such as hurricanes and tornados havethe potential to focus society’s collective attention andspark conversations about climate change. A person’s be-lief in climate change is often correlated with the weatheron the day the question is asked [17, 25, 44]. A studyusing “climate change” and “global warming” tweetsshowed that both weather and mass media coverage heav-ily influence belief in anthropogenic climate change [20].In this section, we analyze tweets during three naturaldisasters: Hurricane Irene, Hurricane Sandy, and a mid-west tornado outbreak that damaged many towns includ-ing Moore, Oklahoma and Rozel, Kansas. Fig. 7 givesthe frequencies of the words “hurricane” and “tornado”within tweets that contain the word “climate”. Each plotlabels several of the spikes with the names of the hurri-canes (top) or the locations (state abbreviations) of thetornado outbreaks (bottom). This figure indicates thatbefore Hurricane Irene in August 2011, hurricanes werenot commonly referenced alongside climate, and beforethe April 2011 tornado outbreak in Alabama and Missis-sippi, tornados were not commonly referenced alongsideclimate.

6

−20 −10 0 10 20

1

5

10

15

20

25

30

35

40

45

50

no −↓united +↑

play +↑hopes +↑

−↑crisis+↓love−↑loss

−↑fever−↑fight

−↑down−↑can’t

trees +↑−↑suck

leading +↑−↑chaos

high +↑not −↓

−↑wronggood +↑great +↑

−↑restlessdon’t −↓successful +↑

−↑toughbad −↓die −↓warm +↑won +↑

−↑taxes−↑buried

never −↓−↑worst

−↑shame+↓christmas

computer +↑+↓me

−↑least+↓happy

hate −↓+↓fun

perfectly +↑−↑fucked

perfect +↑−↑forget

progress +↑+↓you

soon +↑old −↓

−↑darknessinteresting +↑

Per word average happiness shif t δh av g, r (%)

Word

rankr

T r e f : 2008-12-28 (havg=6.06)Tcom p: cl imate (havg=6.27)

Tex t s i z e :T r e f T c omp

+↓ +↑

−↑ −↓

Balanc e :

−271 : +371

0 100

100

101

102

103

104

!ri=1δ hav g, i

a)

−40 −20 0 20 40

1

5

10

15

20

25

30

35

40

45

50

book +↑energy +↑

+↓love−↑never

prize +↑−↑pressure

solution +↑−↑closed

conservation +↑no −↓

−↑severewins +↑not −↓science +↑don’t −↓

−↑risks+↓me

−↑pollutionbad −↓

−↑victims−↑attacked−↑useless−↑doesn’t

die −↓garden +↑

−↑crisispositive +↑

−↑trash+↓happy

weekend +↑good +↑

−↑poor+↓lol

live +↑−↑decline

like +↑hate −↓

−↑fight−↑difficult

−↑avoid+↓you

−↑madwait −↓

−↑runaway−↑dire

shit −↓−↑apart

−↑missingperfect +↑

+↓home

Per word average happiness shif t δh av g, r (%)

Word

rankr

T r e f : 2009-04-09 (havg=6.07)Tcom p: cl imate (havg=6.27)

Tex t s i z e :T r e f T c omp

+↓ +↑

−↑ −↓

Balanc e :

−297 : +397

0 100

100

101

102

103

104

!ri=1δ hav g, i

b)

−20 0 20

1

5

10

15

20

25

30

35

40

45

50

dear +↑new +↑

protect +↑forest +↑

save +↑please +↑

no −↓+↓love

don’t −↓shit −↓

+↓me+↓like+↓you

hate −↓+↓lol

not −↓+↓haha−↑crisis

our +↑can’t −↓never −↓bitch −↓bad −↓

+↓goodass −↓

−↑combat+↓hahaha

damn −↓dont −↓

+↓happy+↓my

farms +↑miss −↓hell −↓energy +↑down −↓die −↓

+↓life−↑against

mad −↓wait −↓

+↓friends−↑pollution

+↓thanks−↑failing

−↑fighting+↓girl

didn’t −↓+↓free

+↓birthday

Per word average happiness shif t δh av g, r (%)

Word

rankr

T r e f : 2012-04-30 (havg=5.95)Tcom p: cl imate (havg=6.36)

Tex t s i z e :T r e f T c omp

+↓ +↑

−↑ −↓

Balanc e :

−135 : +235

0 100

100

101

102

103

104

!ri=1δ hav g, i

c)

Figure 4. Word shift graphs for three of the happiest days in the climate tweet time series.

