key: cord-0143177-ny8ptuzr
authors: Israeli, Abraham; Tsur, Oren
title: Going Extreme: Comparative Analysis of Hate Speech in Parler and Gab
date: 2022-01-27
journal: nan
DOI: nan
sha: 6a1e4ac6bd38a628f42014ee1640e8c0cd71c958
doc_id: 143177
cord_uid: ny8ptuzr

Social platforms such as Gab and Parler, branded as `free-speech' networks, have seen a significant growth of their user base in recent years. This popularity is mainly attributed to the stricter moderation enforced by mainstream platforms such as Twitter, Facebook, and Reddit. In this work we provide the first large scale analysis of hate-speech on Parler. We experiment with an array of algorithms for hate-speech detection, demonstrating limitations of transfer learning in that domain, given the illusive and ever changing nature of the ways hate-speech is delivered. In order to improve classification accuracy we annotated 10K Parler posts, which we use to fine-tune a BERT classifier. Classification of individual posts is then leveraged for the classification of millions of users via label propagation over the social network. Classifying users by their propensity to disseminate hate, we find that hate mongers make 16.1% of Parler active users, and that they have distinct characteristics comparing to other user groups. We find that hate mongers are more active, more central and express distinct levels of sentiment and convey a distinct array of emotions like anger and sadness. We further complement our analysis by comparing the trends discovered in Parler and those found in Gab. To the best of our knowledge, this is among the first works to analyze hate speech in Parler in a quantitative manner and on the user level, and the first annotated dataset to be made available to the community.

[Warning: Some of the readers may find the language in the examples provided in this manuscript offensive.] Social platforms like Twitter, Facebook, and Reddit have become a central communication channel for billions of users 1 . However, the immense popularity of social platforms resulted in a significant rise in the toxicity of the discourse, ranging from cyber-bullying to explicit hate speech and calls for violence against individuals and groups (Waseem and Hovy 2016; Mondal, Silva, and Benevenuto 2017; Laub 2019; Ziems et al. 2020) . Women, people of color, the LGBT community, Muslims, immigrants, and Jews are among the most targeted groups. Recent studies report on a surge in Islamophobia (Akbarzadeh 2016; Sunar 2017; Osman 2017;  examples is presented in Table 1 . Notably, some posts are more explicit than others -using vulgar language (e.g., posts #1-#3), explicitly mentioning the targeted individual/group (e.g., #1,#3,#6,#7), while other posts are using nick-names, codes and implicit references (e.g., #2,#4,#5,#8).

Striking the right balance between contradicting values (e.g., the freedom of speech vs. public safety of members of protected groups) is a walk on a tightrope. We believe, however, that a data-oriented analysis may help individuals and policy maker alike at reaching an informed balance.

In this work we focus on Parler social platform, investigating the proliferation of hate speech on the platform, both on the post level and on the user level. We identify three distinct groups of users (hate mongers, regular users and hate flirts) and show significant differences between them in terms of language, emotion, activity level and role in the network. We further compare our result to the hateful dynamics observed in the Gab platform.

Contribution Our contribution in this paper is fourfold: (i) We compare an array of state-of-the-art algorithms for hate detection, showing they all fail to accurately identify nuanced and novel manifestations of hate speech found on Parler, (ii) We share the first annotated Parler dataset, containing 10K Parler posts, each post labeled by the level of hate it conveys, (iii) We fine-tune a BERT-based classifier to achieve accurate classification, and modify DeGroot's diffusion model (Golub and Jackson 2010) in order to allow analysis on the platform level, and finally (iv) We provide the first large scale analysis of the proliferation of hate in Parler and compare it to the user dynamics in Gab.

The remainder of the paper is organized as follows: Section 2 provides a brief review of the relevant literature. A detailed description of the datasets and the annotation procedure are given in Section 3. In Section 4 we present the computational methods we use for the post and user level classification, and results follow in Section 5. A detailed analysis of hate levels and user propensity for hate speech in Parler and Gab is provided in Section 6. Finally, Section 7 offers some discussion regarding some of the observations, including ethical considerations.

