False information processing method and device, storage medium, product and equipment

Abstract

The invention discloses a false information processing method and device, a storage medium, a product and equipment, and the method comprises the steps: constructing a word graph based on a to-be-detected text, the word graph taking words in the to-be-detected text as nodes, and taking semantic association between the words as edges; extracting multi-dimensional heterogeneous features of nodes in the word graph by using a natural language model; performing feature embedding on the nodes in the word graph according to the multi-dimensional heterogeneous features of the nodes to obtain feature embedding representation of the word graph; inputting the feature embedded representation of the word graph into a graph neural network model for false information detection to obtain a detection result; and detecting the misleading degree of the misleading effect of the nodes in the word graph on the detection result by using a mask mechanism to obtain a word misleading ranking result. According to the method, false information can be efficiently and accurately detected, higher interpretability is achieved, the detection result is more transparent and credible, and the requirements of field experts and common users for interpretability can be met at the same time.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, computer-readable storage medium, computer program product, and terminal device for processing false information. Background Technology

[0002] False information (such as fake news) is fabricated content intended to mislead the public. It typically appears in the form of government propaganda and news editing, and its production and dissemination require financial support and careful coordination. However, in recent years, with the widespread adoption and significant reduction in costs of text manipulation and generative artificial intelligence technologies, society is flooded with false information, causing widespread negative social impacts. For example, fake news deliberately misleads the public into accepting false or biased information to achieve certain political goals or manipulate financial markets. Furthermore, fake news undermines the credibility of real news and affects the public's ability to distinguish between facts and false information. Therefore, the detection and identification of false information has become a research hotspot.

[0003] Existing research on misinformation faces two main challenges: First, the spread and creation of misinformation often involve the coexistence of multiple heterogeneous characteristics, increasing the difficulty of detection and even allowing misinformation to masquerade as genuine information, affecting detection efficiency and accuracy. For example, even a single word can have drastically different meanings in different contexts; "apple" can refer to a fruit in the food industry but a company in the technology field. Second, interpretability is a key factor in understanding and trusting misinformation detection solutions, helping to further improve the accuracy and transparency of detection results. Although some existing misinformation detection solutions have made progress in detection accuracy, many remain black-box models, lacking effective explanation mechanisms and failing to clearly explain the specific reasons why information is marked as misinformation. For example, not every sentence in fake news is false; how to clarify the decision-making basis of the detection solution in an interpretable way is crucial for improving existing detection solutions. Therefore, there is an urgent need for a new technology that can accurately and efficiently detect misinformation in its early stages before large-scale dissemination, while providing an explanation mechanism that can be understood by both domain experts and ordinary users. Summary of the Invention

[0004] The purpose of this invention is to provide a method, apparatus, computer-readable storage medium, computer program product, and terminal device for processing false information, which can efficiently and accurately detect false information and has stronger interpretability, making the detection results more transparent and credible, and can simultaneously meet the interpretability needs of both domain experts and ordinary users.

[0005] To achieve the above objectives, embodiments of the present invention provide a method for processing false information, including:

[0006] A word graph is constructed based on the text to be detected; wherein, the word graph uses words in the text to be detected as nodes and semantic relationships between words as edges;

[0007] Multidimensional heterogeneous features of nodes in the word graph are extracted using a natural language model;

[0008] Based on the multidimensional heterogeneous features of the nodes, feature embedding is performed on the nodes in the word graph to obtain the feature embedding representation of the word graph;

[0009] The feature embedding representation of the word graph is input into a graph neural network model for false information detection to obtain the detection result;

[0010] The degree of misleading effect of nodes in the word graph on the detection results is detected by using a masking mechanism, and the misleading ranking results of words are obtained.

[0011] Furthermore, the construction of a word map based on the text to be detected specifically includes:

[0012] The process involves acquiring multiple words from the text to be detected and determining the semantic relationships between the words.

[0013] Each word is treated as a node in the word graph, and edges are created between words that are semantically related based on a sliding window.

[0014] Further, the step of embedding features into the nodes in the word graph based on the multidimensional heterogeneous features of the nodes to obtain the feature embedding representation of the word graph specifically includes:

[0015] The PageRank algorithm is used to obtain the PageRank score of each node in the word graph;

[0016] The feature embedding representation of the corresponding node is obtained based on the multidimensional heterogeneous features of the node and the PageRank score of each node;

[0017] The feature embedding representation of the word graph is obtained based on the feature embedding representation of all nodes of the word graph.

[0018] Furthermore, the step of obtaining the feature embedding representation of the corresponding node based on the multidimensional heterogeneous features of the node and the PageRank score of each node specifically includes:

[0019] For the i-th node in the word graph, according to formula h i =X·PR(i) TObtain the feature embedding representation of the i-th node; where i is the node index, X is the multidimensional heterogeneous feature of the node, and PR(i) is the PageRank score of the i-th node.

[0020] Further, obtaining the feature embedding representation of the word graph based on the feature embedding representations of all nodes of the word graph specifically includes:

[0021] According to the formula u=pool(h) i |i∈{1,2,...,n}) to obtain the feature embedding representation of the word graph; where pool is the pooling operation and n is the total number of nodes in the word graph.

[0022] Furthermore, the graph neural network model uses cross-entropy as the loss function during training, and the definition of cross-entropy is: Where j is the training sample index, y j z is the label for the j-th training sample, which indicates whether the j-th training sample is false information. j This represents the false information detection result for the j-th training sample.

[0023] Furthermore, the step of using a masking mechanism to detect the degree of misleading effect of nodes in the word graph on the detection results, and obtaining a word misleading ranking result, specifically includes:

[0024] The nodes in the word graph are masked using a masking mechanism, which transforms the word graph into a new word graph, and a new feature embedding representation of the new word graph is obtained.

[0025] The new feature embedding representation of the new word graph is input into the graph neural network model for false information detection to obtain new detection results;

[0026] Based on the new detection results and the detection results, compare the difference in correct prediction probability between the new feature embedding representation of the new word graph and the feature embedding representation of the word graph to obtain the degree of misleading effect of the masked nodes on the detection results;

[0027] The misleading ranking result of words is obtained based on the degree of misleading effect of each node in the word graph after it is masked on the detection result.

[0028] Furthermore, obtaining the new feature embedding representation of the new word graph specifically includes:

[0029] Based on the topological changes of the new word graph and the PageRank score of each node in the word graph, obtain the new PageRank score of each node in the new word graph;

[0030] Based on the multidimensional heterogeneous features of the nodes and the new PageRank score of each node, obtain the new feature embedding representation of the corresponding node;

[0031] The feature embedding representation of the new word graph is obtained based on the new feature embedding representation of all nodes in the new word graph.

