Social Processes in the Intensification of Online Hate: The Effects of Verbal Replies to Anti-Muslim and Anti-Jewish Posts Following 7 October 2023

Abstract

Online hate messaging targeting Muslims and Jews increased dramatically following Hamas’s attack on Israelis on 7 October 2023 and Israel’s military response in Gaza. This study examined anti-Jewish and anti-Muslim hate posts on X.com and the verbal replies, Likes, and reposts they acquired over the following month. It tests a theory explaining the propagation of hate messages in social media based on the social approval posters garner from other users. The analysis involved replies to 6388 anti-Muslim or anti-Jewish hate posts in terms of their semantic convergence or divergence with the content of original posts. No differences between patterns of anti-Jewish and anti-Muslim discourse arose. Convergent replies to one’s hate posts led individuals to post more hatefully in their next post, and more quickly. Likes also accelerated hate posting, while reposts decelerated them. Divergent replies led to less hateful and slower subsequent hate postings. Conclusions address implications for the social approval theory of online hate, and the relative influence of verbal replies due to their costliness.

Keywords

social media hate speech anti-Semitism Islamophobia social approval linguistic convergence/divergence

The attack on Israel by Hamas on 7 October 2023 and Israel’s violent incursion into Gaza thereafter led to a dramatic increase in both anti-Jewish and anti-Islamic hate messaging on social media (Ortiz, 2024). On Facebook, for instance, the posts in groups that were critical of both Israelis/Jews and Palestinians/Arabs/Muslims shrank in proportion to the number of posts targeting one respective side or the other (Nefriana et al., 2024). The virulent and debasing postings following these tragedies offer an unfortunately active field for the study of online hate.

The spread of hate messaging on social media had already become a global issue. Concerns over it include the psychological harm to hate victims (see Tong, 2024), desensitization and a reduction of empathy among observers (Pluta et al., 2023; Soral et al., 2018), and its larger potential to incite intolerance and violence (Benesch, 2013). Online hate is different from traditional hate speech in many ways (Brown, 2018), and it is the focus of an increasing level of international institutional and academic concern (Rice, 2024). Online hate messages “convey intolerance or aversion towards specific groups, such as ethnic, religious, sexual, or gender minorities and immigrants. It may include derogatory comments, threats, and incitements to violence” (Oliveira et al., 2023, p. 1), including “‘discriminatory’ (biased, bigoted, or intolerant) or ‘pejorative’ (prejudiced, contemptuous, or demeaning) language” (United Nations, n.d.). Among the most frequent foci are Jews and Muslims.

Recent theorizing over the propagation of online hate makes central the role of social interactions among hate posters that social media facilitate (Walther, 2024a). Particularly, a social approval theory of online hate argues that people create and deliver hate messages online primarily to garner signals of admiration and approval, in a variety of forms, from like-minded others and appreciative onlookers. The reception of that approval increases the frequency and extremity of individuals’ subsequent hate expressions, according to the theory (Walther, 2024b).

Extant research examining social approval of hate messages, in the form of single-click Likes and reposts (retweets), shows that reception of greater levels of these reinforcers leads to increased toxicity and outrage in one’s subsequent hate messages (Brady et al., 2021; Frimer et al., 2023; Jiang et al., 2023). However, Likes and reposts are not the only form of engagement people garner from their social media posts that have the potential to exacerbate hate.

The verbiage of reply messages also potentially provides social approval to encourage hate messaging. Research in other domains shows that comments one receives expressing agreement or appreciation of one’s postings increase message posters’ satisfaction (Sannon et al., 2017). The approval value of message-consistent conversational responses to hate messages potentially also promotes and elevates hate messaging behavior over time. But comments on one’s posts are not consistently approving or appreciative. Replies may alternatively affirm, undermine, or even counter-attack the hate message to which they respond (Weigel & Gitomer, 2024), making replies less superficially interpretable than Likes and reposts insofar as their social approval value goes. Despite advances in automated analyses of sentiment and wording (for review, see Dehghani & Boyd, 2022), verbal messages are notoriously difficult to characterize automatically with respect to their semantic and attitudinal relationship to prior utterances. If analyses could detect the similarity of other users’ replies to hate messages, it is possible to test the effect of those verbal replies to an original hate message on the original posters’ subsequent posting.

This study examines how the semantic similarity of replies to hate messages on X.com, and other engagement signals, influences the hatefulness of subsequent messages in the context of activity immediately following 7 October 2023, in Israel and Gaza. It provides a nuanced and inclusive view of the degree to which social processes among hate posters and their audience promulgate online hate.

Social Approval Theory

Most research about online hate presumes that those who generate it manifest their internal animosity toward the targets of the message and the desire to harm that person or group (e.g., Zhang, 2023), and/or that individual difference characteristics propel it (Bührer et al., 2024). The social approval theory of online hate side-steps this possibility and focuses instead on social processes among hate posters themselves, and their sympathetic audience, that facilitate additional and increasingly hateful messages. It asserts that hate posters’ primary audience is frequently not the targets implicated in those messages, but one another. Their primary incentive for hate posting lies not in generating Schadenfreude toward its victims. Rather, their incentive may lie in garnering the signals of adulation that they draw from others who appear to agree with and/or appreciate their messages. This is the central assumption of the theory, leading to the hypothesis that the more one receives social approval for one’s hate message, the more frequently and more hatefully one posts thereafter.

Previous empirical studies suggest some support for the hypothesis, but their data were limited to the effects of Likes and reposts, and not the content of verbal replies to hate. In an analysis of Twitter users who posted messages deriding immigrants, Jiang et al. (2023) found that when hate posters received a significantly greater number of Likes or reposts than they usually received, their toxicity subsequently increased. When Likes and reposts were fewer than usual, a significant decrease in toxicity followed. Elsewhere, an analysis of elected US congressional representatives’ Twitter activity over a decade not only found a 23% increase in incivility over time but also that the more Likes and reposts that uncivil posts received, the more uncivil were Tweets thereafter (Frimer et al., 2023). Other studies as well, including both big data analyses and field experiments, have demonstrated similar effects of Likes and Upvotes by observers on posters’ toxicity, outrage, and hatefulness (Brady et al., 2021; Shmargad et al., 2022).

