Knowledge graph completion method and device based on attention penalty and noise sampling
By introducing attention punishment and noise sampling technologies into the multimodal knowledge graph completion method, the robustness and multimodal hallucination problems of the multimodal knowledge graph completion method are solved, and the accuracy of knowledge graph completion is significantly improved. Sex and integrity.
Patent Information
- Application Number
- CN202411230179.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-04
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-09-04
AI Technical Summary
The existing multimodal knowledge graph completion method lacks robustness and has multimodal hallucinations problems, which affects the accuracy of the knowledge graph completion task.
Using an approach based on attention punishment and noise sampling, adaptive noise sampling and attention punishment processing are carried out by obtaining embedded vectors of multimodal data, a knowledge graph completion model is constructed to enhance model robustness and avoid multimodal illusion.
Through adaptive noise sampling and attention punishment mechanisms, the robustness of the model is enhanced, and the accuracy and completeness of knowledge graph completion are improved, which is specifically manifested as an accuracy increase of 12% on the DB15k dataset and MKG-W data The accuracy on set is improved by 14%, and 8% on the DB15k dataset and 11% on the MKG-W dataset.
Smart Images

Figure CN118734947B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of knowledge graph technology, and in particular to a knowledge graph completion method and device based on attention penalty and noise sampling. Background Art
[0002] A knowledge graph is a directed graph structure that represents entities and their relationships in the real world. It consists of nodes (entities) and edges (relationships), where each node represents an entity and each edge represents a relationship between entities. Knowledge graphs are widely used in search engines, recommendation systems, natural language processing, and other fields. Through structured knowledge representation and reasoning capabilities, they improve the efficiency of data understanding and information retrieval.
[0003] With the continuous increase in data volume and the diversification of information needs, the traditional knowledge graph that only relies on text and structured data has gradually exposed its limitations. For example, for some specific fields, pure text data cannot fully reflect the multidimensional characteristics and relationships of entities. Therefore, it has become a trend to introduce non-text modal data such as images and videos to build multimodal knowledge graphs. Multimodal data includes information in different forms such as text, images, videos, and audio. Each modality of data contains unique and valuable features, which play a complementary role in multimodal knowledge graphs. Knowledge graph completion refers to the process of predicting and adding new entities and relationships based on existing knowledge graphs through algorithms and models to supplement and improve knowledge graphs. Since there will inevitably be missing and incomplete data in the construction process of knowledge graphs, completion technology is particularly important. Multimodal knowledge graph completion technology uses multimodal data such as text and images to predict and add missing entities and relationships to improve the completeness and accuracy of knowledge graphs.
[0004] However, in most existing multimodal knowledge graph completion methods, the acquired data does not include external noise data, resulting in a lack of robustness in the model. In addition, the multimodal language large models commonly used in the field of multimodal knowledge graphs currently have the problem of multimodal hallucination, which will affect the accuracy of the knowledge graph completion task. Summary of the invention
[0005] The purpose of the present invention is to provide a knowledge graph completion method and device based on attention penalty and noise sampling to address the lack of robustness of existing multimodal knowledge graph completion methods and the multimodal hallucination problem therein.
[0006] The above-mentioned purpose of the present application is achieved through the following technical solutions:
[0007] S1: Obtain multimodal data of the knowledge graph, fuse the text embedding, image embedding and relationship embedding of the entity, and obtain the first embedding vector;
[0008] S2: Perform adaptive noise sampling on the first embedding vector to obtain a second embedding vector;
[0009] S3: Encode the second embedding vector through an encoder with an attention penalty mechanism to obtain a prediction vector ;
[0010] S4: Build a knowledge graph completion model; calculate the prediction vector through the score function The predicted entity scores are used to train the knowledge graph completion model.
[0011] S5: Obtain the knowledge graph to be completed; perform the knowledge graph completion task on the knowledge graph to be completed through the trained knowledge graph completion model.
[0012] Optionally, step S1 includes:
[0013] S11: Obtaining a multimodal knowledge graph , Represents the set of all entities in the knowledge graph, represents the set of all relations, represents the set of all correct triples, is a set representing multimodal data; Get triplets of multimodal data ,in Represents the head entity, Indicates relationship, Represents the tail entity;
[0014] S12: Fusion of head entities through feature processing Text Embedding and image embedding , get the relation embedding , get the first embedding vector ,as follows:
[0015] .
