Analyzed the data: PSD CMD. Wrote the paper: PSD. Data collection and curation: CMD KDH IMK CAB PSD.
The authors have declared that no competing interests exist.
Individual happiness is a fundamental societal metric. Normally measured through self-report, happiness has often been indirectly characterized and overshadowed by more readily quantifiable economic indicators such as gross domestic product. Here, we examine expressions made on the online, global microblog and social networking service Twitter, uncovering and explaining temporal variations in happiness and information levels over timescales ranging from hours to years. Our data set comprises over 46 billion words contained in nearly 4.6 billion expressions posted over a 33 month span by over 63 million unique users. In measuring happiness, we construct a tunable, real-time, remote-sensing, and non-invasive, text-based hedonometer. In building our metric, made available with this paper, we conducted a survey to obtain happiness evaluations of over 10,000 individual words, representing a tenfold size improvement over similar existing word sets. Rather than being ad hoc, our word list is chosen solely by frequency of usage, and we show how a highly robust and tunable metric can be constructed and defended.
One of the great modern scientific challenges we face lies in understanding macroscale sociotechnical phenomena–i.e., the behavior of decentralized, networked systems inextricably involving people, information, and machine algorithms–such as global economic crashes and the spreading of ideas and beliefs
While there are undoubtedly limits to that which may eventually be quantified regarding human behavior, recent studies have demonstrated a number of successful and diverse methodologies, all impossible (if imaginable) prior to the Internet age. Three examples relevant to public health, markets, entertainment, history, evolution of language and culture, and prediction are (1) Google's digitization of over 15 million books and an initial analysis of the last two hundred years, showing language usage changes, censorship, dynamics of fame, and time compression of collective memory
Out of the many possibilities in the ‘Big Data’ age of social sciences, we focus here on measuring, describing, and understanding the well-being of large populations. A measure of ‘societal happiness’ is a crucial adjunct to traditional economic measures such as gross domestic product and is of fundamental scientific interest in its own right
Our overall objective is to use web-scale text analysis to remotely sense societal-scale levels of happiness using the singular source of the microblog and social networking service Twitter.
Our contributions are both methodological and observational. First, our method for measuring the happiness of a given text, which we introduced in
Second, using Twitter as a data source, we are able to explore happiness as a function of time, space, demographics, and network structure, with time being our focus here. Twitter is extremely simple in nature, allowing users to place brief, text-only expressions online–‘status updates’ or ‘tweets’–that are no more than 140 characters in length. As we will show, Twitter's framing tends to yield in-the-moment expressions that reflect users' current experiences, making the service an ideal candidate input signal for a real time societal ‘hedonometer’
There is an important psychological distinction between an individual's current, experiential happiness
We further focus our present work on our essential findings regarding temporal variations in happiness including: the overall time series; regular cycles at the scale of days and weeks; time series for subsets of tweets containing specific keywords; and detailed comparisons between texts at the level of individual words. We also compare happiness levels with measures of information content, which we show are, in general, uncorrelated quantities (see 7.2). For information, as we explain below, we employ an estimate of lexical size (or effective vocabulary size) which is related to species diversity for ecological populations and is derived from generalized entropy measures
Our methods and findings complement a number of related efforts undertaken in recent years regarding happiness and well-being including: large-scale surveys carried out by Gallup
We structure our paper as follows: in Sec. 1, we describe our data set; in Secs. 2 and 3, we detail our methods for measuring happiness and information content, demonstrating in particular the robustness of our hedonometer while uncovering some intriguing aspects of the English language's emotional content; in Sec. 4, we present and discuss the overall time series for happiness and information; in Secs. 5 and 6, we examine the average weekly and daily cycles in detail; in Sec. 7, we explore happiness and information time series for tweets containing keywords and short phrases; and in Sec. 8, we offer some concluding remarks.
Since its inception, Twitter has provided various kinds of dedicated data feeds for research purposes. For the results we present here, we collected tweets over a three year period running from September 9, 2008 to September 18, 2011. To the nearest million, our data set comprises 46.076 billion words contained in 4.586 billion tweets posted by over 63 million individual users. Up until November 6, 2010, our collection represents approximately 8% of all tweets posted to that point in time
Our rate of gathering tweets was not constant over time, with regions of stability connected by short periods of considerable fluctuations (shown later in detail). These changes were due to periodic alterations in Twitter's feed mechanism as the company adjusted to increasing demand on their service
Each tweet delivered by Twitter was accompanied by a basic set of informational attributes; we list the salient ones in
Tweet attributes: |
Tweet text |
Unique tweet ID |
Date and time tweet was posted |
UTC offset (from GMT) |
User's location |
User ID |
Date and time user's account was created |
User's current follower count |
User's current friends count |
User's total number of tweets |
In-reply-to tweet ID |
In-reply-to user ID |
Retweet (Y/N) |
Information regarding the time of posting was altered (
User location is available for some tweets in the form of either current latitude and longitude, as reported for example by a smartphone, or a static, free text entry of a home city along with state and country. For measures of social interactions, we have a user's current follower and friend counts (but no information on who the followers and friends are), and if a tweet is made in reply to another tweet, we also have the identifying number (ID) of the latter. Finally, a ‘retweet’ flag (‘RT’) indicates if a tweet is a rebroadcasting of another tweet, encoding an important kind of information spreading in the Twitter network.
