This paper presents a novel framework for next-generation semantic communication. Semantic communication prioritizes the conveyance of the meaning of a message over the transmission of raw data, and the main challenge is to accurately identify and extract important semantic information without performance degradation when the receiver's goals change over time. The proposed framework can flexibly adapt to various receiver tasks by jointly capturing task-specific and contextual information, and demonstrates superior performance (improved downstream task performance, improved generalization performance, ultra-high bandwidth efficiency, and low reconstruction latency) to existing studies through experiments on image datasets and computer vision tasks.
