A fake news detection method based on heuristic heterogeneous graph reasoning network

Abstract

The present application relates to a kind of false news detection method based on heuristic heterogeneous graph reasoning network, comprising the following steps: the news sequence to be detected, table, procedure are represented in context;For news sequence, table and procedure three types of data, three subgraphs are respectively constructed by means of different methods;The matching connection of program subgraph and table subgraph is guided by news sequence, to form program-table associated graph;The matching connection of table subgraph is guided by program information news sequence subgraph, to form news-table associated graph;Design local multi-hop knowledge reasoning network MKR to learn evidence with more context features, and MKR is executed on news-table associated subgraph and program-table associated subgraph to capture language evidence and logic evidence respectively;Design global double-end attention network DAN, to globally excavate the association between language evidence and logic evidence.

Description

Technical Field

[0001] The present invention belongs to the field of electronic information technology, and in particular relates to a false news detection method based on a heuristic heterogeneous graph reasoning network. Background Art

[0002] With the rapid development of the internet, social media has become a vital platform for information sharing and exchange, thanks to its advantages of convenient publishing, rapid dissemination, and low access costs. The vast amount of information disseminated on social media not only contains accurate and reliable real information, but also contains a large amount of deliberately fabricated false information or unintentionally disseminated misinformation. Therefore, how to timely and effectively detect false information disseminated on social media has become a pressing issue in the social media field.

[0003] Existing research primarily focuses on unstructured data, such as news text. Detecting unstructured false information primarily involves collecting large amounts of objective, factual information to verify the authenticity of the information being detected, and significant progress has been made in this area. However, social media contains not only unstructured information but also a large amount of structured data (such as tabular information). There is a huge demand and promising future for extracting useful information from tabular information as evidence to verify the authenticity of false information. Furthermore, false news detection based on structured information is relatively rare and challenging in current research, making it a key research topic.

[0004] Currently, research on disinformation detection using structured data falls into two main categories. The first category is pre-training-based methods, which are suitable for scenarios where both structured and unstructured data coexist. They typically leverage pre-trained models to learn deep contextual semantics to improve detection performance. However, the limitation of these methods is that the generality of the semantics learned by pre-trained models makes them difficult to achieve accurate reasoning. The second category of structured disinformation detection methods is graph-based. These methods typically first construct a reasonable heterogeneous graph, then focus on linguistic evidence through semantic reasoning and learn logical evidence through symbolic reasoning. Finally, they fuse these two types of evidence to improve detection performance. While both of these approaches can achieve relatively good detection performance, they still suffer from several key issues: 1) Traditional graph construction suffers from data sparsity. In traditional graph construction, subgraphs are typically connected by connecting nodes with the same content, which can easily lead to data sparsity, especially when semantics are limited across different subgraphs. 2) There is a lack of effective correlation between linguistic and logical evidence. Existing models typically learn linguistic and logical evidence independently across the entire graph structure, which can lead to a lack of correlation and fusion between the two types of evidence and difficulty learning valuable features between them. Therefore, how to overcome the problem of data sparsity when constructing the graph and establish effective fusion and correlation communication between multiple evidences during evidence learning is a key issue in fake news detection based on structured data. Summary of the Invention

[0005] Technical problems to be solved

[0006] To address the shortcomings of current false information detection methods based on structured data, this paper proposes a false news detection method based on a heuristic heterogeneous graph reasoning network (H2GRN). This method first constructs a heuristic heterogeneous graph network that not only expands the connection between the information to be detected and structured tables, but also strengthens the connection between the news to be detected and the program subgraph, effectively overcoming the sparsity of connections between subgraphs. It then constructs a multi-perspective structured reasoning method that learns linguistic and logical evidence from both local and global perspectives and strengthens the correlation and fusion between the two types of evidence, thereby capturing consistent shared evidence between the two for table-based false information detection.