A growing body of research studies the magnitude and the different manifestations of hate speech in social media (Knuttila 2011; Chandrasekharan et al. 2017; Zannettou et al. 2018; Zampieri et al. 2020; Ranasinghe and Zampieri 2020) , among others. Here, we present an overview of the current literature through three different perspectives: (i) The detection of hate speech on the post level, (ii) The detection of hate-promoting users, and (iii) The characterization of hate speech on the platform level.

Post-level classification Most previous works address the detection of hate in textual form. Keywords and sentence structure in Twitter and Whisper were used in (Mondal, Silva, and Benevenuto 2017; Saleem et al. 2017) , demonstrating the limitations of a lexical approach.

The use of code words, ambiguity and dog-whistling, and the challenges they introduce to text-based models were studied by (Davidson et al. 2017; Ribeiro et al. 2017; Arviv, Hanouna, and Tsur 2021) . The detection of implicit forms of hate speech is addressed by Magu, Joshi, and Luo (2017) which detects the use of hate code words (e.g., google, skype, bing and skittle to refer to Black people, Jews, Chinese, and Muslims, respectively) using SVM classifier based on bag-of-words feature vectors. ElSherief et al. (2021) introduced a benchmark corpus of 22.5K tweets to study implicit hate speech. The authors presented baseline results over this dataset using Jigsaw Perspective 5 , SVM, and different variants of BERT (Devlin et al. 2018) .

The use of demographic features such as gender and location in the detection of hate speech is explored by Waseem and Hovy (2016) , and user meta features, e.g., account age, posts per day, number of followers/friends, are used by Ribeiro et al. (2017) .

Computational methods for the detection of hate speech and abusive language range from SVM and logistic regression (Davidson et al. 2017; Waseem and Hovy 2016; Nobata et al. 2016; Magu, Joshi, and Luo 2017) , to neural architectures such as RNNs and CNNs (Zhang, Robinson, and Tepper 2016; Gambäck and Sikdar 2017; Del Vigna12 et al. 2017; Park and Fung 2017) . Transformer-based architectures achieved significant improvements, see (Mozafari, Farahbakhsh, and Crespi 2019; Aluru et al. 2020; Samghabadi et al. 2020; Salminen et al. 2020; Qian et al. 2021; Kennedy et al. 2020; Arviv, Hanouna, and Tsur 2021) , among others. In an effort to mitigate the need for extensive annotation some works use transformers to generate more samples, e.g., (Vidgen et al. 2020b; Minkov 2020, 2021) . Zhou et al. (2021) integrate features from external resources to support the model performance.

In order to account for the sometimes elusive and coded language and the unfortunate variety of targeted groups (Schmidt and Wiegand 2017; Ross et al. 2017 ), a set of functional test was suggested by Röttger et al. (2020) , allowing an quick evaluation of hate-detection models.

Classification of hate users Characterizing accounts that are instrumental in the propagation of hate and violence is gaining interest from the research community and industry alike, whether in order to better understand the phenomena or in order to suspend major perpetrators instead of removing sporadic content. Detection and characterization of hateful Twitter and Gab users was tackled by Ribeiro et al. (2018) ; Mathew et al. (2018 Mathew et al. ( , 2019 ; Arviv, Hanouna, and Tsur (2021) , among others. An annotated dataset of a few hundreds of Twitter users was released as part of a shared task in CLEF 2021, see (Bevendorff et al. 2021) for an overview of the data and the submissions. An annotated dataset of Twitter users using the ambiguous ((())) ('echo') symbol was released by Arviv, Hanouna, and Tsur (2021) .

Hate speech on Parler and Gab While most prior work focus on the manifestations of hate in the mainstream platforms, a number of works do address alternative platforms such as Gab and Parler. Two annotated Gab datasets were introduced by Kennedy et al. (2018) and by Qian et al. (2019) . We use these datasets in this work as we compare Parler to Gab.

Focusing on users, rather than posts, Das et al. (2021) experiment with an array of models for hate users classification. Lima et al. (2018) aims to understand what users join the platform and what kind of content they share, while Jasser et al. (2021) conduct a qualitative analysis studying Gab's platform norms, given the lack of moderation. Gallacher and Bright (2021) explore whether users seek out Gab in order to express hate, or that the toxic attitude is adopted after joining the platform. The spread of hate speech and the diffusion dynamics of the content posted by hateful and non-hateful Gab users is modeled by Mathew et al. (2019) and Mathew et al. (2020) .

