An Early Look at the Parler Online Social Network
Max Aliapoulios, Emmi Bevensee, Jeremy Blackburn, Barry Bradlyn, Emiliano De Cristofaro, Gianluca Stringhini, Savvas Zannettou
AAn Early Look at the Parler Online Social Network
Max Aliapoulios , Emmi Bevensee , Jeremy Blackburn , Barry Bradlyn ,Emiliano De Cristofaro , Gianluca Stringhini , and Savvas Zannettou New York University, SMAT, Binghamton University, University of Illinois at Urbana-Champaign, University College London, Boston University, Max Planck Institute for [email protected], [email protected], [email protected], [email protected], [email protected]@bu.edu, [email protected]
Abstract
Parler is as an “alternative” social network promoting itselfas a service that allows to “speak freely and express yourselfopenly, without fear of being deplatformed for your views.”Because of this promise, the platform become popular amongusers who were suspended on mainstream social networksfor violating their terms of service, as well as those fearingcensorship. In particular, the service was endorsed by sev-eral conservative public figures, encouraging people to migratefrom traditional social networks. After the storming of the USCapitol on January 6, 2021, Parler has been progressively de-platformed, as its app was removed from Apple/Google Playstores and the website taken down by the hosting provider.This paper presents a dataset of 183M Parler posts made by4M users between August 2018 and January 2021, as well asmetadata from 13.25M user profiles. We also present a basiccharacterization of the dataset, which shows that the platformhas witnessed large influxes of new users after being endorsedby popular figures, as well as a reaction to the 2020 US Presi-dential Election. We also show that discussion on the platformis dominated by conservative topics, President Trump, as wellas conspiracy theories like QAnon.
Over the past few years, social media platforms that caterspecifically to users disaffected by the policies of mainstreamsocial networks have emerged. Typically, these tend not to beterribly innovative in terms of features, but instead attract usersbased on their commitment to “free speech.” In reality, theseplatforms usually wind up as echo chambers, harboring dan-gerous conspiracies and violent extremist groups. A case inpoint is Gab, one of the earliest alternative homes for peoplebanned from Twitter [32, 11]. After the Tree of Life terroristattack, it was hit with multiple attempts to de-platform the ser-vice, essentially erasing it from the Web. Gab, however, hassurvived and even rolled out new features under the guise offree speech that are in reality tools used to further evade andcircumvent moderation policies put in place by mainstreamplatforms [27].Worryingly, Gab as well as other more fringe platforms likeVoat [22] and TheDonald.win [25] have shown that not onlyis it feasible in technical terms to create a new social media platform, but marketing the platform towards specific polar-ized communities is an extremely successful strategy to boot-strap a user base. In other words, there is a subset of userson Twitter, Facebook, Reddit, etc., that will happily migrateto a new platform, especially if it advertises moderation poli-cies that do not restrict the growth and spread of political po-larization, conspiracy theories, extremist ideology, hateful andviolent speech, and mis- and dis-information.
Parler.
In this paper, we present an extensive dataset collectedfrom Parler. Parler is an emerging social media platform thathas positioned itself as the new home of disaffected right-wingsocial media users in the wake of active measures by main-stream platforms to excise themselves of dangerous communi-ties and content. While Parler works approximately the sameas Twitter and Gab, it additionally offers an extensive set ofself-serve moderation tools. For instance, filters can be set toplace replies to posted content into a moderation queue re-quiring manual approval, mark content as spam, and even au-tomatically block all interactions with users that post contentmatching the filters.After the events of January 6, 2021, when a violent mobstormed the US capitol, Parler came under fire for letting threatof violence unchallenged on its platform. The Parler app wasfirst removed from the Google Play and the Apple App stores,and the website was eventually deplatformed by the hostingprovider, Amazon AWS. At the time of writing, it is unclearwhether Parler will come back online and when.
Data Release.
Along with the paper, we release a dataset [1]including 183M posts made by 4M users between August2018 and January 2021, as well as metadata from 13.25M userprofiles. Each post in our dataset has the content of the postalong with other metadata (creation timestamp, score, hash-tags, etc.). The profile metadata include bio, number of follow-ers, how many posts the account made, etc. Our data releasefollows the FAIR principles, as discussed later in Section 4.We warn the readers that we post and analyze the datasetunfiltered; as such, some of the content might be toxic, racist,and hateful, and can overall be disturbing.
