This paper proposes an OFDM-based digital semantic communication (SemCom) system that is compatible with existing digital communication infrastructures. The system extracts semantics from source data, quantizes them using scalar quantizers, and transforms them into OFDM signals for transmission over frequency-selective channels. A semantic importance measurement method is introduced to establish the relationship between target tasks and semantic features. Based on semantic importance, a sub-carrier and bit allocation problem is formulated to maximize communication performance. However, the optimization objective function cannot be accurately characterized due to the neural network-based semantic codec. To address this, a low-complexity sub-carrier allocation method is proposed, assigning sub-carriers with better channel conditions to more critical semantics. Additionally, a deep reinforcement learning-based bit allocation algorithm with dynamic action space is introduced. Simulation results show that the proposed system achieves 9.7% and 28.7% performance gains compared to analog SemCom and conventional bit-based communication systems, respectively. The system model includes a semantic encoder, scalar quantizers, a channel encoder, and an OFDM transmitter for the transmitter side, and an OFDM receiver, channel decoder, de-quantizers, and semantic decoder for the receiver side. The semantic features are quantized and encoded to generate OFDM symbols for transmission. The received signal is demodulated, decoded, and de-quantized to recover the semantics, which are then used for task execution. The quality of SemCom is evaluated based on semantic distortion and task performance, which are influenced by bit allocation strategy and sub-carrier allocation policy. The optimization problem is formulated to minimize semantic distortion and maximize task performance under the maximum bit constraint. A semantic importance evaluation method is proposed, consisting of two modules: the Semantics Task Relevance (STR) module and the Inter-Semantics Relevance (ISR) module. The STR module computes the gradient of the task result with respect to the semantic features to determine their relevance to the task. The ISR module calculates the cosine similarity between semantic features to determine their inter-semantic relevance. The final semantic importance is computed as the product of the STR and ISR values. A dynamic proximal policy optimization (DPPO) algorithm is proposed for bit allocation, incorporating a dynamic action space. The DPPO algorithm consists of an Actor network and a Critic network, which work together to determine the action selection process and evaluate the quality of the chosen action. The algorithm is trained offline, with the trajectory sampling and parameter updating phases. The complexity of the DPPO algorithm is analyzed, showing that it depends on the number of semantics, the size of each semantics, and the overall number of bits. Simulation results demonstrate that the proposed OFDM-based digital SemCom system outperforms analog SemCom and conventional bit-based communication systems in terms of compression and task performance. The proposed DPPO algorithm achieves significant improvements in task performance across different signal-to-noise ratio (SNR) regimes and varying bit constraints. TheThis paper proposes an OFDM-based digital semantic communication (SemCom) system that is compatible with existing digital communication infrastructures. The system extracts semantics from source data, quantizes them using scalar quantizers, and transforms them into OFDM signals for transmission over frequency-selective channels. A semantic importance measurement method is introduced to establish the relationship between target tasks and semantic features. Based on semantic importance, a sub-carrier and bit allocation problem is formulated to maximize communication performance. However, the optimization objective function cannot be accurately characterized due to the neural network-based semantic codec. To address this, a low-complexity sub-carrier allocation method is proposed, assigning sub-carriers with better channel conditions to more critical semantics. Additionally, a deep reinforcement learning-based bit allocation algorithm with dynamic action space is introduced. Simulation results show that the proposed system achieves 9.7% and 28.7% performance gains compared to analog SemCom and conventional bit-based communication systems, respectively. The system model includes a semantic encoder, scalar quantizers, a channel encoder, and an OFDM transmitter for the transmitter side, and an OFDM receiver, channel decoder, de-quantizers, and semantic decoder for the receiver side. The semantic features are quantized and encoded to generate OFDM symbols for transmission. The received signal is demodulated, decoded, and de-quantized to recover the semantics, which are then used for task execution. The quality of SemCom is evaluated based on semantic distortion and task performance, which are influenced by bit allocation strategy and sub-carrier allocation policy. The optimization problem is formulated to minimize semantic distortion and maximize task performance under the maximum bit constraint. A semantic importance evaluation method is proposed, consisting of two modules: the Semantics Task Relevance (STR) module and the Inter-Semantics Relevance (ISR) module. The STR module computes the gradient of the task result with respect to the semantic features to determine their relevance to the task. The ISR module calculates the cosine similarity between semantic features to determine their inter-semantic relevance. The final semantic importance is computed as the product of the STR and ISR values. A dynamic proximal policy optimization (DPPO) algorithm is proposed for bit allocation, incorporating a dynamic action space. The DPPO algorithm consists of an Actor network and a Critic network, which work together to determine the action selection process and evaluate the quality of the chosen action. The algorithm is trained offline, with the trajectory sampling and parameter updating phases. The complexity of the DPPO algorithm is analyzed, showing that it depends on the number of semantics, the size of each semantics, and the overall number of bits. Simulation results demonstrate that the proposed OFDM-based digital SemCom system outperforms analog SemCom and conventional bit-based communication systems in terms of compression and task performance. The proposed DPPO algorithm achieves significant improvements in task performance across different signal-to-noise ratio (SNR) regimes and varying bit constraints. The