A knowledge base architecture and adaptive construction method for semantic communication

Abstract

The application discloses a knowledge base framework and an adaptive construction method for semantic communication and belongs to the technical field of semantic communication. The application considers link level and system level communication. The link level comprises task, signal source and channel semantic knowledge bases. The system level comprises scheduling semantic knowledge bases. The application fully considers the performance of existing communication equipment. Before data transmission starts, a transceiver selects a corresponding task node from a hardware-matched task semantic knowledge base based on a graph database according to a communication task, and is matched to an optimal signal source semantic knowledge base. Then, the transceiver extracts, recovers and encodes and decodes semantics according to prior information provided by the signal source semantic knowledge base to obtain semantic code words. In order to resist the influence of wireless fading, the application introduces a channel semantic knowledge base before and after the transceiver transmits and receives the semantic code words, perceives channel semantics and realizes semantic enhancement. The application comprehensively considers hardware performance, efficiently matches communication tasks, perceives a wireless channel environment, reduces communication bandwidth and simultaneously improves multi-scene and multi-task performance.

Description

Technical Field

[0001] This invention relates to a knowledge base architecture and adaptive construction method for semantic communication, belonging to the field of semantic communication technology, and is applicable to smart devices in 6G mobile communication systems. Background Technology

[0002] Compared to 5G, 6G mobile communication systems represent a paradigm shift from "Internet of Everything" to "Intelligent Internet of Everything," an evolution that demands native intelligence in communication devices. Thanks to the rapid development of artificial intelligence and its breakthroughs in semantic extraction, semantic communication technology has become one of the key enabling technologies supporting future natively intelligent devices. Unlike traditional communication technologies that focus on the accurate and efficient transmission of symbols, semantic communication emphasizes the efficient transmission of semantic information from the source. This new communication paradigm can significantly reduce the bandwidth requirements of the system under equivalent task performance. From a system architecture perspective, a semantic communication system mainly consists of core modules such as a semantic extractor, a semantic restorer, a semantic codec, and a semantic knowledge base. The semantic knowledge base provides prior information for the semantic communication system and is a key technology for realizing semantic communication. Essentially, it utilizes the storage and computational overhead of the transceiver to exchange for communication resources, including power and bandwidth. Simultaneously, the semantic knowledge base provides semantic recovery capabilities for the semantic communication system, improving task performance in extreme wireless environments.

[0003] Current research on semantic knowledge bases mainly focuses on two directions: task semantic knowledge bases and source semantic knowledge bases in point-to-point communication scenarios. Task semantic knowledge bases often rely on large-scale foundational models such as large language models, using prompt word engineering to provide task descriptions and contextual information for task selection. However, these solutions place high demands on the storage and computing power of communication devices, have long inference times, and perform poorly in task selection accuracy and multi-task switching scenarios. Research on source semantic knowledge bases is mostly limited to single tasks and single data modalities, lacking universal support for diverse semantic communication tasks. Furthermore, effectively suppressing the interference of wireless multipath fading channels on semantic information is a key issue that urgently needs to be addressed to improve the performance of semantic communication systems. Therefore, there is an urgent need to propose a more comprehensive knowledge base architecture and adaptive construction method for semantic communication, capable of quickly adapting to different communication scenarios, reducing bandwidth overhead, resisting wireless fading, and improving task accuracy while reducing the storage and computing costs of communication devices. This innovative semantic knowledge base solution will provide important support for the comprehensive development of semantic communication technology and promote the practical deployment and application of intelligent communication systems in the 6G era.

