近日，SIGIR 2025 公布了录用论文列表，博士生卢俊宇关于仇恨模因检测的研究成果被录用为长文。SIGIR 是信息检索与数据挖掘领域的顶级会议，被CCF推荐为A类国际学术会议，在学术界及业界均享有盛誉。这次会议共收到1105篇长文投稿，其中有238篇论文被录用，录用率为21.5%。

题目：Is Having Rationales Enough? Rethinking Knowledge Enhancement for Hateful Meme Detection

Abstract: Hateful memes are prevalent on the Internet, raising the urgent need for effective detectors.  Given their implicit nature, incorporating rationales with background knowledge is essential for enhancing detectors' understanding of memes.  However, current methods often struggle with the limited quality of external rationales and the lack of alignment with original meme information.  These challenges can hinder the detector's understanding capability, resulting in reduced accuracy and diminished decision explainability. To address these challenges, we propose a novel Multimodal Multi-agent Knowledge Enhanced (M2KE) framework for hateful meme detection. M2KE introduces a multi-agent rationale discovery mechanism to extract high-quality rationales relevant to meme content and an adaptive knowledge interaction mechanism to ensure alignment between the original meme information and external rationales. Specifically, multi-agent rationale discovery mechanism improves the reliability of rationales by collaboratively verifying and refining them with multiple agents, supported by large language models (LLMs) due to their extensive encoded knowledge. And adaptive knowledge interaction mechanism uses information entropy to dynamically balance the detector's attention between original meme information and external rationales, preventing over-reliance on external rationales and enabling a more comprehensive understanding of meme content. Experimental results on three public hateful meme datasets demonstrate that M2KE significantly outperforms existing state-of-the-art models.  Further analysis underscores the importance of effectively integrating accurate rationales to enhance model performance.

中文摘要：互联网上充斥着仇恨模因，亟需构建高效的检测模型。由于模因的含义通常比较隐晦，因此结合带有背景知识的理据对于提升模型的理解能力至关重要。然而，现有方法普遍面临理据质量有限，以及缺乏与原始模因信息进行对齐的问题。这些挑战会影响模型的理解能力，从而降低检测精度，同时削弱决策的可解释性。为应对上述挑战，我们提出了一种新颖的多模态多智能体知识增强（M2KE）框架，用于仇恨模因检测。M2KE 引入了一个多智能体理据发现机制，以提取与模因内容高度相关的高质量理据，并设计了一个自适应知识交互机制，以确保原始模因信息与外部理据之间的对齐。具体而言，多智能体理据发现机制通过多个智能体的协同验证与优化来提升理据的可靠性，这些智能体基于大模型构建，充分利用其丰富的背景知识和推理能力。而自适应知识交互机制则利用信息熵动态平衡模型对原始模因信息与外部理据的关注度，避免模型对外部理据的过度依赖，从而实现对模因内容的全面理解。在三个公开的仇恨模因数据集上的实验结果表明，M2KE 显著优于现有的主流方法。进一步的分析也强调了准确整合理据对于提升模型性能的重要性。