In high-speed railway communication scenarios, most of the computing tasks generated by trains are compute-intensive and delay-sensitive. Multi-access mobile edge computing (MEC) technology is very promising in resolving the contradiction between this situation and the limited local resources. Traditionally, MEC servers are deployed at base stations along the railway. We introduce dynamic MEC servers assisted by unmanned aerial vehicles (UAVs) additionally to deal with excessive server loads and the real-time nature of trains and tasks. Moreover, to complete real-time tasks, trains usually need abundant resources to transmit a large amount of raw data to MECs. However, limited local resources cannot meet the needs. Therefore, we propose a semantic-assisted communication model, introducing semantic encoding, decoding, extraction and feature aggregation to reduce resource overhead and computing delay while ensuring the accuracy of semantic understanding. Next, we conduct research on multi-dimensional resource allocation with the goal of minimizing delays and maximizing cache resource utilization by presenting our problem formulation. We propose a scheme based on Multi-Agent Deep Deterministic Policy Gradient (MADDPG) to enable MEC agents to collaborate in offloading tasks and allocate resources. Lastly, we conduct simulation experiments and compare with two schemes: centralized resource management and random offloading resource allocation, verifying that the proposed scheme can achieve higher resource utilization and QoS satisfaction rate.