A Mongolian rumor detection method based on subjective and objective interpretable bidirectional graph neural network

Abstract

A Mongolian rumor detection method based on a subjective and objective interpretable bidirectional graph neural network is proposed. The method uses an evidence retrieval module to retrieve evidence, including document retrieval, evidence retrieval and claim verification. Two heterogeneous graphs, a bidirectional tree graph and an evidence star graph, are constructed. The bidirectional tree graph uses a graph convolutional neural network with a top-down rumor propagation directed graph to learn the rumor propagation pattern. The rumor diffusion pattern is captured using a graph convolutional neural network with a bottom-up rumor diffusion graph. In the evidence star graph, the source post, i.e., the detection object, is located at the center, and the evidence nodes surround it, representing an angle in the star structure. A rumor detection model is constructed based on the bidirectional tree graph. After K iterations of information transmission based on the bidirectional tree structure and the star structure, the final representation of the two embedding results is obtained, which are connected and passed to a multi-layer perceptron for final prediction. The model can be used to detect Mongolian rumors and obtain classification labels and evidence explanations.

Description

Technical Field

[0001] The invention belongs to the technical field of artificial intelligence, and in particular relates to a Mongolian rumor detection method of a subjectively and objectively interpretable bidirectional graph neural network. Background Art

[0002] Rumor detection is known as an indispensable part of purifying the Internet ecosystem. Existing rumor detection methods include manual rumor detection methods, rumor detection methods based on machine learning, and rumor detection methods based on deep learning.

[0003] In recent years, with the rapid development of deep learning, various model structures have emerged. Compared with the deep propagation model of tweets, which is widely mentioned in rumor detection, the widely dispersed structure of rumors is often ignored. In addition, existing rumor detection strategies usually provide detection labels but ignore their explanations, but providing evidence to explain why suspicious articles are rumors is essential.

[0004] For Mongolian, rumor propagation includes not only the propagation pattern but also the structure of rumor diffusion. Secondly, rumor detection requires not only a label of whether it is a rumor, but also the evidence to support the label of whether it is a rumor. Therefore, the propagation pattern analysis and explainability in rumor detection will face more difficulties. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a Mongolian rumor detection method based on a subjective and objective interpretable bidirectional graph neural network, which detects rumors by adopting a dual-task approach of evidence retrieval subtask and rumor prediction subtask.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A Mongolian rumor detection method based on a subjective and objective interpretable bidirectional graph neural network includes the following steps:

[0008] Step 1: Retrieve evidence using the evidence retrieval module, which includes three steps: document retrieval, evidence retrieval, and claim verification;

[0009] Step 2, construct two heterogeneous graphs, a bidirectional tree graph and an evidence star graph. The bidirectional tree graph is a bottom-up and top-down bidirectional graph neural network, which uses a graph convolutional neural network with a top-down rumor propagation directed graph to learn the rumor propagation pattern; uses a graph convolutional neural network with a bottom-up rumor diffusion graph to capture the rumor diffusion pattern; in the evidence star graph, the source post, i.e., the detection object, is located at the center, and all evidence nodes surround it, representing an angle in the star structure;

[0010] Step 3, constructing a rumor detection model based on the bidirectional tree diagram;

[0011] Step 4: After K iterations of information transmission based on the bidirectional tree structure and the star structure, the final representations of the two embedding results are obtained, the final representations of the two embedding results are connected together, and then passed to the multi-layer perceptron for final prediction;

[0012] Step 5: Use the model to detect Mongolian rumors and obtain classification labels and evidence explanations.

[0013] In one embodiment, the document retrieval in step 1 is to crawl relevant evidence articles in encyclopedia data using a keyword matching algorithm; the evidence retrieval is to extract sentence-level evidence from the retrieved evidence articles; and the claim verification is based on sentence-level evidence to predict the relationship between the claim and the evidence as "support", "refute" or "irrelevant".

[0014] In one embodiment, step 1 utilizes a semantic NLP toolkit to extract potential entities from a given statement, and by analyzing the entities, the evidence articles are filtered using an API. Then, from the retrieved documents, ERM extracts objective facts in the form of sentences related to the statement as prediction evidence. Finally, a verification component in ERM makes predictions for the given statement and the retrieved evidence, and verifies whether the relationship between the statement and the evidence is "support", "refute", or "irrelevant".

[0015] In one embodiment, in the bidirectional tree diagram described in step 2, the conversation structure is in the form of a tree, the root node of the tree is the source post, each node represents a comment, and the nodes are connected through their reply relationships.