[0032] Further, the step of obtaining a new PageRank score for each node in the new word graph based on the topological changes of the new word graph and the PageRank score of each node in the word graph specifically includes:

[0033] For the i-th node in the new word graph, according to the formula Obtain the new PageRank score for the i-th node; where i is the node index, PR(i) is the PageRank score of the i-th node, α is the damping factor, M' is the column random transition matrix of the new word graph, and PR(i) pushout The change in PageRank score of the i-th node caused by the topological change of the new word graph.

[0034] Furthermore, the formula for calculating the change in the PageRank score of the i-th node is: PR(i) pushout =α(M'-M)PR(i); where M is the column random transition matrix of the word graph.

[0035] To achieve the above objectives, embodiments of the present invention also provide a false information processing device, comprising:

[0036] A word graph construction module is used to construct a word graph based on the text to be detected; wherein, the word graph uses words in the text to be detected as nodes and semantic relationships between words as edges;

[0037] The feature extraction module is used to extract multidimensional heterogeneous features of nodes in the word graph using a natural language model;

[0038] The feature embedding module is used to embed features of nodes in the word graph based on the multidimensional heterogeneous features of the nodes, so as to obtain the feature embedding representation of the word graph;

[0039] The fake information detection module is used to input the feature embedding representation of the word graph into the graph neural network model to detect fake information and obtain the detection result;

[0040] The detection result interpretation module is used to detect the degree of misleading effect of nodes in the word graph on the detection results using a masking mechanism, and to obtain the word misleading ranking results.

[0041] This invention also provides a computer-readable storage medium including a stored computer program, which, when running, controls the device where the computer-readable storage medium is located to execute any of the above-described methods for processing misinformation.

[0042] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the false information processing method described above.

[0043] This invention also provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the false information processing method described above when executing the computer program.

[0044] Compared with existing technologies, embodiments of the present invention provide a method, apparatus, computer-readable storage medium, computer program product, and terminal device for processing false information. First, a word graph is constructed based on the text to be detected, with words in the text as nodes and semantic relationships between words as edges. Next, a natural language model is used to extract multidimensional heterogeneous features of the nodes in the word graph. Then, feature embedding is performed on the nodes in the word graph based on the multidimensional heterogeneous features to obtain a feature embedding representation of the word graph. Next, the feature embedding representation of the word graph is input into a graph neural network model for false information detection to obtain detection results. Finally, a masking mechanism is used to detect the degree of misleading effect of nodes in the word graph on the detection results, obtaining a ranking result of word misleadingness. Embodiments of the present invention can efficiently and accurately detect false information and have stronger interpretability, making the detection results more transparent and credible, and simultaneously meeting the interpretability needs of domain experts and ordinary users. Attached Figure Description

[0045] Figure 1 This is a flowchart of a preferred embodiment of a method for processing false information provided by the present invention;

[0046] Figure 2 This is a framework diagram of a preferred embodiment of a method for processing false information provided by the present invention;

[0047] Figure 3 This is a structural block diagram of a preferred embodiment of a false information processing device provided by the present invention;

[0048] Figure 4 This is a structural block diagram of a preferred embodiment of a terminal device provided by the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] This invention provides a method for processing false information, see [link to relevant documentation]. Figure 1 The diagram shown is a flowchart of a preferred embodiment of a method for processing false information provided by the present invention, the method comprising steps S11 to S15:

[0051] Step S11: Construct a word graph based on the text to be detected; wherein the word graph uses words in the text to be detected as nodes and semantic relationships between words as edges;

[0052] Step S12: Extract multidimensional heterogeneous features of nodes in the word graph using a natural language model;

[0053] Step S13: Based on the multidimensional heterogeneous features of the nodes, perform feature embedding on the nodes in the word graph to obtain the feature embedding representation of the word graph;

[0054] Step S14: Input the feature embedding representation of the word graph into the graph neural network model for false information detection and obtain the detection result;

[0055] Step S15: Use a masking mechanism to detect the degree of misleading effect of nodes in the word graph on the detection results, and obtain the word misleading ranking results.

[0056] Specifically, firstly, a word graph is constructed based on the text to be detected (e.g., a news article), resulting in a word graph of the text to be detected. The nodes in this word graph represent the words in the text to be detected, and the edges between nodes represent the semantic relationships between words in the text (two nodes connected by an edge represent a pair of semantically related words). Next, based on the word relationships constructed in the graph data structure, a natural language model is used to extract multidimensional heterogeneous features from each node (i.e., each word in the text to be detected) in the word graph. Then, feature embedding is performed on the corresponding nodes based on the extracted multidimensional heterogeneous features, generating the feature embedding representation of the word graph. Finally, the feature embedding representation of the word graph is input into a pre-trained graph neural network. The system performs fake information detection in a Network (GNN) model to obtain the detection result of whether the text to be detected is fake information. Finally, the system uses a masking mechanism to detect the degree of misleading effect of each node in the word graph on the fake information detection result (used to explain the contribution of each word in the text to be detected in the final decision of fake information detection, i.e., the contribution to the fake information detection result), and obtains the word misleading ranking result of all words in the text to be detected.

[0057] It should be noted that when extracting the multidimensional heterogeneous features of each node in the word graph using a natural language model, existing pre-trained models such as Word2Vec and BERT can be directly selected. Through feature extraction by these models, the multidimensional heterogeneous features x∈R of each node in the word graph can be constructed. 1×d d represents the dimension of the heterogeneous features of a node, and R represents the real number field. The multidimensional heterogeneous features of a node refer to the multidimensional heterogeneous feature vectors corresponding to words extracted using natural language processing models. These vectors capture information such as the semantics, part of speech, and contextual dependencies of words. For example, when using the Word2Vec model for feature extraction, the output of the Word2Vec model is a 300-dimensional (i.e., d = 300) static word vector. When using the BERT model for feature extraction, the output of the BERT model is a high-dimensional embedding vector that dynamically changes based on the context.

[0058] It should be noted that when using graph neural network models for fake information detection, the deployment of the graph neural network model can be independent of the model itself. That is, the construction of the feature embedding representation of the word graph is independent of the detection by the graph neural network model. The method of constructing the feature embedding representation does not limit the type of graph neural network model used. Various types of graph neural networks can be flexibly applied according to different needs and settings. For example, graph convolutional networks (GCNs) can be used. Furthermore, in this embodiment of the invention, a simple 32*2 multilayer perceptron (MLP) can also be used to perform instance analysis on the graph neural network model to obtain effective performance.

