A one-to-many multi-user semantic communication model and communication method

Abstract

The present invention discloses a one-to-many multi-user semantic communication model and a communication method. The model is integrated by a sending end and a plurality of receiving ends that establish communication relationships with the sending end and are independent of each other. The method comprises the following steps: collecting text sentences of various types according to preset user needs; combining text sentences of various types into text sequences and converting them into digital ID sequences as sending information of the sending end; generating a communication signal for channel transmission at the sending end and sending the signal to each receiving end; each receiving end performs channel decoding and semantic decoding on the received communication signal to restore the original sentence sent by the sending end; and inputting the signal into a semantic recognizer based on a distilled bidirectional language representation pre-training model to output corresponding sentences according to user needs. Through the system model and communication method of the present invention, the transmission procedure of multi-user communication is simplified and the efficiency of information transmission is improved; and the receiving end is trained in combination with a transfer learning method to improve the training efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a one-to-many multi-user semantic communication model and a communication method. Background Art

[0002] With the rapid development of artificial intelligence and natural language processing technologies, semantic communication is considered a promising technical solution to address the transmission bottlenecks of traditional communication systems. Semantic communication refers to a communication method that extracts and encodes semantic information from a source signal, then transmits it over a noisy channel. Traditional communication requires that the decoded information at the receiver and the encoded information at the transmitter be strictly consistent, meaning that bit-level error-free transmission is achieved. However, semantic communication does not require strict consistency between the decoded and encoded sequences; it only requires that the semantic information recovered by the receiver match the semantic information extracted by the transmitter. Because semantic communication relaxes the error-free transmission requirement, it avoids the transmission of redundant data. Therefore, when transmitting the same amount of source information, the amount of data transmitted is smaller, without compromising transmission accuracy, resulting in higher system efficiency.

[0003] In recent years, semantic communication has attracted considerable attention due to its enormous potential in the future of communications, and many researchers have begun designing semantic communication systems. In their paper "Deep learning for joint source-channel coding of text" (IEEE Int'l. Conf. Acoustics Speech Signal Process. (ICASSP), Algary, AB, Canada, pp. 2326–2330, 2018), Farsad N, Rao M, Goldsmith A, et al. proposed an end-to-end semantic communication system that jointly designs source and channel coding for transmitting fixed-length text sentences over simple channels. In their paper "Deep learning enabled semantic communication systems" (IEEE Trans. on Signal Processing, vol. 69, pp. 2663–2675, 2021), Xie H, Qin Z, Li GY, et al. developed a more powerful joint semantic channel coding based on the Transformer model, enabling it to more flexibly process text sentences of varying lengths in complex channel environments. The system also compared its performance with that of traditional communication systems. Experiments demonstrated that the semantic communication system outperformed traditional communication systems in terms of word error rate. However, communication systems often involve multiple users in practical applications. Therefore, Xie H, Qin Z, Tao X et al. designed a task-oriented multi-user semantic communication system in their published paper "Task-Oriented Multi-User Semantic Communications" (arXiv preprint arXiv:2112.10255, 2021), and developed many-to-one and many-to-many communication systems for different tasks. Experimental results show that the designed system outperforms traditional communication systems in terms of channel robustness, transmission delay, and execution performance under various task-specific indicators. However, their system does not consider broadcast communication with a single transmitter and multiple receivers, which makes this system have certain limitations in practical applications. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the deficiencies in the prior art, the present invention provides a one-to-many multi-user semantic communication model and communication method; while improving the system model training speed, it also improves the performance gain of the communication system, and directly utilizes different semantic features (emotion, purpose, person, language, etc.) in the source information to distinguish different users, thereby simplifying the transmission procedure of multi-user communication and improving information transmission efficiency.

