Design method of future perceptual cognitive semantic communication system based on large language model

Abstract

The present invention relates to the field of semantic communication technology, and in particular to a method for designing a future perception cognitive semantic communication system based on a large language model. The method comprises the following steps: obtaining a large language model and a source text set, dividing the source text set into a training set and a test set, preprocessing the training set and the test set respectively, and forming a future perception quadruple knowledge graph based on the large language model; designing a loss function to train a quadruple knowledge graph embedding model in combination with the future perception quadruple knowledge graph to obtain high-dimensional natural language triples; guiding the large language model to perform language verification and correction based on the high-dimensional natural language triples to obtain a future perception cognitive semantic communication system based on the large language model; verifying the effectiveness of the future perception cognitive semantic communication system based on the large language model through a test set; and achieving quantitative representation of the semantic stability of future scenarios by the knowledge graph through future perception reasoning to solve the instability problem of long-term transmission.

Description

Technical Field

[0001] The present invention relates to the field of semantic communication technology, and in particular to a design method for a future perceptual cognitive semantic communication system based on a large language model. Background Art

[0002] Advances in artificial intelligence and natural language processing have driven the development of semantic communication, a new communication paradigm that breaks through the transmission bottlenecks of traditional communication systems. By extracting and encoding semantic information from the source and requiring only the receiver to restore the semantic content, semantic communication reduces redundant data, reduces transmission volume, and improves efficiency. In recent years, semantic communication technology has achieved remarkable results. For example, in their 2021 paper, "Deeplearning enabled semantic communication systems" (IEEE Trans. on SignalProcessing, vol. 69, pp. 2663–2675, 2021), researchers Xie H, Qin Z, Li GY, et al. proposed a deep learning-based joint semantic channel coding scheme. This scheme, leveraging a neural network architecture, enables flexible processing of text information over complex channels. The semantic communication system significantly outperforms traditional communication methods in terms of word error rate. However, these end-to-end deep learning schemes rely on massive amounts of high-quality annotated data, and the black-box nature of the models leads to insufficient interpretability, making them difficult to meet the demands for transparent decision-making in industrial scenarios.

[0003] To address these shortcomings, Zhou F, Li YH, Yuan L and other scholars introduced knowledge graph technology in their published paper "Cognitive Semantic Communication Systems Driven by Knowledge Graph: Principle, Implementation, and Performance Evaluation" (IEEE Trans. on Communications, vol. 72, pp. 193–208, 2024). They achieved structured compression of semantic information through its triple structure (head entity, relationship, tail entity). The designed system successfully achieved significant data compression, and on this basis enhanced the interpretability of the semantic communication system and possessed error correction capabilities. However, there are still some technical bottlenecks. First, mainstream knowledge graph embedding models are mostly based on static graphs or dynamic modeling driven by historical data, which makes it difficult to capture the temporal evolution characteristics of entities and relationships, namely sudden events, trend changes, etc., resulting in insufficient long-term semantic transmission stability. Second, traditional cognitive semantic communication needs to repeatedly transmit relationship embedding vectors when transmitting large-scale triples, resulting in excessive redundant communication energy consumption, which restricts its application in massive data scenarios. Summary of the Invention

[0004] To address the technical issues that existing cognitive semantic communication systems have difficulty capturing the temporal evolution characteristics of entities and relationships, resulting in insufficient long-term semantic transmission stability, excessive redundant communication energy consumption, and unsuitability for large-scale data scenarios, the present invention aims to provide a design method for a future-perception cognitive semantic communication system based on a large language model. The technical solutions adopted are as follows:

[0005] Obtain a large language model and a source text collection, divide the source text collection into a training set and a test set, preprocess the training set and the test set respectively, and form a future-aware quadruple knowledge graph based on the large language model;

[0006] Construct a quadruple knowledge graph embedding model, design a loss function, and train the quadruple knowledge graph embedding model based on the future-perceived quadruple knowledge graph to obtain high-dimensional natural language triplets.

[0007] Based on high-dimensional natural language triples, the large language model is guided to perform language verification and correction, and a future perceptual cognitive semantic communication system based on the large language model is obtained;

[0008] The effectiveness of the future perceptual cognitive semantic communication system based on the large language model is verified through a test set.

[0009] Preferably, a large language model and a source text set are obtained, and the source text set is divided into a training set and a test set. The training set and the test set are preprocessed respectively, and a future-perceived quadruple knowledge graph is formed based on the large language model, including:

[0010] Structural semantic information is extracted based on the training set and the test set respectively, and a triple is generated for each source text, which is recorded as , build a triple knowledge graph, recorded as ,in, represents the triple knowledge graph, Represents the number of triples in the triple knowledge graph;

[0011] The time stability of each triple is evaluated through the large language model, and the stability score of each triple is determined. The corresponding quadruple is generated by combining the triples and recorded as , establish the quadruple knowledge graph of future perception, recorded as ,in, represents the quadruple knowledge graph, Represents the number of quadruple in the quadruple knowledge graph, corresponding to the triple knowledge graph.

