Event false information detection method and device based on prior knowledge guidance

Abstract

The application discloses an event false information detection method and device based on prior knowledge guidance, and relates to the technical field of network information security. The method comprises the following steps: obtaining a description text and an associated picture of an event to be detected, extracting a relationship triple, a text feature and an image feature from the description text and the associated picture, verifying and optimizing the triple relationship feature based on prior knowledge, ensuring the judgment ability of a model on a small sample event relationship, fusing the triple relationship feature, the text feature and the image feature, realizing sufficient expression of multi-modal information and enhancement of discrimination ability, and training a prediction model in combination with a relationship loss and a prediction loss, enhancing the prediction accuracy of the model, and improving the credibility and interpretability of the detection result.

Description

Technical Field

[0001] The present invention relates to the field of network information security technology, and in particular to a method and device for detecting false event information based on prior knowledge guidance. Background Art

[0002] In the field of network security technology, network information security is particularly important. The openness of the Internet leads to it being filled with various false information. The spread of such false information not only interferes with public judgment, but may also cause confusion in social public opinion and have a profound negative impact on social stability and development.

[0003] To improve the credibility of information on social platforms, industry experts and scholars have begun actively exploring effective governance strategies to reduce the impact of false information. Knowledge guidance is one effective approach, leveraging external knowledge sources to analyze the discrepancies between the content of events to be detected on social platforms and real data to detect the authenticity of events. These knowledge sources include text and images obtained from other web pages, as well as online encyclopedias and fact-checking databases. Detection models include hierarchical multi-head attention networks, which combine text and evidence attention for false event detection; the CompareNet model, which detects fake news by comparing event information with external knowledge; and a preference-aware false event recognition framework, which combines pattern recognition and joint detection with external knowledge bases. These studies have shown that knowledge guidance can effectively improve the accuracy of false event detection.

[0004] Mainstream approaches in the field of false event detection rely on deep learning techniques to analyze the semantic content of events to determine authenticity. However, due to the limited number of small sample events, the model lacks effective prior knowledge during training, making it difficult to clearly determine which text features or descriptive information have a key impact on decision-making, resulting in insufficient interpretability. This interpretability issue reduces the credibility and reliability of related research, thereby hindering the practical application and promotion of the model.

[0005] Therefore, how to effectively use external knowledge bases for fact verification to improve the interpretability of the model's judgment when facing small sample events has become a key issue that needs to be urgently addressed in the current field of false event detection. Summary of the Invention

[0006] The present invention provides a method and device for detecting false information about events based on prior knowledge guidance, so as to solve the problem that due to the limited prior knowledge of small sample events, the existing false information detection methods based on knowledge guidance cannot obtain sufficient prior information, resulting in reduced credibility of the detection results.

[0007] The present invention is achieved through the following technical solutions:

[0008] A first aspect of the present invention provides a method for detecting false information about an event based on prior knowledge, comprising:

[0009] Get the description text and associated pictures of the event to be detected;

[0010] Extracting relation triples and text features from the description text, and extracting image features from the associated pictures;

[0011] A heterogeneous graph is constructed based on the relationship triples, entities are used as nodes of the heterogeneous graph, node feature vectors of each node are calculated, and the relationship triples are input into a pre-established relationship weight prediction model to obtain relationship weights; the relationship weight prediction model is trained based on the loss between the predicted value and the prior value of the known knowledge base;

[0012] Processing the node feature vectors and relationship weights using a graph convolutional network to obtain relationship features of the relationship triples;

[0013] Fusing the relationship features, text features, and image features to obtain fused features;

[0014] The fused features are input into a pre-trained discriminator to obtain a detection result of the event to be detected.

[0015] This invention optimizes triple relationship features by introducing an external knowledge base, ensuring that the model's judgment of event relationships is more accurate under small sample conditions. It also fuses graph structure features with multimodal features, achieving full expression of multimodal information and enhancing discriminative capabilities, further improving the model's ability to judge small sample events. Furthermore, because this invention focuses on knowledge guidance and multimodal information fusion, model training is not limited to data from specific fields or small sample events, expanding the application of false information detection in multiple fields and scenarios.

[0016] In one embodiment, extracting text features from the description text includes: extracting global features and local features from the description text using a Bert model, wherein the text features include concatenated text local features and text global features.

