Posts of Peril: Detecting Information About Hazards in Text

TL;DR

This work introduces a multilingual hazard detector trained on a newly annotated X-post corpus to identify information about hazards in social media text. It demonstrates that hazard signals are partially independent from negative affect indicators and shows the detector scales to millions of posts, outperforming simple dictionary baselines and approaching LLM baselines in accuracy while offering far higher throughput. The model is applied to 3.6M Israel-Hamas war posts and 5.9M French election posts to examine hazard usage in real-world information campaigns, revealing that coordinated actors often frame hazards to support weaker sides and influence civilian perceptions. The authors release hazard-annotated data and code as an open-source Python package, enabling researchers and journalists to analyze hazard content at scale and to study hazard-based information operations in multilingual, geopolitical contexts.

Abstract

Socio-linguistic indicators of affectively-relevant phenomena, such as emotion or sentiment, are often extracted from text to better understand features of human-computer interactions, including on social media. However, an indicator that is often overlooked is the presence or absence of information concerning harms or hazards. Here, we develop a new model to detect information concerning hazards, trained on a new collection of annotated X posts. We show that not only does this model perform well (outperforming, e.g., dictionary approaches), but that the hazard information it extracts is not strongly correlated with common indicators. To demonstrate the utility of our tool, we apply it to two datasets of X posts that discuss important geopolitical events, namely the Israel-Hamas war and the 2022 French national election. In both cases, we find that hazard information, especially information concerning conflict, is common. We extract accounts associated with information campaigns from each data set to explore how information about hazards could be used to attempt to influence geopolitical events. We find that inorganic accounts representing the viewpoints of weaker sides in a conflict often discuss hazards to civilians, potentially as a way to elicit aid for the weaker side. Moreover, the rate at which these hazards are mentioned differs markedly from organic accounts, likely reflecting information operators' efforts to frame the given geopolitical event for strategic purposes. These results are first steps towards exploring hazards within an information warfare environment. The model is shared as a Python package to help researchers and journalists analyze hazard content. The model, along with data and annotations, is available in the following repository: https://github.com/KeithBurghardt/DetectHazards.

Figures (14)

• Figure 1: Performance of models on human-annotated X posts. (a) We show the ROC-AUC of XG-Boost (XGB) chen2016xgboost, a neutral network (NN) tensorflow2015-whitepaper, random forest (RF) ho1995random, and a support vector machine (SVM) hearst1998support trained on human-annotated posts for multi-lingual embedding models described in the Scientific Methods section. We also show the ROC-AUC of GPT-3.5, GPT-4, and GPT-5 with zero-shot, two-shot, and five-shot predictions. Finally, we also show performance of the Threat Dictionary plus synonyms ("T.D. + Synonyms") at predicting threat text (if a post contained a word that was in this set of words, we labeled the data "hazard"; otherwise it was not). Gray bars represent standard deviations across 50 evaluations. (b) Best model performance and baseline performance on posts from the 2023 Israel-Hamas war dataset. (b) Best model performance and baseline performance on posts from the 2022 French election dataset.
• Figure 2: Understanding hazards. Spearman correlation between hazards moral outrage Brady2021, sentiment hutto2014vader, emotions Chochlakis2023, Threat Dictionary lilienfeld2014threat and threat synonyms for (a) the Israel-Hamas war, (b) the 2022 French election. All values are statistically significant (p-value $<0.05$).
• Figure 3: Hazards in representative large coordinated account clusters for (a) the Israel-Hamas war, (b) the 2022 French election. Mean hazard confidence per account is shown as a color from yellow to purple. Top hashtags and example posts are shown next to each cluster.
• Figure 4: Hazards and threats over time for the Israel-Hamas war dataset. The plots show mean hazard confidences each day as well as the overall mean proportion of posts with at least one word from the Threat Dictionary Choi2022 + Synonyms baseline for (a) authentic accounts and (b) inauthentic coordinated accounts. Vertical lines correspond to the Hamas attack on October 7th, 2023, and the Israel Defense Forces (IDF) entering Gaza on October 27th.
• Figure S1: Post frequency over time for the Israel-Hamas dataset. Authentic and inauthentic coordinated accounts based on (a) hashtags (main text), and (b) merged similarity networks luceri2023unmasking. Vertical lines correspond to the Hamas attack on October 7th, 2023, and the Israel Defense Force (IDF) entering Gaza on October 27th.