[0012] Preferably, a quadruple knowledge graph embedding model is constructed, and a loss function is designed to train the quadruple knowledge graph embedding model in combination with the future-perceived quadruple knowledge graph to obtain high-dimensional natural language triples, including:

[0013] Design a loss function and obtain a set of quadruple embeddings based on the quadruple knowledge graph mapping;

[0014] Design a relation-aware classification transmission scheme, and determine the transmission quadruple embedding set through the quadruple embedding set;

[0015] Channel coding and transmission are performed based on the transmission quadruple embedding set, and decoding processing is performed. The decoded transmission quadruple embedding set is mapped and restored to obtain high-dimensional natural language triples.

[0016] Preferably, a loss function is designed to obtain a quadruple embedding set based on the quadruple knowledge graph mapping, including:

[0017] Design a loss function and divide the quadruple knowledge graph of future perception based on the loss function to determine positive and negative samples;

[0018] Initialize the quadruple knowledge graph, perform sampling, and filter positive samples;

[0019] Calculate the scores of positive and negative samples, optimize the quadruple knowledge graph and minimize the loss function, and perform iterative training until the quadruple knowledge graph embedding model converges or reaches the maximum number of iterations;

[0020] After training is completed, the quadruple knowledge graph is mapped to a low-dimensional vector space to obtain a quadruple embedding set.

[0021] Preferably, a loss function is designed, and the positive and negative samples are determined by dividing the quadruple knowledge graph of future perception based on the loss function, including:

[0022] Design the loss function, and the corresponding calculation formula is:

[0023]

[0024]

[0025]

[0026] in, represents the loss function, Represents a quadruple knowledge graph; represents a set of quadruple embeddings; 、 Represent the score functions of positive samples and negative samples respectively; represents the boundary hyperparameter; Represents function operation, which is used to ensure that the loss function is non-negative; 、 、 、 Respectively represent the head entity, relationship, tail entity and stability score corresponding to any quadruple in the quadruple knowledge graph; Indicates the operation of taking the real part of a complex number;

[0027] Positive samples represent unprocessed quadruples; negative samples represent quadruples generated by randomly replacing the head entity or tail entity in the positive sample.

[0028] Preferably, a relation-aware classification transmission scheme is designed, and the transmission quadruple embedding set is determined by the quadruple embedding set, including:

[0029] Obtain the corresponding triple knowledge graph based on the quadruple knowledge graph, and extract the directed adjacency matrix from the triple knowledge graph;

[0030] Embed the quadruple into a set, denoted as , reconstructed into an initial ordered quadruple embedding set classified by relationship, recorded as ,in, To express a relationship; represents the total number of relations in the initial ordered quadruple embedding set;

[0031] The preceding data samples of each data sample in the initial ordered quadruple embedding set are deleted, the currently analyzed data sample is reshaped and reconstructed to obtain a new data sample, and all new data samples are integrated to determine the final transmission quadruple embedding set.

[0032] Preferably, a directed adjacency matrix is ​​extracted from the triple knowledge graph, and the corresponding calculation formula is:

[0033]

[0034] in, represents a directed adjacency matrix, , Represents the number of all entities in the triple knowledge graph; 、 They all represent the corresponding entities in the triple knowledge graph.

[0035] Preferably, channel coding and transmission are performed based on the transmission quadruple embedding set, and decoding processing is performed, and mapping and recovery are performed on the decoded transmission quadruple embedding set to obtain a high-dimensional natural language triple, including:

[0036] Based on the transmission quadruple embedding set, low-density parity check code is used for channel coding. The corresponding calculation formula is:

[0037]

[0038] in, represents the output of channel coding; represents the channel encoder operation; represents the transmission quad-tuple embedding set;

[0039] The output after channel coding is combined with the channel environment for transmission. The channel environment includes additive white Gaussian noise and fading channel to determine the channel output. The corresponding calculation formula is:

[0040]

[0041] in, Indicates channel output; represents the channel gain of the fading channel; represents complex Gaussian noise, represents the noise power;

[0042] According to the channel output, the low-density parity check code is used for decoding. The corresponding calculation formula is:

[0043]

[0044] in, Indicates decoded output; represents a channel decoder;

[0045] The decoded output is restored through the nonlinear mapping of a multi-layer perceptron to obtain a high-dimensional natural language triple.

[0046] Preferably, the decoded output is restored through nonlinear mapping of a multi-layer perceptron to obtain a high-dimensional natural language triple. The corresponding calculation formula is:

[0047]

[0048] in, Represents high-dimensional natural language triples; Represents the multilayer perceptron operation; Indicates decoded output; Represents function operation; 、 Both represent weight matrices; 、 Both represent bias terms; Represents the activation function.

[0049] To solve the above problems, the present application also provides: an electronic device, characterized in that it includes: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus, and the processor calls the logical instructions in the memory to execute any one of the aforementioned methods for designing a future perceptual cognitive semantic communication system based on a large language model.