[0016] Optionally, step S2 includes:
[0017] S21: Exploiting a Learnable Noise Scaling Factor , get the noise sample of the first embedding vector ;
[0018] Noise sampling includes two sampling methods, namely Gaussian noise sampling and average noise sampling;
[0019] The calculation formula for Gaussian noise sampling is as follows:
[0020]
[0021] in Indicates matrix dot multiplication operation. Represents a learnable noise scaling factor, which adaptively changes the proportion of noise in the embedding during training. represents the standard deviation of the original noise embedded data, represents the average value of the original noise embedded data; represents a mean vector with mean 0 and a unit covariance matrix Gaussian distribution value of ; This means that from the Gaussian distribution The sample values obtained in ; represents the noise input vector obtained after Gaussian noise sampling, that is, the noise sampling of the first embedded vector;
[0022] The formula for average noise sampling is as follows:
[0023]
[0024] in Indicates matrix dot multiplication operation. represents the learnable noise scaling factor, Indicates the maximum value of the random variable interval set, Indicates the minimum value of the random variable interval set, Is subject to the set interval A random variable uniformly distributed on represents the noise input vector obtained after averaging the noise sampling, that is, the noise sampling of the first embedded vector;
[0025] S22: Setting random factors , compare the random factors The sampling factor of the first embedding vector , determine whether to add noise samples to the first input vector , generate a judgment result;
[0026] S23: Add noise samples to the first embedding vector according to the judgment result , expressed as:
[0027]
[0028] in, represents the second embedding vector obtained after adaptive noise sampling.
[0029] Optionally, step S3 includes:
[0030] S31: Through the second embedding vector , perform attention calculation and obtain attention score;
[0031] Through the attention penalty mechanism, the attention score is reconstructed to obtain the reconstructed attention penalty embedding vector , the calculation formula is as follows:
[0032]
[0033] in Represents the calculation function of the attention penalty mechanism;
[0034] S32: Construct an encoder, which includes a feedforward neural network, a fully connected layer and Function layer;
[0035] S33: Attention Penalty Embedding Vector After the feedforward neural network layer, the fully connected layer, The function layer is encoded to obtain the prediction vector , the calculation formula is as follows:
[0036]
[0037] in Represents the encoding calculation of the encoder.
[0038] Optionally, step S31 includes:
[0039] S31a: Calculate the attention score of each feature in the second embedding vector through the attention penalty mechanism. The calculation formula is as follows:
[0040]
[0041] in, Indicates The second embedding vector of the entity The attention score of each feature, represents the set of all feature attention scores, represents the number of features of the second embedding vector, Indicates the attention calculation operation;
[0042] S31b: Compare the attention scores of each feature. If the attention score is greater than the set penalty limit, an attention penalty is applied to the attention score of the feature. The calculation formula is as follows:
[0043]
[0044] in, Indicates the penalty limit set. Indicates the attention score that is greater than the set penalty limit, represents the penalty attention factor, represents the attention score after punishment;
[0045] S31c: Attention score After the penalty, the attention score in the set The attention scores of the previously preset number of features are redistributed, and the calculation process is as follows:
[0046]
[0047] in, represents the amount of attention reallocation, represents the attention score to be redistributed, Redistribute the attention factor, represents the attention score after reallocation;
[0048] S31e: Repeat steps S32b-S32c to obtain the attention penalty embedding vector as shown below:
[0049] .
[0050] Optionally, step S4 includes:
[0051] S41: Construct a scoring function, the formula is as follows:
[0052]
[0053] in Represents the prediction vector The real part score of It represents the imaginary score of the prediction vector, Represents the final predicted entity score;
[0054] The real part score of the prediction vector is calculated as follows:
[0055]
[0056] in represents the prediction vector, Indicates the cosine value operation. represents the boundary value of the predicted embedding vector, represents the real part embedding of the prediction vector;
[0057] Regarding the imaginary score of the prediction vector, the calculation formula is as follows:
[0058]
[0059] in Indicates the operation of finding the sine value. is the imaginary part embedding of the prediction vector;
[0060] S42: Train the knowledge graph completion model by predicting entity scores;
[0061] The loss function of the knowledge graph completion model is as follows:
[0062]
[0063] in Indicates the sigmoid activation function operation. represents the boundary value of the loss function, Represents the loss value.
