Hierarchical semantic communication method and system based on dependency syntax analysis

Abstract

The present application relates to a hierarchical semantic communication method and system based on dependency syntactic analysis in the field of intelligent agent semantic communication technology. The method introduces dependency syntactic analysis into semantic communication, uses dependency syntactic analysis at the sending end to semantically stratify the words in the transmitted sentence, selectively transmits words of different semantic levels in the transmitted sentence according to the channel quality, receives the information sent by the sending end, performs channel decoding and source decoding on the received information in turn, and obtains the received sentence; and uses the ERNIE model to semantically restore the received sentence. The present invention introduces dependency syntactic analysis into semantic communication at the sending end to overcome the problems of unclear semantic characterization and overly abstract semantic representation in existing semantic communication solutions. After receiving the information, the receiving end uses the ERNIE language model to perform semantic recovery. Compared with traditional communication methods, the present invention has significantly improved transmission reliability.

Description

Technical Field

[0001] The present application relates to the technical field of intelligent agent semantic communication, and in particular to a hierarchical semantic communication method and system based on dependency syntactic analysis. Background Art

[0002] Traditional communication systems don't consider the underlying meaning of the content being transmitted; their primary purpose is to accurately transmit communication symbols. Semantic communication focuses on different issues than traditional communication. Rather than focusing on the symbols themselves, it prioritizes the meaning they convey. Therefore, semantic communication systems are closer to the essence of communication, focusing more on the transmission of semantic meaning, and offer a deeper level of communication.

[0003] Traditional communication systems require error-free transmission at the bit level, meaning that the sent and received content must be identical. When communication conditions are poor, even after channel decoding and error correction, erroneous codewords or bits may still be present in the received information, leading to source decoding failure. In such cases, traditional communication systems employ retransmission technology, but this approach is time-consuming and consumes significant transmission resources. Semantic communication systems, on the other hand, strive for semantically indiscriminate transmission. Therefore, the sending end can extract and transmit the semantic information of the transmitted content, which can then be recovered at the receiving end after receiving the semantic information. Summary of the Invention

[0004] Based on this, it is necessary to provide a hierarchical semantic communication method and system based on dependency syntactic analysis to address the above technical problems, so as to overcome the problems of unclear semantic characterization and overly abstract semantic representation in existing semantic communication schemes.

[0005] A hierarchical semantic communication method based on dependency syntactic analysis, the method comprising:

[0006] Sending end:

[0007] Get the sentence to be transmitted, and use the Jieba model to perform Chinese word segmentation on the sentence to be transmitted.

[0008] Perform dependency syntactic analysis on the sentence to be transmitted after word segmentation to obtain a dependency syntactic tree corresponding to the sentence to be transmitted.

[0009] The sentence to be transmitted is semantically layered according to the dependency syntax tree, and the words in the sentence to be transmitted are divided into different semantic levels.

[0010] The transmission words are determined based on the principle of selectively transmitting words of different semantic levels in the sentence to be transmitted according to the channel quality.

[0011] Source coding is performed on the transmission words, the coding result is converted into bit information, channel coding is performed on the bit information, and the obtained coding result is sent to the receiving end.

[0012] Receiver:

[0013] Receive information sent by the transmitting end to obtain received information, and perform channel decoding and source decoding on the received information in sequence to obtain a received statement.

[0014] The ERNIE model is used to perform semantic recovery on the received sentence.

[0015] In one embodiment, dependency syntactic analysis is performed on the sentence to be transmitted after word segmentation to obtain a dependency syntactic tree corresponding to the sentence to be transmitted, including:

[0016] A training data set is obtained, where the training data set includes Chinese words in the segmented transmission sentences.

[0017] The word vectors of each Chinese word in the training data set are calculated using the Glove network and the Char_CNN network, and the two word vectors obtained are spliced.

[0018] The concatenated word vectors are used to calculate the hidden layer feature vectors corresponding to the words using the BiLSTM network.

[0019] The hidden layer feature vectors corresponding to the two words are input into the dependency edge recognition module to calculate the embedding representation of the dependency edge between each two words; and based on the embedding representation of each dependency edge in the set of dependency edges between words in the training dataset, the dependency edges are scored using the dependency edge scoring function, and the probability of the existence of a dependency edge between words is determined based on the obtained dependency edge score; and the correct dependency edge between words is determined based on the probability of the existence of a dependency edge between words.