[0050] The present invention has the following beneficial effects:

[0051] 1. The future-aware cognitive semantic communication system based on a large language model has achieved a breakthrough in dynamic modeling capabilities. Through the future-aware reasoning of the large language model, it achieves a quantitative representation of the semantic stability of future scenarios in the knowledge graph, solving the instability problem of long-term transmission of traditional static models. The system trains a four-tuple knowledge graph embedding model and reduces the total communication energy consumption during large-scale triple transmission by approximately 48% through relationship-aware classification transmission, indicating that it is suitable for IoT scenarios with massive repetitive relationships. The system then guides the large language model to perform language verification and correction based on high-dimensional natural language triples. That is, through the future-aware reasoning and triple correction mechanism driven by the large language model, the transmission accuracy in low signal-to-noise ratio scenarios is improved by 45% compared to traditional methods, breaking through the constraints of channel noise on semantic recovery. That is, focusing on the three dimensions of future-aware feature modeling, redundant transmission optimization, and semantic error correction enhancement, the system achieves a coordinated improvement in semantic transmission stability, energy efficiency, and accuracy in complex scenarios.

[0052] 2. The electronic device provided by the present invention has the same beneficial effects as the future perception cognitive semantic communication system design method based on a large language model provided by the present invention, and will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0054] Figure 1 A flowchart of a method for designing a future-aware cognitive semantic communication system based on a large language model provided by one embodiment of the present invention;

[0055] Figure 2 A schematic diagram of the overall framework of a method for designing a future perception cognitive semantic communication system based on a large language model provided by one embodiment of the present invention;

[0056] Figure 3 A comparison chart of the total energy consumption of computing and communication when transmitting different numbers of triples between a design method for a future perceptual cognitive semantic communication system based on a large language model provided by one embodiment of the present invention and a traditional method;

[0057] Figure 4 A comparison chart of transmission accuracy under different signal-to-noise ratios between a design method for a future perception-cognitive semantic communication system based on a large language model provided by one embodiment of the present invention and a traditional method;

[0058] Figure 5 A comparison chart of the transmission accuracy of a future perceptual cognitive semantic communication system design method based on a large language model provided by one embodiment of the present invention and traditional methods for different relationship categories. DETAILED DESCRIPTION

[0059] To further illustrate the technical means and effects employed by the present invention to achieve the intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the specific implementation, structure, features, and effects of a design method for a future perceptual cognitive semantic communication system based on a large language model proposed by the present invention. In the following description, different references to "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0060] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0061] The following describes in detail a specific solution of a method for designing a future perceptual cognitive semantic communication system based on a large language model provided by the present invention in conjunction with the accompanying drawings.

[0062] Please combine Figure 1 and Figure 2 , which respectively show a step flow chart and an overall framework schematic diagram of a method for designing a future-aware cognitive semantic communication system based on a large language model provided by the first embodiment of the present invention, the method comprising:

[0063] Step S1: Obtain a large language model and a source text set, divide the source text set into a training set and a test set, preprocess the training set and the test set respectively, and form a future-perceived quadruple knowledge graph based on the large language model;

[0064] Step S2: Construct a quadruple knowledge graph embedding model, design a loss function and train the quadruple knowledge graph embedding model based on the future-perceived quadruple knowledge graph to obtain high-dimensional natural language triples;

[0065] Step S3: guiding the large language model to perform language verification and correction based on the high-dimensional natural language triples, thereby obtaining a future perceptual cognitive semantic communication system based on the large language model;

[0066] Step S4: Verify the effectiveness of the future perceptual cognitive semantic communication system based on the large language model through the test set.

[0067] To better explain, future perception and cognition refers to humans using various sensory organs and intelligent technologies to predict and understand possible future changes and development trends in advance; semantic communication is a way of information transmission, which is to extract and encode semantic information in the source, and only requires the receiving end to restore the consistency of the semantic content, rather than precise bit-by-bit matching; future perception and cognition semantic communication means achieving efficient communication at the semantic level through future perception and cognition, so as to break through the bottlenecks of existing semantic communication systems in dynamic semantic modeling capabilities and communication energy efficiency optimization.

[0068] As an optional implementation method, the large language model can be flexibly deployed according to the specific experimental scenario to adapt to different research needs and environmental conditions; preferably, in this embodiment, the large language model is LlaMa-3.1-8B, which has powerful computing power and rich corpus support, can efficiently process complex natural language tasks, and provide accurate language understanding and generation capabilities; and the future perceptual cognitive semantic communication system based on the large language model includes a transmitting end and a receiving end, which are used to work together to improve communication energy efficiency.

[0069] Furthermore, step S1 includes:

[0070] Step S11: Extract structured semantic information based on the training set and the test set respectively, and generate a triple for each source text, which is recorded as , build a triple knowledge graph, recorded as ,in, represents the triple knowledge graph, Represents the number of triples in the triple knowledge graph.

[0071] Specifically, natural language processing technology is used to extract structured semantic information from the source text input from the sender, and any source text is converted into triples. The format of the triple is (head entity, relationship, tail entity). All triples are integrated to establish a triple knowledge graph to realize the conversion of the source text into a semantic structure. It can be explained that the source text refers to the sentence sent by the sender. For example, the source text is "Albert Einstein was born in Germany in 1879, lived in the United States for most of his life, and died in the United States in 1955." The obtained triples are (Albert Einstein, born in Germany), (Albert Einstein, lived in the United States), and (Albert Einstein, died in the United States).