[0004] In summary, traditional communication technologies focus on the accurate transmission of symbols, while semantic communication emphasizes the efficient delivery of semantic information, significantly reducing bandwidth requirements. Existing semantic knowledge base research suffers from problems such as limited scenario scope, slow task switching, and poor anti-interference capabilities. This invention addresses these issues by introducing a dynamic perception mechanism for channel semantic knowledge bases, combined with a hierarchical task knowledge base design, significantly improving semantic fidelity and reducing storage overhead on resource-constrained devices, thus adapting to the semantic communication needs of multiple scenarios and multiple modalities. Summary of the Invention

[0005] Purpose of the invention: This invention provides a knowledge base framework and adaptive construction method for semantic communication. It addresses the problems of existing semantic knowledge bases, such as limited scenario, slow task switching, poor anti-interference ability, difficulty in updating the knowledge base, and limited hardware performance, for semantic communication tasks with multiple scenarios, multiple modalities, and multiple requirements, thereby improving the development of semantic knowledge bases.

[0006] Technical solution: The technical solution adopted in this invention specifically includes the following steps:

[0007] Step 1: The framework proposed in this invention divides the semantic knowledge base architecture for semantic communication into link-level and system-level architectures. The link-level semantic knowledge base includes task SKB. T Source SKB S and Channel SKB C Semantic knowledge base. The task semantic knowledge base stores semantic communication tasks, forming a task network with tasks as nodes and modalities and types between tasks as edges. Each task node carries the source semantic knowledge base responsible for processing that task, which is divided into explicit and implicit types. The explicit source semantic knowledge base mainly includes a set of feature vectors and a set of triples, while the implicit source semantic knowledge base mainly includes pre-trained neural network parameters. The channel semantic knowledge base is used to perceive channel semantic information, thereby improving semantic fidelity at the receiving end. The system-level semantic knowledge base includes the scheduling semantic knowledge base SKB. D It is responsible for tasks such as resource scheduling during system-level communication;

[0008] Step 2: This invention fully considers the hardware conditions of communication equipment and designs a hierarchical task semantic knowledge base architecture. When hardware resources are sufficient, the task semantic knowledge base is divided into multiple levels according to the task category, such as a three-level knowledge base divided by modality, reconstruction / goal orientation, and task specific description. Modality, reconstruction / goal orientation, and task specific description are respectively called level one, level two, and level three tasks, and each level three task node is equipped with a corresponding source semantic knowledge base. When hardware resources are limited and the device is located at the transmitting end, the task semantic knowledge base is divided into fewer levels according to the task category, such as a two-level knowledge base divided by modality and reconstruction / goal orientation. Different high-level tasks belonging to the same low-level task node use the same source semantic knowledge base at the resource-limited transmitting end.

[0009] Step 3: Based on the performance resources of communication equipment, such as storage and computing power, equip different devices with matching semantic knowledge bases. The communication system equipped with the framework semantic knowledge base described in Step 1 allocates resources such as power and spectrum to both parties in the link-level communication based on the scheduling semantic knowledge base and the current system state.

[0010] Step 4: Before communication, the two communicating parties equipped with the link-level semantic knowledge base described in Step 1 determine the communication task T according to a predetermined protocol, and respectively index task nodes from the knowledge graph-based task semantic knowledge base, selecting the best source semantic knowledge base to complete the current task. The above process is denoted as SKB. S =f(T;SKB T );

[0011] Step 5: The transmitter utilizes the source semantic knowledge base SKB Sen The provided prior information and semantic encoder efficiently extract and encode semantic information Z related to the current task from the information source S. The above process is denoted as Z = g en (S;SKB Sen );

[0012] Step 6: The transmitter utilizes the channel semantic knowledge base SKB Cen The provided predefined mapping relationships map the task-related semantic information Z extracted in step 5 into floating-point training symbols, used to indicate source semantics with a small amount of data, and simultaneously represent the semantic information Z as floating-point semantic codewords suitable for transmission on the current wireless channel. Furthermore, the channel semantic knowledge base provides a quantizer q. en (·), this quantizer quantizes the floating-point semantic codeword and training symbol to obtain the bit-form semantic codeword X and training symbol P, adapting to modern wireless communication systems. The above process is denoted as X, P = q. en (h en (Z;SKB Cen ));

[0013] Step 7: The transmitter uses modulation, MIMO-OFDM, up-conversion, and other technologies to convert the semantic codeword X and training symbol P into electromagnetic waves for transmission into the wireless channel. The receiver then performs down-conversion, sampling, synchronization, channel estimation, signal detection, demodulation, and other techniques to obtain the damaged semantic codeword. and training symbols Additional prior information provided by the knowledge base is transmitted via the control channel in the form of an index.