Parler, launched in August 2018 and experiencing its impressive expansion of user base from late in 2020, is only beginning to draw the attention of the research community. Early works analysed the language in Parler in several aspects such as QAnon content (Sipka, Hannak, and Urman 2021) , COVID-19 vaccine (Baines, Ittefaq, and Ab-wao 2021) , and the 2021 Capitol riots (Esser 2021) . The first dataset of Parler messages was introduced by Aliapoulios et al. (2021) , along with a basic statistical analysis of the data, e.g., the number of posts and the number of registered users per month, along with the most popular tokens, bigrams, and hashtags in the data. Ward (2021) used a list of predefined keywords (hate terms), assessing the level of hate-speech on the platform.

Our work differs from these works in a number of fundamental aspects. First, we combine textual and social (network) signals in order to detect both hateful posts and hatepromoting accounts. Second, We suggest models that rely on state-of-the-art neural architectures and computational methods, while previous work detects hate speech by matching a fixed set of keywords from a predefined list of hate terms. Furthermore, we provide a thorough analysis of the applicability of different algorithms, trained and fine-tuned on various datasets and tasks. Third, we provide a broader context to our analysis of the proliferation of hate in Parler, as we compare and contrast it to trends observed on Gab.

In this section we describe the datasets used for this workstarting with a general overview of the platforms, then providing a detailed description of the datasets and the annotation procedure.

Parler Alluding to the french verb 'to speek', Parler was launched on August 2018. The platform brands itself as "The World's Town Square" a place in which users can "Speak freely and express yourself openly, without fear of being "deplatforme" for your views." 6 .

Parler users post texts (called parlays) of up to 1,000 characters. Users can reply to parlays and to previous replies. Parler supports a reposting mechanism similar to Twitters retweets (referred to as 'echos'). Throughout this paper we refer to echo posts as reposts, not to confuse with the ((())) (echo) hate symbol.

Parler's official guidelines 7 explicitly allow "trolling" and "not-safe-for-work" content, include only two "Principles" prohibiting "unlawful acts", citing "Obvious examples include: child sexual abuse material, content posted by or on behalf of terrorist organizations, intellectual property theft." and spamming.

By January 2021, 13.25M users have joined Parler and its mobile application was the most downloaded app in Apple's App Store. This growth is attributed to celebrities and political figures promoting the platform (see Section 1) and the stricter moderation enforced by Facebook and Twitter, culminating with the suspension of the @realDonaldTrump account from Twitter and Facebook.

Gab Gab, launched on August 2016, was created as an alternative to Twitter and it positioning itself as putting "people and free speech first" and welcoming users suspended from other social networks (Zannettou et al. 2018) . Gab posts (called gabs) are limited to 300-characters, and users can repost, quote or reply to previously created gabs. Gab permits pornographic and obscene content, as long as it is labeled NSFW (Not-Safe-For-Work). Previous research finds that Gab is a politically oriented system -while many users who use the platform are extremists, the majority of users are Caucasians-conservatives-males (Lima et al. 2018) . For more details about gab usage, users and manifestations of hate see references at Section 2.

We use the Parler and Gab datasets published by Aliapoulios et al. (2021) and Zannettou et al. (2018) , respectively. The Parler dataset is unlabeled, therefore annotation is required. We describe the annotation procedure and label statistics in Section 3.3.

Both datasets include posts and users' meta data, though the Parler dataset is richer, containing more attributes such as registration time and total number of likes. Each of the datasets is composed of millions of posts and replies, see Table 2 . The Parler dataset is bigger, containing more posts and more users, however, on average, Gab users post more content per user. We note that there is no temporal overlap between the two datasets. We discuss this point and its impact on the analysis and comparison in Section 7.

