False news detection method and device based on semantic-structure double-teacher migration

Abstract

The invention discloses a false news detection method and device based on semantic-structure double-teacher migration, and belongs to the technical field of false news detection. The method comprises the following steps: acquiring propagation data of news in a social network, and constructing a local propagation graph between the news and comments and a global propagation graph between the news based on the propagation data; respectively learning news content features and structural features on the local propagation graph and the global propagation graph to obtain a trained content teacher model and a trained propagation teacher model; performing multi-channel transfer learning on the student model according to the trained content teacher model and the trained propagation teacher model; and constructing a global propagation graph of the to-be-detected data set, and obtaining a false news detection result of each piece of news in the to-be-detected data set based on the trained student model. The performance of the false news detection task can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fake news detection, and particularly relates to a fake news detection method and device based on semantic-structure double-teacher migration. BACKGROUND

[0002] In the new media era, social media websites provide great convenience for users to obtain information, express opinions and communicate with each other. More and more users are keen to participate in the discussion of hot topics in social media, and due to the convenience of the platform, user-generated content can quickly reach a wide audience. However, a large number of rumors have also been generated and spread in social media, which has brought harm to society and seriously affected people's daily life and the healthy development of society. Therefore, how to timely and accurately detect rumors is an urgent research topic in the field of social network analysis, and has important research significance and practical application significance.

[0003] With the passage of time, news forms a specific propagation structure through a series of forwarding or replying propagation relationships. Existing researches have found that, compared with real information, rumors are propagated faster, wider and deeper. These structural characteristics provide the possibility for researchers to detect rumors using propagation data. With the development of deep learning and data mining technologies, many methods model the propagation data of source posts as tree structure or graph structure, learn the feature representation of source posts using deep learning tools, and construct classifiers for detection.

[0004] However, in the increasingly complex modern society, user-driven information propagation has a certain degree of randomness. Due to the writing and expression habits of network users, some network jargons and non-textual content usually appear as feature noise in the collected news data set. And the mistakes in the human collection process, user post deletion and the existence of social robots make the collected news propagation structure have noise or be incomplete. When the collected news data has both text features and propagation structure noise, the mutual interference between them will cause the performance of existing models to decline, because when learning the graph, semantic noise will propagate on the news propagation graph containing noise, and then exacerbate the error in feature learning, which is easy to cause serious error accumulation in the model learning process, and cannot capture more useful structural features, thereby limiting the performance of existing fake news detection methods. SUMMARY

[0005] The main technical problem addressed by the present application is how to alleviate the mutual interference between actual text noise and propagation noise. The present application provides a fake news detection method and device based on semantic-structure double-teacher migration to improve the performance of the fake news detection task.

[0006] To achieve the above object, the technical scheme of the present application is as follows:

[0007] A false news detection method based on semantic-structure double teacher migration, the method comprising:

[0008] Obtain the propagation data of news in a social network, and based on the propagation data, construct a local propagation graph between news and comments and a global propagation graph between news and news;

[0009] Learn news content features and structure features on the local propagation graph and the global propagation graph respectively to obtain a trained content teacher model and a trained propagation teacher model;

[0010] According to the trained content teacher model and the trained propagation teacher model, perform multi-channel transfer learning on a student model;

[0011] Construct a global propagation graph of a to-be-detected data set, and based on the trained student model, obtain a false news detection result of each news in the to-be-detected data set.

[0012] Further, the propagation data includes a set of text information and a set of propagation relationships in the news propagation process, users commenting on the news and the number of times the users comment on the news; wherein the set of text information includes news and comments on the news.

[0013] Further, the nodes in the local propagation graph include news nodes and comment nodes, the initial features of the news nodes and the comment nodes are text features of the news and the comments respectively, and the edges in the local propagation graph are generated based on the set of propagation relationships.

[0014] Further, learning news content features on the local propagation graph to obtain a trained content teacher model comprises:

[0015] Obtain the initial features of the news nodes in the local propagation graph;

[0016] Use a multi-layer perception machine to learn semantic features of the initial features of the news nodes to obtain a classification result

[0017] Based on the classification result Calculate the cross-entropy loss of the content teacher model and perform back propagation to obtain a trained content teacher model.