[0014] Step 8: The receiver selects a dequantizer q with the same quantization order as the transmitter. de (·) For damaged semantic codewords and training symbols Dequantization is performed to obtain the damaged semantic codewords and training symbols in floating-point form. The receiver retrieves the channel semantic knowledge base (SKB) based on the predefined training symbol index received from the control channel. Cde The training symbol P is used, and the received damaged training symbol P is used as input to the semantic enhancer in the channel semantic knowledge base. This is used to simultaneously perceive channel semantic information and enhance the damaged semantic information to obtain enhanced semantic information. The above process is denoted as

[0015] Step 9: The receiver utilizes the source semantic knowledge base SKB. Sde The provided semantic decoder decodes the damaged semantic information, and then, with the current communication task as the objective, uses the optimal task executor or source reconstructor to complete the semantic communication task. This process is denoted as...

[0016] As a further technical solution of the present invention, in step 1, the semantic knowledge base is mainly stored in the communication device in the form of graph structure, database, neural network parameters, etc.

[0017] As a further technical solution of the present invention, in step 2, the updating of the task semantic knowledge base mainly involves adding or removing task nodes. When the semantic communication task is not a task node in the transceiver task semantic knowledge base, the knowledge base adds or removes task nodes through node addition / removal instructions to update the task semantic knowledge base. Then, the dataset involved in the task is collected, and the task performance and bandwidth overhead are used as loss functions to train and update the source semantic knowledge base in an online or offline manner. The task semantic knowledge base is presented in the form of a knowledge graph or other graph databases involving entities, relationships, and attributes.

[0018] As a further technical solution of the present invention, in step 4, the semantic communication task is presented in the form of natural language, that is, the communication task is described in the form of text or voice. If it is in the form of voice, it is first converted into text after being entered into the communication device, and then the corresponding task index is searched in the task semantic knowledge base using text, a syntax analyzer, and a keyword search algorithm.

[0019] As a further technical solution of the present invention, in step 5, the organization of the explicit source semantic knowledge base can involve not only feature vector sets and triple sets, but also large-scale basic models. The unified token generated by these basic models after modal alignment of different modal data can also serve as a source semantic knowledge base for a specific task.

[0020] As a further technical solution of the present invention, in step 6, the neural network provided by the channel semantic knowledge base maps source semantics to training symbols in units of OFDM symbols or resource blocks, and maps all or part of the semantic information in the unit to rule training symbols according to task performance and user requirements. The parameters in the mapping function are initialized in a random manner, and the index of the current mapping result is searched in the training symbol codebook as a small data volume indicator of semantic information.

[0021] As a further technical solution of the present invention, in step 8, the semantic enhancer provided by the channel semantic knowledge base enhances the semantic codewords on the current unit subcarrier based on OFDM symbols or resource blocks, according to task performance and user requirements. The semantic enhancer is presented in the form of a neural network, with its input being the damaged semantic codewords, the original training symbols, and the damaged training symbols, and its output being the enhanced semantic information. The loss function is to maximize task performance and minimize bandwidth overhead. The semantic enhancer embeds a channel semantic perceptron, which can be further used to display and feedback channel semantics, guiding the online updating of the source and the channel semantic knowledge base.

[0022] As a further technical solution of the present invention, in step 9, the network parameters of the task executor and the source reconstructor are trained end-to-end offline and directly deployed in the communication equipment. During communication, they can be updated online according to task requirements and channel environment.

[0023] Beneficial effects: Compared with the prior art, the present invention, by adopting the above technical solution, has the following beneficial effects:

[0024] 1. This invention comprehensively considers both link-level and system-level communication scenarios, constructing a relatively comprehensive semantic knowledge base architecture that can serve different modules of semantic communication. This architecture can directly adapt to link-level and system-level communication needs, and compared with existing semantic knowledge base research, it is more comprehensive and practical in terms of coverage and application scenarios.

