A method for detecting false topics based on dynamic graph transformer representation

Abstract

The application belongs to the technical field of natural language processing, and particularly relates to a false topic detection method based on dynamic graph Transformer representation; the method comprises the following steps: obtaining social communication data; constructing an influence function of a user position by means of a multiple linear regression algorithm, and then introducing a game theory to quantify a group position in a communication space; constructing a graph structure of the communication space according to a friend relationship between users and a comment relationship between users, and obtaining a communication space relationship matrix; based on a communication space feature matrix and a relationship matrix of a shortest communication path, aggregating group and individual features through a modified multi-head attention mechanism; quantifying communication space graph structure features based on the modified multi-head attention mechanism and a traditional Transformer architecture; quantifying time sequence communication features based on a gate neural network, then fusing topic text features and inputting the topic text features into two linear layers for processing to obtain false topic prediction results; the application can mine dynamic graph structure features, thereby improving the accuracy and generalization of false topic detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing, and specifically relates to a false topic detection method based on dynamic graph Transformer representation. Background Art

[0002] Social networks are a double-edged sword. On the one hand, they provide convenience, allowing people from all corners of the world to communicate face-to-face. On the other hand, they can also greatly facilitate the generation and spread of false information. This is because it is difficult to distinguish true from false information on social platforms. Major social platforms have now actively launched their own online false information identification platforms, publicly publicizing confirmed false information. Mainstream social networking platforms, such as WEIBO and TikTok, still rely on a manual application-approval model to identify false information. This model consumes enormous manpower and resources, but the results are limited. However, the unique nature of social networks makes automatically identifying these false information a very difficult task.

[0003] The process of spreading false topics uses graph structures as the transmission carrier, and the relationships in the graph mainly involve interactive relationships such as user comments and forwarding. Traditional false topic detection models, based on graph neural networks, focus on the graph structure differences between different topics to identify false topics. However, the propagation graph structure constructed based on comment or forwarding relationships is not stable and cannot effectively measure the dynamic characteristics of groups and individuals, long-term and short-term. In order to improve the accuracy and generalization of false topic detection, it is necessary to extract the propagation graph information more deeply and mine the temporal dynamic feature information in the graph structure. At the same time, existing false topic detection models overly rely on static propagation graphs and ignore the role of implicit relationships in dynamic propagation graphs, which cannot meet the complex and changing false topic detection tasks. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a false topic detection method based on dynamic graph Transformer representation, comprising:

[0005] Obtain social communication data and input it into the trained dynamic graph Transformer representation of the false topic detection model for processing to obtain the false topic detection results;

[0006] The training process for false topic detection based on dynamic graph Transformer representation includes:

[0007] S1: Obtain social communication data, including topic tags, topic text, all user comment texts participating in the topic conditioning thread, friend relationships between users, and comment relationships between users;

[0008] S2: Using the multivariate linear regression algorithm, we construct an influence function for user positions and introduce game theory to quantify the positions of groups in the communication space. This yields position features, which are used as feature vectors of nodes to construct a feature matrix for the communication space.

[0009] S3: Build a graph structure of the propagation space based on the friendship relationships between users and the comment relationships between users, obtain the propagation space relationship matrix, and calculate the relationship matrix of the shortest propagation path based on the shortest path algorithm;

[0010] S4: Based on the propagation space feature matrix and the relationship matrix of the shortest propagation path, group and individual features are aggregated through a modified multi-head attention mechanism;

[0011] S5: Based on the modified multi-head attention mechanism and the traditional Transformer architecture, the structural features of the propagation space graph are quantified to obtain the temporal propagation features;

[0012] S6: Quantify the temporal propagation features based on a gated neural network, then fuse the topic text features and input them into two linear layers for processing to obtain the false topic prediction results;

[0013] S7: Calculate the cross-entropy loss of the model based on the topic tags and false topic prediction results, and adjust the model parameters based on the cross-entropy loss to obtain the false topic detection model represented by the trained dynamic graph Transformer.

[0014] Beneficial effects of the present invention:

[0015] The present invention can effectively take into account the representation of group behavior in the propagation space and the dynamic modeling of propagation, thereby solving the problem of early false topics. First, a relative position encoding, masking and multi-head attention mechanism are introduced to realize the local and global feature representation of graph nodes, so as to more effectively mine the group adversarial information of the propagation subgraph; secondly, considering the huge difference between the incubation period and the outbreak period of event propagation, an event propagation representation method based on temporal memory units is proposed; finally, the group behavior analysis and the propagation representation method are combined to solve the false topic detection task; the present invention is of great significance for the false topic detection task based on graph data, can cope with complex and changeable cross-domain social false topic tasks, and help to suppress the spread of false topics. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flowchart for training the false topic detection model represented by the dynamic graph Transformer in the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] A method for detecting false topics using a dynamic graph Transformer representation, including the following:

[0019] Obtain social communication data and input it into the trained dynamic graph Transformer representation of the false topic detection model for processing to obtain the false topic detection results;

[0020] The training process of the false topic detection model represented by the dynamic graph Transformer, such as Figure 1 Shown, including:

[0021] S1: Obtain social communication data, including topic tags, topic text, all user comment texts participating in the topic conditioning thread, friend relationships between users, and comment relationships between users.

