Method of semantic communication based on information bottleneck

Abstract

The application discloses a semantic communication method based on information bottleneck, and mainly solves the problem that the existing technology cannot consider communication overhead and bandwidth demand in semantic communication, and is difficult to deploy in a resource-constrained scene.The implementation scheme is as follows: a semantic channel joint encoder and a semantic channel joint decoder are respectively constructed at a sending end and a receiving end;a dictionary is constructed according to words in communication transmission of the sending end and the receiving end;a text vector is obtained by using the dictionary;the text vector is subjected to semantic channel joint encoding to obtain a semantic feature vector;the semantic feature vector is subjected to semantic channel joint decoding to output a received text sentence;an objective function based on information bottleneck is designed and calculated;the objective function is minimized, weights of modules in the encoder and the decoder are iteratively updated, and a final received text sentence is output.The application can reduce semantic communication overhead and bandwidth demand, improve communication performance in a resource-constrained semantic communication scene, and can be used for data communication in a resource-limited Internet of Things scene.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wireless communication, and further relates to a semantic communication method, which can be used for data communication under resource restriction in an Internet of Things scenario. BACKGROUND

[0002] With the development of wireless communication and artificial intelligence, the intelligent needs of users will be further explored and realized. The huge data generated by "intelligent connection of all things" will bring serious challenges to wireless communication, and wireless communication systems urgently need technological breakthroughs. Shannon and Weaver proposed three levels of communication. The first level is the technical problem of communication, mainly studying how to accurately transmit the symbols of communication. The second level is the semantic problem of communication, mainly studying whether the transmitted symbols accurately express the intended meaning. The third level is the effectiveness problem of communication, mainly solving the problem of how to effectively affect behavior according to the expected way of the received semantics. Semantic communication is a new communication architecture that expands the level of information research from syntax information to semantic information. By extracting user semantic information, it compresses data and provides a new research angle for communication system optimization.

[0003] A semantic communication method for intelligent tasks is disclosed in the patent document with application number 202210658290 by Beijing University of Posts and Telecommunications. The implementation scheme is as follows: first, the sending device collects image data, extracts feature maps through a convolutional neural network, and obtains the importance weight of the feature map with respect to the semantic concept by taking the average of the gradient of the feature map with respect to the semantic concept corresponding to the neuron activation value. The semantic concept and the feature map are sorted; second, the obtained semantic relationship is used to crop and compress the feature map, and then the feature map with high importance weight is retained. The compressed feature map is transmitted through the wireless channel; then the received signal is demodulated by the receiving end, the feature map data is recovered, and the subsequent neural network is inputted to complete the intelligent task and return the result to the sending device. This method adopts a scheme of separating semantic encoding and channel encoding. Since the joint encoding of semantic channels is not considered, the performance of semantic encoding and channel encoding is not jointly optimal, thereby limiting the performance of semantic communication.

[0004] Qin Zhijin and Geoffrey Ye Li proposed a deep learning-based semantic communication system in their published paper "Deep Learning Enabled Semantic Communication Systems" (IEEE Transactions on Signal Processing, vol. 69, pp. 2663-2675, 2021). It consists of a sending end, a channel and a receiving end. The sending end includes a semantic encoder and a channel encoder, and the receiving end includes a channel decoder and a semantic decoder. Among them, the semantic encoder and the semantic decoder are respectively composed of multiple Transformer encoding layers and Transformer decoding layers, the channel encoder and the channel decoder are composed of fully connected layers with different neurons, and the communication channel is modeled as an untrainable neural network layer. This method explicitly defines the concept of semantic information at the sentence level. First, the channel encoder and decoder are trained based on maximizing mutual information to maximize the transmission rate, and then the entire system is trained based on minimizing cross-entropy to minimize semantic error. However, since the overhead and bandwidth requirements of semantic communication are not considered, it is difficult to deploy in resource-constrained situations and cannot be widely applied to the Internet of Everything scenario. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art and provide a semantic communication method based on information bottleneck to improve communication performance in resource-constrained semantic communication scenarios and expand application scenarios.

[0006] The technical idea for achieving the purpose of the present application is: through the constructed semantic channel joint encoding module, the input text sentence is subjected to text vectorization, semantic encoding and channel encoding, and the encoded semantic feature vector is selected and compressed, and the dimension of the redundant semantic feature vector is pruned to reduce the overhead and bandwidth requirements of semantic communication; through the constructed semantic channel joint decoding module, the semantic feature vector is subjected to channel decoding, semantic decoding and text recovery, and the received text sentence is output; through the design of the target function based on information bottleneck, the semantic channel joint encoder and the semantic channel joint decoder are optimized to reduce the overhead and bandwidth requirements of semantic communication and improve the communication performance in resource-constrained semantic communication scenarios.