[0005] Technical solution: In the first aspect, the present invention provides a one-to-many multi-user semantic communication method, comprising:

[0006] Collect various types of text sentences according to preset user needs;

[0007] Combine text sentences of different types into a text sequence and convert it into a digital ID sequence as the sending information of the sender;

[0008] The information sent is semantically encoded and channel-coded based on the attention mechanism and linear neural network at the sending end to generate a communication signal for channel transmission;

[0009] Transmit the communication signal one-to-many to each receiving end;

[0010] Each receiving end performs channel decoding and semantic decoding on the received communication signal to recover the original sentence sent by the sending end;

[0011] The recovered original sentence is input into the semantic recognizer based on the distilled bidirectional language representation pre-training model, and the corresponding sentence is output according to the user's needs;

[0012] Among them, the semantic channel encoder at the sending end and the channel semantic decoder at each receiving end are trained multiple times based on the transfer learning method to adjust the effectiveness of information transmission in the one-to-many semantic communication model.

[0013] In a further embodiment, the method of combining text sentences of different types into a text sequence and converting it into a digital ID sequence as the sending information of the sending end is:

[0014] Perform word segmentation on all text sentences to obtain text sentences without stop words;

[0015] Each word or symbol obtained after word segmentation is assigned a digital ID to construct a vocabulary as a background knowledge base; the expression set of the background knowledge base is: K = {"a":1, "b":2..., word:num,...}, word is the word or symbol obtained after word segmentation, and num is the digital ID corresponding to the word;

[0016] The sentences sent to different users are combined into a long text sequence, and the order of the sentences of all users is randomly shuffled during the combination process; the expression set of the text sequence is: Represents the words in the sentences sent to different users, <sep>It is the separator between each sentence; then, by querying each word and the corresponding ID in the background knowledge base, the text sequence is converted into a digital ID sequence as the sending information.

[0017] In a further embodiment, the method for performing semantic encoding and channel encoding on the transmitted information based on the attention mechanism and the linear neural network at the transmitting end to generate a communication signal for channel transmission is:

[0018] Input the digital ID sequence into the embedding layer, and convert each word ID into a vector through the embedding layer to obtain a vector symbol sequence;

[0019] The attention-based encoding layer uses a multi-head attention mechanism to obtain the important dependencies between the word vector at each position in the vector symbol sequence and the word vectors at other positions, thereby obtaining the meaning of the word at each position in the sentence and outputting the semantic symbol sequence of the sentence.

[0020] The output semantic symbol sequence is imported into a linear neural network with a linear rectified unit activation function to obtain a communication signal that is convenient for channel transmission;

[0021] Among them, the expression of the linear rectification unit activation function is:

[0022] f(z)=max(0,z)

[0023] Where max(·) represents the maximum value of the object in the brackets, and z represents the input of the activation function of the linear rectification unit.

[0024] In a further embodiment, a method for performing one-to-many transmission of a communication signal to each receiving end, wherein each receiving end performs channel decoding and semantic decoding on the received communication signal to recover the original sentence sent by the sending end includes:

[0025] The transmitting end sends a communication signal to each receiving end; wherein the transmitting end establishes a transmission relationship with each receiving end through a channel;

[0026] The communication signal received by each receiving end is introduced into a linear neural network with a linear rectified unit activation function for conversion to obtain a semantic symbol sequence with increased dimension and reconstruction;

[0027] The dimensionality-enhanced and reshaped semantic symbol sequence is imported into the encoding layer based on the attention mechanism to perform inverse decoding on the semantic features and restore the original sentence sent by the sender.

[0028] In a further embodiment, the recovered original sentence is input into a semantic recognizer based on a distilled bidirectional language representation pre-training model, and a method for outputting a corresponding sentence according to the user's needs is as follows:

[0029] The recovered original sentences are imported into the distilled bidirectional language representation pre-training model. The original sentences contain sentences sent to two different users. The distilled bidirectional language representation pre-training model, which has been pre-trained on billions of sentences, calculates the overall semantic features of each sentence through the attention mechanism and compares the overall semantic features of different sentences to obtain the differences in the overall semantic features between the sentences.

[0030] Output corresponding sentences based on the differences in overall semantic features between sentences and preset user needs.