[0017] In one embodiment, extracting image features from the associated pictures includes: extracting global features and local features of the associated pictures using a VGG model, wherein the image features include spliced ​​local image features and global image features.

[0018] In one embodiment, the relationship features, text features, and image features are fused to obtain fused features, including:

[0019] Splicing the text features and image features to obtain multimodal joint features;

[0020] The multimodal joint feature and the relationship feature are fused to obtain a fused feature.

[0021] In one embodiment, fusing the multimodal joint feature and the relationship feature to obtain a fused feature includes:

[0022] Performing dimensionality reduction processing on the multimodal joint features through a multilayer perceptron to obtain a reduced-dimensional joint feature;

[0023] Processing the relationship features through global mean pooling to obtain global semantic features;

[0024] The dimensionality reduction joint feature and the global semantic feature are fused to obtain a fused feature.

[0025] In one embodiment, the node feature vector is calculated as:

[0026]

[0027] in, Represents entity node e i The node feature vector of Represents entity node e i a collection of indices of word positions in the text, Indicates the length of the set, S j The feature vector representing the corresponding position in the text.

[0028] In one embodiment, the relationship weight prediction model and the training are performed through a joint loss, and the joint loss includes a triple loss function and a cross-entropy loss function; the triple loss function is used to evaluate the difference between the predicted relationship weight of the relationship triple of the sample data and the relationship weight provided by the known knowledge base, and the cross-entropy function is used to evaluate the difference between the predicted result of the sample data and the annotation label.

[0029] A second aspect of the present invention provides an event false information detection system guided by prior knowledge, comprising:

[0030] The data acquisition module is used to obtain the description text and associated pictures of the event to be detected;

[0031] A feature extraction module is used to extract relationship triples and text features from the description text and extract image features from the associated pictures; it is also used to construct a heterogeneous graph based on the relationship triples, use entities as nodes of the heterogeneous graph, and calculate the node feature vector of each node;

[0032] A relationship weight prediction module, used to predict the relationship weight of the relationship triple;

[0033] The feature extraction module is further used to process the node feature vectors and relationship weights to obtain the relationship features of the relationship triples;

[0034] A multimodal fusion module, configured to fuse the relationship features, text features, and image features to obtain fused features;

[0035] The discriminator module is used to obtain the detection result of the event to be detected based on the fusion feature.

[0036] The third aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored on the memory and runnable on the processor, wherein when the processor executes the computer program, the method for detecting false event information based on prior knowledge guidance as described in any one of the first aspects of the present invention is implemented.

[0037] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the method for detecting false event information based on prior knowledge guidance as described in any one of the first aspects of the present invention is implemented.

[0038] Compared with the existing technology, the present invention has the following advantages and beneficial effects: a knowledge-guided event false information detection method is proposed, which adopts a heterogeneous graph model, external knowledge audit, and an optimized design of triple loss and cross entropy loss, effectively improving the generalization ability of the model and the accuracy of false information detection in small sample scenarios, enhancing the ability to integrate different modal data, and improving the credibility of detection results, which has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the examples. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. A person of ordinary skill in the art can also derive other relevant drawings based on these drawings without inventive effort. In the drawings:

[0040] Figure 1 This is a flow chart of a method for detecting false event information based on prior knowledge according to an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram illustrating the processing flow of a false information detection device according to an embodiment of the present invention;

[0042] Figure 3 It is a structural diagram of a false information detection system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the objects, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with embodiments and drawings, the schematic embodiments and their descriptions are only used to explain the present application, and do not limit the present application.

[0044] It should be noted that the terms "comprising" and "having" in the specification and claims of the present application and the above drawings are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units does not have to be limited to or inherent to other steps or units.

[0045] The terms used in various embodiments of the present application are only used for the purpose of describing specific embodiments and are not intended to limit various embodiments of the present application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which various embodiments of the present application belong. The terms (such as those defined in a commonly used dictionary) will be interpreted to have the same meaning as the contextual meaning in the relevant technical field and will not be interpreted to have an idealized or overly formal meaning, unless clearly defined in various embodiments of the present application.

[0046] The present application aims at the deficiencies in the prior art, and proposes an event false information detection method and device based on prior knowledge guidance, which is suitable for false message prediction based on prior information guidance, especially for small sample event detection, and is beneficial to improve the credibility and reliability of the detection result and the generalization ability of the detection model.

