Great! News! People! Fake! Donald’s Tweets: 18 January 2017 to 18 January 2018

Trump Simplest Words image Image via The Telegraph
Image via The Telegraph

In two days it will be a year since the inauguration of Twitter user ID 25073877.  Time flies when things are beyond ridiculous, right?

Some of you may remember I’ve published before other posts looking into various aspects of this user’s tweetage. I have already detailed the methodology I have followed (as well as its acknowledged limitations) on some of those previous posts. This has been a work in progress. See for example this, or this, or even this. There’s more if you follow the links.

Anyway, as the anniversary of the inauguration approaches I wanted to share with you, for what it’s worth, some quick numbers from a whole year’s worth of Twitter data.

The dataset I worked with for the purpose of this post is based on a larger Twitter archive I’ve been collecting and studying.

The dataset that I looked into in this occasion is composed by 2,587 tweets posted between 18/01/2018 08:49 AM EST (GMT -5) and 18/01/2017 06:53 AM EST (GMT-5).

As usual I did some basic text analysis, and some quick comparative quant stuff.

20 Most Tweeted Terms

Term Count
great 473
news 190
people 182
fake 166
thank 162
just 160
today 158
president 151
big 145
tax 140
trump 137
america 134
country 128
u.s 125
jobs 116
american 115
time 110
foxandfriends 98
media 98
new 97

 

Other Twitter Data Numeralia

Twitter Text Counts

Number of ! 1,261
Number of Characters (no spaces, including URLs and usernames) 275,964
Number of Pages (single space, 12pt) 109
Number of Words 50,176

Follower Growth

User followers as of  18/01/2018 08:49 46,815,170
User followers as of 18/01/2017 06:53 20,227,768
Gained followers in the period 26,587,402

Tweets About the Mexico Border Wall

id_str time (EST)
9.53979E+17 18/01/2018 08:16
9.53264E+17 16/01/2018 08:54
9.51229E+17 10/01/2018 18:07
9.50884E+17 09/01/2018 19:16
9.49066E+17 04/01/2018 18:53
9.46732E+17 29/12/2017 08:16
9.38391E+17 06/12/2017 07:53
9.20425E+17 17/10/2017 19:03
9.18063E+17 11/10/2017 06:36
9.08274E+17 14/09/2017 06:20
9.01803E+17 27/08/2017 09:44
8.97833E+17 16/08/2017 10:51
8.97045E+17 14/08/2017 06:38
8.85279E+17 12/07/2017 19:24
8.78014E+17 22/06/2017 18:15
8.56849E+17 25/04/2017 08:36
8.56485E+17 24/04/2017 08:28
8.56172E+17 23/04/2017 11:44
8.56171E+17 23/04/2017 11:42
8.30406E+17 11/02/2017 08:18
8.24617E+17 26/01/2017 08:55
8.24084E+17 24/01/2017 21:37
8.23147E+17 22/01/2017 07:35

[hydrate tweets using twarc]

The susual caveats apply. Numbers must be taken with a pinch of salt: the Twitter Search API is not a complete index of all Tweets, but instead an index of recent Tweets– my archive has collected Tweets every hour, which means, for instance, that Tweets that are promptly deleted in between collections do not get archived.

I have attempted refining the dataset, but duplicated Tweets might have stubbornly survived, which in turn logically would have affected the counts. However, in spite of these limitations, the data is indicative and potentially useful and/or interesting as documentation of current and recent historical events. For what it’s worth.

We’ve lived with this user’s tweets daily, and we are very much aware of the kind of discourse developed through the constant, reliably exasperating tweetage. So these basic numbers are most likely not to tell you anything you weren’t aware of already. A simile occurs to me: we are all aware of the daily, accumulative effects of stress, or, say, ageing, but sometimes it is only until we compare snapshots that we realise the true extent of its effects.

On UK Labour and Conservatives Tweet Sources

 

I‘ve been tracking the Twitter accounts of the UK Labour, Conservative, Green, and LibDem parties as we approach June the 8th (General Election). I am interested in what they are saying on Twitter through their official Twitter accounts, how they are saying it, how often and what apps they choose to do so.

Unfortunately there are still some duplicates in my Twitter data collection, but I can at least share at this point the sources used to tweet from the UK Labour and Conservatives Twitter accounts, as well as some indicative numbers, bearing in mind they may vary slightly, for tweets per source in a sample of 500 Tweets per account from 12/05/2017 to 01/06/2017 so far.

 

from_user
UKLabour
Source Count
MediaStudio 279
SproutSocial 106
TweetDeck 55
Twibbon 1
Twitter for Android 3
Twitter for iPhone 8
Twitter Web Client 48
500

 

from_user
Conservatives
Source Count
MediaStudio 25
TweetDeck 222
Twitter for iPhone 73
Twitter Web Client 180
500

Even bearing in my mind the sample of 500 tweets from each account may still contain some duplicates, the list of sources alone provides objective indication of each account’s social media management tool preferences. Something that stands out is that in comparison to, say, the realdonaldtrump account, none of these tweets were posted from Twitter Ads.

The source list indicates to me that UK Labour has attempted a more professional social media management strategy, with a reduced number of tweets from Android, iPhone and the Web Client, whereas the Conservatives have a majority of tweets coming from free & anyone-can-use apps, with no shortage of tweets coming from an iPhone (but no Android at all).

This short update is part of an ongoing lunchtime pet project for which I wish I had more time, but hey.  I also have data from the other political parties, but no time right now. Anyway, for what it’s worth, I thought I’d share.


N.B. Dear Guardian Data, in case you like what you see here and you ‘borrow’ the idea or any data… please kindly attribute and link back. It’s only polite to do so. Thank you!

Exeunt Android; Enter Ads: An Update on the Sources of Presidential Tweetage

 

A quick update as something I consider interesting has emerged from the ongoing archiving of the, er, current ‘Trumpian’ tweetage (see a previous post here). In case you do follow this blog you may be aware I’ve been keeping an eye on the ‘source’ of the Tweets, which is information (a metadata field) pertaining to each published Tweet which is made publicly visible by Twitter to anyone through certain applications like TweetDeck and directly through Twitter’s API (for Twitter’s ‘Field Guide’, see this).

Given the diversity of sources detected on the Tweets from the account under scrutiny in the past, hypotheses have been proposed suggesting correlations between type of content and source (application used to post Tweets); others have suggested that it is also indication of different people behind the account (though as we have said previously it is also possible that the same person tweets from different devices and applications).

