Enhancing a JSCC Architecture with Retransmission Based on Semantic Feedback

Traditional communication systems have long prioritized the accurate transmission of bit sequences, focusing more on symbol-level precision than on the actual meaning of the transmitted information. This approach, while effective in many technical scenarios, can be limiting when the goal is to ensure that the receiver understands the intended message rather than its exact form. Semantic communication introduces a shift in perspective by aiming to convey meaning instead of replicating every transmitted symbol. Recent advancements in deep learning, particularly through Transformer-based architectures, have made it possible to develop systems that optimize for semantic understanding rather than syntactic accuracy. This thesis proposes the enhancement of a semantic communication system based on a Joint Source–Channel Coding (JSCC) framework for text transmission, with several modifications aimed at improving its robustness and adaptability in real-world conditions. The proposed solution includes a feedback mechanism that enables selective retransmissions and leverages reference-free quality metrics to evaluate decoding success without needing a ground truth comparison. In addition, the thesis presents an experimental evaluation of the system, carried out under different wireless channel conditions. The goal is not only to test its performance but also to observe how the feedback mechanism responds in each scenario, offering insights into its behavior. The results show the impact achieved with the proposed enhancements and provide hints towards further approaches for an effective semantic-based retransmission system.