A multi-task rumor detection method based on bidirectional propagation graph

Abstract

The present invention discloses a multi-task rumor detection method based on a bidirectional propagation graph, which can be used for rumor detection of social network posts and position detection of comment information. The present invention first generates a text feature matrix, a user feature matrix and a text statistical feature matrix according to the content of the rumor post, then constructs a bidirectional propagation graph of the rumor, extracts the propagation features of the rumor by calculating the bidirectional graph convolution and performing root node feature enhancement, and finally trains a softmax classifier after average pooling and feature integration of the propagation features to obtain rumor detection and position detection results. The present invention can effectively obtain the sequential propagation features and breadth spread features of the rumor, and introduces the position detection of user comments as an auxiliary task, thereby improving the generalization of the model and further improving the accuracy of the rumor detection task.

Description

Technical Field

[0001] The present invention relates to a multi-task rumor detection method based on a bidirectional propagation graph. The method can be used for rumor detection of social network posts and position detection of comment information, and belongs to the technical field of Internet and natural language processing. Background Art

[0002] In recent years, social networks have developed rapidly and have quickly become one of the important ways for people to obtain news and information. Due to the huge amount of information on social networks and the speed of transmission far exceeding traditional media, a large number of unconfirmed rumors have been able to spread in cyberspace, which has become an increasingly serious problem. The Internet has become the "fifth space" after land, sea, sky, and space, and the use of social network platforms to spread rumors, take the opportunity to commit cyber crimes, spread violent terrorist information, incite color revolutions and other illegal and criminal acts have posed a serious threat to the security of cyberspace.

[0003] Existing rumor detection methods for social networks mainly include manual rumor detection methods, machine learning-based rumor detection methods, and deep learning-based rumor detection methods. Most mainstream social network platforms still use manual rumor detection methods as the mainstream mechanism for identifying suspected rumor information. Although they have high discrimination accuracy, they are lagging and are greatly affected by human factors, so they cannot meet the needs of social network rumor detection. Machine learning-based methods reduce rumor detection work to a binary classification problem and use supervised learning methods to automatically detect rumor information. They have high requirements for feature engineering and cannot obtain the deep features of rumors during the propagation and dissemination process, so they cannot obtain a high recognition rate. Many rumor detection methods based on deep learning start from the temporal propagation structure of rumors. They detect suspected rumor information by extracting deep features of rumors during the propagation process. They often only focus on the sequential propagation features of rumors and ignore the breadth dissemination features. Therefore, there is still a certain loss of accuracy in the rumor detection task.

[0004] In view of the serious threat posed by social network rumors to the security of cyberspace and the problem that existing rumor detection methods cannot effectively extract the characteristics of rumor propagation, the present invention proposes a multi-task rumor detection method based on a bidirectional propagation graph. An improved bidirectional graph convolutional neural network (Bi-GCN) is used to extract the sequential propagation characteristics and breadth spread characteristics of rumors, and the stance detection of comment texts is introduced as an auxiliary task to improve the performance and generalization of the rumor detection task, thereby improving the accuracy of rumor detection. Summary of the invention

[0005] In view of the problems and shortcomings of the prior art, the present invention provides a multi-task rumor detection method based on a bidirectional propagation graph. The method can detect the stance of suspected rumor information and comment text in real time. Compared with the existing methods, the method can more effectively obtain the sequential propagation characteristics and breadth distribution characteristics of rumors, thereby improving the problem of insufficient accuracy of existing rumor detection methods.

[0006] In order to achieve the above purpose, the technical solution of the present invention is as follows: a multi-task rumor detection method based on a two-way propagation graph, which covers the entire process of social network rumor detection, mainly including feature extraction, feature fusion, comment stance type classification and rumor type classification, etc., and can effectively extract the propagation characteristics of rumors, thereby improving the accuracy of rumor detection tasks. The method mainly includes three steps, as follows:

[0007] Step 1: Construct rumor and stance sample datasets. First, use the user uids in the Twitter15 and Twitter16 datasets to obtain user information through the Twitter open API for rumor detection task training. At the same time, the PHEME dataset is used for user comment stance detection task training. The two together constitute the training dataset for the multi-task model.

[0008] Step 2: Multi-task classification model training. For each rumor post in the dataset, we first use the TF-IDF algorithm to extract its text features, generate user features and text statistical features, and then build a two-way propagation graph of the rumor. Then, we use the text features, user features, and text statistical features as inputs, extract the rumor propagation features by calculating the bidirectional graph convolution and performing root node feature enhancement, and finally train the softmax classifier after average pooling and feature integration of the propagation features.