[0059] For example, when using a graph neural network model for fake information detection, the fake information detection result obtained refers to the prediction result of the graph neural network model on whether the input text to be detected is "real" or "fake".

[0060] It should be noted that when an article is detected as fake or real, the contribution of each word in the article to the detection result varies. For example, some words may help fake information hide its falsity, making it difficult for the detection model to identify. If these words are removed, the detection model's prediction can become more certain, i.e., the prediction probability will increase. These words that mislead the detection results can be called "misleading words." In this embodiment of the invention, in order to improve the interpretability of fake information detection results, a masking mechanism can be introduced. By masking the nodes in the word graph, the influence of the nodes on the feature embedding representation of the word graph can be observed, and then the degree of contribution to the detection results can be analyzed to obtain the word misleading ranking results. Based on the word misleading ranking results, it is possible to effectively explain why the detection model makes a certain prediction, especially for the explanation of false detections. At the same time, it also improves the transparency of the detection results, enabling users to better understand and trust the model's detection results.

[0061] The method for processing false information provided in this invention constructs a word graph data structure, extracts heterogeneous features of nodes, embeds features into nodes, uses a graph neural network model for detection, and combines a word misleading detection mechanism. This method can efficiently and accurately detect false information and has stronger interpretability, improving the interpretability of false information and making the detection results more transparent and credible. It can simultaneously meet the interpretability needs of domain experts and ordinary users.

[0062] In one optional embodiment, the construction of a word map based on the text to be detected specifically includes:

[0063] The process involves acquiring multiple words from the text to be detected and determining the semantic relationships between the words.

[0064] Each word is treated as a node in the word graph, and edges are created between words that are semantically related based on a sliding window.

[0065] Specifically, in conjunction with the above embodiments, when constructing a word graph based on the text to be detected, the text to be detected can first be preprocessed (e.g., word segmentation, removal of stop words, part-of-speech tagging, etc.) to obtain multiple words in the text to be detected and determine the semantic relationships between words. Then, each word in the text to be detected is taken as a node in the word graph, and based on the semantic relationships between words, edges of the word graph are established between words with semantic relationships using a sliding window method.

[0066] It should be noted that when performing text preprocessing, some existing text preprocessing techniques can be used. For example, the n-gram method can be used to segment the text to be detected into multiple words.

[0067] It should be noted that when building the edges of the word graph, the sliding window can be set to a fixed step size. For example, the default sliding window size is 3, which means that for each word in the text to be detected (i.e., each node in the word graph), an edge can be built between it and the three words before and after it. When dealing with words at the beginning or end of a sentence, an asymmetric sliding window mechanism can be used to build edges only with its valid context words to ensure the integrity and validity of the word graph construction.

[0068] It should be noted that a word graph can be represented as an undirected graph G = (V, E), where V is the set of all nodes in the undirected graph G, E is the set of semantic association edges between nodes in the undirected graph G, and V = {v1, v2, ..., v...} n}, E={(v1,v2),(v2,v3),...}, where v1, v2, v3, ... are the words in the text, and (v1,v2), (v2,v3), ... correspond to the edges between v1 and v2, between v2 and v3, ... This graph structure can capture the short-distance dependencies between words and construct the adjacency matrix A of the word graph required for the graph neural network model.

[0069] For example, assuming the input sentence is: "The quick brown fox jumps over the lazydog," and a word graph corresponding to this sentence is constructed using a sliding window with a step size of 3, then the constructed word graph contains the following set of nodes:

[0070] V={The,quick,brown,fox,jumps,over,the,lazy,dog},

[0071] By constructing edges using a sliding window with a step size of 3, where each node represents a word and the edges represent the contextual relationships between these words in a sentence, the constructed word graph contains the following set of edges:

[0072] E={(The,quick),(The,brown),(The,fox),(quick,brown),(quick,fox),...,(lazy,dog)}.

[0073] In one optional embodiment, the step of embedding features of nodes in the word graph based on the multidimensional heterogeneous features of the nodes to obtain the feature embedding representation of the word graph specifically includes:

[0074] The PageRank algorithm is used to obtain the PageRank score of each node in the word graph;

[0075] The feature embedding representation of the corresponding node is obtained based on the multidimensional heterogeneous features of the node and the PageRank score of each node;

[0076] The feature embedding representation of the word graph is obtained based on the feature embedding representation of all nodes of the word graph.

[0077] Specifically, in conjunction with the above embodiments, when embedding features of nodes based on the multidimensional heterogeneous features of each node in the word graph, the PageRank algorithm can be used to generate feature embedding representations for each node in the word graph. That is, the PageRank algorithm can be used to obtain the PageRank score of each node in the word graph, and then the feature embedding representation of the same node can be obtained based on the extracted multidimensional heterogeneous features of the nodes and each PageRank score. After that, the feature embedding representation of the word graph can be obtained based on the feature embedding representations of all nodes in the word graph.

[0078] It should be noted that the core idea of ​​the PageRank algorithm is to calculate the importance of each node through random walks. In this embodiment of the invention, the PageRank algorithm is used to measure the global importance of each word in the propagation path of the text to be detected. The personalized PageRank vector (i.e., PageRank score) of each node can be calculated using the following formula. Assuming the random walk starts from node v... i Initially, the probability distribution of its reaching other nodes is calculated as follows:

[0079] Among them, PR(vi ) is node v i The PageRank score; α is the damping factor, which can also be understood as the jump probability, that is, the probability of node v i The probability of jumping to the next node and forming an edge when there are no external factors interfering; N(v i ) is node v i The set of neighboring nodes of node v, that is, all nodes that are related to node v. i A set of nodes that are directly connected by edges; PR(v j ) is node v j PageRank score, PR(v) j The value of d(v) is usually determined through iterative calculation. Before starting the iteration, an initial value needs to be set for the PageRank score of each node. A common practice is to initialize the PageRank scores of all nodes to equal values. For example, if there are n nodes in the word graph, then the PageRank score of each node can be initialized to 1 / n; j ) represents the degree of a node. In an undirected graph, edges have no direction, and the degree of a node is the total number of edges connected to that node; e i For personalized jump vectors for nodes, e i ∈R n×1 , where n is the total number of nodes in the word graph, e i =1 or 0, when the random walk starts from node v i At the beginning, e i The corresponding v in the middle i The value of one position is 1, and the values ​​of the others are 0. For example, suppose there are 5 nodes v1, v2, v3, v4, and v5. If the random walk starts from node v3, then e3 = 1, while e1 = e2 = e4 = e5 = 0. At this time, e = (0,0,1,0,0). T .