[0007] According to the above idea, the technical scheme of the present application includes the following steps:

[0008] (1) constructing a semantic channel joint encoder composed of a text embedding module, a semantic encoding module, a channel encoding module and a feature selection module in the sending end;

[0009] (2) constructing a semantic channel joint decoder composed of a channel decoding module, a semantic decoding module and a text recovery module in the receiving end:

[0010] (3) Construct a dictionary according to the words that the sending end and the receiving end can transmit in the communication process;

[0011] (4) Obtain the text sequence of the text sentence to be transmitted by using the dictionary, and perform word embedding and position embedding on the text sequence of the text sentence by a text embedding module of the sending end to obtain a text vector;

[0012] (5) The text vector obtained in step (4) is sequentially encoded by a semantic encoding module and a channel encoding module of the sending end, and the semantic features are selected, compressed, and pruned by a feature selection module to obtain a semantic feature vector to be transmitted;

[0013] (6) The semantic feature vector to be transmitted is transmitted to a semantic channel joint decoder of the receiving end through a wireless channel, and the semantic feature vector is sequentially decoded by a channel decoding module and a semantic decoding module in the decoder, and is converted into corresponding words by a text recovery module to output a received text sentence;

[0014] (7) Design a target function L based on the information bottleneck theory, which is composed of a semantic error and a semantic redundancy weighted and summed by two mutual information terms IB , and calculate the semantic error L e and the semantic redundancy L r based on the text sentence to be transmitted and the received text sentence to obtain the target function value;

[0015] (8) Reduce the target function based on the gradient descent method, and iteratively update the weights of each module in the semantic channel joint encoder and the semantic channel joint decoder. After updating the weights in each round, return to step (5) to prune the redundant semantic features by the feature selection module, reduce the semantic communication overhead and bandwidth demand, and obtain the received text sentence through step (6). The iteration process is stopped when the target function reaches the minimum value, and the final received text sentence is obtained.

[0016] Compared with the prior art, the present application has the following advantages:

[0017] First, the present application constructs a semantic channel joint encoder, and performs semantic encoding and channel encoding based on the deep learning method, and optimizes jointly through end-to-end back propagation, which overcomes the problem of system performance reduction caused by the separate semantic encoding and channel encoding error propagation in the prior art during semantic communication, so that the joint optimization of the semantic encoding module and the channel encoding module can be optimal, and the performance of the semantic communication is improved.

[0018] Secondly, the system optimization scheme based on the information bottleneck theory is designed, the balance between the semantic communication performance and the overhead is described by using mutual information to obtain the target function, the communication overhead is reduced while the communication performance is ensured by minimizing the target function, the problem that the communication overhead and bandwidth demand are not considered in the semantic communication in the prior art is overcome, the application scenarios are expanded, and the communication performance can be ensured under the semantic communication resource constraint. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is an implementation flowchart of the present application;

[0020] Figure 2 is a structure schematic diagram of the joint encoder of the semantic channel constructed in the present application;

[0021] Figure 3 is a structure schematic diagram of the joint decoder of the semantic channel constructed in the present application;

[0022] Figure 4 is a comparison performance diagram of data communication simulation by using the present application and three existing methods respectively. DETAILED DESCRIPTION

[0023] The embodiments and effects of the present application will be further described in detail below with reference to the drawings.

[0024] Referring to Figure 1 , the implementation steps of the present example are as follows:

[0025] Step 1, constructing a joint encoder of a semantic channel.

[0026] Referring to Figure 2 , the implementation of the present step is as follows:

[0027] 1.1) establishing a text embedding module composed of a word embedding layer and a position embedding layer, the output dimension of the word embedding layer is set to 128, and the position embedding layer is realized by using the following sine and cosine functions:

[0028]

[0029]

[0030] wherein PE(pos,2i) represents the position information of the even dimension of the word vector, PE(pos,2i+1) represents the position information of the odd dimension of the word vector, pos represents the position of the word in the text sentence, i represents the dimension sequence number of the word vector, d model represents the dimension value of the word vector;

[0031] 1.2) Establish a semantic encoding module composed of N multi-head self-attention layers and a feedforward neural network, which includes two fully connected layers with output dimensions of 512 and 128, respectively, and the multi-head self-attention layer is implemented using the following self-attention function:

[0032]

[0033] wherein Q, K, and V represent the query matrix, key matrix, and value matrix after different linear transformations of the input text vector, T represents the transpose operation, QK T represents the relevance matrix of the query matrix Q and the key matrix K after the transpose operation, d K represents the number of columns of the key matrix K, and softmax(·) represents the softmax function.