[0064] An electronic device includes a processor, a memory, a user interface and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device performs a knowledge graph completion method based on attention penalty and noise sampling.
[0065] A computer-readable storage medium stores instructions, which, when executed, execute a knowledge graph completion method based on attention penalty and noise sampling.
[0066] The beneficial effects of the technical solution provided by this application are:
[0067] The multimodal knowledge graph completion model performs adaptive noise sampling on the multimodal data of the knowledge graph entities, simulating real-world scenarios. The introduction of noise allows the model to avoid the negative impact of a small amount of data, thereby enhancing the robustness of the model. The attention penalty mechanism is used to penalize over-reliance on input data to avoid excessive attention weights for a single feature, thereby alleviating the multimodal hallucination problem, so that the trained learning model can perform more accurate multimodal knowledge graph completion tasks. BRIEF DESCRIPTION OF THE DRAWINGS
[0068] The present application will be further described below with reference to the accompanying drawings and embodiments, in which:
[0069] Figure 1 This is a step diagram of a knowledge graph completion method based on attention penalty and noise sampling in an embodiment of the present application;
[0070] Figure 2 Schematic diagram of the attention penalty mechanism of the knowledge graph completion method based on attention penalty and noise sampling in an embodiment of the present application;
[0071] Figure 3 is a schematic diagram of the structure of an electronic device of a knowledge graph completion method based on attention penalty and noise sampling in an embodiment of the present application;
[0072] Figure 4 This is a schematic diagram of triples to be completed in the knowledge graph completion method based on attention penalty and noise sampling in an embodiment of the present application. DETAILED DESCRIPTION
[0073] In order to have a clearer understanding of the technical features, purposes and effects of the present application, the specific implementation methods of the present application are now described in detail with reference to the accompanying drawings.
[0074] An embodiment of the present application provides a knowledge graph completion method based on attention penalty and noise sampling.
[0075] Please refer to Figure 1 , Figure 1 It is a step diagram of a knowledge graph completion method based on attention penalty and noise sampling in an embodiment of the present application, including:
[0076] S1: Obtain multimodal data of the knowledge graph, fuse the text embedding, image embedding and relationship embedding of the entity, and obtain the first embedding vector;
[0077] Step S1 includes:
[0078] S11: Obtaining a multimodal knowledge graph , Represents the set of all entities in the knowledge graph, represents the set of all relations, represents the set of all correct triples, is a set representing multimodal data; Get triplets of multimodal data ,in Represents the head entity, Indicates relationship, Represents the tail entity;
[0079] S12: Fusion of head entities through feature processing Text Embedding and image embedding , get the relation embedding , get the first embedding vector ,as follows:
[0080] .
[0081] Specifically, for image embedding , a pre-trained and then frozen visual encoder is used to extract the visual feature parts of each entity with associated image data. For image embedding , we use Sentence-BERT, a pre-trained text encoder, to obtain text features from the description of each entity.
[0082] S2: Perform adaptive noise sampling on the first embedding vector to obtain a second embedding vector;
[0083] Step S2 includes:
[0084] S21: Exploiting a Learnable Noise Scaling Factor , get the noise sample of the first embedding vector ;
[0085] Noise sampling includes two sampling methods, namely Gaussian noise sampling and average noise sampling;
[0086] The calculation formula for Gaussian noise sampling is as follows:
[0087]
[0088] in Indicates matrix dot multiplication operation. Represents a learnable noise scaling factor, which adaptively changes the proportion of noise in the embedding during training. represents the standard deviation of the original noise embedded data, represents the average value of the original noise embedded data; represents a mean vector with mean 0 and a unit covariance matrix Gaussian distribution value of ; This means that from the Gaussian distribution The sample values obtained in ; represents the noise input vector obtained after Gaussian noise sampling, that is, the noise sampling of the first embedded vector;
[0089] The formula for average noise sampling is as follows:
[0090]
[0091] in Indicates matrix dot multiplication operation. represents the learnable noise scaling factor, Indicates the maximum value of the random variable interval set, Indicates the minimum value of the random variable interval set, Is subject to the set interval A random variable uniformly distributed on represents the noise input vector obtained after averaging the noise sampling, that is, the noise sampling of the first embedded vector;
[0092] S22: Setting random factors , compare the random factors The sampling factor of the first embedding vector , determine whether to add noise samples to the first input vector , generate a judgment result;
[0093] S23: Add noise samples to the first embedding vector according to the judgment result , expressed as:
[0094]
[0095] in, represents the second embedding vector obtained after adaptive noise sampling.