[0031] In a further embodiment, the semantic channel encoder at the transmitting end and the channel semantic decoder at each receiving end are trained multiple times iteratively based on a transfer learning method, and the method for adjusting the effectiveness of information transmission in the one-to-many semantic communication model includes:

[0032] In the training network, the preset training set is used to train the transmitter and the first receiver multiple times to obtain the training errors of the transmitter and the first receiver respectively. The number of epochs of network training is initialized to 1, the upper limit of the number of iterations is 80, the learning rate is 0.001, and the Adam optimization algorithm is used as the network training optimizer.

[0033] The training errors of the transmitter and the first receiver are introduced into the error loss function for calculation and back-propagation to obtain the network optimization parameters of the transmitter and the first receiver respectively; wherein, during the iterative process, it is determined whether the error loss function has converged or whether the number of iterative training times of the transmitter and the first receiver has reached the upper limit;

[0034] The second receiving end is trained multiple times based on the judgment result to obtain the training error of the second receiving end. When training the second receiving end, the training parameters of the final iteration of the first receiving end are imported, and the training parameters of the sending end are frozen.

[0035] The training error of the second receiving end is introduced into the error loss function for calculation and back propagation to obtain the network optimization parameters of the second receiving end respectively; wherein, the network optimization parameters of the second receiving end are introduced based on the number of receiving ends and the network optimization parameters of several receiving ends are obtained by cyclic training in sequence;

[0036] The one-to-many semantic communication model is adjusted based on network optimization parameters of a sending end, a first receiving end, a second receiving end, and a plurality of receiving ends, so that the adjusted one-to-many semantic communication model remains effective;

[0037] Among them, the error loss function adopts the cross entropy function, and the calculation formula is as follows:

[0038]

[0039] Among them, s represents the input sentence of the encoder, represents the output sentence of the decoder, α represents the trainable parameters in the semantic encoder, β represents the trainable parameters in the channel encoder, χ represents the trainable parameters in the channel decoder, δ represents the trainable parameters in the semantic decoder, q(w l ) is the lth word w l The actual probability of appearing in the input sentence s, p(w l ) is the lth word w l Appears in output statement When all the back propagation of all batches of data in the training data is completed, it is 1 epoch.

[0040] In a second aspect, the present invention provides a one-to-many multi-user semantic communication model, comprising:

[0041] A sending end and multiple receiving ends that establish communication relationships with the sending end and are independent of each other.

[0042] In a further embodiment, the transmitter is integrated with an embedding layer, three attention-based encoding layers, and two densely connected layers;

[0043] The embedding layer is used to convert the digital ID sequence into a vector symbol sequence, thereby increasing the semantic feature input dimension to capture the relationship between each word and other words;

[0044] The encoding layer is used for semantic extraction. The encoding layer calculates the important dependencies between the word vector at each position in the vector symbol sequence and the word vectors at other positions through a multi-head self-attention mechanism, obtains the meaning of each word in the sentence, and thus extracts the semantic symbol sequence of the entire sentence.

[0045] The dense connection layer is used as a channel coder, and the dense connection layer reshapes the semantic features output by the coding layer into a symbol sequence through a linear neural network dimension reduction.

[0046] In a further embodiment, the multiple receiving ends are in different channel environments, so the expressions of the signals received by each receiving end are respectively:

[0047] Y1=H1X+W1

[0048] …

[0049] Y k =H k X+W k

[0050] …

[0051] Y n =H n X+W n

[0052] Where X is the input signal of the channel, Y1 represents the signal received by the first receiving end, and Y k represents the signal received by the kth receiver, Y n Indicates the n The signal received by the receiving end; H k 、H n are the coefficients of the linear channels associated with different receivers, W k 、W n are independent and identically distributed Gaussian noises.

[0053] In a further embodiment, the receiving end is integrated with two densely connected layers, three attention-based decoding layers and one classification layer;

[0054] Among them, two densely connected layers serve as channel decoding, and the densely connected layers reshape the received communication signal into a semantic symbol sequence through a linear neural network.

[0055] The decoding layer performs an inverse operation on the semantic symbol sequence through a multi-head autonomous mechanism, and outputs the original sentence through the classification layer operation.

