Semantic communication method and system based on knowledge graph
Through the semantic communication method based on knowledge graph, the problem of semantic representation abstraction and logical rules is solved. The knowledge graph is used to represent semantic information, the importance of triplets is evaluated and selective transmission is performed. Combined with Huffman and LPDC encoding, the low High reliability and effective transmission under signal-to-noise ratio.
Patent Information
- Application Number
- CN202210461166.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-28
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2042-04-28
AI Technical Summary
The existing semantic communication scheme has the problem of abstracting semantic representations and losing the logic rules of human languages, resulting in inaccurate information semantics during transmission.
Using a semantic communication method based on knowledge graphs, the semantic importance of triples is used to construct knowledge graphs, evaluate the semantic importance of triples, determine the transmission triples based on channel quality, and use Huffman encoding and LPDC encoding for channel encoding. The receiver uses GAT network and attention Force mechanism restores semantic information.
In a low signal-to-noise ratio environment, it improves the reliability and effectiveness of communication, ensures the semantic correctness of information transmission, and removes redundancy in the transmitted information.
Smart Images

Figure CN114817565B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of semantic communication technology, and in particular to a semantic communication method and system based on knowledge graph. Background Art
[0002] The purpose of traditional communication systems is to accurately transmit communication symbols, but the meaning of the content expressed is not considered during the transmission process. Different from traditional communication systems, semantic communication systems pay more attention to how to accurately convey the meaning of communication symbols. Therefore, semantic communication systems are closer to the essence of communication and are a deeper communication method.
[0003] Traditional communication systems pursue bit-level indiscriminate transmission, requiring all sent content to be transmitted indiscriminately to the receiving end. Semantic communication systems pursue semantic indiscriminate transmission, so the sending end can extract the semantic information of the sent content for transmission, and the receiving end can restore the transmitted information based on the received semantic information. Existing semantic communication solutions have defects such as abstract representation of semantics and loss of the logical laws of human language. Summary of the invention
[0004] Based on this, it is necessary to provide a semantic communication method and system based on knowledge graph to address the above technical problems, which can overcome the defects of existing semantic communication solutions such as abstract representation of semantics and loss of logical laws of human language.
[0005] A semantic communication method based on knowledge graph, the method comprising:
[0006] Sending side:
[0007] The information to be transmitted is obtained, and a knowledge graph is constructed according to the information to be transmitted; the knowledge graph includes a plurality of triples.
[0008] Semantic similarity is used to evaluate the semantic importance of each triple in the knowledge graph.
[0009] According to the channel quality and the semantic importance of the triples, the transmission triples are determined and the knowledge graph is reconstructed.
[0010] The reconstructed knowledge graph is encoded using Huffman coding, and channel coding is performed using LPDC coding, and the encoded information is sent to the receiving end.
[0011] Receiver:
[0012] Receive the information sent by the sender, perform channel decoding and source decoding on the received information in sequence, and obtain the knowledge graph of the receiving end.
[0013] The GAT network is used to encode the receiving end knowledge graph to obtain the embedded representation of the receiving end knowledge graph.
[0014] According to the embedded representation of the knowledge graph at the receiving end, the attention mechanism and multi-layer perceptron are used to restore the semantic information.
[0015] In one embodiment, obtaining information to be transmitted and constructing a knowledge graph according to the information to be transmitted includes:
[0016] The information to be transmitted is obtained, and the entity extraction model NER is used to extract the entity list in the information to be transmitted.
[0017] A multi-label classification model is used to predict the relationship between entities in the entity list and obtain an entity relationship set.
[0018] Build a knowledge graph based on entity relationship sets.
[0019] In one embodiment, a multi-label classification model is used to predict the relationship between entities in an entity list to obtain an entity relationship set, including:
[0020] According to the word and its context encoding, the word embedding vector is obtained.
[0021] The embedding vector of each word in the entity is weighted averaged to obtain the embedding vector of the entity; the embedding vector expression of the entity is:
[0022]
[0023] In the formula, emb(w j ) is the embedding vector of the word, w j is the jth word in the i-th entity.
[0024] The entity embedding vector is input into the multi-label classification layer to predict the relationship between entities and obtain the entity relationship set.