(a) (b)

(c) (d)

(e) (f)

(g)

(h)

(i) (j)

Figure 5. Example tweets on the happiest and saddest daysfor climate conversation on Twitter

Fig. 7 shows that the largest peak in the word “hurri-cane” occurred during Hurricane Sandy in October 2012.Fig. 8 provides a deeper analysis for the climate time se-ries during hurricane Sandy. The time series of the words“hurricane” and “climate” as a fraction of all tweets be-fore, during, and after Hurricane Sandy hit are given inFig. 8(a) and (c). Spikes in the frequency of usage ofthese words is evident in these plots. The decay of eachword is fitted with a power law in Fig. 8(b) and (d). Apower law is a functional relationship of the followingform:

f(t− tevent) = α(t− tevent)−γ (1)

Here, t is measured in days, and tevent is the day Hurri-cane Sandy made landfall. f(t) represents the relativefrequency of the word “hurricane” (top) or “climate”(bottom), and α and γ are constants.

Using the power law fit, we calculate the first three halflives of the decay. Letting M equal the maximum relativefrequency, the time at which the first half life of the powerlaw relationship occurs is calculated by equation 2:

t 12

=

(M

2α

)− 1γ

(2)

The first three half lives of the decay in the frequency ofthe word “hurricane” during hurricane Sandy are 1.57,0.96, and 1.56 additional days. Since the decay is not ex-ponential, these half lives are not constant. The first half

7

−50 0 50

1

5

10

15

20

25

30

35

40

45

50

−↑threatened−↑breaking−↑pressure−↑diseases

−↑bad−↑crisis−↑tough

−↑turmoil−↑bill−↑flu

−↑worried−↑attacks

−↑not−↑losing

−↑sad−↑fever−↑fight−↑risk

−↑deadly−↑stress−↑ouch

−↑warning−↑misses

good +↑health +↑don’t −↓

−↑nothing+↓love

international +↑easier +↑

+↓megreat +↑wins +↑

−↑poverty−↑can’t

no −↓wildlife +↑die −↓trust +↑

−↑lastincrease +↑wow +↑down −↓true +↑friends +↑new +↑never −↓hate −↓movie +↑style +↑

Per word average happiness shif t δh av g, r (%)

Word

rankr

T r e f : 2008-10-09 (havg=6.00)Tcom p: cl imate (havg=5.29)

Tex t s i z e :T r e f T c omp

+↓ +↑

−↑ −↓

Balanc e :

−190 : +90

−100 0

100

101

102

103

104

!ri=1δ hav g, i

a)

−20 0 20

1

5

10

15

20

25

30

35

40

45

50

−↑poor−↑assault

−↑battle−↑bill

no −↓+↓happy

−↑death−↑bad+↓love+↓haha

+↓me+↓easter

+↓chocolate−↑attack−↑denial

−↑hurt−↑cut+↓lol

−↑not−↑stupid

−↑fails−↑chaos−↑failure−↑killed−↑dies

sun +↑energy +↑

−↑war−↑losses−↑dying−↑crisis

−↑ban+↓hahaha

dont −↓don’t −↓

+↓likebrilliant +↑

−↑declinedamn −↓

−↑combat−↑idiot

−↑avoid−↑problems

science +↑earned +↑

−↑global−↑risk

earthquake −↓miss −↓

−↑costs

Per word average happiness shif t δh av g, r (%)

Word

rankr

T r e f : 2010-04-04 (havg=6.11)Tcom p: cl imate (havg=5.37)

Tex t s i z e :T r e f T c omp

+↓ +↑

−↑ −↓

Balanc e :

−165 : +65

−100 0

100

101

102

103

104

!ri=1δ hav g, i

b)

