Rate matching puncturing method based on semantic importance

Abstract

The application provides a rate matching and deleting method based on semantic importance, a sending end rate matcher is added between a sending end semantic encoder and a channel encoder, semantic rate is compressed by reserving more important symbols in the output result of the semantic encoder, semantic loss is minimized, and a receiving end rate matcher is added between a channel decoder and a semantic decoder at the receiving end to reconstruct the deleted semantic slice. The technical scheme can effectively identify and reserve semantic slices containing more important semantic information, thereby reducing semantic loss caused in the rate matching process, and the scheme does not need to retrain the original semantic communication system and can be embedded in the existing semantic communication system.

Description

Technical Field

[0001] This invention relates to the fields of communication and artificial intelligence applications, and in particular to a rate matching pruning method based on semantic importance. Background Technology

[0002] With the deep integration of artificial intelligence and communication technologies, semantic encoding and decoding technologies based on deep neural networks (DNNs) are becoming a research hotspot in the field of communication. Unlike traditional communication systems that aim to transmit bit streams without loss, semantic communication systems aim to encode and decode information at the semantic level, thereby efficiently extracting and transmitting the "meaning" of information. This allows communication systems to ensure effective information transmission even under conditions of low signal-to-noise ratio or limited bandwidth. However, when coupling DNN-based semantic codecs with traditional channel codecs, a prominent technical challenge arises.

[0003] Existing DNN-based semantic codecs typically employ pre-trained models, with their output bitstream rate fixed after training. In practical applications, the transmission rate of wireless channels is significantly affected by noise, interference, fading, and other factors, exhibiting substantial dynamic variation. This mismatch between the fixed semantic rate and the dynamic maximum information transmission rate (i.e., Shannon channel capacity) leads to the following problems: when the channel capacity is lower than the semantic rate, semantic information becomes congested in the transmission link, causing not only partial semantic data loss and incomplete transmission, but also preventing the receiving semantic decoder from accurately reconstructing the original semantic information due to missing key semantic information, ultimately significantly reducing the quality of semantic decoding.

[0004] Existing research offers several solutions to address the aforementioned problems. One approach involves introducing an additional DNN layer to perform secondary compression on the semantic encoder output, thereby reducing the semantic rate. However, this approach is inherently static; the semantic rate is fixed once trained and cannot be adaptively adjusted in real-time according to changes in channel capacity. Another simple method is to randomly delete a portion of the semantic encoder output at the transmitter and then train at the receiver to mitigate the impact of semantic loss, thus achieving dynamic rate adjustment. However, this approach has a fatal flaw: it treats the semantic encoder output as equally important, undifferentiated data, completely ignoring the inherent differences in the importance of semantic information. This could potentially lead to the loss of crucial semantic information carrying core meaning, causing irreversible and catastrophic damage to the final semantic decoding quality. Furthermore, this method requires retraining the original semantic codec, resulting in higher implementation costs. Summary of the Invention

[0005] The present invention aims to at least partially solve one of the technical problems in the related art.

[0006] Therefore, the purpose of this invention is to propose a rate matching pruning method based on semantic importance.

[0007] To achieve the above objectives, embodiments of the present invention propose a rate matching pruning method based on semantic importance, comprising: At the transmitting end, a semantic encoder composed of DNN extracts and compresses semantic features from the input signal to generate a high-dimensional semantic representation. The transmitter rate matcher dynamically adjusts the output data volume according to the importance of semantic information to achieve a match between semantic rate and physical channel capacity; The processed data is then encoded and modulated by a channel encoder, converted into a channel signal, and sent into the channel for transmission. The receiver receives the received signal and sequentially performs channel decoding, rate matching, and semantic decoding on the received signal to reconstruct the original information.

[0008] Optionally, the transmitting end includes a semantic encoder, a transmitting end rate matcher, and a channel encoder, and the receiving end includes a channel decoder, a receiving end rate matcher, and a semantic decoder, wherein: The channel decoder corresponds to the channel encoder in function and is responsible for performing decoding operations; The receiver rate matcher functions as the transmitter rate matcher and is responsible for performing the reverse reconstruction operation. The semantic decoder corresponds to the semantic encoder and is responsible for performing the decoding operation.