[0025] In one embodiment, the semantic importance of each triple in the knowledge graph is evaluated by using semantic similarity, including:
[0026] The cosine similarity is used to calculate the semantic importance of each triple in the knowledge graph. The calculation formula for the semantic importance of the triple is:
[0027]
[0028] Among them, sim v (s,g i ) is the i-th triple g in sentence s i The semantic importance of v(s), v(g i ) are the sentence s and the i-th triple g respectively i The vector representation of ||v(s)||, ||v(gi )||represents the vector modulus of sentence s and triple g respectively i The vector modulus of .
[0029] In one embodiment, determining the transmission triplet according to the channel quality and the semantic importance of the triplet, and reconstructing the knowledge graph includes:
[0030] The triplets are sorted in descending order according to their semantic importance to obtain a new knowledge graph.
[0031] The number of transmission triplets is determined according to the channel quality, and the transmission triplets are determined according to the new knowledge graph and the number of transmission triplets.
[0032] The knowledge graph is reconstructed according to the transmission triples to obtain a reconstructed knowledge graph, and the reconstructed knowledge graph expression is:
[0033] G'=](g'1,g'2,…,g' t )1,(g'1,g'2,…,g' t )2,…,(g'1,g'2,…,g' t ) n / t ]
[0034] Among them, G' is the reconstructed knowledge graph, (g'1,g'2,…,g' t ) are t transmitted triplets, and n is the total number of triplets in the knowledge graph.
[0035] A semantic communication system based on a knowledge graph, the system comprising a transmitting end, a receiving end and a channel, the transmitting end comprising a source coding module and a channel coding module, the receiving end comprising a source decoding module and a channel decoding module, the transmitting end further comprising a semantic extraction module, and the receiving end further comprising a semantic recovery module.
[0036] The semantic extraction module is used to construct a knowledge graph based on the acquired information to be transmitted, and use semantic similarity to evaluate the semantic importance of each triple in the knowledge graph, and determine the transmission triple based on the channel quality and the semantic importance of the triple, and reconstruct the knowledge graph.
[0037] The semantic recovery module is used to encode the knowledge graph of the receiving end using the GAT network, and restore the semantic information using the attention mechanism and the multi-layer perceptron according to the obtained results.
[0038] In one of the embodiments, the semantic extraction module includes: a knowledge graph extraction module, a knowledge graph sorting module and a knowledge graph reconstruction module.
[0039] The knowledge graph extraction module is used to receive the information to be transmitted, use the entity extraction model NER to extract the entity list in the information to be transmitted, and use the multi-label classification model to predict the relationship between the entities in the entity list to obtain an entity relationship set, and construct a knowledge graph based on the entity relationship set.
[0040] The knowledge graph sorting module is used to determine the semantic importance of each triple in the knowledge graph based on semantic similarity, and sort the triples according to the semantic importance.
[0041] The knowledge graph reconstruction module is used to determine the transmission triplet based on the channel quality and the sorted triplet, and reconstruct the knowledge graph based on the transmission triplet.
[0042] In one of the embodiments, the semantic importance of each triple in the knowledge graph sorting module is calculated using cosine similarity.
[0043] In one of the embodiments, the knowledge graph reconstruction module is further used to determine the number of transmission triplets according to the channel quality, and determine the transmission triplets according to the sorted knowledge graph and the number of transmission triplets; reconstruct the knowledge graph according to the transmission triplets to obtain a reconstructed knowledge graph, and the reconstructed knowledge graph expression is:
[0044] G'=[(g'1,g'2,…,g' t )1,(g'1,g'2,…,g' t )2,…,(g'1,g'2,…,g' t ) n / t ]
[0045] Among them, G' is the reconstructed knowledge graph, (g'1,g'2,…,g' t ) are t transmitted triplets, and n is the total number of triplets in the knowledge graph.
[0046] In one of the embodiments, the source coding module is used to encode the reconstructed knowledge graph using Huffman coding; the channel coding module is used to encode the bit information after source coding using LPDC coding to resist channel influence.