Anyway, here’s some recent new insights emerging from the data since the last post:

  • Since Inauguration Day (20 January 2017), the last Tweet coming from Twitter for Android so far was timestamped 25/03/2017 10:41 (AM; DC time).  No Tweets from Android have been posted since that Tweet until the time of writing of this.
  • The last Tweet coming from the Twitter Web Client so far was timestamped 25/01/2017  19:03:33. No more Tweets with the Web Client as source have been posted (or collected by my archive) since then.
  • Since Inauguration Day, the Tweet timestamped 31/03/2017  14:30:38 was the first one to come from Twitter Ads. Since then 21 Tweets have been posted from Twitter Ads, the last one so far timestamped 17/05/2017  16:36:02.
  • During April and May 2017 Tweets have only come from Twitter for iPhone or Twitter Ads. The account in question has tweeted every sincle day throughout May until today 18 May 2017, a total of 90 Tweets so far (including a duplicated one in which a typo was corrected). Below a breakdown per source:

 

from_user Month Source Count
realDonaldTrump May Twitter for iPhone 81
Twitter Ads 9

 

As a keen Twitter user I personally find it interesting Twitter for Android has stopped being used by the account in question and that Twitter Ads has been used recently (instead?) in alternation with the Tweets from iPhone. Eyeballing the dataset quickly appears to indicate there might be a potential correlation between Tweets with links and official announcements (rather than statements/opinions) and Twitter Ads, but that requires looking into more closely and I will have to leave that for another time.


*Public note to self: I need to get rid of this habit of capitalising ‘Tweets’ as a noun… it becomes annoying.

Android vs iPhone: Source Counts and Trends in a Bit More than a Year’s Worth of Trumpian Tweetage

Last month I took a quick look at a month’s worth of Trumpian tweetage (user ID 25073877)  using text analysis. Using a similar methodology I have now prepared and shared a CSV file containing Tweet IDs and other metadata of 3,805 Tweets  from user ID 25073877 posted publicly between Thursday February 25  2016 16:35:12 +0000  to Monday April 03 2017 12:51:01 +0000. I deposited the file on figshare, including notes on motivation and methodology, here:

3805 Tweet IDs from User 25073877 [Thu Feb 25 16:35:12 +0000 2016 to Mon Apr 03 12:51:01 +0000 2017].

figshare. https://doi.org/10.6084/m9.figshare.4811284.v1

The dataset allows us count the sources for each Tweet (i.e. the application used to publish each Tweet according to the data provided by the Twitter Search API). The resulting counts are:

Source Tweet Count
Twitter for iPhone 1816
Twitter for Android 1672
Twitter Web Client 287
Twitter for iPad 22
Twitter Ads 3
Instagram 2
Media Studio 2
Periscope 1

As we have seen in previous posts, the account has alternated between iPhone and Android since the Inauguration. I wanted to look at relative trends throughout the dataset. Having prepared the main dataset I performed the text analysis of a document comprising the source listing arranged in chronological order according to the date and time of Tweet publication, and the listing corresponds to Tweets published between 25 February 2016 and Monday 3 April 2017. Using the Trends tool in Voyant, I divided the document in 25 segments, with the intention to roughly represent each monthly period covered in the listing and highlight source relative frequency trends in the period covered per segment.

The Trends tool shows a line graph depicting the distribution of a word’s occurrence across a corpus or document; in this case each word represents the source of a Tweet in the document. Each line in the graph is coloured according to the word it represents, at the top of the graph a legend displays which words are associated with which colours. I only included the most-used sources, leaving iPad there as reference.

The resulting graph looks like this:

Line Graph of Relative Frequencies of four most used sources by realdonaldtrump visualised in 25 segments of a document including Twe3,805 Tweets  from user ID 25073877 posted publicly between Thursday February 25  2016 16:35:12 +0000  to Monday April 03 2017 12:51:01 +0000. Data collected and analysed by Ernesto Priego. CC-BY. Chart made with Trends, Voyant Tools by Stéfan Sinclair & Geoffrey Rockwell (CC-BY 2017).
Line graph of the relative frequencies of the four most used sources visualised in 25 segments of a document including 3,805 Tweets from user ID 25073877 dated between Thursday February 25 2016 16:35:12 +0000 and Monday April 03 2017 12:51:01 +0000. Data collected and analysed by Ernesto Priego. CC-BY. Chart made with Trends, Voyant Tools by Stéfan Sinclair & Geoffrey Rockwell (CC-BY 2017).

I enjoyed this article by Christopher Ingraham (Washington Post Weblog, 3 April 2017), and I envy the access to the whole Trupian tweetage dataset, that would be essential to attempt to reproduce the analysis presented. The piece focuses on the use of exclamation marks (something I took an initial look at on my 6 February 2017 post), but it would be useful to take a closer look at any potential significant correlations between use of language in specific Tweets and the sources used to post those Tweets.

The article also has an embedded video titled ‘When it’s actually Trump tweeting, it’s way angrier’, repeating claims that there is a clear difference between those Tweets the account in question published from an iPhone and those published from an Android. I briefly referred to this issue on my 15 March 2017 post already, and I have not seen evidence yet that it is a staffer who actually posts from Twitter for iPhone from the account. I may be completely wrong, but I am still not convinced there is data-backed evidence to say for certain that Tweets from different sources are always tweeted by two or more different people, or that the differences in language per source are predictable and reliably attributable to a single specific person (the same people can after all tweet from the same account using different devices and applications, and indeed potentially. use different language/discourse/tone).  Anecdotal, I know, but I have noticed that sometimes my tweetage from the Android mobile app is different from my tweetage from TweetDeck on my Mac, but no regular patterns can be inferred there.

I do not necessarily doubt there is more than one person using the account, nor that the language used may vary significantly depending on the Tweets’ source.  What I’d like to see however is more robust studies demonstrating and highlighting correlations between language use in Tweets- texts and Tweets’ sources from the account in question taking into consideration that the same users can own different devices and use different language strategies depending on a series of contextual variables. Access to the source data of said studies should be consider essential for any assessment of any results or conclusions provided. Limitations and oppostion to more open sharing of Twitter data for research reproducibility are just one hurdle on the way for more scholarship in this area.

Android vs iPhone: Trends in a Month’s Worth of Trumpian Tweetage

What’s in a month’s worth of presidential tweetage?