[0072] Step S12: Use the large language model to evaluate the temporal stability of each triple, determine the stability score of each triple, and generate a quadruple corresponding to the triple and combined, which is recorded as , establish the quadruple knowledge graph of future perception, recorded as ,in, represents the quadruple knowledge graph, Represents the number of quadruple in the quadruple knowledge graph, corresponding to the triple knowledge graph.

[0073] Specifically, the zero-sample reasoning capability of the large language model is introduced to quantify the stability of triples in future scenarios. The large language model is used to evaluate the temporal stability of each triple and determine the stability score of each triple. That is, the large language model generates a stability score for each triple, which represents the effective probability of the triple in the future scenario. The stability score is recorded as , , 1 means completely stable, that is, the fact will not change in the future; 0 means inevitable change, that is, the fact will change in the future; corresponding to and combined with the triples to generate quadruple, the format of the quadruple is (head entity, relationship, tail entity, stability score), all quadruple are integrated to form the quadruple knowledge graph of future perception, providing a temporal feature basis for subsequent dynamic modeling; it can be explained that, taking the aforementioned triple (Albert Einstein, born, Germany) as an example, the determined quadruple is (Albert Einstein, born, Germany, 0.9)

[0074] Furthermore, step S2 includes:

[0075] Step S21: Design a loss function and obtain a quadruple embedding set based on the quadruple knowledge graph mapping.

[0076] For better explanation, in this embodiment, the four-tuple knowledge graph embedding model is a future-aware four-tuple knowledge graph embedding model, which enhances the perception of future events and dynamic changes based on the traditional four-tuple knowledge graph embedding model.

[0077] It can be understood that a new loss function that integrates the stability score is designed to train the quadruple knowledge graph embedding model through the loss function, mapping the natural language quadruple in the high-dimensional space, that is, the quadruple knowledge graph of future perception, to the low-dimensional embedding space to generate entity embedding containing future perception features. and relation embedding , to improve the stability of long-term semantic transmission.

[0078] Furthermore, step S21 includes:

[0079] Step S211: Design a loss function, and divide the future perceived quadruple knowledge graph based on the loss function to determine positive samples and negative samples; that is, learn entity and relationship embedding by minimizing the loss function.

[0080] Furthermore, step S211 includes:

[0081] Design the loss function, and the corresponding calculation formula is:

[0082]

[0083]

[0084]

[0085] in, represents the loss function, Represents a quadruple knowledge graph; represents a set of quadruple embeddings; 、 Represent the score functions of positive samples and negative samples respectively; represents the boundary hyperparameter; Represents function operation, which is used to ensure that the loss function is non-negative; 、 、 、 Respectively represent the head entity, relationship, tail entity and stability score corresponding to any quadruple in the quadruple knowledge graph; Indicates the operation of taking the real part of a complex number;

[0086] Positive samples represent unprocessed quadruples; negative samples represent quadruples generated by randomly replacing the head entity or tail entity in the positive sample.

[0087] Make an explanation, Represents the set of quadruple embeddings, i.e., the embedding vectors corresponding to the head entity, relation, and tail entity; the loss function contains boundary constraints to integrate the stability score; the positive sample is a quadruple without any processing, i.e., (head entity, relation, tail entity, stability score); the negative sample represents the sample generated by randomly replacing the head entity or tail entity, i.e., randomly extracting a head entity or tail entity from the entire quadruple knowledge graph to replace the head entity or tail entity in the positive sample, and the score function is , recorded as damaged quadruple sample; represents the boundary hyperparameter that controls the separation between positive and negative sample scores; Used to ensure that the loss function is non-negative and only contributes when the bound constraints are violated.

[0088] Step S212: Initialize the four-tuple knowledge graph, perform sampling, and filter positive samples; that is, randomly initialize the complex-valued embedding vectors of all entities and relationships in the four-tuple knowledge graph, which represents the position of the sample data in the high-dimensional space, and filter positive samples according to the aforementioned step S211, and generate negative samples at the same time.

[0089] Step S213: Calculate the scores of positive samples and negative samples, optimize the four-tuple knowledge graph and minimize the loss function, and perform iterative training until the four-tuple knowledge graph embedding model converges or reaches the maximum number of iterations.

[0090] To explain, according to the scoring function in step S211 and Determine the scores of positive and negative samples respectively, and minimize the loss function by optimizing the complex-valued embedding vector through gradient descent , and repeat the process until the model converges or reaches the maximum number of iterations, where the maximum number of iterations is usually set to 1000; that is, by calculating the gradient of the loss function, the direction and size of each complex-valued embedding vector are determined so that it moves in the direction of reducing the loss in each iteration, ensuring that the change amplitude is moderate when updating the embedding vector at each step to avoid excessive jumps that lead to unstable convergence. As the iteration proceeds, the loss function gradually decreases and eventually reaches a local or global minimum, achieving effective optimization of the complex-valued embedding vector.

[0091] Step S214: After the training is completed, the quadruple knowledge graph is mapped to a low-dimensional vector space to obtain a quadruple embedding set.

[0092] It can be explained that after the training is completed, the quadruple knowledge graph embedding model maps the quadruple knowledge graph in the high-dimensional natural language space to the low-dimensional vector space to obtain the quadruple embedding set, which is recorded as ,in, , Indicates the The low-dimensional space embedding representation corresponding to the quadruple is 、 、 Represent the head entity, relationship and tail entity respectively, represents the embedding dimension. Preferably, in this embodiment, the embedding dimension ; It effectively captures the semantic relations and stability relations in the triple knowledge graph and realizes the joint representation of semantic features and stability scores.

