Rumor detonation point detection method and device, computer device and medium

Abstract

The application relates to the field of deep learning and is applied to rumor ignition point detection in the fields of finance, medical treatment and the like. The application discloses a rumor ignition point detection method, which comprises the following steps: acquiring a data set, wherein the data set comprises node texts participating in rumor propagation and relationships between nodes; pretraining the data set by using a Bert language representation model to obtain a vector matrix of the node texts and an adjacency matrix based on the relationships between the nodes; inputting the vector matrix and the adjacency matrix into a first graph convolutional neural network for calculation to obtain a first high-order representation vector matrix; transposing the adjacency matrix to obtain a transposed adjacency matrix; inputting the vector matrix and the transposed adjacency matrix into a second graph convolutional neural network for calculation to obtain a second high-order representation vector matrix; and inputting the high-order representation vector matrices of the nodes into a full connection layer of a third graph convolutional neural network for classification to obtain classification results of the nodes.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a rumor detonation point detection method and device, computer equipment and medium. BACKGROUND

[0002] In recent years, the rapid development of the Internet has promoted the widespread use of social media, and online platforms such as microblogs and Twitter have become important media for users to obtain information and communicate with each other. Users of these platforms express different opinions and views on various hot issues, and at the same time, various rumors have also appeared. For example, some rumors about the financial field are widely and rapidly spread on social media, not only causing great harm to society, but also causing a large amount of economic loss; or rumors about the medical field can even endanger the safety of patients' lives.

[0003] Rumor detonation point detection belongs to the research category of Internet information credibility and is a new direction of Internet information credibility research. Current technology mainly uses rumor propagation trees for rumor detection. This group intelligence method has achieved leading results in the field of rumor detection.

[0004] This method constructs the rumor propagation process into a tree structure, uses the text features of events and user comments as the nodes of the tree, and then uses a graph neural network to learn the representation of the rumor propagation tree to build a rumor classifier, and then determines whether each node in the rumor event is a detonation point. If it is a detonation point, determine its detonation point type, such as a certain forwarding causing a huge discussion, which can be considered as a detonation point, and then determine how it triggers subsequent discussions. The current typical method of this kind is to carefully design a specific encoder or training mode for the rumor detection task. This method belongs to the combination of natural language processing and graph neural networks.

[0005] However, the existing rumor detonation point classification method does not consider the relationship between nodes, has poor performance, and cannot accurately and effectively determine the rumor detonation point. SUMMARY

[0006] The main purpose of the present application is to provide a rumor detonation point detection method, device, computer equipment and medium, which aims to solve the technical problem of the existing rumor detonation point classification method not considering the relationship between nodes and having poor performance.

[0007] The present application discloses the following technical solutions:

[0008] A rumor detonation point detection method, comprising:

[0009] Obtaining a data set, wherein the data set contains node text participating in rumor propagation and the relationship between nodes;

[0010] pre-training the data set by using a Bert language representation model to obtain a vector matrix of the node text and an adjacency matrix based on the relationship between each node;

[0011] inputting the vector matrix and the adjacency matrix into a first graph convolutional neural network for calculation to obtain a first high-order representation vector matrix;

[0012] transposing the adjacency matrix to obtain a transposed adjacency matrix;

[0013] inputting the vector matrix and the transposed adjacency matrix into a second graph convolutional neural network for calculation to obtain a second high-order representation vector matrix;

[0014] splicing the first high-order representation vector matrix and the second high-order representation vector matrix to obtain a high-order representation vector matrix of each node;

[0015] inputting the high-order representation vector matrix of each node into a fully connected layer of a third graph convolutional neural network for classification to obtain a classification result of each node, wherein the classification label is divided into four types: opening a topic, opposing and questioning, clarifying and explaining, and no substantive meaning;

[0016] determining a rumor detonation point based on the classification result of each node.

[0017] Further, the step of inputting the vector matrix and the adjacency matrix into a first graph convolutional neural network for calculation to obtain a first high-order representation vector matrix comprises:

[0018] inputting the vector matrix and the adjacency matrix into a first layer of the first graph convolutional neural network for calculation by using the following formula to obtain a first calculation result:

[0019] H 11 =GCN 11 (X,A)

[0020] wherein: H 11 is the first calculation result, X is the vector matrix, and A is the adjacency matrix;

[0021] inputting the first calculation result and the adjacency matrix into a second layer of the first graph convolutional neural network for calculation by using the following formula to obtain a first high-order representation vector matrix:

[0022] H 12 =GCN 12 (H 11 ,)

[0023] wherein: H 12 is the first high-order representation vector matrix.