[0009] Optional, also includes: Assuming the information source requires in Transmission length within s is The text sequence, ,in Indicates the transmission sequence number Embedding vectors of words, Indicates the dimension of the embedded vector; At the sending end, The output after passing through the DNN-based semantic encoder is ,in Represents the dimension of the hidden state. This is the input to the transmitter rate matcher; In the transmitter rate matcher, Each column in the code is denoted as a semantic slice, serving as the smallest unit for deletion. The slices are then processed according to their semantic importance. In Selective deletion is performed on each semantic slice, retaining only the most important ones. Each slice has the following output: , where N represents the number of semantic slices retained after deletion, i.e. the average code length per word; The semantic rate output by the semantic encoder is defined as... ,because If it is variable, then the semantic rate can be dynamically changed to adapt to the channel capacity. The semantic compression rate is defined as... , to indicate the degree of compression of semantic encoding; Output of the transmitter rate matcher After encoding and modulation by the channel encoder, the length is obtained as follows: Complex symbol vector The channel coding rate output by the channel encoder ; Through channel capacity Channel transmission; At the receiving end, a noisy signal is received. The channel decoder decodes it into semantic representations. Then the receiver rate matcher reconstructs it as Finally, a semantic decoder based on a DNN decodes it into... .

[0010] Optionally, the steps for constructing the sender rate matcher include: Calculation of dynamic maximum code length constraint for semantic encoder output based on instantaneous channel capacity; Semantic slice importance quantification and transmission priority ranking based on gradient analysis; Selective pruning and mask generation of semantic slices based on transmission priority and maximum code length constraints.

[0011] Optionally, the calculation of the dynamic maximum code length constraint of the semantic encoder output based on the instantaneous channel capacity includes: The information transmission rate of the channel is defined as follows:

[0012] According to Shannon's second theorem, the information transmission rate of a channel must satisfy the following constraint:

[0013]

[0014] Based on real-time signal-to-noise ratio and channel bandwidth Estimate channel capacity That is, ; Average code length per word after further pruning by the rate matcher The maximum value can be determined as ,in This represents the floor function.

[0015] Optionally, semantic slice importance quantification and transmission priority ranking based on gradient analysis include: Assume the semantic decoder output is (Before softmax), among which Indicates the vocabulary size, and calculates the output vector at the l-th position of the semantic decoder. Compared to The e-th semantic slice The gradient, i.e. ; The semantic importance weight of the e-th semantic slice to the l-th output is obtained by taking the global average of the gradient. :

[0016] in In the middle The activation value in the l-th row and i-th column. express The activation value in the j-th row and e-th column; Similarly, the importance weight vector of the e-th semantic slice with respect to each position in the semantic decoder output is derived. :

[0017] The semantic importance weight of the e-th semantic slice to the output is obtained by weighted summation of the importance weights of the L positions. :

[0018] in Let represent the weight of the i-th position, and have ; Accordingly, the semantic importance weight vector obtained for all semantic slices is: :

[0019] right The transmission priority vectors of each semantic slice are obtained by sorting them in descending order. ,in The transmission priority of the i-th semantic slice is 1. Bit.

[0020] Optionally, semantic slice selective pruning and mask generation based on transmission priority and maximum code length constraints includes: Based on the maximum code length and transmission priority vector Determine the binary mask vector For each ,have:

[0021] Therefore, the rate matching pruning operation at the sending end is represented as:

[0022] in This represents the dot product operation, and only if... hour, The i-th semantic slice is selected to form Part of, namely yes Medium priority ranking A set of semantic slices of bits Medium priority ranking Semantic slices will be discarded; mask vector The transmission decision basis of the sending end is synchronously sent to the receiving end, where it is assumed that... Transmitted in an error-free manner, and without considering errors during transmission scheduling. Required communication resources.