[0020] The multi-label classification method is used to mark each correct dependency edge, and the marked correct dependency edges are classified using the dependency edge classification module to obtain the dependency syntax tree prediction result.

[0021] A total loss function of a dependency parsing model is constructed, wherein the total loss function includes a dependency edge recognition loss function and a dependency edge classification loss function; the dependency parsing model is a network model composed of a Glove network, a Char_CNN network, a BiLSTM network, a dependency edge recognition module, and a dependency edge classification module.

[0022] According to the dependency syntax tree prediction result, the total loss function and the training data set, the dependency syntax analysis model is trained using the gradient descent method to obtain a trained dependency syntax analysis model.

[0023] The sentence to be transmitted after word segmentation is input into the trained dependency syntax analysis model to obtain a dependency syntax tree corresponding to the sentence to be transmitted.

[0024] In one embodiment, hidden layer feature vectors corresponding to two words are input into a dependency edge recognition module to calculate an embedding representation of the dependency edge between each two words; and based on the embedding representation of each dependency edge in the set of dependency edges between words in the training data set, the dependency edges are scored using a dependency edge scoring function; based on the obtained dependency edge scores, the probability of the existence of a dependency edge between the words is determined, and based on the probability of the dependency edge, the correct dependency edge between the words is determined, including:

[0025] The embedded representation of the dependency edge between the two words is calculated using the hidden layer vectors of the two words. The expression of the embedded representation is:

[0026]

[0027] in is the embedding representation of the dependency edge, i ,w j > is the i-th word w i and the jth word w j The dependency edges between are the i-th word w i and the jth word w j The corresponding hidden layer feature vector.

[0028] According to the embedded representation of each dependency edge in the set of dependency edges between words in the training dataset, the dependency edge is scored using the dependency edge scoring function, which is:

[0029]

[0030] where f edge (w i ,w j ) is the i-th word w i and the jth word w j The score of the dependency edge between them, ε is the set of all dependency edges in the transmitted sentence.

[0031] According to the obtained dependency edge scores, the probability of the existence of dependency edges between words is determined, and the correct dependency edges between words are determined according to the probability of dependency edges. The expression of the probability of dependency edges is:

[0032]

[0033] Among them, p(w j ∣w i ) is the i-th word w i and the jth word wj The probability of the dependency edge between them, n is the total number of words in the transmitted sentence.

[0034] In one embodiment, a multi-label classification method is used to label each correct dependency edge, and the labeled correct dependency edges are classified using a dependency edge classification module to obtain a dependency syntax tree prediction result, including:

[0035] For each correct dependency edge, the dependency relationship of the dependency edge is introduced, and the score of whether the dependency relationship of the dependency edge is correct is calculated based on the embedding representation of the correct dependency edge. The scoring function for evaluating whether the dependency relationship is correct is:

[0036]

[0037] Where f label (·) is a scoring function to evaluate whether the dependency relationship is correct, and ε(r) is the set of dependency edges in ε with relationship r.

[0038] According to the score of whether the dependency relationship of each dependency edge is correct, the probability of determining that the dependency relationship of the dependency edge is r is:

[0039]

[0040] Among them, p(r|w i ,w j ) is the i-th word w i and the jth word w j The probability that the dependency relationship between the dependency edges is r, is the set of all dependency types, r′ is Any dependency relationship in .

[0041] The dependency edges are classified according to the probability that their dependency relationship is r, and the dependency syntax tree prediction result is obtained.

[0042] In one embodiment, the total loss function includes a dependency edge identification loss function and a dependency edge classification loss function.

[0043] Construct the total loss function of the dependency parsing model, including:

[0044] Construct a dependency edge recognition loss function and a dependency edge classification loss function, and use the sum of the dependency edge recognition loss function and the dependency edge classification loss function as the total loss function of the dependency parsing model; the dependency edge recognition loss function is:

[0045]

[0046] Where, is the dependency edge identification loss, is the dataset used for network training, represents the i-th word in the l-th sentence in the dataset, represent The governing word of .

[0047] The dependency edge classification loss function is:

[0048]

[0049] in, is the dependency edge classification loss, yes and The dependencies between them.