[0018] Further, the nodes in the global propagation graph include news nodes and user nodes, initial features of the news nodes and the user nodes are generated based on location encoding, edges in the global propagation graph are generated based on a set of comment relationships between users and news, and weights of the edges are obtained according to the number of times of commenting on the news by the users; wherein the set of comment relationships between the users and the news is generated based on the users commenting on the news.

[0019] Further, the global propagation graph is used to learn news content features to obtain a trained propagation teacher model, including:

[0020] An adjacency matrix A of the global propagation graph is obtained Global ;

[0021] A degree-based propagation structure correction matrix M is used to correct the adjacency matrix A Global to obtain a corrected adjacency matrix

[0022] Based on the corrected adjacency matrix and initial features of nodes in the global propagation graph, structure feature learning is performed by using a graph convolution network to obtain a classification result

[0023] According to the classification result , a cross-entropy loss of the propagation teacher model is calculated, and back propagation is performed to obtain the trained propagation teacher model.

[0024] Further, according to the trained content teacher model and the trained propagation teacher model, multi-channel transfer learning is performed on a student model, including:

[0025] An adjacency matrix A of the local propagation graph is obtained loc ;

[0026] Based on the adjacency matrix A loc and initial features of nodes in the local propagation graph, local features h Loc are calculated by using a graph convolution neural network.

[0027] The local features h Loc of all news are propagated on the global propagation graph to obtain hidden representations h s output by the student model and classification results

[0028] According to the classification results , a cross-entropy loss L CLS of the student model is calculated.

[0029] Hidden representations output by the trained content teacher model and classification results are obtained. and the hidden representation output by the trained propagation teacher model and the classification result

[0030] using the classification result the classification result of the student model performing knowledge distillation to obtain a distillation loss wherein tea∈{CT,PT};

[0031] by aligning the hidden representation h s with the hidden representation to obtain an alignment loss

[0032] based on the cross-entropy loss L CLS , the distillation loss and the alignment loss performing back propagation to obtain a trained student model.

[0033] Further, the distillation loss wherein D KL represents the Kullback-Leibler divergence between the distribution of the classification result and the classification result and p represents a temperature parameter.

[0034] Further, the alignment loss wherein |V news | represents the number of news, h s,i represents the hidden representation h s of the i-th news calculated by the student model, represents the hidden representation of the i-th news calculated by the trained content teacher model or the trained propagation teacher model

[0035] A false news detection device based on semantic-structure double teacher migration, the device comprises:

[0036] a graph construction module configured to obtain propagation data of news in a social network, and construct a local propagation graph between news and comments and a global propagation graph between news and news based on the propagation data;

[0037] a first training module configured to learn news content features and structural features on the local propagation graph and the global propagation graph respectively to obtain a trained content teacher model and a trained propagation teacher model;

[0038] The second training module is configured to perform multi-channel transfer learning on the student model according to the trained content teacher model and the trained propagation teacher model.

[0039] The detection classification module is configured to construct a global propagation graph of the to-be-detected data set, and obtain a fake news detection result of each news in the to-be-detected data set based on the trained student model.

[0040] A storage medium, wherein the storage medium stores a computer program, and the computer program is configured to execute the above method when running.

[0041] An electronic device, comprising a memory and a processor, wherein the memory stores a program for executing the above method.

[0042] In summary, the fake news detection method for solving mixed noise by using transfer learning provided by the application can be applied to the fake news detection problem of a social media platform, and has the following advantages:

[0043] 1) The double-teacher model is used to learn reliable news semantic knowledge and structural knowledge respectively, and the double-channel transfer learning is used to transfer the knowledge to the student model, so that the student model learns the news representation without interference of semantic noise and propagation noise;

[0044] 2) The training framework based on double-channel transfer learning can improve the transfer efficiency, alleviate the mutual interference of semantic noise and propagation noise, and effectively improve the accuracy of fake news detection. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The implementation flowchart of the fake news detection method for solving mixed noise by using transfer learning provided by the application is provided.

[0046] Figure 2 The implementation flowchart of the graph neural network method based on local propagation graph and global propagation graph provided by the application is provided.