[0023] Optionally, the steps for establishing receiver rate matching include: Initialize a matrix of all zeros As a basic template for reconstruction; Based on the binary mask vector shared by the sender ,Will The actual amount received in China Semantic slices are sequentially filled into The middle corresponds to The index position is 1 in the middle, while for If the index position is 0, then keep it. The corresponding column is the zero column vector, and this process can be described as follows:

[0024] Where k is The sequence numbers that are filled into the semantic slices in sequence; Output , where the output It maintains the input dimensions required by the semantic decoder and restores the structure of the semantic representation at the sending end to the greatest extent, providing a foundation for subsequent semantic recovery.

[0025] The technical solution provided by the embodiments of the present invention brings at least the following beneficial effects: it can effectively identify and retain semantic slices containing more important semantic information, thereby reducing semantic loss caused in the rate matching process. At the same time, the solution does not require retraining of the original semantic communication system and can be easily embedded into the existing semantic communication system.

[0026] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0027] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a rate matching pruning method based on semantic importance provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the framework of the rate matching semantic communication system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the transmitter rate matcher provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of a receiver rate matcher provided in an embodiment of the present invention; Figure 5 This is a comparison chart of BLEU scores for different pruning methods under AWGN channels provided in this embodiment of the invention. Detailed Implementation

[0028] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0029] To address the semantic loss caused by the mismatch between channel capacity and semantic rate in semantic communication systems under dynamic channel conditions, this invention provides a rate matching pruning method based on semantic importance. This method achieves rate matching while taking into account the intrinsic value of semantic information, thereby greatly improving the robustness and transmission efficiency of the semantic communication system.

[0030] Figure 1 This is a flowchart illustrating a rate matching pruning method based on semantic importance provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps: S1, At the transmitting end, the semantic features of the input signal are extracted and compressed by a semantic encoder composed of DNN to generate a high-dimensional semantic representation. S2, The transmitter rate matcher dynamically adjusts the output data volume according to the importance of semantic information to achieve the matching of semantic rate and physical channel capacity; S3, the processed data is then encoded and modulated by the channel encoder, converted into a channel signal, and sent into the channel for transmission; S4, responding to the receiving end to obtain the received signal, and sequentially performing channel decoding, rate matching and semantic decoding on the received signal to reconstruct the original information.

[0031] In this embodiment of the invention, the method is achieved through... Figure 2 The proposed rate-matching semantic communication system is implemented as shown. Compared with existing systems, this system adds semantic-channel rate matchers between the semantic encoder and channel encoder at the transmitting end, and between the channel decoder and semantic decoder at the receiving end. At the transmitting end, the proposed rate matcher minimizes semantic loss by compressing semantic rates with more important symbols in the semantic encoder output. At the receiving end, the proposed rate matcher reconstructs the censored semantic representation through zero-padding.

[0032] Specifically, the system sequentially includes a semantic encoder, a transmitter rate matcher, and a channel encoder at the transmitting end, and sequentially includes a channel decoder, a receiver rate matcher, and a semantic decoder at the receiving end, thus forming a complete rate-matched semantic communication link. Compared with typical communication systems, this system has significant structural differences and innovations. At the transmitting end, the semantic encoder is composed of a DNN, and its function is not traditional formal source coding, but rather it extracts and compresses semantic features from the input signal to generate a high-dimensional semantic representation. Subsequently, the semantic representation is input into the core module of this invention—the transmitter rate matcher. This module dynamically adjusts the output data volume according to the importance of semantic information. Its function is analogous to the rate adaptation unit in a traditional system, but the decision-making basis changes from bit-level error correction capability to semantic-level importance measurement, thereby achieving the matching of semantic rate and physical channel capacity. The processed data is then encoded and modulated by the channel encoder, converted into a channel signal, and sent into the channel for transmission. At the receiving end, the system first performs channel decoding on the received signal to recover the noisy coded signal. The signal then enters the rate matcher at the receiver, whose function corresponds to that at the transmitter. It performs the inverse operation to reconstruct the complete semantic representation sequence for subsequent semantic decoding. Finally, the DNN-based semantic decoder reconstructs the original information based on the recovered semantic representation.