[0009] Step 3: Classify and predict the posts to be detected. For the posts to be detected, first extract their text features, user features, and text statistical features, then build a bidirectional propagation graph of the post, extract propagation features by calculating bidirectional graph convolution and performing root node feature enhancement, and use the trained softmax classifier to obtain the results of rumor detection and stance detection.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0011] 1. This method uses an improved Bi-GCN network to obtain the propagation characteristics of rumors, which can effectively obtain the sequential propagation characteristics and breadth dispersion characteristics of rumors, thereby improving the accuracy of rumor detection tasks. Compared with the detection accuracy of 84%-86% of the Bi-GCN model before the improvement, this improved model has improved it by 1-1.5 percentage points.

[0012] 2. This method introduces stance detection of user comments as an auxiliary task, which can effectively improve the generalization of the model and further improve the accuracy of the rumor detection task, reaching the level of 86%-88.5%. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0014] Figure 2 It is a general framework diagram of the method of an embodiment of the present invention;

[0015] Figure 3 A bidirectional graph convolution calculation flow chart involved in an embodiment of the present invention;

[0016] Figure 4 This is a framework diagram of an improved bidirectional graph convolutional network involved in an embodiment of the present invention;

[0017] Figure 5 A schematic diagram of the attention mechanism in TD-GCN involved in an embodiment of the present invention;

[0018] Figure 6 Schematic diagram of the attention mechanism in BU-GCN involved in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to deepen the recognition and understanding of the present invention, the present invention is further illustrated below in conjunction with specific embodiments.

[0020] Embodiment: The specific process and overall framework of the present invention are as follows: Figure 1 and Figure 2 As shown, a multi-task rumor detection method based on a bidirectional propagation graph, the specific implementation steps of the present invention are as follows:

[0021] Step 1, construct rumor and stance sample datasets. The present invention first obtains user uids from Twitter15 and Twitter16 datasets, and obtains feature information of 430,000 users through Twitter open API for training rumor detection tasks. At the same time, the PHEME dataset is used for training user comment stance detection tasks. The two together constitute the training dataset of the multi-task model.

[0022] Step 2: Multi-task classification model training. For each rumor post in the dataset, we first use the TF-IDF algorithm to extract its text feature X s , and generate user features X m＝1 and text statistical features X m＝2 , then construct the two-way propagation graph of rumors, and then transform the text feature X s 、User Characteristics X m＝1 and text statistical features Xm＝2 As input, the rumor propagation feature S is extracted by calculating the bidirectional graph convolution and performing root node feature enhancement. Finally, the propagation feature S is averaged and integrated before training the softmax classifier. The implementation process of this step is divided into 7 sub-steps:

[0023] Sub-step 2-1, use the TF-IDF algorithm to generate the text feature matrix X s For each rumor post, use the TF-IDF algorithm to generate a text feature vector for each tweet it contains. And concatenate them one by one to generate the text feature matrix X s The calculation formula of TF-IDF is as follows:

[0024]

[0025]

[0026] TF-IDF(i,w j )=TF(i,w j )·IDF(w j ) (3)

[0027] Among them, TF(i,w j ) represents word w in tweet i j The frequency of word w in tweet i j The number of occurrences is given by Count(i,w j ) indicates that Count(i) indicates the total number of words in tweet i; IDF(w j ) represents word w j The inverse frequency of |T| represents the total number of tweets, CountTweets(w j ) means it contains word w j Total number of tweets.

[0028] Sub-step 2-2, generate user feature matrix X m＝1 For each rumor post, extract the corresponding user's feature information (whether there is a profile, gender, location, etc.) from each tweet it contains, generate a vector representation of the user's features, and concatenate them one by one to generate the user feature matrix X m＝1 The specific content of user characteristic information is shown in Table 1:

[0029] Table 1 User characteristics

[0030]

[0031] Sub-step 2-3, generate text statistical feature matrix X m＝2For each rumor post, extract the statistical feature information of the corresponding tweet (the number of negative words, the number of curse words, the number of periods, etc.) from each tweet it contains, generate a vector representation of the text statistical features, and concatenate them one by one to generate the text statistical feature matrix X m＝2 The specific content of the text statistical feature information is shown in Table 2. The negative words set in the present invention include not, no, nobody, nothing, none, never, neither, nor, nowhere, hardly, scarcely, barely, don't, isn't, wasn't, shouldn't, wouldn't, couldn't, doesn't; the swear words refer to the English swear word dictionary in the open source NLP tool textfilter.