[0024] Further, the step of inputting the vector matrix and the adjacency matrix into a first layer of the first graph convolutional neural network for calculation to obtain a first calculation result, comprises:

[0025] According to the adjacency matrix, a first diagonal matrix is obtained by using the following formula:

[0026]

[0027] Wherein: A ij is the adjacency matrix, d ii is the first diagonal matrix, i and j are the number of rows and columns of the matrix, and b is the number of columns of the adjacency matrix;

[0028] According to the first diagonal matrix, the adjacency matrix is normalized by using the following formula:

[0029]

[0030] Wherein: is the normalized adjacency matrix, I is the unit matrix, and D is the first diagonal matrix;

[0031] The normalized adjacency matrix and the vector matrix are substituted into an activation function for calculation by using the following formula to obtain a first calculation result:

[0032]

[0033] Wherein: W 11 is the weight matrix of the first layer of the first graph convolutional neural network.

[0034] Further, the step of inputting the vector matrix and the transposed adjacency matrix into a second graph convolutional neural network for calculation to obtain a second high-order representation vector matrix, comprises:

[0035] The vector matrix and the transposed adjacency matrix are input into a first layer of the second graph convolutional neural network for calculation to obtain a second calculation result by using the following formula:

[0036] H 21 = GCN 21 (X, A T )

[0037] Wherein: H 21 is the second calculation result, X is the vector matrix, and A T is the transposed adjacency matrix;

[0038] The second calculation result, the transposed adjacency matrix are input into a second layer of the second graph convolutional neural network to obtain a second high-order representation vector matrix by using the following formula:

[0039] H 22 =GCN 22 (H 21 ,A T )

[0040] wherein H 22 is the second high-order representation vector matrix.

[0041] Further, the step of inputting the vector matrix and the transposed adjacency matrix into a first layer of the second graph convolutional neural network to obtain a second calculation result by using the following formula comprises:

[0042] According to the transposed adjacency matrix, a second diagonal matrix is obtained by using the following formula:

[0043]

[0044] wherein: is the transposed adjacency matrix, d′ ii is the second diagonal matrix, i and j are the number of rows and columns of the matrix, and n is the number of columns of the transposed adjacency matrix;

[0045] According to the second diagonal matrix, the adjacency matrix is normalized by using the following formula:

[0046]

[0047] According to: is the normalized transposed adjacency matrix, I is an identity matrix, and D′ is the second diagonal matrix;

[0048] The first calculation result is obtained by using the following formula to input the normalized adjacency matrix and the vector matrix into an activation function:

[0049]

[0050] wherein W 21 is a weight matrix of the first layer of the second graph convolutional neural network.

[0051] Further, the step of inputting the high-order representation vector matrix of each node into a fully connected layer of a third graph convolutional neural network to obtain a classification result of each node comprises:

[0052] The high-order representation vector matrix of each node is substituted into a softmax function for calculation to obtain the probability of each node in four classification labels;

[0053] The probabilities of each node in the four classification labels are combined to obtain the classification result of each node.

[0054] Further, the step of combining the probabilities of each node in the four classification labels to obtain the classification result of each node comprises:

[0055] The probabilities of each node in the four classification labels are substituted into a loss function using the following formula to calculate the gap between the probability distribution and the true distribution of each node in the four classification labels:

[0056]

[0057] wherein y1 is the true distribution of each node in the classification label of starting a topic, is the probability distribution of each node in the classification label of starting a topic, y2 is the true distribution of each node in the classification label of opposing and questioning, is the probability distribution of each node in the classification label of opposing and questioning, y3 is the true distribution of each node in the classification label of clarifying and explaining, is the probability distribution of each node in the classification label of clarifying and explaining, y4 is the true distribution of each node in the classification label of having no substantive meaning, is the probability distribution of each node in the classification label of having no substantive meaning, and alpha1 is the weight of the classification label of starting a topic, alpha2 is the weight of the classification label of opposing and questioning, alpha3 is the weight of the classification label of clarifying and explaining, and alpha4 is the weight of the classification label of having no substantive meaning.

[0058] According to the gap between the probability distribution and the true distribution of each node in the four classification labels, the first graph convolutional neural network, the second graph convolutional neural network and the third graph convolutional neural network are trained and fitted to optimize the parameters of the first graph convolutional neural network, the second graph convolutional neural network and the third graph convolutional neural network.

[0059] The application also provides a rumor explosion point detection device, which comprises:

[0060] An acquisition module is configured to acquire a data set, wherein the data set contains node texts participating in rumor transmission and relationships between nodes;

[0061] A pre-training module is configured to pre-train the data set by using a Bert language representation model to obtain a vector matrix of the node texts and an adjacency matrix based on the relationships between nodes;

[0062] The first calculation module is configured to input the vector matrix and the adjacency matrix into a first graph convolutional neural network for calculation to obtain a first high-order representation vector matrix.