Against the many benefits of using a data source such as Twitter, there are a number of reasonable concerns to be raised, notably representativeness. First, in terms of basic sampling, tweets allocated to data feeds by Twitter were effectively chosen at random from all tweets. Our observation of this apparent absence of bias in no way dismisses the far stronger issue that the full collection of tweets is a non-uniform subsampling of all utterances made by a non-representative subpopulation of all people
In sum, we see two main arguments for pursuing the massive data stream of Twitter: (1) the potential for describing universal human patterns, whether they be emotional, social, or otherwise; and (2) the current and growing importance of Twitter
A preliminary glance at the data set shows that the raw word content of tweets does appear to reflect people's current circumstances. For example,
For purposes of comparison, each curve is normalized so that the count fraction represents the fraction of times a word is mentioned in a given hour relative to a day. The numbers in parentheses indicate the relative overall abundance normalized for each set of words by the most common word. Data for these plots is drawn from approximately 26.5 billion words collected from May 21, 2009 to December 31, 2010 inclusive, with the time of day adjusted to local time by Twitter from the former date onwards. The words ‘food’ and ‘dinner’ appeared a total of 2,994,745 (0.011%) and 4,486,379 (0.016%) times respectively.
We use a simple, fast method for measuring the happiness of texts that hinges on two key components: (1) human evaluations of the happiness of a set of individual words, and (2) a naive algorithm for scaling up from individual words to texts. We substantially improve here on the method introduced by two of the present authors in
For a single text, we would naturally rank the
For human evaluations of happiness, we used Amazon's Mechanical Turk
We drew on four disparate text sources: Twitter, Google Books (English)
By simply employing frequency as the measure of a word's importance, we naturally achieve a number of goals: (1) Precision: we have evaluations for as many words in a text as possible, given cost restrictions (the number of unique ‘words’ being tens of millions); (2) Relevance: we tailor our instrument to our focus of study; and (3) Impartiality: we do not a priori decide if a given word has emotional or meaningful content. Our word set consequently involves multiple languages, all parts of speech, plurals, conjugations of verbs, slang, abbreviations, and emotionless, or neutral, words such as ‘the’ and ‘of’.
For the evaluations, we asked users on Mechanical Turk to rate how a given word made them feel on a nine point integer scale, obtaining 50 independent evaluations per word. We broke the overall assignment into 100 smaller tasks of rating approximately 100 randomly assigned words at a time. We emphasized the scores 1, 3, 5, 7, and 9 by stylized faces, representing a sad to happy spectrum. Such five point scales are in widespread use on the web today (e.g., Amazon) and would likely be familiar with users. The four intermediate scores of 2, 4, 6, 8 allowed for fine tuning of assessments. In using this scheme, we remained consistent with the 1999 Affective Norms for English Words (ANEW) study by Bradley and Lang
Some illustrative examples of average happiness we obtained for individual words are:
Note that in analysing texts, we avoid stemming words, i.e., conflating inflected words with their root form, such as all conjugations of a specific verb. For verbs in particular, by focusing on the most frequent words, we obtained scores for those conjugations likely to appear in texts, obviating any need for stemming. Moreover, while we observe stemming works well in some cases for happiness measures, e.g.,
In the Supplementary Information, we provide happiness averages and standard deviations for all 10,222 words, along with other information.
An immediate and reassuring sign of the robustness of the word happiness scores we obtained via Mechanical Turk is that our results agree very well with that of the earlier ANEW study which consisted of 1034 words
The ANEW study words were also broadly chosen for their emotional and meaningful import rather than usage frequency, and we show below that our larger frequency-based word set affords a much greater coverage of texts. (By coverage, we mean the percentage of words in a text for which we have individual happiness estimates.) Note that in the ANEW study and our earlier work
We now show that our hedonometer can be improved by considering the effects of taking subsets of the overall list of 10,222 words. Clearly, truly neutral words such as ‘the’ and ‘of’ should be omitted, especially because of their high relative abundance, thereby forming a list of excluded words commonly referred to as stop words
Because we have filtered by frequency in selecting our word list, we are able to determine stop word lists in a principled way, leading to a feature of tunability. Here, we exclude words whose average happiness
We explore and demonstrate our hedonometer's behavior [Eq. (1)] with respect to different stop word lists by varying
To measure the happiness of a given text, we first compute frequencies of all words; we then create an overall happiness score, Eq. (1) , as a weighted average of subsets of 10,222 individual word happiness assessments on a 1 to 9 scale, obtained through Mechanical Turk (see main text and
As a test case and as shown in
We quantify the similarity between time series by computing Pearson's correlation coefficient for each pair of time series with
The striking congruence for all time series generated with
For the purposes of this paper, it is most useful if we choose a specific value of
We support the robustness of our choice with evidence provided in
As a final testament to the quality of our hedonometer, we note that in an earlier version of the present paper
We address several key aspects and limitations of our measurement. First, as with any sentiment analysis technique, our instrument is fallible for smaller texts, especially at the scale of a typical sentence, where ambiguity may render even human readers unable to judge meaning or tone
Second, we are also effectively extracting a happiness level as perceived by a generic reader who sees only word frequency. Indeed, our method is purposefully more simplistic than traditional natural language processing (NLP) algorithms which attempt to infer meaning (e.g., OpinionFinder
Third, we quantify only how people appear to others; as should be obvious, our method cannot divine the internal emotional states of specific individuals or populations. In attempting to truly understand a social system's potential dynamical evolution, we would have to account for publicly hidden but accessible internal ranges and states of emotions, beliefs, etc. However, a person's exhibited emotional tone, now increasingly filtered through the signal-limiting medium of written interactions (e.g., status updates, emails, and text messages), is that which other people evidently observe and react to.