[0025] 2. This invention considers the performance limitations of communication equipment and constructs a multi-level task semantic knowledge base based on a graph database. Each node in the graph is a task node, and the hardware matching source semantic knowledge base for processing that task is attached. When a performance-constrained device acts as a transmitter, semantic extraction and encoders for the same type of task use the same structure and parameters; in other cases, different structures and parameters are used. Compared with existing methods, the task and source semantic knowledge base construction method of this invention greatly reduces the storage and computing power requirements of the device, while reducing task switching time and semantic knowledge base update costs, fully leveraging the role of the source semantic knowledge base. The channel semantic knowledge base construction method of this invention is a flexible and feasible channel semantic awareness and semantic enhancement scheme, effectively resisting wireless fading and improving semantic fidelity, and adapting to devices with different hardware performance through hierarchical task knowledge base design. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the semantic knowledge base architecture for semantic communication proposed in this invention.

[0027] Figure 2 This is a schematic diagram of the hierarchical division of the task semantic knowledge base proposed in this invention;

[0028] Figure 3 This is a schematic diagram of the link-level semantic knowledge base organization form for semantic communication proposed in this invention. Detailed Implementation

[0029] This invention designs a knowledge base framework and adaptive construction method for semantic communication. The specific implementation of this invention will be described below with reference to the accompanying drawings.

[0030] Figure 1 The semantic knowledge base architecture designed for this invention considers both link-level and system-level communication. The semantic knowledge base involved in link-level communication includes the task SKB. T Source SKB S and Channel SKB C The semantic knowledge base includes a task semantic knowledge base that stores semantic communication tasks. Tasks are organized into a task network as nodes, with edges formed by the modality and type relevance of the tasks. The source semantic knowledge base provides prior information for specific tasks, with different task source semantic knowledge bases mounted on different task nodes. The channel semantic knowledge base provides prior information for channel semantics, primarily generating semantic training symbols for the source at the transmitting end and performing channel semantic awareness and semantic enhancement using impaired semantic training symbols at the receiving end. The semantic knowledge bases involved in system-level communication include the scheduling semantic knowledge base (SKB). D It provides efficient resource scheduling support for various components and devices in the communication system by intelligently managing system information such as base station location, user distribution, and total power, and combining this with the current system status.

[0031] Because system-level scheduling semantic knowledge bases involve many factors and complex situations, this invention designs and implements a link-level semantic knowledge base {SKB}. T SKB S SKB C The implementation process of a link-level semantic knowledge base includes:

[0032] Step 1: Design and implement the task semantic knowledge base SKB T . Figure 2This diagram illustrates the hierarchical division of the hardware matching task semantic knowledge base designed for this invention. Taking communication between a base station and a user as an example, the base station equipment by default has large storage and strong computing power, while the user equipment's performance is relatively weak. During uplink communication, the user equipment acts as the transmitter, performing semantic extraction and semantic encoding, while the base station acts as the receiver, performing semantic decoding and task execution. The user equipment's task semantic knowledge base is divided into two levels: modality and reconstruction / goal-oriented. The base station's task semantic knowledge base is divided into three levels: modality, reconstruction / goal-oriented, and task-specific description. During downlink communication, the base station acts as the transmitter, performing semantic extraction and semantic encoding, while the user equipment acts as the receiver, performing semantic decoding and task execution. The base station's task semantic knowledge base is divided into three levels: modality, reconstruction / goal-oriented, and task-specific description. The user equipment's task semantic knowledge base is divided into two levels: modality and reconstruction / goal-oriented. Other communicating parties with similar relative hardware performance can also design their task semantic knowledge bases according to this hierarchy. When both communicating parties use a three-level task semantic knowledge base, each party indexes the corresponding three-level task nodes in the task semantic knowledge base for different tasks, and extracts the source semantic knowledge base supporting the task from the task nodes to complete the task. When the transmitter uses a two-level task semantic knowledge base and the receiver uses a three-level task semantic knowledge base, the transmitter indexes the corresponding two-level task nodes in the task semantic knowledge base according to the task requirements, and the receiver indexes the corresponding three-level task nodes, and extracts the source semantic knowledge base from the task nodes respectively to complete the task. In this scenario, tasks belonging to the same two-level node in the transmitter use the same source semantic knowledge base. Although this design increases the number of task nodes in the task semantic knowledge base of lower-performance devices, the method of sharing the same source semantic knowledge base for tasks belonging to the same two-level node greatly reduces storage overhead, because the storage overhead of the source semantic knowledge base is much higher than that of the task semantic knowledge base. This invention uses Neo4j to build the task semantic knowledge base. Neo4j is a high-performance graph database management system used to process highly correlated data. First, the Neo4j system is installed on the communication device, and the database service is started through commands or configuration files. Then, based on existing semantic communication tasks and their relationships, task nodes are created using the Cypher language supported by the database. Edges are created based on relationships such as inter-task modalities, and attribute descriptions are created for tasks and edges. Next, a syntax-based question parser is constructed, which parses the natural language descriptions of tasks into entities and relationships and converts them into Cypher query language. Finally, the Neo4j knowledge graph is queried using Cypher query language to obtain the task node index. When a task is not included in the task semantic knowledge base, task nodes are added using Cypher language, and syntax rules are added as needed to update the question parser, thereby updating the task semantic knowledge base.

[0033] Step 2: Design and implement the source semantic knowledge base (SKB) for specific tasks. S The source semantic knowledge base is mounted under the corresponding task node. It is divided into explicit and implicit types. The explicit source semantic knowledge base is mainly presented in the form of feature vector sets and triplet sets, while the implicit source semantic knowledge base is mainly presented in the form of network parameters of the semantic codec. Taking uplink natural language sentiment analysis, topic classification, and downlink image pixel-level restoration as examples, in the uplink task, the transmitter first needs to preprocess the natural language to obtain word-based natural language. Then, using a predefined dictionary and a pre-trained word embedding model such as GloVe, the processed natural language is converted into a two-dimensional matrix. Next, a Long Short-Term Memory (LSTM) network and a Convolutional Neural Network (CNN) are used to obtain the temporal correlation of the first-dimensional semantic data and compress it. Finally, the compressed semantic information is transmitted to the next module. After receiving the damaged compressed semantic information, the receiver first decompresses the data using a Transposed Convolutional Neural Network (TCNN), and then designs a Fully Connected Neural Network (FCNN) of the corresponding dimension according to the specific task to output the probabilities of different categories. For the sentiment analysis task, end-to-end training is performed, with the loss function being the cross-entropy loss of the transmit and receive class probabilities. After training, the transmitter's predefined dictionary, pre-trained word embedding model GloVe, and trained LSTM and CNN network parameters together serve as the transmitter's source semantic knowledge base for this task, where the LSTM and CNN network parameters are the implicit knowledge base, and the predefined dictionary and GloVe are the explicit knowledge base. The receiver's TCNN and FCNN network parameters serve as the receiver's implicit knowledge base for this task. For the downlink topic classification task, the transmitter's source semantic knowledge base for this task is the same as that of the sentiment analysis task transmitter. Therefore, during end-to-end training, the transmitter network parameters are fixed, and only the receiver network parameters are trained to obtain the receiver's implicit knowledge base TCNN and FCNN network parameters for downlink topic classification. In the downlink image pixel-level restoration task, the transmitter uses CNN to extract image semantics and achieve semantic compression, while the receiver uses TCNN to achieve semantic decompression and image restoration. The transmit and receive network parameters are obtained through end-to-end training, serving as the transmitter and receiver's implicit knowledge bases. There are various ways to construct a source semantic knowledge base. In specific implementation, the optimal paradigm can be selected based on task requirements: either the existing best solution can be migrated and deployed, a customized architecture design can be carried out, or functional expansion can be achieved by integrating a lightweight basic model.