One effort to examine the effect of the frequency of verbal replies and quote tweets on hate posters’ subsequent hate messaging found no effect (Jiang et al., 2023). Its authors speculated that it may not be the case that no effects exist, but rather that the effects of verbal responses systematically depend on the relationship of the reply content to the position of the original message. Thumbs-up and heart responses are rather unambiguous affirmation signals. Verbal replies, however, may affirm, negate, converge, or diverge from an original message’s content. If a reply message pragmatically resembles or affirms the original hate message, it may function as an approval signal and reflect the patterns predicted in the hypothesis above. If the reply is dissimilar or negates the original hate messages, it may function as a disapproval message and also affect hate messaging.

This proposition is operationally not as simple as it may first look. It requires the determination of what is a similar or affirming message and what is a dissimilar or negating message. Determination of a message’s “sentiment” is not useful. Many hate messages and their replies express negative sentiments, which is unsurprising given the definitions and typical content of such messages. Nevertheless, negative replies may have a positive complementary functional relationship with the original message to which they reply. Consider the couplet, “Those people should die immediately,” followed by a reply, “No, those people should die a slow painful death.” Both messages’ sentiment may be classified as negative, but the reply pragmatically converges as an affirmation with respect to the original message. An alternative reply, “I agree,” has similarly affirming pragmatic value, but its sentiment may be positive. For this reason, labels of positive versus negative and sentiment analysis are insufficient to detect the functional relationship of such messages to one another.

Convergence

The analysis of linguistic convergence overcomes these analytic conundrums. Convergence, simply defined, is “a similar way of speaking with other members of the same group” (Rosen & Dale, 2024, p. 3140). Without necessarily copying or imitating the same words, “interlocutors must construct novel utterances serving to advance an ongoing dialog while simultaneously signaling their social proximity to their fellow discursive partners” (p. 3141).

Previous research related to online hate has also explored linguistic convergence. Studies examining the lexical changes in the White Supremacist discussions on Stormfront found that new members gradually adopted and reflected the nomenclature of long-time members over time; those who did not often left the site after a short while (P. Törnberg & Törnberg, 2022). According to A. Törnberg and Törnberg (2024, p. 104), “members picked up different slang and vernacular that is particular to the community. . .The language changed from mainstream terms to community-specific vernacular and themes. . .embracing the White supremacist ideology of the forums.”

Convergence can signal not only ingroup similarity but interpersonal affirmation as well (Bradac et al., 1988). A broad literature supports the notion that as communicators’ nonverbal and verbal behavior becomes more reciprocal and similar, they experience affection and reinforcement, from the nonverbal “chameleon effect” in face-to-face communication (Chartrand & Bargh, 1999; see also Bavelas et al., 1988) to verbal mimicry (Kulesza et al., 2014) and “linguistic style matching” in person (e.g., Ireland et al., 2010) and online (Gonzales et al., 2010). Convergence in message exchanges has been studied through the concepts of “reciprocity, behavioral matching, mirroring, . . .and their opposites of compensation, complementarity, and divergence” (Burgoon et al., 1993, p. 296), as well as synchrony and its opposite, dissynchrony. Divergence and dissynchrony “involves the participant adjusting their behaviors to become more dissimilar to those of the partner. . .and divergence can be used to show that the participant does not desire affiliation with the partner” (Dunbar et al., 2023, p. 2).

Consistent with these principles, linguistic convergence by other users to an individual’s hate message is hypothesized to operate as an affirmation and a signal of social approval. It differs from Likes and single-click reinforcers, however, in that a Like may have value as a compliment, whereas the social approval value of linguistic convergence is that it forms a conversational complement. Divergence, on the other hand, is its antithesis. Divergence may signal disliking and disapproval. The social approval theory of online hate contends that approval of an individual’s hate message, in this case, convergence with it, predicts subsequently greater hate messaging from that individual. In the opposite manner, linguistic divergence from an individual’s hate message may constitute a signal of disapproval. Operationally, convergence has been operationalized regarding the level of entropy in a conversation. Based on information theoretic principles from Shannon (1948), entropy is lower (i.e., convergence is greater) when message exchanges are more similar, that is, the comments are more predictable within a given conversation or thread (Rosen & Dale, 2024).

The measurement of convergence differs from other measures of linguistic style matching in important ways. Unlike dictionary-based approaches to linguistic matching using tools like Linguistic Inquiry and Word Count (LIWC), convergence detection captures both similarities and “stylistic differences in lexical choices between individuals and groups [in situ] without needing to start from a hand-picked dictionary.” It relies on the BERT transformer model, by which it can create “a word vector for each word in a sentence by processing a series of weighted sums of the adjacent word vectors in a large, deep neural network” (Rosen & Dale, 2024, p. 3142). In this manner, analysis detects whether statements within a corpus reflect not only the similarity of the words across statements, but the degree to which the semantic usages of words suggest convergence or divergence. Its measurement has been demonstrated in previous analyses of social media discourse on Reddit. It showed that different, yet ideologically similar, subreddits demonstrated convergent semantic similarities despite their participants’ independence from one another across subreddit discussions (Rosen & Dale, 2024).

Hypotheses and Research Questions

The following hypotheses are posited with respect to both anti-Semitic and anti-Islamic messages and responses.

Hypothesis 1a (H1a): Convergence of replies to hate posts increases the hatefulness of posters’ next posts.

Hypothesis 2a (H2a): Convergence of replies to hate posts increases the frequency of posters’ next posts.

Previous research finds that above-average frequencies of Likes and reposts to hate messages on Twitter had a positive relationship with the toxicity of one’s subsequent posts (Jiang et al., 2023). It is important to establish whether that relationship reappears in this topical context and whether it coexists with, overrides, or moderates the hypothesized effects of convergence on subsequent hatefulness

Hypothesis 3 (H3): Single-click responses to hate posts, including (l) Likes and (r) reposts, increase the hatefulness of posters’ next posts.

Hypothesis 4 (H4): Single-click responses to hate posts, including (l) Likes and (r) reposts, increase the frequency of posters’ next posts.

The social approval theory also recognizes that insufficient social approval can lead hate posters to attempt greater extremity in their subsequent posts in an effort to acquire the degree of social approval they failed to acquire. There is a good reason to expect divergence in replies, since X.com offers no other way to signal disapproval, disagreement, or opposition other than by writing such responses. There are no “dislike” icons to append to someone’s post, as there are on other platforms such as Gab, YouTube, or others. Therefore, it is possible that while Likes and reposts signal approval, users relegate the expression of disapproval to replies. If such disapproving replies fail to satisfy the posters’ needs for approval, they may lead to more hateful, verbally aggressive responses in their next posting out of frustration or retaliation. This relationship is expressed in the following additional hypotheses:

Hypothesis 1b (H1b): Divergence of replies to hate posts increases the hatefulness of posters’ next posts.