[0063] The adjacency matrix transposition module is configured to transpose the adjacency matrix to obtain a transposed adjacency matrix.

[0064] The second calculation module is configured to input the vector matrix and the transposed adjacency matrix into a second graph convolutional neural network for calculation to obtain a second high-order representation vector matrix.

[0065] The high-order representation vector matrix calculation module is configured to splice the first high-order representation vector matrix and the second high-order representation vector matrix to obtain a high-order representation vector matrix of each node.

[0066] The classification module is configured to input the high-order representation vector matrix of each node into a fully connected layer of a third graph convolutional neural network for classification to obtain a classification result of each node, wherein the classification label is divided into four types: opening a topic, opposing a question, clarifying an explanation, and no substantive meaning.

[0067] The rumor ignition point determination module is configured to determine a rumor ignition point based on the classification result of each node.

[0068] The application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method according to any one of the preceding embodiments when executing the computer program.

[0069] The application also provides a computer readable storage medium, which stores a computer program, and the computer program implements the steps of the method according to any one of the preceding embodiments when executed by a processor.

[0070] Advantages:

[0071] In the application, the data set is pre-trained by using a Bert language representation model, so that the node text and the relationship between nodes can be more effectively represented. Two double-layer graph convolutional neural networks with the same structure but independently trained parameters are used to calculate the data set, so that the rumor propagation and diffusion process can be effectively modeled, the relationship between nodes can be modeled, and the rumor ignition point can be more accurately and effectively determined. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 is a flowchart of a rumor ignition point detection method according to an embodiment of the application;

[0073] Figure 2 is a structural schematic block diagram of a rumor ignition point detection device according to an embodiment of the application;

[0074] Figure 3 is a structural schematic block diagram of a computer device according to an embodiment of the present application.

[0075] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0076] It should be understood that the specific embodiments described herein are merely exemplary and are not intended to limit the present application.

[0077] With reference to Figure 1 , the present application provides an embodiment of a rumor detonation point detection method, comprising:

[0078] S1: obtaining a data set, wherein the data set contains node texts participating in rumor propagation and relationships between nodes;

[0079] S2: pre-training the data set using a Bert language representation model to obtain a vector matrix of the node texts and an adjacency matrix based on the relationships between the nodes;

[0080] S3: inputting the vector matrix and the adjacency matrix into a first graph convolutional neural network for calculation to obtain a first high-order representation vector matrix;

[0081] S4: transposing the adjacency matrix to obtain a transposed adjacency matrix;

[0082] S5: inputting the vector matrix and the transposed adjacency matrix into a second graph convolutional neural network for calculation to obtain a second high-order representation vector matrix;

[0083] S6: concatenating the first high-order representation vector matrix and the second high-order representation vector matrix to obtain a high-order representation vector matrix of each node;

[0084] S7: inputting the high-order representation vector matrix of each node into a fully connected layer of a third graph convolutional neural network for classification to obtain a classification result of each node, wherein the classification labels are divided into four types: starting a topic, opposing and questioning, clarifying, and no substantive meaning;

[0085] S8: determining a rumor detonation point based on the classification result of each node.

[0086] In the above embodiment, the data set is pre-trained using a Bert language representation model, which can more effectively represent the node texts and the relationships between the nodes. Moreover, two double-layer graph convolutional neural networks with the same structure but independently trained parameters are used to calculate the data set, which can effectively model the propagation and diffusion process of rumors and the connection between nodes, and more accurately and effectively determine the detonation point of rumors.

[0087] Social networks are currently important information dissemination media, however, its relatively free and open speaking environment, also helps the rumor spread. For example, some rumors in the field of finance are widely and rapidly spread on social media, not only bring great harm to society, but also cause a lot of economic losses; or rumors about medical aspects can even endanger the safety of patients' lives. Current research believes that rumors in social networks are self-justifying with public participation, that is, the comments or forwarded contents generated in the transmission process can determine the authenticity of the message.

[0088] The present application is based on the exploration of rumor transmission mechanism in social science research, and constructs a rumor trigger identification task. The information trigger point refers to the message that has a significant driving effect on rumor transmission or has a clear indication on rumor judgment, which is divided into the following four categories: (1) amplify, which refers to the message that introduces new ideas or expands the scope of discussion; (2) deny, which refers to the message that opposes or questions the previous message; (3) clarify, which refers to the message that contains factual basis to explain the idea; (4) null, which refers to the message without indication.