[0093] Step S22: Design a relationship-aware classification transmission scheme, and determine the transmission quadruple embedding set through the quadruple embedding set; that is, group the triples according to the relationship category, and only transmit one shared relationship embedding and entity embedding for each group, omitting the redundant transmission of repeated relationships in the same category. By reducing the repeated transmission of relationship vectors, the energy consumption in the triple transmission scenario is reduced, and the communication energy efficiency is improved.

[0094] Furthermore, step S22 includes:

[0095] Step S221: Obtain the corresponding triple knowledge graph based on the quadruple knowledge graph, and extract the directed adjacency matrix from the triple knowledge graph.

[0096] Furthermore, in step S221, a directed adjacency matrix is ​​extracted from the triple knowledge graph, and the corresponding calculation formula is:

[0097]

[0098] in, represents a directed adjacency matrix, , Represents the number of all entities in the triple knowledge graph; 、 They all represent the corresponding entities in the triple knowledge graph.

[0099] Explanation: According to the quadruple knowledge graph, the corresponding triple knowledge graph is obtained, that is, based on the aforementioned triple knowledge graph , extract the directed adjacency matrix , to identify different entities, the value of each element in the corresponding matrix represents the relationship from the row entity to the column entity, which is used to represent the topological relationship between entities. Since all entities can be used as head entities or tail entities, the matrix size is .

[0100] Step S222: embed the quadruple into a set, denoted as , reconstructed into an initial ordered quadruple embedding set classified by relationship, recorded as ,in, To express a relationship; Represents the total number of relations in the initial ordered quadruple embedding set; that is, through the directed adjacency matrix , embed the unordered quadruple into the set Reconstructed into an initial ordered quadruple embedding set classified by relationship , where unordered represents the set of quadruple embeddings before classification according to relation categories.

[0101] Step S223: Delete the previous several data samples of each data sample in the initial ordered quadruple embedding set, reshape and reconstruct the currently analyzed data sample to obtain a new data sample, and integrate all the new data samples to determine the final transmission quadruple embedding set.

[0102] It can be shown that, assuming the relationship class Contains several The quadruple embedding vector is The corresponding logical expression is:

[0103]

[0104] Based on the The set of quadruple embeddings of relation categories , move forward The relation embedding in the embedding vector is 、 、 、 Delete and re- The new data sample is obtained by reshaping and reconstruction, that is, the four-tuple embedding set, which is recorded as , then for all After the same processing is performed on the four-tuple embedding sets, the final transmission four-tuple embedding set is determined, which is recorded as .

[0105] It can be understood that the semantic encoding part consisting of the construction of the sent triple knowledge graph, the generation of the future-aware quadruple knowledge graph, the training of the future-aware quadruple knowledge graph embedding model, and the relationship-aware classification transmission has been fully realized. At this point, the overall logical correspondence is expressed as:

[0106]

[0107] in, represents the transmission quad-tuple embedding set; Indicates that the parameter set is The future-aware quadruple knowledge graph embedding model of , that is, the parameter set is set to 0 at the beginning of the large language model training. After each iteration of gradient propagation, the parameter set changes accordingly until the loss function reaches convergence. The optimal parameter set corresponding to the large language model is ; Represents the original input, that is, the parameters of the initial input large model parameters.

[0108] Step S23: Channel coding and transmission are performed based on the transmission quadruple embedding set, and decoding is performed, and mapping and recovery are performed on the decoded transmission quadruple embedding set to obtain high-dimensional natural language triples.

[0109] Furthermore, step S23 includes:

[0110] Step S231: low-density parity check code is used for channel coding based on the transmission quadruple embedding set. The corresponding calculation formula is:

[0111]

[0112] in, represents the output of channel coding; represents the channel encoder operation; Represents the set of transport quadruple embeddings.

[0113] It is explained that the traditional coding method of low-density parity check code is used based on the transmission quadruple embedded set to perform channel coding, which effectively improves the reliability of data transmission; wherein, It represents the output of channel coding, that is, the symbol sequence suitable for transmission over the physical channel.

[0114] Step S232: The output after channel coding is combined with the channel environment for transmission. The channel environment includes additive white Gaussian noise and fading channel. The channel output is determined. The corresponding calculation formula is:

[0115]

[0116] in, Indicates channel output; represents the channel gain of the fading channel; represents complex Gaussian noise, Represents the noise power.

[0117] It is explained that during the transmission process, the signal will be subject to various interferences, such as random fluctuations of electromagnetic waves and multipath effects, which will lead to changes in signal strength and phase, and convert the signal into a signal with superimposed noise and distortion. That is, the actual channel environment is simulated by additive white Gaussian noise and fading channels to ensure the authenticity of the data and obtain the channel output.

[0118] Step S233: Decoding is performed using a low-density parity check code according to the channel output. The corresponding calculation formula is:

[0119]

[0120] in, Indicates decoded output; Represents a channel decoder.

[0121] It is explained that decoding processing is performed through low-density parity-check codes to effectively identify and correct bit errors during transmission, thereby improving the reliability and efficiency of data transmission.