Hypothesis 2b (H2b): Divergence of replies to hate posts increases the frequency of posters’ next posts.

Method

Focus Population

Hate against Jews and Muslims, Arabs, Palestinians, and Israelis, online and off, is a long-standing and rampant problem (for review, see Levin, 2015; Weimann, 2024). In the month following the inception of war between Hamas and Israel in October 2023, “Antisemitic content soared more than 919 percent on X. . . .Anti-Muslim hate speech on X jumped 422 percent on Oct. 7 and Oct. 8, and rose 297 percent over the next five days” (Frenkel & Myers, 2023). Although one study detected an increase in hate messages on every social media platform it examined, “hate was more prevalent on X (Twitter) than on other platforms” (Oboler et al., 2024, p. 3).¹ In the past, most such postings could be traced to White Supremacists who issued hate messages toward both Jews and Muslims (Hobbs et al., 2023). In this time interval, however, anti-Semitic and anti-Islamic hate postings were emitted from a variety of user types, even those who identify themselves as not being anti-Semitic or anti-Islamic (Kressel, 2024).

Sample

The sample of messages in this study comprised an original collection of anti-Semitic and anti-Islamic posts, replies, and engagement data, posted on X between 8 October 2023 and 4 November 2023. The process of data collection was two-fold. The first step required the identification of a number of X.com users who posted anti-Semitic and/or anti-Islamic hate messages during this period.

To identify qualified users, we queried the X.com platform to find users who had posted messages featuring any of 61 specific anti-Islamic words or 22 specific anti-Semitic words catalogued in the Weaponized Word (2024) database of hate terms. These keywords became part of a search script implemented through the Brandwatch (2024) platform to access, search, and retrieve relevant data from X. Brandwatch offers access to all historical and contemporaneous interaction records on X.com. The final data records include hate posters’ original posts and all engagements with their posts, including verbal replies, Likes, and reposts. (For similar keyword approaches and time spans using X and other platforms, see Nefriana et al., 2024; Oboler et al., 2024.)

To score hatefulness of the posts and replies, all messages were analyzed using the Roberta Hate Speech DynaBench classifier (HSD classifier: Vidgen et al., 2021). Originally developed by Facebook, the tool generates a score estimating the degree to which a text contains hate content. The HSD classifier was trained on data in which one or more the following features appeared in a given utterance: language that (1) is explicitly derogatory toward individuals belonging to a minority group, (2) expresses animosity toward a group identity, (3) threatens individuals or groups based on their minority status, (4) expresses support for hate proliferating organizations and known, named entities, and (5) is dehumanizing toward minorities or individuals holding a minority identity (Vidgen et al., 2021). The HSD classifier does not produce ratings for any of these individual features, but reports an aggregate rating as to whether an example meets any of these criteria. Although other models, such as Google’s Perspective API, are also widely used, such models generate sets of scores on oblique dimensions such as toxicity, extreme toxicity, and others, the relationships among which are unclear and less useful than a model trained specifically for hate detection.

The final sample contains 17,013 original posts and replies by 7137 individual users. Posts received a total of 10,718 replies (M = 1.68, SD = 3.05) and 65,331 Likes (M = 10.21, SD = 48.37). The hatefulness ratings of these messages ranged from .00013 to 1.0 (M = .406, SD = .334; see Figure 1). Most subsequent posts came relatively quickly, although intervals between posts varied somewhat. The amount of time between original posts ranged from 1 s to 1,677,717 s (19 days), M = 20,596.82 s (16.96 hr), SD = 61,068.54 s.

Figure 1.

Histogram of distribution of hate post scores calculated with the HSD classifier.

Computational Approach

The measurement of convergence/divergence scores involved the linguistic, convergence-entropy measurement (CEM) described in Rosen and Dale (2024). CEM measures how predictable the semantic meaning of each lexical unit is in some utterance x when compared to another utterance y using a mixture of Shannon entropy (Shannon 1948) and contextually informed word vectors (i.e., transformer neural networks; Devlin et al., 2019). CEM is operationalized as follows: The total entropy for all the lexical units comprising an utterance x is calculated by finding the Shannon entropy between the word vectors generated for x and the word vectors generated for the utterance y,

H [x; y] = - \sum_{i} P (E_{x i} | E_{y}) \log P (E_{x i} | E_{y})

(1)

where the probability of any individual word vector for the ith token in x (E_xi) is found by taking the cosine error (CoE) of E_xi and the word vector most similar to it from any of the word vectors in the utterance y (the set of word vectors for the jth token in the utterance y is denoted E_yj and the set of all tokens is indicated via E_y); see Rosen and Dale (2024). The CoE for this closest match is passed to a half-Gaussian function with µ = 0 and a scale σ to convert the raw CoE value to a probability. Analysis assessed convergence among a given tweet and all replies to that tweet, as well as replies to those replies.

Contextually aware word embeddings (i.e., E_x and E_y in the equations) were derived using the RoBERTa-base model (Liu et al., 2019) available via the HuggingFace transformer model library (Wolf et al., 2020). The final computation of convergence/divergence was then residualized based on the number of tokens in both the utterances x and y. When the residual was negative (i.e., the predicted CEM value was higher than the observed CEM value), this was counted as an example of a convergent reply.

Hypothesis tests involved a custom, stochastic linear regression model to assess how various responses influenced the hatefulness observed in subsequent posts (see Supplemental Appendix A). We observed at time t the hatefulness of a post, S_t, and the hatefulness expressed in their next post, S_t ₊ ₁. The change in hatefulness was affected by feedback that the original post received in the form of (1) how many replies to a post posted at time $t$ converged with the conceptual content proposed in that post ( $n_{c}$ ), (2) how many replies diverged with the conceptual content proposed in that same post ( $n_{d}$ ), (3) the number of Likes the comment received ( $n_{l}$ ), and (4) the number of reposts the comment received ( $n_{r}$ ). This approach produced the following equation:

S_{t + 1} ~ P_{N} (β_{c} n_{c} + β_{d} n_{d} + β_{l} n_{l} + β_{r} n_{r} + S_{t}; σ_{t})

(2)

where the hatefulness rating of the comment written by a user at time $t + 1$ ( $S_{t + 1}$ ) is sampled from a Gaussian distribution, the mean of which is determined via $β_{c} n_{c} + β_{d} n_{d} + β_{l} n_{l} + β_{r} n_{r} + S_{t}$ and whose variance $σ_{t}$ is estimated at time $t$ .