[0047] Figure 3 The fake news detection method training framework based on double-channel transfer learning provided by the application is provided.

[0048] Figure 4 The structure diagram of the fake news detection system for mixed noise provided by the application is provided. DETAILED DESCRIPTION

[0049] The application will be described in detail below in combination with the drawings and embodiments, and it should be pointed out that the described embodiments are only intended to facilitate the understanding of the application and do not limit the application.

[0050] Figure 1The implementation flowchart of the method for detecting fake news by using transfer learning to solve mixed noise is provided, and specifically includes the following steps:

[0051] Step A: Obtain the propagation data of news in a social network, and based on the propagation data, construct a local propagation graph between news and comments and a global propagation graph between news and news.

[0052] Firstly, the propagation data of sample i is defined as Wherein, represents the text information set in the propagation process, r i is news, is the subsequent comment; represents the sample i propagation relationship set; U i represents the user participating in the news; S i ={s u ,u∈U i} indicates the number of times the user participates in the news.

[0053] Then, for each sample, the BERT text features of news r and comments x1,...,x n-1 are extracted, which are defined as Wherein, is the BERT text feature of news r, is the BERT text feature of comments x1,...,x n-1 .

[0054] Next, based on the text information set V and the propagation relationship set E, a local propagation graph is constructed, denoted as G loc . The adjacency matrix corresponding to the graph structure is defined as A loc , A Global , wherein A loc ∈R n×n corresponds to a sample, n is the number of texts in sample i, and A loc The initial value is defined as:

[0055]

[0056] The initial feature representation of all nodes in the local propagation graph is constructed based on the text feature, denoted as

[0057] Finally, based on the text information set V, the user participating in the news U and the number of times the user participates in the news S, a global propagation graph is constructed, denoted as A Global ∈R N×N , N is the number of news in the data set (i.e., the number of samples), and the element a Global in A ijdenotes the number of common users in sample i and sample j, where A Global The construction process of is as follows:

[0058] 1) Construct the user participation matrix C ∈ R N×|U| of the data set, where U is the union set of U i in all samples, that is, the set of all users in the data set, where the element c ij denotes the number of times user j participates in news i.

[0059] 2) Obtain the adjacency matrix A Global of the global propagation graph by A T Global .

[0060] Step B: Learning of news content features and structural features on the local propagation graph and the global propagation graph respectively to obtain a trained content teacher model and a trained propagation teacher model.

[0061] Step B1: Design a content teacher model for noisy news text features Use a multi-layer perceptron to learn semantic features:

[0062]

[0063]

[0064] where MLP(·) denotes a multi-layer perceptron; Relu(·) denotes a Relu activation function; and denote the hidden representation and classification result output by the content teacher model, respectively.

[0065] Step B2: Design a propagation teacher model for noisy propagation structure, and perform structural feature learning based on the global propagation graph. First, set initial features for nodes in the global propagation graph based on position encoding:

[0066] X PE = WX one-hot +b

[0067] where W is a learnable parameter matrix, b denotes a learnable bias vector, and X one-hot is a one-hot vector representing the position of news in the global propagation graph.

[0068] Then, design a degree-based propagation structure correction matrix M to correct the adjacency matrix A Global of the global propagation graph:

[0069]

[0070] ​Where I is the identity matrix, and the calculation process of M is as follows:

[0071]

[0072] Where d i ,d j Represents the global propagation graph A Global China News i and news r j The degree of the nodes.

[0073] like Figure 2 As shown, a two-layer graph convolutional network is used for structural feature learning:

[0074]

[0075]

[0076] Where GCN(·) represents a graph convolutional neural network, and These represent the implicit representation and classification result output by the propagation teacher model, respectively.

[0077] Step B3: Train the teacher model according to the following training objectives:

[0078]

[0079] Where L tea Let CE(·) represent the training loss of the content teacher, and let y represent the cross-entropy loss. news The labels represent the true nature of the news being judged, CT represents the content teacher model, and PT represents the dissemination teacher model.

[0080] Step C: Based on the trained content teacher model and the trained dissemination teacher model, perform multi-channel transfer learning on the student model.

