Skip to main navigation Skip to search Skip to main content

CAMC: A Conflict-Aware Multimodal Clue Model for Sarcasm Recognition

  • Xidian University
  • Northwestern Polytechnical University
  • Harbin Institute of Technology
  • Zhejiang University
  • State Key Laboratory of Extreme Photonics and Instrumentation
  • University of Liverpool

Research output: Contribution to journal › Article › peer-review

Abstract

Social-media sarcasm may be conveyed by discourse context, speaker expectations, cultural norms, or purely linguistic irony; in image–text posts, however, a frequent and experimen tally accessible cue is tension between the literal text and visible evidence. We present CAMC, a conflict-aware multimodal clue model for visually grounded multimodal sarcasm recognition. CAMC does not treat image–text contradiction as a complete theory of sarcasm; rather, it exposes contradiction as an explicit intermediate signal when sarcasm is grounded in visual context. The Cross-Modal Conflict Localizer (CMCL) extracts text-token and image-token states from the same projected Qwen2.5-VL decoder space and computes spatial conflict maps using coverage weighted residuals between text tokens and visual patches. The Gaussian Clue Generator (GCG-K) converts salient conflict peaks into scale-adaptive soft-focus views and region-anchored textual clues. A reinforcement-learning objective further optimizes prediction accuracy, output compliance, and conflict-grounded reasoning structure. On MMSD2.0, CAMC improves Qwen2.5 VL baselines by 4.65 and 4.24 points in accuracy and macro-F1 at the 3B scale, and by 4.40 and 4.98 points at the 7B scale. The ablation experiments show that explicit contradiction localization improves multimodal sarcasm recognition while retaining interpretable visual and textual evidence.

Original languageEnglish
JournalIEEE Transactions on Affective Computing
DOIs
Publication statusAccepted/In press - 2026

Keywords

  • Affective computing
  • Reinforcement learning
  • Sarcasm recognition
  • Vision language models

Cite this