Relevance.
We are confident that our dataset will be useful tothe research community in several ways. Parler gained quickpopularity at a very crucial time in US History, following therefusal of a sitting President to concede a lost election, aninsurrection where a mob stormed the US Capitol building,1 a r X i v : . [ c s . S I] F e b nd, perhaps more importantly, the unprecedented ban of theUS President platforms like Facebook and Twitter. Thus, thedataset will constitute an invaluable resource for researchers,journalists, and activists alike to study this particular momentfollowing a data-driven approach.Moreover, Parler attracted a large migration of users onthe basis of fighting censorship, reacting to deplatformingfrom mainstream social networks, and overall an ideology ofstriving toward unrestricted online free speech. As such, thisdataset provides an almost unique view into the effects of de-platforming as well as the rise of a social network specificallytargeted to a certain type of users. Finally, our Parler datasetcontains a large amount of hate speech and coded languagethat can be leveraged to establish baseline comparisons as wellas to train classifiers. Parler (usually pronounced “par-luh” as in the French wordfor “to speak”) is a microblogging social network launchedin August 2018. Parler markets itself as being “built upona foundation of respect for privacy and personal data, freespeech, free markets, and ethical, transparent corporate pol-icy” [15]. Overall, Parler has been extensively covered in thenews for fostering a substantial user-base of Donald Trumpsupporters, conservatives, conspiracy theorists, and right-wingextremists [18].
Basics.
At the time of our data collection, to create an ac-count, users had to provide an email address and phone num-ber that can receive an activation SMS (Google Voice/VoIPnumbers are not allowed). Users interact on the social net-work by making posts of maximum 1,000 characters, called“parlays,” which are broadcasted to their followers. Users alsohave the ability to make comments on posts and on other com-ments.
Voting.
Similar to Reddit and Gab, Parler also has a votingsystem designated for ranking content, following a simple up-vote/downvote mechanism. Posts can only be upvoted, thusmaking upvotes functionally similar to likes on Facebook.Comments to posts, however, can receive both upvotes anddownvotes. Voting allows users to influence the order in whichcomments are displayed, akin to Reddit score.
Verification.
Verification on Parler is opt-in; users can will-ingly make a verification request by submitting a photographof themselves and a photo-id card. According to the website,verification–in addition to giving users a red badge–evidently“unlocks additional features and privileges.” They also declarethat the personal information required for verification is nevershared with third parties, and that after verification such in-formation is deleted except for “encrypted selfie data.” At thetime of writing, only 240,666 (2%) users on Parler are verified.
Moderation.
The Parler platform has the capability to per-form content moderation and user banning through adminis-trators. We explore these functionalities, from a quantitativeperspective, in Section 5. Note that there are several modera-tion attributes put in place per account, which are visible in anaccount’s settings. For instance, there is a field for whether the account “pending.” It appears that new accounts show up as“pending” until they are approved by automated moderation.Each account has a “moderation” panel allowing users toview comments on their own content and perform moderationactions on them. A comment can fall into any of five mod-eration categories: review , approved , denied , spam , or muted .Users can also apply keyword filters, which will enact one ofseveral automated actions based on a filter match: default (pre-vent the comment), approve (require user approval), pending , ban member notification , deny , deny with notification , denydetailed , mute comment , mute member , none , review , and tem-porary ban . These actions are enforced at the level of the userconfiguring the filters, i.e., if a filter is matched for temporaryban, then the user making the comment matching the filter isbanned from commenting on the original user’s content.There are several additional comment moderation settingsavailable to users. For example, users can allow only verifiedusers to comment on their content; there are tools to handlespam, etc. Overall, Parler allows for more individual contentmoderation compared to other social networks; however, re-cent reports have highlighted how global moderation is ar-guably weaker, as large amounts of illegal content has been al-lowed on the platform [6]. We posit this may be due to globalmoderation being a manual process performed by a few ac-counts. Monetization.
Parler supports “tipping,” allowing users to tipone another for content they produce. This behavior is turnedoff by default, both with respect to accepting and being able togive out tips. An additional monetization layer is incorporatedwithin Parler, which is called “Ad Network” or “Influence Net-work” [8]. Users with access to this feature are able to pay foror earn money for hosting ad campaigns. Users set their rateper thousand views in a Parler specified currency called “Par-ler Influence Credit.”