I prepared a dataset containing a total of 123 public Tweets and corresponding metadata from user_id_str 25073877 between 15 February 2017 06:40:32 and 15 March 2017  08:14:20 Eastern Time (this figure does not factor in any tweets the user may have deleted shortly after publication). Of the 123 Tweets 68 were published from Android; 55 from iPhone. The whole text of the Tweets in the dataset accounts for 2,288 words, or 12,364 characters (no spaces; including URLs).

Using the Trends tools from Voyant Tools by Stéfan Sinclair & Geoffrey Rockwell I visualised the raw frequencies of the terms ‘Android’ and ‘iPhone’ in this dataset over 30 segments (more or less corresponding to the length of the month covered in the dataset) where each timestamped Tweet, sorted in chronological order, had its corresponding source indicated.

The result looked like this:

Raw frequency of Tweets per source in 30 segments by realdonaldtrump between 15 February 2017 06:40:32 and 15 March 2017 08:14:20 Eastern Time. Total: 123 Tweets: 68 from Android; 55 from iPhone. Data collected and analysed by Ernesto Priego. CC-BY. Chart made with Trends, Voyant Tools by Stéfan Sinclair & Geoffrey Rockwell (CC 2017).
Raw frequency of Tweets per source in 30 segments by realdonaldtrump between 15 February 2017 06:40:32 and 15 March 2017 08:14:20 Eastern Time. Total: 123 Tweets: 68 from Android; 55 from iPhone. Data collected and analysed by Ernesto Priego. CC-BY. Chart made with Trends, Voyant Tools by Stéfan Sinclair & Geoffrey Rockwell (CC 2017).

The chart does indeed reflect the higher number of Tweets from Android, and it also shows how over the whole document both sources are, in spite of more frequent absences from Tweets from iPhone, present throughout. The question as usual is what does this tell us. Back in 9 August 2016 David Robinson published an insightful analysis where he concludes that “he [Trump] writes only the (angrier) Android half”. With the source data I have gathered so far it would be possible (given the time and right circumstances) to perform a content analysis of Tweets per source, in order to confirm or reject any potential corelations between types of Tweets (re: tone, function, sentiment, time of day) and source used to post them.

Eyeballing the data, specifically since Inauguration Day until the present, does not seem to provide unambiguous evidence that the Tweets are undoubtedly written by two different persons (or more). What it is factual is that the Tweets do come from different sources (see my previous post), but at the moment, like with everything else this administration has been doing, my cursory analysis has only found conflicting insights, where for example a Tweet one would perhaps have expected to have been posted from iPhone (attributable hypothetically to a potentially less inflammable aide) was in fact posted from Android, and viceversa.

I may be wrong, but at the moment I cannot see any evidence there is any kind of predictable pattern, let alone strategy, behind the alternation between Android and iPhone (the only two type of sources used to publish Tweet from the account in question in the last month). Most of the times Tweets by source type will come in sequences of four or more Tweets, but sometimes a random lone Tweet from a different source will be sandwiched in between.

More confunsigly, all of the Tweets published between 08/03/2017 18:50 and 15/03/2017  08:14:20 have only had iPhone as source, without exception. Attention to detail is required to run robust statistical and content analyses that consider complete timestamps and further code the Tweet text and time data into more discrete categories, attempting a high level of granularity at both the temporal (time of publishing; ongoing documented events) and textual (content; discourse) levels. (If you are reading this and would like to take a look at the dataset, DM me via Twitter).

Anyway. In case you are curious, here’s the top 20 most frequent words in the text of the tweets, per source, in this dataset ( 15 February 2017 06:40:32 and 15 March 2017  08:14:20 Eastern Time). Analysis courtesy of Voyant Tools, applying a customised English stop words list (excluding Twitter-specific terms like rt, t.co, https, etc, but leaving terms in hashtags).

Android iPhone
Term Count Trend Term Count Trend
fake 11 0.007795889 great 16 0.016129032
great 11 0.007795889 jobs 14 0.014112903
media 10 0.007087172 america 6 0.006048387
obama 10 0.007087172 trump 6 0.006048387
election 9 0.006378455 american 5 0.005040322
just 9 0.006378455 join 5 0.005040322
news 9 0.006378455 big 4 0.004032258
big 8 0.005669738 healthcare 4 0.004032258
failing 6 0.004252303 meeting 4 0.004032258
foxandfriends 6 0.004252303 obamacare 4 0.004032258
president 6 0.004252303 thank 4 0.004032258
russia 6 0.004252303 u.s 4 0.004032258
democrats 5 0.003543586 whitehouse 4 0.004032258
fbi 5 0.003543586 address 3 0.003024194
house 5 0.003543586 better 3 0.003024194
new 5 0.003543586 day 3 0.003024194
nytimes 5 0.003543586 exxonmobil 3 0.003024194
people 5 0.003543586 investment 3 0.003024194
white 5 0.003543586 just 3 0.003024194
american 4 0.002834869 make 3 0.003024194

Android vs iPhone: Most Frequent Words from_user_id_str 25073877 Per Source

I have archived 3,603 public Tweets from_user_id_str 25073877 published between 27/02/2016 00:06 and 27/02/2017 12:06 (GMT -5, Washington DC Time). This is almost exactly a year’s worth of Tweets from the account in question.

Eight source types were detected in the dataset. Most of the Tweets were published either from iPhone (46%) or an Android (45%).

The Tweet counts per source are as follows:

 

Instagram 2
MediaStudio 1
Periscope 1
Twitter Ads 1
Twitter for Android 1629
Twitter for iPad 22
Twitter for iPhone 1660
Twitter Web Client 287
 Total 3603

 

The table above visualised as a bar chart, just because:

 

Source of 3603 Tweets from_user_id_str 25073877 (27/02/2016 00:06 to 27/02/2017 12:06) Bar chart.