While the social approval theory explicitly predicts changes in a user’s frequency of hate postings, we tested a related variable in the assessment of H2a,b and H4: how Likes, reposts, and the convergence or divergence of replies affected the amount of time between posting the original post and a user’s subsequent post discussing the same topic. This approach was warranted because analyses were done on a per-post basis rather than a between-users basis; that is, the unit of analysis was the average change between users’ posts (with respect to hatefulness or time lag), from post to post, due to Likes, reposts, and the convergence/divergence of replies to the initial post. While it would be reasonable to assume that a reduction in the time between posts renders more frequent posts within a certain time interval, that assumption does not take into account variations in the responses to a user’s various posts across time. A user may post several messages within a time span, some of which garner greater approval and others that garner less. Therefore, a total frequency count of a user’s posting across time is a relatively less precise indicator of the effects of social approval, post by post. The present approach, in contrast, offers a more detailed view of the consequences of social approval on a user’s likelihood of posting again more or less immediately, even if it is not a perfect representation of frequency per se.

To detect these effects, we used a custom linear regression model, where the time between the original post at time t and a subsequent post on the same topic at time t + 1 ( $\nabla_{t + t}$ ) was estimated using a Gaussian distribution with a mean $β_{0} + β_{c, \nabla_{t}} n_{c} + β_{d, \nabla_{t}} n_{d} + β_{n_{l}, \nabla_{t}} n_{l} + β_{n_{r}, \nabla_{t}} n_{r}$ and variance $σ_{Δ_{t}}$ .

\nabla_{t + t} \sim P_{N} (\begin{array}{l} β_{0} + β_{c, \nabla_{t}} n_{c} + β_{d, \nabla_{t}} n_{d} + β_{n_{l}, \nabla_{t}} n_{l} \\ + β_{n_{r}, \nabla_{t}} n_{r}; σ_{\nabla_{t}} \end{array})

(3)

We estimated the parameters in our models using Gibbs sampling. Both models (equations 2 and 3) were implemented in PyJAGS and run using Google CoLab. The full model formalism with defined priors appears in the Supplemental Materials (Appendix A). A Jupyter notebook file (.ipynb) containing PyJAGS scripts for executing both models appears in the Supplemental Materials (Appendix B).

Results

Initial analyses examined the moderating effect of whether posts and replies were anti-Semitic or anti-Islamic. No differences were found in any relationships presented below. Analyses continued by pooling these subsamples.

Results indicated that both the number of convergent replies ( $6.96 e^{- 3}$ , z = 2.34, p = .0096) and the number of divergent replies ( $- 6.78 e^{- 3}$ , z = −2.26, p = .012) are significant linear predictors of subsequent hatefulness (see Table 1). Convergent replies are associated with significantly greater hatefulness in respective posters’ subsequent posts. This finding supports H1a, that approving replies to hate posts leads to even more extreme hatefulness. However, increasing amounts of divergent replies are associated with less, rather than more subsequent hatefulness. This finding rejects H1b, which entertained an oppositional effect based on the notion that insufficient approval drives haters to more extreme hate messaging.

Table 1.

Parameter Estimates for JAGS Model Used to Fit Parameters Referenced in Equations (2) and (3).

	Parameter estimate	Parameter STD	Parameter Z-statistic
Predicted change to hatefulness at time t + 1
β_c	6.96e⁻³	2.97e⁻³	2.34*
β_d	−6.78e⁻³	3.00e⁻³	−2.26*
β_l	3.03e⁻⁵	1.00e⁻⁵	.30
β_r	−8.38e⁻⁴	9.89e⁻⁴	−.85
Predicted time to subsequent post (∇_t ₊ ₁)
β_c∇t	−364.72	.37	−993.24**
β_d∇t	345.35	.42	825.36**
β_l∇t	−11.20	.01	−942.34**
β_r∇t	180.08	.14	1281.25**
β ₀	3723.11	.48	7694.53**

p < .05, **p < 1e⁻⁵.

The number of Likes and reposts attributed to a hate post did not significantly predict the hatefulness of users’ subsequent posts. These results suggest the failure of H3 (getting more single-click approval signals increases subsequent hatefulness). In contrast to a previous study (Jiang et al., 2023, where single-click responses and the frequency but not the content of written replies were considered), single-click responses did not affect hatefulness one way or another in these conversations.

The number of convergent replies ( $- 264.72,$ z = −993.24, $p < 1 e^{- 5}$ ) and divergent replies ( $345.35$ , z = 82,536, $p < 1 e^{- 3}$ ) are also significant, linear predictors of differences in the amount of time between users’ original posts that targeted either Jewish or Muslim persons or groups, in opposite directions. Results support H2a: The more replies that were convergent with original hate messages, the more quickly a hate poster added a new hate message, which leads to greater amounts of hate messages within a time interval. In contrast, divergent replies led to longer time intervals before a poster added a new hate message, rejecting H2b.

In addition to the findings that convergent and divergent replies affect the amount of time between posts, the number of Likes accrued by a post was a significant, linear predictor of a reduction in the amount of time before a user posted a subsequent hate post, $- 11.19,$ z = −942.34, $p < 1 e^{- 5}$ , leading to more hate postings overall, as predicted by H4. However, the number of reposts significantly predicted a greater time lag before a user posted a new post, $180.08,$ z = 1281.25, $p < 1 e^{- 5}$ , which was inconsistent with H4. H4 was partially confirmed and partially refuted due to the contrary effects of Likes versus reposts on the time lag between hate posts.

Discussion

The results of this study provide evidence that the overt expression of hatefulness by X.com users comprises a stochastic process that is sensitive to convergence and divergence expressed in the replies to users’ posts. The social approval theory of online hate (Walther, 2024b) predicts that as users receive greater positive feedback—in this case, more verbal replies indicating approval of the ideas expressed in their original posts—they post more, and more extreme hate messages: The more convergent replies users’ hate posts receive, the more they amplify hate in their next posts.

