Parler’s Swan Song: preliminary insights into 672k comments from 291k users in Parler’s final six days

January 15th, 1:30pm

Background

 

On Wednesday, January 6th, the U.S. Capitol was subject to a violent siege by supporters of President Donald Trump. In the aftermath, one alt-tech social media platform, Parler, became the subject of intense scrutiny for its role in facilitating the organization and execution of these acts of protest and sedition. In the five days following the capital attack major web hosting and app marketplaces have banned the site, leading to its eventual shutdown on 12am PST, Sunday, Jan 11. There are some extraordinary cases of site data being preserved through web scraping and white hat hacking. Our data was collected through Parler’s fairly permissive API; we prioritized posts that were shared by popular users, as well as those listed in the discover section of the website. We traversed the comments from these posts in reverse chronological order, regularly checking for additional new comments. We collected the user profiles for each post and comment creator. What follows is our initial analysis of the data with some high level insights.

Position Overview

We collected a dataset from the popular alt-tech social media website, Parler, comprising 672,454 comments made by 291,950 users on the top 1,377 most popular posts from January 6 to 12. Our initial n-gram analysis revealed that the most frequent trigram within the comments is “we need to,” used in contexts where users are calling for change or action. We further investigated what actions and desires users sought within the comments. We found that users’ attitudes were split on topics such as the need for a violent revolution, as well as hate speech. We investigated these attitudes toward certain topics within both user biographies and comments, and found that sentiment toward organized activity started to change after January 6. Finally, we found several bot accounts that were able to operate within the social network; while the most obvious bot accounts typically aimed to promote a product, service, or external website, it leaves the door open for more insidious bots that artificially amplify conversations or sentiments.

Figure 1. Word cloud formed from all collected comments

‘We need to start...”

We investigated n-gram frequencies across all comments, without a stop-word filter. While the top 1- and 2- gram frequencies were stop-words (“the”, “of the”), the top trigram was “we need to;” comments containing this trigram mostly called for action, answers, or political wish.

Figure 2. Ten most frequent phrases in the comments, ranked.

While threats of physical harm was present on the social network, much of the calls to action involved boycotts or virtual messaging campaigns (aka harassment). In the comments we encountered, “We need to start blowing up …” referred exclusively to sending massive amounts of messages to social media accounts or emails, not physical acts of terrorism. However, we do not preclude the possibility of physical threats being phrased in indirect ways, which is a common pattern in these networks. We are investigating this possibility further.

Controversial opinions on aftermath, hate speech, and need for violence

We used AuCoDe’s patent pending Contention Score to find the top 10 most controversial posts.

Figure 3. Ranked Top 10 posts with highest Contention Scores.

Hate Speech

We encountered a few instances of racially insensitive terms. While conventional social networks usually do not tolerate such language, posts containing these terms proved divisive on Parler. Use of these slurs did not pervade the platform (existing in less than 1% of posts) but nevertheless accounted for hundreds of votes.

Figure 4. Percent upvoted is computed from total votes across all posts containing a particular slur.

Violent Speech

We investigated calls for violence by testing for known phrases. We looked into posts made by 37 users whose profile biographies all contained the “blood of patriots and tyrants” quote from Thomas Jefferson. The accounts did not obviously overlap (which would suggest sockpuppeting); however, we also have no way of verifying that these accounts belong to individual people. Their posts did not contain enough votes to make sweeping claims about Parler users, but none of their posts had majority downvotes. We include a sample of these comments below.

Figure 5. A sample of comments from users who quoted Jefferson in their biographies, along with the total upvotes and downvotes for those comments.

Proposed organized activity leading up to Inauguration Day

Many posts call for protest on or leading up to Inauguration Day either explicitly through peaceful means or they are deliberately vague:

Figure 6. The top upvoted posts that explicitly mention action on, around, or as a result of the inauguration

Shifting sentiment for certain keywords

To analyse sentiment in these posts, we used the TextBlob library, and calculated the polarity score of each post. A negative score indicates a negative sentiment, positive score shows a positive sentiment and score of zero shows a neutral sentiment. Our preliminary results on sentiment analysis of certain keywords over the six days show that sentiment towards specific keywords show different trends. For instance, for the “stopthesteal” and “fightback” keyword we found a constantly increasing positive sentiment score over the six days, while for the keyword “stand down” we found a downward trend with negative sentiment. It is important to note that these are our preliminary findings and we intend to investigate this further with more advanced methods and analysis.

User Biographies

92093 of 291k users had filled out their biographies. On our initial glance at the user biographies, we found a series of key words and phrases that were shared across multiple bios. We then conducted searches on this sample of words and phrases to find the total number of user biographies that mention them. While a user may include a profession, religion, political stance, or other descriptive words, it cannot be assumed that the user is associated with, identifies as, supports, or opposes the mentioned word or phrase.

Figure 7. A sample of frequently mentioned keywords and phrases in user biographies with a focus on professions along with the percent of the 92093 user biographies that mentioned them.

Further research


Bot activity

Bots were present on the platform, though the most obvious ones were only there to sell services or products such as a Trump challenge coin. These bots posted the exact text with several accounts in different posts, and even had these posts upvoted. The oldest bot was from Nov. 19, 2020, and the majority were created within a week of their posts. We found many phrases that could potentially come from bots, such as “This is communism.” The lack of moderation around bots leaves enormous potential for disinformation campaigns to be carried out on these networks. We are investigating this further.

Time analysis of Contention and Sentiment

We believe that with further experiments we will be able to find insights into interesting trends such as “invitations for violence/peace”, “protests or riots” and others as well as analyzing the responses of different users to these actions. Additionally we are interested in investigating the level of violent rhetoric in these posts.

4mabht-aucode_logo_jg-resize21_white.png