−50 0 50

1

5

10

15

20

25

30

35

40

45

50

−↑don’t−↑stop

no −↓+↓love

−↑doesn’t−↑not+↓me

shit −↓+↓haha

−↑hell+↓lol

hate −↓−↑corruption

−↑fail−↑fight+↓like

+↓hahaha−↑attack+↓happy

never −↓+↓you

heaven +↑−↑tax

die −↓bad −↓

−↑avoiddont −↓

−↑violent−↑failure

−↑destructiondamn −↓ass −↓care +↑peace +↑miss −↓

−↑dying−↑crisis

−↑fearcan’t −↓

−↑shootingbible +↑

+↓photo+↓my

−↑fighting−↑severe

−↑fools−↑debt

last −↓−↑dead

wait −↓

Per word average happiness shif t δh av g, r (%)

Word

rankr

T r e f : 2011-08-06 (havg=5.98)Tcom p: cl imate (havg=5.38)

Tex t s i z e :T r e f T c omp

+↓ +↑

−↑ −↓

Balanc e :

−198 : +98

−100 0

100

101

102

103

104

!ri=1δ hav g, i

c)

Figure 6. Word shift graphs for 3 of the saddest days in the climate tweet time series.

2009 2010 2011 2012 2013 20140

100

200

300

400

Coun

t

Frequency of "hurricane" in Climate Tweets

irene issac

sandy

sandy 1 yr

2009 2010 2011 2012 2013 20140

20

40

60

80

100

Coun

t

Frequency of "tornado" in Climate Tweets

AL and MS

OK and KS

IL

Figure 7. Frequency of the word “hurricane’” (top) and “tor-nado” (bottom) within tweets containing the word “climate”.Several spikes have been identified with the hurricane or tor-nado that took place during that time period.

life indicates that after about a day and a half, “hurri-cane” was already tweeted only half as often. The secondhalf life indicates that after one more day, “hurricane”was tweeted only one fourth as often, and so on. Thus,it did not take long for the discussion of the hurricane todecrease. The half lives, however, of the word “climate”are much larger at 8.19, 22.58, and 84.85 days.

Fig. 9 gives happiness time series plots for three natu-

10/05/12 10/29/12 11/22/12

0.2

0.4

0.6

0.8

1 x 10−3

(a)

hurricane

Rel

ativ

e Fr

eque

ncy

10/05/12 10/29/12 11/22/12

0.5

1

1.5

2 x 10−5

(c)

climate

Rel

ativ

e Fr

eque

ncy

10/29/12 11/22/12

10−4

10−3

Rel

ativ

e Fr

eque

ncy

Exp = −1.45

(b)

10/29/12 11/22/12

10−4

10−3

Rel

ativ

e Fr

eque

ncy

Exp = −0.52

(d)

dataloglog fit

dataloglog fit

Figure 8. Decay rates of the words “hurricane” (top) and“climate” (bottom). The left plots gives the time series ofeach word during hurricane Sandy. The right plots gives thepower law fit for the decay in relative frequency, x-axes arespaced logarithmically. The power law exponents are given inthe titles of the figures.

ral disasters occurring in the United States. These plotsshow that there is a dip in happiness on the day that thedisasters hit the affected areas, offering additional evi-dence that sentiment is depressed by natural disasters [1].The word shift graphs indicate which words contributedto the dip in happiness. The circles on the bottom rightof the word shift plots indicate that for all three disas-

8

08/23 08/25 08/27 08/29 08/31 09/025.5

5.6

5.7

5.8

5.9

6H

appi

ness

Day (2011)

Hurricane Irene

10/24 10/26 10/28 10/30 11/01 11/035.6

5.65

5.7

5.75

5.8

5.85

5.9

5.95

6

6.05

Hap

pine

ss

Day (2012)

Hurricane Sandy

05/16 05/18 05/20 05/22 05/24 05/265.4

5.5

5.6

5.7

5.8

5.9

6

6.1

6.2

Happ

ines

s

Day (2013)

Midwest Tornado Outbreak

−100 −50 0 50 100

1

5

10

15

20

25

30

35

40

45

50

−↑threatathletic +↑no −↓

−↑hurricane+↓love

−↑worry−↑against−↑illness

shit −↓+↓haha

+↓me−↑worse

+↓lol−↑fight

don’t −↓basketball +↑

−↑nothate −↓never −↓

+↓hahaha+↓you

−↑blamedamn −↓

+↓good+↓like

ass −↓die −↓new +↑

+↓happylast −↓

−↑ignorebitch −↓bad −↓science +↑health +↑

+↓mydown −↓

−↑povertydont −↓

−↑problems−↑drowned

mad −↓−↑tornado

−↑fraudmiss −↓

−↑warwait −↓

−↑risk−↑crisis

−↑denying

Per word average happiness shif t δh av g, r (%)