We use three Gab annotated datasets which are all sampled from the unlabeled Gab corpus we use: (i) The Gab Hate Corpus -27.5K Gab posts published by Kennedy et al. (2018) , (ii) 9.5K Gab posts published by Qian et al. (2019) , and (iii) 5K posts published by (Arviv, Hanouna, and Tsur 2021) . In total, we collect a corpus of 42.1K annotated Gab posts. 7.7K (18.4%) of the posts are tagged as hateful.

Hate speech takes different forms in different social platforms (Wiegand, Ruppenhofer, and Kleinbauer 2019) and across time (Florio et al. 2020) . It is often implicit , targeting a variety of groups. Consequently, transfer learning remains a challenge for hate-speech detection, and annotated Parler data is needed in order to achieve accurate classification. This challenges and the significant improvements in performance achieved by proper fine-tuning are demonstrated through extensive experimentation, see Section 4.1. In the remainder of this section we describe the annotation procedure and the annotated dataset we use.

The annotation task was designed as follows: 10K posts were sampled from the full Parler corpus. All posts met the following criteria: (i) Primary language is English; (ii) A post should be at least 10 characters long; (iii) The post does not contain a URL; and (iv) The post is neither a repost nor a comment.

The 10K annotated posts were not randomly selected from the Parler corpus. A random selection of posts would have led to an extremely imbalanced dataset as most of the posts do not contain hate speech. Hence, we opt to stratified sampling. This sampling process relies on an approximation of the likelihood of each post to include hateful content. We used a pretrained hate speech prediction model to approximate this likelihood.

Annotation was done by 112 student (more than half of them are graduate students), who were provided detailed guidelines and training involving the various types of hate speech, the elusiveness of hate expressions using coded language, how to detect it, and a number of examples of different types. Each of the annotators was prompted with a list of 300 posts and had to assign each with a Lickert score ranging from 1 (not hate) to 5 (extreme or explicit hate). We provided annotators only with the textual content of the post. Each of the 10K posts was annotated by three annotators. Annotators presented a satisfying agreement level of 72% and a Cohen's Kappa of 0.44. Labels of posts with a low agreement level 8 were ignored (∼7% of the annotated posts). We define a post as hateful (non-hateful) if its average score is higher (lower) than three. We omit posts with an average score of exactly three. Accordingly, 3224 of the 10K posts (32.8%) were labeled as hateful and 6053 (59.8%) as non-hateful.

We make this annotated corpus available in the project's repository 9 -the first public annotated corpus of Parler.

In this work we are interested in the detection of hate, both on the post level and the account level. Our interest in the post level classification is twofold. Given an accurate classifier, we can: (a) Approximate the hate degree in different aggregation levels -e.g., over all social network, and per user, and (b) Use the post-level predictions to support training a user level classifier. A review of the various post level classifiers is provided in Section 4.1 and our modifications to a diffusion-based model for user classification are presented in Section 4.2. Ethical considerations related to user classification are discussed at the end of Section 7.

We fine-tune the DistilBERT (Sanh et al. 2019 ) transformer on each of the datasets, obtaining two fine-tuned models (referred to as Our-FT BERT). We compare the performance of Figure 1 : An illustration of the diffusion model over three nodes. Self loops represent the total number of posts per node. In step (a) we build the repost network and assign each node with an initial belief -seed hate mongers with a value of one and others with a value of zero. In steps (b) and (c) we convert the network to a belief network -reversing the edges' direction and normalizing their weight. In step (d) we run the diffusion process and get a belief score per node, which is indicated in the graph by the darkness of each node. the models on the respective datasets against four competitive models:

1. Jigsaw Perspective: A widely used commercial model to detect hate and toxic content, developed by Google. Jigsaw was found to perform well in an array of tasks related to hate-speech detection (Röttger et al. 2020) . Jigsaw implementation is not public and the service is provided as a black-box through an online API 10 . HateBase is a multilanguage vocabulary of hate terms that is maintained on order to assist in content moderation and research. We use 68 explicit hate terms that were used in prior works Mathew et al. (2018 Mathew et al. ( , 2019 . These terms were mainly selected from HateBase's English lexicon and is composed only of explicit hate terms like 'kike' (slur targeting Jews, see post #2 in Table 1 ), 'paki' (slur against Muslims, especially with Pakistani roots), and 'cunt' (see post #1 in Table 1 ).