[0080] It should be noted that the above formula can be used to obtain the PageRank score of each node, which represents the importance of the node in the word graph (i.e. the importance of the corresponding word in the text to be detected). The PageRank scores of these nodes will be used as feature embeddings for the input of the subsequent graph neural network model.

[0081] For example, assuming the input sentence is: "The quick brown fox jumps over the lazydog," the PageRank algorithm can be used to obtain the PageRank score of each node in the word graph of this sentence. Assuming "fox" is selected as the seed node, the process of using the PageRank algorithm to calculate the importance of the seed node among the nodes in the word graph is as follows:

[0082] Step 1: Initialize the PageRank vector

[0083] For each node in the word graph, the initial PageRank vector can be set to a uniform distribution; for example, if the input sentence above contains 9 words, then the initial PageRank score of each node can be set to 1 / 9.

[0084] Step 2: Random Walk

[0085] The PageRank algorithm simulates a random walk process starting from the seed node "fox". Each time, it jumps from the current node to its neighboring nodes with a certain probability, and calculates the probability of each node being visited over a long period of time (the calculation formula is shown above).

[0086] Ste:3: Reaching a steady-state distribution (Stationary Distribution)

[0087] As the iterations continue, PageRank eventually converges to a steady-state distribution, which is the PageRank score of each node. This score reflects the probability of randomly walking to the node after multiple random jumps, starting from the seed node "fox". This probability represents the global importance of the word in the context of the article. For example, "fox" as a seed node will have a higher PageRank score on words that are semantically related to it (such as "quick" and "brown") because they are more connected to each other.

[0088] It should be noted that this feature embedding process incorporates the importance of nodes into their feature embedding representation through the PageRank algorithm, thereby providing more accurate input features for subsequent detection by graph neural network models.

[0089] In one optional embodiment, obtaining the feature embedding representation of the corresponding node based on the multidimensional heterogeneous features of the node and the PageRank score of each node specifically includes:

[0090] For the i-th node in the word graph, according to formula h i =X·PR(i) T Obtain the feature embedding representation of the i-th node; where i is the node index, X is the multidimensional heterogeneous feature of the node, and PR(i) is the PageRank score of the i-th node.

[0091] Specifically, in conjunction with the above embodiments, the principle for obtaining the feature embedding representation of each node is the same. That is, the PageRank score of a node calculated by the PageRank algorithm is used as an important feature of that node. Combined with the multidimensional heterogeneous features of all nodes in the word graph, the two are combined to form a feature vector, which is the feature embedding representation of that node. Here, we take the i-th node in the word graph as an example for specific explanation:

[0092] In the word graph data structure of text, the PageRank score of the corresponding node can be defined as:

[0093] PR(i)=αAD -1 PR(i)+(1-α)e i In this context, PR(i) is essentially the same as PR(vi) in the formula above; both represent the personalized PageRank vector (i.e., PageRank score) of the i-th node. The only difference is the meaning of the index. PR(i) takes the i-th element from the PageRank vectors of all nodes (i.e., the vector composed of the PageRank scores of all nodes) to obtain the PageRank score of the i-th node, while PR(vi) takes the i-th element from the vectors of all nodes. i The PageRank score of the i-th node is calculated directly using the PageRank algorithm; A is the adjacency matrix of the word graph, and D is the degree matrix of the adjacency matrix (the degree matrix is ​​a diagonal matrix, and the diagonal elements are the degree of each node in the word graph).

[0094] Furthermore, assuming an article contains n distinct words, combining the above examples, a word graph G with n nodes can be constructed, and the multidimensional heterogeneous features (matrix form) of all nodes in the word graph G can be extracted as X∈R. n×d That is, X(i,:) = x i Then, the feature embedding representation h of the i-th node is... i ∈R 1×d For i∈{1,2,...,n}, it can be obtained through the following formula:

[0095] h i =X·PR(i) T Where PR(i) is the personalized PageRank vector (i.e., PageRank score) of the i-th node, PR(i)∈R n×1 ,PR(i) T It is the transpose of PR(i).

[0096] It should be noted that through this geometric feature extraction, the heterogeneous semantics of words are jointly modeled. Specifically, PR(i) encodes the static distribution of random walks starting from the i-th node, which can be interpreted as the relevance weights of other nodes (i.e., other words) to the seed node (i.e., the selected word) in this word graph G (i.e., the input article). Since the heterogeneous features of the nodes are extracted from a pre-trained natural language model, the input node features x i It is universal, that is, x i being h i Deep specialization means that the meaning of the selected words is specialized by the adjacent words within the range of the input article.

[0097] In one optional embodiment, obtaining the feature embedding representation of the word graph based on the feature embedding representations of all nodes of the word graph specifically includes:

[0098] According to the formula u=pool(h) i |i∈{1,2,...,n}) to obtain the feature embedding representation of the word graph; where pool is the pooling operation and n is the total number of nodes in the word graph.

[0099] Specifically, in conjunction with the above embodiments, after obtaining the feature embedding representation of each node in the word graph, in order to obtain the graph-level embedding representation of the word graph (i.e., the feature embedding representation of the word graph, also known as the text-level embedding representation) u∈R 1×d All word-level embeddings (i.e., node feature embeddings) can be read out as follows: i :

[0100] u = pool(h) i |i∈{1,2,...,n}); where pool is the pooling operation (e.g., it can be an addition or average pooling operation), and n is the total number of nodes in the word graph.

[0101] It should be noted that the feature embedding generation proposed in this embodiment of the invention uses aggregation based on static distribution to replace the traditional message passing scheme of stacked graph neural networks. When the topology of the word graph changes, the new static distribution can be quickly tracked. Therefore, multiple nodes in the input word graph can be directly masked without retraining and detection can be performed directly.

[0102] In one optional embodiment, the graph neural network model uses cross-entropy as a loss function during training, and the cross-entropy is defined as: Where j is the training sample index, y j z is the label for the j-th training sample, which indicates whether the j-th training sample is false information. jThis represents the false information detection result for the j-th training sample.

[0103] Specifically, in conjunction with the above embodiments, the graph neural network model used in this embodiment of the invention employs a loss function during training that is applied to the label information y of the j-th information carrier. j In the feature embedding representation of word graph u j To achieve this, a graph neural network model needs to be called to deploy u. j Corresponding to z j Finally, the graph neural network model is trained using the cross-entropy loss function. Here, the j-th information carrier refers to the j-th input training sample (e.g., the j-th article or the j-th text unit in the training dataset). Model training follows the batch size; if a batch contains 128 training samples, then j = 1, 2, ..., 128. The j-th training sample has its own word graph G. j Word-image G j Feature embedding representation u j Detection label y j (e.g., false or true information), output detection result z j (i.e., graph neural network model for u) j (The predicted results).