[0007] Technical Solution

[0008] A fake news detection method based on a heuristic heterogeneous graph reasoning network is characterized by the following steps:

[0009] S1: Context Representation Module

[0010] The pre-trained BERT model is used to represent the context of the news sequence, table, and program to be detected;

[0011] S2: Subgraph building blocks

[0012] Three subgraphs are constructed for news sequence, table, and program data using different methods. The news sequence subgraph is constructed using the dependency parse tree method, the table subgraph is constructed by connecting the table header nodes with the table body nodes, and the program subgraph is constructed using the latent program algorithm (LPA).

[0013] S3: Heterogeneous Graph Building Blocks

[0014] The news sequence guides the matching connection between the program subgraph and the table subgraph, thereby forming a program-table association graph; the program information guides the matching connection between the news sequence subgraph and the table subgraph, thereby forming a news-table association graph;

[0015] S4: Multi-view reasoning module

[0016] A local multi-hop knowledge reasoning network (MKR) is designed to learn evidence with more contextual features, and MKR is executed on the news-table association subgraph and the program-table association subgraph to capture linguistic evidence and logical evidence respectively. A global two-terminal attention network (DAN) is designed to globally mine the association relationship between linguistic evidence and logical evidence.

[0017] A further technical solution of the present invention is as follows: the S1 program is generated by the news information to be detected and the table through logical symbols.

[0018] A further technical solution of the present invention: S2 comprises the following steps:

[0019] S21: News subgraph construction: Use the dependency parse tree to convert the news sequence into a news dependency parse subgraph. In particular, each node in the parse subgraph represents a word or phrase, which depends on the specific word content in the table or the keyword in the program;

[0020] S22: Construction of table subgraph: In order to convert the table sequence into a graph form, the table nodes are converted into table header nodes and table content nodes. Each node in the table content node is connected not only to its table header node but also to other nodes in the row.

[0021] S23: Construction of program subgraph: Use the latent program algorithm to construct program synthesis, where the program consists of two types of data: parameters and functions. Because the LPA algorithm cannot generate completely accurate programs, this program uses a selection strategy to screen the top two best candidate programs as synthetic programs. The selection strategy is to train a two-way encoder based on a converter to obtain label-consistent programs as candidate programs.

[0022] A further technical solution of the present invention: S3 comprises the following steps:

[0023] S31: News sequence guides the matching connection between program subgraph and table subgraph: first, nodes with logical information in the news are retrieved, and then the information connected to the node is used as additional knowledge to expand the connection between program and table association subgraph;

[0024] S32: Program information guides the matching connection between the news sequence subgraph and the table subgraph: The program subgraph is a tree structure composed of leaf nodes and parent nodes. Two types of nodes are designed as external knowledge to enhance the connection of the news-table association subgraph, namely, leaf-leaf association nodes and association nodes between parent nodes and leaf nodes; the leaf nodes in the leaf-leaf association nodes have the same parent node; a leaf node in the association nodes between parent nodes and leaf nodes is associated with its parent node.

[0025] A further technical solution of the present invention: S4 comprises the following steps:

[0026] S41: For a node v i , not only learn its one-hop neighbor nodes Also learn k-hop neighbor nodes In this way, the graph attention network GAN considers the neighbor vertex vectors and multi-hop neighbor vertex vectors Thus updating v i Each new vertex vector Among them I k Represents node v i The number of k-th hop neighbors of

[0027]

[0028] Among them, λ1,…,λ k-1 Represents node v i Hyperparameters affected by k-hop nodes, W is a trainable parameter;

[0029] S42: Node v i The one-hop neighbor weight γ ij and k-hop neighbor weights The calculation is as follows:

[0030]

[0031] Among them, f(·) is a LeakyReLU function. When X ij =γ ij When x i with x j Respectively represent s i and s j ;when When xi with x j Respectively represent s i and is the k-th hop node updated by the graph attention network;

[0032] S43: Tight graph representation for the entire graph network by updating all nodes The pooling operation obtains:

[0033]