Ideally, an account should be classified as a hate account based on the content it posts (or likes). However, this seemingly straight forward approach is severely limited by ambiguity, vagueness, dog-whistling, and emerging idioms and racial slurs. For example, defining a threshold of k hateful posts is still not well defined. How explicit these k posts should be? would 2k less explicit posts make the cut? is one post enough to declare a user a hate-monger? Moreover, defining a threshold does not account for networked aspect of the data and the fact that "birds of a feather flock together" (Himelboim, McCreery, and Smith 2013) .

In order to leverage the network structure, we view each platform as a social network with users as nodes and reposts as directed edges. Edges are weighted to reflect levels of engagement, as illustrated in Figure 1 We build on the diffusion-based approached for the detection of hate mongers, proposed by Mathew et al. (2019) , modifying it in order to achieve a more accurate classification. The basic diffusion-based classification is achieved in two stages: (a) Identifying a seed group of hate mongers.

(b) Applying a diffusion model over the social network. We use the DeGroot's hate diffusion model (Golub and Jackson 2010) which outputs an estimated belief value (i.e., "hate") per user, over the [0,1] range. A toy example of the diffusion process is illustrated in Figure 1 . In our experiments we set the number of diffusion iterations to three. One clear advantage of this approach over fully supervised methods is that it does not require a large dataset annotated on the user level.

We modified the diffusion model used by Ribeiro et al. (2018) and Mathew et al. (2019) in two ways: (i) Seed definition. Instead of taking a lexical approach in order to identify users posting more than k hateful posts, we use our fine-tuned Transformers. We argue that fine-tuning the classifiers for each social network significantly improves the classification on the post level (as demonstrated in Section 5.1), and ultimately, improves the performance of the diffusion model; and (ii) Hateful users definition. In the original diffusion process, hate (as well as "not-hate") labels are diffused through the network. This way, seed hate mongers may end with a low belief (hate) score, which in turn propagates to their neighbours. However, seed users were chosen due to the fact that they post a significant number of undoubtedly hateful posts. Fixing the hate score of these users results in a more accurate labeling of the accounts in the network.

We use the annotated corpora (see Section 3.3) to fine-tune the pretrained Transformer on each social platform, splitting the labeled data to train (60%), validation (20%), and test (20%) sets. The precision-recall curves of the Parler and Gab models are presented in Figure 2 . Our fine-tuned models significantly outperforms the other models in both datasets. We wish to point out that while the popular keyword base approach (HateBase) achieves a high precision and a moderate recall on the Gab data, outperforming all Transformer models except the platform fine-tuned ones, it collapses in both measures on the newer Parler dataset. These results revalidate the limitations of lexical approaches, and of neural methods that are not fine-tuned for the specific dataset (even though they were fine-tuned for a similar task -hate speech detection in another microblogging platform).

As described in Section 4.2, in order to classify accounts we use a diffusion model. The diffusion process is seeded with a set of hateful accounts. The choice of seed accounts involves the following steps: (i) After establishing the accuracy of the fine-tuned models (Section 5.1) we use these models to label all the posts in the respective dataests. (ii) Opting for a conservative assignment of seed users, we consider only posts with hate score (likelihood) over 0.95 (0.9) in the Parler (Gab) dataset to be hateful. Finally, (iii) Users posting 10 or more hateful posts are labeled as seed accounts. We take the conservative approach in steps (ii) and (iii) in order to control the often noisy diffusion process.

Simulating the modified diffusion process described in Section 4.2 we obtain a hate score per user. For analysis purposes we divide users to three distinct groups -hate mongers (denoted HM), composed of the users making the top quartile of hate scores; normal users (denoted N) making the bottom quartile; the rest of the users (denoted HM) suspected as "flirting" with hate mongers and hate dissemination. Users with a low level of activity (less than five posts or users joining the network in the last 60 days) were not considered 11 . The distribution of active users by type is presented in Fig 3. Evaluation of the diffusion model A user-level annotated dataset of 798 Gab users was shared by Das et al. (2021) . We use this dataset to validate the performance of the diffusion models -both the standard and our modified models (see Section 4.2). We find our modified model to outperform the standard models, achieving precision/recall/F1-scores of 0.9/0.54/0.678, comparing to of 0.95/0.34/0.5. Therefore, results and analysis in the remainder of the paper are based on the modified diffusion model.

