Abstract: The coexistence of multimodal semantic communication and ultra-reliable low-latency communication (URLLC) creates a resource management challenge for next-generation wireless networks. Semantic communication reduces transmission redundancy by delivering task-relevant information rather than complete bit streams, whereas URLLC services require immediate scheduling and strict latency guarantees. When these two types of traffic share the same downlink orthogonal frequency-division multiple access (OFDMA) system, bursty URLLC packets may need to puncture the resource blocks originally allocated to semantic users. Although such puncturing can satisfy the urgent URLLC transmission requirements, it may also damage semantic features and degrade the performance of downstream intelligent tasks. Therefore, an effective scheduling mechanism should not only respond to URLLC traffic in real time but also distinguish the unequal importance of semantic information carried by different physical resources. This study investigates the dynamic resource allocation problem in a multimodal semantic communication and URLLC coexistence network, with the aim of maximizing the quality of experience (QoE) of semantic users while satisfying URLLC latency constraints. To this end, an importance-aware dynamic resource-puncturing scheme based on deep reinforcement learning is proposed. To overcome the limitations of conventional bit-level resource allocation, an integrated gradient (IG)-based attribution method is introduced to evaluate the contribution of semantic features to the final task output. The obtained feature importance values are mapped onto physical resource blocks to form a semantic importance weight matrix. This matrix enables the scheduler to identify the resource blocks carrying critical semantic features and avoid puncturing them whenever possible. The original optimization problem was formulated by jointly considering the semantic transmission rate, semantic distortion, channel allocation, power constraints, puncturing exclusivity, and URLLC delay requirements. Because the problem is nonconvex and combinatorial, it is decomposed into two subproblems. In the first stage, the channel and power resources are pre-allocated to semantic users according to their rate requirements. In the second stage, the dynamic puncturing decision caused by URLLC packet arrival is modeled as a Markov decision process. The system state includes the URLLC queue length, waiting time of the head-of-line packet, semantic importance matrix, and remaining puncturing tolerance of semantic users. Based on this state information, a proximal policy optimization (PPO) agent learns whether to puncture each minislot and which frequency resource should be selected for URLLC transmission. A reward function was designed to reflect the tradeoff between semantic service protection and URLLC latency satisfaction. It penalizes excessive puncturing, unacceptable semantic distortion, and URLLC packet timeouts, thereby guiding the agent to avoid highly important semantic resource blocks while transmitting urgent URLLC packets within the required delay bounds. In addition, by observing the remaining puncturing budget, the agent learns to distribute the puncturing pressure among different semantic users instead of repeatedly damaging the same user’s resources. Simulation results demonstrate that the proposed IG-assisted PPO method achieves a better overall performance than random puncturing, several greedy heuristics, double deep Q-network, PPO without IG, and attention-assisted PPO. Under different URLLC arrival probabilities and semantic compression ratios, the proposed method maintains the lowest semantic outage percentage and achieves the highest average total reward. The training curves also exhibit stable convergence behavior in terms of policy loss, value loss, Kullback–Leibler (KL) divergence, and policy entropy. Furthermore, image classification experiments on the STL-10 dataset confirm that the proposed method preserves the downstream task accuracy with increasing URLLC traffic intensity better than the other methods. These results indicate that combining semantic feature attribution with reinforcement learning is an effective solution for dynamic resource allocation in semantic–URLLC coexistence networks.