[0033] In terms of parameters, it would be: assuming the information source requires... Transmission length within s is The text sequence, ,in Indicates the transmission sequence number Embedding vectors of words, This represents the dimension of the embedding vector. At the sending end, The output after passing through the DNN-based semantic encoder is ,in Represents the hidden state dimension. This is the input to the transmitter rate matcher. Each column in the matrix is ​​denoted as a semantic slice, serving as the smallest unit for pruning. The sender rate matcher selects the slice based on semantic importance. In Selective deletion is performed on each semantic slice, retaining only the most important ones. Each slice has the following output: Where N represents the number of semantic slices retained after deletion, i.e., the average code length per word. The semantic encoder output rate (hereinafter referred to as semantic rate) is defined as... .because It is variable, so the semantic rate can be dynamically changed to adapt to channel capacity. The semantic compression rate is defined as... This indicates the degree of compression in semantic coding. The output of the sender rate matcher. After encoding and modulation by the channel encoder, the length is obtained as follows: Complex symbol vector The rate output by the channel encoder (hereinafter referred to as the channel coding rate). . Through channel capacity The channel transmission. At the receiving end, a noisy signal is received. The channel decoder decodes it into semantic representations. Then the receiver rate matcher reconstructs it as Finally, a semantic decoder based on a DNN decodes it into... .

[0034] In one embodiment of the present invention, the construction of the transmitting-end rate matcher mainly includes the following steps, the schematic diagram of which is shown below. Figure 3 : Step 1: Calculate the dynamic maximum code length constraint of the semantic encoder output based on the instantaneous channel capacity.

[0035] Shannon's second theorem states that, given the bandwidth and signal-to-noise ratio, the channel capacity has a theoretical maximum value. If the information transmission rate Exceeding this capacity will make information transmission unreliable and increase the probability of errors, meaning it needs to meet certain requirements. .

[0036] In this embodiment of the invention, the information transmission rate is defined as...

[0037] Therefore, the following restrictions must be met.

[0038]

[0039] Based on real-time signal-to-noise ratio and channel bandwidth Channel capacity It can be estimated that, i.e., there is .

[0040] In one embodiment of the present invention, for an AWGN channel, there is Channel coding rate Determined by the channel coding scheme, it can also be considered a constant.

[0041] Therefore, the average code length per word after further pruning by the transmitter rate matcher The maximum value can be determined as ,in This represents the floor function.

[0042] Step 2: Semantic slice importance quantification and transmission priority ranking based on gradient analysis.

[0043] This step jointly considers semantic information and channel conditions to evaluate the input of the transmitter rate matcher. middle The semantic importance of each semantic slice is used to determine the transmission priority vector, where semantic slices with higher importance correspond to higher priority. Inspired by sensitivity analysis based on partial derivatives, the first-order partial derivative of the output with respect to the target variable can reflect the sensitivity of the output to that variable. The larger the first-order partial derivative, the more significant the impact of that variable on the model output. Therefore, that variable is assigned higher importance. Based on this, this invention proposes a gradient-based semantic importance evaluation algorithm. The core idea is to calculate the semantic importance of the semantic decoder output relative to the target variable. The first-order partial derivative of each semantic slice is used to quantify the importance of each semantic slice.

[0044] Assuming the semantic decoder output is (Before softmax), among which This represents the vocabulary size. First, the output vector at the l-th position of the semantic decoder is calculated. Compared to The e-th semantic slice The gradient, i.e. Next, the semantic importance weight of the e-th semantic slice to the l-th output is obtained by taking the global average of the gradients. :

[0045] in In the middle The activation value in the l-th row and i-th column. express The activation value in the j-th row and e-th column of the semantic slice. Similarly, the importance weight vector of the e-th semantic slice with respect to each position in the semantic decoder output can be obtained. :

[0046] Furthermore, the semantic importance weight of the e-th semantic slice to the output can be obtained by weighted summing of the importance weights of the L positions. :

[0047] in Let represent the weight of the i-th position, and have .

[0048] Accordingly, the semantic importance weight vector obtained for all semantic slices is: :

[0049] right The transmission priority vectors of each semantic slice are obtained by sorting them in descending order. ,in The transmission priority of the i-th semantic slice is 1. Bit.