[0089] As described in the above step S1, a data set is obtained, wherein the data set contains node texts participating in rumor transmission and relationships between nodes;

[0090] Sina Weibo (hereinafter referred to as Weibo) is an online social media platform serving the Chinese and global Chinese community. Users can connect and use through various mobile terminals to share, disseminate and interact information in the form of text, pictures, videos and other multimedia. Weibo has a large number of users in China, with a monthly active user number close to 600 million. It is necessary to study the automatic detection method of Weibo rumors. The method in this paper mainly faces the automatic detection of Weibo rumors, and the data used are all from the Weibo platform.

[0091] The data set contains all the information participating in the rumor transmission event, which contains node texts participating in rumor transmission and relationships between nodes. The nodes of the data set include source nodes, child nodes participating in the event comments and forwarding, each node is a Weibo post, the source node is the Weibo post that initially publishes the rumor; the child nodes participating in the event comments and forwarding are the Weibo that forwards the rumor Weibo, and the source node is the father node of these forwarding child nodes; the relationship between nodes is the forwarding information between nodes, which is the forwarding relationship between different nodes, for example, A forwards B's microblog, the relationship between A and B is B→A, then B forwards C's microblog, the relationship between A, B and C is C→B→A.

[0092] Specifically, one rumor event in the data set is represented as:

[0093] Source node: own number 1 text 1 father node number (the source node has no father node, so this place can be regarded as no number)

[0094] Node 1: own number 2 text 2 father node number

[0095] Node 2: own number 3 text 3 father node number

[0096] Node 3: own number 4 text 4 father node number…

[0097] In the data set, the propagation process of the rumor is constructed into a tree structure, and the text features of the event and user comments are taken as the nodes of the tree, and the reply relationship of the post is taken as the edge of the tree.

[0098] As described in step S2 above, the data set is pre-trained using the Bert language representation model to obtain a vector matrix of the node text and an adjacency matrix based on the relationship between nodes;

[0099] The full name of BERT is Bidirectional Encoder Representation from Transformer (Bidirectional Encoder Representation from Transformer), that is, BERT is an encoder based on the Transformer model. BERT is a context-based embedding model, unlike other embedding models such as word2vec, which are context-free. The Bert language representation model understands the context of each word in the node text through the multi-head attention mechanism, and then outputs the embedding vector of each word, and then each word is composed into a vector matrix according to the order of the node text. The relationship between nodes is represented using an adjacency matrix.

[0100] As described in step S3, the vector matrix and the adjacency matrix are input into the first graph convolutional neural network for calculation to obtain a first high-order representation vector matrix;

[0101] The graph convolutional neural network is a feature extractor, which processes graph data. In the present application, the graph data is the tree graph formed by the text of the nodes participating in rumor propagation and the relationship between the nodes. The graph convolutional neural network designs a method for extracting features from graph data, so that these features can be used for node classification, graph classification, link prediction of graph data, and also obtain the embedding representation of the graph. In the present application, the first graph convolutional neural network can be used to classify the nodes of the rumor tree graph.

[0102] The first graph convolutional neural network calculates the vector matrix and the adjacency matrix to obtain a first high-order representation vector matrix.

[0103] As described in steps S4 and S5, the adjacency matrix is transposed to obtain a transposed adjacency matrix.

[0104] The vector matrix and the transposed adjacency matrix are input into the second graph convolutional neural network for calculation to obtain a second high-order representation vector matrix.

[0105] The second graph convolutional neural network and the first graph convolutional neural network are two graph convolutional neural networks with the same structure but different parameters, wherein the parameters of the second graph convolutional neural network and the first graph convolutional neural network are trained separately. The transposed adjacency matrix is obtained by transposing the adjacency matrix, and then the transposed adjacency matrix and the vector matrix are input into the second graph convolutional neural network for calculation.

[0106] As described in steps S6 and S7, the first high-order representation vector matrix and the second high-order representation vector matrix are spliced to obtain a high-order representation vector matrix of each node.

[0107] The high-order representation vector matrix of each node is input into the fully connected layer of the third graph convolutional neural network for classification to obtain the classification result of each node, wherein the classification label is divided into four types: amplify, deny, clarify, and null.

[0108] The second and first convolutional neural networks compute the first and second higher-order representation vector matrices for each node in the dataset. After concatenating these matrices, the higher-order representation vector matrices for each node are obtained. These matrices are then processed by the fully connected layers of the third convolutional neural network to classify the nodes and obtain the classification results. Specifically, the third convolutional neural network has the same structure as the second and first convolutional neural networks.

[0109] An "amplify" message introduces a new perspective or broadens the scope of discussion; when a node is categorized as an "amplify," it signifies that the node is the trigger point for a rumor. A node categorized as a "deny" message indicates that the node raises objections or questions about the rumor. A node categorized as a "clarify" message indicates that the node provides a factual explanation and clarification of the rumor. A node categorized as "null" message indicates that the node contains no indicative message.