 

As a follow/up to a previous post, I share in the table below the top 50 most frequent word forms per source (iPhone and Android) in this set of 3,603 Tweets  from_user_id_str 25073877, courtesy of a quick text analysis (applying a customised English stop word list globally) made with Voyant Tools:

 

Android iPhone
Term Count Trend Term Count Trend
great 276 0.008124816 thank 417 0.015241785
hillary 252 0.00741831 trump2016 215 0.007858475
trump 184 0.005416544 great 190 0.006944698
crooked 162 0.004768914 makeamericagreatagain 165 0.006030922
people 160 0.004710038 join 160 0.005848167
just 151 0.004445099 rt 144 0.00526335
clinton 120 0.003532529 hillary 119 0.004349574
big 107 0.003149838 clinton 118 0.004313023
media 106 0.0031204 america 111 0.004057166
thank 94 0.002767148 trump 104 0.003801309
bad 89 0.002619959 make 89 0.003253043
president 88 0.002590521 new 88 0.003216492
make 86 0.002531646 tomorrow 82 0.002997186
america 85 0.002502208 people 75 0.002741328
cnn 85 0.002502208 maga 73 0.002668226
country 72 0.002119517 today 73 0.002668226
like 72 0.002119517 americafirst 69 0.002522022
u.s 72 0.002119517 draintheswamp 68 0.002485471
time 71 0.00209008 tonight 67 0.00244892
said 67 0.001972329 ohio 66 0.002412369
jobs 66 0.001942891 vote 63 0.002302716
vote 63 0.001854578 just 61 0.002229614
win 63 0.001854578 florida 59 0.002156512
new 62 0.00182514 crooked 52 0.001900654
going 59 0.001736827 going 49 0.001791001
news 58 0.001707389 imwithyou 49 0.001791001
bernie 56 0.001648513 president 49 0.001791001
foxnews 55 0.001619076 votetrump 49 0.001791001
good 54 0.001589638 tickets 46 0.001681348
wow 53 0.0015602 american 43 0.001571695
job 50 0.001471887 time 43 0.001571695
nytimes 50 0.001471887 pennsylvania 42 0.001535144
republican 50 0.001471887 poll 41 0.001498593
0 49 0.001442449 soon 41 0.001498593
today 49 0.001442449 support 41 0.001498593
totally 49 0.001442449 enjoy 38 0.00138894
enjoy 48 0.001413012 campaign 37 0.001352389
cruz 46 0.001354136 rally 37 0.001352389
election 46 0.001354136 carolina 35 0.001279287
look 46 0.001354136 north 35 0.001279287
want 46 0.001354136 live 34 0.001242735
obama 44 0.001295261 speech 33 0.001206184
dishonest 41 0.001206947 california 18 0.000657919
can’t 39 0.001148072 hillaryclinton 18 0.000657919
night 39 0.001148072 honor 18 0.000657919
really 39 0.001148072 job 18 0.000657919
show 39 0.001148072 nevada 18 0.000657919
way 39 0.001148072 right 18 0.000657919
ted 38 0.001118634 supertuesday 18 0.000657919

 

I thought you’d like to know.

Donald’s Followers Going Up and Up…

In the context of popular calls to unfollow it (there’s a hashtag too), I  thought it would be interesting to look at how the number of followers of said Twitter account has been changing recently.

I looked at a dataset of all the Tweets from the account linked above timestamped between 04/11/2016 14:56 and 13/02/2017 22:30 (Washington DC time).

The change in followers (user_followers_count) in that period of time looks like this:

user_follower_count growth from:realdonaldtrump in tweets timestamped between 04/11/2016 14:56 and 13/02/2017 22:30 (Washington DC time)

 

The world appears to be collapsing, but his follower count keep going up… I thought you’d like to know.

We’ll keep an eye on this.

If you want to be able to read the account’s tweets without following it directly, there are many options. In case it’s useful, here’s a live searchable archive of recent tweets. (It’s bandwidth and Tweet volume dependent, so the resource may not always load).

 

Words Donald Likes So Far!

Trump Simplest Words image Image via The Telegraph
Image via The Telegraph, 20 September 2016

Since 20/01/2017 07:31:53 AM Eastern Time until 06/02/2017  07:07:55 AM Eastern Time Donald has…

  • …published 106 Tweets with his realDonaldTrump Twitter account. (He has published at least two more in the time I’ve been drafting this).
  • In this collection only one was published from the Twitter Web Client (the first one in this set)
  • 34 Tweets were published from Twitter for iPhone (mostly for Tweets between 20/01/2017 23:56 and 02/02/2017  12:29:16)
  • the remaining 71 Tweets were published from Twitter for Android.
  • All his latest Tweets, between 03/02/2017  06:24:51 and 06/02/2017  07:07:55, were published from Twitter for Android.
  • In this corpus he typed about 2,096 words or word forms (including URLs, Twitter account mentions and hashtags). This is about 5 pages.
  • 67 of his 106 Tweets include exclamation marks (!).
  • 31 of this 106 Tweets have included at least one word in all caps.

Sorry for all the bold type above.

Finally, these are the top 50 most frequent words (and emojis) in this set of 106 realDonaldTrump Tweets, courtesy of a quick text analysis (applying a customised English stop word list globally) made with Voyant Tools:

Term Count Trend
people 19 0.008796296
country 13 0.006018519
great 12 0.005555556
u.s 10 0.00462963
america 9 0.004166667
news 9 0.004166667
bad 8 0.003703704
fake 8 0.003703704
security 7 0.003240741
american 6 0.002777778
court 6 0.002777778
decision 6 0.002777778
enjoy 6 0.002777778
jobs 6 0.002777778
judge 6 0.002777778
just 6 0.002777778
meeting 6 0.002777778
today 6 0.002777778
ban 5 0.002314815
going 5 0.002314815
iran 5 0.002314815
make 5 0.002314815
states 5 0.002314815
thank 5 0.002314815
tonight 5 0.002314815
beginning 4 0.001851852
big 4 0.001851852
bring 4 0.001851852
coming 4 0.001851852
day 4 0.001851852
deal 4 0.001851852
election 4 0.001851852
illegal 4 0.001851852
interview 4 0.001851852
interviewed 4 0.001851852
like 4 0.001851852
long 4 0.001851852
nytimes 4 0.001851852
obama 4 0.001851852
p.m 4 0.001851852
party 4 0.001851852
president 4 0.001851852
supreme 4 0.001851852
united 4 0.001851852
whitehouse 4 0.001851852
yesterday 4 0.001851852
ºðÿ 4 0.001851852
abc 3 0.001388889
administration 3 0.001388889

So these are the most frequent presidential words so far.

I thought you would like to know.

Inauguration Day

Photo CC-BY Marc Nozell, Flick Commons
Photo CC-BY Marc Nozell, Flickr

One can give nothing whatever without giving oneself – that is to say, risking oneself. If one cannot risk oneself, then one is spimply incapable of giving.”