[0050] In one embodiment, semantic stratification is performed on the sentence to be transmitted according to the dependency syntax tree, and words in the sentence to be transmitted are divided into different semantic levels, including:

[0051] According to the dependency syntax tree, the sentence to be transmitted is divided into four different semantic levels; wherein, the first semantic level includes the subject, predicate and object of the sentence to be transmitted; the second semantic level includes all words that depend on the subject, predicate and object of the sentence to be transmitted; the third semantic level includes relational modifiers; and the fourth semantic level includes structural auxiliary words.

[0052] In one embodiment, the principle of selectively transmitting words at different semantic levels in a sentence to be transmitted based on channel quality is used to determine the words to be transmitted, including:

[0053] When the channel quality is poor, only words of the first semantic level are transmitted, and the communication resources of the untransmitted words are used to protect the transmitted words, ensuring that the important semantics conveyed by the transmitted words are not affected by channel noise.

[0054] As the channel quality gradually improves, words at other semantic levels are transmitted.

[0055] In one embodiment, source encoding is performed on the transmission word, the encoded result is converted into bit information, the bit information is channel encoded, and the obtained encoded result is sent to the receiving end, including:

[0056] The transmission words are source-coded using Huffman coding, and the coded result is converted into bit information.

[0057] The bit information is channel-coded using LDPC coding, and the obtained coding result is sent to a receiving end.

[0058] In one embodiment, the semantic recovery of the received sentence using the ERNIE model includes:

[0059] The missing words in the received sentence are filled with [MASK] and input into the ERNIE model. Based on the context information in the received sentence, the [MASK] in the received sentence is filled with words that are appropriate in the context to obtain predicted words. The predicted words filled in are:

[0060]

[0061] in, To predict words, To predict the conditional probability corresponding to the context information of the word, is the lth predicted word; Emb(context) is the embedding representation of the entire predicted word set context.

[0062] A layered semantic communication system based on dependency syntactic analysis, the system comprising a sending end and a receiving end; the sending end and the receiving end communicate by executing any of the above-mentioned layered semantic communications based on dependency syntactic analysis.

[0063] The above-mentioned layered semantic communication method and system based on dependency syntactic analysis introduces dependency syntactic analysis into semantic communication. At the transmitting end, dependency syntactic analysis is used to semantically stratify the words in the transmitted sentence. Words at different semantic levels in the transmitted sentence are selectively transmitted based on channel quality. The receiving end receives the information sent by the transmitting end, sequentially performs channel decoding and source decoding on the received information to obtain the received sentence. The ERNIE model is then used to semantically recover the received sentence. This invention introduces dependency syntactic analysis into semantic communication at the transmitting end to overcome the problems of unclear semantic characterization and overly abstract semantic representation in existing semantic communication schemes. After receiving the information, the receiving end uses the ERNIE language model for semantic recovery. Compared to traditional communication methods, this invention significantly improves transmission reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 1 is a flow chart of a hierarchical semantic communication method based on dependency parsing in one embodiment;

[0065] Figure 2 A schematic diagram of a flow chart of the step of determining a dependency syntax tree corresponding to a sentence to be transmitted in one embodiment;

[0066] Figure 3 A block diagram of a hierarchical semantic communication principle based on dependency parsing in one embodiment;

[0067] Figure 4 The following is a simulation result of transmitting semantic similarity scores of words at different semantic levels based on BERT over a Gaussian white noise channel in another embodiment;

[0068] Figure 5 The simulation results of BLEU scores of the system of the present invention and the classical communication system on a Gaussian white noise channel in another embodiment are shown;

[0069] Figure 6 The following is a simulation result of a BERT-based semantic similarity score between a hierarchical semantic communication system based on dependency parsing and a classical communication system over a Gaussian white noise channel in another embodiment;

[0070] Figure 7 The figure shows the simulation results of BLEU scores of a hierarchical semantic communication system based on dependency parsing and a classical communication system on a Rayleigh fading channel in another embodiment;

[0071] Figure 8 The figure shows the simulation results of the semantic similarity score based on BERT between a hierarchical semantic communication system based on dependency parsing and a classical communication system over a Rayleigh fading channel in another embodiment. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0073] In one embodiment, Figure 1 As shown, a hierarchical semantic communication method based on dependency syntactic analysis is provided, which includes the following steps:

[0074] Sending end:

[0075] Step 100: Get the sentence to be transmitted and use the Jieba model to perform Chinese word segmentation on the sentence to be transmitted.

[0076] Step 102: Perform dependency syntactic analysis on the sentence to be transmitted after word segmentation to obtain a dependency syntactic tree corresponding to the sentence to be transmitted.