These findings raise implications beyond the increase in hatefulness. Positive responses to hate posts seem to incentivize riskier forms of hate expression. The logic of this claim is as follows: The classifier that this study employed to analyze hatefulness is a “state-of-the-art” metric, commonly deployed, freely available, peer-reviewed, and developed for industry to detect unacceptably extreme messages to trigger content moderation and the removal of posts. Its developers were engineers from Meta’s (formerly Facebook’s) own labs (Vidgen et al., 2021). Online hate posters try to evade such detection to avoid having their posts deleted by these devices (Parvaresh, 2023; Rieger et al., 2021). To do so, they often compose hate messages and slurs in heavily encoded terms and covert expressions that evade detection by industry classifiers (Bhat & Klein, 2020). Therefore, since the hate classifier employed in this study scores the likelihood that a message violates industry standards, the evidence indicates not only increases in the hatefulness of users’ posts, but also the degree to which they are willing to engage in more overt, riskier, extreme expressions of hate. Thus, positive feedback in the form of more people² converging with the ideas expressed by haters in discussions of Jews and Muslims incentivizes these users to “remove the mask” of coded language. Social approval triggers hate posters to “say the quiet part out loud” (see Rae et al., 2024, p. 180), perhaps due to the sense that others are more receptive to such messages.

A second finding related to hate posters’ behavior also confirms hypotheses from the social approval theory of online hate. The observed decrease in the amount of time between posts, due to the influence of convergent replies, implies a greater frequency of hate posts within a time interval.

In contrast to the effects of convergent replies, the effects of single-click responses, such as Likes and reposts, did not significantly affect the hatefulness of users’ next posts in this study. This finding contrasts with a number of other, recent reports that analyzed the effect of single-click responses only, without inclusion of the content of verbal replies, and showed significant positive associations between Likes (Brady et al., 2021; Frimer et al., 2023) or Likes and retweets (Jiang et al., 2023) on increased hatefulness. Whether these differences are due to the analytic approach (the inclusion of verbal replies in addition to single-click responses), the intense topic (anti-Semitic and anti-Islamic sentiments during an active war), or are consequences of different content moderation practices attributable to changes in the platform’s ownership (Ortiz, 2024) cannot be discerned.

The lack of effects from Likes or reposts on the hatefulness of subsequent posts might be due to the kinds of users included in this study. These X users were not “influencers”; they did not have large followings as do many celebrities or politicians who use the platform for public outreach. It seems possible that accounts that enjoy much greater popularity may respond differently to sheer numbers of easily-summed single-click approval signals than in the dialogues their posts generate (see, e.g., Frimer et al., 2023). Future research and different methods are required to discern whether that is the case.

Another, simpler possible explanation for the lack of effect of reposts and Likes on users’ subsequent hatefulness, relative to the effect of verbal replies, could be the sparsity of the former set of signals. While 44.4% of posts received at least one Like, the distribution of Likes drops precipitously from there. In this dataset, users could expect to receive no Likes at all, and if they did get a Like, it is highly unlikely that they would get more than a single one. Similarly, only 15.8% of posts were reposted at least once, after which there is another precipitous drop off. In both cases, the modal response to a post is one repost and one Like.

Importantly, few posts received verbal replies, with only 21% receiving at least one comment. Despite the greater scarcity of replies than Likes, replies nevertheless were significantly more influential when they appeared. Writing a reply at all requires greater effort, and is thus a more costly marker of approval or disapproval. Costly markers of individuals’ positions are often more strongly attended to than ones that cost less effort to generate (Donath, 2007), consistent with other findings that the more people perceive that the Likes they get from others are automatic and mindless reactions, the less emotional influence they have (Carr et al., 2016). As has been shown in the impact of friendly comments and Likes (Burke & Kraut, 2016), replies signal effortful engagement with users’ ideas. When replies demonstrate convergence, users are likely to experience greater social approval for the ideas they expressed in their post. This appears to be the case in hate posting as well as friendly interactions.

The different effects of reposts versus replies to hate messages, as seen in the results of H4, may be due to missing data of a kind: The Brandwatch application’s protocols do not distinguish between reposts and “quote tweets,” that is, reposts of a hate message to which the reposting user appends an original message of their own. Brandwatch does not capture the original content contained in quote tweets. Prior research (see Metzger et al., 2021) shows that a significant amount of message re-sharing, like quote tweets, adds content that defies, ridicules, or otherwise contradicts the original messages to which they are appended.³ Thus, it is unknowable, among the sparse amount of reposts included in these data, whether and what portion of the reposts were appended with divergent or convergent original content. This gap represents an important limitation to the present work.

It is interesting that the frequency of replies that diverged from the hate posts correlated with decreased subsequent hatefulness. Social approval theory offers conflicting alternative hypotheses about divergence, i.e., social disapproval responses to hateful behavior. It proposed that social disapproval renders no direct effect on subsequent hate posting. However, an implication of the social approval hypothesis is that the less approval one receives, the less subsequent hate should be expected. If more frequent convergence is interpreted as greater social approval, greater divergence may constitute less approval. In the present case, the independent measurement of convergence (approval) and divergence (disapproval) prevents the conflation of effects. However, additional considerations about the effects of these oppositional signals on an individual’s balance of approval-to-disapproval warrant further research consideration.

The theory’s most specific hypothesis about disapproval responses considers the effect of disapproval to be mediated by other, contingent factors. In this case, results demonstrate that the null hypothesis—that divergence does not affect subsequent behavior—is not consistent with the data. Additional theorizing and testing need to be performed to determine both why divergence appears to predict less subsequent hatefulness and what mechanisms facilitate that effect. The results of this study clearly indicate that the greater the positive verbal responses to a hate post a user accrues, the more the user posts hate more extremely and more quickly (see Stsiampkouskaya et al., 2021). At the same time, increased negative feedback causes such users to disengage from hate posting to some extent (see Erişti & Akbulut, 2019; Sannon et al., 2017).