[0056] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0057] (1) By designing a one-to-many multi-user semantic communication model, the system model training speed is improved while the performance gain of the communication system is improved. Different semantic features (emotion, purpose, person, language, etc.) in the source information are directly used to distinguish different user needs and users, simplifying the transmission process of multi-user communication, improving the efficiency of information transmission, and further expanding the application of semantic communication in multi-user communication.

[0058] (2) The system model is combined with the transfer learning method to train the receiver in different channel environments, which reduces the training cost and improves the training speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a flow chart of the present invention;

[0060] Figure 2 This is a model framework diagram of the sending end in the present invention;

[0061] Figure 3 This is a model framework diagram of the receiving end in the present invention;

[0062] Figure 4 It is a holistic model framework for one-to-many multi-user semantic communication;

[0063] Figure 5 It is the training framework diagram of the system model;

[0064] Figure 6 This is a diagram showing an embodiment of the comparison of BLEU scores of two receiving ends under different channel conditions using the present invention and other existing benchmarks;

[0065] Figure 7 This is a second embodiment of the comparison of BLEU scores of two receiving ends under different channel conditions using the present invention and other existing benchmarks;

[0066] Figure 8 This is a comparison chart of training speed using the transfer method and not using the transfer method. DETAILED DESCRIPTION

[0067] In order to more fully understand the technical content of the present invention, the technical solution of the present invention is further introduced and illustrated in conjunction with specific embodiments below, but is not limited thereto.

[0068] Example 1: In this example, there are two receiving terminals and one transmitting terminal. The user of the first receiving terminal is set as user 1, and the user of the second receiving terminal is preset as user 2. The communication method of this example specifically includes:

[0069] Collect two different types of text sentences based on preset user needs;

[0070] Combine two different types of text sentences in pairs and convert them into digital ID sequences as the sending information of the sender;

[0071] The sender performs semantic encoding and channel encoding on the transmitted information based on the attention mechanism and linear neural network to generate a communication signal for channel transmission;

[0072] The communication signal is transmitted one-to-many to each receiving end. Each receiving end performs channel decoding and semantic decoding on the received communication signal to recover the original sentence sent by the sending end.

[0073] The recovered original sentence is input into the semantic recognizer based on the distilled bidirectional language representation pre-training model, and the corresponding sentence is output according to the user's needs;

[0074] Among them, the semantic channel encoder at the sending end and the channel semantic decoder at each receiving end are trained multiple times based on the transfer learning method to adjust the effectiveness of information transmission in the one-to-many semantic communication model.

[0075] In this embodiment, two different types of text sentences are combined in pairs and converted into a text sequence. The method for sending information as a sending end is as follows:

[0076] Perform word segmentation on all text sentences to obtain text sentences without stop words. Assign each word or symbol obtained after word segmentation a numerical ID to construct a vocabulary table as the background knowledge base K = {"a":1, "b":2..., word:num,...}, where word is the word or symbol obtained after word segmentation and num is the numerical ID corresponding to the word.

[0077] Combine the sentences sent to user 1 and user 2 in pairs, and randomly shuffle the order of the two sentences to generate the following text sequence: in represents the L1th word in the sentence sent to user 1, represents the L2th word in the sentence sent to user 2, <sep>It is the separator between two sentences. Then, by querying each word and the corresponding ID in the background knowledge base K, the text sequence is converted into a digital ID sequence as the sent information.

[0078] The sending end performs semantic encoding and channel encoding on the transmitted information based on the attention mechanism and linear neural network, thereby generating the communication signal for channel transmission as follows:

[0079] Input the sent information into the embedding layer, and obtain a vector symbol sequence by converting each word ID into a vector;

[0080] The attention-based encoding layer uses a multi-head attention mechanism to obtain the important dependencies between the word vector at each position in the vector symbol sequence and the word vectors at other positions, thereby obtaining the meaning of the word at each position in the sentence and outputting the semantic symbol sequence of the sentence.

[0081] The output semantic symbol sequence is imported into a linear neural network with a linear rectified unit activation function to obtain a communication signal that is convenient for channel transmission;

[0082] Among them, the expression of the linear rectification unit activation function is:

[0083] f(z)=max(0,z)

[0084] Where max(·) represents the maximum value of the object in the brackets, and z represents the input of the activation function of the linear rectification unit.