Last, by using a simple kind of text analysis, we are able to non-invasively, remotely sense the exhibited happiness of very large numbers of people via their written, open, web-scale output. Crucially, we do not ask people how happy they are, we merely observe how they behave online. As such, we avoid the many difficulties associated with self-report
In quantifying a text's information content, we use concepts traditionally employed for estimating species diversity in ecological studies
A first observation is that the sheer number of distinct words in a text is not a good representation of lexical size. Because natural texts generally exhibit highly skewed distributions of word frequencies, such a measure discards much salient information, and moreover is difficult to estimate if a text is subsampled.
To arrive at a more useful and meaningful quantity, we consider generalized entropy:
These generalized entropies can be seen as direct measures of information but their values can be hard to immediately interpret. To make comparisons between the information content of texts more understandable, if by adding an extra step, we use these information measures to compute an equivalent lexical size,
We observe that the lexical sizes
Using
We observe a variety of temporal trends in happiness and information content across timescales of hours, days, months, and years. In
Looking at the complete time series, we see that after a gradual upward trend that ran from January to April, 2009, the overall time series has shown a gradual downward trend, accelerating somewhat over the first half of 2011. We also see that average happiness gradually increased over the last months of 2008, 2009, and 2010, and dropped in January of the ensuing years. Moving down to timescales less than a month, we see a clear weekly signal with the peak generally occurring over the weekend, and the nadir on Monday and Tuesday (c.f.,
At the scale of a day, we find a number of dates which strongly deviate in their happiness levels from nearby dates, and we indicate these in
In the following section, we look more closely at several dates, showing how individual words contribute to their anomalous measurements.
For the outlying happy dates, in 2008, 2009, and 2010, Christmas Day returned the highest levels of happiness, followed by Christmas Eve. Other relatively positive dates include New Year's Eve and Day, Valentine's Day, Thanksgiving, Fourth of July, Easter Sunday, Mother's Day, and Father's Day. All of these observations are sensible, and reflect a strong (though not universal) degree of social synchrony. The spikes for Thanksgiving and Fourth of July reflects the fact that while Twitter is a global service, the majority of users still come from the United States
Over the entire time span, we see substantial, system-wide drops in happiness in response to a range of disparate events, both exogenous and endogenous in nature. Working from the start of our time series, we first see the Bailout of the U.S. financial system, which induced a multi-week depression in our time series. The lowest point corresponds to Monday, September 29, 2008, when the U.S. government agreed to an unprecedented purchase of toxic assets in the form of mortgage backed securities.
Following the 2008 Bailout, we see the overall time series rebound well through the end of 2008, suffer the usual post New Year's dip, and begin to rise again until an extraordinary week long drop due to the onset of the 2009 swine flu or H1N1 pandemic.
The next decline occurred with Michael Jackson's death, the largest single day drop we observed. His memorial on July 7, 2009 induced another clear negative signal. The death of actor Patrick Swayze on September 14, 2009 also left a discernible negative impact on the time series. In between, Twitter itself was the victim of a large-scale distributed denial of service attack, leading to an outage of the service; upon resumption, tweets were noticeably focused on this internal story.
Several natural disasters registered as days with relatively low happiness: the February, 2010 Chilean earthquake, the October, 2010 record size storm complex across the U.S., and the March, 2011 earthquake and tsunami which devastated Japan.
Reports of the killing of Osama Bin Laden on May 2, 2011 resulted in the day of the lowest happiness across the entire time frame. And global sport left one identifiable drop: the 4–1 victory of Germany over England in the 2010 Football World Cup. Spain's ultimate victory in the tournament was detectable in terms of word usage but did not lead to a significant change in overall happiness.
One arguably false finding of a cultural event being negative was the finale of the last season of the highly rated television show ‘Lost’, marked by a drop in our time series on May 24, 2010, and in part due to the word ‘lost’ having a low happiness score of
A number of these departures for specific dates qualitatively match observations we made in our earlier work on blogs
When comparing two or more texts using a single summary statistic, as we have here with average happiness, we naturally need to look further into why a given measure shows variation. In
Words are ranked by their percentage contribution to the change in average happiness,
Consider two texts
Whether or not the
Whether or not the
We will signify a word's happiness relative to text
+↑: Increased usage of relatively positive words–If a word is happier than text
−↓: Decreased usage of relatively negative words–If a word is less happy than text
+↑: Decreased usage of relatively positive words–If a word is happier than text
+↓: Increased usage of relatively negative words–If a word is less happy than text
For the convenience of visualization, we normalize the summands in Eq. (2) and convert to percentages to obtain:
Finally, in comparing two texts, we rank words by their absolute contribution to the change in average happiness,
With these definitions in hand, we return to
The primary element of our word shift graphs is a central bar graph showing a desired number of highest ranked labMT 1.0 words (
In each graph in
By contrast, we see the happiness spike of the Royal Wedding is due to higher prevalence of positive words such as ‘wedding’, ‘beautiful’, ‘kiss’, ‘prince’, ‘princess’, ‘dress’, and ‘gorgeous’ (all
Beyond these dominant stories, our word shifts allow us to make a number of supporting and clarifying observations. First, since we have chosen to compare specific dates to the surrounding 14 days, nearby anomalous events appear in each word shift. For example, the Royal Wedding (2011/4/29) has less ‘Easter’ and ‘chocolate’ because Easter occurred five days earlier and less ‘dead’ and ‘killed’ because of Bin Laden's death three days later (2011/5/2). The Bin Laden graph in turn shows less ‘wedding’, ‘happy’, and ‘Mother's’ (due to the Royal Wedding and Mother's Day, 2011/5/8). Other reference texts can be readily constructed for comparisons (e.g., tweets on all days or matching weekdays). However, we find that the main words contributing to word shifts reliably appear as we consider alternative, reasonable reference texts.