[0047] Referring to Figure 1 , Figure 2 as shown, Figure 1 is a flowchart of an event false information detection method based on prior knowledge guidance according to an embodiment of the present application, Figure 2 is a processing flowchart of a false information detection device, which includes the following steps.

[0048] S1, obtaining a description text and an associated picture of an event to be detected.

[0049] S2, extracting a triple relationship feature, a text feature and an image feature.

[0050] When extracting triples from text, the Bert model is used to embed the text and extract the relationship triples (entity, relationship, entity). The entities are used as node features in the graph model, and the node feature vectors and relationship weights of each entity node are calculated. The node feature vectors and relationship weights are then processed using a graph convolutional network (GCN) to obtain the relationship features of the relationship triples.

[0051] Conventional feature extraction models can be used for text features and image features. Text features and image features can be local features, global features, spliced ​​features of local features and global features, or fusion features in other ways.

[0052] In one embodiment, the Bert model is used to extract global features and local features of the description text to obtain text feature representations of different granularities, and the local features of the text and the global features of the text are concatenated to obtain the text features in S2.

[0053] For the input text content t, t={x1,x2...x m}, where m represents the number of words in the text, and the converted features are represented as S = {s1, s2...s m}, where s i is corresponding to x i The conversion features of the word are calculated by the pre-trained model Bert to obtain the fine-grained local features of the text content. The calculation formula is expressed as:

[0054] S={s1,s2...s m}=BERT(t)

[0055] For the global semantic information output by Bert, the calculation formula is expressed as:

[0056] h CLS =BERT(t)

[0057] Among them, h CLS Represents the global semantic information of the CLS tag corresponding to the Bert output. This vector is usually regarded as the global representation of the entire text t. Therefore, the global semantic information t output by BERT g It can be expressed as:

[0058] t g =h CLS

[0059] In one embodiment, a VGG (Visual Geometry Group) model is used to extract global features and local features of the associated image, and the image features in S2 are obtained by splicing the local features and the global features of the image.

[0060] For the input image content I, I is represented as a pixel matrix, and the converted feature is represented as F = {f1,f2,…,f n}, n represents the total number of pixels in the image, and f corresponds to pixel p i The pixel features are calculated by the pre-trained model VGG, which extracts the fine-grained local features of the image content. The calculation formula is expressed as:

[0061] F={f1,f2,…,f n}=VGG(I)

[0062] For the global semantic information output by VGG, the calculation formula is expressed as:

[0063] h global =VGG(I)

[0064] Among them, h global Represents the global semantic information output from VGG, and this vector is usually regarded as the global representation of the entire image I. Therefore, the global semantic information I output by VGG g Expressed as:

[0065] I g =h global

[0066] S3, relational features based on prior knowledge triples.

[0067] For triple relationship features, the entities in the relationship triples are used as nodes of the heterogeneous graph. The node features are obtained by calculating the mean of the feature vectors corresponding to the entity positions in the text. For entities e1 and e2 in the triples, the word positions in the original text are index sets and The node characteristics of the entity are calculated as follows:

[0068]

[0069] in, and Represent the feature vectors of the entity nodes e1 and e2 of the triple, and They correspond to the length of the entity node word set, S j and S k They represent the feature vectors of the corresponding characters in the text, and j and k represent the indexes of the corresponding characters.

[0070] The relationship features of triples are verified and optimized with the help of external prior knowledge to improve the model's accuracy in determining event relationships. Specifically, a relationship weight prediction model is built based on prior knowledge. The relationship triples are first input into a pre-built relationship weight prediction network for processing to obtain predicted relationship weights. These predicted relationship weights are then compared with the relationship weights in the known knowledge base to optimize the model's prediction of the relationship weights, thereby optimizing the triple relationship features.

[0071] First, the node feature vectors are spliced ​​and input into a neural network to establish a relationship weight prediction model to obtain the relationship weight W edge The expression is:

[0072]

[0073] in, Represents the concatenation operation of feature vectors, W nn is the weight matrix of the neural network, σ is the activation function, ensuring that the output value remains between (0,1), W edge Represents the predicted relationship weight.

[0074] The relation triplets extracted from S2 are fed into a large-scale external knowledge base, phi3, leveraging its powerful knowledge graph and relational reasoning capabilities to assess the credibility of relationships between entities. The relational features in the triples are further optimized and adjusted by comparing the relation weights predicted by the model with the weights of known relations in the knowledge base.