[0034] S44: Through MKR, we continuously update and learn on the news-table association subgraph and the program-table association subgraph, thereby capturing language evidence H=H with more contextual semantics. lg And logical evidence H=H lo ;

[0035] S45: Considering that each node in a graph network may contain two types of neighboring nodes, namely nodes in the same subgraph and nodes not in the same subgraph; the global two-terminal attention network DAN designs two reasoning mechanisms: intra-graph attention mechanism and inter-graph attention mechanism:

[0036] S46: In-graph attention mechanism: Formally, given a specific node v and its neighbor node v′, based on the node representation s v and s v′ Calculate the in-graph attention score:

[0037] x vv′ =LeakyReLU(W x [W v s v ||s v′ W v′ ]) (4)

[0038]

[0039]

[0040] Among them, all W are trainable parameters, α vv′ is the attention weight of node v and its neighbor node v′, N v represents the number of neighbor nodes of node v, ';' represents the splicing operation, σ(·) is the activation function, u v It is the learnable internal structure of the subgraph where node v is located;

[0041] S47: Design an inter-graph attention mechanism to measure the association between adjacent nodes between subgraphs:

[0042] z ij=LeakyReLU(W z γ i [W i s i ||s j W j ]) (7)

[0043] S47: Exploring Node v Using Graph Attention Network i The correlation feature ρ between the corresponding subgraphs i Finally, the bidirectional attention mechanism between graphs and within graphs is integrated into the graph convolutional network:

[0044] H (1) =Pool(u1; ...; u v ;…;ρ i ;…;ρ I ) (8)

[0045]

[0046]

[0047] Among them, V and I represent the node v i The number of nodes within and between subgraphs, H (1) Represents node v i The average pooling representation of the first layer in the entire network graph, v l is the regularization weight value of each layer, L is the number of superimposed layers, all W and b are trainable parameters, and finally, the global evidence obtained by the present invention is H ho ;

[0048] S47: A consistency fusion layer consisting of three identical modules is designed. lg and global evidence H ho The details of each module are as follows:

[0049] Use two BiLSTMs to encode the language evidence H lg and global evidence H ho is m lg and m ho ; Then, the attention mechanism is used to match the salient features between the two pieces of evidence; for the jth word of the global evidence The calculation of significant features is as follows:

[0050]

[0051]

[0052] in, and Respectively The original association measure and the standardized association measure;

[0053] Language-guided global representation v ho It can be calculated by the following formula:

[0054]

[0055] Using the element-wise summation method V ho and the encoded global evidence m ho Combined with m lg The concatenation is performed and input into the fully connected layer to obtain a low-dimensional consistent representation of s between the language evidence and the global evidence. loho :

[0056]

[0057] s lgho =MLP([n ho ;m lg ]) (15)

[0058] Accordingly, the consistent representation of logical evidence and global evidence s loho are also calculated in the same way; finally, these two consistent representations s loho With s loho Spliced ​​together as consistent shared evidence of the three types of evidence lglo ;

[0059] S48: The Softmax function predicts the probability distribution of the model and performs cross entropy training through the global loss function:

[0060] p=softmax(W p s lglo +b p ) (16)

[0061] loss=-∑ylogp (17)

[0062] Among them, y is the true label.

[0063] A computer system, characterized in that it includes: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method.

[0064] A computer-readable storage medium is characterized by storing computer-executable instructions, which are used to implement the above method when executed.

[0065] A computer program, characterized by comprising computer executable instructions, wherein the instructions are used to implement the above method when executed.

[0066] Beneficial effects

[0067] This paper proposes a fake news detection method based on a heuristic heterogeneous graph reasoning network. By novelly proposing a heuristic graph construction and building hierarchical heterogeneous graph reasoning, it strengthens the connection between linguistic and logical evidence and exploits the shared evidence between them to detect fake news based on structured data. This method not only improves fake news detection performance but also provides effective evidence for interpreting detection results from structured data. Compared with existing technologies, this paper offers the following innovations:

[0068] 1: The innovative heuristic graph construction model proposed in this paper improves the data sparsity in traditional graph construction and enhances the connectivity between the language subgraph and the logical subgraph;

[0069] 2: The local multi-hop knowledge reasoning network designed by this invention not only captures the contextual semantics of adjacent one-hop nodes, but also learns the contextual information of multi-hop related nodes. It can effectively mine linguistic evidence and logical evidence when executed on the news-table association subgraph and the program-table association subgraph respectively.