Word

rankr

T r e f : 2011-08-28 (havg=5.97)Tcom p: cl imate (havg=5.53)

Tex t s i z e :T r e f T c omp

+↓ +↑

−↑ −↓

Balanc e :

−238 : +138

−100 0

100

101

102

103

104

!ri=1δ hav g, i

−50 0 50

1

5

10

15

20

25

30

35

40

45

50

−↑not−↑hurricane

−↑wrath−↑sucks

enjoy +↑−↑no

−↑slow−↑sorry

liked +↑+↓me

scientists +↑shit −↓

−↑worsenew +↑don’t −↓

+↓lol+↓haha

+↓like−↑warning

hate −↓−↑ignored

bitch −↓+↓my

can’t −↓bad −↓never −↓

+↓happyhealth +↑

−↑flood−↑punishment

−↑horriblereal +↑

+↓hahahaass −↓we +↑miss −↓

−↑disasterdont −↓

+↓lifeglad +↑damn −↓

−↑politiciansnature +↑believe +↑

−↑risksdie −↓you +↑

−↑warill −↓

−↑monster

Per word average happiness shif t δh av g, r (%)

Word

rankr

T r e f : 2012-10-29 (havg=5.92)Tcom p: cl imate (havg=5.60)

Tex t s i z e :T r e f T c omp

+↓ +↑

−↑ −↓

Balanc e :

−271 : +171

−100 0

100

101

102

103

104

!ri=1δ hav g, i

−20 0 20

1

5

10

15

20

25

30

35

40

45

50

−↑dammit−↑never

−↑no−↑sorry

−↑tornado+↓love

like +↑science +↑

+↓meshit −↓

−↑denyingdon’t −↓

−↑stop+↓you

−↑blame+↓haha

hate −↓+↓lol

−↑disaster−↑fighting+↓happy

bitch −↓+↓hahaha

+↓my−↑accident−↑tragedy

−↑risk−↑hurricane

miss −↓−↑worse

−↑bill−↑trouble

ass −↓+↓good

can’t −↓−↑hell

−↑attack−↑deaths

−↑crisis−↑poor

−↑againstdont −↓

−↑dangerouslast −↓we +↑ill −↓

−↑denial−↑awful

heaven +↑+↓friends

Per word average happiness shif t δh av g, r (%)

Word

rankr

T r e f : 2013-05-21 (havg=5.94)Tcom p: cl imate (havg=5.47)

Tex t s i z e :T r e f T c omp

+↓ +↑

−↑ −↓

Balanc e :

−213 : +113

−100 0

100

101

102

103

104

!ri=1δ hav g, i

1

Figure 9. Happiness time series plots for tweets containing the word “climate” one week before and one week after three naturaldisasters in the United States (top) and word shift graphs indicating what words contributed most to the drop in happinessduring the natural disasters (bottom). The word shift graphs compare the climate tweets to unfiltered tweets on the day of thenatural disaster.

ters, the dip in happiness is due to an increase in negativewords, more so than a decrease in positive words. Duringa natural disaster, tweets mentioning the word “climate”use more negative words than tweets not mentioning theword “climate”.

Forward on Climate Rally

In this section, we analyze tweets during the Forwardon Climate Rally, which took place in Washington D.C.on February 17, 2013. The goal of the rally, one of thelargest climate rallies ever in the United States, was to

convince the government to take action against climatechange. The proposed Keystone pipeline bill was a par-ticular focus. Fig. 10 shows that the happiness of climatetweets increased slightly above the unfiltered tweets dur-ing this event, which only occurs on 8% of days in Fig. 2.

Despite the presence of negative words such as“protestors”, “denial”, and “crisis”, the Forward on Cli-mate Rally introduced positive words such as “live”,“largest”, and “promise”. The Keystone pipeline bill waseventually vetoed by President Obama.