[0096] Specifically, for different knowledge graph datasets, the noise sampling method is flexibly selected according to its characteristics. For knowledge graph datasets with more relational data, that is, more edges, such as the DB15k dataset, Gaussian noise sampling is used to simulate the impact of cluttered environments in the real world on prediction. For knowledge graph datasets with more entity data, that is, more nodes, such as the MKG-W dataset, average noise sampling is used to model the complex situations between nodes. And according to experiments, the noise scale factor is And the sampling factor Setting it as a learnable parameter and allowing it to change adaptively during training is beneficial to increasing the accuracy of knowledge graph completion. Experiments have confirmed that this strategy improves the accuracy of knowledge graph completion by 12% on the DB15k dataset and 14% on the MKG-W dataset.
[0097] S3: Encode the second embedding vector through an encoder with an attention penalty mechanism to obtain a prediction vector ;
[0098] Step S3 includes:
[0099] S31: Through the second embedding vector , perform attention calculation and obtain attention score;
[0100] Through the attention penalty mechanism, the attention score is reconstructed to obtain the reconstructed attention penalty embedding vector , the calculation formula is as follows:
[0101]
[0102] in Represents the calculation function of the attention penalty mechanism;
[0103] Step S31 includes:
[0104] S31a: Calculate the attention score of each feature in the second embedding vector through the attention penalty mechanism. The calculation formula is as follows:
[0105]
[0106] in, Indicates The second embedding vector of the entity The attention score of each feature, represents the set of all feature attention scores, represents the number of features of the second embedding vector, Indicates the attention calculation operation;
[0107] S31b: Compare the attention scores of each feature. If the attention score is greater than the set penalty limit, an attention penalty is applied to the attention score of the feature. The calculation formula is as follows:
[0108]
[0109] in, Indicates the penalty limit set. Indicates the attention score that is greater than the set penalty limit, represents the penalty attention factor, represents the attention score after punishment;
[0110] S31c: Attention score After the penalty, the attention score in the set The attention scores of the previously preset number of features are redistributed, and the calculation process is as follows:
[0111]
[0112] in, represents the amount of attention reallocation, represents the attention score to be redistributed, Redistribute the attention factor, represents the attention score after reallocation;
[0113] S31e: Repeat steps S32b-S32c to obtain the attention penalty embedding vector as shown below:
[0114] .
[0115] S32: Construct an encoder, which includes a feedforward neural network, a fully connected layer and Function layer;
[0116] S33: Attention Penalty Embedding Vector After the feedforward neural network layer, the fully connected layer, The function layer is encoded to obtain the prediction vector , the calculation formula is as follows:
[0117]
[0118] in Represents the encoding calculation of the encoder.
[0119] Specifically, for the self-attention operation, a transformer encoder-like operation is used to perform a multi-head attention operation on the input vector using the attention bias value. Multiple attention scores are obtained. For such attention scores, after experimental analysis, there is a problem of over-reliance on attention, that is, too high attention scores are given to some labels. The designed attention penalty module can effectively reduce this attention dependence. It is beneficial to increase the accuracy of knowledge graph completion. Experiments have confirmed that this strategy has improved the accuracy of knowledge graph completion by 8% on the DB15k dataset and 11% on the MKG-W dataset. Figure 2 This is a schematic diagram of the attention penalty mechanism of the knowledge graph completion method based on attention penalty and noise sampling.
[0120] S4: Build a knowledge graph completion model; calculate the prediction vector through the score function The predicted entity scores are used to train the knowledge graph completion model.