In this section we provide a comprehensive analysis of the propensity for hate speech on Parler and Gab

Taking our conservative approach, we find that hate posts are more frequent in Parler (3.29%) than in Gab (2.13%). However, we find that 13.95% of Parler users share at least one hateful post -significantly lower number compared to Gab (18.58%). We find that 65.5% of the hate content in Parler is posted as a reply to other parlays. This reflects a significant over-representation of replies compared with full corpus distribution (46.2% of posts are replies, see Table 2 ). Similarly, 38.9% of the hate content on Gab are replies.

We provide an analysis of the characteristics of the HM, HM and N accounts on an array of attributes, ranging from activity levels to centrality, sentiment and the emotions they convey.

Activity level Activity levels are compared via four features -number of posts, replies, reposts, and users' age (measured in days).

HM are the most active user group in both platforms across all activity types (see Figure 4 ). We find that the HM users have similar characteristics in both platforms -they share less content than the HM users, repost more content than the N group, and their tendency to reply is lower compared to the N users. User Age (days from account creation to the most recent post in the data), is an exceptional feature. We observed only insignificant differences between the three user groups. This observation holds for both platforms. However, collapsing the groups -we do find a significant difference between the two platforms. Gab users are "older" with an average age of 323.9 compared to 189.6 of the Parler users. We hypothesize that the difference is a result of the way both platform evolve over time, given the unfolding of events driving users to these platforms (see Sections 1 and 3.1).

We quantify the popularity level of users based on the number of followers they have. Figure 5 presents numbers for both platforms. On both platforms hate mongers (HM) are significantly more popular compared to all other user groups. In Parler, the median number of followers is 121 compared to 15 and 12 of HM and N, respectively. The same holds for Gab -a median value of 160 for HM users compared to 43 and 41 of the other two user groups. Interestingly, although Parler is a much larger social platform (mainly in terms of registered users, see Section 3 and Table 2 ) we do not see a significant higher number of followers in Parler. Moreover, when calculating the number of followers over the whole population, the median in Gab is three times higher -48 vs. 16. Engagement level is measured by the number of followees each account has (the number of accounts a user follows). We find that hate mongers are highly engaged in both platforms compared to other user groups. In Parler, the median followees number of HM users is 106 -significantly higher than 46 and 36 median values of the HM and N users, respectively. A similar pattern holds for Gab.

Account's self description Analogue to the account's description in Twitter, Parler users can provide a short descriptive/biographical text to appear next to the user's avatar. For example, the biography that is associated with a specific Parler user is: "Conservative banned by mainstream social media outlets for calling the leftists out for what they really are! Been awake for YEARS! #trump2020 #maga". We use this content to further assess users commitment to the platform, assuming more engaged users are, the more likely they add the description to their profile. We find that while only 35.8% of the N users use the biography field, 59.6% of the HM users provide the description in their profile. We also find that the average (median) biographical text length of HM users is 128.6 (134) -considerably longer, compared to HM and N users who included the description in their profile, with an average (median) of 99.4 (90) and 94.6 (84), respectively.

Social Structure We further analyze the differences between Parler and Gab platforms over the different user groups from a social network analysis (SNA) perspective, based on the reposts network. Table 3 provides an overview of a number of centrality measures. The HM users have a significantly higher values in all measures in both platforms. Interestingly, the full order between the different user groups is kept only for the 'betweeness' centrality, while other centrality measures a less stable comparing the HM and N groups.

Analysing the degree distribution of users provides an interesting difference between the platforms. In line with the numbers in Table 3 , HM users have the most distinctive distribution in both Parler and Gab. However, while the HM and the N group distributions are inseparable in Gab, the Parler user groups have distinct distributions (see Figure 6 ). These distributions highlight the distinctiveness of the location of HM users in the network, as well the role of the HM compared to N users.