9

02/12 02/14 02/16 02/18 02/20 02/225.7

5.8

5.9

6

6.1

6.2

Hap

pine

ss

Day (2013)

Forward on Climate Rally

02/12 02/14 02/16 02/18 02/20 02/220.5

1

1.5

2

2.5 x 10−5

Rel

ativ

e Fr

eque

ncy

of c

limat

e

Day (2013)

−500 0 500

1

5

10

15

20

25

30

35

40

45

50

live +↑−↑protesters

largest +↑+↓me

−↑denialus +↑

shit −↓−↑bad

−↑crisis−↑protest

thirsty −↓don’t −↓hate −↓

+↓like+↓love−↑fight

no −↓+↓haha

+↓yourelax +↑

+↓lolnever −↓

−↑pressure+↓happy

−↑warning−↑against

usa +↑−↑tax

bitch −↓thousands +↑

+↓hahaha+↓good

miss −↓+↓my

not −↓ass −↓dont −↓can’t −↓today +↑

−↑threat+↓birthday

damn −↓−↑demanding

+↓lifepromise +↑helped +↑die −↓last −↓ill −↓dead −↓

Per word average happiness shif t δh av g, r (%)

Word

rankr

T r e f : 2013-02-17 (havg=5.95)Tcom p: cl imate (havg=5.97)

Tex t s i z e :T r e f T c omp

+↓ +↑

−↑ −↓

Balanc e :

−3288 : +3388

−100 0 100

100

101

102

103

104

!ri=1δ hav g, i

Figure 10. Left: Happiness time series plot for unfilteredtweets (red dashed) and tweets containing the word “climate”(blue solid) one week before and one week after the Forwardon Climate Rally. Right: word shift plot for climate tweetsversus unfiltered tweets on the day of the rally.

CONCLUSION

We have provided a general exploration of the senti-ment surrounding tweets containing the word “climate”in response to natural disasters and climate change newsand events. The general public is becoming more likelyto use social media as an information source, and dis-cussion on Twitter is becoming more commonplace. Wefind that tweets containing the word “climate” are lesshappy than all tweets. In the United States, climatechange is a topic that is heavily politicized; the words“deny”, “denial”, and “deniers” are used more often intweets containing the word “climate”. The words thatappear in our climate-related tweets word shift suggestthat the discussion surrounding climate change is dom-inated by climate change activists rather than climatechange deniers, indicating that the twittersphere largelyagrees with the scientific consensus on this issue. Thepresence of the words “science” and “scientists” in al-most every word shift in this analysis also strengthensthis finding (see also [1]). The decreased “denial” of cli-mate change is evidence for how a democratization ofknowledge transfer through mass media can circumvent

the influence of large stakeholders on public opinion.In examining tweets on specific dates, we have deter-

mined that climate change news is abundant on Twitter.Events such as the release of a book, the winner of agreen ideas contest, or a plea to a political figure canproduce an increase in sentiment for tweets discussingclimate change. For example, the Forward on ClimateRally demonstrates a day when the happiness of climateconversation peaked above the background conversation.On the other hand, consequences of climate change suchas threats to certain species, extreme weather events, andclimate related legislative bills can cause a decrease inoverall happiness of the climate conversation on Twitterdue to an increase in the use words such as “threat”,“crisis”, and “battle”.

Natural disasters are more commonly discussed withinclimate-related tweets than unfiltered tweets, implyingthat some Twitter users associate climate change withthe increase in severity and frequency of certain naturaldisasters [14, 19, 29]. During Hurricane Irene, for exam-ple, the word “threat” was used much more often withinclimate tweets, suggesting that climate change may beperceived as a bigger threat than the hurricane itself.The analysis of Hurricane Sandy in Fig. 8 demonstratesthat while climate conversation peaked during HurricaneSandy, it persisted longer than the conversation aboutthe hurricane itself.

While climate change news is prevalent in traditionalmedia, our research provides an overall analysis of cli-mate change discussion on the social media site, Twitter.Through social media, the general public can learn aboutcurrent events and display their own opinions aboutglobal issues such as climate change. Twitter may be auseful asset in the ongoing battle against anthropogenicclimate change, as well as a useful research source forsocial scientists, an unsolicited public opinion tool forpolicy makers, and public engagement channel for scien-tists.

ACKNOWLEDGMENTS

The authors are grateful for the computational re-sources provided by the Vermont Advanced ComputingCore which is supported by NASA (NNX 08A096G),and the Vermont Complex Systems Center. CMD, AJR,EMC, and LM were supported by National Science Foun-dation (NSF) grant DMS-0940271 to the Mathematics &Climate Research Network. PSD was supported by NSFCAREER Award #0846668. Thanks to Mary Lou Zee-man for helpful discussions.