―James Baldwin, The Fire Next Time (1963)

As in previous years, my unwritten new year resolution is to write here more. I started blogging 18 years ago, and about 7 or 8 years ago my ‘personal’ blogging was reduced drastically, no doubt related to the rise of microblogging and new chapters in my life with greater public pressures, workloads and responsibilities.

The sad news of the death of Mark Fisher last week shook me deeply and reminded me how important someone’s public writing can be for others, or at least how much his blogging influenced me and helped me shape my own work ethic and politics.

I have blogged about trying to blog more meaningfully, only to be defeated by heavy workloads and my inability to not sleep or sacrifice even more personal and family wellbeing, so my blogging has been in the past few years largely dominated by updates, announcements and data research related posts.

I look up to colleagues like Martin Eve (and others) who reliably and periodically contribute to meaningful, public intellectual debate through their blog posts. Some of the most thought-provoking writing today is written not to meet targets, not to fulfill mandates, not as part of job descriptions. It’s urgent writing; it works as an invitation for collective thought and discussion. A space of reflection, shared generously, despite constraints. Good writing brings down walls.

Today I cannot but exercise my right to write about what I consider important.

Later today the world will witness Donald Trump’s inauguration as the new President of the United States of America. This morning I cannot but make a pause in everything urgent that I am working on to write this brief blog post where I express my profound concern and total rejection of everything that this political figure represents, both for the US and the rest of the world.

I am a Mexican and British citizen. I have friends and family living and working around the world. I aspire to being a ‘citizen of the world’. I  grew up witnessing at different levels of personal proximity the effects of social polarisation, corruption, stark inequality, discrimination, poverty and even civil war. The level of civic disempowerment that we are experiencing at this stage of the 21st century seems unparalleled; the more access we seem to have to means of producing and disseminating information the more defenseless we seem to become. This contradiction is painful to those of us who grew up being told that information was power and that education would lead to greater equality and with it, greater chances of wellbeing if not peace.

I want to be optimistic, not succumb to paranoia and keep hoping that everything will be all right. However we must also not lie to ourselves and pretend that the ‘values’ (tropes and motifs would be better terms) of the extreme right are not mainstream now. ‘The Brexit Bad Boys’ (sic), Trump, the alt-right, the ‘post-truth’ media ecosystem are not mere multimedia simulacra. It’s not just a dystopian fiction. It’s very much a tangible reality already affecting directly the lives of millions, within and outside the United States.

Those of us who believe in humanism and liberal, democratic values, who crave and work for equality and justice, cannot sit in front of the TV, sigh and merely hope for the better. Politically, it cannot be business as usual. We cannot be shy about publicly expressing our rejection of the politics of division, bullying and hate.

The revolution won’t be blogged, it won’t be tweeted, it won’t happen either on the streets. The Angel of History won’t come for us. In an age of total surveillance (not just from the State, but from everyone around us, including ourselves) the temptation is to continue remaining silent, our heads down, keeping calm and carrying on. (Others opt for constant commentary, noise silencing the signal).

Saying what we feel and believe, however, leaves a testimony, albeit a limited and fragile one. Saying what we feel and believe, openly, also works as a greeting, an expression of friendship and solidarity. When we write we give ourselves, and as such writing (not as an administrative requirement or task, but as a human need to share) is a risk. The risks of remaining silent seem much worse.

One cannot but hope for peace, tolerance, equality, respect. The near future looks very challenging. Information literacy, critical thinking, education are our armory.  Resisting won’t be futile.

‘BBCDebate’ on Twitter. A First Look into an Archive of #BBCDebate Tweets

[For the previous post in this series, click here].

The BBC Debate

The BBC’s Great Debate” was broadcasted live in the UK by the BBC on Tuesday 21 June 2016 between 20:00 and 22:00 BST. It saw activity on Twitter with the #BBCDebate hashtag.

I collected some of the Tweets tagged with #BBCDebate using a Google Spreadsheet. (See the methodology section below). I have shared an anonymised dataset on figshare:

Priego, E. (2016) “The BBC’s Great Debate”: Anonymised Data from a #BBCDebate Archive. figshare. https://dx.doi.org/10.6084/m9.figshare.3457688.v1

[Note: figshare DOIs are not resolving or there are delays in resolving; it should be fixed soon…]

Archive Summary (#BBCDebate)

Number of links 16826
Number of RTs 32206 <-estimate based on occurrence of RT
Number of Tweets 38116
Unique tweets 38066 <-used to monitor quality of archive
First Tweet in Archive 14/06/2016 22:03:18 BST
Last Tweet in Archive 22/06/2016 09:12:32 BST
In Reply Ids 349
In Reply @s 456
Tweet rate (tw/min) 62 Tweets/min (from last archive 10mins)
Unique Users in archive:

                      20, 243

Tweets from StrongerIn in archive:

16

Tweets from vote_leave in archive:

15

The raw data was downloaded as an Excel spreadsheet file containing 38,166 Tweets (38,066 Unique Tweets) publicly published with the queried hashtag (#BBCDebate) between 14/06/2016 22:03:18 and 22/06/2016 09:12:32 BST.

Due to the expected high volume of Tweets only users with at least 10 followers were included in the archive.

As indicated above the BBC Debate was broadcasted live on UK national television on Tuesday 21 June 2016 between 20:00 and 22:00 BST. This means the data collection covered the real-time broadcasting of the live debate (see the chart below).

#BBCDebate Activity in the last 3 days
#BBCDebate Activity in the last 3 days. Key: blue: Tweet; red: Reply

The data collected indicated only 12 Tweets in the whole archive contained geolocation data. A variety of user languages (user_lang) were identified.

Number of Different User Languages (user_lang)

Note this is not the language of the Tweets’ text, but the language setting in the application used to post the Tweet. In other words user_lang indicates the language the Twitter user selected from the drop-down list on their Twitter Settings page. This metadata is an indication of a user’s primary language but it might be misleading. For example, a user might select ‘es’ (Spanish) as their preferred language but compose their Tweets in English.

The following list ranks  user_lang  by number of Tweets in dataset in descending order. Specific counts can be obtained by looking at the dataset shared.

user_lang
en
en-gb
fr
de
nl
es
it
ja
ru
pt
ar
sv
pl
tr
da
ca
fi
id
ko
th
el
cs
no
en-IN
he
zh-cn
hi
uk

If you are interested in user_lang, GET help/languages returns the list of languages supported by Twitter along with the language code supported by Twitter. At the time of writing the language code may be formatted as ISO 639-1 alpha-2 (en), ISO 639-3 alpha-3 (msa), or ISO 639-1 alpha-2 combined with an ISO 3166-1 alpha-2 localization (zh-tw).