[0110] As described in step S8 above, the rumor trigger point is determined based on the classification results of each node.

[0111] When a node is identified as a tag that initiates a topic, it is considered a rumor tipping point. For example, this could be a tipping point for rumors in the financial or medical fields. After identifying these tipping points, the relevant Weibo posts can be deleted, the associated accounts banned, or offline investigations conducted. This helps prevent users from suffering losses to their lives and property.

[0112] In one embodiment, step S3, which involves inputting the vector matrix and the adjacency matrix into a first graph convolutional neural network for calculation to obtain a first higher-order representation vector matrix, includes:

[0113] S301: Using the following formula, input the vector matrix and the adjacency matrix into the first layer of the first graph convolutional neural network for calculation to obtain the first calculation result:

[0114] H 11 =GCN 11 (X,A)

[0115] Wherein: H 11 The first calculation result is given by X, where X is the vector matrix and A is the adjacency matrix.

[0116] S302: input the first calculation result and the adjacency matrix into a second layer of the first graph convolutional neural network to calculate, using the following formula, to obtain a first high-order representation vector matrix:

[0117] H 12 = GCN 12 (H 11 , A)

[0118] wherein H 12 is the first high-order representation vector matrix.

[0119] In the above embodiment, the first graph convolutional neural network has two layers, and the two-layer graph convolutional neural network is used. Compared with one layer, the model is deeper, the structure is more complex, and the classification result is more accurate.

[0120] In an embodiment, the step S301 of inputting the vector matrix and the adjacency matrix into a first layer of the first graph convolutional neural network to calculate, using the following formula, to obtain a first calculation result comprises:

[0121] S311: according to the adjacency matrix, calculate using the following formula to obtain a first diagonal matrix:

[0122]

[0123] wherein A ij is the adjacency matrix, d ii is the first diagonal matrix, i and j are the number of rows and columns of the matrix, and n is the number of columns of the adjacency matrix.

[0124] S312: according to the first diagonal matrix, normalize the adjacency matrix using the following formula:

[0125]

[0126] wherein: is the normalized adjacency matrix, I is the unit matrix, and D is the first diagonal matrix.

[0127] S313: input the normalized adjacency matrix and the vector matrix into an activation function to calculate, using the following formula, to obtain a first calculation result:

[0128]

[0129] wherein W 11 is a weight matrix of the first layer of the first graph convolutional neural network.

[0130] In the above embodiment, as described in step S311, the first diagonal matrix is a degree matrix, the adjacency matrix represents the relationship between nodes, and the first diagonal matrix represents the number of edges connected to the node. The first diagonal matrix can be obtained by substituting the adjacency matrix into the formula.

[0131] As described in step S312, according to the first diagonal matrix, the adjacency matrix is normalized by using the following formula:

[0132]

[0133] wherein: is the normalized adjacency matrix, I is the unit matrix, and D is the first diagonal matrix;

[0134] If the matrix is not normalized, the adjacency matrix and the vector matrix X are directly multiplied, which means that the influence of the dimension of the vectors between nodes is not eliminated, which will change the distribution of the original node characteristics and greatly affect the final effect. Therefore, normalization or standardization is needed to limit all data within a reasonable range and eliminate the adverse effects of singular sample data. Balance the importance of nodes with large degrees.

[0135] Before normalization, the adjacency matrix is added to a unit matrix. At this time, if only the adjacency matrix A is used, the diagonal elements are all 0, so after the adjacency matrix A and the vector matrix X are multiplied, only the weighted sum of the features of all neighbor nodes of the node is calculated, and the features of the node itself are ignored. The unit matrix is a square matrix, and the elements on the diagonal from the top left corner to the bottom right corner (called the main diagonal) are all 1. Except for this, all are 0. In this way, the diagonal elements of the adjacency matrix will become 1.

[0136] As described in step S313, the normalized adjacency matrix and the vector matrix are substituted into the activation function to calculate the first calculation result by using the following formula:

[0137]

[0138] wherein: W 11 is the weight matrix of the first layer of the first graph convolutional neural network.

[0139] ReLU represents the activation function, represents the extracted features of each node after passing through the first layer of the first graph convolutional neural network.

[0140] In an embodiment, the step S5 of inputting the vector matrix and the transposed adjacency matrix into the second graph convolutional neural network to calculate a second high-order representation vector matrix comprises:

[0141] S501: input the vector matrix and the transposed adjacency matrix into the first layer of the second graph convolutional neural network for calculation by using the following formula to obtain a second calculation result:

[0142] H 21 = GCN 21 (X, A T )

[0143] wherein H 21 is the second calculation result, X is the vector matrix, and A T is the transposed adjacency matrix;

[0144] S502: input the second calculation result and the transposed adjacency matrix into the second layer of the second graph convolutional neural network for calculation by using the following formula to obtain a second high-order representation vector matrix:

[0145] H 22 = GCN 22 (H 21 , A T )

[0146] wherein H 22 is the second high-order representation vector matrix.

