Research map / Large language models / AI and society
Tracking and Identifying International Propaganda and Influence Networks Online
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an LLM/embedding-based system that crawls ~40,000 international news sites, clusters stories with multilingual Matryoshka embeddings, detects stance toward topics, and infers influence networks via NETINF to track propaganda and misinformation.
Problem
Identifying, tracking, and analyzing the spread of propaganda, misinformation, and news narratives across thousands of international news websites in near real time, which manual qualitative methods cannot scale to.
Approach
Combines multilingual Matryoshka embeddings whose upper, middle, and lower dimensions encode event, topic, and theme similarity with a hierarchical online Reciprocal Nearest Neighbor clustering, then applies zero-shot multilingual stance detection and NETINF network inference over time cascades to map slanted-content influence networks.
Key results
- Matryoshka embedding approach achieves Pearson ρ = 0.833 with human similarity labels on SemEval-22-t8 test set, reported as state of the art.
- Clustering approach reported as state of the art on Miranda et al. (2018) news clustering dataset.
- Case studies: anti-vaccine websites pandemic.news and childrenshealthdefense.org found most influential by eigenvector centrality in prior work.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use multilingual Matryoshka embeddings for news article similarity | Proposed approach · SemEval-22-t8 test set | Pearson correlation | Helped |
Limitations
- Stance detection currently only works for English; multilingual model still under development.
- Keyword extraction via pointwise mutual information only works for monolingual texts.
- Many components are reported from prior or ongoing work without full details in this extended abstract.
Experimental setup
Datasets: SemEval-22-t8 Miranda et al. (2018) news clustering dataset CommonCrawl website index
Benchmarks: SemEval-22-t8 Miranda et al. (2018) news clustering dataset
Metrics: Pearson correlation
Baselines: traditional cosine-similarity embedding approaches
Builds on: Matryoshka embeddings NETINF Online Level-wise Hierarchical Clustering Reciprocal Nearest Neighbor clustering
Closest papers by meaning
- Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (AAAI 2024 · similarity 0.82)
- Learning Complex Heterogeneous Multimodal Fake News via Social Latent Network Inference (AAAI 2025 · similarity 0.82)
- Contradicted in Reliable, Replicated in Unreliable: Dual-Source Reference for Fake News Early Detection (AAAI 2025 · similarity 0.81)
- IOHunter: Graph Foundation Model to Uncover Online Information Operations (AAAI 2025 · similarity 0.81)
- Drifting Away from Truth: GenAI-Driven News Diversity Challenges LVLM-Based Misinformation Detection (AAAI 2026 · similarity 0.81)
- Harnessing Network Effect for Fake News Mitigation: Selecting Debunkers via Self-Imitation Learning (AAAI 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.