[0081] Step C1: Adjacency matrix A based on the local propagation graph of sample i Loc and node features The local representation of the student model is obtained by computing the graph convolutional neural network in the student model:

[0082] h Loc =Relu(GCN(X,A) Loc ))

[0083] Then we obtain the local representations of N news items. The prediction is propagated across the global propagation graph, and finally, the prediction result is output.

[0084] h s =W1(A news HLoc )+b1

[0085]

[0086] where W1 and W2 are parameter matrices, and b1 and b2 are bias terms.

[0087] Step C2: Based on the hidden representation and classification results output by the student model, transfer learning is performed, first aligning the output results of the teacher model with the classification results of the student model Knowledge distillation is performed:

[0088]

[0089] where, represents the distillation loss of a certain teacher model on the student model, D KL (·) represents the Kullback-Leibler divergence between the distributions of and , and p is a temperature parameter that controls the smoothness of the output of the teacher model.

[0090] Then the hidden representation h s output by the student model is aligned with the hidden representation output by the teacher model:

[0091]

[0092] where, represents the alignment loss of the student model with a certain teacher model, V news represents the set of all news samples.

[0093] The final training objective of the student model is:

[0094]

[0095] L CLS is the cross-entropy loss between the prediction results of the student model and the true labels of the news, and l and b control the importance of L sup and L tar to the student model by the content teacher model and the propagation teacher model, respectively.

[0096] Step D: Construct a global propagation graph for the data set to be detected, and based on the trained student model, obtain the fake news detection results of each news in the data set to be detected.

[0097] Based on the final feature representation of the news by the trained model, the rumor detection is completed using a classifier. Based on the final feature representation C of the sample, the probability of the fake news category label of the sample is calculated, i.e.

[0098]

[0099] where W c ,b c are trainable parameters.

[0100] As Figure 3 shown, the application provides a multi-channel transfer learning fake news detection method training framework, specifically comprising the following steps:

[0101] Step a: obtain a labeled data set, label the authenticity class label of each sample, and divide the training set and the test set. The rumor class label includes true news (True News, T) and fake news (Fake News, F).

[0102] Step b: calculate the transfer learning loss of the training set, specifically:

[0103] The multi-channel transfer learning loss is calculated based on the hidden representation and the classification result output by the student model, and the specific calculation formula is as follows:

[0104]

[0105]

[0106] where D KL (·) represents the KL divergence of two distributions, which is a quantity used to measure similarity.

[0107] Step c: calculate the supervised classification cross-entropy loss of the training set, specifically, the supervised learning loss based on the cross-entropy loss can be defined as:

[0108]

[0109] where y i represents the vector representation distribution of the true label of sample i.

[0110] Step d: adjust the model parameters by minimizing the loss functions in steps b and c to train the optimal classification model, defined as follows:

[0111]

[0112] where λ and β are the balance coefficients of the importance of the content teacher model and the propagation teacher model to the student model in L sup and L tar .

[0113] Step e: use the optimal student model to classify the news in the test set to obtain the classification result.

[0114] As Figure 4 shown, the application also provides a false news detection device based on semantic-structure dual teacher migration, characterized in that the device comprises:

[0115] a graph construction module, configured to acquire propagation data of news in a social network, and based on the propagation data, construct a local propagation graph between news and comments and a global propagation graph between news and news;

[0116] a first training module, configured to learn news content features and structural features on the local propagation graph and the global propagation graph respectively, to obtain a trained content teacher model and a trained propagation teacher model;

[0117] a second training module, configured to perform multi-channel migration learning on a student model according to the trained content teacher model and the trained propagation teacher model;

[0118] a detection and classification module, configured to construct a global propagation graph of a to-be-detected data set, and based on the trained student model, acquire a false news detection result of each news in the to-be-detected data set.

[0119] The above is only the preferred embodiment of the application, and is not used to limit the application, and those skilled in the art should understand that any modification, equivalent replacement or improvement within the spirit and principle of the application should be included in the protection scope of the application, and the protection scope is defined by the claims.