We now discuss our methodology to build the dataset releasedalong with this paper. We use a custom-built crawler that ac-cesses the (undocumented, but open) Parler API. This crawlerwas based on Parler API discoveries that allowed for fastercrawling [7].
Crawling.
Our data collection procedure works as follows.First, we populate users via an API request that maps a mono-tonically increasing integer ID (modulo a few exceptions) toa universally unique ID (UUID) that serves as the user’s IDin the rest of the API. Next, for each UUID we discover,we query for its profile information, which includes metadatasuch as badges, whether or not the user is banned, bio, pub-lic posts, comments, follower and following counts, when theuser joined, the user’s name, their username, whether or notthe account is private, whether or not they are verified, etc.Note that, to retrieve posts and comments, we use an API end-point that allows for time-bounded queries; i.e., for each user,we retrieve the set of post/comments since the most recentpost/comment we have already collected for that user.
Data.
Overall, we collect all user profile information for the2 ount
Comments
Total
Table 1:
Dataset Statistics.
Limitations of Sampling.
As mentioned above, posts andcomments in our dataset are from a sample of users; moreprecisely, 183.063M and 4.08M users, respectively. Although,numerically, this should in theory provide us with a good rep-resentation of the activities of Parler’s user base, we acknowl-edge that our sampling might not necessarily be representa-tive in a strict statistical sense. In fact, using a two-sample KStest, we reject the null hypothesis that the distribution of com-ments reported in the profile data from all users is the same asthe distribution of those we actually collect from 1.1M users( p < . ). We speculate that this could be due to the pres-ence of a small number of very active users which were notcaptured in our sample. Moreover, we have posts from manyfewer users than we have comments for; this is due to users’tendency to make more posts than comments, which increasesthe wall clock time it takes to collect posts.Therefore, we need to take these possible limitations intoaccount when analyzing user content—e.g., as we do in Sec-tion 6. Nonetheless, we believe that our sample does ultimatelycapture the general trends measured from profile data, and thuswe are confident our sample provides at least a reasonable rep-resentation of content posted to Parler. Ethical Considerations.
We only collect and analyze publiclyavailable data. We also follow standard ethical guidelines [26],not making any attempts to track users across sites or de-anonymize them. Also, taking into account user privacy, weremove from the data the names of the Parler accounts in ourdataset.
This section presents the structure of the data, available at [1].Overall, the data consists of newline-delimited JSON files( .ndjson ), obtained by crawling three main Parler API end-points, /v1/post , /v1/comment , and /v1/user . EachJSON consists of key/value pairs returned by their respectiveAPI endpoint.Due to space limitations, in the following, we only list thekeys used in the analysis in this paper. The complete key/valuelist as well as their definitions are available at [1]. Key/Values from /v1/post[comment].
From the post andcomment endpoints, we have:• id : Parler generated UUID of the post/comment.• createdAt : Timestamp of the post/comment in UTC. • upvotes : Number of upvotes that a post/comment re-ceived.• score : Number of upvotes minus the sum of the down-votes a post/comment received.• hashtags : List of strings that corresponds to the hashtagsused in a post/comment.• urls : List of dictionaries correspond to URLs and theirrespective metadata used in a post/comment.We also enriched the data with fields from the user profilewho produced the content, like username and verified , in or-der to provide additional context to the post or comment. Ad-ditionally, we formatted the following fields from strings tointegers to facilitate numerical analysis: depth , impressions , reposts , upvotes , and score . Key/Values from /v1/user . From the user endpoint, wehave:• id : Parler generated UUID for the user.• badges : List of numeric values corresponding to the userprofile badges.• bio : Biography string written by a user for their profile.• ban : Boolean field as to whether or not the user is cur-rently banned.• user_followers : Numeric field corresponding to the num-ber of followers the user profile has.• user_following : Numeric field corresponding to the num-ber of users the user profile follows.• posts : Number of posts a user has made.• comments : Number of comments a user has made.• joined : Timestamp of when a user joined in UTC.We renamed the following fields because they are reservedfield names in our datastore: followers to user_followers and following to user_following . We also reformatted the follow-ing fields from strings to integers to facilitate numerical anal-ysis: comments , posts , following , media , score , followers . FAIR Principles.