[0070] 3: The global two-terminal attention network constructed by this invention designs an intra-graph attention mechanism and an inter-graph attention mechanism to specifically strengthen the importance of nodes within the same subgraph and nodes in different subgraphs, thereby learning global consistency evidence between linguistic evidence and logical evidence through different information reasoning mechanisms;

[0071] Extensive experiments on two competitive fake news detection datasets demonstrate the effectiveness and interpretability of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0073] Figure 1 This is a diagram of the architecture of the present invention;

[0074] Figure 2 This is the experimental performance diagram of the present invention under the TABFACT and FEVEROUS data sets;

[0075] Figure 3 This is a comparison chart of the separation performance of different modules of the present invention under the TABFACT and FEVEROUS data sets. DETAILED DESCRIPTION

[0076] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0077] In response to the defects in the current false information detection methods based on structured data, the present invention proposes a false news detection method based on a heuristic heterogeneous graph reasoning network (H2GRN), which first constructs a heuristic heterogeneous graph network, which not only expands the connection between the information to be detected and the structured table, but also strengthens the connection between the news to be detected and the program subgraph, effectively overcoming the sparsity of the connection between the subgraphs. Then, a multi-perspective structured reasoning method is constructed, which learns language evidence and logical evidence from local and global perspectives and strengthens the correlation and fusion between the two types of evidence, thereby capturing the consistent shared evidence between the two for table-based false information detection. The architecture diagram of the present invention is shown in FIG. Figure 1 As shown, it includes the following four modules:

[0078] Module 1. Context representation module.

[0079] The invention uses a pre-trained BERT model to learn the contextual representation of news sequences, tables, and programs to be detected (the program is generated by logical symbols from the information to be detected and the table);

[0080] Module 2. Subgraph construction module.

[0081] This module constructs three subgraphs for three types of data: news sequence, table, and program using different methods. The first is to construct a news sequence subgraph using the dependency parse tree method, to construct a table subgraph by connecting the table header node with the table body node, and to construct a program subgraph using the LPA algorithm.

[0082] Module 3. Heterogeneous graph construction module.

[0083] Based on the traditional heterogeneous graph construction, this module uses news sequences to guide the connection between program subgraphs and table subgraphs, thus forming a program-table association graph. It also uses program information to guide the construction of news sequence subgraphs and table subgraphs, thus forming a news-table association graph. This mutual heuristic guidance method effectively strengthens the closeness of subgraph connections and overcomes the defect of data sparsity.

[0084] Module 4. Multi-view reasoning module.

[0085] This module designs a local multi-hop knowledge reasoning network (MKR) to learn evidence with more contextual features. MKR is executed on the news-table association subgraph and the program-table association subgraph to capture linguistic evidence and logical evidence, respectively. This module also designs a global dual-end attention network (DAN) to globally mine the association between linguistic and logical evidence.

[0086] The process of the present invention is specifically as follows:

[0087] Phase 1: Context Representation Module

[0088] Step 1: This invention uses a pre-trained BERT model to contextually represent the news sequence to be tested, the table, and the program (using symbolic information to convert the news and table into a program). Specifically, for learning table sequences, to effectively learn the contextual semantics of the table's main content and column names, this invention uses a template-based approach (column names and content concatenated) to convert the table into a table sequence.

[0089] Step 2: The present invention inputs the news sequence, table sequence, and program into the BERT model in the form of [[CLS], S, [SEP], T, [SEP], F, [SEP]], where [CLS] is a special character used for classification and [SEP] is a terminator. S, T, and F represent the news sequence, table sequence, and program, respectively.