Claims

1. A method for fake news detection based on semantic-structure dual teacher transfer, characterized in that, The method comprises: obtaining propagation data of news in a social network, and constructing a local propagation graph between news and comments and a global propagation graph between news based on the propagation data; learning news content features and structural features on the local propagation graph and the global propagation graph respectively to obtain a trained content teacher model and a trained propagation teacher model; performing multi-channel transfer learning on a student model according to the trained content teacher model and the trained propagation teacher model; constructing a global propagation graph of a to-be-detected data set, and obtaining a fake news detection result of each news in the to-be-detected data set based on the trained student model.

2. The method of claim 1, wherein, The propagation data comprises a text information set and a propagation relationship set in a news propagation process, and a user commenting on the news and the number of times of commenting on the news by the user; wherein the text information set comprises news and comments on the news.

3. The method of claim 2, wherein, The nodes in the local propagation graph comprise news nodes and comment nodes, and initial features of the news nodes and the comment nodes are text features of the news and the comments, respectively; edges in the local propagation graph are generated based on the propagation relationship set.

4. The method of claim 3, wherein, The learning of news content features on the local propagation graph to obtain the trained content teacher model comprises: obtaining initial features of news nodes in the local propagation graph; The initial features of the news nodes are subjected to semantic feature learning by using a multilayer perceptron to obtain a classification result based on the classification result The cross-entropy loss of the content teacher model is calculated, and back propagation is performed to obtain a trained content teacher model.

5. The method of claim 2, wherein, The nodes in the global propagation graph comprise news nodes and user nodes, and initial features of the news nodes and the user nodes are generated based on position encoding; edges in the global propagation graph are generated based on a user-news comment relationship set, and weights of the edges are obtained according to the number of times of commenting on the news by the user; wherein the user-news comment relationship set is generated based on the user commenting on the news.

6. The method of claim 1, wherein, The learning of news content features on the global propagation graph to obtain the trained propagation teacher model comprises: obtaining an adjacency matrix A of the global propagation graph Global ; The degree-based propagation structure correction matrix M corrects the adjacency matrix A Global The correction is performed to obtain a corrected adjacency matrix based on the corrected adjacency matrix and the initial features of the nodes in the global propagation graph, using a graph convolution network to learn structural features, to obtain a classification result According to the classification result The cross-entropy loss of the propagation teacher model is calculated and back-propagated to obtain a trained propagation teacher model.

7. The method of claim 1, wherein, performing multi-channel transfer learning on the student model according to the trained content teacher model and the trained propagation teacher model comprises: obtaining an adjacency matrix A of the local propagation graph loc ; based on the adjacency matrix A loc and the initial features of the nodes in the local propagation graph, a local feature h is calculated by a graph convolutional neural network Loc ; h = h (x, y, z, t) Loc h = h (x, y, z, t) s and classification results According to the classification result Computing a cross-entropy loss L of the student model CLS ; obtaining a hidden representation output by the trained content teacher model and a classification result obtaining a hidden representation output by the trained propagation teacher model and a classification result Utilizing the classification result Classification result of the student model Knowledge distillation is performed to obtain a distillation loss Wherein, tea∈{CT,PT}; By aligning the latent representation h s with the latent representation halign, we obtain an alignment loss based on the cross-entropy loss L CLS , the distillation loss and the alignment loss are back-propagated to obtain a trained student model.

8. The method of claim 7, wherein, the distillation loss where D KL denotes the Kullback-Leibler divergence between the distribution of the computed classification result and the classification result and p denotes the temperature parameter.

9. The method of claim 7, wherein, The alignment loss where |V news | represents the number of news, h s,i represents the hidden representation of the i-th news calculated by the student model s , represents the hidden representation of the i-th news calculated by the trained content teacher model or the trained propagation teacher model 10. A false news detection device based on semantic-structural dual teacher transfer, characterized in that, The device comprises: a graph construction module configured to obtain propagation data of news in a social network, and construct a local propagation graph between news and comments and a global propagation graph between news based on the propagation data; a first training module configured to learn news content features and structural features on the local propagation graph and the global propagation graph respectively to obtain a trained content teacher model and a trained propagation teacher model; a second training module configured to perform multi-channel transfer learning on a student model according to the trained content teacher model and the trained propagation teacher model; a detection and classification module configured to construct a global propagation graph of a to-be-detected data set, and obtain a fake news detection result of each news in the to-be-detected data set based on the trained student model.

Images