The data released along with this paperaligns with the FAIR guiding principles for scientific data. First, we make our data
Findable by assigning a uniqueand persistent digital object identifier (DOI): 10.5281/zen-odo.4442460. Second, our dataset is
Accessible as it can bedownloaded, for free. It is in JSON format, which is widelyused for storing data and has an extensive and detailed doc-umentation for all of the computer programming languagesthat support it, thus enabling our data to be
Interoperable . Fi-nally, our dataset is extensively documented and described inthis paper and in [1], and released openly, thus our dataset is
Reusable . This section analyzes the data from the 13.25M Parler userprofiles collected between November 25, 2020 and January 11,2021.
We analyze the user bios of all the Parler users in ourdataset. We extract the most popular words and bigrams, re- ord (%) Bigram (%) conservative 1.23% trump supporter 0.26%god 0.99% husband father 0.24%trump 0.96% wife mother 0.19%love 0.88% god family 0.17%christian 0.78% trump 2020 0.17%patriot 0.76% proud american 0.17%wife 0.74% wife mom 0.16%american 0.7% pro life 0.15%country 0.65% christian conservative 0.14%family 0.62% love country 0.13%life 0.58% love god 0.13%proud 0.57% family country 0.13%maga 0.55% president trump 0.12%mom 0.54% god bless 0.12%father 0.54% business owner 0.12%husband 0.52% jesus christ 0.1%jesus 0.45% conservative christian 0.1%freedom 0.43% american patriot 0.1%retired 0.42% maga kag 0.1%america 0.41% god country 0.09% Table 2:
Top 20 words and bigrams found in Parler users bios. ported in Table 2. Several popular words indicate that a sub-stantial number of users on Parler self identify as conserva-tives (1.3% of all users include the word “conservative” it intheir bios), Trump supporters (1% include the word “trump”and 0.27% the bigram “trump supporter”), patriots (0.79% ofall users include the word “patriot”), and religious individu-als (1.05% of all users include the word “god” in their bios).Overall, these results indicate that Parler attracts a user basesimilar to the one that exists on Gab [32].
Parler profile data includes a flag that is set when a user isbanned. We find this flag set for 252,209 (2.09%) of users. Al-most all of these banned accounts, 252,076 (99.95%) are alsoset to private, but for those that are not, we can observe theirusername, name and bio attributes, and even retrieve com-ments/posts they might have made. While not a thorough anal-ysis of the ban system, when exploring the 157 non-privatebanned accounts, we notice some interesting things. In gen-eral, there appears to be two classes of banned users. The firstare accounts banned for impersonating notable figures, e.g.,the name “Donald J Trump” and a bio that describes the useras the “45th President of the United States of America,” or a“ParlerCEO” username with “John Matza” as the user’s dis-play name. These actions violate the Parler guideline around“Fraud, IP Theft, Impersonation, Doxxing” suggesting Parlerdoes in fact enforce at least some of their moderation policies.For the second class of banned user, it is harder to determinethe guideline violation that led to the ban. For example, an ac-count named “ConservativesAreRetarded” whose profile pic-ture was a hammer and sickle made a comment (the account’sonly comment) in response to a post by a Parler employee. Thecomment,“You look like garbage, at least take a decent photoof the shirt,” was made in reply to a post that included an image
Badge
Table 3:
Badges assigned to user profiles. Users are given an array ofbadges to choose from based on their profile parameters. the Parler employee wearing a Parler t-shirt. While the com-ment by “ConservativesAreRetarded” is certainly not nice, itwas not clear to us which of the Parler guidelines it violated.We do not believe this would be considered violent, threaten-ing, or sexual content, which are explicitly noted in the guide-lines. Although outside the scope of this paper, it does call intoquestion how consistently Parler’s moderation guidelines arefollowed.
There are several badges that can be awarded to a Parleruser profile. These badges correspond to different types of ac-count behavior. Users are able to select which badges they optto appear on their user profile.We detail the large variety ofbadges available in Table 3. A user can have no badges or mul-tiple badges. For each user, our crawler returns a set of badgenumbers; we then looked up users with specific badges in theParler UI in order to see the badge tag and description whichare displayed visually.