[0085] The method of performing one-to-many transmission of a communication signal to each receiving end, wherein each receiving end performs channel decoding and semantic decoding on the received communication signal to restore the original sentence sent by the sending end includes:

[0086] The transmitter transmits the communication signal to each receiver through the channel;

[0087] The communication signal received by each receiving end is introduced into a linear neural network with a linear rectified unit activation function for conversion to obtain a semantic symbol sequence with increased dimension and reconstruction;

[0088] The dimensionality-enhanced and reshaped semantic symbol sequence is imported into the encoding layer based on the attention mechanism to perform inverse decoding on the semantic features and restore the original sentence sent by the sender.

[0089] The recovered original sentence is input into the semantic recognizer based on the distilled bidirectional language representation pre-training model, and the method of outputting the corresponding sentence according to the user's needs is as follows:

[0090] The recovered original sentence is imported into the distilled bidirectional language representation pre-training model. The original sentence contains sentences sent to two different users. The distilled bidirectional language representation pre-training model, which has been pre-trained on billions of sentences, calculates the overall semantic features of each sentence through the attention mechanism.

[0091] According to the differences in the overall semantic features between sentences and the needs of users, the corresponding sentences are output.

[0092] The semantic channel encoder at the transmitter and the semantic channel decoder at each receiver are trained iteratively multiple times based on the transfer learning method. The method for adjusting the effectiveness of information transmission in the one-to-many semantic communication model includes:

[0093] The first step is to train the transmitter and the first receiver. First, the trainable parameters of the network model are randomly initialized. The number of network training iterations is initialized to 1, the maximum number of iterations is 80, the learning rate is 0.001, and the Adam optimization algorithm is used as the network training optimizer. The training data is fed into the network training in batches, with adjustable batch size. The training error of each batch is backpropagated to optimize the network parameters. The error loss function uses the cross entropy function, calculated as follows:

[0094]

[0095] Among them, s represents the input sentence of the encoder, represents the output sentence of the decoder, α represents the trainable parameters in the semantic encoder, β represents the trainable parameters in the channel encoder, χ represents the trainable parameters in the channel decoder, δ represents the trainable parameters in the semantic decoder, q(w l ) is the lth word w l The actual probability of appearing in the input sentence s, p(w l ) is the lth word w l Appears in output statement The predicted probability in , when all the batches of data in the training data are fully back-propagated, it is 1 epoch.

[0096] Determine whether the loss function of the model training has converged or the number of iterations has reached the upper limit. If so, proceed to the second step. If not, increase the number of training iterations by one and continue training the model.

[0097] In the second step, train the second receiver. First, load the pre-trained transmitter and the first receiver from the first step. Then, freeze the transmitter parameters and import the parameters of the first receiver into the second receiver. Finally, repeat the steps in the first step to train the second receiver.

[0098] In Example 2, there are two receiving terminals and one sending terminal. Figures 2 to 5 Further describing a one-to-many multi-user semantic communication model in this embodiment, the communication model includes:

[0099] The sending end and two receiving ends that establish communication relationships with the sending end and are independent of each other;

[0100] The sending end is integrated with an embedding layer, three attention-based encoding layers, and two densely connected layers;

[0101] The embedding layer is used to convert each word ID into a vector symbol sequence, thereby increasing the semantic feature input dimension to capture the relationship between each word and other words;

[0102] The encoding layer is used for semantic extraction. The encoding layer uses a multi-head self-attention mechanism to calculate the important dependencies between the word vector at each position in the vector symbol sequence and the word vectors at other positions, obtain the meaning of each word in the sentence, and thus extract the semantic symbol sequence of the entire sentence.

[0103] The dense connection layer is used as a channel coder. The dense connection layer reshapes the semantic features output by the coding layer into a symbol sequence through linear neural network dimensionality reduction.