Second, in all text comparisons, we find some words go against the main trend. For example, we see more ‘money’, ‘weekend’, and ‘billion’ (all
The three insets in the word shift graphs of
Second, on the bottom left of each word shift graph, the inset line graph shows the cumulative sum of the individual word contributions,
The third and final inset on the bottom right is a key one. An increase in happiness may be due to the use of more positive words, an avoidance of negative words, or a combination of both, and we need to quantify this in a simple way. The inset's four circles show the relative total contributions of the four classes of words to the overall shift in average happiness. For example, the area of the top right (yellow) circle represents the sum of all contributions due to relatively positive words that increase in frequency in
The two numbers above the circles give the total percentage change toward and away from the reference text's average happiness. For the Bailout example, there is a drop in happiness of −165% of
For the Bailout and Bin Laden graphs, we see similar overall patterns: the more frequent use of negative words (
To complete our analysis of the overall time series, we turn to information content (
By examining shifts in word usage, we are able to attribute the more than doubling of
As we saw in
To make the average weekly cycle more clear, we repeat the pattern for a second week. The crosses indicate happiness scores based on all data, while the filled circles show the results of removing the outlier days indicated in
We take the reasonable step of focusing on the data with outlier days removed. We see Saturday has the highest average happiness (
While the weekend peak in the cycle conforms with everyday intuition, the minimum on Tuesday goes against standard notions of the Monday blues with its back-to-work nature, and Wednesday's middle-of-the-week labeling as the work week's hump day
In our earlier work on blogs using the ANEW study word list
With any observed pattern, a fundamental issue is universality. Is the three day midweek low followed by a peak around Saturday a pattern we always see, given enough data? Further inspection of our Twitter data set shows a constancy in the weekly cycle occurring over time. In
As per
In
Each day of the week's word frequency distribution was generated by averaging normalized distributions for each instance of that week day in May 21, 2009 to December 31, 2010, with outlier dates removed. See
The insets of
The bottom right inset shows that the overall positive shift from Tuesdays to Saturdays is due to the more frequent use of positive words (
The average Simpson lexical size
We compute
To see further into these changes between days, we can generate word shift graphs for Simpson lexical size
Using Eq. (5) , we find Friday's larger value of
We next examine how average happiness levels change throughout the day at the resolution of an hour. As shown in
As for days of the week in
We also find that usage rates of the most common profanities are remarkably similar and are roughly anticorrelated with the observed happiness cycle.
To give a deeper sense of the underlying moods reflected in the low and high of the day, we explore the word shift graph in
Days given equal weighting with outlier dates removed. (See
The balance plot (bottom right inset) shows that 5 to 6 am is happier because of an overall preponderance of less abundant negative words and more abundant positive words, the former's contribution marginally larger than the latter. As the lower left inset cumulative plot shows, the first 50 words account for approximately 70% of the total shift. Thereafter, word shifts gradually bring the overall difference up to 100%, requiring all words to do so.