[0047] The above-mentioned semantic communication method and system based on knowledge graph, the method introduces knowledge graph into semantic communication, uses knowledge graph to represent the semantics of transmitted information, calculates the semantic importance of triples in the knowledge graph, and sorts them by semantic importance. In a poor communication environment, only triples with high semantic importance are transmitted, and transmission resources are used to protect the most important semantic information, so that the transmitted information can express correct semantics. This method is more reliable than traditional communication methods under low signal-to-noise ratio; after the receiving end receives the information and decodes and recovers the knowledge graph, it combines text generation technology to describe the knowledge graph, and then recovers the semantic information. This method introduces knowledge graph into semantic communication, which provides a new idea for semantic representation and semantic recovery in the process of semantic communication. In addition, this method uses knowledge graph to represent semantic information, removes redundancy in transmitted information, and thus improves the effectiveness of semantic communication. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 A schematic diagram of a flow chart of a semantic communication method based on a knowledge graph in one embodiment;
[0049] Figure 2 A block diagram of a knowledge graph-based semantic communication system in one embodiment;
[0050] Figure 3 A block diagram of a knowledge graph-based semantic communication system in one embodiment;
[0051] Figure 4 is a comparison diagram of BERT-based semantic similarity score simulation results of different triplet transmission strategies on a Gaussian white noise channel in another embodiment;
[0052] Figure 5 is a comparison diagram of average transmission bit numbers of different triplet transmission strategies of the method in another embodiment;
[0053] Figure 6 A comparison diagram of BLEU score simulation results between the method and the classical communication system on a Gaussian white noise channel in another embodiment;
[0054] Figure 7 A comparison diagram of the simulation results of semantic similarity scores based on BERT between the method and the classical communication system on a Gaussian white noise channel in another embodiment;
[0055] Figure 8 In another embodiment, a comparison diagram of BLEU score simulation results of the present method and a classical communication system on a Rayleigh fading channel is shown;
[0056] Fig. 9 This is a comparison chart of the simulation results of the semantic similarity score based on BERT between the method and the classical communication system on the Rayleigh fading channel in another embodiment. DETAILED DESCRIPTION
[0057] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0058] In one embodiment, Figure 1 As shown, a semantic communication method based on a knowledge graph is provided, and the method includes the following steps:
[0059] Sending side:
[0060] Step 100: Obtain information to be transmitted, and construct a knowledge graph based on the information to be transmitted; the knowledge graph includes multiple triples.
[0061] Specifically, the Pipeline method is used to extract the knowledge graph corresponding to the information to be transmitted, and the extracted knowledge graph is used to represent the semantics of the transmitted information.
[0062] The use of knowledge graphs to represent semantic information removes redundancy in the transmitted information, thereby improving the effectiveness of the present invention. This method introduces knowledge graphs into semantic communication, providing a new idea for semantic representation and semantic recovery in the process of semantic communication.
[0063] Step 102: Use semantic similarity to evaluate the semantic importance of each triple in the knowledge graph.
[0064] Specifically, the semantic importance of a triple represents the importance of the triple in the completeness of the semantic information expressed by the sentence.
[0065] Step 104: Determine the transmission triples based on the channel quality and the semantic importance of the triples, and reconstruct the knowledge graph.
[0066] Specifically, channel quality refers to the signal-to-noise ratio of the channel. The higher the signal-to-noise ratio, the better the channel quality.
[0067] The triples in the knowledge graph are selectively transmitted according to the channel quality. When the channel quality is poor, only the triples with the highest semantic importance are transmitted. As the channel quality gradually improves, the number of transmitted triples is appropriately increased. After determining the transmission triples, the knowledge graph is reconstructed, and the communication resources of the untransmitted triples are used to protect the transmission triples to ensure the correct transmission of important semantics.
[0068] Step 106: Use Huffman coding to encode the reconstructed knowledge graph, use LPDC coding to perform channel coding, and send the encoded information to the receiving end.
[0069] Specifically, Huffman coding is used to convert the reconstructed knowledge graph into bit information, and LPDC coding is used to resist the influence of channel noise, and the encoded information is sent to the receiving end;
[0070] Receiver:
[0071] Step 108: Receive information sent by the transmitter, perform channel decoding and source decoding on the received information in sequence, and obtain a knowledge graph of the receiver;
[0072] Step 110: Use the GAT network to encode the receiving-end knowledge graph to obtain an embedded representation of the receiving-end knowledge graph.