[0034] 1.3) Establish a channel encoding module composed of a multi-layer perceptron, which includes two fully connected layers with output dimensions of 256 and 16, respectively.

[0035] 1.4) Establish a feature selection module composed of a custom fully connected layer with an output dimension of 16, and the expression of this module is as follows:

[0036]

[0037] wherein e and z represent the semantic feature vectors before and after feature selection, respectively, γ represents the semantic feature importance vector to be updated, represents the augmented weight matrix of the custom fully connected layer, ||·||2 represents the matrix 2-norm operation, and Tanh(·) represents the Tanh activation function.

[0038] 1.5) Cascade the above text embedding module, semantic encoding module, channel encoding module, and feature selection module in sequence to form a semantic channel joint encoder of the sending end.

[0039] Step 2, construct a semantic channel joint decoder.

[0040] Referring to Figure 3 , this step is implemented as follows:

[0041] 2.1) Establish a channel decoding module composed of a multi-layer perceptron, which includes two fully connected layers with output dimensions of 512 and 128, respectively.

[0042] 2.2) Establish a semantic decoding module composed of N multi-head self-attention layers and a feedforward neural network, which includes two fully connected layers with output dimensions of 512 and 128, respectively, and the multi-head self-attention layer is implemented using the same self-attention function as in step 1.2).

[0043] 2.3) Establish a text recovery module, which is composed of a fully connected layer and a softmax layer, the output dimension of the fully connected layer is set to the size of the dictionary, and the softmax layer is implemented by the following softmax function:

[0044]

[0045] wherein z i represents the i-th element of the input vector, n represents the dimension of the input vector, e (·) represents the exponential operation;

[0046] 2.4) The above channel decoding module, semantic decoding module and text recovery module are sequentially cascaded to form a semantic channel joint decoder of the receiving end.

[0047] Step 3, constructing a dictionary

[0048] In this example, the European Parliament Parallel Corpus is used as the common knowledge of the sending end and the receiving end, which is extracted from the meeting records of the European Parliament and consists of 2 million sentences and 53 million words.

[0049] According to the words that the sending end and the receiving end may transmit in the communication process, a dictionary is constructed:

[0050] 3.1) Tokenize all text sentences to be transmitted, count the frequency of each word, and number each word according to the frequency to obtain an initial dictionary;

[0051] 3.2) Add the start “ <start>"end" <end>"filling" <pad>" characters and their corresponding numbers, and words with a deletion frequency of less than 3, to obtain a final dictionary containing vocab_size words.

[0052] Step 4: Vectorize the text to obtain a text vector with position information.

[0053] 4.1) Align the length of the sentences of the text to be transmitted. If the sentence is more than 30 words, the first 30 words are truncated. If it is less than 30 words, padding characters are added to the end of the sentence to 30 words.

[0054] 4.2) Query the dictionary constructed in step (3) to obtain the corresponding number of each word in the text to be transmitted, and form a text sequence.

[0055] 4.3) Input the text sequence into the text embedding module of the semantic encoder, obtain the word vector and position vector through the word embedding layer and position embedding layer in it respectively, and add the two vectors to output the text vector x with position information.

[0056] Step 5: Jointly encode the text vector x in the semantic channel.

[0057] The communication overhead and bandwidth requirement of semantic communication is determined by the dimension of the semantic feature vector output by the semantic channel joint encoder. If more dimensions of symbols are transmitted, the receiver will receive a high-quality semantic feature vector, thereby improving the communication performance, but it will also result in higher communication overhead and bandwidth requirement. Therefore, the text vector needs to be semantically encoded and channel encoded, and the encoded vector needs to be selected and pruned.

[0058] 5.1) The text vector x obtained in step (4) is sequentially encoded by the semantic encoding module and the channel encoding module of the sending end to obtain the encoded semantic feature vector e:

[0059] e = C α (S β (x))

[0060] Where S β represents the semantic encoding module with parameter β, and C α represents the channel encoding module with parameter α.