It is interesting to note the variety of European user_lang selected by those tweeting about #BBCDebate.

Notes on Methodology

The Tweets contained in the Archive sheet were collected using Martin Hawksey’s TAGS 6.0.

Given the relatively large volume of activity expected around #BBCDebate and the public and political nature of the hashtag, I have only shared indicative data. No full tweets nor any other associated metadata have been shared.

The dataset contains a metrics summary as well as a table with column headings labeled  created_at,  time,    geo_coordinates (anonymised; if there was data YES has been indicated; if no data was present the corresponding cell has been left blank), user_lang and user_followers_count data corresponding to each Tweet.

Timestamps should suffice to prove the existence of the Tweets and could be useful to run analyses of activity on Twitter around a real-time media event.

Text analysis of the raw dataset was performed using Stéfan Sinclair’s & Geoffrey Rockwell’s Voyant Tools. I may share results eventually if I find the time.

The collection and analysis of the dataset complies with Twitter’s Developer Rules of the Road.

Some basic deduplication and refining of the collected data was performed.

As in all the previous datasets I have created and shared it must be taken into account this is just a sample dataset containing the tweets published during the indicated period and not a large-scale collection of the whole output. The data is presented as is as a research sample and as the result of an archival task. The sample’s significance is subject to interpretation.

Again as in all the previous cases please note that both research and experience show that the Twitter search API is not 100% reliable. Large Tweet volumes affect the search collection process. The API might “over-represent the more central users”, not offering “an accurate picture of peripheral activity” (González-Bailón, Sandra, et al. 2012). Google spreadsheet limits must also be taken into account. Therefore it cannot be guaranteed the dataset contains each and every Tweet actually published with the queried Twitter hashtag during the indicated period. [González-Bailón et al have done very interesting work regarding political discussions online and their work remains an inspiration].

Only data from public accounts was included and analysed. The data was obtained from the public Twitter Search API. The analysed data is also publicly available to all Twitter users via the Twitter Search API and available to anyone with an Internet connection via the Twitter and Twitter Search web client and mobile apps without the need of a Twitter account.

Each Tweet and its contents were published openly on the Web, they were explicitly meant for public consumption and distribution and are responsibility of the original authors. Any copyright belongs to its original authors.

No Personally identifiable information (PII), nor Sensitive Personal Information (SPI) was collected nor was contained in the dataset.

I have shared the dataset including the extra tables as a sample and as an act of citizen scholarship in order to archive, document and encourage open educational and historical research and analysis. It is hoped that by sharing the data someone else might be able to run different analyses and ideally discover different or more significant insights.

For the previous post in this series, click here. If you got all the way here, thank you for reading.

References
[vote_leave]. (2016) [Twitter account]. Retrieved from https://twitter.com/vote_leave. [Accessed 21 June 2016].

González-Bailón, S., Banchs, R.E. and Kaltenbrunner, A. (2012) Emotions, Public Opinion and U.S. Presidential Approval Rates: A 5 Year Analysis of Online Political Discussions. Human Communication Research 38 (2) 121-143.

González-Bailón, S. et al (2012) Assessing the Bias in Communication Networks Sampled from Twitter (December 4, 2012). DOI: http://dx.doi.org/10.2139/ssrn.2185134

Hawksey, M. (2013) What the little birdy tells me: Twitter in education. Published on November 12, 2013. Presentation given from the LSE NetworkED Seminar Series 2013 on the use of Twitter in Education. Available from http://www.slideshare.net/mhawksey/what-the-little-birdy-tells-me-twitter-in-education [Accessed 21 June 2016].

Priego, E. (2016) “Vote Leave”. A Dataset of 1,100 Tweets by vote_leave with Archive Summary, Sources and Corpus Terms and Collocates Counts and Trends. figshare. URL: DOI: https://dx.doi.org/10.6084/m9.figshare.3452834.v1

Priego, E. (2016) “Stronger In”. A Dataset of 1,005 Tweets by StrongerIn with Archive Summary, Sources and Corpus Terms and Collocates Counts and Trends. figshare.
https://dx.doi.org/10.6084/m9.figshare.3456617.v1

Priego, E. (2016) “Stronger In”: Looking Into a Sample Archive of 1,005 StrongerIn Tweets. 21 June 2016. Available from https://epriego.wordpress.com/2016/06/21/stronger-in-looking-into-a-sample-archive-of-1005-strongerin-tweets/. [Accessed 21 June 2016].

Priego, E. (2016) “The BBC’s Great Debate”: Anonymised Data from a #BBCDebate Archive. figshare. https://dx.doi.org/10.6084/m9.figshare.3457688.v1

“Stronger In”: Looking Into a Sample Archive of 1,005 StrongerIn Tweets

If you haven’t been there already, please start here. An introduction and a detailed methodological note provide context to this post.

I have now shared a spreadsheet containing an archive of 1,005 @StrongerIn Tweets publicly published by the queried account between12/06/2016 13:34:35 and 21/06/2016 13:11:34 BST.

The spreadsheet contains four more sheets containing a data summary from the archive, a table of tweets’ sources, and tables of corpus term and trend counts and collocate counts.

This will hopefully allow to compare two similar samples from the output of two homologous Twitter accounts, both officially representing the ‘Leave’ and ‘Remain’ sides of the UK EU Referendum. The collected period is the same and if desired it is possible to edit the sets to have for example 1,000 Tweets each.

Following the structrue of my previous post on the ‘Vote Leave‘ dataset, here’s some quick insights from the @StrongerIn account for comparison.

Archive (from:StrongerIn)

Number of links 735
Number of RTs 409 <-estimate based on occurrence of RT
Number of Tweets

1005

Unique tweets 1004 <-used to monitor quality of archive
First Tweet in Archive

12/06/2016 13:34:35

BST
Last Tweet in Archive

21/06/2016 13:11:34

BST
In Reply Ids

9

In Reply @s 0
Tweet rate (tw/min)

0.1

Tweets/min (from last archive 10mins)

Like the @vote_leave account, @StrongerIn is used for mainly broadcasting Tweets and no @ Replies to users were collected during the period represented in the dataset.