[0073] Specifically, the receiving end receives the transmitted knowledge graph After that, the GAT network is used to encode the knowledge graph to obtain the embedded representation of the knowledge graph, that is,
[0074] The received knowledge graph is processed to generate a reasonable text structure for it to help the receiver understand the semantics.
[0075] Step 112: Based on the embedded representation of the knowledge graph at the receiving end, the attention mechanism and the multi-layer perceptron are used to restore the semantic information.
[0076] Specifically, according to the embedded representation of the knowledge graph at the receiving end, the attention mechanism and multi-layer perceptron are used to generate the recovery text word by word.
[0077] In the process of restoring text, the knowledge graph embedding vector is first decoded using a recurrent neural network to obtain the hidden state b of the i-th step of the recurrent neural network. i , and then use the attention mechanism to aggregate the context information c i , and finally b i With c i Input them into the multilayer perceptron to calculate the generation probability of the i-th word. The probability of the i-th word is p(w i |w <i ,g)∝exp(m([b i ;c i ])), where m(·) is the multi-layer perceptron calculation function, and ∝ indicates that the probability of generating the i-th word is proportional to the exponential value of the multi-layer perceptron output value. In this process, since a recurrent neural network is used for decoding, the generation of each word in the process of generating the restored text word by word takes into account all the words w generated in the previous decoding process. <i , further increasing the accuracy of the recovered text. After obtaining the probability of the i-th word, the loss function can be constructed using the cross entropy method, and finally the gradient descent method is used for training. The loss function can be expressed as:
[0078]
[0079] in, is the loss function, s and g are the training data sets respectively. The sentences in and their corresponding triples, are the network parameters of the semantic recovery method, It means to find the mathematical expectation.
[0080] In the above-mentioned semantic communication method based on knowledge graph, the method introduces knowledge graph into semantic communication, uses knowledge graph to represent the semantics of transmitted information, calculates the semantic importance of triples in knowledge graph, and sorts them by semantic importance. In a poor communication environment, only triples with high semantic importance are transmitted, and transmission resources are used to protect the most important semantic information so that the transmitted information can express correct semantics. This method is more reliable than traditional communication methods under low signal-to-noise ratio. After the receiving end receives the information and decodes and recovers the knowledge graph, it combines text generation technology to describe the knowledge graph and then recover the semantic information. This method introduces knowledge graph into semantic communication, which provides a new idea for semantic representation and semantic recovery in the process of semantic communication. In addition, this method uses knowledge graph to represent semantic information, removes redundancy in transmitted information, and thus improves the effectiveness of semantic communication.
[0081] In one embodiment, step 100 includes: obtaining information to be transmitted, and using an entity extraction model NER to extract an entity list in the information to be transmitted; using a multi-label classification model to predict the relationship between entities in the entity list to obtain an entity relationship set; and constructing a knowledge graph based on the entity relationship set.
[0082] Specifically, named entity recognition is first performed on the transmission information, and a mature entity extraction model NER is used to obtain a list of entities in the transmission information. Then the relationship between entities is predicted and a knowledge graph is formed.
[0083] In one embodiment, step 100 further includes: obtaining an embedding vector of a word according to the word and the context encoding of the word; performing weighted averaging on the embedding vector of each word in the entity to obtain the embedding vector of the entity; the embedding vector expression of the entity is:
[0084]
[0085] In the formula, emb(w j ) is the embedding vector of the word, w j is the jth word in the i-th entity;
[0086] The entity embedding vector is input into the multi-label classification layer to predict the relationship between entities and obtain the entity relationship set.
[0087] Specifically, after using the entity extraction model to obtain the entity list of the transmitted information, the embedding vector of each word in the entity is weighted averaged to obtain the embedding vector of the entity; the embedding vector expression of the entity is shown in formula (1). j ) is based on the word w j The word embedding vector emb(w j ) is:
[0088] emb(w j )=encode(w j ,w <j ,w >j )
[0089] The embedding vector of a word can be obtained by encoding the word using a long short-term memory network or a bidirectional long short-term memory network.