[0016] In one embodiment, the creation of the evidence star graph in step 2 is as follows:

[0017] G e =(V e ,G e )

[0018] V e =[c,e1,e 2, ,…e k ]

[0019] E e ={(c,e1),(c,e2),…(c,e k )}

[0020] Among them, G e is a vertex set V e and edge set E e Evidence graph object composed of vertex set V e Including detection object c and evidence sentences e1, e2, …, e k , e krepresents the kth evidence sentence, and the edge set E e Represents the relationship between each evidence sentence and the detection object.

[0021] In one embodiment, in step 3, in the rumor detection model, rumor propagation features are obtained through a top-down graph convolutional neural network, and rumor diffusion features are obtained through a bottom-up graph convolutional neural network; the bottom-up graph convolutional neural network recursively visits each node from the bottom leaf node to the root node, and generates a feature vector for each subtree, so that subtrees with similar contexts are projected to adjacent areas in the representation space; the top-down graph convolutional neural network generates an enhanced feature vector for each source post that takes into account its propagation path, and the representation of each node is calculated by combining its own input and its parent node, recursively from the root node to its child nodes until all leaf nodes are reached, where the root node represents the source post, the child nodes represent comments and reposts made to the parent node, and each node is connected by its reply relationship.

[0022] In one embodiment, step 4 is based on K iterations of information transmission between a bidirectional tree structure and a star structure, and each iteration updates the parameters of the two structures once; a maximum aggregator is used to aggregate and connect the representations of the rumor detection embedding results and the evidence embedding results, and pass them to a multi-layer perceptron for final prediction to obtain a rumor label and evidence explanation.

[0023] Compared with the prior art, the present invention explains the rumor detection strategy through subjective inference and objective facts, and provides evidence to explain rumors. By constructing two graph object structures: an evidence star graph and a bidirectional tree graph. The bidirectional tree graph also takes into account the two modes of rumor propagation, and extracts two features of rumors through bottom-up and top-down graph convolutional neural networks. The present invention can demonstrate better functions and advantages in Mongolian rumor detection, and the trained model significantly improves the accuracy of rumor detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is the overall framework diagram of the present invention.

[0025] Figure 2 It is a bidirectional tree graph structure.

[0026] Figure 3 It is the structure of the evidence star graph. DETAILED DESCRIPTION

[0027] The embodiments of the present invention are described in detail below with reference to the accompanying drawings and examples.

[0028] The present invention provides a Mongolian rumor detection method based on a subjective and objective interpretable bidirectional graph neural network, wherein the subjective represents subjective facts, source posts, forwarding and comments based on posts. The objective represents objective facts, which refers to relevant evidence articles in encyclopedia data such as Wikipedia. The bidirectional graph neural network refers to bottom-up and top-down graph neural networks.

[0029] like Figure 1 As shown, the present invention specifically includes

[0030] Step 1: Use the Evidence Retrieval Module (ERM) to retrieve evidence. The Evidence Retrieval Module is based on a three-step pipeline approach. Evidence retrieval consists of three steps: document retrieval, evidence retrieval, and claim verification. The Evidence Retrieval Module is a retrieval module officially provided by FEVER and is one of the most widely used evidence retrieval methods.

[0031] Specifically:

[0032] Document retrieval uses keyword matching algorithms to crawl relevant evidence articles from encyclopedia data such as Wikipedia.

[0033] Evidence retrieval is to extract sentence-level evidence from the retrieved evidence articles.

[0034] Claim verification is based on sentence-level evidence and predicts the relationship between the claim and the evidence as “support”, “refute” or “irrelevant”.

[0035] The evidence retrieval module first extracts potential entities from a given statement using a semantic NLP toolkit, and filters evidence articles such as Wikipedia by analyzing the entities using an API; then from the retrieved documents, the module extracts objective facts in the form of sentences related to the statement as predictive evidence; finally, a verification component in the module makes predictions for a given statement and the retrieved evidence, and verifies whether the relationship between the statement and the evidence is "support", "refute", or "irrelevant".

[0036] Step 2, reference Figure 2 ,Constructing heterogeneous graphs: bidirectional tree graphs.

[0037] The bidirectional tree graph is a special reply relationship topology that is naturally formed by social media and provides important clues for rumor detection. Specifically, the bidirectional tree graph is a bottom-up and top-down bidirectional graph neural network that uses a graph convolutional neural network with a top-down rumor propagation directed graph to learn rumor propagation patterns and uses a graph convolutional neural network with a bottom-up rumor diffusion graph to capture rumor diffusion patterns. In the bidirectional tree graph, the conversation structure is tree-shaped, with the root node of the tree being the source post (detection object), each node representing a comment, and the nodes connected by their reply relationships.