Though this dataset, collected over slightly different timings but covering the same number of days, contains 60 fewer Tweets than the Vote Leave one; this @StrongerIn dataset reflects the account shared 235 links more than its @vote_leave counterpart.

Sources

Unlike @vote_leave, the dataset does not indicate that @StrongerIn uses Buffer nor Twitter for iPhone. However TweetDeck (423) and the Twitter Web Client (591) appear as the main sources. There’s even an interestingly strange Tweet, linking to a StrongerIn 404 web site page, published from NationBuilder.

Source Count
Nationbuilder

1

TweetDeck

413

Twitter Web Client

591

Total

1,005

Most Frequent Words

Removing Twitter data-specific stopwords from the raw data (e.g. t.co, amp, rt) the 10 most frequent words in the corpus are:

Term Count Trend
eu

287

0.013906387

remain

224

0.010853765

bbcqt

216

0.01046613

europe

209

0.01012695

vote

170

0.008237232

strongerin

167

0.00809187

uk

159

0.0077042347

jobs

148

0.0071712374

leave

148

0.0071712374

eudebate

113

0.0054753367

Compare them with the 10 most frequent words in the vote_leave data. Anything interesting?

 Let’s compare the top 10 terms from each account side by side:

 

Top 10 Terms in 1,100 vote_leave Tweets over 7 days vote_leave count Top 10 Terms in 1,005 StrongerIn Tweets over 7 days StrongerIn count
voteleave 558 eu 287
eu 402 remain 224
bbcqt 398 bbcqt 216
gove 165 europe 209
takecontrol 146 vote 170
immigration 133 strongerin 167
control 95 uk 159
cameron 89 jobs 148
turkey 84 leave 148
uk 72 eudebate 113

The terms in red are those appearing in both datsets; the terms in blue correspond to the name of each campaign. It’s interesting that though the StrongerIn account has 182 fewer mentions of ‘bbcqt’ (bear in mind the StrongerIn dataset has 95 fewer Tweets), ‘bbqt’ remains in third place on both sets.

The differences between the ranking of mentions of each campaign’s name are noticeable; as well as the fact that the vote_leave campaign has the name of the Prime Minister (himself a Remain campaigner) in its top 10 (as well as that of Gove; a Leave campaigner), while StrongerIn has no names of politicians on its 10 most frequent words.

There are other potentially interesting or noticeable differences when we compare these two top 10s. Can you spot them?  Do they tell us anything or not?

Digging into data and creating datasets does not necessarily tell us new things, but it does allow us to pinpoint otherwise moving objects. We don’t need to pin butterflies to recognise they are indeed butterflies, but the intention is to create new settings for observation.

References

González-Bailón, S., Banchs, R.E. and Kaltenbrunner, A. (2012) Emotions, Public Opinion and U.S. Presidential Approval Rates: A 5 Year Analysis of Online Political Discussions. Human Communication Research 38 (2) 121-143.

González-Bailón, S. et al (2012) Assessing the Bias in Communication Networks Sampled from Twitter (December 4, 2012). DOI: http://dx.doi.org/10.2139/ssrn.2185134

Hawksey, M. (2013) What the little birdy tells me: Twitter in education. Published on November 12, 2013. Presentation given from the LSE NetworkED Seminar Series 2013 on the use of Twitter in Education. Available from http://www.slideshare.net/mhawksey/what-the-little-birdy-tells-me-twitter-in-education [Accessed 21 June 2016].

Priego, E. (2016) “Vote Leave” Looking Into a Sample Archive of 1,100 vote_leave Tweets. 21 June 2016. Available from https://epriego.wordpress.com/2016/06/21/vote-leave-looking-into-a-sample-archive-of-1100-vote_leave-tweets/. [Accessed 21 June 2016].

Priego, E. (2016) “Vote Leave” A Dataset of 1,100 Tweets by vote_leave with Archive Summary, Sources and Corpus Terms and Collocates Counts and Trends. figshare. URL: DOI: https://dx.doi.org/10.6084/m9.figshare.3452834.v1

Priego, E. (2016) “Stronger In” A Dataset of 1,005 Tweets by StrongerIn with Archive Summary, Sources and Corpus Terms and Collocates Counts and Trends. figshare. DOI:
https://dx.doi.org/10.6084/m9.figshare.3456617.v1

[StrongerIn]. (2016). [Twitter account].Retrieved from https://twitter.com/StrongerIn. [Accessed 21 June 2016].

[vote_leave]. (2016) [Twitter account]. Retrieved from https://twitter.com/vote_leave. [Accessed 21 June 2016].

“Vote Leave”: Looking Into a Sample Archive of 1,100 vote_leave Tweets

In two days the United Kingdom will be voting in a Referendum that is very likely to change its destiny. More importantly, it is likely to change the destiny of everyone else who has a relationship with the UK.

This is a political event that is not only of national, internal or local interest, but one that is likely to have direct and immediate repercussions well beyond its borders. If one has ever lived in one of the EU member countries recently one does not need to be a political scientist to feel that these repercussions will not only be of a merely economic nature– already, even before the vote is cast, the UK’s social tissue has been undoubtedly transformed and deeply, even tragically affected.

Needless to say one of the arenas where political activity is taking place is on the media (TV, Radio) and social media. As the date to vote in person approaches, I collected and shared a dataset of tweets published by the official Leave campaign Twitter account, @vote_leave, between 12/06/2016 09:06:22 – 21/06/2016 09:29:29 BST. The dataset contains 1,100 tweets.

I did a quick text analysis of the Tweets themselves to get a quick insight into the most frequent terms and collocates in the corpus, and also looked at the tweets’ sources (the services used to publish the Tweets, i.e. the Twitter Web Client, Buffer, the Twitter iPhone app).

Some quick insights from the data:

Archive Summary (from:vote_leave)

Number of links 500
Number of RTs 592 <-estimate based on occurrence of RT
Number of Tweets 1100
Unique tweets 1099 <-used to monitor quality of archive
First Tweet in Archive

12/06/2016 09:06:22

BST
Last Tweet in Archive

21/06/2016 09:29:29

BST
Tweet rate (tw/min) 0.1 Tweets/min (from last archive 10mins)
In Reply Ids 3
In Reply @s 2
@s 90
RTs 54%

It is interesting that the account mostly broadcasts and RTs Tweets, but does a minimal interaction with other users via Reply @s, at least according to this sample dataset. (A larger dataset could corroborate or not if that is a trend indicating a media/content strategy or not).

