CN114357160A - Early rumor detection method and device based on generation propagation structure characteristics - Google Patents

Abstract

The present invention provides an early rumor detection method and device based on generating propagation structure features. The method includes: acquiring a first source text of an event to be detected; inputting the first source text into a conditional variational autoencoder to obtain a conditional variation The first propagation structure feature of the event to be detected generated by the autoencoder; the conditional variational autoencoder is trained and obtained according to the second source text and the second propagation structure feature of the sample event; the second propagation structure feature is based on the graph convolutional neural network The model performs feature extraction on the propagation structure diagram of the sample event; the propagation structure diagram is constructed and obtained according to the second source text of the sample event and the response tweet corresponding to the second source text; the first source text and the first propagation structure feature are input into the classification In the model, the rumor detection result of the event to be detected output by the classification model is obtained. The invention realizes that after acquiring the source text of the event to be detected, the rumor detection result of the event to be detected can be quickly and accurately acquired.

Description

Method and device for early rumor detection based on generated propagation structure features

technical field

The present invention relates to the technical field of data processing, and in particular, to a method and device for detecting early rumors based on generating propagation structure features.

Background technique

With the rise and development of the Internet era, the speed of information flow and dissemination is getting faster and faster. People can freely publish information on social media, which also leads to the rapid spread of rumors. Rumors refer to things that go against the truth or fabricate out of thin air, are deceptive to a certain extent, and are speeches that disrupt social order and even bring serious panic and economic losses to the society. Therefore, how to distinguish rumors from the huge amount of information on social media is a serious challenge.

Rumor detection technology is the process of modeling and identifying rumors through relevant computer knowledge and skills. The existing rumor detection technology has the following problems: (1) The existing rumor detection research is limited to the detection of rumors after a certain period of time, that is, rumors may only be identified after a large number of exposures, and the early rumor detection is lacking It lacks certain timeliness; (2) Most of the existing researches use text information to detect rumors in early rumor detection, but due to the diversification of information structures, it is difficult to obtain accurate rumor detection results only based on text information.

SUMMARY OF THE INVENTION

The present invention provides an early rumor detection method and device based on generating a propagation structure feature, which is used to solve the defects of poor timeliness and low accuracy of the rumor detection method in the prior art, and realizes that when the source text of the rumor is obtained, it can be timely Accurately detect rumors.

The present invention provides an early rumor detection method based on generating a propagation structure feature, comprising:

Obtain the first source text of the event to be detected;

Inputting the first source text into a conditional variational self-encoder, and obtaining the first propagation structure feature of the to-be-detected event generated by the conditional variational self-encoder; wherein, the conditional variational self-encoder is based on The second source text of the sample event and the second propagation structure feature are obtained by training; the second propagation structure feature is obtained by feature extraction on the propagation structure graph of the sample event based on the graph convolutional neural network model; the propagation structure graph is obtained according to the The second source text of the sample event and the response tweet corresponding to the second source text are constructed and acquired;

The first source text and the first propagation structure feature are input into the classification model, and the rumor detection result of the to-be-detected event output by the classification model is obtained; the classification model is based on the second source text information and the second Propagation structural features, and rumor detection labels of the sample events are obtained by training.

According to an early rumor detection method based on generating a propagation structure feature provided by the present invention, the first source text and the first propagation structure feature are input into a classification model, and the event to be detected outputted by the classification model is obtained. rumour detection results, including:

Inputting the first source text into the feature extraction layer of the classification model to obtain text features of the first source text;

After splicing the text feature and the first propagation structure feature, input the classification layer of the classification model to obtain the rumor detection result of the to-be-detected event.

According to an early rumor detection method based on generating a propagation structure feature provided by the present invention, the feature extraction layer of the classification model is constructed and generated based on a cyclic neural network model;

The classification layer of the classification model is constructed and generated based on the softmax function.

According to an early rumor detection method based on generating a propagation structure feature provided by the present invention, the propagation structure diagram is constructed based on the following steps:

Taking the second source text and each response tweet in the sample event as nodes, the propagation structure diagram of the sample event is constructed according to the response relationship between any two nodes.