[0034] Step 3: Design the Channel Semantic Knowledge Base (SKB)C This invention perceives and stores channel semantics, and performs semantic enhancement at the receiving end. Due to the complexity and diversity of wireless transmission environments, defining channel semantics uniformly is very difficult. Therefore, this invention designs a scheme where semantic training symbols are generated at the transmitting end, and changes in these semantic training symbols are perceived at the receiving end to perceive channel semantics and perform semantic enhancement. After processing in steps 1 and 2, the transmitter obtains semantic information in floating-point format with varying lengths. Using OFDM symbols or resource blocks as units, the semantic information is mapped to a small number of floating-point semantic training symbols through a randomly initialized FCNN network. Then, the index of this semantic training symbol is searched in a predefined codebook. Since the number of bits required for the index is small, the index is transmitted through the control channel, while the semantic codewords and semantic training symbols obtained from the semantic information processing are quantized before being transmitted through the service channel. At the receiving end, based on the received index, the original semantic training symbols are retrieved from the predefined codebook and sent together with the damaged semantic training symbols transmitted in the service channel into the FCNN network to perceive channel semantics. Then, the perceived channel semantics and the damaged semantic codewords transmitted in the service channel are input into the FCNN network to perform semantic enhancement, resulting in enhanced semantic information. Finally, this semantic information is input into the task executor or reconstructor in step 2 of the task to complete the communication task. Since embedding the complex form of the MIMO-OFDM wireless channel into the neural network results in gradient truncation, when training the FCNN network in the channel semantic knowledge base, the parameters of the transmitter's FCNN network are randomly initialized to the same order of magnitude as the parameters of the transmitter's implicit source semantic knowledge base. Only the receiver's FCNN network is optimized to obtain the parameters of the channel semantic perceptron and semantic enhancer, which are stored in the channel semantic knowledge base. Although the channel semantic knowledge base designed in this invention achieves effective perception of wireless channel semantics, its construction process is constrained by the diversity of source modalities and the specificity of task types. This technical characteristic indicates that the current methods for constructing channel semantic knowledge bases still require further research.

[0035] Figure 3 This diagram illustrates the link-level semantic knowledge base organization for semantic communication proposed in this invention. Taking the semantic communication between a mobile phone and a car as an example, assuming the mobile phone's hardware performance is inferior to that of the car, the semantic knowledge bases deployed in the mobile phone and car are A and B, respectively. The complete communication process after deploying the semantic knowledge base is as follows:

[0036] Step 1: The mobile phone and the car determine the communication task as "Task 1" (denoted as T) through a predetermined protocol and establish a directional communication link, taking the transmission of data from the mobile phone to the car as an example.

[0037] Step 2: The mobile phone and the car respectively retrieve the task semantic knowledge base SKB deployed therein. T The task 1 node was found.

[0038] Step 3: The mobile phone and the car respectively extract the source semantic knowledge base 1 mounted on the task 1 node, denoted as SKB. S ={SKB Sen SKB Sde}, of which SKB Sen SKB serves as the source semantic knowledge base within the transmitting end, i.e., the mobile phone. Sde For the source semantic knowledge base in the receiving end, i.e., the car, this process is denoted as SKB. S =f(T;SKB T The mobile phone uses a source semantic knowledge base as prior information to perform task-related semantic extraction, semantic encoding, and compression to obtain source semantic information Z. This process is denoted as Z = g. en (S;SKB Sen );

[0039] Step 4: The mobile phone extracts a randomly initialized training symbol generator from the channel semantic knowledge base according to the task mode and type, and generates semantic training symbols P. It then retrieves the training symbol index from the codebook and maps the source semantic information into semantic codewords suitable for channel transmission. The mobile phone uses a quantizer q... en (·) Quantize the floating-point semantic codewords and training symbols into bit-form semantic codewords X and training symbols P. This process is denoted as X, P = q. en (h en (Z;SKB Cde ));

[0040] Step 5: Semantic codewords and semantic training symbols are efficiently transmitted via the service channel using technologies such as MIMO-OFDM, while the index of the training symbols is accurately transmitted via the control channel.