[0122] Step S234: The decoded output is restored through the nonlinear mapping of the multilayer perceptron to obtain a high-dimensional natural language triple; that is, the received decoded output is mapped to the high-dimensional natural language space using the multilayer perceptron to generate a preliminary received triple; wherein the multilayer perceptron includes multiple layers of fully connected layers to capture complex features and patterns through each fully connected layer, and to extract information to lay the foundation for subsequent analysis.

[0123] Furthermore, in step S234, the corresponding calculation formula is:

[0124]

[0125] in, Represents high-dimensional natural language triples; Represents the multilayer perceptron operation; Indicates decoded output; Represents function operation; 、 Both represent weight matrices; 、 Both represent bias terms; Represents the activation function.

[0126] Specifically, the decoded output , which is the embedding vector After the first fully connected layer , to achieve dimensionality improvement, among which, , , represents the embedding dimension; Represents the fully connected layer; then the activation function is used Enhance the nonlinear expression capability of the output of the first fully connected layer. Preferably, in this embodiment, the activation function The ReLU function is used; then, the dimension is restored to the triple space dimension through the second fully connected layer, where , , represents the total number of relations in the initial ordered quadruple embedding set; and through The function gives the probability distribution of the output of the second fully connected layer, and obtains the high-dimensional natural language triples recovered at the receiving end , which means the received high-dimensional natural language triples affected by channel impairment.

[0127] It is understandable that although the nonlinear mapping recovery of the decoded output realizes the mapping from low-dimensional semantic features to high-dimensional natural language structure, it may cause errors such as entity-relationship misalignment and logical contradictions due to the influence of channel noise, so it needs to be further corrected through the triple correction module enabled by the large language model.

[0128] It is explained that in step S3, the large language model is guided to perform language verification and correction based on high-dimensional natural language triples, and a future perceptual cognitive semantic communication system based on the large language model is obtained; specifically, a two-stage prompt template is designed to guide the large language model to perform semantics and correction, that is, prompt words are used to guide the large language model to use internal knowledge bases, such as Wikipedia facts, common sense graphs, etc. to identify errors, thereby enhancing its understanding and processing capabilities of complex contexts; then prompt words are used to guide the large language model to format the output according to the reasoning results, so as to prevent the large language model from outputting unstructured text and improve the adaptability of downstream tasks.

[0129] It can be understood that the relationship represents all relationship categories in the knowledge base, and the task is to understand the relationship between all elements in the high-dimensional natural language triples, and then determine whether the entities and relationships in any triples are correctly matched. If an error occurs, it is corrected; for example, the recovered high-dimensional natural language triple is (Albert Einstein, located in, Germany), the relationship is determined to be {born, located}, and the triple output after verification and correction is (Albert Einstein, born, Germany).

[0130] It is explained that in step S4, the effectiveness of the future perception cognitive semantic communication system based on the large language model is verified through a test set; that is, the total communication, energy consumption and transmission accuracy are calculated through the test set, and then the effectiveness of the overall constructed future perception cognitive semantic communication system based on the large language model is verified to improve its applicability.

[0131] Understandably, the future perception cognitive semantic communication system based on the large language model has achieved a breakthrough in dynamic modeling capabilities. Through the future perception reasoning of the large language model, the knowledge graph can be used to quantitatively represent the semantic stability of future scenarios, solving the instability problem of long-term transmission of traditional static models. The four-tuple knowledge graph embedding model is trained, and through relationship-aware classification transmission, the total communication energy consumption during large-scale triple transmission is reduced by about 48%, indicating that it is suitable for IoT scenarios with massive repeated relationships. Then, based on high-dimensional natural language triples, the large language model is guided to perform language verification and correction, that is, through the future perception reasoning and triple correction mechanism driven by the large language model, the transmission accuracy in low signal-to-noise ratio scenarios is improved by 45% compared with traditional methods, breaking through the constraints of channel noise on semantic recovery. That is, around the three dimensions of future perception feature modeling, redundant transmission optimization and semantic error correction enhancement, the coordinated improvement of semantic transmission stability, energy efficiency and accuracy in complex scenarios is achieved.

[0132] For better explanation, in order to determine the feasibility of the future perceptual cognitive semantic communication system based on the large language model proposed in this application, a simulation experiment is conducted. Specifically, in this embodiment, the experiment is conducted based on Python 3.12 and Pytorch 2.2.0 simulation platform, where the embedding dimension , the learning rate is set to 0.0004, the maximum number of iterations is 100, the margin is 1, and the experimental dataset uses YAGO3-10, which covers rich semantic information, including 123182 entities and 37 relations. It is divided and the number of triplets in the training set and test set is determined to be 1079040 and 5000 respectively.

[0133] Specifically, the simulation experiment was compared with the traditional method. The first one is the traditional triple method, which uses the traditional ComplEx model in the encoder and directly transmits all triples during transmission without compression based on relationship categories. At the same time, the receiving end does not correct the triples; the second one is a method that only uses classification transmission, which only uses the future-aware four-tuple knowledge graph embedding model and the relationship-aware classification transmission mechanism at the sending end, that is, it runs according to step S2, and the receiving end does not correct the triples.