[0077] Specifically, dependency syntactic analysis adopts a two-stage process, that is, first predicting whether there is a dependency relationship between the words in the transmitted sentence, then predicting the specific type of dependency relationship between different words, and generating the dependency syntactic tree corresponding to the transmitted sentence.

[0078] Step 104: semantically stratify the sentence to be transmitted according to the dependency syntax tree, and divide the words in the sentence to be transmitted into different semantic levels.

[0079] Specifically, the higher the word is in the semantic hierarchy, the more important the semantics it conveys.

[0080] Step 106: Determine the transmission words based on the principle of selectively transmitting words at different semantic levels in the sentence to be transmitted according to the channel quality.

[0081] Specifically, the system selectively transmits words at different semantic levels within a sentence based on channel quality. When channel quality is poor, only the words at the top of the semantic hierarchy are transmitted. As channel quality improves, words at other semantic levels are transmitted. Furthermore, when channel quality is poor, communication resources used for untransmitted words are used to protect transmitted words, ensuring that the important semantics conveyed by the transmitted words are not affected by channel noise.

[0082] When the channel quality is very poor, instead of transmitting erroneous information and failing to express the correct semantics, it is better to use transmission resources to protect the most important semantic information so that it can be transmitted correctly.

[0083] Step 108 performs source coding on the transmission words, converts the coding result into bit information, performs channel coding on the bit information, and sends the obtained coding result to the receiving end.

[0084] Specifically, channel coding is performed to resist the influence of channel noise.

[0085] Receiver:

[0086] Step 110: Receive the information sent by the transmitter to obtain received information, perform channel decoding and source decoding on the received information in sequence to obtain a received statement.

[0087] Step 112: Use the ERNIE model to perform semantic recovery on the received sentence.

[0088] The above-mentioned layered semantic communication method based on dependency syntactic analysis introduces dependency syntactic analysis into semantic communication. At the transmitting end, dependency syntactic analysis is used to semantically stratify the words in the transmitted sentence. Words at different semantic levels in the transmitted sentence are selectively transmitted based on channel quality. The receiving end receives the information sent by the transmitting end, sequentially performs channel decoding and source decoding on the received information to obtain the received sentence. The ERNIE model is then used to semantically recover the received sentence. This method introduces dependency syntactic analysis into semantic communication at the transmitting end, overcoming the problems of unclear semantic characterization and overly abstract semantic representation in existing semantic communication schemes. After receiving the information, the receiving end uses the ERNIE language model for semantic recovery. Compared to traditional communication methods, this method significantly improves transmission reliability.

[0089] In one embodiment, the process of determining the dependency syntax tree corresponding to the sentence to be transmitted is as follows: Figure 2 As shown, step 102 specifically includes the following steps:

[0090] Step 200: Obtain a training data set, where the training data set includes Chinese words in the transmitted sentences after word segmentation.

[0091] Step 202: Calculate the word vector of each Chinese word in the training data set using the Glove network and the Char_CNN network respectively, and concatenate the two obtained word vectors.

[0092] Specifically, after obtaining the segmented transmission sentence, calculate the word vector of each Chinese word in the transmission sentence:

[0093]

[0094] Where Glove(w i ) is the word w calculated by the Glove model i The word vector of Char_CNN(w i ) is the character-level embedding vector calculated by the Char_CNN model.

[0095] Step 204: The concatenated word vectors are used to calculate the hidden layer feature vectors corresponding to the words using a BiLSTM network.

[0096] Specifically, get w i After concatenating the word vectors, e i Feed into the BiLSTM network to calculate w i The corresponding hidden layer feature vector:

[0097] v i =BiLSTM(e i )

[0098] Among them, v i For the word w i The corresponding hidden layer feature vector.

[0099] Step 206: Based on the hidden layer feature vectors corresponding to the two words, the dependency edge recognition module is input to calculate the embedding representation of the dependency edge between each two words; and based on the embedding representation of each dependency edge in the set of dependency edges between words in the training data set, the dependency edges are scored using the dependency edge scoring function, and the probability of the existence of a dependency edge between words is determined based on the obtained dependency edge score; and the correct dependency edge between words is determined based on the probability of the existence of a dependency edge between words.

[0100] Step 208: Use a multi-label classification method to mark each correct dependency edge, and classify the marked correct dependency edges using a dependency edge classification module to obtain a dependency syntax tree prediction result.