The user profile objects returned by the Parler API containa “verified” field that corresponds to a boolean value. All userswith this value set to “True” have a gold badge and vice versa.We assume that these users are actually the small set of truly“verified” users in the more widely adopted sense of the word,akin to “blue check” users on Twitter. There are only 5964 C D F OtherVerifiedGold badge (a) C D F OtherVerifiedGold badge (b)
Figure 1:
CDFs of the number of posts and comments of verified, gold badge, and other users. (Note log scale on x-axis). C D F OtherVerifiedGold badge (a) C D F OtherVerifiedGold badge (b)
Figure 2:
CDF of the number of following and followers of gold badge, verified, and other users. (Note log scale on x-axis). gold badge users on Parler, which is less than 1% of the en-tire user count. This is in contrast to the red “Verified” badgetag (Badge
There are two numbers related to the underlying social net-work structure available in the user profile objects: A user’sfollowers corresponds to how many individuals are currently following that user, whereas their followings corresponds tohow many users they follow. Figure 2(a) shows the cumulativedistribution function (CDF) of the followers per user split bybadge type, while Figure 2(b) shows the same for the follow-ings of each user. First we note that standard users are less pop-ular, as they have fewer followers; see Figure 2(a). Gold badgeusers on the other hand have a much larger number of follow-ers. We see that about 40% of the typical users have more thana single follower, whereas about 40% of gold badge users havemore than 10,000 followers; verified users fall somewhere inthe middle. As seen in Figure 2(b), typical users also followa smaller number of accounts compared to gold and verifiedusers.
The number of users on Parler grew throughout the courseof the platform’s lifetime. We notice several key events corre-lated with periods of user growth. Parler originally launchedin August 2018. Figure 3 plots cumulative users growth sinceParler went live. Parler saw its first major growth in number ofusers in December 2018, reportedly because conservative ac-tivist Candace Owens tweeted about it [21]. The second largenew user event occurred in June 2019 when Parler reportedthat a large number of accounts from Saudi Arabia joined [9].In 2020 there were two large events of new users. The firstoccurred in June 2020, where on June 16th 2020 conservativecommentator Dan Bongino announced he had purchased anownership stake in the platform [28]. At the same time, Parler5 / / / / / / / / / / / / / / o f u s e r s ( c u m u l a t . ) Figure 3:
Cumulative number of users joining daily. (Note log scaleon y-axis). Table 4 reports the events annotated in the figure. o f u s e r s j o i n e d (a) November 2020
05 k10 k15 k20 k25 k30 k o f u s e r s j o i n e d (b) January 2021 Figure 4:
Number of users joining per day in November 2020 (themonth of the US Elections) and January 2021 (the month of the Jan-uary 6 insurrection). also received a second endorsement from Brad Parscale, thesocial media campaign manager for Trump’s 2016 campaign.The last major user growth event in 2020 occurred around thetime of the United States 2020 election and some cite [10]this growth as a result of Twitter’s continuous fact-checking ofDonald Trump’s tweets. As we show in Figure 4(a), a substan-tial number of new accounts were created during November2020, while the outcome of the US 2020 Presidential Elec-tion was determined. In addition, we saw additional substan-tial user growth in January 2021, as shown in Figure 4(b), es-pecially in the days after the Capitol insurrection.Finally, Figure 5 shows account creations for users that havea “gold” badge, “verified” badge, and “other” users that do nothave either badge. We observe that throughout the course oftime, Parler attracts new users that become verified and goldusers.
We now analyze the content posted by Parler users, focusingon activity volume, voting, hashtags used, and URLs sharedon the platform.
We begin our analysis by looking at the volume of postsand comments over time. Figure 6 plots the weekly number ofposts and comments in our dataset. We observe that the shape / / / / / / / / / / / / / / o f u s e r s j o i n e d OtherVerifiedGold badge
Figure 5:
Number of users joining daily split by gold badge, verified,and other users. (Note log scale on y-axis.)