[0134] See also Figure 3, which shows a comparison of the total energy consumption of computing and communication when transmitting different numbers of triplets between a future-aware cognitive semantic communication system design method based on a large language model provided by the first embodiment of the present invention and a traditional method; it is explained that as the number of triples increases, the energy consumption of all schemes shows an upward trend. Among them, due to the relationship-aware classification transmission scheme, which omits redundant relationship embedding, the energy consumption of the future-aware cognitive semantic communication system based on a large language model proposed in this application and the method using only classification transmission are almost consistent, so there is an overlap in the linear presentation; and the growth rate is slower than that of the traditional triple method, so that the communication energy consumption decreases with the proportion of the number of triples, and the computational overhead is constrained by the limited relationship categories. It can be shown that as the number of triples increases, the differences between the three methods become obvious. When the number of triples reaches 1000, the energy consumption of the future-aware cognitive semantic communication system based on a large language model is reduced by 48% compared with the other two groups of methods, which illustrates the effectiveness of the relationship-aware classification transmission scheme in large-scale communication scenarios with repeated relationships.

[0135] See also Figure 4 , which shows a comparison of the transmission accuracy of a future-perceptual cognitive semantic communication system design method based on a large language model, provided by the first embodiment of the present invention, and traditional methods at different signal-to-noise ratios. It illustrates that as the signal-to-noise ratio increases, the accuracy of the classification-only transmission method and the traditional triplet method improves. However, the future-perceptual cognitive semantic communication system based on the large language model consistently outperforms the other two methods. At -6dB, the transmission accuracy improves by 45% compared to the traditional triplet method, demonstrating that the proposed system can achieve more accurate semantic encoding and recovery under dynamic signal-to-noise ratio conditions. It can be shown that due to the inherent hallucination problem of the large language model, namely, incorrectly correcting triples while maintaining factual consistency (for example, (Albert Einstein, lived in the United States) is corrected to (Albert Einstein, died in the United States), the transmission accuracy of the future-perceptual cognitive semantic communication system based on the large language model does not strictly increase with increasing signal-to-noise ratio. For example, at -3dB, its accuracy is lower than at -6dB. Furthermore, the classification-only transmission method exhibits a significant signal-to-noise ratio threshold at approximately -6dB. Because its classification-based transmission scheme lacks the operation of correcting the triplet after transmission, the adverse effects of channel noise become more significant under poor channel conditions. Therefore, when the signal-to-noise ratio is lower than this threshold, the accuracy of the classification-based transmission method alone is 12% to 18% lower than that of the traditional triplet method, indicating the necessity of correcting the recovered triplet for large-scale reliable transmission.

[0136] See also Figure 5, which shows a comparison chart of the transmission accuracy of a future perception cognitive semantic communication system design method based on a large language model provided by the first embodiment of the present invention and the traditional method for different relationship categories; it explains that in actual application, 30 triples are transmitted for each relationship category, where the stability score represents the average stability score of the 30 triples; the future perception cognitive semantic communication system based on the large language model performs better than the other two methods in each relationship category, and compared with the triples with low stability, the classification transmission method alone shows a more significant accuracy improvement on the triples with high stability relationships compared with the traditional triple method, which further verifies the ability of the future perception cognitive semantic communication system based on the large language model proposed in this application to prioritize stable semantic patterns.

[0137] In summary, this application introduces a future-aware four-tuple knowledge graph embedding set, a relationship-aware classification transmission mechanism, and correction of the restored triples to form a future-aware cognitive semantic communication system based on a large language model. It can significantly reduce communication energy consumption when transmitting large amounts of repetitive text under low signal-to-noise ratio conditions, and at the same time exhibit excellent performance in terms of transmission accuracy.

[0138] The second embodiment of the present invention proposes an electronic device, including: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus, and the processor calls the logic instructions in the memory to execute the future perception cognitive semantic communication system design method based on the large language model provided by any one of the aforementioned embodiments; this system has the same beneficial effects as the aforementioned future perception cognitive semantic communication system design method based on the large language model provided, and will not be repeated here.

[0139] It can be understood that when a module of an electronic device is in operation, it is necessary to utilize the future perception cognitive semantic communication system design method based on a large language model provided by the aforementioned embodiment. Therefore, whether the method is integrated with program data or different hardware is configured to produce functions similar to the effects achieved by the present invention, it falls within the scope of protection of the present invention.

[0140] It should be noted that the order in which the embodiments of the present invention are described above is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0141] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