[0104] The two receiving ends are in different channel environments, so the expressions of the signals received by the two receiving ends are:

[0105] Y1=H1X+W1

[0106] Y2=H2X+W2

[0107] Where X is the input signal of the channel, Y1 and Y2 represent the signals received by the two receiving ends respectively, H1 and H2 are the coefficients of the linear channel, and W1 and W2 are independent and identically distributed Gaussian noises.

[0108] The receiving end is integrated with two densely connected layers, three attention-based decoding layers and a classification layer;

[0109] Among them, the two densely connected layers serve as channel decoding. The densely connected layers reshape the received communication signals into semantic symbol sequences through linear neural networks.

[0110] The decoding layer performs inverse operations on the semantic symbol sequence through a multi-head autonomous mechanism, and outputs the original sentence through the classification layer operation.

[0111] Example 3:

[0112] Combine Figures 6 to 8 The application of this embodiment in the simulation experiment is further described as follows:

[0113] Simulation conditions and parameter settings:

[0114] Simulation experiments were conducted on a Python 3.7, PyTorch 1.8.0 platform. The semantic encoder and decoder consist of three Transformer encoding layers and three Transformer decoding layers, respectively, with eight self-attention heads in each layer. The densely connected layers of the channel encoder and decoder each have 16 and 128 units, respectively. The network was trained with a maximum iteration count of 80 and a learning rate of 0.001. The Adam optimization algorithm was used as the network training optimizer, and the batch size was 256 data points.

[0115] Simulation content:

[0116] In the simulation of the present invention, all benchmarks have two receivers, where the first receiver is tested in an AWGN channel and the second receiver is tested in a Rayleigh fading channel. Figure 6 and Figure 7 The relationship between the BLEU score and the signal-to-noise ratio of two receivers with different benchmarks in different channel environments is shown. The horizontal axis in the two figures represents different signal-to-noise ratios (dB), and the vertical axis represents the BLEU score. The broken line marked with a rectangle represents the BLEU score curve using the method of the present invention. The broken line marked with a diamond and the broken line marked with a triangle represent the BLEU score curve using the traditional communication model, in which the source coding uses Huffman coding, the channel coding uses RS coding, the modulation method is 64-QAM, and the multiple access technology is CDMA. The broken line marked with a circle represents the BLEU score curve of the semantic communication model using TextCNN to construct a semantic recognizer. The signal-to-noise ratio ranges from 0dB to 18dB. It can be observed that in the AWGN channel and the Rayleigh fading channel, when the signal-to-noise ratio changes, the BLEU score of the present invention changes more smoothly than that of the traditional communication model. This shows that the semantic communication system based on DNN is more robust to different channel conditions, especially under low signal-to-noise ratio conditions. It can also be seen that in AWGN channels and Rayleigh fading channels, the BLEU scores obtained by the present invention and another semantic communication model are higher than those of the traditional communication model. This is because only relevant semantic information is transmitted in the semantic communication system. Although the BLEU score of the traditional communication model in the AWGN channel is slightly higher than that of the proposed model when the SNR is higher than 12dB, the present invention performs better than the traditional communication model in the more complex Rayleigh fading channel. In addition, because the DistilBERT pre-trained model used in the present invention has been trained on billions of data, while the TextCNN model has only been trained on less data, the present invention is able to obtain a higher BLEU score than the semantic communication model using the Text CNN recognizer.

[0117] Attachment Figure 8 This is a comparison chart of the training speed when the second receiving end is retrained using the training method of the present invention and not using the transfer learning method. Figure 7 The horizontal axis represents the number of training iterations, and the vertical axis represents the loss function value. The rectangular line represents the convergence curve for the transfer learning method, while the circular line represents the convergence curve for the method without transfer learning. Comparing the training speed convergence curves of the two methods shows that the method with transfer learning trains significantly faster than the method without transfer learning. The method with transfer learning converges after only about 5 training epochs, while the method without transfer learning does not reach loss convergence until 20 training epochs.

[0118] In summary, the present invention designs a one-to-many multi-user semantic communication model, which further expands the application of semantic communication in multi-user communication; secondly, the system model directly utilizes the semantic features of each user's information to distinguish different users, simplifies the transmission procedure of multi-user communication, and improves the efficiency of information transmission; finally, the system model combines the transfer learning method to train the receiving ends in different channel environments, reducing the training cost and improving the training speed.