[0101] Step 210: Construct a total loss function of the dependency parsing model, which includes a dependency edge recognition loss function and a dependency edge classification loss function; the dependency parsing model is a network model composed of a Glove network, a Char_CNN network, a BiLSTM network, a dependency edge recognition module, and a dependency edge classification module.

[0102] Step 212: Based on the dependency syntax tree prediction result, the total loss function, and the training data set, the dependency syntax analysis model is trained using the gradient descent method to obtain a trained dependency syntax analysis model.

[0103] Step 214: Input the segmented sentence to be transmitted into the trained dependency syntax analysis model to obtain a dependency syntax tree corresponding to the sentence to be transmitted.

[0104] In one embodiment, step 206 includes: using the hidden layer vectors of the two words to calculate the embedding representation of the dependency edge between the two words, and the expression of the embedding representation is:

[0105]

[0106] in is the embedding representation of the dependency edge, i ,w j > is the i-th word w i and the jth word w j The dependency edges between are the i-th word w i and the jth word w j The corresponding hidden layer feature vector.

[0107] Specifically, Represents the dependency edge between two words.

[0108] According to the embedded representation of each dependency edge in the set of dependency edges between words in the training dataset, the dependency edge is scored using the dependency edge scoring function. The dependency edge scoring function is:

[0109]

[0110] where f edge (w i ,w j ) is the i-th word w i and the jth word w j The score of the dependency edge between them, ε is the set of all dependency edges in the transmitted sentence.

[0111] According to the obtained dependency edge score, the probability of the existence of dependency edges between words is determined, and the correct dependency edges between words are determined according to the probability of dependency edges. The expression of the probability of dependency edges is:

[0112]

[0113] Among them, p(w j ∣w i ) is the i-th word w i and the jth word w j The probability of the dependency edge between them, n is the total number of words in the transmitted sentence.

[0114] In one embodiment, step 208 includes: introducing a dependency relation of each correct dependency edge, and calculating a score of whether the dependency relationship of the dependency edge is correct based on the embedded representation of the correct dependency edge. The scoring function for evaluating whether the dependency relationship is correct is:

[0115]

[0116] Where f label (·) is a scoring function to evaluate whether the dependency relationship is correct, and ε(r) is the set of dependency edges in ε with relationship r.

[0117] According to the score of whether the dependency relationship of each dependency edge is correct, the probability of determining that the dependency relationship of the dependency edge is r is:

[0118]

[0119] Among them, p(r|w i ,w j ) is the i-th word w i and the jth word w j The probability that the dependency relationship between the dependency edges is r, is the set of all dependency types, r′ is Any dependency relationship in .

[0120] The dependency edges are classified according to the probability that their dependency relationship is r, and the dependency syntax tree prediction result is obtained.

[0121] In one embodiment, the total loss function includes a dependency edge recognition loss function and a dependency edge classification loss function; step 210 includes: constructing the dependency edge recognition loss function and the dependency edge classification loss function, and taking the sum of the dependency edge recognition loss function and the dependency edge classification loss function as the total loss function of the dependency parsing model; the dependency edge recognition loss function is:

[0122]

[0123] Where, is the dependency edge identification loss, is the dataset used for network training, represents the i-th word in the l-th sentence in the dataset, represent The governing word of .

[0124] The dependency edge classification loss function is:

[0125]

[0126] in, is the dependency edge classification loss, r i (l) yes and The dependencies between them.

[0127] In one embodiment, step 104 includes: dividing the sentence to be transmitted into four different semantic levels according to the dependency syntax tree; wherein the first semantic level includes the subject, predicate and object of the sentence to be transmitted; the second semantic level includes all words that depend on the subject, predicate and object of the sentence to be transmitted; the third semantic level includes relational modifiers; and the fourth semantic level includes structural auxiliary words.

[0128] Specifically, semantic stratification is performed based on the dependency syntax tree of the transmitted sentence, which is divided into four different semantic levels. The first semantic level mainly includes the subject, predicate, and object of the transmitted sentence, representing the core semantics of the transmitted sentence. To construct the first semantic level, we first need to query the predicate in the transmitted sentence, and then search for the subject and object that depend on the predicate. The word set of the first semantic level can be expressed as:

[0129]

[0130] Among them, w i ←root represents all words that depend on the root node (root), w j ←w i Represents all dependencies on w i words, is the word set of the first semantic level, and s is the sentence to be transmitted.