According to an early rumor detection method based on generating a propagation structure feature provided by the present invention, the response relationship includes a forwarding relationship, a time dimension relationship and a commenting relationship.

According to an early rumor detection method based on generating a propagation structure feature provided by the present invention, the second propagation structure feature is obtained based on the following steps:

Input the propagation structure diagram of the sample event into the graph convolutional neural network model, and obtain the second propagation structure feature of the sample event output by the feature extraction layer of the graph convolutional neural network model;

Wherein, the graph convolutional neural network model is obtained by training according to the propagation structure diagram of the sample event and the rumor detection label of the sample event.

According to an early rumor detection method based on generating a propagation structure feature provided by the present invention, the first source text is input into a conditional variational auto-encoder, and the to-be-varied auto-encoder generated by the conditional variational auto-encoder is obtained. Detect first propagation structural features of events, including:

preprocessing the first source text;

Wherein, the preprocessing includes removing garbled data and text vectorization;

Inputting the preprocessed first source text into the conditional variational autoencoder to obtain the first propagation structure feature of the event to be detected.

The present invention also provides an early rumor detection device based on the generated propagation structure feature, comprising:

A text acquisition module for acquiring the first source text of the event to be detected;

A propagation structure feature generation module, configured to input the first source text into a conditional variational autoencoder, and obtain the first propagation structure feature of the to-be-detected event generated by the conditional variational autoencoder; wherein, the The conditional variational autoencoder is obtained by training according to the second source text of the sample event and the second propagation structure feature; the second propagation structure feature is obtained by feature extraction on the propagation structure graph of the sample event based on the graph convolutional neural network model. ; The propagation structure diagram is constructed and obtained according to the second source text of the sample event and the response tweet corresponding to the second source text;

A rumor detection module, configured to input the first source text and the first propagation structure feature into a classification model, and obtain the rumor detection result of the to-be-detected event output by the classification model; the classification model is based on the second The source text information, the second propagation structure feature, and the rumor detection label of the sample event are acquired through training.

The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, when the processor executes the program, the generation-based propagation as described in any of the above is realized Steps of an early rumor detection method for structural features.

The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, implements the steps of any of the above-mentioned methods for detecting early rumors based on generating propagation structure features .

The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of any of the above-mentioned methods for detecting early rumors based on generating propagation structure features.

The method and device for early rumor detection based on generating propagation structure features provided by the present invention construct a propagation structure graph according to the second source text and response tweets of sample events, and perform feature extraction on the propagation structure graph based on a graph convolutional neural network model. Obtaining the second propagation structure feature of the sample event, and iterating according to the second source text and the second propagation structure feature of the sample event, to obtain a conditional variational autoencoder that can simulate and generate the propagation structure feature according to the source text of the event; When the first source text of the event to be detected is obtained, the first propagation structure feature of the event to be detected can be automatically simulated and generated based on the conditional variational autoencoder; then, the first propagation structure feature of the event to be detected is combined with the first The source text can be used to detect rumors of the event to be detected; on the one hand, in the entire rumor detection process, only the first source text of the event to be detected needs to be obtained, and the corresponding communication structure features can be quickly generated, reducing the collection of a large number of response tweets. Only then can the delay caused by the characteristics of the communication structure be obtained, and the rumors can be detected at the source of the release of the rumors, which effectively improves the timeliness of rumor detection; on the other hand, the source text information of the event and the communication structure characteristics are combined. , which enriches the features related to rumor detection and makes the rumor detection results more accurate.

Description of drawings

In order to explain the present invention or the technical solutions in the prior art more clearly, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are the For some embodiments of the invention, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without any creative effort.

Fig. 1 is one of the schematic flow charts of the early rumor detection method based on the generated propagation structure feature provided by the present invention;

2 is a schematic structural diagram of a conditional variational autoencoder, a graph convolutional neural network model and a classification model in the early rumor detection method based on the generated propagation structure feature provided by the present invention;

3 is the second schematic flow chart of the early rumor detection method based on the generated propagation structure feature provided by the present invention;

Fig. 4 is the third schematic flow chart of the early rumor detection method based on the generated propagation structure feature provided by the present invention;

5 is a schematic structural diagram of an early rumor detection device based on generating a propagation structure feature provided by the present invention;

FIG. 6 is a schematic structural diagram of an electronic device provided by the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention. , not all examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

In the description of this embodiment, it should be understood that the terms "first" and "second" are only used to distinguish different descriptions, and cannot be understood as indicating or implying relative importance or hierarchical order.