[0119] The embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0120] The embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0121] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0122] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0123] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0124] The above is only a preferred embodiment of the present invention. Without departing from the technical principles of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be considered as the scope of protection of the present invention.< / sep> < / sep>

Claims

1. A one-to-many multi-user semantic communication method, characterized in that: include: Collect various types of text sentences according to preset user needs; Combine text sentences of different types into a text sequence and convert it into a digital ID sequence as the sending information of the sender; The information sent is semantically encoded and channel-coded based on the attention mechanism and linear neural network at the sending end to generate a communication signal for channel transmission; Transmit the communication signal one-to-many to each receiving end; Each receiving end performs channel decoding and semantic decoding on the received communication signal to recover the original sentence sent by the sending end; The recovered original sentence is input into the semantic recognizer based on the distilled bidirectional language representation pre-training model, and the corresponding sentence is output according to the user's needs; The semantic channel encoder at the transmitter and the semantic channel decoder at each receiver are trained iteratively multiple times based on the transfer learning method to adjust the effectiveness of information transmission in the one-to-many semantic communication model. The method of combining text statements of different types into a text sequence and converting it into a digital ID sequence as the sending information of the sender is as follows: Perform word segmentation on all text sentences to obtain text sentences without stop words; Each word or symbol obtained after word segmentation is assigned a numerical ID to construct a vocabulary as a background knowledge base. The expression set of the background knowledge base is: K = { "a": 1, "b": 2..., word: num,...}, where word is the word or symbol obtained after word segmentation and num is the numerical ID corresponding to the word. The sentences sent to different users are combined into a long text sequence, and the order of the sentences of all users is randomly shuffled during the combination process; the expression set of the text sequence is: , 、 、 Represents the words in the sentences sent to different users, <sep> It is the separator between each sentence; then, by querying each word and the corresponding ID in the background knowledge base, the text sequence is converted into a digital ID sequence as the sent information;< / sep> The method of performing one-to-many transmission of a communication signal to each receiving end, wherein each receiving end performs channel decoding and semantic decoding on the received communication signal to restore the original sentence sent by the sending end includes: The transmitting end sends a communication signal to each receiving end; wherein the transmitting end establishes a transmission relationship with each receiving end through a channel; The communication signal received by each receiving end is introduced into a linear neural network with a linear rectified unit activation function for conversion to obtain a semantic symbol sequence with increased dimension and reconstruction; The semantic symbol sequence reshaped by dimensionality increase is imported into the encoding layer based on the attention mechanism to perform inverse operation decoding on the semantic features and restore the original sentence sent by the sender; The recovered original sentence is input into the semantic recognizer based on the distilled bidirectional language representation pre-training model, and the method of outputting the corresponding sentence according to the user's needs is as follows: The recovered original sentences are imported into the distilled bidirectional language representation pre-training model. The original sentences contain sentences sent to two different users. The distilled bidirectional language representation pre-training model, which has been pre-trained on billions of sentences, calculates the overall semantic features of each sentence through the attention mechanism and compares the overall semantic features of different sentences to obtain the differences in the overall semantic features between the sentences. Output corresponding sentences based on the differences in overall semantic features between sentences and preset user needs.

2. The one-to-many multi-user semantic communication method according to claim 1, characterized in that: The method for sending information by performing semantic encoding and channel encoding based on the attention mechanism and linear neural network at the sending end to generate the communication signal for channel transmission is as follows: Input the digital ID sequence into the embedding layer, and convert each word ID into a vector through the embedding layer to obtain a vector symbol sequence; The attention-based encoding layer uses a multi-head attention mechanism to obtain the important dependencies between the word vector at each position in the vector symbol sequence and the word vectors at other positions, thereby obtaining the meaning of the word at each position in the sentence and outputting the semantic symbol sequence of the sentence. The output semantic symbol sequence is imported into a linear neural network with a linear rectified unit activation function to obtain a communication signal that is convenient for channel transmission; Among them, the expression of the linear rectification unit activation function is: ； Where, Indicates calculating the maximum value of the objects in the brackets. Represents the input to the rectified linear unit activation function.