[0131] The second semantic level includes all words that are dependent on the subject, predicate, and object of the transmission sentence. These words play a role in supplementing the core semantics in the transmission sentence. The word set of the second semantic level can be expressed as:

[0132]

[0133] in, It is a word set of the second semantic level.

[0134] Similarly, the word sets of the third and fourth semantic levels can be expressed as:

[0135]

[0136]

[0137] in, is the word set of the third semantic level, It is a word set of the fourth semantic level.

[0138] The third semantic level usually contains relational modifiers, which make the semantic relationship between the first and second semantic levels more clear, help eliminate semantic ambiguity, and facilitate semantic understanding. The fourth semantic level is mostly structural particles and does not affect the semantic transmission of the sentence.

[0139] In one embodiment, step 106 includes: when the channel quality is poor, only transmitting words of the first semantic level, and using the communication resources of the non-transmitted words to protect the transmitted words, to ensure that the important semantics conveyed by the transmitted words are not affected by channel noise; as the channel quality gradually improves, transmitting words of other semantic levels.

[0140] In one embodiment, step 108 includes: performing source coding on the transmission words using Huffman coding, and converting the coding result into bit information; performing channel coding on the bit information using LDPC coding, and sending the obtained coding result to the receiving end.

[0141] In one embodiment, step 112 includes: filling in the missing words in the received sentence with [MASK] and inputting it into the ERNIE model; filling in the [MASK] in the received sentence with a word that is appropriate in the context according to the context information in the received sentence to obtain a predicted word; the predicted word is:

[0142]

[0143] in, To predict words, To predict the conditional probability corresponding to the context information of the word, is the lth predicted word; Emb(context) is the embedding representation of the entire predicted word set context.

[0144] It should be understood that although Figure 1-Figure 2 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1-Figure 2At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0145] In one embodiment, a hierarchical semantic communication system based on dependency syntax analysis is provided. The system includes a sending end and a receiving end.

[0146] The sender and the receiver communicate by executing any of the above-mentioned hierarchical semantic communications based on dependency parsing.

[0147] In a specific embodiment, Figure 3 The system provides a hierarchical semantic communication system based on dependency syntactic analysis. Based on the traditional communication system, the system adds a dependency syntactic analysis module at the sending end and semantically hierarchizes the transmitted sentences according to their dependency syntactic tree; the ERNIE language model is used for semantic recovery at the receiving end.

[0148] Figure 2 The simulation results of the method of the present invention for transmitting semantic information of different levels under additive white Gaussian noise channel are given.

[0149] In a simulation example, four transmission strategies were selected to transmit semantic information of different levels using layered semantic communication based on dependency syntactic analysis, namely, transmitting only the semantic information of the first layer, transmitting the semantic information of the first two layers, transmitting the semantic information of the first three layers, and transmitting the semantic information of the first four layers. The simulation results of using this system to transmit semantic information of different levels under additive white Gaussian noise channel are shown in Figure 2. Figure 4 shown. Figure 4 The evaluation criteria used in the evaluation are: semantic similarity evaluation criteria based on the BERT model, which reflects the degree of semantic similarity by calculating the angle between word vectors. Figure 4 As shown, in a Gaussian white noise channel, when the signal-to-noise ratio is less than -2dB, transmitting only the first level of semantic information scores the highest; when the signal-to-noise ratio is between -1dB and 1dB, transmitting the first two levels of semantic information scores the highest; when the signal-to-noise ratio is between 2dB and 4dB, transmitting the first three levels of semantic information scores the highest; and when the signal-to-noise ratio is greater than 4dB, transmitting the first four levels of semantic information is suitable. The simulation results are consistent with the expectations of the proposed method and also verify the effectiveness of the proposed method's scheme of selectively transmitting semantic information based on its importance.