[0121] Step S4 includes:
[0122] S41: Construct a scoring function, the formula is as follows:
[0123]
[0124] in Represents the prediction vector The real part score of It represents the imaginary score of the prediction vector, Represents the final predicted entity score;
[0125] The real part score of the prediction vector is calculated as follows:
[0126]
[0127] in represents the prediction vector, Indicates the cosine value operation. represents the boundary value of the predicted embedding vector, represents the real part embedding of the prediction vector;
[0128] Regarding the imaginary score of the prediction vector, the calculation formula is as follows:
[0129]
[0130] in Indicates the operation of finding the sine value. is the imaginary part embedding of the prediction vector;
[0131] S42: Train the knowledge graph completion model by predicting entity scores;
[0132] The loss function of the knowledge graph completion model is as follows:
[0133]
[0134] in Indicates the sigmoid activation function operation. represents the boundary value of the loss function, Represents the loss value.
[0135] Specifically, for the scoring function, a scoring method similar to the RotatE model is used as a scoring function to evaluate the rationality of the triple. It transforms the embedding of the head entity into an embedding that approximates the tail entity using a relationship. In order to give a higher score to the correct predicted triple, a sigmoid-based loss function is used to optimize the embedding, which makes the model more accurate in prediction. Experiments have shown that the use of a sigmoid-based loss function improves the accuracy of knowledge graph completion by 17% on the DB15k dataset and 23% on the MKG-W dataset.
[0136] S5: Obtain the knowledge graph to be completed; perform the knowledge graph completion task on the knowledge graph to be completed through the trained knowledge graph completion model.
[0137] Specifically, the triples to be completed are extracted from the knowledge graph to be completed, including the head entity and relationship , where the tail entity To be completed; in this embodiment, the header entity provided by the user and relationship They are: the first leader and the leader.
[0138] In this embodiment, if Figure 4As shown, taking the triple (first leader, leader, company A) as an example, the head entity As the first leader, relationship For the leader, the tail entity For company A. And get the head entity The features of the text to be completed are obtained from the text description of as part of the text embedding of the tail entity, so as to supplement the text information of the tail entity. Figure 4 As shown in the figure, taking the triple (Company A, located in, City B) as an example, add the entity description of the head entity Company A: "Company A is the largest online e-commerce company in City B, located in City B in a certain province." as the text embedding of the tail entity Through the knowledge graph completion model, combined with the description, the tail entity in the triple to be completed is completed.
[0139] The present application also discloses an electronic device. Figure 3 , Figure 3 The electronic device 500 may include: at least one processor 501 , at least one network interface 504 , a user interface 503 , a memory 505 , and at least one communication bus 502 .
[0140] The communication bus 502 is used to realize the connection and communication between these components.
[0141] The user interface 503 may include a display screen, and the optional user interface 503 may also include a standard wired interface or a wireless interface.
[0142] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).
[0143] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure.
[0144] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.
Claims
1. A knowledge graph completion method based on attention penalty and noise sampling, characterized in that: The method comprises the following steps: S1: Obtain multimodal data of the knowledge graph, fuse the text embedding, image embedding and relationship embedding of the entity, and obtain the first embedding vector; S2: Perform adaptive noise sampling on the first embedding vector to obtain a second embedding vector; Step S2 includes: S21: Exploiting a Learnable Noise Scaling Factor , get the noise sample of the first embedding vector ; Noise sampling includes two sampling methods, namely Gaussian noise sampling and average noise sampling; The calculation formula for Gaussian noise sampling is as follows: in Indicates matrix dot multiplication operation. Represents a learnable noise scaling factor, which adaptively changes the proportion of noise in the embedding during training. represents the standard deviation of the original noise embedded data, represents the average value of the original noise embedded data; represents a mean vector with mean 0 and a unit covariance matrix Gaussian distribution value of ; This means that from the Gaussian distribution The sample values obtained in ; represents the noise input vector obtained after Gaussian noise sampling, that is, the noise sampling of the first embedded vector; The formula for average noise sampling is as follows: in Indicates matrix dot multiplication operation. represents the learnable noise scaling factor, Indicates the maximum value of the random variable interval set, Indicates the minimum value of the random variable interval set, Is subject to the set interval A random variable uniformly distributed on represents