[0147] In the above embodiment, the second graph convolutional neural network has two layers, and the two-layer graph convolutional neural network is used. Compared with one layer, the model is deeper, the structure is more complex, and the classification result is more accurate.

[0148] In an embodiment, the step S501 of inputting the vector matrix and the transposed adjacency matrix into the first layer of the second graph convolutional neural network for calculation by using the following formula to obtain a second calculation result comprises:

[0149] S511: calculate a second diagonal matrix according to the transposed adjacency matrix by using the following formula:

[0150]

[0151] wherein A ii is the transposed adjacency matrix, d′ 21 is the second diagonal matrix, i and j are the row number and column number of the matrix, and n is the column number of the transposed adjacency matrix;

[0152] S512: normalize the adjacency matrix according to the second diagonal matrix by using the following formula:

[0153]

[0154] According to: is a unit matrix, and D' is the second diagonal matrix;

[0155] S513: The normalized adjacency matrix and the vector matrix are substituted into an activation function for calculation by using the following formula to obtain a first calculation result:

[0156]

[0157] wherein W 21 is a weight matrix of the first layer of the second graph convolutional neural network.

[0158] In the above embodiment, the second graph convolutional neural network and the first graph convolutional neural network have the same structure, and thus the vector matrix and the transposed adjacency matrix are input into the first layer of the second graph convolutional neural network for calculation to obtain the second calculation result. The calculation of the first layer of the first graph convolutional neural network is similar, and specific explanations are given in the above embodiment.

[0159] ReLU represents an activation function, represents the extracted features of each node after passing through the first layer of the second graph convolutional neural network.

[0160] In an embodiment, the step S7 of inputting the high-order representation vector matrix of each node into a fully connected layer of a third graph convolutional neural network for classification to obtain a classification result of each node comprises:

[0161] S701: The high-order representation vector matrix of each node is substituted into a softmax function for calculation to obtain the probability of each node belonging to four classification labels;

[0162] S702: The probabilities of each node belonging to the four classification labels are combined to obtain the classification result of each node.

[0163] In the above embodiment, the calculation of the high-order representation vector matrix by the softmax function obtains the probability of each node belonging to the four classification labels, i.e., the probability of each node belonging to the amplify label, the probability of each node belonging to the deny label, the probability of each node belonging to the clarify label, and the probability of each node belonging to the Null label.

[0164] Then, the probabilities of each node belonging to the four classification labels are combined, which can be weighted summation or maximum value calculation, etc., to finally obtain the label of the node, i.e., the rumor type.

[0165] In an embodiment, after the step S702 of combining the probabilities of each node belonging to the four classification labels to obtain the classification result of each node, the method further comprises:

[0166] S711: Calculate the difference between the probability distribution of the four classification labels of the node and the true distribution by substituting the probability of the node in the four classification labels into the loss function using the following formula:

[0167]

[0168] wherein y1 is the true distribution of the node in the classification label of the opening topic, is the probability distribution of the node in the classification label of the opening topic, y2 is the true distribution of the node in the classification label of the denial and doubt, is the probability distribution of the node in the classification label of the denial and doubt, y3 is the true distribution of the node in the classification label of the clarification, is the probability distribution of the node in the classification label of the clarification, y4 is the true distribution of the node in the classification label of the null, is the probability distribution of the node in the classification label of the null, a1 is the weight of the classification label of the opening topic, a2 is the weight of the classification label of the denial and doubt, a3 is the weight of the classification label of the clarification, and a4 is the weight of the classification label of the null.

[0169] S712: Train and fit the first graph convolutional neural network, the second graph convolutional neural network, and the third graph convolutional neural network according to the difference between the probability distribution of the four classification labels of the node and the true distribution, and optimize the parameters of the first graph convolutional neural network, the second graph convolutional neural network, and the third graph convolutional neural network.

[0170] In the above embodiment, the loss function is used to calculate the difference between the predicted value and the true value for a specific data set. The predicted value is the label classification result calculated by the graph convolutional neural network, and the true value is the label data of each node contained in the training set. In the training process of the graph convolutional neural network, the training data is used to minimize the loss function, thereby changing the parameters of the graph convolutional neural network. By modifying the parameters of the graph convolutional neural network, the predicted value calculated by the graph convolutional neural network is closer to the true value.

[0171] In the scheme of the application, there is a label imbalance problem. In a rumor event, most of the nodes are not the trigger point, so the number of null labels is large, while the number of amplify, deny, and clarify labels is small. Therefore, in the training, a weighted loss function is used for calculation, which can effectively alleviate the precision decline caused by label imbalance and reduce the negative impact of label imbalance, thereby improving the accuracy of rumor trigger point classification.