[0075] During model training, we use phi3's feedback to update relational features by calculating the triple loss between the predicted and query relations. Specifically, during model training, we optimize relational features through backpropagation, enabling the model to effectively learn the underlying relational patterns in the knowledge base, significantly improving the model's ability to understand and discriminate relations. The resulting relational weight labels are as follows:

[0076] R={r1,r2,…,r k}

[0077] Among them, R represents the set of relationship weight labels, r i Represents the weight label corresponding to each triple, and k represents the number of triplets contained in the event content.

[0078] The graph convolutional network is used to process the node feature vector and the relationship weight to obtain the relationship feature of the triple. Specifically, let the node feature be n is the number of nodes, and the weight matrix is ​​W edge , then the calculation formula of GCN can be expressed as:

[0079]

[0080] Among them, H (l) Represents the node feature matrix of the lth layer, which is obtained by concatenating V. It is a normalized adjacency matrix, which contains the connection relationship and weight information W between nodes edge , W (l) is the learnable weight of the lth layer, and σ is the activation function. After a layer of GCN, the implicit feature H of the node is obtained (1) .

[0081] S4, fuses the triple's relational features, text features, and image features to obtain fused features.

[0082] First, the text features are spliced ​​with the global image features to generate a multimodal joint feature representation, which is then combined with the triple relationship features to obtain the final fusion feature.

[0083] Furthermore, before fusion, the joint feature representation is reduced in dimension through a multi-layer perceptron to obtain a more compact feature representation, which is then fused with the triple relationship feature.

[0084] Specifically, the global feature t extracted by Bert g Global features extracted from VGG I g Splice to form a joint feature vector C, expressed as:

[0085] C=[t g ;I g ]

[0086] In order to improve the effectiveness of feature representation and reduce computational complexity, a multi-layer perceptron (MLP) neural network is used to reduce the dimension of the concatenated joint feature c to obtain a more compact and discriminative feature representation c′. The calculation formula can be expressed as:

[0087] c′=MLP(c)

[0088] Ultimately, the obtained feature representation C′ will be used in the subsequent event false information detection model to provide stronger support for the accuracy and robustness of the model.

[0089] For relational features, global mean pooling is performed to obtain global semantic features, which are then fused with the multimodal joint features after dimensionality reduction.

[0090] To extract global semantic information, the calculation formula is:

[0091] h global =global_mean_pool(H (l+1) )

[0092] Among them, hglobal Represents the global semantic information output by GCN, and global_mean_pool represents the global mean pooling process.

[0093] Next, the joint feature representation C′ after dimensionality reduction is combined with the global semantic information h extracted by GCN global Direct addition is performed to obtain the final feature representation:

[0094]

[0095] S5, input the fused features into the pre-trained discriminator to obtain the detection result of the event to be detected.

[0096] The discriminator is trained by event information that has been marked as false news. In the present invention, the discriminator uses a neural network model to perform binary classification.

[0097] In one embodiment, when jointly training the relationship weight prediction model and the discriminator, the loss function adopts the joint loss of the triple loss function and the cross entropy loss function, wherein the triple loss is used to optimize the model's prediction of triple relationships, and the cross entropy loss is used to constrain the judgment of the authenticity of events. By introducing the joint loss of the triple loss function and the cross entropy loss function, the model is optimized to improve the accuracy and performance of relationship prediction and true and false event discrimination, and ultimately a trained multimodal small sample event false information detection model is obtained. The content description text and associated images of the event to be detected are input into the trained small sample event false information detection model to obtain the small sample event false information detection results.

[0098] The detailed training process is described as follows.

[0099] S51, the relationship loss is obtained by calculating the difference between the relationship edge weight predicted by the edge weight prediction model and the relationship weight provided in the external knowledge base.

[0100] Specifically, given the relationship edge weight W predicted by the model edge The relationship weight labels R provided in the external knowledge base are calculated by calculating the distance between these weights to obtain the relationship loss. The loss is calculated as follows: for each sample, if its label weight is greater than 0.7, it is considered a positive sample, and samples with label weights less than or equal to 0.7 are considered negative samples. For each pair of positive and negative samples, the distance between their weights is calculated and evaluated using the ternary loss formula. This ensures that the distance between positive samples is smaller than that between negative samples. A margin is added, here set to 1.0, to enhance the model's discriminative ability and improve prediction accuracy:

[0101] loss=max(0,d pos -dneg +margin)

[0102] where d pos represents the distance between the positive sample and its label weight, d neg represents the distance between the negative sample and its label weight. Finally, the final relationship loss is obtained by taking the average of all sample losses, denoted as:

[0103]

[0104] This relationship loss is used to optimize the model parameters, N represents the number of samples, and improve the accuracy and robustness of the model in the relationship weight prediction task.