[0090] Phase 2: Subgraph building blocks

[0091] Step 3: The present invention constructs different subgraphs to describe the internal semantics of three types of data: news, tables, and programs.

[0092] Step 4: Construction of news subgraph. The present invention uses dependency parse tree to convert the news sequence into a news dependency parse subgraph. In particular, each node in the parse subgraph represents a word or phrase, which depends on the specific word content in the table or the keyword in the program. For example, if the word "less than" in the news sequence is a keyword in the program, then this word will be treated as an independent node in the parse subgraph without further segmentation. Representation of node i It is learned by BERT.

[0093] Step 5: Construction of table subgraph. In order to convert the table sequence into a graph, the present invention converts the table nodes into table header nodes and table content nodes, which is inspired by the graph structure of table-based question-answering. Figure 1 As shown, each node in the table content node is connected not only to its table header node, but also to other nodes in the row.

[0094] Step 6: Construction of program subgraph. In order to explore the internal relationship between news sequences and table sequences from the perspective of logical symbols, the present invention converts the two into a program with a logical form. Specifically, the latent program algorithm (LPA algorithm) is used to construct program synthesis, where the program consists of two types of data: parameters (such as string, bool or sub-table, etc.) and functions (the algorithm includes more than 50 functions, such as count(·), argmax(·)). Because the LPA algorithm cannot generate completely accurate programs, this program uses a selection strategy to screen the top two optimal candidate programs as synthetic programs. Among them, the selection strategy is to train a two-way encoder based on a converter to obtain label-consistent programs as candidate programs.

[0095] Phase 3: Heterogeneous Graph Building Blocks

[0096] Step 7: The present invention proposes a heuristic matching mechanism based on the traditional heterogeneous graph construction model to strengthen the association between language information and logical information.

[0097] Step 8: Traditional heterogeneous graph construction. This invention integrates news sequences, tables, and program subgraphs to construct a traditional heterogeneous graph. This heterogeneous graph construction is primarily performed from the following aspects: 1) News-table association subgraph: news nodes are associated with nodes with the same content in the table; 2) Program-table association subgraph: program nodes are associated with nodes with the same content in the program. To enable more flexible information transfer between different subgraphs, the edges of these graphs are all undirected.

[0098] Step 10: Heuristic matching mechanism. The present invention finds that connections between subgraphs in traditional heterogeneous graphs only connect nodes with the same content, which easily leads to data sparsity and hinders communication between different subgraphs. To this end, the present invention constructs a heuristic heterogeneous graph, which primarily includes two heuristic matching mechanisms: news content guides matching connections in the program-table association subgraph, and program guides matching connections in the news-table subgraph.

[0099] Step 11: Matching connection between news content guidance program and table association subgraph. The present invention first retrieves the node with logical information in the news, and then uses the information connected to the node as additional knowledge to expand the connection between the program and the table association subgraph. Figure 1In the example, the "less than" node in the news, which carries logical information, and its connected dependent nodes, "score" and "270," serve as additional knowledge to guide the connection of the program-table subgraph. This allows the "score" node in the program to connect not only to the "score" node in the table, but also to the "270" node in the table. This heuristic approach allows the connection of the program-table association subgraph to incorporate the dependencies of the news, strengthening the relationship between news semantics and program logic information.

[0100] Step 12: The program guides the matching connection of the news-table subgraph. The program subgraph is a tree structure consisting of leaf nodes and parent nodes. Therefore, we design two types of nodes as external knowledge to enhance the connection of the news-table association subgraph, namely leaf-leaf association nodes (two leaf nodes have the same parent node) and parent-leaf association nodes (a leaf node is associated with its parent node). Specifically, in Figure 1 For leaf-leaf nodes, the node "2004" in the news is not only connected to the node "2004" in the table, but also connected to the node "year"; for parent-leaf nodes, the node "score" in the news is not only connected to the node "score" in the table, but also directly connected to the nodes "272, 270, 273, 279" in the table.