[0061] 5.2) Select and compress the semantic feature vector e through the feature selection module, and according to the semantic feature importance vector γ in step 1.4), prune the semantic feature vector dimension less than the set threshold γ0 to reduce the semantic communication overhead and bandwidth requirement, and obtain the semantic feature vector z to be transmitted.

[0062] Step 6: Jointly decode the semantic feature vector z in the semantic channel.

[0063] The semantic feature vector z can be output only after channel decoding and semantic decoding and text recovery according to the decoded text vector, so the semantic feature vector to be transmitted needs to be transmitted to the semantic channel joint decoder at the receiving end through a wireless channel for decoding, which is implemented as follows:

[0064] 6.1) The semantic feature vector z to be transmitted obtained in step (5) is transmitted to the semantic channel joint encoder at the receiving end through a wireless channel to obtain the received semantic feature vector

[0065]

[0066] wherein ε represents a Gaussian white noise, which is subject to a Gaussian distribution with a mean of 0 and a variance of σ 2 , i.e. ε ~ N(0, σ 2 I);

[0067] 6.2) The received semantic feature vector is sequentially decoded by the channel decoding module and the semantic decoding module in the decoder to obtain the decoded text vector y:

[0068]

[0069] wherein represents the channel decoding module with a parameter of δ, represents the semantic decoding module with a parameter of χ;

[0070] 6.3) The decoded text vector y is subjected to text recovery by the text recovery module to be converted into corresponding words, and the received text sentence is output.

[0071] Step 7, design and calculate the target function based on the information bottleneck.

[0072] 7.1) According to the text sentence to be transmitted and the received text sentence, the semantic error L e is calculated:

[0073]

[0074] wherein Y represents a set of decoded text vectors y, represents a set of received semantic feature vectors , I(·) represents the mutual information function, E p(x,y) (·) represents the expectation under the joint distribution p(x, y), represents the expectation under the conditional distribution , represents the conditional distribution of y under .

[0075] 7.2) Calculate semantic redundancy L based on the text sentence to be transmitted and the received text sentence r :

[0076]

[0077] where X represents the set of input text vectors x, D KL (·) represents the divergence function, represents the conditional distribution of x under represents the distribution of ;

[0078] 7.3) Weighted addition of the calculated semantic error L e and semantic redundancy L r to obtain the value of the objective function L IB :

[0079] L IB = L e + β·L r

[0080] where β > 0 represents the weighting weight.

[0081] Step 8, minimize the objective function to update the weights of each module.

[0082] 8.1) Based on the following formula, the weights of each module in the semantic channel joint encoder and the semantic channel joint decoder are back propagated, and the weights of each module are iteratively updated:

[0083]

[0084] where W t+1 represents the updated weight, W t represents the weight before updating, η represents the learning rate of back propagation, represents the derivative of the objective function L IB with respect to the weight variable W t ;

[0085] 8.2) Return to step (5), encode the text vector based on the updated semantic channel joint encoder to obtain the updated semantic feature vector, then decode it through the updated semantic channel joint decoder in step (6) to obtain the updated received text sentence, and then recalculate the value of the objective function L IB through (7);

[0086] 8.3) Repeat step 8.2) until the objective function L IB The iteration process is stopped when the minimum value is reached, the final weight of each module is obtained, and the final received text sentence is output.

[0087] The effect of the application will be further described below in combination with a simulation experiment:

[0088] 1. Simulation experiment conditions.

[0089] The hardware platform of the simulation experiment of the application is: the processor is Intel(R) Xeon(R) Gold 5218 CPU, the main frequency is 2.30 GHz, and the memory is 128 GB. The graphics card is NVIDIA GeForce RTX 3090, and the memory is 64 GB.

[0090] The software platform of the simulation experiment of the application is: Windows 10 operating system and python 3.6.

[0091] The data set used in the simulation experiment of the application is the European Parliament Parallel Corpus, which is extracted from the meeting records of the European Parliament and published by the European Commission on May 15, 2012. There are 21 versions of European languages, the annotation format of the corpus is XML language, and it includes 2 million sentences and 53 million words.

[0092] 2. Simulation content and result analysis.

[0093] The simulation experiment of the application is to encode, transmit and decode the input text sentence by using the application and three prior art Huffman+Turbo traditional communication methods, LSTM semantic communication method and DeepSC semantic communication method, and to compare the input text sentence and the decoded text sentence; the bilingual evaluation substitution score (BLEU) evaluation index is used The communication performance of the four methods is evaluated, wherein P n is the accuracy rate of n words, w n is the weight coefficient. The higher the accuracy rate of consecutive n words in BLEU, the more accurate the semantic information recovery. The simulation results are shown in the accompanying Figure 4 .

