Abstract: Aiming at the spectrum resource competition caused by the coexistence of multi-modal semantic communication and Ultra-Reliable and Low-Latency Communications (URLLC) services in downlink Orthogonal Frequency Division Multiple Access (OFDMA) systems for next-generation wireless networks, this paper investigates how to maximize the Quality of Experience (QoE) of semantic users while satisfying the strict latency constraints of URLLC. An importance-aware dynamic resource puncturing scheme based on Deep Reinforcement Learning (DRL) is proposed. First, to address the inability of traditional bit-level resource allocation to adapt to semantic task characteristics, an attribution analysis method based on Integrated Gradients is introduced. This establishes a fine-grained semantic-to-physical resource mapping model, quantifies the contribution of semantic features to the accuracy of downstream intelligent tasks, and constructs a semantic feature importance weight matrix. Second, a coexistence model for semantic communication and URLLC is established, decoupling the complex non-convex combinatorial optimization problem into two sub-problems: channel allocation and dynamic puncturing. In the channel allocation stage, a water-filling algorithm is adopted for pre-allocation based on semantic user rate requirements. In the puncturing stage, the problem is modeled as a Markov Decision Process (MDP), and an agent based on the Proximal Policy Optimization (PPO) algorithm is designed. Based on real-time semantic feature importance, remaining allowable puncture counts, and URLLC queue status, the agent dynamically determines the transmission positions of URLLC packets, seeking an optimal strategy to balance avoiding the destruction of critical semantic information and guaranteeing URLLC latency. Simulation results demonstrate that, compared with random puncturing and greedy strategies, the proposed algorithm accurately identifies and avoids high-weight semantic resource blocks. Under varying URLLC traffic intensities and semantic compression ratios, it maintains the lowest semantic interruption percentage and a high average total reward, achieving efficient dynamic resource allocation between heterogeneous services.