[0022] Obtain social communication data on social platforms, including topic tags (whether they are false topics), topic text, all user comment texts participating in the topic adjustment thread, friend relationships between users, and comment relationships between users.

[0023] S2: Use the multivariate linear regression algorithm to construct the influence function of user positions, and then introduce game theory to quantify the group positions in the communication space.

[0024] User positions on social platforms are often hidden, and relying solely on content features cannot effectively uncover their true stances. For example, "Haha!" represents different user positions in different contexts. However, user positions often share similarities and consistency with those of their friends. Therefore, we combine internal individual drivers with external group influencers and construct an influence function for user positions using a multivariate linear regression algorithm. We then introduce game theory to quantify users' true positions. The specific calculation process is as follows:

[0025]

[0026] in, and Represents user u i influences that hold positive, negative, and neutral positions at time t; and represent the internal individual driving factors of users holding positive, negative and neutral positions at time t, respectively; and Represents user u i The external group influencing factors that hold positive, negative, and neutral positions at time t; ρ0, ρ1, and ρ2 are the partial regression coefficients of the trained multiple linear regression algorithm; ρ1 and ρ2 reflect the proportion of internal driving factors and external influencing factors in the influence of user positions. The calculation process for the group influencing factors corresponding to positive positions is as follows:

[0027]

[0028] in, Represents user u in the communication space at time t i The number of friends with positive attitudes; Represents user u in the communication space at time t i The number of friends who participated in the event dissemination. Subsequently, the calculation process for group influence factors corresponding to negative and neutral positions is similar. Next, the calculation process for individual driving factors corresponding to positive positions is as follows:

[0029]

[0030] Among them, LSTM represents the pre-trained long short-term memory neural network; c j Represents user u i Comments posted at time t.

[0031] Then, we use game theory to quantify the mutual influence of user positions. The formula is as follows:

[0032]

[0033] in, and They represent the mutual influence of three user positions quantified by game theory; and Represents two adjustable parameters, used to express the different degrees of influence of positive and negative stances on no stance, and Represents user u i The influence of positive, negative and neutral positions at time t.

[0034] Then, the position characteristics is used as the feature vector of the node. Therefore, the propagation space feature matrix at time t is as follows:

[0035]

[0036] Among them, A t represents the propagation space feature matrix at time t. s represents the number of users.

[0037] S3: Construct the graph structure of the propagation space based on the friend relationships between users and the comment relationships between users, and obtain the propagation space relationship matrix.

[0038] A friend relationship graph is constructed based on the friend relationships between users, and a comment relationship graph is constructed based on the comment relationships between users.

[0039] Combine the friend relationship graph and the comment relationship graph, that is, combine all the relationships in the two relationship graphs to generate the relationship matrix of the propagation space. The formula is as follows:

[0040]

[0041] Among them, R t Represents the relationship matrix corresponding to the propagation space at time t, a ij represents the element in the i-th row and j-th column of the relationship matrix, and s represents the number of users.

[0042] Then, the relationship matrix of the shortest propagation path is calculated based on the shortest path algorithm (Dijkstra), and the calculation formula is:

[0043]

[0044] in, Represents the relationship matrix of the shortest propagation path corresponding to the propagation space at time t, R t represents the relationship matrix corresponding to the propagation space at time t. Dijkstra represents the Dijkstra algorithm.

[0045] S4: Based on the relationship matrix and feature matrix of the shortest propagation path at time t, group and individual features are aggregated with the help of the attention mechanism.

[0046] Modify the calculation method of the multi-head attention mechanism (MHA) in the Transformer architecture to adapt to the task of social false topic detection. The formula is as follows:

[0047]

[0048] Among them, Q, K and V represent the characteristic matrix A of the propagation space at time t of event t The corresponding query, key, and value vectors; K T represents the key vector K transposed; d k represents the dimension of the key vector K; B R Represents the relative position encoding matrix; B M Represents the relative position mask matrix; the encoding matrix calculation process is shown in the following formula:

[0049]

[0050] Among them, TR Represents a trainable relative position bias table; subsequently, the mask matrix calculation process is as follows:

[0051]

[0052] in, Represents the mask table, size and bias table T R Same; when the mask table index ε> When , the mask table value is -100. On the contrary, the index When , the mask table value is 0. Representation mask table When the value is -100, the soft function returns the final attention to 0 regardless of the relationship between nodes. However, the result remains unchanged when the value is 0. Therefore, combining the masking mechanism with attention calculation can limit the receptive field of the current node, thereby capturing the characteristics of local individual competition behavior.