[0041] Step 6: The car receives the damaged semantic codewords. Semantic training symbols And training symbol index. The car is retrieved from the channel semantic knowledge base SKB based on task modality and type. Cde The channel semantic perceptron and semantic enhancer are extracted. The damaged semantic training symbols and the original training symbols are input into the channel semantic perceptron to perceive the channel semantics. Then, the channel semantics and the dequantized damaged semantic codewords are input into the semantic enhancer to obtain the enhanced semantic information. The above process is denoted as

[0042] Step 7, the car will use the enhanced semantic information Input the source semantic knowledge base 1SKB selected in step 3. Sde To complete the semantic communication task, this process is denoted as...

[0043] The above description, in conjunction with the accompanying drawings, is merely one of the preferred embodiments of the present invention and should not be construed as limiting the scope of the claims. It should be noted that any equivalent modifications made in accordance with the principles of the present invention are covered within the scope of protection of the claims.

Claims

1. A knowledge base architecture and adaptive construction method for semantic communication, characterized in that, Includes the following steps: Step 1: Divide the semantic knowledge base into link-level and system-level. The link-level includes task, source, and channel semantic knowledge bases, while the system-level includes the scheduling semantic knowledge base. The scheduling semantic knowledge base is used for system-level resource scheduling, including base station location, user distribution, and system status management; the channel semantic knowledge base generates semantic training symbols at the transmitting end, and realizes channel semantic awareness and semantic enhancement at the receiving end by sensing changes in the training symbols; the semantic training symbols are transmitted through a predefined codebook index, the index is transmitted through the control channel, and the semantic codewords are transmitted through the service channel. Step 2: The communication system allocates spectrum and power resources to the two parties in the current link-level communication based on the prior information provided by the scheduling semantic knowledge base; Step 3: The transceiver determines the current semantic communication task according to the protocol, retrieves the task in the task semantic knowledge base based on the graph database, and extracts the best source semantic knowledge base if it exists; otherwise, it updates the task knowledge base and trains the source semantic knowledge base for the new task. Step 4: The transmitter uses the selected source semantic knowledge base as prior information to extract the task-related semantic information of the source and obtains the semantic information in floating-point form through the semantic encoder; Step 5: The transmitter uses the channel semantic knowledge base as prior information to process semantic information to obtain floating-point semantic codewords and training symbols suitable for the current wireless channel transmission, and quantizes them into bit-level form; Step 6: The transmitter converts the semantic codewords and training symbols into electromagnetic waves and transmits them to the wireless channel. The receiver obtains the damaged semantic codewords and training symbols through demodulation technology. Step 7: The receiver dequantizes the damaged semantic codewords and training symbols, and performs channel semantic perception and semantic enhancement in conjunction with the channel semantic knowledge base to obtain enhanced semantic information; Step 8: The receiver uses the semantic decoder to decode the enhanced semantic information, and uses the source semantic knowledge base as a priori to recover the original source or perform communication tasks.

2. The method according to claim 1, characterized in that, The task semantic knowledge base is stored in the form of a graph database. The task nodes are equipped with source semantic knowledge bases adapted to hardware performance. Devices with sufficient hardware performance adopt a three-level task division, while devices with limited performance adopt a two-level task division.

3. The method according to claim 1, characterized in that, The source semantic knowledge base is divided into explicit and implicit types. The explicit type includes a set of feature vectors or a set of triples, while the implicit type includes pre-trained semantic encoder-decoder parameters.

4. The method according to claim 1, characterized in that, The semantic knowledge base is updated by adding or removing task nodes, and the new task's dataset is used to update the source semantic knowledge base through offline or online training.

5. The method according to claim 1, characterized in that, The semantic enhancement is implemented in the form of a neural network, with the input being damaged semantic codewords, original training symbols, and damaged training symbols, and the output being enhanced semantic information.

Images