[0050] Step 3: Selective pruning and mask generation of semantic slices based on transmission priority and maximum code length constraints.

[0051] It should be noted that determining the maximum code length and transmission priority vector The binary mask vector can then be determined. For each ,have:

[0052] Therefore, the rate matching pruning operation at the sending end can be expressed as:

[0053] in The dot product operation is defined if and only if hour, The i-th semantic slice will be selected to form a composition. Part of it. In other words, yes Medium priority ranking A set of semantic slices of bits Medium priority ranking Semantic slices will be discarded.

[0054] Specifically, the mask vector The transmission decision criteria used by the sending end are synchronously sent to the receiving end to ensure that the receiving end can accurately identify the location of the discarded semantic slices, thereby reconstructing the semantic representation more accurately. In this embodiment of the invention, it is assumed that... Transmitted in an error-free manner, and without considering during transmission scheduling. Required communication resources, because Compared to It requires fewer bits, so lossless transmission can be easily achieved.

[0055] In one embodiment of the present invention, the construction of the receiver rate matcher mainly includes the following steps, the schematic diagram of which is shown below. Figure 4 : At the receiving end, because the semantic decoder requires a fixed-size input (i.e., a... (The matrix), the receiver rate matcher needs to perform a corrupted semantic representation on the received data. Perform structural restoration and rebuild it into a complete structure. 3D semantic representation The reconstruction process is achieved through zero-filling: First, initialize a matrix of all zeros. This serves as the basic template for reconstruction. Subsequently, based on the binary mask vector shared by the sending end... (Where "1" indicates that the semantic slice at that position is preserved, and "0" indicates that it is deleted), will The actual amount received in China Semantic slices are sequentially filled into The middle corresponds to The index position is 1 in the middle, while for If the index position is 0, then keep it. The corresponding column is the zero column vector. This process can be described as:

[0056] Where k is The semantic slices are filled with sequential numbers. Finally, the output... It maintains the input dimensions required by the semantic decoder and restores the structure of the semantic representation at the sending end to the greatest extent, providing a foundation for subsequent semantic recovery.

[0057] The key features of this invention are as follows: On the one hand, by adding a semantic importance-aware rate matcher between the semantic encoder and channel encoder at the transmitting end, a gradient-based semantic importance evaluation algorithm is used to calculate the transmission priority of each semantic slice. This allows for the prioritization of pruning semantic slices with lower importance at the transmitting end, minimizing semantic loss while compressing the semantic rate. On the other hand, to reduce the impact of transmitting end processing on the semantic decoder, a rate matcher is added between the channel decoder and semantic decoder at the receiving end. This process involves zero-padding the pruned semantic slices in sequence to reconstruct the fixed-shape semantic representation. This invention can be seamlessly embedded into any semantic communication system to solve the rate mismatch problem between the semantic encoder and channel encoder inputs, while ensuring minimal semantic loss during rate matching.

[0058] To demonstrate the technical effectiveness of the embodiments of the present invention, an example is given below, which demonstrates how the rate matcher proposed in this invention can be embedded and perform effectively on the basis of an existing semantic communication system, and verifies its effectiveness under dynamic channel conditions.

[0059] This example uses the Europarl dataset of European Parliament meeting minutes for training and testing. Specifically, the English portion is selected, which contains approximately 2 million sentences and 53 million words. After preprocessing, sentences with a length between 4 and 30 words are selected, and the training and test sets are divided in a 9:1 ratio.

[0060] First, the existing semantic communication system is trained using a training set to obtain the structure and parameters of the trained semantic communication system. In this system, the semantic codec is built based on the pre-trained language model BART, where the BART encoder serves as the semantic encoder and the BART decoder serves as the semantic decoder. Channel encoding / decoding employs 16-QAM modulation / demodulation, and the channel model is additive white Gaussian noise (AWGN). Furthermore, to meet the requirements of digital modulation, the output of the semantic encoder is quantized into 16 levels using a scalar quantizer with a quantization order of 16. The channel coding rate... When set to 1, the quantizer output passes directly through the 16-QAM modulator. During training, Adam is used as the optimizer, with a learning rate of 5e-5, weight decay of 1e-8, and batch size of 64. It should be noted that the rate matcher module proposed in this invention is not included in the training process; this module is only added to the system during the testing phase to verify its ability to be embedded into existing systems without retraining.

