# Detection of Manipulative Political Narratives Based on Large Language Models: Identifying Information Manipulation in Social Media

> This article introduces a new framework for detecting and structuring manipulative political narratives using large language models. Through a combination of few-shot prompt filtering, UMAP dimensionality reduction, HDBSCAN clustering, and narrative extraction via reasoning models, it successfully identified 41 distinct manipulative narrative patterns from 1.2 million social media posts.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-05-14T04:30:21.000Z
- 最近活动: 2026-05-15T04:20:46.949Z
- 热度: 136.2
- 关键词: 大语言模型, 政治叙事检测, 社交媒体分析, 无监督聚类, HDBSCAN, UMAP, 信息操控, 自然语言处理
- 页面链接: https://www.zingnex.cn/en/forum/thread/llm-arxiv-2605-14354v1
- Canonical: https://www.zingnex.cn/forum/thread/llm-arxiv-2605-14354v1
- Markdown 来源: floors_fallback

---

## Introduction: Framework for Detecting Manipulative Political Narratives Based on Large Language Models

This article proposes an innovative framework combining the reasoning capabilities of large language models with unsupervised learning. Through a combination of few-shot prompt filtering, UMAP dimensionality reduction, HDBSCAN clustering, and narrative extraction, it successfully identified 41 manipulative narrative patterns from 1.2 million social media posts, aiming to address the severe challenge of information manipulation in the social media era.

## Background: New Challenges of Political Discourse in the Social Media Era

As political discussions shift to social media, manipulative political narratives are rampant. Characterized by distorted facts and emotional incitement, they blur the line with legitimate criticism and their forms are constantly evolving, making traditional keyword or rule-based methods difficult to identify them effectively.

## Research Objectives and Core Challenges

The core objective is to build a computational framework for automatically detecting and structuring manipulative narratives. Three key challenges need to be addressed: distinguishing manipulative narratives from legitimate criticism, discovering unknown narrative types, and having the ability to process data at the million-scale level.

## Technical Solution: Multi-stage Intelligent Processing Pipeline

A four-stage process is adopted: 1. Few-shot prompt filtering (using positive and negative example contrastive learning to screen manipulative posts); 2. Semantic embedding + UMAP dimensionality reduction (preserving local structure); 3. HDBSCAN clustering (automatically determining categories and identifying noise); 4. Narrative theme extraction via large language models (enhancing interpretability).

## Experimental Validation: Practical Results on Million-scale Data

Applied to 1.2 million social media posts, it successfully identified 41 manipulative narrative clusters and discovered new, undefined narrative patterns, proving the framework's effectiveness and the potential for human-machine collaboration.

## Application Value and Social Significance

It contributes to platform governance (content moderation), academic research (analyzing political discourse patterns), and public education (improving media literacy), protecting users from information manipulation at multiple levels.

## Limitations and Future Directions

Limitations include language dependence, cultural context differences, and adversarial evolution. Future directions: multilingual expansion, real-time stream processing, integration with platform policies, and improving system transparency and interpretability.

## Conclusion: Technology and Society Collaborate to Build a Healthy Information Ecosystem

It emphasizes the potential of artificial intelligence technology to address information manipulation, which needs to be coordinated with social governance, public education, and platform responsibility. Large language models are tools that require human wisdom and ethical judgment to guide their application.