[0053] Subsequently, the multi-head attention calculation process is shown in the formula below:

[0054]

[0055] MHA(Q,K,V)=Concat(head1,...,head h )W M

[0056] Among them, head i Indicates the calculation of the attention of the i-th head; and The table shows the parameter matrices corresponding to the Q, K and V vectors; W M represents the parameter matrix.

[0057] S5: Quantitatively propagating spatial graph structural features based on a modified multi-head attention mechanism and traditional Transformer architecture.

[0058] To capture group behavior characteristics in the propagation subspace, this section uses two Transformer Blocks (TBs) to focus on group and individual behaviors, respectively. Subsequently, the feature representations are fused to extract propagation graph features. The specific formula is as follows:

[0059]

[0060] Among them, H t The characteristic representation matrix representing the group behavior in the propagation space at time t; A Transformer block that focuses on group competition behavior; Represents a Transformer block that will restrict the node receptive field to capture local individual competitive behavior features. t Represents the propagation space feature matrix at time t. The architecture of TB is as follows:

[0061]

[0062] in, Represents the feature matrix after the Transformer block, A t is the eigenvector at time t, is the intermediate variable. MLP stands for Multilayer Linear Perceptron. LN stands for Layer Normalization. Represents the modified multi-head attention mechanism. Indicates the mask rate.

[0063] Then, the topic propagation space sequence {G1,G2,...,G l} is transformed into a feature sequence {H1,H2,...,H l},and A single H is a feature of a short period of propagation time. Here it is a sequence, including time series, that is, H1 to H l It satisfies the timing relationship.

[0064] S6: Based on the gated neural network, the temporal propagation features are quantified, and then the topic text features are fused and input into two linear layers for processing to obtain the false topic prediction results.

[0065] Based on the gated neural network to quantify the time series propagation characteristics, the formula is as follows:

[0066] H=GRU({H1,H2,...,H l})

[0067] Where H represents the final representation of the propagated features and GRU represents the gated neural network.

[0068] Subsequently, the topic text features are integrated to construct the final feature representation of the topic and input into two linear layers for processing to obtain the false topic prediction result. The formula is:

[0069]

[0070] P=softmax(W2P1+b2)

[0071] Among them, P represents the prediction vector, A I Represents the feature matrix A oThe vector sequence after one-dimensional expansion, P1 represents the output vector of the first linear layer, W1 and W2 represent trainable linear matrices, and b1 and b2 represent trainable bias matrices. represents the final features, c0 represents the topic text, and Bert represents the pre-trained Bert model. Concat represents the feature concatenation algorithm, using the TensorFlow framework as an example. Relu and softmax represent activation functions. To prevent overfitting during training, the present invention uses the Dropout algorithm to further enhance the generalization of the model.

[0072] S7: Calculate the total model loss based on the topic tags and false topic prediction results and adjust the model parameters based on the total loss to obtain a trained reinforced multi-relation false topic detection model.

[0073] The total loss function of the model uses cross entropy loss:

[0074]

[0075] Among them, Loss represents the total loss of the model, y i represents the true label of the i-th topic, P i Represents the false topic prediction result of the i-th topic.

[0076] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A false topic detection method based on dynamic graph Transformer representation, characterized by: include: Obtain social communication data and input it into the trained dynamic graph Transformer representation of the false topic detection model for processing to obtain the false topic detection results; The training process for false topic detection based on dynamic graph Transformer representation includes: S1: Obtain social communication data, including topic tags, topic text, all user comment texts participating in the topic conditioning thread, friend relationships between users, and comment relationships between users; S2: Using the multivariate linear regression algorithm, we construct an influence function for user positions and introduce game theory to quantify the positions of groups in the communication space. This yields position features, which are used as feature vectors of nodes to construct a feature matrix for the communication space. S3: Build a graph structure of the propagation space based on the friendship relationships between users and the comment relationships between users, obtain the propagation space relationship matrix, and calculate the relationship matrix of the shortest propagation path based on the shortest path algorithm; S4: Based on the propagation space feature matrix and the relationship matrix of the shortest propagation path, group and individual features are aggregated through a modified multi-head attention mechanism; Based on the propagation space feature matrix and the shortest propagation path relationship matrix, a modified multi-head attention mechanism is used to aggregate group and individual features, including: Among them, Q, K and V represent the characteristic matrix A of the propagation space at time t respectively. t The corresponding query, key, and value vectors; T represents the matrix transpose; B R Represents the relative position encoding matrix; B M Represents the relative position mask matrix; d k represents the dimension of the key vector K; S5: Based on the modified multi-head attention mechanism and the traditional Transformer architecture, the structural features of the propagation space graph are quantified to obtain the temporal propagation features; Based on the modified multi-head attention mechanism and the traditional Transformer architecture, the spatial graph structural features of the propagation are quantified, including: In order to obtain the group behavior characteristics of the propagation subspace, two Transformer blocks are used, Transformer Block and TB, which focus on the group stance and individual stance behaviors respectively, and fuse the feature representations to extract the propagation graph features; The topic propagation space sequence {G1,G2,...,G l } is transformed into a feature sequence {H1,H2,...,H l },and S6: Quantify the temporal propagation features based on a gated neural network, then fuse the topic text features and input them into two linear layers for processing to obtain the false topic prediction results; S7: Calculate the cross-entropy loss of the model based on the topic tags and false topic prediction results, and adjust the model parameters based on the cross-entropy loss to obtain the false topic detection model represented by the trained dynamic graph Transformer.