[0038] Step 3, reference Figure 3 ,Constructing heterogeneous graphs: Evidence star graphs.

[0039] Based on the retrieved evidence information, a deep BiLSTM is used to extract the information between words after word embedding and generate sentence representations to construct an evidence star graph. The evidence star structure indicates that each piece of evidence is a supplementary description of the source post, so each evidence sentence directly related to the source post forms a star topology. In the evidence star graph, the source post (detection object) is located in the center, and all evidence nodes surround it, representing an angle in the star structure.

[0040] In one embodiment of the present invention, the creation of the evidence star graph is as follows:

[0041] G e =(V e ,G e )

[0042] V e =[c,e1,e 2, ,…e k ]

[0043] E e ={(c,e1),(c,e2),…(c,e k )}

[0044] Among them, G e is a vertex set V e and edge set E e Evidence graph object composed of vertex set V e Including detection object c and evidence sentences e1, e2, …, e k , e k represents the kth evidence sentence, and the edge set E e Represents the relationship between each evidence sentence and the detection object.

[0045] Step 4: Construct a rumor detection model based on the bidirectional tree diagram.

[0046] The rumor detection model adopts a bottom-up and top-down bidirectional graph neural network structure to capture the two propagation characteristics of rumors. Specifically, the rumor propagation feature is obtained by a top-down graph convolutional neural network, and the rumor diffusion feature is obtained by a bottom-up graph convolutional neural network. The core idea of ​​the bottom-up graph convolutional neural network is to generate a feature vector for each subtree by recursively visiting each node from the bottom leaf node to the root node. In this way, subtrees with similar contexts, such as those with rejected parent nodes and child node support, will be projected to neighboring areas in the representation space. Therefore, this local rumor indication feature is aggregated along different branches into some global representations of the entire tree; the top-down graph convolutional neural network aims to capture the complex propagation pattern of rumors by using the top-down structure. The idea of ​​this top-down approach is to generate an enhanced feature vector for each post that takes into account its propagation path. The representation of each node is calculated by combining its own input and its parent node. This process is recursively performed from the root node to its child nodes until all leaf nodes are reached. The root node represents the source post (detection object), the child nodes represent the comments and reposts made to the parent node, and each node is connected through its reply relationship.

[0047] The present invention uses DropEdge to reduce the overfitting of the model, that is, at each training stage, it randomly removes edges from the input graph and generates different deformed copies at a certain rate. Therefore, this method increases the randomness and diversity of the input data, just like randomly rotating or flipping an image. Formally, suppose the total number of edges in graph A is N e , the drop rate is d, then the adjacency matrix A′ after DropEdge is calculated as follows:

[0048] A′=AA drop

[0049] Among them A drop is to use N randomly sampled from the original edge set e ×d edges, and A′ is the adjacency matrix after using DropEdge.

[0050] Step 5: After K iterations of information transmission based on the bidirectional tree structure and the star structure, the final representation of the conversation embedding result and the evidence embedding result is obtained as follows:

[0051]

[0052]

[0053] Where p, e are the reply and evidence of the i-th event.

[0054] Since the parameters of the two graph information structures are updated once each iteration, the maximum aggregator is used to aggregate the information into a fixed size, and then the two parts of information are connected together and then passed to the multi-layer perceptron for final prediction to obtain the rumor label and evidence explanation. The formula is as follows:

[0055]

[0056] where y r is the rumor prediction label, p,e is the reply and evidence of the ith event related to the source post, V and b y are the parameters in the output layer of the multilayer perceptron.

[0057] Step 6: Using this model, Mongolian rumor detection can be performed and classification labels and evidence explanations can be obtained.

[0058] For example, for the statement (Zhang San is a singer in China.)”, and its related replies and forwardings are fed into the model. First, the evidence retrieval module (ERM) extracts potential entities from the given statement. (Artist name: Zhang San)" (singer)", then analyze the entity and filter the evidence articles in the encyclopedia with the API interface. Then, from these retrieved documents, ERM extracts objective facts in the form of sentences related to the statement as predictive evidence. ERM gives a "support" label and uses the retrieved evidence as the star structure of the evidence star graph, and enters the rumor detection module together. You can get the label "not a rumor" for the statement "Zhang San is a Chinese singer." and the evidence articles in the encyclopedia that support this statement.