[1] X. An, A. R. Ganguly, Y. Fang, S. B. Scyphers, A. M.Hunter, and J. G. Dy. Tracking climate change opinionsfrom Twitter data. Workshop on Data Science for SocialGood, 2014.

[2] W. R. Anderegg, J. W. Prall, J. Harold, and S. H. Schnei-der. Expert credibility in climate change. Proceedings ofthe National Academy of Sciences, 107(27):12107–12109,2010.

10

[3] E. Aramaki, S. Maskawa, and M. Morita. Twitter catchesthe flu: detecting influenza epidemics using Twitter. InProceedings of the conference on empirical methods innatural language processing, pages 1568–1576. Associa-tion for Computational Linguistics, 2011.

[4] C. A. Bliss, I. M. Kloumann, K. D. Harris, C. M. Dan-forth, and P. S. Dodds. Twitter reciprocal reply networksexhibit assortativity with respect to happiness. Journalof Computational Science, 3(5):388–397, 2012.

[5] M. T. Boykoff. Who speaks for the climate?: Makingsense of media reporting on climate change. CambridgeUniversity Press, 2011.

[6] M. T. Boykoff and J. M. Boykoff. Climate change andjournalistic norms: A case-study of US mass-media cov-erage. Geoforum, 38(6):1190–1204, 2007.

[7] N. Cao, Y.-R. Lin, X. Sun, D. Lazer, S. Liu, and H. Qu.Whisper: Tracing the spatiotemporal process of infor-mation diffusion in real time. Visualization and Com-puter Graphics, IEEE Transactions on, 18(12):2649–2658, 2012.

[8] D. Carli. Twitter, 2008 (accessed March 19, 2015).https://twitter.com/dcarli/status/953288121.

[9] J. W. Dearing and E. M. Rogers. Agenda-setting, vol-ume 6. Sage Publications, 1996.

[10] C. Doctorow. Twitter, 2009 (accessed March 19, 2015).https://twitter.com/doctorow/status/1482803994.

[11] P. S. Dodds, K. D. Harris, I. M. Kloumann, C. A. Bliss,and C. M. Danforth. Temporal patterns of happiness andinformation in a global social network: Hedonometricsand Twitter. PLoS ONE, 6(12):e26752, 2011.

[12] P. T. Doran and M. K. Zimmerman. Examining the sci-entific consensus on climate change. Eos, TransactionsAmerican Geophysical Union, 90(3):22–23, 2009.

[13] C. Field and M. Van Aalst. Climate change 2014: Im-pacts, adaptation, and vulnerability, volume 1. IPCC,2014.

[14] C. B. Field. Managing the risks of extreme events anddisasters to advance climate change adaptation: specialreport of the intergovernmental panel on climate change.Cambridge University Press, 2012.

[15] E. M. Fischer and R. Knutti. Anthropogenic contribu-tion to global occurrence of heavy-precipitation and high-temperature extremes. Nature Climate Change, advanceonline publication, April 2015.

[16] O. V. Gardens. Twitter, 2011 (accessed March19, 2015). https://twitter.com/openviewgardens/

status/99975488293978112.[17] L. C. Hamilton and M. D. Stampone. Blowin’ in the

wind: Short-term weather and belief in anthropogenicclimate change. Weather, Climate, and Society, 5(2):112–119, 2013.

[18] P. D. Howe, M. Mildenberger, J. R. Marlon, and A. Leis-erowitz. Geographic variation in opinions on climatechange at state and local scales in the usa. Nature Cli-mate Change, 2015.

[19] D. G. Huber and J. Gulledge. Extreme weather and cli-mate change: Understanding the link, managing the risk.Pew Center on Global Climate Change Arlington, 2011.

[20] A. P. Kirilenko, T. Molodtsova, and S. O. Stepchenkova.People as sensors: Mass media and local temperatureinfluence climate change discussion on Twitter. GlobalEnvironmental Change, 30:92–100, 2015.

[21] I. M. Kloumann, C. M. Danforth, K. D. Harris, C. A.Bliss, and P. S. Dodds. Positivity of the English language.