Sources

The data indicates that most Tweets are published from the Twitter Web Client (496!), which I would have thought any marketing professional would find clunky if not really unfit for purpose.

Not suprisingly however Buffer is used (411 buffered Tweets), which indicates they are likely to have been scheduled in advance. Surprisingly for me, most of the Tweets in the dataset did not have TweetDeck as a source (only 4 according to the collected data in the given period), but it is possible that TweetDeck was used to ‘buffer’ the Tweets, as TweetDeck allows for Buffer integration.

Twitter for iPhone emerges as a significant source, well above Tweetdeck. Personally, I picture such an important political campaigning being done from a mobile phone as kind of scary. Influencing a nation’s destiny from the train home after the pub!

Source Count
Tweetdeck

4

Buffer

411

Twitter for iPhone

189

Twitter Web Client

496

Total

1100

Most Frequent Words

I was not surprised to see that ‘immigration’ was one of the most frequent words appearing in the corpus. However it was interesting to see the centrality of the hashtag ‘bbcqt’ (BBC Question Time). Even if we take into account the specific context of the data’s time period, the prevalence of bbcqt as a term in the corpus could be potentially interpreted as an indication of the importance that television, and specifically the BBC, has had in defining voting trends and public discourse regarding the Referendum.

Removing Twitter data-specific stopwords from the raw data (e.g. t.co, amp, rt) the 10 most frequent words in the corpus are:

Term Count Trend
voteleave

558

0.026160337

eu

402

0.018846694

bbcqt

398

0.018659165

gove

165

0.0077355835

takecontrol

146

0.0068448195

immigration

133

0.0062353495

control

95

0.004453821

cameron

89

0.0041725268

turkey

84

0.003938115

uk

72

0.0033755274

(voteleave, bbcqt, takecontrol were hashtags).

It is not clear how much of a social media/content strategy might be behind a Twitter account like @vote_leave, nor how many account managers are behind the tweetage. Apart from the obvious prevalence of ‘immigration’ as a term, it is nevertheless interesting to see that in 8 days of Tweets in the final countdown to the Referendum there would be a clear interest in tapping into televised debate and influence (bbcqt), to the point that the term would get such a high ranking. Bear in mind that ‘voteleave’ is their standard campaign hashtag, and that ‘eu’ would be expected to be a very frequent word, to the point that it could be considered a stop word in the specific context of this corpus. Perhaps for all the onus on social media as an autonomous medium it is still traditional mainstream media, in this case the BBC, which has the greatest influence in public opinion?

Notes on Methodology

The Tweets contained in the Archive sheet were collected using Martin Hawksey’s TAGS 6.0.

The text analysis was performed using Stéfan Sinclair’s & Geoffrey Rockwell’s Voyant Tools.

The collection and analysis of the dataset complies with Twitter’s Developer Rules of the Road.

The data was collected as an Excel spreadsheet file containing an archive of 1,100 @vote_leave Tweets publicly published by the queried account between 12/06/2016 09:06:22 – 21/06/2016 09:29:29 BST.

I prepared a spreadsheet and added four more sheets to add a data summary from the archive, a table of tweets’ sources, and tables of corpus term and trend counts and collocate counts.

It must be taken into account this is just a sample dataset containing the tweets published during the indicated period and not a large-scale collection of the whole output. The data is presented as is as a research sample and as the result of an archival task. The sample’s significance is subject to interpretation.

Please note that both research and experience show that the Twitter search API is not 100% reliable. Large Tweet volumes affect the search collection process. The API might “over-represent the more central users”, not offering “an accurate picture of peripheral activity” (González-Bailón, Sandra, et al. 2012). Therefore it cannot be guaranteed the dataset contains each and every Tweet actually published by the queried Twitter account during the indicated period. [González-Bailón et al have done very interesting work regarding political discussions online and their work remains an inspiration].

Only content from public accounts was included and analysed. The data was obtained from the public Twitter Search API. The analysed data is also publicly available to all Twitter users via the Twitter Search API and available to anyone with an Internet connection via the Twitter and Twitter Search web client and mobile apps without the need of a Twitter account.

Each Tweet and its contents were published openly on the Web, they were explicitly meant for public consumption and distribution and are responsibility of the original authors. Any copyright belongs to its original authors.

No Personally identifiable information (PII), nor Sensitive Personal Information (SPI) was collected nor was contained in the dataset.

I have shared the dataset including the extra tables as a sample and as an act of citizen scholarship in order to archive, document and encourage open educational and historical research and analysis. It is hoped that by sharing the data someone else might be able to run different analyses and ideally discover different or more significant insights.

For the next post on this series, click here.

References
[vote_leave]. (2016) [Twitter account]. Retrieved from https://twitter.com/vote_leave. [Accessed 21 June 2016].

González-Bailón, S., Banchs, R.E. and Kaltenbrunner, A. (2012) Emotions, Public Opinion and U.S. Presidential Approval Rates: A 5 Year Analysis of Online Political Discussions. Human Communication Research 38 (2) 121-143.

González-Bailón, S. et al (2012) Assessing the Bias in Communication Networks Sampled from Twitter (December 4, 2012). DOI: http://dx.doi.org/10.2139/ssrn.2185134

Hawksey, M. (2013) What the little birdy tells me: Twitter in education. Published on November 12, 2013. Presentation given from the LSE NetworkED Seminar Series 2013 on the use of Twitter in Education. Available from http://www.slideshare.net/mhawksey/what-the-little-birdy-tells-me-twitter-in-education [Accessed 21 June 2016].

Priego, E. (2016) “Vote Leave”. A Dataset of 1,100 Tweets by vote_leave with Archive Summary, Sources and Corpus Terms and Collocates Counts and Trends. figshare. URL: DOI: https://dx.doi.org/10.6084/m9.figshare.3452834.v1

Priego, E. (2016) “Stronger In”. A Dataset of 1,005 Tweets by StrongerIn with Archive Summary, Sources and Corpus Terms and Collocates Counts and Trends. figshare.
https://dx.doi.org/10.6084/m9.figshare.3456617.v1

Priego, E. (2016) “Stronger In”: Looking Into a Sample Archive of 1,005 StrongerIn Tweets. 21 June 2016. Available from https://epriego.wordpress.com/2016/06/21/stronger-in-looking-into-a-sample-archive-of-1005-strongerin-tweets/. [Accessed 21 June 2016].