We compare the sentiment expressed and the emotions conveyed by different user groups. We use pretrained BERT models for both the sentiment 12 and emotion 13 predictions. Results are presented in Table 4 (a) Parler (b) Gab Figure 6 : Social networks degree distribution. We present the indegree distributions. Network is based on reposts. p(k) (y-axis) is the probability value per a each node's degree (x-axis). We use a log-scale over both the axis.

at the Parler users, we find a small though significant (pvalue < 10 −3 ) tendency of HM to express a more negative sentiment. The same holds for Gab, although the sentiment expressed by HM is closer to the sentiment of the HM users, rather to that of the N users. Aggregating the emotion predictions, we find that HM users tend to convey more Anger and Sadness than the other groups. This observation holds for both Parler and Gab, although Anger is more prominent. 

Seed hate mongers One design choice critical to this work, affecting the user-level analysis, is the way we define seed hate mongers (see Section 4.2). Previous works used lexicon based solutions. We decided to use our post level classification model which significantly outperformed other alternatives. However, both solutions rely on counting the number of hate posts per user.This binary definition lacks the sensitivity to mark hate mongers in a more nuanced way. Alternative methods to mark seed hate mongers should be considered in future work. Two possible directions are summation of the probabilities yielded by the hate post classifier, and averaging the number of hate posts per user are two optional alternatives. However, we wish to stress that opting for a conservative labeling of seed users achieves a cleaner diffusion process -a process that is usually prone to noise.

Parler users In this work, we make use of a Parler dataset introduces by Aliapoulios et al. (2021) . One limitation of this dataset is that it includes only part of the Parler full corpus. However, data were not sampled at random -the authors retrieved data based on users' identity (i.e., all data for 4.08M users out of 13.25M), providing a decent coverage of a significant part of the network.

Time span Given that we provide a comparison between trends in Parler and Gab, it is important to note the datasets span different and non-overlapping time-frames (see Table  2 ). Therefor, the comparison we provide should be read cautiously. We do note, however, that each of the datasets was crawled from the early days of the social platform and spans over a similar range of time (17 months). Moreover, the time disparity between the dataset could be considered as an advantage -allowing to examine the generalization performance of hate speech models, as we report in Section 5.1.

Analysing and modeling hate speech in a new social platform such as Parler is of great importance. However, classifying users as hate mongers, based on the output of an algorithm, may result in marking users falsely (which may result in suspension or other measures taken against them). While we always opted for a conservative approach, as well as focusing on aggregate measures characterizing the trends of a platform, we note that user labeling should be used in a careful manner, ideally involving a 'man-in-the-loop'. Considering the annotation task -the annotation process did not include any information about the identity of the users. In addition, we warned our human annotators about the possible inappropriate content of the posts.

To the best of our knowledge, we present the first large-scale computational analysis of hate speech on Parler, and provide a comparison to trends observed in the Gab platform.

We annotate and share a the first Parler dataset, containing 10K posts labeled by the level of hate they convey. We used this dataset to fine-tune a transformer model to be used to mark a seed set of users in a diffusion model, resulting in user-level classification. We find significant differences between hate mongers (HM) and other user groups: HM represent only 16.1% and 10% of the active users in Parler and Gab respectively. However, they create 41.23% of the content in Parler and 71.38% of the content in Gab. We find that HM are show higher engagement and they have significantly more followers and followees. Other differences are manifested through the sentiment level expressed and the emotions conveyed.

ADL. 2020. ANTISEMITIC INCIDENTS HIT ALL-TIME HIGH IN 2019

The Muslim Question in Australia: Islamophobia and Muslim Alienation

An Early Look at the Parler Online Social Network

Deep learning models for multilingual hate speech detection

Predicting Anti-Asian Hateful Users on Twitter during COVID-19

It'sa Thin Line Between Love and Hate: Using the Echo in Modeling Dynamics of Racist Online Communities

# Scamdemic,# Plandemic, or# Scaredemic: What Parler Social Media Platform Tells Us about COVID-19 Vaccine

Tweeteval: Unified benchmark and comparative evaluation for tweet classification

Overview of PAN 2021: Authorship Verification,Profiling Hate Speech Spreaders on Twitter,and Style Change Detection

A Virus Has No Religion": Analyzing Islamophobia on Twitter During the COVID-19 Outbreak

