📄 中文摘要
讽刺检测因其对细致上下文理解、世界知识和多方面语言线索的依赖而面临巨大挑战，这些线索在不同讽刺表达中差异显著。现有方法，无论是微调Transformer模型还是大型语言模型，都对所有输入采用统一的推理策略，难以应对讽刺表达多样化的分析需求。这些需求涵盖了从建模反讽和语境不一致性到理解情感和意图之间的复杂交互。为解决这一问题，RAM-SD（Retrieval-Augmented Multi-agent framework for Sarcasm Detection）框架被提出，它将讽刺检测任务分解为多个子任务，并为每个子任务分配一个专门的智能体。这些智能体包括一个上下文理解智能体，负责分析输入文本的语境信息；一个世界知识检索智能体，用于获取与文本内容相关的外部世界知识；一个情感分析智能体，以识别文本中蕴含的情感极性；以及一个反讽识别智能体，专门识别言语与实际意图之间的反差。每个智能体都配备了检索增强机制，能够从外部知识库中动态获取相关信息，从而提升其对特定子任务的处理能力。智能体之间通过协作机制进行信息共享和决策聚合，最终生成一个综合的讽刺检测结果。这种模块化、智能体协作和检索增强的设计，使得RAM-SD能够更灵活、更深入地处理讽刺表达的复杂性，显著优于传统的统一推理方法，特别是在处理语境依赖性强、需要丰富世界知识和多层语言分析的讽刺案例时表现出卓越性能。
📄 English Summary
RAM-SD: Retrieval-Augmented Multi-agent framework for Sarcasm Detection
Sarcasm detection remains a significant challenge due to its reliance on nuanced contextual understanding, world knowledge, and multi-faceted linguistic cues that vary substantially across different sarcastic expressions. Existing approaches, from fine-tuned transformers to large language models, apply a uniform reasoning strategy to all inputs, struggling to address the diverse analytical demands of sarcasm. These demands range from modeling irony and contextual incongruity to understanding complex interactions between sentiment and intent. To address this, the Retrieval-Augmented Multi-agent framework for Sarcasm Detection (RAM-SD) is proposed, which decomposes the sarcasm detection task into multiple sub-tasks, assigning a specialized agent to each. These agents include a Context Understanding Agent responsible for analyzing contextual information within the input text, a World Knowledge Retrieval Agent for acquiring external world knowledge relevant to the text content, a Sentiment Analysis Agent to identify the emotional polarity embedded in the text, and an Irony Recognition Agent specifically designed to detect the contrast between literal speech and actual intent. Each agent is equipped with a retrieval-augmented mechanism, enabling dynamic acquisition of relevant information from external knowledge bases, thereby enhancing its processing capability for specific sub-tasks. Information sharing and decision aggregation occur through a collaborative mechanism among agents, ultimately generating a comprehensive sarcasm detection result. This modular, agent-collaborative, and retrieval-augmented design allows RAM-SD to handle the complexities of sarcastic expressions more flexibly and deeply, significantly outperforming traditional uniform reasoning methods, especially in cases of sarcasm that are highly context-dependent, require rich world knowledge, and demand multi-layered linguistic analysis.