[0032] Table 2 Text statistical features

[0033]

[0034] Sub-steps 2-4, construct a bidirectional propagation graph. For a rumor post, construct an adjacency matrix A based on the forwarding / reply relationships between all the tweets it contains. If A ts =1, it means there is a directed edge from node t to node s. Then use the DropEdge strategy to randomly extract some edges from A:

[0035] A′=AA drop (4)

[0036] Among them A drop is the adjacency matrix corresponding to the edge set randomly selected from A with probability p. Therefore, for the top-down propagation graph, its adjacency matrix is ​​A TD =A′, and for the bottom-up propagation graph, its adjacency matrix is ​​A BU =A′ T .

[0037] Sub-steps 2-5, extract rumor propagation features. Figure 3 As shown, the text feature matrix X generated by sub-steps 2-1, 2-2 and 2-3 is s , user feature matrix X m＝1 And text statistical feature matrix X m＝2 Input such as Figure 4 In the improved Bi-GCN network shown in Figure 1, the propagation characteristics of rumors are extracted by calculating bidirectional graph convolution. Taking the top-down graph convolutional neural network (TD-GCN) as an example, the hidden matrix corresponding to the first layer of TD-GCN in the parameter sharing layer is Defined as:

[0038]

[0039] in is a trainable parameter matrix.

[0040] The hidden feature matrix corresponding to the first layer of TD-GCN in the specific task layer Defined as:

[0041]

[0042] in, is the normalized adjacency matrix, defined as:

[0043]

[0044] From the adjacency matrix and the diagonal matrix I N It is calculated and defined as:

[0045]

[0046] is the degree matrix corresponding to the adjacency matrix, defined as:

[0047]

[0048] σ(·) is the nonlinear ReLU activation function.

[0049] Then use the same method to calculate the hidden feature matrix of the second layer TD-GCN of the specific task layer

[0050]

[0051] Similarly, the same method can be used to calculate the hidden feature matrix of the bottom-up graph convolutional neural network (BU-GCN) and

[0052] Sub-steps 2-6, root node feature enhancement. For TD-GCN, calculate the hidden feature matrix The feature vector for node i Perform linear transformation Convert it to a high-dimensional representation, where U is a trainable parameter matrix.

[0053] Next, use Figure 5 The attention mechanism shown calculates the AttentionScore of the root node root for node i:

[0054]

[0055] in, is the feature vector corresponding to the node root, is a trainable parameter vector.

[0056] Then calculate the root node feature enhancement vector corresponding to node i, which is defined as:

[0057]

[0058] Where σ(·) is the nonlinear sigmoid activation function.

[0059] Therefore, the feature matrix generated by TD-GCN after root node enhancement is:

[0060]

[0061] Combining formula (5) and formula (6), we get the hidden matrix

[0062]

[0063] By transforming the characteristic matrix in formula (10) and Replace with and Get the hidden feature matrix The calculation formula is updated to:

[0064]

[0065] Combining formula (13) to get the hidden matrix

[0066]

[0067] For the bottom-up graph convolutional network BU-GCN, calculate the hidden feature matrix Referring to formula (11) in the process of calculating the enhanced features of the root node of TD-GCN, the Attention Score of node i to the root node root in BU-GCN is calculated:

[0068]

[0069] Feature enhancement vector of the root node root definition:

[0070]

[0071] Therefore, the feature matrix generated by BU-GCN after root node enhancement is:

[0072]

[0073] in, for , referring to formula (14), formula (15) and formula (16), the hidden matrix of BU-GCN is and

[0074]

[0075]

[0076]

[0077] Sub-steps 2-7: feature fusion and classification result output. For the rumor detection task, the average pooling operation is used to integrate the sequential propagation feature S of TD-GCN. TD And the breadth spread feature S extracted by BU-GCN BU :

[0078]

[0079]

[0080] After the concatenation operation, the integrated feature S is obtained, which is defined as:

[0081] S = concat(S TD ,S BU ) (25)

[0082] After the integrated feature S passes through a fully connected layer and a softmax classifier in sequence, the rumor detection result is obtained Defined as:

[0083]

[0084]

[0085] in, is the normalized probability, is the probability that the rumor event belongs to label i. The final result of the rumor detection task is the label with the highest probability Defined as:

[0086]

[0087] For the stance detection task, the hidden layer features extracted by TD-GCN and BU-GCN are concatenated to obtain an integrated feature matrix Z:

[0088]

[0089] Defining vectors is the feature vector corresponding to the feature matrix Z of the i-th tweet in the event. After sending it to a fully connected layer and a softmax classifier in turn, the stance detection result of the i-th tweet can be obtained. Defined as:

[0090]

[0091]

[0092] in, is the normalized probability, is the probability that the tweet stance belongs to label j. The final result of the stance detection task is the label with the highest probability Defined as:

[0093]

[0094] Step 3: Classify and predict the posts to be tested. For the posts to be tested, first use the TF-IDF algorithm to extract their text features X s , and generate user features X m＝1 and text statistical features X m＝2 , then construct the two-way propagation graph of rumors, and then transform the text feature X s 、User Characteristics X m＝1 and text statistical features X m＝2 As input, the rumor propagation features are extracted by calculating bidirectional graph convolution and performing root node feature enhancement. Finally, after average pooling and feature integration, the propagation features are input into the softmax classifier trained in step 2 to obtain the rumor detection and stance detection results.

[0095] In summary, the present invention first generates a text feature matrix, a user feature matrix and a text statistical feature matrix according to the content of the rumor post, then constructs a two-way propagation graph of the rumor, extracts the propagation features of the rumor by calculating the two-way graph convolution and performing root node feature enhancement, and finally trains the softmax classifier after averaging the propagation features and integrating the features to obtain the rumor detection and stance detection results. The present invention can effectively obtain the sequential propagation and breadth distribution features of the rumor, and introduces the stance detection of user comments as an auxiliary task, thereby improving the generalization of the model and further improving the accuracy of the rumor detection task.

[0096] Based on the same inventive concept, an embodiment of the present invention discloses a multi-task rumor detection method and device based on a two-way propagation graph, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the multi-task rumor detection method and device based on a two-way propagation graph are implemented.

[0097] It should be noted that the above embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent modifications of the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