[0104] Furthermore, the definition of cross-entropy is:

[0105] Where j is the training sample index, y j This refers to the label information for the j-th training sample. The label information is used to indicate whether the j-th training sample contains false or true information. j The result of false information detection is obtained after the j-th training sample is detected by the graph neural network model.

[0106] For example, in completing the word graph G j Feature embedding u j Subsequently, a multilayer perceptron (MLP) can be used to detect false information in each training sample. This is done by mapping the samples through a neural network containing two fully connected layers, outputting z. j , z j This represents the probability prediction result of whether the training sample is true or false. Then, the obtained output z... j The true label y of the training samples j Cross-entropy loss is used for model optimization. This design decouples the embedded features from the detection model, possessing high versatility and flexibility, and facilitating the replacement and deployment of subsequent detection models.

[0107] It should be noted that, for the training dataset of the model, existing publicly available fake information detection datasets, such as the ISOT fake information detection dataset, can be directly selected.

[0108] In one optional embodiment, the step of using a masking mechanism to detect the degree of misleading effect of nodes in the word graph on the detection results, and obtaining a word misleading ranking result, specifically includes:

[0109] The nodes in the word graph are masked using a masking mechanism, which transforms the word graph into a new word graph, and a new feature embedding representation of the new word graph is obtained.

[0110] The new feature embedding representation of the new word graph is input into the graph neural network model for false information detection to obtain new detection results;

[0111] Based on the new detection results and the detection results, compare the difference in correct prediction probability between the new feature embedding representation of the new word graph and the feature embedding representation of the word graph to obtain the degree of misleading effect of the masked nodes on the detection results;

[0112] The misleading ranking result of words is obtained based on the degree of misleading effect of each node in the word graph after it is masked on the detection result.

[0113] Specifically, in conjunction with the above embodiments, when using the masking mechanism to detect the degree of misleading effect of each node in the word graph on the false information detection results, the detection principle for the degree of misleading effect of each node is the same. Here, any node in the word graph is taken as the node to be masked, and the process of detecting the degree of misleading effect of the node to be masked is used as an example: First, the node to be masked is masked using the masking mechanism (the adjacent edges of the node to be masked in the word graph are deleted), and the topology of the word graph changes, transforming the word graph into a new word graph, and a new feature embedding representation of the new word graph is obtained; then, the new feature embedding representation of the new word graph is input into a pre-trained graph neural network model for false information detection, and a new detection result is obtained accordingly; finally, based on the obtained new detection result and the detection result obtained above by performing false information detection on the feature embedding of the word graph, the difference between the new feature embedding representation of the new word graph and the feature embedding representation of the word graph in the correct prediction probability is compared, and the degree of misleading effect of the masked node on the false information detection results can be obtained.

[0114] Understandably, after obtaining the degree of misleading effect of each node in the word graph on the false information detection results, the nodes in the word graph are sorted according to the degree of misleading effect (equivalent to sorting the words), and the word misleading ranking results of all words in the text to be detected can be obtained.

[0115] It should be noted that the new feature embedding representation of the new word graph reflects the overall changes in the graph structure and semantic space of the text to be detected after a certain word is missing. Therefore, the new feature embedding representation of the new word graph can be compared with the feature embedding representation of the original word graph to analyze whether the removed word affects the false information detection results of the text to be detected.

[0116] In one optional embodiment, obtaining the new feature embedding representation of the new word graph specifically includes:

[0117] Based on the topological changes of the new word graph and the PageRank score of each node in the word graph, obtain the new PageRank score of each node in the new word graph;

[0118] Based on the multidimensional heterogeneous features of the nodes and the new PageRank score of each node, obtain the new feature embedding representation of the corresponding node;

[0119] The feature embedding representation of the new word graph is obtained based on the new feature embedding representation of all nodes in the new word graph.

[0120] Specifically, in conjunction with the above embodiments, when obtaining the new feature embedding representation of the new word graph, we can first obtain the new PageRank score of each node in the new word graph based on the topological changes of the new word graph and the PageRank score of each node in the word graph obtained above. Then, based on the multidimensional heterogeneous features of the nodes extracted above and the new PageRank score of each node in the new word graph, we can obtain the new feature embedding representation of the same node in the new word graph (the principle is the same as the method of obtaining the feature embedding representation of the node in the word graph described above). After that, we can obtain the feature embedding representation of the new word graph based on the new feature embedding representation of all nodes in the new word graph (the principle is the same as the method of obtaining the feature embedding representation of the word graph described above).

[0121] In one optional embodiment, obtaining the new PageRank score of each node in the new word graph based on the topological changes of the new word graph and the PageRank score of each node in the word graph specifically includes:

[0122] For the i-th node in the new word graph, according to the formula Obtain the new PageRank score for the i-th node; where i is the node index, PR(i) is the PageRank score of the i-th node, α is the damping factor, M' is the column random transition matrix of the new word graph, and PR(i) pushoutThe change in PageRank score of the i-th node caused by the topological change of the new word graph.

[0123] In one optional embodiment, the formula for calculating the change in the PageRank score of the i-th node is: PR(i) pushout =α(M'-M)PR(i); where M is the column random transition matrix of the word graph.

[0124] Specifically, in conjunction with the above embodiments, when obtaining the new PageRank score for each node in the new word graph based on the topological changes of the new word graph and the PageRank score of each node in the word graph obtained above, since the principle of obtaining the new PageRank score for each node is the same, we will take the i-th node in the new word graph as an example for explanation:

[0125] Based on the static distribution PR(i) (i.e., the PageRank score of the nodes in the word graph G), which represents the static distribution of seed node i in the word graph G, when it is necessary to mask a certain node in the word graph G and transform it into a new word graph G', the new static distribution PR(i)' of the new word graph G' can be quickly tracked by simply changing the topology of the word graph without fine-tuning the parameters of the graph neural network model. Therefore, the calculation formula of PR(i)' is as follows:

[0126] Where PR(i)' is the new PageRank score of the i-th node in the new word graph, PR(i) is the PageRank score of the i-th node in the word graph, α is the damping factor, and after a node is masked, the topology of the word graph G changes from M = AD. -1 It becomes M' (M is the column random transition matrix of the word graph G, and M' is the column random transition matrix of the new word graph G'). Furthermore, when a node is masked, the change in the topology of the word graph due to the deletion of its adjacent edges will lead to the static distribution PR(i) of the new node. pushout It needs to be recalculated; the push process can be represented as:

[0127] PR(i) pushout =α(M'-M)PR(i); where, PR(i) pushout Let G' be the change in PageRank score of the i-th node caused by the change in the topology of the new word graph. This means that the distribution value that needs to be pushed out from the remaining word graph due to the change in the edges in the word graph can be proven to converge to the accurate new static distribution of the new word graph G' through sufficient cumulative power iterations.

