New Chinese Multimodal Sarcasm Detection Benchmark CFMS Introduces Fine-Grained AI Training

A new research benchmark called CFMS has been developed specifically for detecting sarcasm in Chinese social media content. This dataset contains 2,796 carefully selected image-text pairs from online platforms. Unlike previous models, CFMS provides three layers of annotation: identifying sarcasm, recognizing its target, and generating explanations for why content is sarcastic. Researchers found these detailed annotations help artificial intelligence systems create images that clearly convey sarcastic intent. The project also includes a parallel subset of 200 Chinese and 200 English metaphor examples, revealing current AI models struggle significantly with metaphorical reasoning. To address limitations of traditional retrieval approaches, the team proposes a Reinforcement Learning-augmented method. The work aims to advance fine-grained semantic understanding in multimodal AI systems, particularly for culturally specific content like Chinese social media. The research was announced on arXiv with identifier 2604.16372v1.