The existing rumor detection methods cannot detect rumor events in a timely and accurate manner at the source of the rumor release, and the real-time performance is poor; in addition, the amount of data for early detection is small, and rumors can only be detected based on the text information of the event, only considering The single granularity of text information cannot accurately obtain rumor detection results.

In view of the above text, this embodiment proposes an early rumor detection method based on the generated propagation structure feature. According to a large amount of rumor propagation data of historical events, a conditional variational autoencoder is constructed to learn the structural features of rumor propagation, so that when there is only the source text of the event to be detected in the early stage, the trained conditional variational autoencoder can be used to predict the to-be-detected Temporal propagation structure characteristics, and combine it with the source text for rumor detection, so that when the source text of the time to be detected is obtained, the rumor detection result can be quickly and accurately obtained to determine whether the to-be-detected event is a rumor event.

It should be noted that the execution body of the above method may be an electronic device, a component in the electronic device, an integrated circuit, or a chip. The electronic device may be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device may be a mobile phone, a tablet computer, a notebook computer, a palmtop computer, an in-vehicle electronic device, a wearable device, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (personal digital assistant). assistant, PDA), etc., non-mobile electronic devices can be servers, network attached storage (Network Attached Storage, NAS), personal computer (personal computer, PC), smart TV (television, TV), teller machine or self-service machine, etc. The invention is not specifically limited.

The early rumor detection method based on the generated propagation structure feature of the present invention is described below in conjunction with FIG. 1 , and the method includes:

Step 101, obtaining the first source text of the event to be detected;

Optionally, the event to be detected is an event that needs to be detected by rumors. The to-be-detected event may be crawled in real time from the Internet by using a crawler tool, or may be obtained through Internet transmission or download, which is not specifically limited in this embodiment.

The sample event and the to-be-detected event may be events published on various social networking sites, such as posts sent on Weibo or dynamics published on WeChat, which are not specifically limited in this embodiment.

The following sample events and events to be detected are posts sent on Weibo as an example to describe the rumor detection method in this embodiment.

The first source text of the to-be-detected event is text information contained in a source post published for the to-be-detected event.

Step 102, inputting the first source text into a conditional variational auto-encoder, to obtain the first propagation structure feature of the to-be-detected event generated by the conditional variational auto-encoder; wherein, the conditional variational auto-encoder The encoder is trained and obtained according to the second source text of the sample event and the second propagation structure feature; the second propagation structure feature is obtained by feature extraction on the propagation structure diagram of the sample event based on the graph convolutional neural network model; the propagation structure The figure is constructed and obtained according to the second source text of the sample event and the response tweet corresponding to the second source text;

Optionally, the conditional variational autoencoder is obtained by training according to the second source text of the sample event and the second propagation structure feature; the trained conditional variational autoencoder can generate the corresponding propagation structure feature according to the source text. , the propagation structure features and source text are jointly generated in the subsequent rumor detection, and early rumor detection can be performed quickly and accurately.

The response tweet is the text of commenting and forwarding the source text, that is, the response tweet is the text information of the comment post commenting on the source post, and the text information of the forwarding post that forwards the source post. In this embodiment, comment posts and reposts are collectively referred to as reposts.

Conditional Variational AutoEncoder (CVAE) is a deep latent space generation model, which consists of an encoder and a decoder; the encoder and decoder are composed of neural network models, and the type and structure of the model can be Set according to actual requirements, such as a convolutional neural network and a cyclic neural network, which are not specifically limited in this embodiment.

Optionally, before performing step 102, it is necessary to obtain a conditional variational autoencoder capable of generating the propagation structure feature according to the second source text of the sample event and the second propagation structure feature after the training is completed. The specific training steps may include:

First, collect samples and organize the data set;

Optionally, collect a large number of sample posts, and in units of events, aggregate the source posts and reposts of the same event to form a sample event data set C={c ¹ ,c ² ,... ^{ci ,...,c n }} ; Among them, ci represents the ^ith sample event, and n represents the number of sample events included in the sample event data set, and the specific number can be set according to actual requirements, which is not specifically limited in this embodiment.