The first few salient, relatively positive words more abundant between 5 and 6 am (
In
See also
Finally, we find that using alternate averaging schemes to create word frequency distributions for hour of the day yields remarkably little variation in
We turn to our last area of focus: temporal happiness patterns for tweets containing specific text elements. We need not restrict ourselves to words, considering also, for example, short phrases (
To facilitate comparisons, we now measure what we call ‘normalized happiness’
In
We next examine a selection of 100 handpicked keywords and text elements. As mentioned above, the ambient average happiness for tweets containing many of these terms are mostly stable over time, and in
Word |
|
Total Tweets |
|
Word |
|
Total Tweets |
|
1. happy | +0.430 | 1.65e+07 (13) | +1.104 (1) | 51. snow | −0.051 | 2.60e+06 (49) | +0.083 (39) |
2. Christmas | +0.404 | 4.89e+06 (35) | +0.953 (3) | 52. Jon Stewart | −0.052 | 5.21e+04 (97) | −0.024 (48) |
3. vegan | +0.315 | 1.84e+05 (90) | −0.015 (46) | 53. school | −0.056 | 9.26e+06 (24) | +0.050 (42) |
4. :) | +0.274 | 1.04e+07 (20) | +0.630 (12) | 54. Lehman Brothers | −0.078 | 8.50e+03 (100) | −0.721 (79) |
5. family | +0.251 | 5.01e+06 (32) | +0.716 (7) | 55. them | −0.090 | 1.54e+07 (15) | −0.280 (60) |
6. :-) | +0.228 | 1.67e+06 (60) | +0.560 (16) | 56. right | −0.090 | 1.92e+07 (10) | +0.126 (35) |
7. our | +0.207 | 1.41e+07 (16) | +0.159 (33) | 57. woman | −0.115 | 2.54e+06 (51) | +0.202 (30) |
8. win | +0.204 | 7.98e+06 (26) | +0.924 (4) | 58. left | −0.118 | 4.89e+06 (34) | −0.383 (63) |
9. vacation | +0.200 | 9.35e+05 (67) | +0.817 (5) | 59. me | −0.119 | 1.44e+08 (4) | +0.160 (32) |
10. party | +0.170 | 6.44e+06 (29) | +0.679 (9) | 60. election | −0.127 | 5.60e+05 (75) | −0.306 (61) |
11. love | +0.164 | 4.67e+07 (6) | +0.977 (2) | 61. Sarah Palin | −0.128 | 2.26e+05 (87) | −0.681 (76) |
12. friends | +0.155 | 7.67e+06 (27) | +0.685 (8) | 62. no | −0.132 | 9.51e+07 (5) | −1.415 (90) |
13. hope | +0.149 | 1.18e+07 (18) | +0.515 (19) | 63. rain | −0.134 | 3.23e+06 (41) | +0.050 (44) |
14. coffee | +0.147 | 2.80e+06 (46) | +0.518 (18) | 64. climate | −0.135 | 3.64e+05 (80) | −0.160 (51) |
15. cash | +0.146 | 1.28e+06 (63) | +0.601 (14) | 65. gay | −0.152 | 2.73e+06 (47) | −0.552 (72) |
16. sun | +0.144 | 2.39e+06 (52) | +0.737 (6) | 66. lose | −0.157 | 2.06e+06 (55) | −1.181 (86) |
17. income | +0.137 | 5.10e+05 (76) | +0.621 (13) | 67. they | −0.159 | 2.74e+07 (8) | −0.208 (58) |
18. summer | +0.135 | 3.00e+06 (43) | +0.221 (29) | 68. oil | −0.162 | 1.38e+06 (62) | −0.411 (65) |
19. church | +0.131 | 1.81e+06 (58) | −0.016 (47) | 69. cold | −0.162 | 3.67e+06 (36) | −0.546 (71) |
20. Valentine | +0.127 | 2.47e+05 (84) | +0.593 (15) | 70. I feel | −0.173 | 5.17e+06 (31) | −0.129 (50) |
21. Stephen Colbert | +0.126 | 2.38e+04 (99) | +0.001 (45) | 71. man | −0.175 | 1.59e+07 (14) | −0.163 (52) |
22. USA | +0.113 | 2.16e+06 (54) | +0.325 (26) | 72. Republican | −0.181 | 2.30e+05 (86) | −0.539 (70) |
23. ! | +0.106 | 3.44e+06 (40) | +0.195 (31) | 73. sad | −0.187 | 3.56e+06 (38) | −1.366 (89) |
24. winter | +0.101 | 1.26e+06 (64) | +0.050 (43) | 74. gas | −0.193 | 1.02e+06 (65) | −0.471 (67) |
25. God | +0.099 | 8.58e+06 (25) | +0.468 (20) | 75. economy | −0.203 | 6.09e+05 (73) | −0.525 (69) |
26. hot | +0.095 | 7.12e+06 (28) | −0.172 (54) | 76. Obama | −0.205 | 2.98e+06 (44) | −0.173 (55) |
27. ;) | +0.094 | 2.61e+06 (48) | +0.326 (25) | 77. Democrat | −0.226 | 9.32e+04 (93) | −0.384 (64) |
28. Jesus | +0.094 | 2.03e+06 (56) | +0.247 (28) | 78. Congress | −0.231 | 3.92e+05 (79) | −0.580 (74) |
29. today | +0.092 | 2.56e+07 (9) | +0.126 (36) | 79. hell | −0.250 | 6.27e+06 (30) | −1.551 (96) |
30. kiss | +0.072 | 1.70e+06 (59) | +0.632 (11) | 80. sick | −0.262 | 3.58e+06 (37) | −1.630 (97) |
31. yes | +0.056 | 1.16e+07 (19) | +0.321 (27) | 81. Muslim | −0.262 | 2.15e+05 (88) | −0.569 (73) |
32. tomorrow | +0.054 | 1.04e+07 (21) | +0.086 (38) | 82. war | −0.270 | 1.96e+06 (57) | −2.040 (100) |
33. you | +0.052 | 1.73e+08 (3) | +0.111 (37) | 83. Pope | −0.277 | 1.52e+05 (91) | −0.316 (62) |
34. heaven | +0.041 | 7.42e+05 (71) | +0.674 (10) | 84. hate | −0.282 | 9.65e+06 (23) | −1.520 (94) |
35. ;-) | +0.041 | 9.39e+05 (66) | +0.395 (23) | 85. Glenn Beck | −0.282 | 1.14e+05 (92) | −0.776 (82) |
36. we | +0.035 | 3.91e+07 (7) | +0.146 (34) | 86. Islam | −0.299 | 1.87e+05 (89) | −0.710 (78) |
