Toxicidade em Chats de Live Streaming: Um Estudo Comparativo com Coleta Síncrona em Twitch, YouTube e Kick

May combine two claims.

This methodological limitation compromises the validity of comparisons, because observed differences may reflect distinct collection periods, not real structural variations between platforms.

“Essa limitac ¸ão metodológica compromete a validade das comparac ¸ões, pois diferenc ¸as observadas podem refletir períodos distintos de coleta, e não variac ¸ões estruturais reais entre plataformas [Singh et al. 2024].”

• cites Singh et al. 2024 sourceDisputed

  The Data Collection section (page unavailable in our copy) says the authors collected data from both Reddit and Discord for the same period, January–July 2022. The Limitations section does not identify different collection periods as a threat to the comparison. The paper therefore contradicts the premise that its observed cross-platform differences may reflect distinct collection periods. (LLM, conf=0.86)

  Checked by a reasoning model — may be wrong.

  Experimental — accuracy not yet measured

May combine two claims.

The study by Tereshchenko and Hämäläinen does not cross data between distinct platforms nor employ simultaneous collection, limiting multi-platform comparisons under the same temporal window.

“O estudo, contudo, não cruza dados entre plataformas distintas nem emprega coleta simultânea, o que limita comparac ¸ões multiplataforma sob o mesmo recorte temporal.”

• cites Tereshchenko and Hämäläinen 2025 sourcePartly supported

  Section 3.1 describes Dota 2 as the held-out test set and Minecraft and Discord chats as sources for manual validation, rather than a matched cross-platform comparison. Section 6.4.1 says the datasets are snapshots from specific periods, but the paper does not establish whether collection was simultaneous or explicitly assess comparisons within the same temporal window. (LLM, conf=0.81)

  Experimental — accuracy not yet measured

Classified as probably not a substantial claim. May combine two claims.

The lexical distribution of Twitch chats does not obey classical assumptions such as Zipf's law, suggesting that this ecosystem has its own linguistic particularities.

“[Ringer et al. 2020] demonstraram que a distribuic ¸ão lexical de chats da Twitch não obedece a pressupostos clássicos como a Lei de Zipf -segundo a qual a frequência de uma palavra é inversamente proporcional à sua posic ¸ão no ranking de popularidade do vocabulário -sugerindo que esse ecossistema possui particularidades linguísticas próprias.”

• cites Ringer et al. 2020 sourcePartly supported

  The Data Cleaning Process section (page not available) reports that the most and least popular tokens in the cleaned dataset deviate from Zipf's law and says this illustrates the uniqueness of livestream chat. But it also says the distribution appears Zipfian after approximately the 26 most popular tokens, so the unqualified claim that Twitch chat's lexical distribution does not obey Zipf's law overstates the finding. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

May combine two claims.

Systematic comparative analyses between platforms are still scarce, and few studies adopt simultaneous collection strategies.

“No entanto, análises comparativas sistemáticas entre plataformas ainda são escassas, e poucos estudos adotam estratégias de coleta simultânea.”

• cites Singh et al. 2024 sourcePartly supported

  The Related Work section (page unavailable) identifies a gap in cross-platform comparisons of harmful language within communities, but says many studies have examined linguistic differences across platforms. The Data Gathering section describes this paper’s collection of posts from both Reddit and Discord; it does not say that few other studies use simultaneous collection strategies. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Computational efficiency is critical in synchronous environments, where high message flow requires near-real-time analytical processing.

“Os resultados indicam que a eficiência computacional é crítica em ambientes síncronos, onde o alto fluxo de mensagens exige processamento analítico em tempo quase real.”

• cites Tereshchenko and Hämäläinen 2025 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Lobo et al. analyzed toxicity dissemination in Twitch chat cascades, revealing contagion patterns between messages in Brazilian Portuguese.

“No contexto de live streaming, [Lobo et al. 2025] analisaram a disseminac ¸ão de toxicidade em cascatas de chat da Twitch, revelando padrões de contágio entre mensagens em português brasileiro.”

• cites Lobo et al. 2025 sourcePartly supported

  Sections 3.4 and 6 describe toxic messages occurring in sequences and analyze 472 Twitch chat cascades, supporting the claim about dissemination patterns. However, sections 3.1 and 3.3 say the study included channels and messages in both English and Portuguese; the findings are not presented as specific to Brazilian Portuguese messages. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

May combine two claims.

The study by Lobo et al. is limited to a single platform and does not perform simultaneous collection in different environments.

“No entanto, o estudo se limita a uma única plataforma e não realiza coleta simultânea em diferentes ambientes.”

• cites Lobo et al. 2025 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Singh et al. compared toxic language on Reddit and Discord, showing that architecture and norms modulate interaction; neither network operates under live streaming dynamics, and the absence of simultaneity reduces robustness of temporal comparisons.

“[Singh et al. 2024] compararam linguagem tóxica no Reddit e no Discord, mostrando que arquitetura e normas modulam a interac ¸ão; nenhuma dessas redes opera sob dinâmica de live streaming, e a ausência de simultaneidade reduz a robustez de comparac ¸ões temporais.”