[0090] After obtaining the entity embedding vector, it is input into the multi-label classification layer to predict the relationship between entities. The multi-label classification layer converts the multi-label multi-classification problem into multiple single-label binary classifications and merges these classification results, that is, r ij =R(e i ,e j ). After the entity relationship set is constructed, the gradient descent method is used for training in combination with the loss function, where the loss function is Where s and g are the training data sets respectively. The sentences in and their corresponding triples, θ is the network parameter of the semantic extraction method.
[0091] In one embodiment, step 102 includes: using cosine similarity to calculate the semantic importance of each triple in the knowledge graph, and the calculation formula for the semantic importance of the triple is:
[0092]
[0093] Among them, sim v (s,g i ) is the i-th triple g in sentence s i The semantic importance of v(s), v(g i ) are the sentence s and the i-th triple g respectively i The vector representation of ||v(s)||, ||v(g i )||represents the vector modulus of sentence s and triple g respectively i The vector modulus of .
[0094] In one embodiment, step 104 includes: sorting the triples in descending order of the semantic importance of the triples to obtain a new knowledge graph; determining the number of transmission triples according to the channel quality, and determining the transmission triples according to the new knowledge graph and the number of transmission triples; reconstructing the knowledge graph according to the transmission triples to obtain a reconstructed knowledge graph, and the reconstructed knowledge graph expression is:
[0095] G'=[(g'1,g'2,…,g' t )1,(g'1,g'2,…,g' t )2,…,(g'1,g'2,…,g' t ) n / t ]
[0096] Among them, G' is the reconstructed knowledge graph, (g'1,g'2,…,g' t ) are t transmitted triplets, and n is the total number of triplets in the knowledge graph.
[0097] Preferably, when the signal-to-noise ratio of the channel is less than 0dB, only the triplet with the highest semantic importance is transmitted; when the signal-to-noise ratio is between 0dB and 1dB, the triplets ranked in the top 50% of semantic importance are transmitted; when the signal-to-noise ratio is greater than 1dB, all triplets are transmitted.
[0098] It should be understood that although Figure 1 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. Moreover, Figure 1 At 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.
[0099] In one embodiment, a conventional communication system includes a transmitting end, a receiving end, and a channel, wherein the transmitting end includes a source coding module and a channel coding module, and the receiving end includes a source decoding module and a channel decoding module; Figure 2 As shown, this embodiment adds a semantic extraction module at the sending end and a semantic recovery module at the receiving end based on the traditional communication system.
[0100] The semantic extraction module is used to construct a knowledge graph based on the acquired information to be transmitted, and use semantic similarity to evaluate the semantic importance of each triple in the knowledge graph. According to the channel quality and the semantic importance of the triple, the transmission triple is determined and the knowledge graph is reconstructed.
[0101] The semantic recovery module is used to encode the receiving end knowledge graph using the GAT network, and restore the semantic information using the attention mechanism and multi-layer perceptron based on the results obtained.
[0102] In one embodiment, if Figure 3 As shown, the semantic extraction module includes: a knowledge graph extraction module, a knowledge graph sorting module and a knowledge graph reconstruction module; the knowledge graph extraction module is used to receive the information to be transmitted, use the entity extraction model NER to extract the entity list in the information to be transmitted, and use the multi-label classification model to predict the relationship between the entities in the entity list to obtain the entity relationship set, and construct the knowledge graph according to the entity relationship set; the knowledge graph sorting module is used to determine the semantic importance of each triple in the knowledge graph according to the semantic similarity, and sort the triples according to the semantic importance; the knowledge graph reconstruction module is used to determine the transmission triple according to the channel quality and the sorted triples, and reconstruct the knowledge graph according to the transmission triples.
[0103] The semantic extraction module represents the semantic information by extracting the knowledge graph from the transmitted information.
[0104] In one of the embodiments, the semantic importance of each triple in the knowledge graph ranking module is calculated using cosine similarity.