[0101] Stage 4: Multi-view Reasoning Module

[0102] Step 13: Local Multi-hop Knowledge Reasoning Network (MKR): The present invention designs MKR to learn evidence containing more contexts, and executes it on the news-table association subgraph and the program-table association subgraph respectively to learn language evidence and logical evidence with rich contextual semantics.

[0103] Step 14: For a node v i , the present invention not only learns its one-hop neighbor nodes Also learn k-hop neighbor nodes In this way, the Graph Attention Networks (GAN) considers the neighbor vertex vectors and multi-hop neighbor vertex vectors (I k Represents node v i The number of k-th hop neighbors) to update v i Each new vertex vector

[0104]

[0105] Among them, λ1,…,λ k-1Represents node v i Hyperparameters affected by k-hop nodes. Experimental results show that when the performance of the model reaches the optimal value, λ1,…,λ k-1 It usually shows a downward trend. W is a trainable parameter.

[0106] Step 15: Node v i The one-hop neighbor weight γ ij and k-hop neighbor weights The calculation is as follows:

[0107]

[0108] Where f(·) is a LeakyReLU function. ij =γ ij When x i with x j Respectively represent s i and s j .when When x i with x j Respectively represent s i and is the k-th hop node updated by the graph attention network.

[0109] Step 16: In order to fully integrate the semantic features of each node, the dense graph representation of the entire graph network can be obtained by updating all nodes. The pooling operation obtains:

[0110]

[0111] Step 17: Finally, MKR is used to continuously update the learning on the news-table association subgraph and the program-table association subgraph, thereby capturing language evidence H=H with more contextual semantics. lg And logical evidence H=H lo .

[0112] Step 18: Global Dual-Terminal Attention Network (DAN). This paper designs a global dual-terminal attention network (DAN) to globally mine the association between linguistic evidence and logical evidence.

[0113] Step 19: Considering that each node in a graph network may contain two types of neighboring nodes, namely nodes in the same subgraph and nodes outside the same subgraph, DAN designs two reasoning mechanisms: intra-graph attention mechanism and inter-graph attention mechanism:

[0114] Step 20: In-graph attention mechanism. In order to enable each node to learn the internal structure of its subgraph, the present invention designs an in-graph attention mechanism to learn the importance of different adjacent nodes. Formally, given a specific node v and its neighbor node v′, the present invention is based on the node representation s v and s v′ (The representations of these two nodes are obtained by BERT) Calculate the in-graph attention score:

[0115] x vv′ =LeakyReLU(W x [W v s u ||s v′ W v′ ]) (4)

[0116]

[0117]

[0118] Among them, all W are trainable parameters. vv′ is the attention weight of node v and its neighbor node v′, N v Indicates the number of neighbor nodes of node v. ';' indicates the concatenation operation. σ(·) is the activation function. u v It is the learnable internal structure of the subgraph where node v is located.

[0119] Step 21: Inter-graph attention mechanism. The present invention designs an inter-graph attention mechanism to measure the association relationship between adjacent nodes between subgraphs. Different from the attention score in formula (1), for a given specific node s i and its neighbor nodes s in different subgraphs j , this mechanism fully considers the importance between different subgraphs to calculate the inter-graph attention score:

[0120] z ij =LeakyReLU(W z γ i [W i s i ||s j W j ]) (7)

[0121] Step 22: Next, the present invention uses a graph attention network to explore node v i The correlation feature ρ between the corresponding subgraphs i Finally, the bidirectional attention mechanism between graphs and within graphs is integrated into the graph convolutional network:

[0122] H (1) =Pool(u1; ...; uv ;…;ρ i ;…;ρ I ) (8)

[0123]

[0124]

[0125] Among them, V and I represent the node v i The number of nodes within and between subgraphs. (1) Represents node v i Average pooling representation of the first layer in the entire network graph. v l is the regularization weight value of each layer. L is the number of superimposed layers. All W and b are trainable parameters. Finally, the global evidence obtained by the present invention is H ho .