The existence and acceleration of hate discourse targeting religious groups in response to ethnopolitical crises clearly demonstrates an increase in the boundary-spanning spillover of online hatefulness. There has always been some tendency to “holding Jews collectively responsible for actions of the state of Israel” (Weimann & Masri, 2022, p. 170) and holding Muslims responsible for Hamas. Comments on social media in the immediate aftermath of October 7 solidified this tendency and focused on negative stereotyped characterizations of Jews and Muslims in so doing. That is, they repeated historical Islamophobic narratives depicting “Muslims as threats to Western values and security,” whereas antisemitism “usually draws on long-standing conspiracy theories about Jewish influence on politics, economy, and culture”; “The attacks initiated by Hamas and the ensuing political-military aftermath by Israel appear to align with and intensify these existing prejudices, further polarizing online discourse” (Nefriana et al., 2024, p. 1). An analysis of post-October 7 anti-Muslim and anti-Arab postings found that religious vilification of Muslims comprised 91% of the posts; only 9% focused on Palestinian or Arab identities per se. Such features have also been prevalent in anti-Semitic posts originating in the United States (Kressel, 2024) and in “various European countries, where anti-Semitism and anti-Zionism often overlap, complicating the distinction between political criticism and outright hate” (Nefriana et al., 2024, p. 3).

Although the causal relationship of online hate to offline hate and violence is not well-established, it is known that online hate interactions facilitate the logistical organization of offline collective action (Wahlström & Törnberg, 2021), and this is likely to have been the case in the many in-person protests focused on the Israel/Gaza crisis. In addition to logistics, attitudinal messages receiving convergent responses may render psychological effects. Hateful replies and single-click responses in the early days of this event have been credited with exacerbating and polarizing perceptions and attitudes; through interaction, not just observation, social media participation appeared to “fuel extremist thought, perpetuate divisiveness, and influence perceptions” in this specific crisis particularly (Klempner, 2023).

Supplemental Material

sj-pdf-2-sms-10.1177_20563051251383635 – Supplemental material for Social Processes in the Intensification of Online Hate: The Effects of Verbal Replies to Anti-Muslim and Anti-Jewish Posts Following 7 October 2023

Supplemental material, sj-pdf-2-sms-10.1177_20563051251383635 for Social Processes in the Intensification of Online Hate: The Effects of Verbal Replies to Anti-Muslim and Anti-Jewish Posts Following 7 October 2023 by Zachary P. Rosen and Joseph B. Walther in Social Media + Society

Supplemental Material

sj-txt-1-sms-10.1177_20563051251383635 – Supplemental material for Social Processes in the Intensification of Online Hate: The Effects of Verbal Replies to Anti-Muslim and Anti-Jewish Posts Following 7 October 2023

Supplemental material, sj-txt-1-sms-10.1177_20563051251383635 for Social Processes in the Intensification of Online Hate: The Effects of Verbal Replies to Anti-Muslim and Anti-Jewish Posts Following 7 October 2023 by Zachary P. Rosen and Joseph B. Walther in Social Media + Society

Footnotes

Acknowledgements

The authors are very grateful to Prof. Rick Dale for his generous advice on methods and their application and co-development of the convergence-entropy metric; and the UCSB DREAM Lab for assistance in data acquisition.

ORCID iDs

Zachary P. Rosen

Joseph B. Walther

Ethical Considerations

n/a

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported in part by the Institute for Rebooting Social Media, Berkman Klein Center for Internet & Society at Harvard University, and the UC Santa Barbara Bertelsen Presidential Endowment for Technology and Society.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

Upon request to the first author.

Supplemental Material

Supplemental material for this article is available online.

Notes

Author biographies

Zachary P. Rosen is a doctoral student in Communication at the University of California, Los Angeles. He is a computational sociolinguist and data scientist with expertise in natural language processing and predictive modeling. His research focuses on how people communicate and negotiate meaning in group interactions using language, particularly in social media. His previous publications have appeared in the Journal of Language and Social Psychology, Behavior Research Methods, and elsewhere.

Joseph B. Walther holds the Bertelsen Presidential Chair in Technology and Society at the University of California, Santa Barbara, where he is a Distinguished Professor of Communication. He is also a faculty associate at the Berkman Klein Center for Internet & Society at Harvard University.

References

Bavelas

J. B.

Black

Chovil

Lemery

C. R.

Mullett

(1988). Form and function in motor mimicry: Topographic evidence that the primary function is communicative. Human Communication Research, 14(3), 275–299. https://doi.org/10.1111/j.1468-2958.1988.tb00158.x

Benesch

(2013, February 23). Dangerous speech: A proposal to prevent group violence. Dangerous Speech Project. https://dangerousspeech.org/wp-content/uploads/2018/01/Dangerous-Speech-Guidelines-2013.pdf

Bhat

Klein

(2020). Covert hate speech: White nationalists and dog whistle communication on Twitter. In Bouvier

Rosenbaum

J. E.

(Eds.), Twitter, the public sphere, and the chaos of online deliberation (pp. 151–172). Springer. https://doi.org/10.1007/978-3-030-41421-4_7

Bradac

J. J.

Mulac

House

(1988). Lexical diversity and magnitude of convergent versus divergent style shifting-: Perceptual and evaluative consequences. Language & Communication, 8(3), 213–228. https://doi.org/10.1016/0271-5309(88)90019-5

Brady

W. J.

McLoughlin

Doan

T. N.

Crockett

(2021). How social learning amplifies moral outrage expression in online social networks. Science Advances, 7(33), eabe5641. https://doi.org/10.31234/osf.io/gf7t5

Brandwatch. (2024). https://www.brandwatch.com/

Brown

(2018). What is so special about online (as compared to offline) hate speech? Ethnicities, 18(3), 297–326. https://doi.org/10.1177/1468796817709846

Bührer

Koban

Matthes

(2024). The WWW of digital hate perpetration: What, who, and why? A scoping review. Computers in Human Behavior, 159, 108321. https://doi.org/10.1016/j.chb.2024.108321

Burgoon

J. K.

Dillman

Stern

L. A.

(1993). Adaptation in dyadic interaction: Defining and operationalizing patterns of reciprocity and compensation. Communication Theory, 3(4), 295–316. https://doi.org/10.1111/j.1468-2885.1993.tb00076.x

10.

Burke

Kraut

R. E.

(2016). The relationship between Facebook use and well-being depends on communication type and tie strength. Journal of Computer-Mediated Communication, 21(4), 265–281. https://doi.org/10.1111/jcc4.12162

11.

Carr

C. T.

Wohn

D. Y.

Hayes

R. A.

(2016). Thumb up as social support: Relational closeness, automaticity, and interpreting social support from paralinguistic digital affordances in social media. Computers in Human Behavior, 62, 385–393. https://doi.org/10.1016/j.chb.2016.03.087

12.