37. yesterday | +0.033 | 3.08e+06 (42) | −0.168 (53) | 87. George Bush | −0.333 | 3.23e+04 (98) | −0.747 (80) |
38. dark | +0.031 | 1.58e+06 (61) | −0.766 (81) | 88. Goldman Sachs | −0.337 | 5.27e+04 (96) | −0.984 (84) |
39. ? | +0.030 | 2.32e+06 (53) | −0.503 (68) | 89. depressed | −0.339 | 2.81e+05 (82) | −1.541 (95) |
40. RT | +0.028 | 3.39e+08 (1) | −0.443 (66) | 90. Senate | −0.340 | 4.48e+05 (78) | −0.601 (75) |
41. Michael Jackson | +0.018 | 8.26e+05 (70) | −0.213 (59) | 91. BP | −0.355 | 5.82e+05 (74) | −0.902 (83) |
42. night | +0.014 | 1.71e+07 (12) | +0.074 (40) | 92. gun | −0.367 | 6.81e+05 (72) | −1.476 (93) |
43. life | +0.012 | 1.40e+07 (17) | +0.422 (22) | 93. drugs | −0.382 | 5.10e+05 (77) | −1.452 (91) |
44. health | −0.000 | 2.58e+06 (50) | +0.447 (21) | 94. headache | −0.437 | 8.57e+05 (69) | −1.881 (98) |
45. sex | −0.008 | 3.55e+06 (39) | +0.542 (17) | 95. :-( | −0.455 | 3.40e+05 (81) | −1.174 (85) |
46. work | −0.010 | 1.84e+07 (11) | −0.174 (56) | 96. :( | −0.472 | 2.89e+06 (45) | −1.288 (88) |
47. girl | −0.010 | 1.01e+07 (22) | +0.331 (24) | 97. Afghanistan | −0.703 | 2.74e+05 (83) | −1.458 (92) |
48. boy | −0.026 | 4.93e+06 (33) | +0.062 (41) | 98. mosque | −0.709 | 6.98e+04 (95) | −0.694 (77) |
49. I | −0.048 | 3.08e+08 (2) | −0.062 (49) | 99. flu | −0.735 | 9.01e+05 (68) | −1.912 (99) |
50. commute | −0.048 | 9.01e+04 (94) | −0.206 (57) | 100. Iraq | −0.773 | 2.39e+05 (85) | −1.282 (87) |
The number of tweets and the value of normalized happiness
Word |
|
Total Words | Frac top 50K | Word |
|
Total Words | Frac top 50K |
1. RT | 1019.5 | 4.751e+09 (1) | 0.653 (100) | 51. Iraq | 235.5 | 3.722e+06 (84) | 0.832 (68) |
2. ? | 662.1 | 2.608e+07 (58) | 0.731 (98) | 52. Jon Stewart | 234.9 | 7.053e+05 (97) | 0.836 (62) |
3. ! | 621.1 | 3.682e+07 (50) | 0.742 (97) | 53. Senate | 233.7 | 6.791e+06 (78) | 0.826 (71) |
4. USA | 501.5 | 3.150e+07 (54) | 0.751 (94) | 54. happy | 232.8 | 2.041e+08 (17) | 0.834 (65) |
5. no | 487.3 | 1.431e+09 (5) | 0.763 (93) | 55. climate | 231.7 | 5.245e+06 (80) | 0.813 (81) |
6. ;-) | 476.9 | 1.323e+07 (67) | 0.75 (95) | 56. yes | 230.0 | 1.484e+08 (21) | 0.846 (50) |
7. ;) | 389.2 | 3.379e+07 (52) | 0.791 (86) | 57. today | 225.3 | 3.802e+08 (9) | 0.883 (20) |
8. war | 386.2 | 2.901e+07 (56) | 0.785 (88) | 58. election | 220.7 | 8.632e+06 (75) | 0.847 (47) |
9. Goldman Sachs | 379.5 | 7.183e+05 (96) | 0.766 (92) | 59. summer | 219.1 | 4.471e+07 (42) | 0.864 (39) |
10. gay | 377.6 | 3.823e+07 (46) | 0.823 (77) | 60. Christmas | 215.7 | 6.330e+07 (35) | 0.862 (41) |
11. me | 368.4 | 2.136e+09 (4) | 0.829 (70) | 61. rain | 215.1 | 4.620e+07 (41) | 0.836 (61) |
12. :-) | 362.3 | 2.280e+07 (61) | 0.773 (91) | 62. girl | 214.0 | 1.513e+08 (20) | 0.873 (32) |
13. Islam | 355.2 | 2.776e+06 (89) | 0.678 (99) | 63. I feel | 214.0 | 7.141e+07 (34) | 0.901 (4) |
14. :) | 347.1 | 1.313e+08 (24) | 0.775 (90) | 64. kiss | 212.7 | 2.463e+07 (59) | 0.845 (51) |
15. Muslim | 343.9 | 3.327e+06 (86) | 0.779 (89) | 65. God | 211.6 | 1.298e+08 (25) | 0.884 (18) |
16. Michael Jackson | 335.0 | 1.029e+07 (71) | 0.803 (83) | 66. school | 211.2 | 1.328e+08 (23) | 0.88 (25) |
17. Obama | 325.8 | 4.412e+07 (43) | 0.825 (74) | 67. coffee | 209.1 | 3.926e+07 (45) | 0.878 (27) |
18. Lehman Brothers | 324.5 | 1.161e+05 (100) | 0.743 (96) | 68. Afghanistan | 208.8 | 3.898e+06 (83) | 0.793 (85) |
19. :-( | 312.5 | 4.798e+06 (81) | 0.804 (82) | 69. heaven | 208.3 | 1.075e+07 (69) | 0.864 (38) |
20. health | 312.4 | 3.817e+07 (47) | 0.826 (72) | 70. left | 207.8 | 8.017e+07 (31) | 0.873 (31) |
21. gas | 311.8 | 1.580e+07 (65) | 0.822 (78) | 71. family | 207.8 | 7.700e+07 (32) | 0.873 (30) |
22. Jesus | 311.4 | 3.011e+07 (55) | 0.831 (69) | 72. them | 205.1 | 2.672e+08 (12) | 0.893 (9) |
23. :( | 304.5 | 3.802e+07 (48) | 0.798 (84) | 73. sad | 203.6 | 5.482e+07 (36) | 0.886 (17) |
24. hot | 298.3 | 9.826e+07 (28) | 0.847 (46) | 74. night | 203.1 | 2.429e+08 (13) | 0.883 (21) |
25. cash | 298.0 | 1.909e+07 (63) | 0.832 (66) | 75. hell | 202.7 | 9.000e+07 (30) | 0.883 (19) |
26. vegan | 290.9 | 2.696e+06 (90) | 0.845 (54) | 76. mosque | 198.3 | 1.081e+06 (95) | 0.82 (80) |
27. George Bush | 288.0 | 4.546e+05 (98) | 0.847 (48) | 77. tomorrow | 198.1 | 1.516e+08 (19) | 0.892 (11) |
28. BP | 285.2 | 8.957e+06 (74) | 0.791 (87) | 78. friends | 197.5 | 1.242e+08 (27) | 0.886 (16) |
29. man | 283.3 | 2.333e+08 (15) | 0.845 (52) | 79. vacation | 197.1 | 1.341e+07 (66) | 0.876 (28) |
