TL;DR
Sarcasm-R1 is a new framework that treats sarcasm detection as a "System 2" reasoning task rather than a simple classification problem. By decomposing sarcasm into three dimensions—language, context, and emotion—and training with Reinforcement Learning (GRPO), the model achieves state-of-the-art performance across five major benchmarks, proving that emotional contrast is the true heart of sarcastic intent.
What Problem Does This Paper Address?
Detecting sarcasm is notoriously difficult for AI because it requires "System 2" thinking: slow, deliberate, and multi-step reasoning. While Large Language Models (LLMs) excel at "System 1" tasks like basic sentiment analysis, they often take sarcastic text at face value.
The core challenge is that sarcasm isn't just about the words used; it’s about the gap between those words and the speaker's actual intent. Existing methods usually rely on simple supervised learning or basic prompting, which often fail to capture the complex interplay between linguistic cues (like hyperbole), contextual background (like knowing a person just got injured), and the emotional "flip" that defines a sarcastic remark.
Methods
The researchers from the Dalian University of Technology introduced Sarcasm-R1, a model based on Qwen2.5-7B that is specifically trained to "think" through sarcasm using three distinct lenses:
- Linguistic Dimension: Focuses on rhetorical devices, keywords, and punctuation.
- Contextual Dimension: Examines the topic, cultural background, and commonsense knowledge.
- Emotional Dimension: Targets the reversal of emotions (e.g., using positive words to mask a negative situation).
The "R1" Secret Sauce: GRPO
Borrowing a strategy from the DeepSeek-R1 playbook, the authors used Group Relative Policy Optimization (GRPO). Unlike standard Reinforcement Learning that requires a complex "critic" model, GRPO evaluates a group of outputs relative to each other.
To make this work, they built SarGRM (Sarcasm Group Reward Models)—three specialized reward models (LinGRM, ConGRM, and EmoGRM) that judge how well the model reasoned in each of the three dimensions. The training happened in three stages:
- Stage 1: Teaching the model the correct output format.
- Stage 2: Using the specialized reward models to refine reasoning quality.
- Stage 3: A final "boost" combining reasoning quality with a correctness reward for the final classification.
Datasets
The authors didn't just test on existing data; they created a new reasoning-focused benchmark:
- Sarcasm-Reason: A newly annotated dataset derived from five mainstream benchmarks (Ghosh, SARC, IAC-V2, iSarcasm, and SemEval2018).
- Synthesis & Validation: They used GPT-4o to generate reasoning paths, which were then cross-verified by a "jury" of other top-tier models (Claude 3.5 Sonnet, Gemini 2.0 Flash, and Qwen-Max) and human reviewers to ensure high data quality.
Key Insights and Ideas
- Emotional Reversal is the Core: The study found that while linguistic and contextual cues provide the "evidence," the emotional contrast (the "reversal") is the most critical factor in identifying sarcasm. When the emotional reasoning was removed, the model's performance took a significant hit.
- Reasoning Beats Prompting: Simply asking a model to "think step-by-step" via prompting (Chain-of-Thought) often leads to hallucinations in sarcasm detection. By using RL to ground the reasoning process in specific dimensions, Sarcasm-R1 becomes much more reliable.
- Inference-Time Scaling: The model shows "scalability"—meaning if you allow it to generate more reasoning paths (using majority voting with $K=16$ samples), its accuracy continues to improve.
Key Results
Sarcasm-R1 outperformed nearly every baseline, including specialized sarcasm models (like EICR and SensoryT5) and general-purpose giants like GPT-4o.
- iSarcasm Dataset: On this particularly difficult, imbalanced dataset, Sarcasm-R1 achieved 85.4% accuracy and a 72.3% Macro-F1 score, dwarfing GPT-4o’s 76.0% accuracy and 59.4% F1.
- Ablation Success: Removing the "Linguistic" reasoning dimension caused the largest drop in Macro-F1 (7.6%), showing that while emotion is the core, linguistic cues are essential for the model to balance its judgment between sarcastic and non-sarcastic samples.
Limitations and Future Work
The authors acknowledge a few hurdles:
- Subjectivity: Sarcasm is inherently subjective. What one person considers a sarcastic joke, another might see as a literal statement or a direct insult.
- Missing Dimensions: The current three-dimension model might still miss nuances like social hierarchy or specific interpersonal relationships between speakers.
- Text-Only: This model is currently limited to text. Sarcasm in the real world often relies on vocal tone or facial expressions, suggesting that a multimodal version of Sarcasm-R1 is the logical next step.
Summary updated