[0150] In another simulation embodiment, in the channel model, the transmitted and received signals are x and y respectively, and the received signal can be expressed as: y = hx + n, where h is the channel response and Gaussian white noise is Next, we introduce another evaluation metric besides semantic similarity: BLEU (Bilingual Evaluation Understudy), which is used to calculate the semantic similarity between sentences. BLEU (1-gram) evaluates the similarity between two sentences by comprehensively comparing the proportion of n = 1 word tuples that co-occur in the two sentences. We compare the proposed method with classic communication system architectures (Huffman source codec and LDPC channel codec with a code rate of 2 / 3) under two channel models. Figures 5 to 8 The simulation results of the hierarchical semantic communication system based on dependency syntax analysis under Gaussian white noise channel and Rayleigh fading channel model are shown in the figure. Figure 5 and Figure 6 In a Gaussian white noise channel, the BLEU (1-gram) between the sent and recovered information in this semantic communication system and the semantic similarity score based on the BERT model are higher than those of the classical communication system in the entire signal-to-noise ratio range. Figure 7 and Figure 8 The results show that in Rayleigh fading channels, the BLEU score of this method and the semantic similarity score based on the BERT model are higher than those of the classical communication system. This verifies that the hierarchical semantic communication system based on dependency syntactic analysis that adopts the method of the present invention for communication can effectively improve the reliability of the communication process.

[0151] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0152] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A hierarchical semantic communication method based on dependency syntactic analysis, characterized in that: The method comprises: Sending end: Obtain the sentence to be transmitted and use the Jieba model to perform Chinese word segmentation on the sentence to be transmitted; Perform dependency syntactic analysis on the sentence to be transmitted after word segmentation to obtain a dependency syntactic tree corresponding to the sentence to be transmitted; Semantically stratifying the sentence to be transmitted according to the dependency syntax tree, dividing the words in the sentence to be transmitted into different semantic levels; Determine the transmission words based on the principle of selectively transmitting words of different semantic levels in the sentence to be transmitted according to the channel quality; performing source coding on the transmission words, converting the coding result into bit information, performing channel coding on the bit information, and sending the obtained coding result to a receiving end; Receiver: receiving information sent by a transmitting end to obtain received information, and performing channel decoding and source decoding on the received information in sequence to obtain a received statement; The ERNIE model is used to perform semantic recovery on the received sentence.

2. The method according to claim 1, characterized in that Perform dependency syntactic analysis on the sentence to be transmitted after word segmentation to obtain the dependency syntactic tree corresponding to the sentence to be transmitted, including: Acquire a training data set, wherein the training data set includes Chinese words in the transmitted sentence after word segmentation; The Chinese words in the training data set are respectively calculated using the Glove network and the Char_CNN network to calculate the word vector of each Chinese word, and the two obtained word vectors are spliced; The concatenated word vectors are used to calculate the hidden layer feature vectors corresponding to the words using the BiLSTM network; The hidden layer feature vectors corresponding to the two words are input into the dependency edge recognition module to calculate the embedding representation of the dependency edge between each two words; and based on the embedding representation of each dependency edge in the set of dependency edges between words in the training dataset, the dependency edges are scored using the dependency edge scoring function, and the probability of the existence of a dependency edge between words is determined based on the obtained dependency edge score; and the correct dependency edge between words is determined based on the probability of the existence of a dependency edge between words; Use multi-label classification to mark each correct dependency edge, and then use the dependency edge classification module to classify the marked correct dependency edges to obtain the dependency syntax tree prediction result; Constructing a total loss function of a dependency parsing model, wherein the total loss function includes a dependency edge recognition loss function and a dependency edge classification loss function; the dependency parsing model is a network model consisting of a Glove network, a Char_CNN network, a BiLSTM network, a dependency edge recognition module, and a dependency edge classification module; According to the dependency syntax tree prediction result, the total loss function and the training data set, the dependency syntax analysis model is trained using the gradient descent method to obtain a trained dependency syntax analysis model; The sentence to be transmitted after word segmentation is input into the trained dependency syntax analysis model to obtain a dependency syntax tree corresponding to the sentence to be transmitted.

3. The method according to claim 2, characterized in that The hidden layer feature vectors corresponding to the two words are input into the dependency edge recognition module to calculate the embedding representation of the dependency edge between each two words; and the dependency edge is scored using the dependency edge scoring function based on the embedding representation of each dependency edge in the set of dependency edges between words in the training dataset. Based on the obtained dependency edge scores, the probability of the existence of dependency edges between words is determined, and the correct dependency edges between words are determined based on the probability of dependency edges, including: The embedded representation of the dependency edge between the two words is calculated using the hidden layer vectors of the two words. The expression of the embedded representation is: in is the embedding representation of the dependency edge, i ,w j > is the i-th word w i and the jth word w j The dependency edges between are the i-th word w i and the jth word w j The corresponding hidden layer feature vector; According to the embedded representation of each dependency edge in the set of dependency edges between words in the training dataset, the dependency edge is scored using the dependency edge scoring function, which is: where f edge (w i ,w j ) is the i-th word w i and the jth word w j The score of the dependency edge between them, ε is the set of all dependency edges in the transmitted sentence; According to the obtained dependency edge scores, the probability of the existence of dependency edges between words is determined, and the correct dependency edges between words are determined according to the probability of dependency edges. The expression of the probability of dependency edges is: Among them, p(w j ∣w i ) is the i-th word w i and the jth word w j The probability of the dependency edge between them, n is the total number of words in the transmitted sentence.