[0105] S52, by comparing the prediction results of true and false events with the labeled labels, the cross-entropy loss is calculated.

[0106] According to the small sample event false information detection result and the true label, the content true and false classification loss is calculated, and the calculation formula is represented as:

[0107]

[0108] where N represents the number of samples, represents the input joint feature representation, y i is the i-th prediction value output by the discriminator.

[0109] S53, the relationship loss and the cross-entropy loss are weighted and summed to obtain the final loss function.

[0110] Specifically, the relationship loss measures the difference between the edge weight predicted by the model and the relationship weight provided in the external knowledge base, ensuring that the model can more accurately depict the relationship between events. The cross-entropy loss is used to measure the accuracy of the model in event true and false judgment, ensuring that the model can effectively distinguish between false and true events under the fusion of multi-modal information.

[0111] The final loss function is obtained by weighting and summing these two parts of the loss with certain weight coefficients, and its expression is:

[0112]

[0113] where λ represents an adjustment coefficient, which is a hyperparameter for adjusting the balance of loss terms.

[0114] Embodiments of the application also provide an event false information detection device based on prior knowledge guidance, which is used to execute the event false information detection method based on prior knowledge guidance of the above-mentioned embodiments. The device comprises:

[0115] A data acquisition module is configured to acquire a description text and an associated picture of an event to be detected.

[0116] A feature extraction module is used to extract relationship triples and text features from the description text and extract image features from the associated pictures; it is also used to construct a heterogeneous graph based on the relationship triples, use entities as nodes of the heterogeneous graph, and calculate the node feature vector of each node;

[0117] A relationship weight prediction module, used to predict the relationship weight of the relationship triple;

[0118] The feature extraction module is further used to process the node feature vectors and relationship weights to obtain the relationship features of the relationship triples;

[0119] The discriminator module is used to obtain the detection result of the event to be detected based on the fusion feature.

[0120] Furthermore, the present invention also provides an event false information detection system based on prior knowledge guidance, such as Figure 3 The figure shows a schematic diagram of the structure of a false information detection system, which includes the above-mentioned detection device, as well as a user input module and a detection result visualization module. The user input module receives data input by the user, including information about the event to be detected, detection time training samples and other graphic data, as well as instructions input by the user, such as instructions for modifying system parameters, model training setting parameters, etc. Based on the acquired event information and instructions, the system executes the above-mentioned training of the event false information detection method or detection device guided by prior knowledge. The output results are displayed through the visualization module. The detection results include an event structure diagram, relationship verification information and the final true or false judgment. Users can intuitively see how the system processes input data, how it builds a relationship network, how it verifies facts, and quickly understand the detection conclusions. At the same time, key information in the detection process, such as entities, relationships, image features, etc., is also presented through a graphical interface to help users gain a deeper understanding of the detection process.

[0121] An embodiment of the present invention further provides an electronic device comprising a processor and a memory, wherein the number of processors may be one or more. The memory, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules. The processor executes the software programs, instructions, and modules stored in the memory to perform various functional applications and data processing of the electronic device, thereby implementing the method for detecting false event information based on prior knowledge guidance according to any of the above embodiments of the present invention.

[0122] The memory can include a program storage area and a data storage area. The program storage area can store an operating system and applications required for at least one function. The data storage area can store data created according to use of the terminal, etc. In addition, the memory can include a high-speed random access memory, and can further include a nonvolatile memory such as at least one of a magnetic disk storage device, a flash memory device, or other nonvolatile solid state memory device. In some examples, the memory can further include a memory remotely located with respect to the processor, which can be connected to the electronic device through a network. Examples of the network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0123] An embodiment of the present application further provides a computer readable storage medium, having stored thereon a computer program which, when executed by a processor, implements the event false information detection method based on prior knowledge guidance according to any one of the embodiments of the present application.