You can't stay here: The efficacy of reddit's 2015 ban examined through hate speech

You too Brutus! Trapping Hateful Users in Social Media: Challenges, Solutions & Insights

Automated hate speech detection and the problem of offensive language

Hate me, hate me not: Hate speech detection on facebook

Bert: Pre-training of deep bidirectional transformers for language understanding

Anti-Muslim hate crimes increase fivefold since London Bridge attacks. The Guardian 7

The effect of President Trump's election on hate crimes. Available at SSRN 3102652

Latent Hatred: A Benchmark for Understanding Implicit Hate Speech

Refugees welcome? Understanding the regional heterogeneity of anti-foreigner hate crimes in Germany

How does the language of corpora from radicalized communities discovered on Parler compare to online conversations on Twitter regarding the 2021 Capitol riots and election fraud?

Time of your hate: The challenge of time in hate speech detection on social media

Hate Contagion: Measuring the spread and trajectory of hate on social media

Naive learning in social networks and the wisdom of crowds

ASPI explains: 8chan

Birds of a feather tweet together: Integrating network and content analyses to examine cross-ideology exposure on Twitter

Understanding hate crimes against immigrants: C onsiderations for future research

Welcome to# GabFam': Far-right virtual community on Gab. New Media & Society

The Gab Hate Corpus: A collection of 27k posts annotated for hate speech

Contextualizing hate speech classifiers with posthoc explanation

User unknown: 4chan, anonymity and contingency

Report to the nation: hate crimes rise in US cities and counties in time of division and foreign interference

Inside the rightleaning echo chambers: Characterizing gab, an unmoderated social system

Roberta: A robustly optimized bert pretraining approach

Detecting the hate code on social media

Violence begetting violence: An examination of extremist content on deep Web social networks

Spread of hate speech in online social media

Hate begets hate: A temporal study of hate speech

Analyzing the hate and counter speech accounts on twitter

A measurement study of hate speech in social media

A BERT-based transfer learning approach for hate speech detection in online social media

Alt-right pipeline: Individual journeys to extremism online

Abusive language detection in online user content

Retraction: Understanding Islamophobia in Asia: The Cases of Myanmar and Malaysia

One-step and two-step classification for abusive language detection on twitter

Planting Hate Speech to Harvest Hatred: How Does Political Hate Speech Fuel Hate Crimes in Turkey?

Examining The U.S. Capitol Attack: a review of the security planning and response failures

A benchmark dataset for learning to intervene in online hate speech

Lifelong Learning of Hate Speech Classification on Social Media

Multilingual offensive language identification with cross-lingual embeddings

Like Sheep Among Wolves": Characterizing Hateful Users on Twitter

Measuring the reliability of hate speech annotations: The case of the european refugee crisis

Hatecheck: Functional tests for hate speech detection models

A web of hate: Tackling hateful speech in online social spaces

Developing an online hate classifier for multiple social media platforms

Aggression and misogyny detection using bert: A multi-task approach

Dis-tilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter

A survey on hate speech detection using natural language processing

Comparing the Language of QAnon-related content on Parler, Gab, and Twitter

The long history of Islam as a collective "other" of the west and the rise of Islamophobia in the US after Trump

ASPI explains: 8chan

Hatebase: Online database of hate speech. The Sentinal Project

Detecting East Asian prejudice on social media

Learning from the worst: Dynamically generated datasets to improve online hate detection

Parlez-vous le hate?: Examining topics and hate speech in the alternative social network Parler. Master's thesis

Hateful symbols or hateful people? predictive features for hate speech detection on twitter

Detection of abusive language: the problem of biased datasets

Towards hate speech detection at large via deep generative modeling

Fight Fire with Fire: Fine-tuning Hate Detectors using Large Samples of Generated Hate Speech

SemEval-2020 task 12: Multilingual offensive language identification in social media

What is gab: A bastion of free speech or an alt-right echo chamber

A quantitative approach to understanding online antisemitism

Hate speech detection using a convolution-LSTM based deep neural network

Hate Speech Detection Based on Sentiment Knowledge Sharing

Racism is a virus: Anti-asian hate and counterhate in social media during the covid-19 crisis