2. The method for detecting false topics based on dynamic graph Transformer representation according to claim 1, characterized in that: The influence function of user positions is constructed using the multivariate linear regression algorithm, including: in, and Represents user u i influences that hold positive, negative, and neutral positions at time t; and represent the internal individual driving factors of users holding positive, negative and neutral positions at time t, respectively; and Represents user u i The external group influencing factors holding positive, negative and neutral positions at time t; ρ0, ρ1 and ρ2 represent the first, second and third partial regression coefficients of the trained multiple linear regression algorithm, respectively.

3. The method for detecting false topics based on dynamic graph Transformer representation according to claim 1, characterized in that: Game theory is introduced to quantify the positions of groups in the communication space and obtain the characteristics of the positions, including: in, and They represent the mutual influence of three user positions quantified by game theory; and They represent the adjustable parameters of the different degrees of influence of positive and negative stances on no stance, and Represents user u i The influence of positive, negative and neutral positions at time t.

4. The method for detecting false topics based on dynamic graph Transformer representation according to claim 1, characterized in that: The graph structure of the propagation space is constructed based on the friendship relationships between users and the comment relationships between users, and the propagation space relationship matrix is ​​obtained, including: Among them, R t Represents the relationship matrix corresponding to the propagation space at time t; a ij represents the element in row i and column j of the relationship matrix; s represents the number of users; They represent user u at time t respectively i and u j .

5. The method for detecting false topics based on dynamic graph Transformer representation according to claim 1, characterized in that: The relationship matrix of the shortest propagation path is calculated based on the shortest path algorithm, including: in, Represents the relationship matrix of the shortest propagation path corresponding to the propagation space at time t; R t Represents the relationship matrix corresponding to the propagation space at time t; Dijkstra represents the Dijkstra algorithm.

6. The method for detecting false topics based on dynamic graph Transformer representation according to claim 1, characterized in that: To obtain the group behavior characteristics of the propagation subspace, two Transformer blocks are used: Transformer Block and TB, which focus on group stance and individual stance respectively, and fuse the feature representations to extract propagation graph features, including: H t =Concat(TB θ＝1 (A t ),TB θ＝0.5 (A t )) Among them, H t The characteristic representation matrix representing the group behavior in the propagation space at time t; TB θ＝1 (A t ) represents the Transformer block that focuses on group competition behavior; TB θ＝0.5 (A t ) represents the Transformer block that will limit the node receptive field to obtain local individual competition behavior characteristics; A t Represents the propagation space feature matrix at time t; Concat represents feature concatenation.

7. The method for detecting false topics based on dynamic graph Transformer representation according to claim 1, characterized in that: Quantify temporal propagation features based on gated neural networks, including: H=GRU({H1,H2,...,H l }) Where H represents the final representation of the propagation feature; GRU represents the gated neural network; {H1,H2,...,H l } represents a time series feature sequence.

8. The method for detecting false topics based on dynamic graph Transformer representation according to claim 1, characterized in that: The topic text features are integrated to construct the final feature representation of the topic, and then input into two linear layers for processing to obtain the false topic prediction results, including: P=softmax(W2P1+b2) Where P represents the prediction vector, P1 represents the output vector of the first linear layer, W1 and W2 represent the first and second trainable linear matrices, b1 and b2 represent the first and second trainable bias matrices, respectively. Represents the final feature, H represents the final representation of the propagation feature, c0 represents the topic text, Bert represents the pre-trained Bert model, Concat represents feature concatenation, Relu and softmax represent the first and second activation functions.

Images