[0128] Understandably, in addition to using the above methods to obtain the new PageRank score of each node in the new word graph, the PageRank algorithm can still be used to recalculate the importance distribution of each node in the new word graph after starting from a certain node and traversing the entire graph. This new importance distribution is the new static distribution of the new word graph after masking.

[0129] It should be noted that after obtaining the new static distribution of the new word graph, it is also possible to use the formula h. i =X·PR(i) T and u = pool(h) i |i∈{1,2,...,n}) to obtain new feature embedding representations of new word graphs without the need for fine-tuning of the graph neural network model.

[0130] It should be noted that by masking certain nodes in the word graph, we can observe their impact on the feature embedding representation of the word graph, thereby understanding how the detection model makes predictions. This is especially helpful in explaining false detections. This masking operation does not require retraining the detection model; it can be achieved by quickly tracking changes in the topology of the word graph using a formula.

[0131] See Figure 2 The diagram shown is a framework diagram of a preferred embodiment of a method for processing false information provided by the present invention. Assuming the text input is: Taking a hot bath prevents the COVID-19 disease, after the input text is processed by word graph construction, feature embedding generation, graph neural network model detection, and masking mechanism to improve interpretability, two output results can be obtained: [1] detection result and feature vector, [2] word misleading ranking; where the feature vector is the feature embedding representation u of the word graph (i.e., the word graph obtained after the input text is constructed) to be input into the graph neural network model, and the form can be a floating-point vector. The detection result is the prediction label of whether the input text is "real" or "false" output by the graph neural network model after performing false information detection on the input feature vector. The form can be "1" or "0", or a probability distribution like [0.87, 0.13].

[0132] It should be noted that this embodiment of the invention combines the PageRank algorithm with a graph neural network model, which enhances the evaluation of the importance of nodes in the word graph structure. The random walk strategy of the PageRank algorithm can more accurately evaluate important nodes in fake information by ranking the importance of nodes, thus helping to improve the accuracy of fake information detection. This combination enables the detection model to capture complex features in heterogeneous information (such as semantics, syntax, and associations). Especially when dealing with fake information with multiple features, the PageRank algorithm can effectively evaluate the importance of nodes and optimize the structure of the word graph, enabling the graph neural network model to better learn and predict fake information.

[0133] Furthermore, traditional deep learning models often face the "black box" problem in misinformation detection, making it difficult to explain why certain information is labeled as misinformation. To overcome this challenge, this invention introduces a masking mechanism that allows masking of nodes and edges in the word graph, calculating their impact on the detection results. By masking nodes and edges, words or sentences that mislead the final detection results can be identified, thereby analyzing the key parts of misinformation propagation. This masking mechanism not only enhances the interpretability of the detection model but also provides a transparent detection process, enabling users to better understand the decision logic of the detection model.

[0134] For example, the prior art provides a real-time information verification system that utilizes the internet and fact databases. This system assesses the credibility of information submitted by users and provides supporting evidence through the internet and fact databases. It innovatively combines position selection and source credibility evaluation to improve verification accuracy. Furthermore, the system employs a two-step training method to automatically select positions from massive documents and uses the credibility of position sources to adjust position scores, thereby reducing the impact of unreliable sources.

[0135] The embodiments of this invention shift the search process to problem understanding (i.e., detection through heterogeneous feature extraction and graph neural network models), thus enabling real-time responses to open-ended queries. Furthermore, these embodiments are built upon a large corpus containing 3 million different words from numerous articles, providing comprehensive downstream knowledge for specific domains with limited information. They accept open-ended queries and offer credibility analysis. In addition, these embodiments provide explanations based on the contribution (positive or negative) of each word to the text reality prediction.

[0136] For example, existing technology also provides a cross-document misinformation detection method based on contrastive graph learning. This method mainly includes the following steps: receiving a family of documents under the same topic; performing misinformation detection at both the document level and the event level; establishing an intra-document knowledge graph based on the content of each document, then building a cross-document knowledge graph, finally encoding the knowledge graph using a graph neural network, and further fusing information using a high-speed network; performing a nonlinear transformation based on the fused representation obtained in the previous step, and then guiding the model to learn in the vector space with the goal of increasing the similarity between positive samples and the difference between positive and negative samples; and outputting the identification result of the misinformation document.

[0137] Compared to the prior art, this invention employs the PageRank algorithm to enhance the random walk strategy in the graph neural network. PageRank is a ranking algorithm based on node importance, which helps detection models better capture the relationships between nodes in the word graph, especially when dealing with heterogeneous features, allowing for more accurate assessment of node importance and influence. In contrast, the prior art relies on constructing a cross-document knowledge graph and encoding it using a graph neural network, depending on a high-speed network for information fusion. The difference between these two methods lies in the fact that this invention achieves deep optimization of the word graph structure through the random walk strategy of the PageRank algorithm, enhancing the dynamic connections between nodes; while the prior art mainly relies on a cross-document structured graph for information fusion, focusing on the construction and encoding of the knowledge graph.

[0138] Furthermore, this embodiment of the invention employs a masking mechanism to enhance the interpretability of the detection model. By masking the nodes and edges in the word graph structure, nodes that mislead the final decision can be identified, thus more transparently explaining the decision logic of the detection model. This interpretability mechanism allows for rapid analysis of the importance of each node through changes in the topology without altering the parameters of the graph neural network model. In contrast, the aforementioned prior art does not mention any mechanism to enhance model interpretability, primarily focusing on performing a nonlinear transformation after fusion representation to optimize the similarity and differences between samples based on vector space, but without further explanation of how the model interprets the results. Therefore, this embodiment of the invention also has a significant advantage in terms of interpretability.

[0139] Therefore, the existing technologies that are similar to the embodiments of the present invention have the following problems: The existing technology provides a system for real-time information verification using networks and fact databases, which has a certain degree of real-time performance and accuracy. However, this technical solution mainly relies on the comparison of known fact databases and network information, and has significant shortcomings when facing complex false information with multiple heterogeneous features. On the one hand, this technical solution lacks sufficient modeling ability for context-sensitive heterogeneous features. For example, the technical solution cannot effectively capture the difference in meaning of the same word in different contexts, which can easily lead to incorrect judgments. On the other hand, this technical solution is a black box model, lacking a transparent explanation mechanism, and cannot clearly explain its detection results, making it difficult to gain user trust.