Claims

1. A design method for a future perceptual cognitive semantic communication system based on a large language model, characterized by: The method comprises: Obtain a large language model and a source text collection, divide the source text collection into a training set and a test set, preprocess the training set and the test set respectively, and form a future-aware quadruple knowledge graph based on the large language model; Construct a quadruple knowledge graph embedding model, design a loss function, and train the quadruple knowledge graph embedding model based on the future-perceived quadruple knowledge graph to obtain high-dimensional natural language triplets, including: Design a loss function and obtain a quadruple embedding set based on the quadruple knowledge graph mapping, including: Design a loss function and divide the quadruple knowledge graph of future perception based on the loss function to determine positive and negative samples, including: Design the loss function, and the corresponding calculation formula is: ； ； ； in, represents the loss function, Represents a quadruple knowledge graph; represents a set of quadruple embeddings; 、 Represent the score functions of positive samples and negative samples respectively; represents the boundary hyperparameter; Represents function operation, which is used to ensure that the loss function is non-negative; 、 、 、 Respectively represent the head entity, relationship, tail entity and stability score corresponding to any quadruple in the quadruple knowledge graph; Indicates the operation of taking the real part of a complex number; Positive samples represent unprocessed quads; negative samples represent quads generated by randomly replacing the head or tail entity in the positive sample. Initialize the quadruple knowledge graph, perform sampling, and filter positive samples; Calculate the scores of positive and negative samples, optimize the quadruple knowledge graph and minimize the loss function, and perform iterative training until the quadruple knowledge graph embedding model converges or reaches the maximum number of iterations; After training is completed, the quadruple knowledge graph is mapped to a low-dimensional vector space to obtain a quadruple embedding set; Design a relation-aware classification transmission scheme, and determine the transmission quadruple embedding set through the quadruple embedding set; Channel coding and transmission are performed based on the transmission quadruple embedding set, and decoding is performed, and mapping and recovery are performed on the decoded transmission quadruple embedding set to obtain high-dimensional natural language triples; Based on high-dimensional natural language triples, the large language model is guided to perform language verification and correction, and a future perceptual cognitive semantic communication system based on the large language model is obtained; The effectiveness of the future perceptual cognitive semantic communication system based on the large language model is verified through a test set.

2. The method for designing a future-aware cognitive semantic communication system based on a large language model according to claim 1, characterized in that: Obtain a large language model and a source text collection, divide the source text collection into a training set and a test set, preprocess the training set and the test set respectively, and form a future-aware quadruple knowledge graph based on the large language model, including: Structural semantic information is extracted based on the training set and the test set respectively, and a triple is generated for each source text, which is recorded as , build a triple knowledge graph, recorded as ,in, represents the triple knowledge graph, Represents the number of triples in the triple knowledge graph; The time stability of each triple is evaluated through the large language model, and the stability score of each triple is determined. The corresponding quadruple is generated by combining the triples and recorded as , establish the quadruple knowledge graph of future perception, recorded as ,in, represents the quadruple knowledge graph, Represents the number of quadruple in the quadruple knowledge graph, corresponding to the triple knowledge graph.

3. The method for designing a future-aware cognitive semantic communication system based on a large language model according to claim 1, characterized in that: Design a relation-aware classification transmission scheme, and determine the transmission quadruple embedding set through the quadruple embedding set, including: Obtain the corresponding triple knowledge graph based on the quadruple knowledge graph, and extract the directed adjacency matrix from the triple knowledge graph; Embed the quadruple into a set, denoted as , reconstructed into an initial ordered quadruple embedding set classified by relationship, recorded as ,in, To express a relationship; represents the total number of relations in the initial ordered quadruple embedding set; The preceding data samples of each data sample in the initial ordered quadruple embedding set are deleted, the currently analyzed data sample is reshaped and reconstructed to obtain a new data sample, and all new data samples are integrated to determine the final transmission quadruple embedding set.

4. The method for designing a future-aware cognitive semantic communication system based on a large language model according to claim 3, characterized in that: Extracting the directed adjacency matrix from the triple knowledge graph, the corresponding calculation formula is: ； in, represents a directed adjacency matrix, , Represents the number of all entities in the triple knowledge graph; 、 They all represent the corresponding entities in the triple knowledge graph.

5. The method for designing a future-aware cognitive semantic communication system based on a large language model according to claim 1, characterized in that: Channel coding and transmission are performed based on the transmission quadruple embedding set, and decoding is performed. The decoded transmission quadruple embedding set is mapped and restored to obtain a high-dimensional natural language triple, including: Based on the transmission quadruple embedding set, low-density parity check code is used for channel coding. The corresponding calculation formula is: ； in, represents the output of channel coding; represents the channel encoder operation; represents the transmission quad-tuple embedding set; The output after channel coding is combined with the channel environment for transmission. The channel environment includes additive white Gaussian noise and fading channel to determine the channel output. The corresponding calculation formula is: ； in, Indicates channel output; represents the channel gain of the fading channel; represents complex Gaussian noise, represents the noise power; According to the channel output, the low-density parity check code is used for decoding. The corresponding calculation formula is: ； in, Indicates decoded output; represents a channel decoder; The decoded output is restored through the nonlinear mapping of a multi-layer perceptron to obtain a high-dimensional natural language triple.

6. The method for designing a future-aware cognitive semantic communication system based on a large language model according to claim 5, characterized in that: The decoded output is restored through the nonlinear mapping of the multi-layer perceptron to obtain a high-dimensional natural language triple. The corresponding calculation formula is: ； in, Represents high-dimensional natural language triples; Represents the multilayer perceptron operation; Indicates decoded output; Represents function operation; 、 Both represent weight matrices; 、 Both represent bias terms; Represents the activation function.

7. An electronic device, characterized in that: include: A processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus, and the processor calls the logic instructions in the memory to execute the future perception cognitive semantic communication system design method based on the large language model described in any one of claims 1-6.

Images