[0061] Secondly, during the testing phase, the proposed transmitter rate matcher was directly added between the transmitter's BART-based semantic encoder and the 16-QAM modulator, and the receiver rate matcher was added between the receiver's 16-QAM demodulator and the BART-based semantic decoder. The overall process is described below: Assuming the source requires... Transmission length within s is text sequence . The output is generated after passing through a semantic encoder built from BART's encoder. ,in . This is the input to the rate matcher. Each column in the table is denoted as a semantic slice, serving as the smallest unit of pruning. The output of the rate matcher is... Where N represents the number of semantic slices retained after deletion, i.e., the average code length per word. The semantic rate is then... .because It is variable, so the semantic rate can be dynamically changed to adapt to channel capacity. The output of the rate matcher. After 16QAM modulation, the length is obtained as follows: Complex symbol vector Channel coding rate . Through channel capacity AWGN channel transmission, channel bandwidth Signal-to-noise ratio At the receiving end, a noisy signal is received. The channel decoder decodes it into semantic representations. Then the rate matcher reconstructs it as Finally, the semantic decoder constructed by the BART decoder decodes it into... .

[0062] In this example, the construction of the sender rate matcher mainly includes the following steps: Step 1: Calculate the dynamic maximum code length constraint of the semantic encoder output based on the instantaneous channel capacity.

[0063] Given bandwidth and signal-to-noise ratio In AWGN channels, channel capacity From equation (3), the maximum code length allowed for transmission given the channel capacity is... ,Will , Substitute and you will get The calculation results are shown in Table 1.

[0064] Table 1. Numerical relationships between SNR, C, and N under given parameters.

[0065] For example, when the instantaneous channel signal-to-noise ratio At that time, channel capacity The maximum code length allowed to be transmitted at this time .

[0066] Step 2: Semantic slice importance quantification and transmission priority ranking based on gradient analysis.

[0067] This step jointly considers semantic information and channel conditions to evaluate the input of the rate matcher. middle The semantic importance of each semantic slice is used to determine the transmission priority vector, where the semantic slice with higher importance corresponds to higher priority.

[0068] Assuming the semantic decoder output is (Before softmax), among which This represents the vocabulary size. First, the output vector at position 1 of the semantic decoder is calculated. Compared to The first semantic slice The gradient, i.e. Next, the semantic importance weight of the first semantic slice to the first output is obtained by taking the global average of the gradient. :

[0069] in In the middle The activation value in the first row and i-th column. express The activation value of the j-th row and 1-th column. Similarly, the activation value of the first semantic slice with respect to the semantic decoder output can be obtained. Importance weight vector :

[0070] Furthermore, the semantic importance weight of the first semantic slice to the output can be obtained by weighted summation of the importance weights of the 14 positions. :

[0071] in Let represent the weight of the i-th position, and have Accordingly, the semantic importance weight vector obtained for all semantic slices is: :

[0072] right The transmission priority vectors of each semantic slice are obtained by sorting them in descending order. ,in The transmission priority of the i-th semantic slice is 1. For example, the calculation result is: This indicates that the transmission priority of the first semantic slice is bit 445, and the transmission priority of the 768th semantic slice is bit 51.

[0073] Step 3: Selective pruning and mask generation of semantic slices based on transmission priority and maximum code length constraints.

[0074] In real-time signal-to-noise ratio At that time, channel capacity The maximum code length allowed to be transmitted at this time According to the transmission priority vector The binary mask vector can then be determined. For each ,have:

[0075] For example, for and , The rate-matching pruning operation at the sending end can be represented as:

[0076] in The dot product operation is defined if and only if hour, The i-th semantic slice will be selected to form a composition. Part of it. In other words, yes Medium priority ranking A set of semantic slices of bits Medium priority ranking Semantic slices will be discarded.