• cites Singh et al. 2024 sourcePartly supported

  Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

May combine two claims.

Twitch presented the lowest toxicity rates in all considered cuts: global proportion, category distribution, and daily variation.

“A Twitch apresentou as menores taxas de toxicidade em todos os recortes considerados, tanto na proporc ¸ão global quanto na distribuic ¸ão por categoria e na variac ¸ão ao longo do dia.”

May combine two claims.

Twitch maintained the lowest toxicity rates over almost the entire analyzed period, oscillating between 14% and 19%.

“A Twitch manteve as menores taxas de toxicidade ao longo de quase todo o período analisado, oscilando entre 14% e 19%.”

May combine two claims.

Racism varied between 0.10% and 0.12%, sexism between 0.24% and 0.29%, and lgbtqphobia between 0.21% and 0.25% across platforms.

“Racism variou entre 0,10% e 0,12%, sexism entre 0,24% e 0,29%, e lgbtqphobia entre 0,21% e 0,25%.”

May combine two claims.

Kick and YouTube presented values consistently above 20% harmful messages.

“Kick e YouTube, por sua vez, apresentaram valores consistentes acima de 20%.”

May combine two claims.

Kick and YouTube exhibited more pronounced variations throughout the day, with peaks concentrated mainly between 11h and 16h, plus elevations at dawn, when both exceeded 22% harmful messages.

“As duas plataformas também exibiram variac ¸ões mais acentuadas ao longo do dia, com picos concentrados principalmente entre 11h e 16h, além de elevac ¸ões registradas na madrugada, quando ambas superaram 22% de mensagens nocivas.”

May combine two claims.

After preprocessing, the corpus went from 6,811,720 raw messages to 6,158,647 valid messages, corresponding to a retention rate of 90.41%.

“Após essa etapa, o corpus passou de 6.811.720mensagens brutas para 6.158.647mensagens válidas, correspondendo a uma taxa de retenc ¸ão de 90,41%.”

May combine two claims.

Kick had the highest proportion of harmful messages, followed by YouTube, then Twitch.

“Kick presented the highest proportion of harmful messages (21.4%), followed by YouTube (19.6%) and Twitch (16.3%)”

• Also stated in: § Abstract · plain

  “A Kick apresentou a maior proporc ¸ão de mensagens nocivas (21,4%), seguida pelo YouTube (19,6%) e pela Twitch (16,3%)”

May combine two claims.

Kick and YouTube exhibited higher proportions and greater temporal oscillation.

“Kick e YouTube, por sua vez, exibiram proporc ¸ões mais elevadas e maior oscilac ¸ão temporal.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Dreier and Pirker found relationships between increased toxicity on Twitch and factors such as streamer gender, community size, and game niche, showing that even on a single platform harmful behavior is heterogeneous, though the study was limited to about 100,000 messages.

“[Dreier and Pirker 2023] descobriram relac ¸ões entre o aumento da toxicidade na Twitch e fatores como gênero do streamer, tamanho da comunidade e nicho do jogo, mostrando que mesmo em uma única plataforma o comportamento nocivo é heterogêneo, embora o estudo tenha se limitado a cerca de 100 mil mensagens.”

• cites Dreier and Pirker 2023 sourceNot checked — no full text of the source

May combine two claims.

The DTTD model identifies ten categories of toxicity and has precision 0.8180 and accuracy 0.4214.

“o modelo identifica dez categorias que incluem diferentes formas de agressão, discriminac ¸ão e linguagem ofensiva, apresentando uma precisão de 0,8180 e uma acurácia de 0,4214.”

• cites Trajano 2023Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

The selection of channels focused on competitive game streams restricts generalization of results to other live streaming domains; it is not possible to affirm whether the higher toxicity profile observed on Kick would maintain in niches like music or non-competitive live events.

“Entre as limitac ¸ões do estudo destacam-se a selec ¸ão de canais focados em transmissões de jogos competitivos, o que restringe a generalizac ¸ão dos resultados para outros domínios de live streaming -em particular, não é possível afirmar se o maior perfil de toxicidade observado na Kick se manteria em nichos como streams de música ou eventos ao vivo não competitivos.”

May combine two claims.

The combination with the cascade pipeline makes large-scale processing viable without sacrificing categorization granularity, constituting the main methodological advance of this work relative to the literature.

“A combinac ¸ão com o pipeline em cascata torna viável o processamento em larga escala sem sacrificar a granularidade da categorizac ¸ão, constituindo, em conjunto, o principal avanc ¸o metodológico deste trabalho em relac ¸ão à literatura.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Morrier et al. showed that toxicity propagates virally in competitive game streams, amplified by frustrations and rivalries.

“Em transmissões de jogos competitivos, [Morrier et al. 2025] mostraram que a toxicidade se propaga de forma viral, amplificada por frustrac ¸ões e ri-validades, padrão relevante para o corpus deste trabalho.”