[0126] Step 23: Consistency Fusion Layer. To integrate these three types of evidence, the present invention designs a consistency fusion layer consisting of three identical modules, the purpose of which is to weaken the differences between features and thus learn consistent shared evidence. lg and global evidence H ho The details of each module are as follows.

[0127] Step 24: First, the present invention uses two BiLSTMs to encode the language evidence H lg and global evidence H ho is m lg and m ho Then, the attention mechanism is used to match the salient features between the two pieces of evidence. For the jth word of the global evidence The calculation of significant features is as follows:

[0128]

[0129]

[0130] in, and Respectively The original association measure and the standardized association measure.

[0131] Step 25: Language-guided global representation v ho It can be calculated by the following formula:

[0132]

[0133] Step 26: In order to strengthen the global evidence semantics that is closely related to the language evidence semantics, the present invention first uses the element summation method V ho and the encoded global evidence m ho Combined with m lg The concatenation is performed and input into the fully connected layer to obtain a low-dimensional consistent representation of s between the language evidence and the global evidence. loho :

[0134]

[0135] s lgho =MLP([n ho ;m lg ]) (15)

[0136] Step 27: Accordingly, the consistent representation of logical evidence and global evidence s loho Finally, these two consistent representations s loh With s loho Spliced ​​together as consistent shared evidence of the three types of evidence lgloh .

[0137] Step 28: The Softmax function predicts the probability distribution of the model and performs cross entropy training through the global loss function:

[0138]

[0139] loss=-∑ylogp (17)

[0140] The above method is applicable to social network environment and can provide structured table-based evidence in social media network environment.

[0141] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present invention, and these modifications or replacements should all be included in the scope of protection of the present invention.

Claims

1. A fake news detection method based on heuristic heterogeneous graph reasoning network, characterized by Here are the steps: S1: Context Representation Module The pre-trained BERT model is used to represent the context of the news sequence, table, and program to be detected; S2: Subgraph building blocks Three subgraphs are constructed for news sequence, table, and program data using different methods. The news sequence subgraph is constructed using the dependency parse tree method, the table subgraph is constructed by connecting the table header nodes with the table body nodes, and the program subgraph is constructed using the latent program algorithm (LPA). S3: Heterogeneous Graph Building Blocks The news sequence guides the matching connection between the program subgraph and the table subgraph, thereby forming a program-table association graph; the program information guides the matching connection between the news sequence subgraph and the table subgraph, thereby forming a news-table association graph; S4: Multi-view reasoning module We design a local multi-hop knowledge reasoning network (MKR) to learn evidence with more contextual features, and execute MKR on the news-table association subgraph and the program-table association subgraph to capture linguistic evidence and logical evidence, respectively. We also design a global two-terminal attention network (DAN) to globally mine the association between linguistic evidence and logical evidence. S4 specifically includes the following steps: S41: For one node , not only learn its one-hop neighbor nodes , and also learn k-hop neighbor nodes ; In this way, the graph attention network GAN considers the neighbor vertex vectors and multi-hop neighbor vertex vectors Thus update Each new vertex vector ;in Representation node The number of k-th hop neighbors of (1) in, Representation node Hyperparameters affected by k-hop nodes, W is a trainable parameter; S42: Node One-hop neighbor weight and k-hop neighbor weights The calculation is as follows: (2) in, Is a LeakyReLU function, when hour, and Respectively and ;when hour, and Respectively and ; is the k-th hop node updated by the graph attention network; S43: Tight graph representation for the entire graph network by updating all nodes The pooling operation obtains: (3) S44: MKR continuously updates and learns on the news-table association subgraph and the program-table association subgraph, thereby capturing language evidence with more contextual semantics. and logical evidence ; S45: Considering that each node in a graph network may contain two types of neighboring nodes, namely nodes in the same subgraph and nodes not in the same subgraph; the global two-terminal attention network DAN designs two reasoning mechanisms: intra-graph attention mechanism and inter-graph attention mechanism: S46: In-graph attention mechanism: Formally, given a specific node and its neighbor nodes , based on node representation and Calculate the in-graph attention score: (4) (5) (6) Among them, all W are trainable parameters, is a node and its neighbor nodes The attention weight, Representation node The number of neighbor nodes, ';' represents the splicing operation, is the activation function, is a node The internal structure that can be learned in the subgraph; S47: Design an inter-graph attention mechanism to measure the association between adjacent nodes between subgraphs: (7) S47: Exploring Nodes Using Graph Attention Networks Correlation features between corresponding subgraphs Finally, the bidirectional attention mechanism between graphs and within graphs is integrated into the graph convolutional network: (8) (9) (10) Among them, V and I represent nodes respectively. The number of nodes within and between subgraphs of Representation node The average pooling representation of the first layer in the entire network graph, is the regularized weight value of each layer, L is the number of superimposed layers, and all W and b are trainable parameters. ; S47: A consistency fusion layer consisting of three identical modules is designed for language evidence. and global evidence The details of each module are as follows: Use two BiLSTMs to encode language evidence separately and global evidence for and ; Then, the attention mechanism is used to match the salient features between the two pieces of evidence; j words , the calculation of significant features is as follows: (11) (12) in, and Respectively The original association measure and the standardized association measure; Language-guided global representation It can be calculated by the following formula: (13) Using the element-wise summation method Will and global evidence of encoding Combined with concatenate and input into the fully connected layer to obtain a low-dimensional consistent representation between language evidence and global evidence : (14) (15) Accordingly, the consistent representation of logical evidence and global evidence are also calculated in the same way; finally, these two consistent representations and Spliced ​​together as consistent shared evidence of the three types of evidence ; S48: Use the Softmax function to predict the probability distribution of the model and perform cross entropy training through the global loss function: (16) (17) Among them, y is the true label.