[0077] Specifically, the mask vector The transmission decisions made by the sender are synchronously sent to the receiver to ensure that the receiver can accurately identify the location of discarded semantic slices and thus more accurately reconstruct the semantic representation. In this paper, we assume... Transmitted in an error-free manner, and without considering during transmission scheduling. Required communication resources, because Compared to It requires fewer bits, so lossless transmission can be easily achieved.

[0078] The construction of the receiver rate matcher mainly includes the following steps: zero-padding reconstruction based on the semantic representation of the received mask vector: First, initialize a matrix of all zeros. This serves as the basic template for reconstruction. Subsequently, based on the binary mask vector shared by the sending end... (Where "1" indicates that the semantic slice at that position is preserved, and "0" indicates that it is deleted), will The actual amount received in China Semantic slices are sequentially filled into The middle corresponds to The index position is 1 in the middle, while for If the index position is 0, then keep it. The corresponding column is a zero column vector. For example, for , The first and second columns are zero column vectors, and the third column is composed of... The first semantic slice is filled. This process can be described as: (10) Where k is The semantic slices are filled with sequential numbers. Finally, the output... It maintains the input dimensions required by the semantic decoder and restores the structure of the semantic representation at the sending end to the greatest extent, providing a foundation for subsequent semantic recovery.

[0079] To evaluate the effectiveness of the method proposed in this invention, the BLEU score is used as an evaluation metric to measure the effectiveness of the decoded text. With real text The degree of matching between them; the higher the score, the smaller the semantic loss. Figure 5The paper presents a comparison of BLEU scores for different pruning methods in an AWGN channel. The "All" curve (blue) represents the ideal case where no rate matcher is added and all semantic slices participate in transmission; the "Random" curve (green) represents the rate matching method using a random pruning strategy; and the "SI" curve (orange) represents the rate matching pruning method based on semantic importance proposed in this invention. It can be seen that, under the same signal-to-noise ratio (SNR), the BLEU score of the proposed method (SI) is significantly higher than that of the random pruning method (Random), proving that semantic importance-based pruning can effectively preserve key information and minimize semantic loss. When the SNR reaches 9 dB, the performance of the proposed method (SI) is comparable to the ideal case (All). This indicates that at this SNR, the pruned semantic slices are mainly redundant information, and the rate matcher successfully matches the semantic rate to the channel capacity with almost no additional semantic loss. The above results fully verify that the rate matching culling method based on semantic importance proposed in this invention can effectively compress semantic rate in a "plug-and-play" manner under dynamic channel conditions, while ensuring high semantic fidelity, thus improving the robustness and practicality of the semantic communication system.

[0080] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0081] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A rate matching pruning method based on semantic importance, characterized in that, Includes the following steps: At the transmitting end, a semantic encoder composed of DNN extracts and compresses semantic features from the input signal to generate a high-dimensional semantic representation. The transmitter rate matcher dynamically adjusts the output data volume according to the importance of semantic information to achieve a match between semantic rate and physical channel capacity; The processed data is then encoded and modulated by a channel encoder, converted into a channel signal, and sent into the channel for transmission. Upon receiving the received signal, the receiving end sequentially performs channel decoding, rate matching, and semantic decoding on the received signal to reconstruct the original information. The transmitting end includes a semantic encoder, a transmitting end rate matcher, and a channel encoder; the receiving end includes a channel decoder, a receiving end rate matcher, and a semantic decoder, wherein: The channel decoder corresponds to the channel encoder in function and is responsible for performing decoding operations; The receiver rate matcher functions as the transmitter rate matcher and is responsible for performing the reverse reconstruction operation. The semantic decoder corresponds to the semantic encoder and is responsible for performing the decoding operation; Also includes: Assuming the information source requires in Transmission length within s is The text sequence, ,in Indicates the transmission sequence number Embedding vectors of words, Indicates the dimension of the embedded vector; At the sending end, The output after passing through the DNN-based semantic encoder is ,in Represents the dimension of the hidden state. This is the input to the transmitter rate matcher; In the transmitter rate matcher, Each column in the code is denoted as a semantic slice, serving as the smallest unit for deletion. The slices are then processed according to their semantic importance. In Selective deletion is performed on each semantic slice, retaining only the most important ones. Each slice has the following output: , where N represents the number of semantic slices retained after deletion, i.e. the average code length per word; The semantic rate output by the semantic encoder is defined as... ,because If it is variable, then the semantic rate can be dynamically changed to adapt to the channel capacity. The semantic compression rate is defined as... , to indicate the degree of compression of semantic encoding; Output of the transmitter rate matcher After encoding and modulation by the channel encoder, the length is obtained as follows: Complex symbol vector The channel coding rate output by the channel encoder ; Through channel capacity Channel transmission; At the receiving end, a noisy signal is received. The channel decoder decodes it into semantic representations. Then the receiver rate matcher reconstructs it as Finally, a semantic decoder based on a DNN decodes it into... ; The steps for constructing the transmitter rate matcher include: Calculation of dynamic maximum code length constraint for semantic encoder output based on instantaneous channel capacity; Semantic slice importance quantification and transmission priority ranking based on gradient analysis; Selective pruning and mask generation of semantic slices based on transmission priority and maximum code length constraints.