• cites Morrier et al. 2025Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Probably not a substantial claim: may not pass S3 (load-bearing).

DTTD assigns one main category per message, which may simplify cases with multiple simultaneous forms of toxicity.

“o DTTD atribui uma categoria principal por mensagem, o que pode simplificar casos com múltiplas formas simultâneas de toxicidade.”

May combine two claims.

The observed differences in toxicity proportions and temporal variation are compatible with the hypothesis that community norms, moderation mechanisms, and interaction architecture influence users' linguistic behavior, though direct causal relations cannot be established from this study's data.

“Essas diferenc ¸as são compatíveis com a hipótese de que normas comunitárias, mecanismos de moderac ¸ão e a arquitetura de interac ¸ão influenciam o comportamento linguístico dos usuários, embora relac ¸ões causais diretas não possam ser estabelecidas com os dados deste estudo.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The results describe the type of language classified by the DTTD model, not users' intention or the reasons behind messages, which limits inferences about their behavior.

“Ainda assim, os resultados apresentados descrevem o tipo de linguagem classificada pelo modelo DTTD, e não a intenc ¸ão dos usuários ou os motivos por trás das mensagens, o que limita inferências sobre seu comportamento.”

Classified as probably not a substantial claim.

The cascade pipeline structure may fail to identify toxic messages expressed through irony, sarcasm, or semantic ambiguity.

“Outras limitac ¸ões incluem a estrutura em cascata do pipeline de classificac ¸ão, que pode deixar de identificar mensagens tóxicas expressas por meio de ironia, sarcasmo ou ambiguidade semântica.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The intervals of temporal peaks are compatible with periods of greater activity in the Brazilian e-sports ecosystem, especially in time bands associated with regional and international competitive transmissions.

“Esses intervalos são compatíveis com períodos de maior atividade no ecossistema brasileiro de e-sports, especialmente em faixas horárias associadas a transmissões competitivas regionais e internacionais.”

Classified as probably not a substantial claim.

The second channel-selection criterion sought to ensure a representative corpus in terms of engagement and regularity of transmissions.

“O segundo buscou assegurar um corpus representativo em termos de engajamento e regularidade das transmissões.”

May combine two claims.

The observed variations are compatible with the hypothesis that structural aspects of the platforms contribute to the recorded toxicity patterns, although direct causal relations cannot be established from this study's data.

“As variac ¸ões observadas, portanto, são compatíveis com a hipótese de que aspectos estruturais das plataformas contribuam para os padrões de toxicidade registrados, embora relac ¸ões causais diretas não possam ser estabelecidas a partir dos dados deste estudo.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Other uncontrolled factors, such as type of match, audience composition in each time slot, and active presence of moderators, may also contribute to the observed variations.

“Outros fatores não controlados, como tipo de partida, composic ¸ão da audiência em cada faixa de horário e presenc ¸a ativa de moderadores, podem igualmente contribuir para as variac ¸ões observadas.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The interpretation of temporal peaks as e-sports activity should be seen as an explanatory hypothesis, not direct evidence, because the study did not collect audience data by time window.

“No entanto, essa interpretac ¸ão deve ser vista como hipótese explicativa, e não como evidência direta, pois o estudo não coletou dados de audiência por janela horária.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

The less frequent categories should not be treated as irrelevant, given the absolute volume of occurrences accumulated over the monitoring period.

“As categorias menos frequentes não devem ser tratadas como irrelevantes, dado o volume absoluto de ocorrências acumuladas ao longo do período de monitoramento.”

May combine two claims.

The operational threshold between non-harmful and harmful messages (negative polarity plus DTTD category) may generate false positives in non-abusive negative messages and false negatives in ironic or ambiguous messages.

“Essa decisão reduz o custo computacional ao concentrar a inferência mais custosa apenas no subconjunto com indícios de negatividade semântica, tornando viável a análise de grandes volumes de dados, mas pode gerar falsos positivos em mensagens negativas não abusivas e falsos negativos em mensagens irônicas ou ambíguas.”

May combine two claims.

The platforms analyzed differ not only in volume of interaction but also in the proportional profile of toxicity observed in their chats.

“Em conjunto, os resultados mostram que as plataformas analisadas diferem não apenas no volume de interac ¸ão, mas também no perfil proporcional de toxicidade observado em seus chats.”

Classified as probably not a substantial claim.

Control policies, engagement mechanisms, and community rules can have a considerable impact on the production and acceptance of toxic content.

“Políticas de controle, mecanismos de engajamento e regras da comunidade podem ter um impacto considerável na produc ¸ão e na aceitac ¸ão de conteúdos tóxicos [Dreier and Pirker 2023].”

• cites Dreier and Pirker 2023 sourceNot checked — no full text of the source

May combine two claims.

This study brings together synchronous multi-platform collection, large-scale mining, and detailed categorization of harmful discourse, allowing comparison of different platforms under the same observation conditions.

“Este estudo reúne esses três elementos e permite comparar o comportamento de diferentes plataformas sob as mesmas condic ¸ões de observac ¸ão.”