30. sex | 276.2 | 5.186e+07 (37) | 0.844 (57) | 80. snow | 195.6 | 3.698e+07 (49) | 0.881 (22) |
31. Sarah Palin | 275.4 | 3.194e+06 (87) | 0.842 (58) | 81. yesterday | 192.7 | 5.003e+07 (39) | 0.887 (14) |
32. we | 272.4 | 6.434e+08 (6) | 0.869 (34) | 82. right | 190.5 | 2.854e+08 (10) | 0.887 (15) |
33. flu | 270.8 | 1.279e+07 (68) | 0.826 (73) | 83. church | 189.1 | 2.668e+07 (57) | 0.879 (26) |
34. income | 270.7 | 7.681e+06 (76) | 0.835 (63) | 84. cold | 188.4 | 5.116e+07 (38) | 0.9 (5) |
35. I | 269.8 | 4.590e+09 (2) | 0.881 (23) | 85. lose | 187.2 | 3.335e+07 (53) | 0.881 (24) |
36. oil | 267.1 | 2.147e+07 (62) | 0.825 (75) | 86. sick | 186.6 | 4.985e+07 (40) | 0.899 (6) |
37. Democrat | 262.4 | 1.469e+06 (94) | 0.832 (67) | 87. economy | 186.5 | 9.512e+06 (73) | 0.847 (49) |
38. drugs | 261.7 | 7.633e+06 (77) | 0.862 (40) | 88. dark | 186.1 | 2.403e+07 (60) | 0.868 (36) |
39. our | 257.6 | 2.394e+08 (14) | 0.869 (35) | 89. Pope | 185.3 | 2.268e+06 (91) | 0.84 (59) |
40. boy | 256.7 | 7.174e+07 (33) | 0.857 (42) | 90. win | 185.1 | 1.261e+08 (26) | 0.825 (76) |
41. Glenn Beck | 252.3 | 1.740e+06 (92) | 0.851 (44) | 91. life | 180.4 | 2.210e+08 (16) | 0.892 (10) |
42. Stephen Colbert | 251.0 | 2.972e+05 (99) | 0.844 (55) | 92. woman | 178.8 | 4.151e+07 (44) | 0.874 (29) |
43. Valentine | 248.4 | 3.169e+06 (88) | 0.822 (79) | 93. work | 178.3 | 2.791e+08 (11) | 0.898 (7) |
44. party | 242.9 | 9.466e+07 (29) | 0.844 (56) | 94. depressed | 175.2 | 4.108e+06 (82) | 0.906 (2) |
45. gun | 241.9 | 1.030e+07 (70) | 0.836 (60) | 95. sun | 166.9 | 3.622e+07 (51) | 0.849 (45) |
46. winter | 240.2 | 1.871e+07 (64) | 0.854 (43) | 96. commute | 165.0 | 1.470e+06 (93) | 0.887 (13) |
47. Republican | 239.8 | 3.607e+06 (85) | 0.845 (53) | 97. hope | 157.2 | 1.853e+08 (18) | 0.89 (12) |
48. they | 239.8 | 4.749e+08 (8) | 0.896 (8) | 98. love | 149.9 | 6.409e+08 (7) | 0.865 (37) |
49. you | 239.2 | 2.484e+09 (3) | 0.871 (33) | 99. headache | 126.7 | 1.005e+07 (72) | 0.907 (1) |
50. Congress | 236.8 | 6.221e+06 (79) | 0.834 (64) | 100. hate | 106.5 | 1.382e+08 (22) | 0.902 (3) |
Keywords themselves are not included in the calculation of
We observe many interesting patterns and we invite the reader to explore the tables beyond the observations we record here. We begin with the highest and lowest rankings of ambient happiness
An important finding is that the average happiness of text elements as assessed through Mechanical Turk and their ambient happiness correlate very strongly (Spearman's correlation coefficient
We nevertheless find some scores move substantially when the text element's score is included; For example, ‘vegan’ ranks 3rd with
For financial terms, we see tweets mentioning the dissolved firm of ‘Lehmann Brothers’ and ‘Goldman Sachs’ are both negative (more so in the latter's case) while relatively high in lexical size (
Tweets referring to United States politics are below average in happiness with ‘Obama’, ‘Sarah Palin’, and ‘George Bush’ registering
Tweets involving the word ‘war’ rank high in information (
Generally, personal pronouns tell a positive prosocial story with ‘our’ and ‘you’ outranking ‘I’ and ‘me’ in happiness (
The ambient words in tweets containing ‘summer’ are slightly happier than those containing ‘winter’ but are less diverse:
Emoticons in increasing order of happiness are ‘:(’, ‘:-(’, ‘;-)’, ‘;)’, ‘:-)’, and ‘:)’ with
Tweets involving the ‘fake news’ comedian Stephen Colbert are both happier and of a higher information level than those concerning his senior colleague Jon Stewart (
As noted above, the exclamation point garners a positive ambient happiness (
A reflection on the preceding survey suggests that groups of related terms may possess positive, negative, or neutral correlation between happiness and information content. Overall, for our set of 100 keywords and text elements, we measure Spearman's correlation coefficient as
The two quantities show no correlation with Spearman's correlation coefficient measuring
In
Ambient happiness of a keyword is
The word shift graphs are for the time periods March and April, 2010 for ‘Pope’ and January and February, 2010 for ‘Israel.’ See
In
In
In
Our last example,