PLoS ONE, 7(1):e29484, 2012.[22] Y. Kryvasheyeu, H. Chen, N. Obradovich, E. Moro,

P. Van Hentenryck, J. Fowler, and M. Serbian. Now-casting disaster damage. Submitted, April 2015.

[23] H. Kwak, C. Lee, H. Park, and S. Moon. What is Twitter,a social network or a news media? In Proceedings of the19th international conference on World wide web, pages591–600. ACM, 2010.

[24] A. Leiserowitz, E. Maibach, C. Roser-Renouf, G. Fein-berg, and P. Howe. Climate change in the american mind:Americans’ global warming beliefs and attitudes in April,2013. Yale University and George Mason University. NewHaven, CT: Yale Project on Climate Change Communi-cation, 2013.

[25] Y. Li, E. J. Johnson, and L. Zaval. Local warming dailytemperature change influences belief in global warming.Psychological Science, 2011.

[26] Y.-R. Lin, B. Keegan, D. Margolin, and D. Lazer. Ris-ing tides or rising stars?: Dynamics of shared attentionon Twitter during media events. PloS one, 9(5):e94093,2014.

[27] Y.-R. Lin, D. Margolin, B. Keegan, and D. Lazer. Voicesof victory: A computational focus group framework fortracking opinion shift in real time. In Proceedings ofthe 22nd international conference on World Wide Web,pages 737–748. International World Wide Web Confer-ences Steering Committee, 2013.

[28] D. MacKay. Sustainable Energy-without the hot air. UITCambridge, 2008.

[29] M. E. Mann and K. A. Emanuel. Atlantic hurricanetrends linked to climate change. Eos, Transactions Amer-ican Geophysical Union, 87(24):233–241, 2006.

[30] L. Mitchell, M. R. Frank, K. D. Harris, P. S. Dodds,and C. M. Danforth. The geography of happiness: Con-necting Twitter sentiment and expression, demograph-ics, and objective characteristics of place. PLoS ONE,8(5):e64417, 2013.

[31] NewGreenStuff. Twitter, 2008 (accessed March 19,2015). https://twitter.com/NewGreenStuff/status/

953099924.[32] H. News. Twitter, 2010 (accessed March 19,

2015). https://twitter.com/HumanityNews/status/

11612292989.[33] O. News. Twitter, 2008 (accessed March 19,

2015). https://twitter.com/OneWorld_News/status/

1083004712.[34] O. News. Twitter, 2008 (accessed March 19,

2015). https://twitter.com/OneWorld_News/status/

953758970.[35] NewsOnGreen. Twitter, 2010 (accessed March 19,

2015). https://twitter.com/NewsOnGreen/status/

11608867076.[36] M. J. Paul and M. Dredze. You are what you tweet:

Analyzing Twitter for public health. In ICWSM, pages265–272, 2011.

[37] T. Q. Phan and E. M. Airoldi. A natural experiment ofsocial network formation and dynamics. Proceedings ofthe National Academy of Sciences, 2015.

[38] J. T. Ripberger, H. C. Jenkins-Smith, C. L. Silva, D. E.Carlson, and M. Henderson. Social media and severeweather: Do tweets provide a valid indicator of public at-tention to severe weather risk communication? Weather,Climate, and Society, 6(4):520–530, 2014.

11

[39] T. Sakaki, M. Okazaki, and Y. Matsuo. Earthquakeshakes Twitter users: real-time event detection by socialsensors. In Proceedings of the 19th international confer-ence on World wide web, pages 851–860. ACM, 2010.

[40] S. Solutions. Twitter, 2009 (accessed March 19,2015). https://twitter.com/ShiftSolutions/status/

1485975759.[41] M. Twain, M. B. Frank, R. P. Browning, L. Salamo,

F. Anderson, and M. Twain. Mark Twain’s Notebooks& Journals, Volume III:(1883-1891), volume 8. Univ ofCalifornia Press, 1980.

[42] K. M. Wilson. Mass media as sources of global warmingknowledge. Mass Comm Review, 22:75–89, 1995.

[43] WWF. Twitter, 2012 (accessed March 19, 2015). https://twitter.com/WWF/status/196902312797671424.

[44] L. Zaval, E. A. Keenan, E. J. Johnson, and E. U. Weber.How warm days increase belief in global warming. NatureClimate Change, 4(2):143–147, 2014.