[0140] To address the aforementioned shortcomings, this invention, based on graph machine learning technology, can effectively model the heterogeneous features in false information and positively utilize these features to solve the challenges in detection and interpretation. Specifically, this invention utilizes a large-scale pre-trained language model to process heterogeneous features through transfer learning, enabling efficient and high-precision detection in specific domains (such as political news) through feature embedding. Furthermore, this invention employs graph-based augmentation techniques. By introducing a masking mechanism, nodes and edges in the text's word graph structure are masked (i.e., nodes and edges are hidden) to adjust the word graph topology and calculate the influence of words on the detection results. This not only improves detection accuracy but also provides a deeper explanation of the contribution of different units (such as words and sentences) in false information detection and identifies nodes that mislead the final decision of the detection model. This more transparently explains the decision logic of the detection model, enhancing its interpretability. This interpretability mechanism allows for rapid analysis of the importance of each node through changes in the word graph topology without altering the detection model's parameters, making the detection results more transparent and credible. It simultaneously meets the interpretability needs of both domain experts and ordinary users.

[0141] Therefore, compared with the prior art, the method proposed in this embodiment of the invention performs better in handling complex and heterogeneous features, can achieve more efficient and accurate false information detection in multiple contexts, and has stronger interpretability, providing a deeper understanding of the detection results, which helps to further enhance human-machine trust and collaboration.

[0142] This invention also provides a false information processing apparatus for implementing the false information processing method described in any of the above embodiments. See [link to relevant documentation]. Figure 3 The diagram shown is a structural block diagram of a preferred embodiment of a false information processing device provided by the present invention. The device includes:

[0143] The word graph construction module 11 is used to construct a word graph based on the text to be detected; wherein, the word graph uses words in the text to be detected as nodes and semantic associations between words as edges;

[0144] Feature extraction module 12 is used to extract multidimensional heterogeneous features of nodes in the word graph using a natural language model;

[0145] Feature embedding module 13 is used to embed features of nodes in the word graph based on the multidimensional heterogeneous features of the nodes, so as to obtain the feature embedding representation of the word graph;

[0146] The fake information detection module 14 is used to input the feature embedding representation of the word graph into the graph neural network model to detect fake information and obtain the detection result;

[0147] The detection result interpretation module 15 is used to detect the degree of misleading effect of nodes in the word graph on the detection results using a masking mechanism, and to obtain the word misleading ranking results.

[0148] Preferably, the word graph construction module 11 specifically includes:

[0149] The text processing unit is used to acquire multiple words in the text to be detected and determine the semantic relationships between the words;

[0150] The word graph construction unit is used to treat each word as a node in the word graph and to build edges in the word graph between words with semantic relationships based on a sliding window.

[0151] Preferably, the feature embedding module 13 specifically includes:

[0152] The node score acquisition unit is used to acquire the PageRank score of each node in the word graph using the PageRank algorithm.

[0153] The node feature embedding unit is used to obtain the feature embedding representation of the corresponding node based on the multidimensional heterogeneous features of the node and the PageRank score of each node;

[0154] The word graph feature embedding unit is used to obtain the feature embedding representation of the word graph based on the feature embedding representations of all nodes of the word graph.

[0155] Preferably, the node feature embedding unit is specifically used for:

[0156] For the i-th node in the word graph, according to formula h i =X·PR(i) TObtain the feature embedding representation of the i-th node; where i is the node index, X is the multidimensional heterogeneous feature of the node, and PR(i) is the PageRank score of the i-th node.

[0157] Preferably, the word graph feature embedding unit is specifically used for:

[0158] According to the formula u=pool(h) i |i∈{1,2,...,n}) to obtain the feature embedding representation of the word graph; where pool is the pooling operation and n is the total number of nodes in the word graph.

[0159] Preferably, the graph neural network model uses cross-entropy as the loss function during training, and the definition of cross-entropy is: Where j is the training sample index, y j z is the label for the j-th training sample, which indicates whether the j-th training sample is false information. j This represents the false information detection result for the j-th training sample.

[0160] Preferably, the detection result interpretation module 15 specifically includes:

[0161] The masking and feature embedding unit is used to mask the nodes in the word graph using a masking mechanism, so that the word graph is converted into a new word graph, and to obtain a new feature embedding representation of the new word graph.

[0162] The new embedded feature detection unit is used to input the new feature embedded representation of the new word graph into the graph neural network model to detect false information and obtain new detection results.

[0163] The misleading degree detection unit is used to compare the difference in correct prediction probability between the new feature embedding representation of the new word graph and the feature embedding representation of the word graph based on the new detection result and the detection result, and to obtain the misleading degree of the masked node that misleads the detection result;

[0164] The detection result interpretation unit is used to obtain a word misleading ranking result based on the degree of misleading effect of each node in the word graph on the detection result after it is masked.

[0165] Preferably, the masking and feature embedding unit obtains a new feature embedding representation of the new word graph, specifically including:

[0166] Based on the topological changes of the new word graph and the PageRank score of each node in the word graph, obtain the new PageRank score of each node in the new word graph;

[0167] Based on the multidimensional heterogeneous features of the nodes and the new PageRank score of each node, obtain the new feature embedding representation of the corresponding node;

[0168] The feature embedding representation of the new word graph is obtained based on the new feature embedding representation of all nodes in the new word graph.

[0169] Preferably, the masking and feature embedding unit obtains a new PageRank score for each node in the new word graph based on the topological changes of the new word graph and the PageRank score of each node in the word graph, specifically including:

[0170] For the i-th node in the new word graph, according to the formula Obtain the new PageRank score for the i-th node; where i is the node index, PR(i) is the PageRank score of the i-th node, α is the damping factor, M' is the column random transition matrix of the new word graph, and PR(i) pushout The change in PageRank score of the i-th node caused by the topological change of the new word graph.

[0171] Preferably, the formula for calculating the change in the PageRank score of the i-th node is: PR(i) pushout =α(M'-M)PR(i); where M is the column random transition matrix of the word graph.

[0172] It should be noted that the false information processing device provided in this embodiment of the invention can implement all the processes of the false information processing method described in any of the above embodiments. The functions and technical effects of each module and unit in the device are the same as the functions and technical effects of the false information processing method described in the above embodiments, and will not be repeated here.

[0173] This invention also provides a computer-readable storage medium including a stored computer program, which, when running, controls the device where the computer-readable storage medium is located to execute the misinformation processing method described in any of the above embodiments.

[0174] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the false information processing method described in any of the above embodiments.

[0175] This invention also provides a terminal device, see [link to relevant documentation]. Figure 4The diagram shown is a structural block diagram of a preferred embodiment of a terminal device provided by the present invention. The terminal device includes a processor 10, a memory 20, and a computer program stored in the memory 20 and configured to be executed by the processor 10. When the processor 10 executes the computer program, it implements the false information processing method described in any of the above embodiments.