4. The method according to claim 2, characterized in that The multi-label classification method is used to mark each correct dependency edge, and the marked correct dependency edges are classified using the dependency edge classification module to obtain the dependency syntax tree prediction results, including: For each correct dependency edge, the dependency relationship of the dependency edge is introduced, and the score of whether the dependency relationship of the dependency edge is correct is calculated based on the embedding representation of the correct dependency edge. The scoring function for evaluating whether the dependency relationship is correct is: Where f label (·) is a scoring function to evaluate whether the dependency relationship is correct, ε(r) is the set of dependency edges in ε with relationship r; According to the score of whether the dependency relationship of each dependency edge is correct, the probability of determining that the dependency relationship of the dependency edge is r is: Among them, p(r|w i ,w j ) is the i-th word w i and the jth word w j The probability that the dependency relationship between the dependency edges is r, is the set of all dependency types, r′ is Any dependency relationship in The dependency edges are classified according to the probability that their dependency relationship is r, and the dependency syntax tree prediction result is obtained.

5. The method according to claim 2, characterized in that The total loss function includes the dependency edge recognition loss function and the dependency edge classification loss function; Construct the total loss function of the dependency parsing model, including: Construct a dependency edge recognition loss function and a dependency edge classification loss function, and use the sum of the dependency edge recognition loss function and the dependency edge classification loss function as the total loss function of the dependency parsing model; the dependency edge recognition loss function is: Where, is the dependency edge identification loss, is the dataset used for network training, represents the i-th word in the l-th sentence in the dataset, represent the governing word of The dependency edge classification loss function is: in, is the dependency edge classification loss, r i (l) yes and The dependencies between them.

6. The method according to claim 1, characterized in that Semantically stratifying the sentence to be transmitted according to the dependency syntax tree, and dividing the words in the sentence to be transmitted into different semantic levels, including: According to the dependency syntax tree, the sentence to be transmitted is divided into four different semantic levels; wherein, the first semantic level includes the subject, predicate and object of the sentence to be transmitted; the second semantic level includes all words that depend on the subject, predicate and object of the sentence to be transmitted; the third semantic level includes relational modifiers; and the fourth semantic level includes structural auxiliary words.

7. The method according to claim 6, characterized in that Based on the channel quality, the transmission of words at different semantic levels in the sentence is selectively transmitted, and the words to be transmitted are determined, including: When the channel quality is poor, only the words of the first semantic level are transmitted, and the communication resources of the untransmitted words are used to protect the transmitted words, ensuring that the important semantics conveyed by the transmitted words are not affected by channel noise; As the channel quality gradually improves, words at other semantic levels are transmitted.

8. The method according to claim 1, characterized in that Performing source coding on the transmission words, converting the coded results into bit information, performing channel coding on the bit information, and sending the obtained coded results to a receiving end, including: Performing source coding on the transmission words using Huffman coding, and converting the coded result into bit information; The bit information is channel-coded using LDPC coding, and the obtained coding result is sent to a receiving end.

9. The method according to claim 1, characterized in that The ERNIE model is used to perform semantic recovery on the received sentence, including: The missing words in the received sentence are filled with [MASK] and input into the ERNIE model. Based on the context information in the received sentence, the [MASK] in the received sentence is filled with words that are appropriate in the context to obtain predicted words. The predicted words filled in are: in, To predict words, To predict the conditional probability corresponding to the context information of the word, is the lth predicted word; Emb(context) is the embedding representation of the entire predicted word set context.

10. A hierarchical semantic communication system based on dependency syntactic analysis, characterized in that: The system includes a transmitting end and a receiving end; The sending end and the receiving end communicate by executing the hierarchical semantic communication based on dependency syntax analysis as described in any one of claims 1 to 9.

Images