Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Grounding Toxicity in Real-World Events across Languages

Wondimagegnhue Tsegaye Tufa, Ilia Markov, Piek Vossen

TL;DR

This study investigates how major real-world events drive the origin and spread of toxicity in multilingual online discussions. It analyzes 4.5 million Reddit comments across six languages (Dutch, English, German, Arabic, Turkish, Spanish) tied to 15 events from 2020–2023, using a lexicon-based toxicity detector supplemented by GPT-4 and Perspective API for evaluation. The authors examine toxicity alongside sentiment and emotion via the NRC lexicon, performing both within-language and cross-language analyses to reveal event- and language-specific patterns, including delayed toxicity peaks and strong cross-language relationships. The work provides a multilingual, event-grounded framework for understanding online toxicity and releases the data and code to enable further research and cross-cultural insights.

Abstract

Social media conversations frequently suffer from toxicity, creating significant issues for users, moderators, and entire communities. Events in the real world, like elections or conflicts, can initiate and escalate toxic behavior online. Our study investigates how real-world events influence the origin and spread of toxicity in online discussions across various languages and regions. We gathered Reddit data comprising 4.5 million comments from 31 thousand posts in six different languages (Dutch, English, German, Arabic, Turkish and Spanish). We target fifteen major social and political world events that occurred between 2020 and 2023. We observe significant variations in toxicity, negative sentiment, and emotion expressions across different events and language communities, showing that toxicity is a complex phenomenon in which many different factors interact and still need to be investigated. We will release the data for further research along with our code.

Grounding Toxicity in Real-World Events across Languages

TL;DR

This study investigates how major real-world events drive the origin and spread of toxicity in multilingual online discussions. It analyzes 4.5 million Reddit comments across six languages (Dutch, English, German, Arabic, Turkish, Spanish) tied to 15 events from 2020–2023, using a lexicon-based toxicity detector supplemented by GPT-4 and Perspective API for evaluation. The authors examine toxicity alongside sentiment and emotion via the NRC lexicon, performing both within-language and cross-language analyses to reveal event- and language-specific patterns, including delayed toxicity peaks and strong cross-language relationships. The work provides a multilingual, event-grounded framework for understanding online toxicity and releases the data and code to enable further research and cross-cultural insights.

Abstract

Social media conversations frequently suffer from toxicity, creating significant issues for users, moderators, and entire communities. Events in the real world, like elections or conflicts, can initiate and escalate toxic behavior online. Our study investigates how real-world events influence the origin and spread of toxicity in online discussions across various languages and regions. We gathered Reddit data comprising 4.5 million comments from 31 thousand posts in six different languages (Dutch, English, German, Arabic, Turkish and Spanish). We target fifteen major social and political world events that occurred between 2020 and 2023. We observe significant variations in toxicity, negative sentiment, and emotion expressions across different events and language communities, showing that toxicity is a complex phenomenon in which many different factors interact and still need to be investigated. We will release the data for further research along with our code.
Paper Structure (23 sections, 3 figures, 2 tables)

This paper contains 23 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Temporal and event analysis
  4. User and community analysis
  5. Methodology
  6. Event selection
  7. Data source and collection
  8. Toxic language detection
  9. Lexicon-based approach
  10. GPT-4
  11. Perspective API
  12. Evaluation
  13. Sentiment and emotion lexicons
  14. Analysis
  15. Comment dynamics per event
  16. ...and 8 more sections

Figures (3)

  • Figure 1: Relationship between temporal event plot and the proportion of toxicity in conversation mentioning those events. The X-axis for the two plots is aligned to facilitate comparison. For visibility, we are only showing plots for five events.
  • Figure 2: Temporal distribution of unique users. The spike in the number of comments in Figure \ref{['fig:count_and_toxicity']} (a) strongly correlates with an increased number of users engaged in commenting on a discussion about a particular event. For visibility, we are only showing plots for six events.
  • Figure 3: Heatmap for toxicity density, negative sentiment density, and negative emotions density. We use a shorter format for the event names to save space as follows- COVID-19:COV, Fall of Kabul:FK, Russia-Ukraine:RUK, Turkey Earthquak:TEQ, Capitol Attack:CAP, Israel Hamas:IHA, European Heatwaves:EUH, Iran Protest:IPR.