ROSE: A Receiver-Oriented Semantic Communication Framework

Semantic communication (SemCom) is an emerging paradigm focusing on the meaning of transmitted symbols for effective communications. Existing SemCom methods necessitate the collaborative construction of a shared common dataset to train semantic encoders and decoders before communications. Unfortunately, the prerequisite limits the application of SemCom due to the difficulty of maintaining a shared dataset. Moreover, the synchronous training paradigm based on the shared dataset also confines existing SemCom to a pairwise mode. In this article, we focus on establishing SemCom in the absence of a shared dataset. We assume that communication participants each possess a non-independent and identically distributed (non-IID) private dataset. We propose a Receiver-Oriented SEmantic communication framework (ROSE). Specifically, in our framework, the receiver first independently trains her encoder and decoder, and publishes the trained encoder and decoder. When a sender expects to communicate with the receiver, he obtains the receiver’s codecs and combines them with his own private dataset to train an encoder for the upcoming communication. Through an implement case study, we demonstrated that the proposed framework can achieve state-of-the-art communication performance without shared common dataset between communication parties. We also present a further discussion on the possibilities and challenges in this new communication paradigm.