[0105] In one of the embodiments, the knowledge graph reconstruction module is further used to determine the number of transmission triplets according to the channel quality, and determine the transmission triplets according to the sorted knowledge graph and the number of transmission triplets; reconstruct the knowledge graph according to the transmission triplets to obtain a reconstructed knowledge graph, and the reconstructed knowledge graph expression is:
[0106] G'=[(g'1,g'2,…,g' t )1,(g'1,g'2,…,g' t )2,…,(g'1,g'2,…,g' t ) n / t ]
[0107] Among them, G' is the reconstructed knowledge graph, (g'1,g'2,…,g' t ) are t transmitted triplets, and n is the total number of triplets in the knowledge graph.
[0108] In one of the embodiments, the source coding module is used to encode the reconstructed knowledge graph using Huffman coding; the channel coding module is used to encode the bit information after source coding using LPDC coding to resist channel influence.
[0109] The semantic recovery module generates text based on the received knowledge graph and then recovers the semantic information.
[0110] In a simulation verification embodiment, Figure 4 The figure shows the simulation results of different triplet transmission strategies of the method of the present invention under additive white Gaussian noise channel. The present invention selects three most representative transmission strategies, namely, transmitting the most important triplet (Send the 1st triplet), transmitting 50% of the triplets (Send 50% triplets) and transmitting all triplets (Send 100% triplets). Figure 4 The evaluation criteria used in the paper are: the 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 0dB, the strategy of transmitting only the most important triples has the highest score; when the signal-to-noise ratio is between 0dB and 1dB, the strategy of transmitting 50% of the triples has the highest score; when the signal-to-noise ratio is greater than 1dB, the strategy of transmitting all triples has the highest score. The simulation results are in line with the expectations of the method of the present invention, and also verify the effectiveness of the scheme proposed by the method of the present invention to select and transmit according to the semantic importance of the knowledge graph. Figure 5 Three triplet transmission strategies and the average number of bits required to transmit information are given. Figure 5 As shown in , compared with the original transmission information, the three transmission strategies can save transmission resources. This verifies the performance of the method of the present invention in improving the effectiveness of the communication process.
[0111] In another simulation verification embodiment, Figure 6-Figure 9 As shown in the figure, the simulation results of the reliable semantic communication system based on the knowledge graph in the Gaussian white noise channel and Rayleigh fading channel model are given. 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, and the channel response H obeys the complex Gaussian distribution. Gaussian white noise vector Next, we introduce another evaluation criterion besides semantic similarity. BLEU (Bilingual evaluation under study) is used to calculate the semantic similarity between sentences. BLEU (1-gram) calculates the similarity of words in two paragraphs by comprehensively comparing the proportion of n = 1 word tuples that appear together in two paragraphs. This method is compared with the classic communication system architecture (Huffman source coding and LDPC channel coding with a code rate of 2 / 3) and semantic extraction without considering channel quality and semantic recovery (Extraction) under two channel models. Figure 6 and Figure 7 As shown in the figure, in a Gaussian white noise channel, the BLEU (1-gram) between the transmitted and recovered information and the semantic similarity score based on the BERT model are higher than the classic communication Huffman source coding plus LDPC channel coding method when the signal-to-noise ratio is less than 4dB. Figure 8 and Fig. 9 It is reflected that in the Rayleigh fading channel, the method of the present invention and the classical source plus channel coding method have higher semantic similarity scores than BLEU and the BERT model, which verifies the effectiveness of the reliable semantic communication system based on the knowledge graph of the present invention.
[0112] The technical features of the above embodiments may 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.
[0113] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A semantic communication method based on knowledge graph, characterized in that: The method comprises: Sending side: Acquire information to be transmitted, and construct a knowledge graph according to the information to be transmitted; the knowledge graph includes multiple triples; Semantic similarity is used to evaluate the semantic importance of each triple in the knowledge graph; the semantic importance of each triple is calculated using cosine similarity; the calculation formula for the semantic importance of the triple is: in, For in the sentence Middle Triples The semantic importance of , Separate sentences s and i Triples The vector representation of Represents the sentences The vector modulus and triplet of The vector modulus of ; According to the channel quality and the semantic importance of the triples, the transmission triples are determined and the knowledge graph is reconstructed; Huffman coding is used to encode the reconstructed knowledge graph, and LPDC coding is used for channel coding, and the encoded information is sent to the receiving end; Receiver: Receive the information sent by the sender, perform channel decoding and source decoding on the received information in sequence, and obtain the knowledge graph of the receiving end; The GAT network is used to encode the knowledge graph of the receiving end to obtain the embedded representation of the knowledge graph of the receiving end; According to the embedded representation of the knowledge graph at the receiving end, the attention mechanism and multi-layer perceptron are used to restore the semantic information; Among them, according to the channel quality and the semantic importance of the triplet, the transmission triplet is determined and the knowledge graph is reconstructed, including: Sorting the triples in descending order of their semantic importance to obtain a new knowledge graph; Determine the number of transmission triplets according to the channel quality, and determine the transmission triplets according to the new knowledge graph and the number of transmission triplets; The knowledge graph is reconstructed according to the transmission triples to obtain a reconstructed knowledge graph, and the reconstructed knowledge graph expression is: in, To reconstruct the knowledge graph, for transmitted triples, and n is the total number of triples in the knowledge graph.