[0124] An embodiment of the present application further provides a computer program product, which, when running on a computer, causes the computer to execute the event false information detection method based on prior knowledge guidance according to any one of the embodiments of the present application.

[0125] The above detailed description merely describes specific implementation of the present application, and is not intended for limiting the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is merely a specific implementation of the present application, and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for detecting false event information based on prior knowledge, characterized in that: include: Get the description text and associated pictures of the event to be detected; Extracting relation triples and text features from the description text, and extracting image features from the associated pictures; A heterogeneous graph is constructed based on the relationship triples, and the entities are used as nodes of the heterogeneous graph. The node feature vector of each node is calculated, and the relationship triples are input into a pre-established relationship weight prediction model to obtain the relationship weight; the relationship weight prediction model is obtained based on the loss training of the predicted value and the prior value of the known knowledge base, specifically: The node feature vectors are spliced ​​and input into the neural network to establish a relationship weight prediction model. Expressed as: in, represents the concatenation operation of feature vectors, and Represent entity nodes respectively and The eigenvector of and are two entities in the relation triple, is the weight matrix of the neural network, Is the activation function, ensuring that the output value remains between (0,1); Processing the node feature vectors and relationship weights using a graph convolutional network to obtain relationship features of the relationship triples; Fusing the relationship features, text features, and image features to obtain fused features; The fused features are input into a pre-trained discriminator to obtain a detection result of the event to be detected.

2. The method for detecting false event information based on prior knowledge according to claim 1, characterized in that: Extracting text features from the description text includes: extracting global features and local features from the description text using a Bert model, wherein the text features include spliced ​​text local features and text global features.

3. The method for detecting false event information based on prior knowledge according to claim 1, characterized in that: Extracting image features from the associated pictures includes: extracting global features and local features of the associated pictures using a VGG model, wherein the image features include spliced ​​local features of the image and global features of the image.

4. The method for detecting false event information based on prior knowledge according to claim 1, characterized in that: The relationship features, text features and image features are fused to obtain fused features, including: Splicing the text features and image features to obtain multimodal joint features; The multimodal joint feature and the relationship feature are fused to obtain a fused feature.

5. The method for detecting false event information based on prior knowledge according to claim 4 is characterized in that: The multimodal joint feature and the relationship feature are fused to obtain a fused feature, including: Performing dimensionality reduction processing on the multimodal joint features through a multilayer perceptron to obtain a reduced-dimensional joint feature; Processing the relationship features through global mean pooling to obtain global semantic features; The dimensionality reduction joint feature and the global semantic feature are fused to obtain a fused feature.

6. The method for detecting false event information based on prior knowledge according to claim 1, characterized in that: The node feature vector is calculated as: in, Represents an entity node The node feature vector of Represents an entity node a collection of indices of word positions in the text, , Indicates the length of the collection, The feature vector representing the corresponding position in the text.

7. The method for detecting false event information based on prior knowledge according to claim 1, characterized in that: The relationship weight prediction model and the discriminator are trained through a joint loss, and the joint loss includes a triple loss function and a cross-entropy loss function; the triple loss function is used to evaluate the difference between the predicted relationship weight of the relationship triple of the sample data and the relationship weight provided by the known knowledge base, and the cross-entropy loss function is used to evaluate the difference between the predicted result of the sample data and the annotation label.

8. An event false information detection device based on prior knowledge guidance, characterized in that: include: The data acquisition module is used to obtain the description text and associated pictures of the event to be detected; A feature extraction module is used to extract relationship triples and text features from the description text and extract image features from the associated pictures; it is also used to construct a heterogeneous graph based on the relationship triples, use entities as nodes of the heterogeneous graph, and calculate the node feature vector of each node; Relationship weight prediction module, used to predict the relationship weight of the relationship triples, relationship weight Expressed as: in, represents the concatenation operation of feature vectors, and Represent entity nodes respectively and The eigenvector of and are two entities in the relation triple, is the weight matrix of the neural network, Is the activation function, ensuring that the output value remains between (0,1); The feature extraction module is further used to process the node feature vectors and relationship weights to obtain the relationship features of the relationship triples; A multimodal fusion module, configured to fuse the relationship features, text features, and image features to obtain fused features; The discriminator module is used to obtain the detection result of the event to be detected based on the fusion feature.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for detecting false event information based on prior knowledge guidance according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for detecting false event information based on prior knowledge guidance according to any one of claims 1 to 7 is implemented.

Images