In order to achieve high-reliability transmission and high-fidelity reception of important semantics over long-distance or unfavorable channel conditions, a Reconfigurable Intelligent Surface (RIS)-assisted semantic communication system (RISSCS) is proposed in this letter. To enhance the performance of RIS-SCS, an optimization problem on maximizing the semantic fidelity is formulated. This requires a cross-layer design of the beamforming in the physical layer and the network training in the semantic layer, which is intractable due to the different mechanisms and mutually coupled variables. To provide insights, an efficient hierarchical optimization framework is proposed, in which the active pre-coding beamforming vector at the source and the passive phase shifts at the RIS are first jointly optimized to form favorable channels, based on which the corresponding semantic neural networks are trained in an end-to-end manner. Simulation results verify the performance gains achieved by the proposed RIS-SCS. Specifically, under unfavorable channel conditions with severe path-loss attenuations or in power constrained regime, the proposed RIS-SCS demonstrates superior performance and robustness compared to the case without RIS.