Claims

1. A multi-task rumor detection method based on bidirectional propagation graph, characterized in that: The method comprises the following steps: Step 1: Construct a dataset of rumor and position samples. Step 2: Multi-task classification model training, Step 3: Classify and predict the posts to be tested; Among them, step 1, constructing rumor and stance sample datasets, specifically as follows: first, using user uids in the Twitter15 and Twitter16 datasets, and obtaining user information through the Twitter open API for rumor detection task training, at the same time, the PHEME dataset is used for user comment stance detection task training, and the two together constitute the training dataset of the multi-task model; Step 2: Multi-task classification model training. The details are as follows: for each rumor post in the dataset, first use the TF-IDF algorithm to extract its text feature X s , and generate user features X m＝1 and text statistical features X m＝2 , then construct the two-way propagation graph of rumors, and then transform the text feature X s 、User Features X m＝1 and text statistical features X m＝2 As input, the rumor propagation feature S is extracted by calculating the bidirectional graph convolution and performing root node feature enhancement. Finally, the propagation feature S is averaged and integrated to train the softmax classifier. Step 2: Multi-task classification model training. The implementation process of this step is divided into 7 sub-steps: Sub-step 2-1, use the TF-IDF algorithm to generate the text feature matrix X s For each rumor post, use the TF-IDF algorithm to generate a text feature vector for each tweet it contains. And concatenate them one by one to generate the text feature matrix X s , the calculation formula of TF-IDF is as follows: TF-IDF(i,w j )=TF(i,w j )·IDF(w j ) (3) Among them, TF(i,w j ) represents word w in tweet i j The frequency of word w in tweet i j The number of occurrences is given by Count(i,w j ) indicates that Count(i) indicates the total number of words in tweet i; IDF(w j ) represents word w j The inverse frequency of |T| represents the total number of tweets, CountTweets(w j ) means it contains word w j Total number of tweets; Sub-step 2-2, generate user feature matrix X m＝1 For each rumor post, extract the corresponding user's feature information from each tweet it contains, generate a vector representation of user features, and concatenate them one by one to generate the user feature matrix X m＝1 ,User characteristic information includes basic characteristics and social influence characteristics. Basic characteristics include whether ,they have a profile, gender, location and registration time. Social influence characteristics include the ,average number of posts per day, the number of fans, the number of ,followers, and whether they are certified users. Sub-step 2-3, generate text statistical feature matrix X m＝2 For each rumor post, extract the statistical feature information of the corresponding tweet from each tweet it contains, generate a vector representation of the text statistical features, and concatenate them one by one to generate the text statistical feature matrix X m＝2 ,The specific contents of text statistical feature information include ,tweet dictionary, punctuation, content length and social influence. ,The tweet dictionary includes the number of negative words and curse words, punctuation includes the number of periods, ,exclamation marks and question marks, content length includes the proportion of capital letters, the number of words and ,characters, and social influence includes the number of likes, forwarding, comments and sharing. Sub-step 2-4, construct a two-way propagation graph. For a rumor post, construct an adjacency matrix A based on the forwarding / reply relationship between all the tweets it contains. If A ts =1, it means there is a directed edge from node t to node s, and then the DropEdge strategy is used to randomly extract some edges from A: A′=A-A drop (4) Among them A drop is the adjacency matrix corresponding to the edge set randomly selected from A with probability p. Therefore, for the top-down propagation graph, its adjacency matrix is ​​A TD =A′, and for the bottom-up propagation graph, its adjacency matrix is ​​A BU =A ′T , Sub-step 2-5, extract rumor propagation features, and convert the text feature matrix X generated by sub-steps 2-1, 2-2 and 2-3 into s , user feature matrix X m＝1 And the text statistical feature matrix X m＝2 Input into the improved Bi-GCN network, by calculating the bidirectional graph convolution, the propagation characteristics of rumors are extracted. For the top-down graph convolution neural network TD-GCN, the hidden matrix corresponding to the first layer of TD-GCN in the parameter sharing layer is Defined as: in is a trainable parameter matrix; The hidden feature matrix corresponding to the first layer of TD-GCN in the specific task layer Defined as: in, is the normalized adjacency matrix, defined as: From the adjacency matrix and the diagonal matrix I N It is calculated and defined as: is the degree matrix corresponding to the adjacency matrix, defined as: σ(·) is the nonlinear ReLU activation function; Then use the same method to calculate the hidden feature matrix of the second layer TD-GCN of the specific task layer The same method is used to calculate the hidden feature matrix of the bottom-up graph convolutional neural network BU-GCN and Sub-steps 2-6, root node feature enhancement, for TD-GCN, calculate the hidden feature matrix The feature vector for node i Perform linear transformation Convert it to a high-dimensional representation, where U is a trainable parameter matrix; Next, use the attention mechanism to calculate the Attention Score of the root node root for node i: in, is the feature vector corresponding to the node root, is a trainable parameter vector, Then calculate the root node feature enhancement vector corresponding to node i, which is defined as: Where σ(·) is the nonlinear sigmoid activation function, Therefore, the feature matrix generated by TD-GCN after root node enhancement is: Combining formula (5) and formula (6), we get the hidden matrix By transforming the characteristic matrix in formula (10) and Replace with and Get the hidden feature matrix The calculation formula is updated to: Combining formula (13) to get the hidden matrix For the bottom-up graph convolutional network BU-GCN, calculate the hidden feature matrix Referring to formula (11) in the process of calculating the enhanced features of the root node of TD-GCN, the Attention Score of node i to the root node root in BU-GCN is calculated: Feature enhancement vector of the root node root definition: Therefore, the feature matrix generated by BU-GCN after root node enhancement is: in, for , referring to formula (14), formula (15) and formula (16), the hidden matrix of BU-GCN is and Sub-steps 2-7, feature fusion and classification result output. For the rumor detection task, the average pooling operation is used to integrate the sequential propagation feature S of TD-GCN. TD And the breadth spread feature S extracted by BU-GCN BU : After the concatenation operation, the integrated feature S is obtained, which is defined as: S=concat(S TD ,S BU ) (25) After the integrated feature S passes through a fully connected layer and a softmax classifier in sequence, the rumor detection result is obtained Defined as: in, is the normalized probability, is the probability that the rumor event belongs to label i. The final result of the rumor detection task is the label with the highest probability. Defined as: For the stance detection task, the hidden layer features extracted by TD-GCN and BU-GCN are concatenated to obtain an integrated feature matrix Z: Defining vectors is the feature vector corresponding to the feature matrix Z of the i-th tweet in the event. After sending it to a fully connected layer and a softmax classifier in turn, the stance detection result of the i-th tweet can be obtained. Defined as: in, is the normalized probability, is the probability that the tweet stance belongs to label j. The final result of the stance detection task is the label with the highest probability. Defined as:

2. The multi-task rumor detection method based on bidirectional propagation graph according to claim 1 is characterized in that: Step 3: Classify and predict the posts to be tested. For the posts to be tested, first use the TF-IDF algorithm to extract their text features X s , and generate user features X m＝1 and text statistical features X m＝2 , then construct the two-way propagation graph of rumors, and then transform the text feature X s 、User Characteristics X m＝1 and text statistical features X m＝2 As input, the rumor propagation features are extracted by calculating bidirectional graph convolution and performing root node feature enhancement. Finally, after average pooling and feature integration, the propagation features are input into the softmax classifier trained in step 2 to obtain the rumor detection and stance detection results.

Images