Scholarlink

<p>The rapid spread of fake news highlights the need for automated detection systems. While current methods show promise, strong performance alone is insufficient, as explainability is equally important. Many scholars have tackled the challenge of explainability, but their solutions often assume users have the technical knowledge to understand them. This paper presents a novel Multi-Level Model-Agnostic Post-Hoc Explanations (MAPE) framework to improve interpretability in fake news detection. MAPE addresses the challenge of explaining complex model predictions by offering tiered explanations, letting users choose their preferred level of detail. It leverages eXplainable Artificial Intelligence (XAI) techniques and Large Language Models (LLMs) to generate explanations ranging from simple feature highlights to rich, context-driven narratives. Evaluation results from both humans and LLMs show that detailed explanations enhance clarity and persuasiveness, outperforming technical ones. LLM assessments also align closely with human judgments across most quality aspects.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explain further: multi-level explanations for fake news detection using large language models

  • 期刊论文

  • 发表日期：

    2025年8月2日

查看全文
  • Dorsaf Sallami,
  • Esma Aïmeur

摘要

The rapid spread of fake news highlights the need for automated detection systems. While current methods show promise, strong performance alone is insufficient, as explainability is equally important. Many scholars have tackled the challenge of explainability, but their solutions often assume users have the technical knowledge to understand them. This paper presents a novel Multi-Level Model-Agnostic Post-Hoc Explanations (MAPE) framework to improve interpretability in fake news detection. MAPE addresses the challenge of explaining complex model predictions by offering tiered explanations, letting users choose their preferred level of detail. It leverages eXplainable Artificial Intelligence (XAI) techniques and Large Language Models (LLMs) to generate explanations ranging from simple feature highlights to rich, context-driven narratives. Evaluation results from both humans and LLMs show that detailed explanations enhance clarity and persuasiveness, outperforming technical ones. LLM assessments also align closely with human judgments across most quality aspects.