Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RAM-SD: Retrieval-Augmented Multi-agent framework for Sarcasm Detection

Ziyang Zhou, Ziqi Liu, Yan Wang, Yiming Lin, Yangbin Chen

TL;DR

The paper introduces RAM-SD, a Retrieval-Augmented Multi-Agent framework for sarcasm detection that dynamically tailors reasoning to input via a meta-planner and a diverse agent ensemble. It grounds analysis with retrieved sarcasm/non-sarcasm exemplars, selects specialized reasoning plans, and synthesizes multi-view evidence into an interpretable final judgment. On four benchmarks, RAM-SD achieves a state-of-the-art Macro-F1 of 77.74%, outperforming strong GPT-4o-based baselines by 7.01 points, while providing reasoning traces that improve transparency and diagnosability. The work highlights the value of plan-conditioned, evidence-grounded reasoning for nuanced language understanding and discusses limitations around retrieval coverage and cross-agent calibration, outlining directions for future enhancements.

Abstract

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 contextual expectation violations to requiring external knowledge grounding or recognizing specific rhetorical patterns. To address this limitation, we introduce RAM-SD, a Retrieval-Augmented Multi-Agent framework for Sarcasm Detection. The framework operates through four stages: (1) contextual retrieval grounds the query in both sarcastic and non-sarcastic exemplars; (2) a meta-planner classifies the sarcasm type and selects an optimal reasoning plan from a predefined set; (3) an ensemble of specialized agents performs complementary, multi-view analysis; and (4) an integrator synthesizes these analyses into a final, interpretable judgment with a natural language explanation. Evaluated on four standard benchmarks, RAM-SD achieves a state-of-the-art Macro-F1 of 77.74%, outperforming the strong GPT-4o+CoC baseline by 7.01 points. Our framework not only sets a new performance benchmark but also provides transparent and interpretable reasoning traces, illuminating the cognitive processes behind sarcasm comprehension.

RAM-SD: Retrieval-Augmented Multi-agent framework for Sarcasm Detection

TL;DR

The paper introduces RAM-SD, a Retrieval-Augmented Multi-Agent framework for sarcasm detection that dynamically tailors reasoning to input via a meta-planner and a diverse agent ensemble. It grounds analysis with retrieved sarcasm/non-sarcasm exemplars, selects specialized reasoning plans, and synthesizes multi-view evidence into an interpretable final judgment. On four benchmarks, RAM-SD achieves a state-of-the-art Macro-F1 of 77.74%, outperforming strong GPT-4o-based baselines by 7.01 points, while providing reasoning traces that improve transparency and diagnosability. The work highlights the value of plan-conditioned, evidence-grounded reasoning for nuanced language understanding and discusses limitations around retrieval coverage and cross-agent calibration, outlining directions for future enhancements.

Abstract

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 contextual expectation violations to requiring external knowledge grounding or recognizing specific rhetorical patterns. To address this limitation, we introduce RAM-SD, a Retrieval-Augmented Multi-Agent framework for Sarcasm Detection. The framework operates through four stages: (1) contextual retrieval grounds the query in both sarcastic and non-sarcastic exemplars; (2) a meta-planner classifies the sarcasm type and selects an optimal reasoning plan from a predefined set; (3) an ensemble of specialized agents performs complementary, multi-view analysis; and (4) an integrator synthesizes these analyses into a final, interpretable judgment with a natural language explanation. Evaluated on four standard benchmarks, RAM-SD achieves a state-of-the-art Macro-F1 of 77.74%, outperforming the strong GPT-4o+CoC baseline by 7.01 points. Our framework not only sets a new performance benchmark but also provides transparent and interpretable reasoning traces, illuminating the cognitive processes behind sarcasm comprehension.
Paper Structure (43 sections, 2 equations, 4 figures, 6 tables)

This paper contains 43 sections, 2 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Traditional Feature Engineering Approaches
  4. The Deep Learning Revolution
  5. Large Language Model Approaches
  6. Methodology
  7. Stage 1: Contextual Retrieval
  8. Dual-Subset Retrieval.
  9. Rationale-Enhanced Contextualization.
  10. Stage 2: Retrieval-Augmented Meta-Planning
  11. Pre-defined Reasoning Plans.
  12. Stage 3: Agent-based Multi-View Reasoning
  13. Stage 4: Synthesis and Final Judgment
  14. Prediction and Explanation.
  15. Experiments
  16. ...and 28 more sections

Figures (4)

  • Figure 1: The four-stage architecture of the RAM-SD framework. Stage 1 grounds the query with retrieved sarcastic and non-sarcastic exemplars. This informs the Stage 2 meta-planner which selects a tailored reasoning plan. Stage 3 executes this plan with an ensemble of specialized agents. The agents' findings are synthesized in Stage 4 to produce a final judgment and explanation.
  • Figure 2: Impact of retrieval parameter $k$ on IAC-V1,V2, SemEval 2018 and MUSTARD datasets. Performance peaks at $k=3$ and plateaus for larger values.
  • Figure 3: LLM-as-Judge evaluation results. Left: Radar chart across four dimensions, RAM-SD (blue) dominates, while GPT-4o+CoC (orange) achieves competitive Specificity, surpassing RAM-SD w/o RAG (purple). Right: Grouped bar chart showing mean scores with standard deviation bars for each dimension.
  • Figure 4: Error distribution showing distinct FP/FN patterns across datasets.