[0059] (Zhang San, born in 1979 in xx city, xx province, native of xx county, xx city, xx province, Chinese pop singer, musician, actor, director, screenwriter.)”

Claims

1. A Mongolian rumor detection method based on a subjective and objective interpretable bidirectional graph neural network, characterized in that: The steps include: Step 1: Retrieve evidence using the evidence retrieval module, which includes three steps: document retrieval, evidence retrieval, and claim verification; Step 2, construct two heterogeneous graphs, a bidirectional tree graph and an evidence star graph. The bidirectional tree graph is a bottom-up and top-down bidirectional graph neural network, which uses a graph convolutional neural network with a top-down rumor propagation directed graph to learn the rumor propagation pattern; uses a graph convolutional neural network with a bottom-up rumor diffusion graph to capture the rumor diffusion pattern; in the evidence star graph, the source post, i.e., the detection object, is located at the center, and all evidence nodes surround it, representing an angle in the star structure; Step 3, constructing a rumor detection model based on the bidirectional tree diagram; Step 4: After K iterations of information transmission based on the bidirectional tree structure and the star structure, the final representations of the two embedding results are obtained, the final representations of the two embedding results are connected together, and then passed to the multi-layer perceptron for final prediction; Step 5: Use the model to detect Mongolian rumors and obtain classification labels and evidence explanations.

2. The Mongolian rumor detection method based on the subjective and objective interpretable bidirectional graph neural network according to claim 1 is characterized in that: Step 1, the document retrieval is to crawl relevant evidence articles in encyclopedia data using a keyword matching algorithm; The evidence retrieval is to extract sentence-level evidence from the retrieved evidence articles; the claim verification is based on the sentence-level evidence to predict the relationship between the claim and the evidence as "support", "refutation" or "irrelevant".

3. The Mongolian rumor detection method based on the subjective and objective interpretable bidirectional graph neural network according to claim 1 is characterized in that: In step 1, a semantic NLP toolkit is used to extract potential entities from a given statement. By analyzing the entities, the API is used to filter evidence articles. Then, from the retrieved documents, ERM extracts objective facts in the form of sentences related to the statement as prediction evidence. Finally, a verification component in ERM makes predictions for the given statement and the retrieved evidence, and verifies whether the relationship between the statement and the evidence is "support", "refute", or "irrelevant".

4. The Mongolian rumor detection method based on the subjective and objective interpretable bidirectional graph neural network according to claim 1 is characterized in that: Step 2: In the bidirectional tree diagram, the conversation structure is in the form of a tree, the root node of the tree is the source post, each node represents a comment, and the nodes are connected through their reply relationships.

5. The Mongolian rumor detection method based on the subjective and objective interpretable bidirectional graph neural network according to claim 1 is characterized in that: Step 2, the creation of the evidence star graph is as follows: G e =(V e ,G e ) V e =[c,e1,e2,,...and k ] AND e ={(c,e1),(c,e2),...(c,e k )} Among them, G e is a vertex set V e and edge set E e Evidence graph object composed of vertex set V e Including detection object c and evidence sentences e1, e2, ..., e k , e k represents the kth evidence sentence, and the edge set E e Represents the relationship between each evidence sentence and the detection object.

6. The Mongolian rumor detection method based on the subjective and objective interpretable bidirectional graph neural network according to claim 1 is characterized in that: In the step 3, in the rumor detection model, rumor propagation features are obtained through a top-down graph convolutional neural network, and rumor diffusion features are obtained through a bottom-up graph convolutional neural network; the bottom-up graph convolutional neural network recursively visits each node from the bottom leaf node to the root node, and generates a feature vector for each subtree, so that subtrees with similar contexts are projected to adjacent areas in the representation space; the top-down graph convolutional neural network generates an enhanced feature vector for each source post that takes into account its propagation path, and the representation of each node is calculated by combining its own input and its parent node, recursively from the root node to its child nodes until all leaf nodes are reached, where the root node represents the source post, and the child nodes represent comments and reposts made to the parent node, and each node is connected by its reply relationship.

7. The Mongolian rumor detection method based on a subjective and objective interpretable bidirectional graph neural network according to claim 1 is characterized in that: The step 4 is based on K iterations of information transmission between the bidirectional tree structure and the star structure. The two structures update their parameters once each iteration. The maximum aggregator is used to aggregate and connect the representations of the rumor detection embedding results and the evidence embedding results, and pass them to the multi-layer perceptron for final prediction to obtain the rumor label and evidence explanation.

Images