3. The one-to-many multi-user semantic communication method according to claim 1, characterized in that: The semantic channel encoder at the transmitter and the semantic channel decoder at each receiver are trained iteratively multiple times based on the transfer learning method. The method for adjusting the effectiveness of information transmission in the one-to-many semantic communication model includes: In the training network, the preset training set is used to train the transmitter and the first receiver multiple times to obtain the training errors of the transmitter and the first receiver respectively. The number of epochs of network training is initialized to 1, the upper limit of the number of iterations is 80, the learning rate is 0.001, and the Adam optimization algorithm is used as the network training optimizer. The training errors of the transmitter and the first receiver are introduced into the error loss function for calculation and back-propagation to obtain the network optimization parameters of the transmitter and the first receiver respectively; wherein, during the iterative process, it is determined whether the error loss function has converged or whether the number of iterative training times of the transmitter and the first receiver has reached the upper limit; The second receiving end is trained multiple times based on the judgment result to obtain the training error of the second receiving end. When training the second receiving end, the training parameters of the final iteration of the first receiving end are imported, and the training parameters of the sending end are frozen. The training error of the second receiving end is introduced into the error loss function for calculation and back propagation to obtain the network optimization parameters of the second receiving end respectively; wherein, the network optimization parameters of the second receiving end are introduced based on the number of receiving ends and the network optimization parameters of several receiving ends are obtained by cyclic training in sequence; The one-to-many semantic communication model is adjusted based on network optimization parameters of a sending end, a first receiving end, a second receiving end, and a plurality of receiving ends, so that the adjusted one-to-many semantic communication model remains effective; Among them, the error loss function adopts the cross entropy function, and the calculation formula is as follows: ； in, represents the input sentence of the encoder, represents the output sentence of the decoder, represents the trainable parameters in the semantic encoder, represents the trainable parameters in the channel encoder, represents the trainable parameters in the channel decoder, represents the trainable parameters in the semantic decoder, It is words Appears in input statement The actual probability of It is words Appears in output statement When all the back propagation of all batches of data in the training data is completed, it is 1 epoch.

4. A one-to-many multi-user semantic communication model, characterized in that The communication method according to claim 1, comprising: A sending end and multiple receiving ends that establish communication relationships with the sending end and are independent of each other.

5. The one-to-many multi-user semantic communication model according to claim 4 is characterized in that: The transmitter is integrated with an embedding layer, three attention-based encoding layers, and two densely connected layers; The embedding layer is used to convert the digital ID sequence into a vector symbol sequence, thereby increasing the semantic feature input dimension to capture the relationship between each word and other words; The encoding layer is used for semantic extraction. The encoding layer calculates the important dependencies between the word vector at each position in the vector symbol sequence and the word vectors at other positions through a multi-head self-attention mechanism, obtains the meaning of each word in the sentence, and thus extracts the semantic symbol sequence of the entire sentence. ; The dense connection layer is used as a channel coder, and the dense connection layer reshapes the semantic features output by the coding layer into a symbol sequence through a linear neural network dimension reduction. .

6. The one-to-many multi-user semantic communication model according to claim 5, characterized in that: The multiple receiving ends are in different channel environments, so the expressions of the signals received by each receiving end are respectively: ； is the input signal of the channel, Indicates the signal received by the first receiving end, represents the signal received by the kth receiver, Indicates the The signal received by the receiving end; 、 are the coefficients of the linear channels associated with different receivers, 、 are independent and identically distributed Gaussian noises.

7. The one-to-many multi-user semantic communication model according to claim 4 is characterized in that: The receiving end is integrated with two densely connected layers, three decoding layers based on the attention mechanism and a classification layer; Among them, two densely connected layers serve as channel decoding, and the densely connected layers reshape the received communication signal into a semantic symbol sequence through a linear neural network. The decoding layer performs an inverse operation on the semantic symbol sequence through a multi-head autonomous mechanism, and outputs the original sentence through the classification layer operation.

Images