the noise input vector obtained after averaging the noise sampling, that is, the noise sampling of the first embedded vector; S22: Setting random factors , compare the random factors The sampling factor of the first embedding vector , determine whether to add noise samples to the first input vector , generate a judgment result; S23: Add noise samples to the first embedding vector according to the judgment result , expressed as: in, represents the second embedding vector obtained after adaptive noise sampling; S3: Encode the second embedding vector through an encoder with an attention penalty mechanism to obtain a prediction vector ; Step S3 includes: S31: Through the second embedding vector , perform attention calculation and obtain attention score; Through the attention penalty mechanism, the attention score is reconstructed to obtain the reconstructed attention penalty embedding vector , the calculation formula is as follows: in Represents the calculation function of the attention penalty mechanism; Step S31 includes: S31a: Calculate the attention score of each feature in the second embedding vector through the attention penalty mechanism. The calculation formula is as follows: in, Indicates The second embedding vector of the entity The attention score of each feature, represents the set of all feature attention scores, represents the number of features of the second embedding vector, Indicates the attention calculation operation; S31b: Compare the attention scores of each feature. If the attention score is greater than the set penalty limit, an attention penalty is applied to the attention score of the feature. The calculation formula is as follows: in, Indicates the penalty limit set. Indicates the attention score that is greater than the set penalty limit, represents the penalty attention factor, represents the attention score after punishment; S31c: Attention score After the penalty, the attention score in the set The attention scores of the previously preset number of features are redistributed, and the calculation process is as follows: in, represents the amount of attention reallocation, represents the attention score to be redistributed, Redistribute the attention factor, represents the attention score after reallocation; S31e: Repeat steps S32b-S32c to obtain the attention penalty embedding vector as shown below: S32: Construct an encoder, which includes a feedforward neural network, a fully connected layer and Function layer; S33: Attention Penalty Embedding Vector After the feedforward neural network layer, the fully connected layer, The function layer is encoded to obtain the prediction vector , the calculation formula is as follows: in represents the encoding calculation of the encoder; S4: Build a knowledge graph completion model; calculate the prediction vector through the score function The predicted entity scores are used to train the knowledge graph completion model. S5: Obtain the knowledge graph to be completed; perform the knowledge graph completion task on the knowledge graph to be completed through the trained knowledge graph completion model.
2. A knowledge graph completion method based on attention penalty and noise sampling as claimed in claim 1, characterized in that: Step S1 includes: S11: Obtaining a multimodal knowledge graph , Represents the set of all entities in the knowledge graph, represents the set of all relations, represents the set of all correct triples, is a set representing multimodal data; Get triplets of multimodal data ,in Represents the head entity, Indicates relationship, Represents the tail entity; S12: Fusion of head entities through feature processing Text embedding and image embedding , get the relation embedding , get the first embedding vector ,as follows: 。 3. A knowledge graph completion method based on attention penalty and noise sampling as claimed in claim 1, characterized in that: Step S4 includes: S41: Construct a scoring function, the formula is as follows: in Represents the prediction vector The real part score of It represents the imaginary score of the prediction vector, Represents the final predicted entity score; The real part score of the prediction vector is calculated as follows: in represents the prediction vector, Indicates the cosine value operation. represents the boundary value of the predicted embedding vector, represents the real part embedding of the prediction vector; Regarding the imaginary score of the prediction vector, the calculation formula is as follows: in Indicates the operation of finding the sine value. is the imaginary part embedding of the prediction vector; S42: Train the knowledge graph completion model by predicting entity scores; The loss function of the knowledge graph completion model is as follows: in Indicates the sigmoid activation function operation. represents the boundary value of the loss function, Represents the loss value.
4. An electronic device, characterized in that: The electronic device comprises a processor (501), a memory (505), a user interface (503) and a network interface (504), wherein the memory (505) is used to store instructions, the user interface (503) and the network interface (504) are used to communicate with other devices, and the processor (501) is used to execute the instructions stored in the memory (505) so that the electronic device executes the method according to any one of claims 1 to 3.
5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed by a computer, the method steps according to any one of claims 1 to 3 are executed.