Center for Countering Digital Hate. (2024) Hate pays: How X accounts are exploiting the Israel-Gaza conflict to grow and profit. https://counterhate.com/blog/new-report-anti-jewish-and-anti-muslim-actors-grow-4x-as-quickly-on-x-exploiting-israel-gaza-conflict-and-elon-musks-policy-changes/

13.

Chartrand

T. L.

Bargh

J. A.

(1999). The chameleon effect: The perception–behavior link and social interaction. Journal of Personality and Social Psychology, 76(6), 893–910. https://doi.org/10.1037/0022-3514.76.6.893

14.

Dehghani

Boyd

R. L.

(Eds.) (2022). Handbook of language analysis in psychology. The Guilford Press.

15.

Devlin

Chang

M.-W.

Lee

Toutanova

(2019). BERT: Pre-training of deep bidirectional transformers for language understanding. In Burstein

Doran

Solorio

(Eds.), Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) (pp. 4171–4186). Association for Computational Linguistics. https://doi.org/10.18653/v1/N19-1423

16.

Donath

(2007). Signals in social supernets. Journal of Computer-mediated Communication, 13(1), 231–251. https://doi.org/10.1111/j.1083-6101.2007.00394.x

17.

Dunbar

N. E.

Fujiwara

Bernhold

Q. S.

(2023). Truth tellers’ and liars’ synchrony during attitude inconsistent conversations. Frontiers in Communication, 8. https://doi.org/10.3389/fcomm.2023.921446

18.

Erişti

Akbulut

(2019). Reactions to cyberbullying among high school and university students. The Social Science Journal, 56(1), 10–20. https://doi.org/10.1016/j.soscij.2018.06.002

19.

Frenkel

Myers

S. L.

(2023, November 15). Antisemitic and anti-Muslim hate speech surges across the internet. The New York Times. https://www.nytimes.com/2023/11/15/technology/hate-speech-israel-gaza-internet.html

20.

Frimer

J. A.

Aujla

Feinberg

Skitka

L. J.

Aquino

Eichstaedt

J. C.

Willer

(2023). Incivility is rising among American politicians on Twitter. Social Psychological and Personality Science, 14(2), 259–269. https://doi.org/10.1177/19485506221083811

21.

Gonzales

A. L.

Hancock

J. T.

Pennebaker

J. W.

(2010). Language style matching as a predictor of social dynamics in small groups. Communication Research, 37(1), 3–19. https://doi.org/10.1177/0093650209351468

22.

Hobbs

Lajevardi

Lucas

(2023). From Anti-Muslim to Anti-Jewish: Target substitution on fringe social media platforms and the persistence of online and offline hate. Political Behavior, 46, 1747–1769. https://doi.org/10.1007/s11109-023-09892-9

23.

Ireland

M. E.

Slatcher

R. B.

Eastwick

P. W.

Scissors

L. E.

Finkel

E. J.

Pennebaker

J. W.

(2010). Language style matching predicts relationship initiation and stability. Psychological Science, 22(1), 39–44. https://doi.org/10.1177/0956797610392928

24.

Jiang

Luceri

Walther

J. B.

Ferrara

(2023). Social approval and network homophily as motivators of online toxicity (No. arXiv:2310.07779). arXiv. http://arxiv.org/abs/2310.07779

25.

Klempner

(2023, October 23). Emojis of terror: Telegram’s role in perpetuating extremism in the October 2023 Hamas-Israel war. Global Network on Extremism & Technology. https://gnet-research.org/2023/10/23/emojis-of-terror-telegrams-role-in-perpetuating-extremism-in-the-october-2023-hamas-israel-war/

26.

Kressel

N. J.

(2024). The psychology of contemporary antisemitism. In Nelson

T. D.

(Ed.), Handbook of prejudice, stereotyping, and discrimination (3rd ed., pp. 327–365). Routledge. https://doi.org/10.4324/9781003399162-14

27.

Kulesza

Dolinski

Huisman

Majewski

(2014). The echo effect: The power of verbal mimicry to influence prosocial behavior. Journal of Language and Social Psychology, 33(2), 183–201. https://doi.org/10.1177/0261927X13506906

28.

Levin

(2015). The original web of hate: Revolution Muslim and American homegrown extremists. American Behavioral Scientist, 59(12), 1609–1630. https://doi.org/10.1177/0002764215588815

29.

Liu

Ott

Goyal

Joshi

Chen

Levy

Lewis

Zettlemoyer

Stoyanov

(2019). RoBERTa: A robustly optimized BERT pretraining approach (No. arXiv:1907.11692). arXiv. https://doi.org/10.48550/arXiv.1907.11692

30.

Metzger

M. J.

Flanagin

A. J.

Mena

Jiang

Wilson

(2021). From dark to light: The many shades of sharing misinformation online. Media and Communication, 9(1), 134–143. https://doi.org/10.17645/mac.v9i1.3409

31.

Nefriana

Yan

Diab

Wheeler

D. L.

Miller

Hwa

Lin

Y.-R.

(2024, September 18–20). Shifting patterns of extremist discourse on Facebook: Analyzing trends and developments during the Israel-Hamas conflict. 17th International Conference on Social Computing, Behavioral-Cultural Modeling, & Prediction and Behavior Representation in Modeling and Simulation, Pittsburgh, PA, United States. https://sbp-brims.org/2024/papers/working-papers/Nefriana_SBP-BRiMS2024_Final_70.pdf

32.

Oboler

Roth

Noonan

Steiner

Mohammed

Beinart

Oliver

McNamara

(2024). Online anti-Muslim hate and racism against Palestinians and Arabs October 2023-February 2024 (No. IR2403; pp. 1–138). Online Hate Prevention Institute. https://doi.org/10.26181/26538670.V1

33.

Oliveira

A. S.

Cecote

T. C.

Silva

P. H. L.

Gertrudes

J. C.

Freitas

V. L. S.

Luz

E. J. S.

(2023). How good is ChatGPT for detecting hate speech in Portuguese? Anais do Simpósio Brasileiro de Tecnologia da Informação e da Linguagem Humana (STIL), 94–103. https://doi.org/10.5753/stil.2023.233943

34.

Ortiz

A. G.

(2024, April 11). Anti-Jewish and anti-Muslim actors grow 4x as quickly on X, exploiting Israel-Gaza conflict and Elon Musk’s policy changes. Center for Countering Digital Hate. https://counterhate.com/blog/new-report-anti-jewish-and-anti-muslim-actors-grow-4x-as-quickly-on-x-exploiting-israel-gaza-conflict-and-elon-musks-policy-changes/

35.