[0172] Reference Figure 2The application further provides a rumor detonation point detection device based on a graph neural network, comprising:

[0173] An acquisition module 10 is configured to acquire a data set, wherein the data set contains node texts participating in rumor propagation and relationships between nodes;

[0174] A pre-training module 20 is configured to pre-train the data set by using a Bert language representation model, so as to obtain a vector matrix of the node texts and an adjacency matrix based on the relationships between nodes;

[0175] A first calculation module 30 is configured to input the vector matrix and the adjacency matrix into a first graph convolutional neural network for calculation, so as to obtain a first high-order representation vector matrix;

[0176] An adjacency matrix transposition module 40 is configured to transpose the adjacency matrix, so as to obtain a transposed adjacency matrix;

[0177] A second calculation module 50 is configured to input the vector matrix and the transposed adjacency matrix into a second graph convolutional neural network for calculation, so as to obtain a second high-order representation vector matrix;

[0178] A high-order representation vector matrix calculation module 60 is configured to splice the first high-order representation vector matrix and the second high-order representation vector matrix, so as to obtain a high-order representation vector matrix of each node;

[0179] A classification module 70 is configured to input the high-order representation vector matrix of each node into a full connection layer of a third graph convolutional neural network for classification, so as to obtain a classification result of each node, wherein the classification labels are divided into four types: amplify, deny, clarify and Null;

[0180] A rumor detonation point determination module is configured to determine a rumor detonation point based on the classification result of each node.

[0181] Reference Figure 3In the embodiments of the present application, a computer device is also provided, which can be a server. The computer device comprises a processor, a memory, a network interface and a database connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store node text, a model and the like. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program, when executed by the processor, can implement the rumor detonation point detection method of any of the above embodiments.

[0182] The embodiments of the present application also provide a computer readable storage medium having a computer program stored thereon. The computer program, when executed by a processor, can implement the rumor detonation point detection method of any of the above embodiments.

[0183] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, the processes of the above-mentioned embodiments can be included. Any reference to memory, storage, database or other medium provided by the present application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM can be available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM) and memory bus dynamic RAM (RDRAM) and the like.

[0184] The above only describes the preferred embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation, or direct or indirect application in other related technical fields, which is based on the content of the specification and drawings of the present application, is also included in the patent protection scope of the present application.

Claims

1. A method for detecting a rumor ignition point, characterized by, The method comprises the following steps: acquiring a data set, wherein the data set contains node texts participating in rumor propagation and relationships between nodes; pre-training the data set by using a Bert language representation model to obtain a vector matrix of the node texts and an adjacency matrix based on the relationships between nodes; inputting the vector matrix and the adjacency matrix into a first graph convolutional neural network for calculation to obtain a first high-order representation vector matrix; transposing the adjacency matrix to obtain a transposed adjacency matrix; inputting the vector matrix and the transposed adjacency matrix into a second graph convolutional neural network for calculation to obtain a second high-order representation vector matrix; splicing the first high-order representation vector matrix and the second high-order representation vector matrix to obtain a high-order representation vector matrix of each node; inputting the high-order representation vector matrix of each node into a fully connected layer of a third graph convolutional neural network for classification to obtain a classification result of each node, wherein the classification labels are divided into four types: opening a topic, opposing and questioning, clarifying and explaining, and no substantive meaning; determining a rumor detonation point based on the classification result of each node; the step of inputting the high-order representation vector matrix of each node into the fully connected layer of the third graph convolutional neural network for classification to obtain the classification result of each node comprises: calculating the probability of each node in the four classification labels by substituting the high-order representation vector matrix of each node into a softmax function; combining the probability of each node in the four classification labels to obtain the classification result of each node; after the step of combining the probability of each node in the four classification labels to obtain the classification result of each node, the following steps are included: substituting the probability of each node in the four classification labels into a loss function by using the following formula to calculate the gap between the probability distribution of each node in the four classification labels and the true distribution: wherein: is a true distribution of the classification label of the node in opening a topic, is a probability distribution of the classification label of the node in opening a topic, is a true distribution of the classification label of the node in opposing a question, is a probability distribution of the classification label of the node in opposing a question, is a true distribution of the classification label of the node in clarifying an explanation, is a probability distribution of the classification label of the node in clarifying an explanation, is a true distribution of the classification label of the node in having no substantive, is a probability distribution of the classification label of the node in having no substantive, is a weight of the classification label of opening a topic, is a weight of the classification label of opposing a question, is a weight of the classification label of clarifying an explanation, is a weight of the classification label of having no substantive. training and fitting the first graph convolutional neural network, the second graph convolutional neural network and the third graph convolutional neural network according to the gap between the probability distribution of each node in the four classification labels and the true distribution to optimize the parameters of the first graph convolutional neural network, the second graph convolutional neural network and the third graph convolutional neural network.