2. The method according to claim 1, characterized in that Acquiring information to be transmitted and constructing a knowledge graph based on the information to be transmitted, including: Obtain the information to be transmitted, and use the entity extraction model NER to extract the entity list in the information to be transmitted; Use a multi-label classification model to predict the relationship between entities in the entity list and obtain an entity relationship set; Build a knowledge graph based on entity relationship sets.
3. The method according to claim 2, characterized in that A multi-label classification model is used to predict the relationship between entities in the entity list, and an entity relationship set is obtained, including: According to the word and its context encoding, the word embedding vector is obtained; The embedding vector of each word in the entity is weighted averaged to obtain the embedding vector of the entity; the embedding vector expression of the entity is: In the formula, is the embedding vector of the word, For the of the entities words; The entity embedding vector is input into the multi-label classification layer to predict the relationship between entities and obtain the entity relationship set.
4. A semantic communication system based on knowledge graph, the system comprising a transmitter, a receiver and a channel, the transmitter comprising a source coding module and a channel coding module, the receiver comprising a source decoding module and a channel decoding module, characterized in that: The transmitting end further includes a semantic extraction module, and the receiving end further includes a semantic recovery module; The semantic extraction module is used to construct a knowledge graph based on the acquired information to be transmitted, and use semantic similarity to evaluate the semantic importance of each triple in the knowledge graph, and determine the transmission triple according to the channel quality and the semantic importance of the triple, and reconstruct the knowledge graph; wherein the semantic importance of each triple is calculated using cosine similarity; the calculation formula for the semantic importance of the triple is: in, For in the sentence Middle Triples The semantic importance of , Separate sentences s and i Triples The vector representation of Represents the sentences The vector modulus and triplet of The vector modulus of ; The semantic recovery module is used to encode the receiving end knowledge graph using the GAT network, and restore the semantic information using the attention mechanism and the multi-layer perceptron according to the obtained results; The semantic extraction module is further used to sort the triples in descending order of the semantic importance of the triples to obtain a new knowledge graph; determine the number of transmission triples according to the channel quality, and determine the transmission triples according to the new knowledge graph and the number of transmission triples; reconstruct the knowledge graph according to the transmission triples to obtain a reconstructed knowledge graph, and the reconstructed knowledge graph expression is: in, To reconstruct the knowledge graph, for transmitted triples, and n is the total number of triples in the knowledge graph.
5. The system according to claim 4, characterized in that The semantic extraction module includes: a knowledge graph extraction module, a knowledge graph sorting module and a knowledge graph reconstruction module; The knowledge graph extraction module is used to receive the information to be transmitted, use the entity extraction model NER to extract the entity list in the information to be transmitted, and use the multi-label classification model to predict the relationship between the entities in the entity list to obtain an entity relationship set, and construct a knowledge graph according to the entity relationship set; The knowledge graph sorting module is used to determine the semantic importance of each triple in the knowledge graph according to the semantic similarity, and sort the triples according to the semantic importance; The knowledge graph reconstruction module is used to determine the transmission triplet based on the channel quality and the sorted triplet, and reconstruct the knowledge graph based on the transmission triplet.
6. The system according to claim 4, characterized in that The source coding module is used to encode the reconstructed knowledge graph using Huffman coding. The channel coding module is used to encode the bit information after source coding using LPDC coding to resist channel influence.