[0094] In the simulation experiment, the three prior art sources are:

[0095] The Huffman+Turbo traditional communication method refers to a traditional communication method combining the source coding method proposed by Huffman et al. in "A method for the construction of minimum-redundancy codes[J]. Proceedings of the IRE, 1952, 40(9): 1098-1101." and the channel coding method proposed by Berrou et al. in "Near Shannon limit error-correcting coding and decoding: Turbo-codes. [C] Proceedings of ICC'93-IEEE International Conference on Communications. IEEE, 1993, 2: 1064-1070.", which is referred to as the Huffman+Turbo traditional communication method for short.

[0096] The prior art LSTM semantic communication method refers to an end-to-end semantic communication method based on LSTM for joint channel coding and decoding of semantics proposed by Farsad et al. in "Deep learning for joint source channel coding of text. 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018: 2326-2330.", which is referred to as the LSTM semantic communication method for short.

[0097] The prior art DeepSC semantic communication method refers to a semantic communication method based on Transformer for joint channel coding and decoding of semantics proposed by Geoffrey Ye Li et al. in "Deep Learning Enabled Semantic Communication Systems. IEEE Transactions on Signal Processing, vol. 69, 2021: 2663-2675.", which is referred to as the DeepSC semantic communication method for short.

[0098] From Figure 4 It can be seen that the performance of the scheme is better than that of the traditional communication using Humffman source coding and Turbo channel coding, especially in the case of low signal-to-noise ratio. This is because semantic communication focuses on information content rather than coded symbols, extracts the semantics of the text, filters out redundant information, further compresses the data, and avoids error propagation through joint decoding of the semantic channel, and can maintain high communication performance in poor communication environment. Compared with the existing semantic communication method, the scheme can effectively extract semantic information while reducing communication overhead and bandwidth requirement. When the communication overhead and bandwidth requirement are the same, the performance of the scheme is better than that of the LSTM and DeepSC semantic communication method.

[0099] The above simulation experiment shows that: the method takes the text sentence as the input, completes text vectorization, semantic information coding and semantic feature selection through the constructed semantic channel joint encoder, completes semantic information decoding and text recovery through the constructed semantic channel joint decoder, and designs the objective function using the information bottleneck theory for systematization. Not only does it overcome the error propagation problem caused by the separated source coding and channel coding in the traditional communication, but also reduces the semantic communication overhead and bandwidth requirement, improves the communication performance under the constraint of semantic communication resources, and expands the application scenarios.< / pad> < / end> < / start>

Claims

1. A semantic communication method based on information bottlenecks, characterized in that, Includes the following steps: (1) Construct a semantic-channel joint encoder at the transmitting end, consisting of a text embedding module, a semantic coding module, a channel coding module, and a feature selection module cascaded together; (2) Construct a semantic-channel joint decoder at the receiving end, consisting of a channel decoding module, a semantic decoding module, and a text recovery module cascaded together; (3) Construct a dictionary based on the words that the sender and receiver may transmit during the communication process; (4) Obtain the text sequence of the text statement to be transmitted using a dictionary, and perform word embedding and position embedding on the text sequence of the text statement through the text embedding module of the sending end to obtain the text vector; (5) The text vector obtained in step (4) is encoded sequentially through the semantic coding module and the channel coding module of the sending end, and selected, compressed and pruned through the feature selection module to obtain the semantic feature vector to be transmitted; the channel coding module is composed of a multilayer perceptron, which includes two fully connected layers, and the output dimensions of the two fully connected layers are set to 256 and 16 respectively. The feature selection module consists of a custom fully connected layer with an output dimension of 16, and the expression of the module is as follows; ； in, and These represent the semantic feature vectors before and after feature selection, respectively. This indicates the introduction of a semantic feature importance vector to be updated. This represents the augmented weight matrix of a custom fully connected layer. This represents the matrix 2-norm operation. express Activation function; (6) The semantic feature vector to be transmitted is transmitted to the semantic channel joint decoder of the receiving end through the wireless channel. The semantic feature vector is decoded by the channel decoding module and the semantic decoding module in the decoder in turn, and then converted into the corresponding words by the text recovery module, and the received text statement is output. (7) Based on the information bottleneck theory, design an objective function composed of the weighted sum of two mutual information terms: semantic error and semantic redundancy. It calculates semantic error based on the text statement to be transmitted and the text statement to be received. and semantic redundancy The objective function value is obtained; the objective function It is expressed as follows: ； Among them, the first item Indicates semantic error, the second term Indicates semantic redundancy. Indicates weight, This represents the input text vector. and These represent the semantic feature vectors before and after wireless channel transmission, respectively. This represents the result of joint decoding of the semantic channels. Represents the mutual information function. Represents the expectation function, Represents the divergence function; (8) Reduce the objective function based on gradient descent, iteratively update the weights of each module in the semantic channel joint encoder and semantic channel joint decoder, return to step (5) after each round of weight update, prune redundant semantic features through the feature selection module, reduce semantic communication overhead and bandwidth requirements, and then obtain the received text statement through step (6) until the objective function reaches the minimum value and the iteration process stops, and the final received text statement is obtained.