EventID Description Date
Table 4:
Events depicted in Figures 3 and 6. / / / / / / / / / / / / / / P o s t s PostsComments
Figure 6:
Number of posts per week. (Note log scale on y-axis). Ta-ble 4 reports the events annotated in the figure. of curves is similar to that in Figure 3, i.e., there are spikes inpost/comment activity at the same dates where there is an in-flux of new users due to external events. Parler was a relativelysmall platform between August 2018 and June 2019, withless than 10K posts and comments per week. Then, by June2019, there is a substantial increase in the volume of posts andcomments, with approximately 100K posts and comments perweek. This coincides with a large-scale migration of Twitterusers originating from Saudi Arabia, who joined Parler due toTwitter’s “censorship” [9]. The volume of posts and commentsremain relatively stable between June 2019 and June 2020,while, in mid-2020, there is another large increase in posts andcomments, with 1M posts/comments per week. This coincideswith when Twitter started flagging President’s Trump tweetsrelated to the George Floyd Protests, which prompted Parlerto launch a campaign called “Twexit,” nudging users to quitTwitter and join Parler [19]. Finally, by the end of our datasetin late 2020, another substantial increase in posts/commentscoincides with a sudden interest in the platform after DonaldTrump’s defeat in the 2020 US Presidential Election. Note thatthe number of posts per user is nearly constant except whenthere is a large influx of new users.6 Upvotes0.00.20.40.60.81.0 C D F (a) Posts Score0.00.20.40.60.81.0 C D F (b) Comments Figure 7:
CDFs of the number of upvotes on posts and scores (upvotes minus downvotes) on comments. (Note log scale on x-axis).
As mentioned in Section 2, posts on Parler can be upvoted,while comments can be upvoted and downvoted, thus yieldinga score (sum of upvotes minus the sum of downvotes). Fig-ure 7 shows the CDFs of the upvotes and score for posts andcomments. We find that 18% of the posts do not receive anyupvotes, and 61% of posts receive at least 10 upvotes. Look-ing at the scores for comments (see Figure 7(b)), we observethat comments rarely have a negative score (only 1.6%), while44% of comments have a score equal to zero, and the rest havepositive scores. Overall, our results indicate that a substantialamount of content posted on Parler is viewed positively by itsusers.
Next, we focus on the prevalence and popularity of hash-tags on Parler. We find that only a small percentage ofposts/comments include hashtags: 2.9% and 3.4% of all postsand comments, respectively. We then analyze the most popularhashtags, as they can provide an indication of users’ interests.Table 5 reports the top 20 hashtags in posts and comments.Among the most popular hashtags in posts (left side of thetable), we find Furthermore, we find several hashtags that arerelated to the alleged election fraud that Trump and his sup-porters claimed occured during the 2020 US Elections (e.g., Where We Go One We Go All (WWG1WGA) is a popular QAnon motto.
Hashtag trump2020 347,799 parlerconcierge 962,207maga 271,379 trump2020 181,219stopthesteal 200,059 newuser 157,186parler 187,363 maga 148,655wwg1wga 176,150 truefreespeech 147,769trump 168,649 stopthesteal 93,927kag 117,894 wwg1wga 52,687qanon 117,134 parler 45,211freedom 108,794 kag 43,396parlerksa 97,004 trump 39,659newuser 87,263 maga2020 31,055news 86,771 usa 28,046usa 84,271 obamagate 26,250trumptrain 82,893 1 22,850thegreatawakening 82,710 wethepeople 22,236meme 82,440 fightback 21,954electionfraud 80,457 blm 19,979maga2020 79,046 qanon 19,758voterfraud 78,793 trump2020landslide 19,179americafirst 75,764 americafirst 19,012
Table 5:
Top 20 hashtags in posts and comments.