2. The fake news detection method based on a heuristic heterogeneous graph reasoning network according to claim 1 is characterized by: The S1 program is generated by the news information to be detected and the table through logical symbols.

3. The fake news detection method based on heuristic heterogeneous graph reasoning network according to claim 2 is characterized by: The S2 comprises the following steps: S21: Construction of news subgraph: Use dependency parse tree to convert the news sequence into a news dependency parse subgraph. Each node in the parse subgraph represents a word or phrase, which depends on the specific word content in the table or the keyword in the program. S22: Construction of table subgraph: In order to convert the table sequence into a graph form, the table nodes are converted into table header nodes and table content nodes. Each node in the table content node is connected not only to its table header node but also to other nodes in the row. S23: Construction of program subgraph: Use the latent program algorithm to construct program synthesis, where the program consists of two types of data: parameters and functions. Because the LPA algorithm cannot generate completely accurate programs, this program uses a selection strategy to screen the top two best candidate programs as synthetic programs. The selection strategy is to train a two-way encoder based on a converter to obtain label-consistent programs as candidate programs.

4. The fake news detection method based on heuristic heterogeneous graph reasoning network according to claim 3 is characterized by S3 includes the following steps: S31: News sequence guides the matching connection between program subgraph and table subgraph: first, nodes with logical information in the news are retrieved, and then the information connected to the node is used as additional knowledge to expand the connection between program and table association subgraph; S32: Program information guides the matching connection between the news sequence subgraph and the table subgraph: The program subgraph is a tree structure composed of leaf nodes and parent nodes. Two types of nodes are designed as external knowledge to enhance the connection of the news-table association subgraph, namely, leaf-leaf association nodes and association nodes between parent nodes and leaf nodes; the leaf nodes in the leaf-leaf association nodes have the same parent node; a leaf node in the association nodes between parent nodes and leaf nodes is associated with its parent node.

5. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method of claim 1.

6. A computer-readable storage medium, characterized in that Computer-executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

7. A computer program product, characterized in that The invention comprises computer executable instructions, which are used to implement the method of claim 1 when the instructions are executed.

Images