[0176] Preferably, the computer program can be divided into one or more modules / units (such as computer program 1, computer program 2, ...), and the one or more modules / units are stored in the memory 20 and executed by the processor 10 to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in the terminal device.

[0177] The processor 10 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor, or the processor 10 may be any conventional processor. The processor 10 is the control center of the terminal device, connecting various parts of the terminal device through various interfaces and lines.

[0178] The memory 20 mainly includes a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc., while the data storage area can store related data, etc. Furthermore, the memory 20 can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard disk, a smart media card (SMC), a secure digital card (SD), and a flash card, or other volatile solid-state storage devices.

[0179] It should be noted that the aforementioned terminal devices may include, but are not limited to, processors and memory, as will be understood by those skilled in the art. Figure 4The structural block diagram shown is merely a structural example of the terminal device described above and does not constitute a limitation on the structure of the terminal device. The terminal device may include more or fewer components than shown, or combine certain components, or use different components.

[0180] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for processing false information, characterized in that, include: A word graph is constructed based on the text to be detected; wherein, the word graph uses words in the text to be detected as nodes and semantic relationships between words as edges; Multidimensional heterogeneous features of nodes in the word graph are extracted using a natural language model; Based on the multidimensional heterogeneous features of the nodes, feature embedding is performed on the nodes in the word graph to obtain the feature embedding representation of the word graph; The feature embedding representation of the word graph is input into a graph neural network model for false information detection to obtain the detection result; The degree of misleading effect of nodes in the word graph on the detection results is detected by using a masking mechanism, and the misleading ranking results of words are obtained.

2. The method for processing false information as described in claim 1, characterized in that, The construction of a word map based on the text to be detected specifically includes: The process involves acquiring multiple words from the text to be detected and determining the semantic relationships between the words. Each word is treated as a node in the word graph, and edges are created between words that are semantically related based on a sliding window.

3. The method for processing false information as described in claim 1, characterized in that, The step of embedding features into the nodes of the word graph based on the multidimensional heterogeneous features of the nodes to obtain the feature embedding representation of the word graph specifically includes: The PageRank algorithm is used to obtain the PageRank score of each node in the word graph; The feature embedding representation of the corresponding node is obtained based on the multidimensional heterogeneous features of the node and the PageRank score of each node; The feature embedding representation of the word graph is obtained based on the feature embedding representation of all nodes of the word graph.

4. The method for processing false information as described in claim 3, characterized in that, The step of obtaining the feature embedding representation of the corresponding node based on the multidimensional heterogeneous features of the node and the PageRank score of each node specifically includes: For the i-th node in the word graph, according to formula h i =X·PR(i) T Obtain the feature embedding representation of the i-th node; where i is the node index, X is the multidimensional heterogeneous feature of the node, and PR(i) is the PageRank score of the i-th node.

5. The method for processing false information as described in claim 4, characterized in that, The step of obtaining the feature embedding representation of the word graph based on the feature embedding representation of all nodes of the word graph specifically includes: According to the formula u=pool(h) i |i∈{1,2,...,n}) to obtain the feature embedding representation of the word graph; where pool is the pooling operation and n is the total number of nodes in the word graph.

6. The method for processing false information as described in claim 1, characterized in that, The graph neural network model uses cross-entropy as the loss function during training, and the definition of cross-entropy is: Where j is the training sample index, y j z is the label for the j-th training sample, which indicates whether the j-th training sample is false information. j This represents the false information detection result for the j-th training sample.

7. The method for processing false information as described in claim 1, characterized in that, The step of using a masking mechanism to detect the degree of misleading effect of nodes in the word graph on the detection results, and obtaining a word misleading ranking result, specifically includes: The nodes in the word graph are masked using a masking mechanism, which transforms the word graph into a new word graph, and a new feature embedding representation of the new word graph is obtained. The new feature embedding representation of the new word graph is input into the graph neural network model for false information detection to obtain new detection results; Based on the new detection results and the detection results, compare the difference in correct prediction probability between the new feature embedding representation of the new word graph and the feature embedding representation of the word graph to obtain the degree of misleading effect of the masked nodes on the detection results; The misleading ranking result of words is obtained based on the degree of misleading effect of each node in the word graph after it is masked on the detection result.

8. The method for processing false information as described in claim 7, characterized in that, The acquisition of the new feature embedding representation of the new word graph specifically includes: Based on the topological changes of the new word graph and the PageRank score of each node in the word graph, obtain the new PageRank score of each node in the new word graph; Based on the multidimensional heterogeneous features of the nodes and the new PageRank score of each node, obtain the new feature embedding representation of the corresponding node; The feature embedding representation of the new word graph is obtained based on the new feature embedding representation of all nodes in the new word graph.

9. The method for processing false information as described in claim 8, characterized in that, The step of obtaining a new PageRank score for each node in the new word graph based on the topological changes of the new word graph and the PageRank score of each node in the word graph specifically includes: For the i-th node in the new word graph, according to the formula Obtain the new PageRank score for the i-th node; where i is the node index, PR(i) is the PageRank score of the i-th node, α is the damping factor, M' is the column random transition matrix of the new word graph, and PR(i) pushout The change in PageRank score of the i-th node caused by the topological change of the new word graph.

10. The method for processing false information as described in claim 9, characterized in that, The formula for calculating the change in the PageRank score of the i-th node is: PR(i) pushout =α(M'-M)PR(i); where M is the column random transition matrix of the word graph.

11. A device for processing false information, characterized in that, include: A word graph construction module is used to construct a word graph based on the text to be detected; wherein, the word graph uses words in the text to be detected as nodes and semantic relationships between words as edges; The feature extraction module is used to extract multidimensional heterogeneous features of nodes in the word graph using a natural language model; The feature embedding module is used to embed features of nodes in the word graph based on the multidimensional heterogeneous features of the nodes, so as to obtain the feature embedding representation of the word graph; The fake information detection module is used to input the feature embedding representation of the word graph into the graph neural network model to detect fake information and obtain the detection result; The detection result interpretation module is used to detect the degree of misleading effect of nodes in the word graph on the detection results using a masking mechanism, and to obtain the word misleading ranking results.

12. A computer-readable storage medium, characterized in that, The device includes a stored computer program that, when executed, controls the device containing the computer-readable storage medium to perform the misinformation processing method as described in any one of claims 1 to 10.

13. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the false information processing method as described in any one of claims 1 to 10.

14. A terminal device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the method for processing false information as described in any one of claims 1 to 10.

Images