2. The method according to claim 1, characterized in that, The structure and parameters of each module in step (1) of the semantic channel joint encoder are as follows: The text embedding module consists of a cascaded word embedding layer and a position embedding layer. The output dimension of the word embedding layer is set to 128, and the position embedding layer is implemented using the following sine and cosine functions: ； in, This represents the positional information of the even-numbered dimensions of the word vector. This represents the positional information of the odd-numbered dimensions of the word vector. Indicates the position of words in a text sentence. The dimension number representing the word vector. The numerical value representing the dimension of the word vector; The semantic encoding module consists of The system consists of a multi-head self-attention layer and a feedforward neural network cascaded together. The feedforward neural network includes two fully connected layers with output dimensions of 512 and 128, respectively. The multi-head self-attention layer is implemented using the following self-attention function: ； in, , , These represent the query matrix, key matrix, and value matrix after the input text vector has undergone different linear transformations, respectively. This indicates the transpose operation. Representing the query matrix and transpose operation after key matrix The correlation matrix Key matrix The number of columns, express function.

3. The method according to claim 1, characterized in that, The structure and parameters of each module in the semantic channel joint decoder constructed in step (2) are as follows: The channel decoding module is composed of a multilayer perceptron, which includes two fully connected layers with output dimensions of 512 and 128, respectively. The semantic decoding module consists of It consists of a multi-head self-attention layer and a feedforward neural network cascaded together. The feedforward neural network includes two fully connected layers with output dimensions of 512 and 128, respectively. The multi-head self-attention layer is implemented using a self-attention function. The text recovery module consists of a fully connected layer and The layers are cascaded, and the output dimension of the fully connected layer is set to the size of the dictionary. Layer by Function implementation.

4. The method according to claim 1, characterized in that, Step (3) constructs a dictionary based on the words that the sending and receiving ends may transmit during the communication process, as follows: (3a) Segment all text sentences to be transmitted, count the frequency of each word, and number each word according to the frequency to obtain the initial dictionary; (3b) Add start, end, and fill characters and their corresponding numbers to the initial dictionary, delete words with too low frequency, and obtain the final containment size. A dictionary of words.

5. The method according to claim 1, characterized in that, In step (4), the text sequence of the text statement to be transmitted is obtained using a dictionary. The text embedding module at the sending end performs word embedding and position embedding on the text sequence to obtain a text vector, as follows: (4a) Align the length of the sentences in the text to be transmitted. If the sentence has more than 30 words, truncate the first 30 words. If it has less than 30 words, add padding characters to the end of the sentence to make it 30 words. (4b) Query the dictionary constructed in step (3) to obtain the number corresponding to each word in the text to be transmitted, and form a text sequence; (4c) Input the text sequence into the text embedding module of the semantic encoder, obtain word vectors and position vectors through the word embedding layer and position embedding layer, respectively, and add the two vectors to output a text vector with position information.

6. The method according to claim 1, characterized in that, In step (8), the objective function is reduced based on gradient descent, and the weights of each module in the semantic channel joint encoder and semantic channel joint decoder are iteratively updated as follows: (8a) Calculate the objective function based on the text statement to be transmitted and the text statement to be received. The value; (8b) Backpropagation is performed on the weights of each module in the semantic channel joint encoder and semantic channel joint decoder based on the following formula, and the weights of each module are iteratively updated: ； in, This represents the updated weights. This represents the weights before the update. This represents the learning rate during backpropagation. Describe the objective function For weight variables The derivative; (8c) Repeat step (8b) until the objective function is achieved. Once the minimum value is reached, the iteration process stops, and the final weights of each module are obtained.

Images