In analysing temporal patterns of happiness and information content for the very large data set generated by Twitter thus far, we have been able to uncover results ranging across many timescales and topics. The weekly and daily cycles in particular appear to be robust and suggestive of universal forms, accepting that the seven day week cycle is an historical and cultural artifact. With our greatly expanded word list as analysed using Mechanical Turk, labMT 1.0 (
An essential part of our comparative analyses is the word shift graph, which we have primarily used here for happiness. These provide us with a detailed view of why two texts differ based on changes in word frequency. These graphs, and their future iterations, should be of use in a range of fields where size distributions are compared through summary statistics (e.g., understanding how species diversity in ecological populations may differ as a result of changes in individual species abundances).
As we have described, the metadata accompanying Twitter messages contains more information than time stamps. Future research will naturally address (and go beyond) geographic variations, particularly for the United States; the change in expressions over time for individuals and the possibility of correlation or contagion of sentiment; effects of popularity as measured by follower count on users' expressions; and the possibility of fine-scale emotional synchronization between individuals based on directed messages
As we have seen in both the work of others and ours, Twitter and similar large-scale, online social networks have thus far provided good evidence that scientifically interesting and meaningful patterns can be extracted from these massive data sources of human behavior. The extent to which small-scale patterns can be elicited, e.g., for rare topics, also remains an open question, as does the true generalizability to the broader population. Whatever the case, Twitter is currently a substantial, growing element of the global media and is worth studying in its own right, just as a study of newspapers would seem entirely valid. And while current evidence suggests ‘instant polls’ created by remote-sensing text analysis methods are valid, and that these instruments complement and may in some cases improve upon traditional surveys, analysts will have to remain cognizant of the ever present problem of users gaming online expression systems to misinform.
Finally, the era of big data social sciences has undoubtedly begun. Rather than being transformed or revolutionized we feel the correct view is that the social sciences are expanding beyond a stable core to become data-abundant fields. In a data-abundant science, the challenge moves first to description and pattern finding, with explanation and experiments following. Instead of first forming hypotheses, we are forced to spend considerable time and effort simply describing. The approaches applicable for a data-scarce science still remain of the same value but new, vast windows into social and psychological behaviour are now open, and new tools are available and being developed to enable us to take in the view.
We defined a word as any contiguous set of characters bounded by white space and/or a small set of punctuation characters. We therefore included all misspellings, words from any language used on Twitter, hyperlinks, etc. All pattern matches we made were case-insensitive, and we did not perform stemming (e.g., ‘love’ and ‘loved’ were counted separately).
The data feed from Twitter was provided in XML and JSON formats
In measuring and comparing information content, a computational difficulty with the Twitter data set lies in accommodating the sheer number of distinct words. We found approximately 230 million unique words (including URLs) from a random sample of 25% of the tweets in our database. We determined that restricting our attention to a more manageable set of the first 50,000 most frequent words would be sufficient for highly accurate estimates of generalized entropy
Consequently, we recorded the frequencies for this specific set of 50,000 words at the level of hours and days. Note that we also always recorded the total number of words for any particular subset of tweets, so that our word probabilities were correctly normalized.
(TXT)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(TIFF)
(PDF)
The authors thank B. O'Connor for discussions and data sharing, as well as N. Allgaier, T. Gray, J. Harris, N. Johnson, P. Lessard, L. Mitchell, E. Pechenick, M. Pellon, A. Reece, N. Riktor, B. Tivnan, M. Tretin, and J. Williams.