2. The method of rumor detonation point detection of claim 1, wherein, the step of inputting the vector matrix and the adjacency matrix into the first graph convolutional neural network for calculation to obtain a first high-order representation vector matrix comprises: substituting the vector matrix and the adjacency matrix into a first layer of the first graph convolutional neural network for calculation by using the following formula to obtain a first calculation result: wherein: is the first calculation result, is the vector matrix, is the adjacency matrix; inputting the first calculation result and the adjacency matrix into a second layer of the first graph convolutional neural network for calculation by using the following formula to obtain a first high-order representation vector matrix: wherein: is a first high order representation vector matrix.

3. The method of rumor detonation point detection of claim 2, wherein, the step of substituting the vector matrix and the adjacency matrix into the first layer of the first graph convolutional neural network for calculation to obtain a first calculation result comprises: calculating a first diagonal matrix according to the adjacency matrix by using the following formula: wherein: is the adjacency matrix, is the first diagonal matrix, i and j is the number of rows and columns of the matrix, is the number of columns of the adjacency matrix; normalizing the adjacency matrix according to the first diagonal matrix by using the following formula: = (A + I) / 2 = (A + I) / 2 wherein: is the normalized adjacency matrix, I is the identity matrix, and D is the first diagonal matrix; The normalized adjacency matrix and the vector matrix are substituted into an activation function for calculation by using the following formula to obtain a first calculation result: wherein: is a weight matrix of a first layer of the first graph convolutional neural network.

4. The method of rumor detonation point detection of claim 1, wherein, The step of inputting the vector matrix and the transposed adjacency matrix into the second graph convolutional neural network for calculation to obtain a second high-order representation vector matrix comprises: The vector matrix and the transposed adjacency matrix are input into a first layer of the second graph convolutional neural network for calculation by using the following formula to obtain a second calculation result: wherein: is the second calculation result, is the vector matrix, is the transpose adjacency matrix; The second calculation result and the transposed adjacency matrix are input into a second layer of the second graph convolutional neural network for calculation by using the following formula to obtain a second high-order representation vector matrix: wherein: is a second higher order representation vector matrix.

5. The method of rumor detonation point detection of claim 4, wherein, The step of inputting the vector matrix and the transposed adjacency matrix into the first layer of the second graph convolutional neural network for calculation to obtain the second calculation result comprises: According to the transposed adjacency matrix, a second diagonal matrix is obtained by calculation using the following formula: wherein: is the transpose adjacency matrix, is the second diagonal matrix, i and j is the number of rows and columns of the matrix, is the number of columns of the transpose adjacency matrix; According to the second diagonal matrix, the adjacency matrix is normalized by using the following formula: = ( +I) According to: is the normalized transposed adjacency matrix, I is the identity matrix, is the second diagonal matrix; The normalized adjacency matrix and the vector matrix are substituted into an activation function for calculation by using the following formula to obtain a first calculation result: wherein: is a weight matrix of a first layer of the second graph convolutional neural network.

6. A device for detecting a rumor ignition point for implementing the method of detecting a rumor ignition point according to any one of claims 1 to 5, characterized by, Comprise: An acquisition module is configured to acquire a data set, wherein the data set contains node texts participating in rumor propagation and relationships between nodes. A pre-training module is configured to pre-train the data set by using a Bert language representation model to obtain a vector matrix of the node texts and an adjacency matrix based on the relationships between nodes. A first calculation module is configured to input the vector matrix and the adjacency matrix into a first graph convolutional neural network for calculation to obtain a first high-order representation vector matrix. An adjacency matrix transposition module is configured to transpose the adjacency matrix to obtain a transposed adjacency matrix. A second calculation module is configured to input the vector matrix and the transposed adjacency matrix into a second graph convolutional neural network for calculation to obtain a second high-order representation vector matrix. A high-order representation vector matrix calculation module is configured to splice the first high-order representation vector matrix and the second high-order representation vector matrix to obtain a high-order representation vector matrix of each node. A classification module is configured to input the high-order representation vector matrix of each node into a fully connected layer of a third graph convolutional neural network for classification to obtain a classification result of each node, wherein the classification labels are divided into four types: opening a topic, opposing and questioning, clarifying and explaining, and no substantive meaning. A rumor ignition point determination module is configured to determine a rumor ignition point based on the classification result of each node. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6. The processor executes the computer program to implement the steps of the method of any one of claims 1 to 5.

8. A computer storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 5.

Images