Parvaresh

(2023). Covertly communicated hate speech: A corpus-assisted pragmatic study. Journal of Pragmatics, 205, 63–77. https://doi.org/10.1016/j.pragma.2022.12.009

36.

Pluta

Mazurek

Wojciechowski

Wolak

Soral

Bilewicz

(2023). Exposure to hate speech deteriorates neurocognitive mechanisms of the ability to understand others’ pain. Scientific Reports, 13(1), 4127. https://doi.org/10.1038/s41598-023-31146-1

37.

Rea

S. C.

Mathew

Kraemer

(2024). “Hate Parties”: Networked antisemitism from the fringes to YouTube. In Walther

J. B.

Rice

R. E.

(Eds.), Social processes of online hate (pp. 168–192). Routledge. https://doi.org/10.4324/9781003472148-8

38.

Rice

R. E.

(2024). Background scholarship and a synthesis of themes in social processes of online hate. In Walther

J. B.

Rice

R. E.

(Eds.), Social processes of online hate (pp. 273–300). Routledge. https://doi.org/10.4324/9781003472148-12

39.

Rieger

Kümpel

A. S.

Wich

Kiening

Groh

(2021). Assessing the extent and types of hate speech in fringe communities: A case study of alt-right communities on 8chan, 4chan, and Reddit. Social Media + Society, 7(4). https://doi.org/10.1177/20563051211052906

40.

Rosen

Z. P.

Dale

(2024). BERTs of a feather: Studying inter- and intra-group communication via information theory and language models. Behavior Research Methods, 56(4), 3140–3160. https://doi.org/10.3758/s13428-023-02267-2

41.

Sannon

Choi

Y. H.

Taft

J. G.

Bazarova

N. N.

(2017). What comments did I get? How post and comment characteristics predict interaction satisfaction on Facebook. Proceedings of the International AAAI Conference on Web and Social Media, 11(1), 664–667. https://doi.org/10.1609/icwsm.v11i1.14930

42.

Shannon

C. E.

(1948). A mathematical theory of communication. Bell System Technical Journal, 27(4), 623–656. https://doi.org/10.1002/j.1538-7305.1948.tb00917.x

43.

Shmargad

Coe

Kenski

Rains

S. A.

(2022). Social norms and the dynamics of online incivility. Social Science Computer Review, 40(3), Article 3. https://doi.org/10.1177/0894439320985527

44.

Soral

Bilewicz

Winiewski

(2018). Exposure to hate speech increases prejudice through desensitization. Aggressive Behavior, 44(2), 136–146. https://doi.org/10.1002/ab.21737

45.

Stsiampkouskaya

Joinson

Piwek

Ahlbom

C.-P.

(2021). Emotional responses to likes and comments regulate posting frequency and content change behaviour on social media: An experimental study and mediation model. Computers in Human Behavior, 124, 1–14. https://doi.org/10.1016/j.chb.2021.106940

46.

Tong

S. T.

(2024). Foundations, definitions, and directions in online hate research. In Walther

J. B.

Rice

R. E.

(Eds.), Social processes of online hate (pp. 37–72). Routledge. https://doi.org/10.4324/9781003472148-3

47.

Törnberg

(2024). From echo chambers to digital campfires: The making of an online community of hate in Stormfront. In Walther

J. B.

Rice

R. E.

(Eds.), Social processes of online hate (pp. 93–119). Routledge. https://doi.org/10.4324/9781003472148-5

48.

Törnberg

(2022). Inside a White power echo chamber: Why fringe digital spaces are polarizing politics. New Media & Society, 26(8), 14614448221122915. https://doi.org/10.1177/14614448221122915

49.

United Nations. (n.d.). Understanding hate speech. https://www.un.org/en/hate-speech/understanding-hate-speech/what-is-hate-speech

50.

Vidgen

Thrush

Waseem

Kiela

(2021). Learning from the worst: Dynamically generated datasets to improve online hate detection. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11Th International Joint Conference on Natural Language Processing, 1: Long Papers, 1, 1667–1682. https://doi.org/10.18653/v1/2021.acl-long.132

51.

Wahlström

Törnberg

(2021). Social media mechanisms for right-wing political violence in the 21st century: Discursive opportunities, group dynamics, and co-ordination. Terrorism and Political Violence, 33(4), Article 4. https://doi.org/10.1080/09546553.2019.1586676

52.

Walther

J. B.

(2024a). Making a case for a social processes approach to online hate. In Walther

J. B.

Rice

R. E.

(Eds.), Social processes of online hate (pp. 9–36). Routledge. https://doi.org/10.4324/9781003472148-2

53.

Walther

J. B.

(2024b). The effects of social approval signals on the production of online hate: A theoretical explication. Communication Research. Advance online publication. https://doi.org/10.1177/00936502241278944

54.

The Weaponized Word. (2024). https://weaponizedword.org/

55.

Weigel

Gitomer

(2024). Hate-sharing: A case study of its prevalence and impact on Gab. New Media & Society, 27(9), 14614448241245349. https://doi.org/10.1177/14614448241245349

56.

Weimann

(2024). New trends in online antisemitism. In Hartleb

(Ed.), Antisemitism on the rise: New ideological dynamics (pp. 43–53). European Institute for Counter Terrorism and Conflict Prevention. https://eictp.eu/wp-content/uploads/2024/05/EICTP_Research_Papers_Antisemitism_FINAL.pdf

57.

Weimann

Masri

(2022). New antisemitism on TikTok. In Hübscher

von Mering

(Eds.), Antisemitism on social media (pp. 168–180). Routledge. https://doi.org/10.4324/9781003200499-11

58.

Wolf

Debut

Sanh

Chaumond

Delangue

Moi

Cistac

Rault

Louf

Funtowicz

Davison

Shleifer

von Platen

Jernite

Plu

Scao

T. L.

Gugger

. . .Rush

A. M.

(2020). HuggingFace’s transformers: State-of-the-art natural language processing (No. arXiv:1910.03771). arXiv. https://doi.org/10.48550/arXiv.1910.03771

59.

Zhang

(2023). A neurocomputational variable on welfare tradeoffs explains the function and form of cyberaggression. Frontiers in Behavioral Neuroscience, 17, Article 1034564. https://www.frontiersin.org/articles/10.3389/fnbeh.2023.1034564

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.22 MB

0.01 MB