2. The method according to claim 1, characterized in that, The calculation of the dynamic maximum code length constraint for the semantic encoder output based on instantaneous channel capacity includes: The information transmission rate of the channel is defined as follows: According to Shannon's second theorem, the information transmission rate of a channel must satisfy the following constraint: Based on real-time signal-to-noise ratio and channel bandwidth Estimate channel capacity That is, ; Average code length per word after further pruning by the rate matcher The maximum value can be determined as ,in This represents the floor function.

3. The method according to claim 2, characterized in that, Semantic slice importance quantification and transmission priority ranking based on gradient analysis include: Assuming the semantic decoder output is (Before softmax), among which Indicates the vocabulary size, and calculates the output vector at the l-th position of the semantic decoder. Compared to The e-th semantic slice The gradient, i.e. ; The semantic importance weight of the e-th semantic slice to the l-th output is obtained by taking the global average of the gradient. : in In the middle The activation value in the l-th row and i-th column. express The activation value in the j-th row and e-th column; Similarly, the importance weight vector of the e-th semantic slice with respect to each position in the semantic decoder output is derived. : The semantic importance weight of the e-th semantic slice to the output is obtained by weighted summation of the importance weights of the L positions. : in Let represent the weight of the i-th position, and have ; Accordingly, the semantic importance weight vector obtained for all semantic slices is: : right The transmission priority vectors of each semantic slice are obtained by sorting them in descending order. ,in The transmission priority of the i-th semantic slice is 1. Bit.

4. The method according to claim 3, characterized in that, Selective pruning and mask generation of semantic slices based on transmission priority and maximum code length constraints, including: Based on the maximum code length and transmission priority vector Determine the binary mask vector For each ,have: Therefore, the rate matching pruning operation at the sending end is represented as: in This represents the dot product operation, and only if... hour, The i-th semantic slice is selected to form Part of, namely yes Medium priority ranking A set of semantic slices of bits Medium priority ranking Semantic slices will be discarded; mask vector The transmission decision basis of the sending end is synchronously sent to the receiving end, where it is assumed that... Transmitted in an error-free manner, and without considering errors during transmission scheduling. Required communication resources.

5. The method according to claim 4, characterized in that, The steps for establishing receiver rate matching include: Initialize a matrix of all zeros As a basic template for reconstruction; Based on the binary mask vector shared by the sender ,Will The actual amount received in China Semantic slices are sequentially filled into The middle corresponds to The index position is 1 in the middle, while for If the index position is 0, then keep it. The corresponding column is the zero column vector, and this process can be described as follows: Where k is The sequence numbers that are filled into the semantic slices in sequence; Output , where the output It maintains the input dimensions required by the semantic decoder and restores the structure of the semantic representation at the sending end to the greatest extent, providing a foundation for subsequent semantic recovery.

Images