Knowledge Graph Based Semantic Communication Networks with Deep Reinforcement Learning Enhancement

With the advancements in physical layer technologies such as source coding and channel coding techniques, the spectral efficiency of traditional communication systems has approached the Shannon limit. Exploring semantic communication paradigms becomes essential to address the challenges of transcending bit-level accuracy, meeting the demands of next-generation intelligent applications such as augmented reality, digital twins, and the tactile internet, which effectively conveys meaning and context in data transmission. Therefore, we propose a Self-Recovery Semantic Communication (SR-SC) architecture, which leverages Knowledge Graph Reasoning (KGR) encoder/decoder, a semantic restorer, and a semantic reducer to enhance the performance of communication networks. To model the correlations between semantic symbols, the KGR employs an efficient semantic encoding scheme based on TransR, enabling the system to represent semantic differences among entities and relationships as semantic distances in a low-dimensional vector space. Additionally, the semantic restorer and reducer are modeled as reinforcement learning tasks, with a Deep Q-Network (DQN) serving as the decision-making agent to incrementally restore and reduce semantic entities and relationships. This approach facilitates subsequent semantic reasoning. Finally, extensive simulation results on public datasets demonstrate that the proposed SR-SC architecture significantly reduces the semantic symbol error rate and enhances the effectiveness and reliability of the semantic communication networks.