Finally, we focus on URLs shared by Parler users: 15.7%and 7.9% of all posts and comments, respectively, includeat least one URL. Table 6 reports the top 20 domains in theshared URLs. Among the most popular domains, we find Par-ler itself, YouTube, image hosting sites like Imgur, links tomainstream social media platforms like Twitter, Facebook, andInstagram, as well as news sources like Breitbart and NewYork Post.Overall, our URL analysis suggests that Parler users aresharing a mixture of both mainstream and alternative contenton the Web. For instance, they are sharing YouTube URLs(mainstream) as well as Bitchute URLs, a “free speech” ori-ented YouTube alternative [30]. The same applies with newssources: Parler users are sharing both alternative news sources(e.g., Breitbart) and mainstream ones (New York Post, aconservative-leaning outlet), with the alternative news sourcesbeing more popular in general.7 omain parler.com 5,017,486 parler.com 2,488,718youtu.be 1,275,127 youtube.com 1,767,928youtube.com 827,145 giphy.com 1,314,282twitter.com 773,041 bit.ly 872,611facebook.com 493,804 youtu.be 449,666thegatewaypundit.com 478,982 imgur.com 163,381imgur.com 353,184 par.pw 50,779breitbart.com 345,700 twitter.com 35,035foxnews.com 336,390 tenor.com 32,922theepochtimes.com 236,278 bitchute.com 30,751giphy.com 87,344 facebook.com 27,460instagram.com 162,769 rumble.com 19,919rumble.com 142,495 thegatewaypundit.com 12,747westernjournal.com 99,271 google.com 12,249t.co 84,633 whitehouse.gov 12,002nypost.com 84,288 blogspot.com 11,575par.pw 78,473 gmail.com 9,267ept.ms 77,069 wordpress.com 9,181bitchute.com 73,970 amazon.com 8,886townhall.com 72,781 foxnews.com 7,987
Table 6:
Top domains on Parler.
In this section, we review related work.
Datasets. [5] present a data collection pipeline and a datasetwith news articles along with their associated sharing activ-ity on Twitter. [11] release a dataset of 37M posts, 24.5Mcomments, and 819K user profiles collected from Gab. [23]present an annotated dataset with 3.3M threads and 134.5Mposts from the Politically Incorrect board (/pol/) of the image-board forum 4chan, posted over a period of almost 3.5 years(June 2016–November 2019). [2] present a large-scale datasetfrom Reddit that includes 651M submissions and 5.6B com-ments posted between June 2005 and April 2019. [14] presenta methodology for collecting large-scale data from WhatsApppublic groups and release an anonymized version of the col-lected data. They scrape data from 200 public groups and ob-tain 454K messages from 45K users. Finally, [13] use crowd-sourcing to label a dataset of 80K tweets as normal, spam, abu-sive, or hateful. More specifically, they release the tweet IDs(not the actual tweet) along with the majority label receivedfrom the crowd-workers.
Fringe Communities.
Over the past few years, a number ofresearch papers have provided data-driven analyses of fringe,alt- and far-right online communities, such as 4chan [17, 4, 31,24], Gab [32], Voat [22], The_Donald and other hateful sub-reddits [12, 20], etc. Prior work has also analyzed their impacton the wider web, e.g., with respect to disinformation [34],hateful memes [33], and doxing [29].
This paper presented our Parler dataset, along with a generalcharacterization. We collected and released user informationfor 13.25M users that joined the platform between 2018 and 2020, as well as a sample of 183M posts by 4M users.Our preliminary analysis shows that Parler attracts the in-terest of conservatives, Trump supporters, religious, and pa-triot individuals. Also, the data reveals that Parler experiencedlarge influxes of new users in close temporal proximity withreal-world events related to online censorship on mainstreamplatforms like Twitter, as well as events related to US poli-tics. Additionally, our dataset sheds light into the content thatis disseminated on Parler; for instance, Parler users share con-tent related to US politics, content that show support to DonaldTrump and his efforts during the 2020 US elections, and con-tent related to conspiracy theories like the QAnon conspiracytheory.Overall, Parler is an emerging alternative platform thatneeds to be considered by the research community that focuseson understanding emerging socio-technical issues (e.g., onlineradicalization, conspiracy theories, or extremist content) thatexist on the Web and are related to US politics. To this end,we are confident that our dataset will pave the way to motivateand assist researchers in studying and understanding extremeplatforms like Parler, especially at a crucial point of US andWorld history.At the time of writing, Parler is being taken down by itshosting provider, Amazon AWS. It is unclear when and howthe service will come back, which potentially makes the snap-shot provided in this paper even more useful to the researchcommunity.
References [1] M. Aliapoulios, E. Bevensee, J. Blackburn, B. Bradlyn,E. De Cristofaro, G. Stringhini, and S. Zannettou. A LargeOpen Dataset from the Parler Social Network. https://doi.org/10.5281/zenodo.4442460, 2021.[2] J. Baumgartner, S. Zannettou, B. Keegan, M. Squire, andJ. Blackburn. The pushshift reddit dataset. In
ICWSM
ICWSM , 2011.[5] G. Brena, M. Brambilla, S. Ceri, M. Di Giovanni, F. Pierri, andG. Ramponi. News Sharing User Behaviour on Twitter: A Com-prehensive Data Collection of News Articles and Social Inter-actions. In
ICWSM , 2019.[6] R. L. Craig Timberg, Drew Harwell. Parler’s got a porn prob-lem: Adult businesses target pro-Trump social network. https://wapo.st/3nlQu9g, 2020.[7] donk_enby. Parler tricks. https://doi.org/10.5281/zenodo.4426283, 2020.[8] donk_enby. Accessing the hidden parts of the Parler iOS app us-ing dynamic runtime instrumentation. https://doi.org/10.5281/zenodo.4426480, 2021.[9] K. P. Elizabeth Culliford. Unhappy with Twitter, thousands ofSaudis join pro-Trump social network Parler. https://reut.rs/3bi7haK, 2019.[10] R. L. Elizabeth Dwoskin. “Stop the Steal” supporters, re-strained by Facebook, turn to Parler to peddle false election ICWSM , 2019.[12] C. Flores-Saviaga, B. Keegan, and S. Savage. Mobilizingthe trump train: Understanding collective action in a politicaltrolling community. In
ICWSM , 2018.[13] A. M. Founta, C. Djouvas, D. Chatzakou, I. Leontiadis,J. Blackburn, G. Stringhini, A. Vakali, M. Sirivianos, andN. Kourtellis. Large scale crowdsourcing and characterizationof twitter abusive behavior. In
ICWSM , 2018.[14] K. Garimella and G. Tyson. WhatsApp Doc? A First Look atWhatsApp Public Group Data. In
ICWSM
ICWSM , 2017.[18] R. Lerman. The conservative alternative to Twitter wants tobe a place for free speech for all. It turns out, rules still apply.https://wapo.st/2MEa2c5, 2020.[19] J. Lewinski. Social Media App Parler Releases “Declara-tion Of Internet Independence” Amidst
ICWSM arXiv preprint 2009.04885 , 2020.[23] A. Papasavva, S. Zannettou, E. De Cristofaro, G. Stringhini, andJ. Blackburn. Raiders of the Lost Kek: 3.5 Years of Augmented4chan Posts from the Politically Incorrect Board. In
ICWSM ,volume 14, 2020.[24] B. T. Pettis. Ambiguity and Ambivalence: Revisiting OnlineDisinhibition Effect and Social Information Processing Theoryin 4chan’s /g/ and /pol/ Boards. https://benpettis.com/writing/2019/5/21/ambiguity-and-ambivalence, 2019.[25] M. H. Ribeiro, S. Jhaver, S. Zannettou, J. Blackburn,E. De Cristofaro, G. Stringhini, and R. West. Does PlatformMigration Compromise Content Moderation? Evidence fromr/The_Donald and r/Incels. arXiv preprint 2010.10397 , 2020.[26] C. M. Rivers and B. L. Lewis. Ethical research standards in aworld of big data.
F1000Research , 2014.[27] E. Rye, J. Blackburn, and R. Beverly. Reading In-Between theLines: An Analysis of Dissenter. In
ACM IMC
ACM IMC , 2017.[30] M. Trujillo, M. Gruppi, C. Buntain, and B. D. Horne. What IsBitChute? Characterizing The. In
ACM HyperText , 2020.[31] M. Tuters and S. Hagen. (((They))) rule: Memetic antagonismand nebulous othering on 4chan.
SAGE NMS , 2019.[32] S. Zannettou, B. Bradlyn, E. De Cristofaro, H. Kwak, M. Siri-vianos, G. Stringini, and J. Blackburn. What is Gab: A Bastionof Free Speech or an Alt-Right Echo Chamber. In
CompanionProceedings of the The Web Conference 2018 , 2018.[33] S. Zannettou, T. Caulfield, J. Blackburn, E. De Cristofaro,M. Sirivianos, G. Stringhini, and G. Suarez-Tangil. On the ori-gins of memes by means of fringe web communities. In
ACMIMC , 2018.[34] S. Zannettou, T. Caulfield, E. De Cristofaro, N. Kourtelris,I. Leontiadis, M. Sirivianos, G. Stringhini, and J. Blackburn.The web centipede: understanding how web communities influ-ence each other through the lens of mainstream and alternativenews sources. In
ACM IMC , 2017., 2017.