代表第i个样本事件的源文本，

代表第i个样本事件的第j个相关评论帖或转发帖，即响应推文，m_i代表第i个事件中包含的响应推文的总数；

到

为第i个样本事件的所有响应推文。sample event

represents the source text of the ith sample event,

represents the jth related comment post or retweet post of the ith sample event, i.e. the response tweet, m _i represents the total number of response tweets contained in the ith event;

arrive

All response tweets for the ith sample event.

其中，c_i代表第i个样本事件的源文本向量，

到

为第i个样本事件的所有响应推文的文本向量。Then, each text of the sample event can be preprocessed to obtain the text vector of each text in the sample event

where c _i represents the source text vector of the ith sample event,

arrive

Text vector of all response tweets for the ith sample event.

Then, construct and obtain a propagation structure diagram according to the second source text of the sample event and the response tweet corresponding to the second source text; and perform feature extraction on the propagation structure diagram of the sample event based on the graph convolutional neural network model to obtain the second Propagation structure characteristics;

代表第i个样本事件的源文本向量，n_i代表源文本向量c_i中包含的单词数。将所有样本事件的传播结构特征H＝{H₀,H₁,…,H_i,…,H_n}，作为条件变分自编码器中的真实样本数据；其中，H_i为第i个样本事件的传播结构特征。Then, the source text vector c= _{ c ₁ , c ₂ . . . , c _i , . in

represents the source text vector of the _ith sample event, and _ni represents the number of words contained in the source text vector ci. The propagation structure feature H={H ₀ , H ₁ ,…,H _i ,…,H _n } of all sample events is taken as the real sample data in the conditional variational autoencoder; where H _i is the ith sample Propagation structure characteristics of events.

Optionally, take the source text vector and the propagation structure features of each sample event as a set of samples, calculate and obtain the loss function of the conditional variational autoencoder according to the multiple sets of samples, and perform the conditional variational autoencoder based on the loss function. Iterative training.

Among them, the learning principle of the conditional variational autoencoder is to encode the original data into a latent variable z that conforms to a specific distribution. Here, the latent variable z is set to conform to the standard normal distribution, that is, p(z)=N(1,0 ), and then decode the latent variable z to generate new data.

For the encoder E(·), input c _i as condition information together with the real sample data Hi into the encoder E(·), and the encoder E(·) calculates the value of p(z|H _i , c _{i )} Mean μ _i and variance δ _i ² .

For the decoder D(·), the hidden variable z is obtained by sampling using the heavy parameter technique, and the _decoder reconstructs the sample data Hi through the hidden variable z.

Among them, the calculation formula of the hidden variable z obtained by sampling using the heavy parameter technique is:

z=μ _i +ε⊙δ _i ;

Among them, ε obeys the standard normal distribution N(0, 1), ⊙ is the dot product operator, μ _i and δ _i ² are the mean and variance of the normal distribution;

The calculation formula of the loss function of the CVAE model is:

为重构损失函数，E(·)为数学期望；

为编码器的分布

与先验分布p_θ(z)之间的KL散度；

为关于z的后验概率，p_θ(H_i|z，c_i)为关于H_i的条件概率；

与θ分别为编码器与解码器的网络参数。in,

is the reconstruction loss function, E( ) is the mathematical expectation;

distribution for the encoder

KL divergence from the prior distribution p _θ (z);

is the posterior probability about z, p _θ (H _i |z, c _i ) is the conditional probability about _Hi ;

and θ are the network parameters of the encoder and decoder, respectively.

Based on the loss function of the CVAE model above, the network parameters of the CVAE model are iteratively updated, and finally the decoder of the CVAE model can reconstruct the simulated propagation structure features that approximate the real sample data.

After the training of the CVAE model is completed, the first source text of the event to be detected is input into the trained CVAE model, and the first propagation structure feature corresponding to the first source text can be generated.

By learning the propagation structure features from the forwarding information, comment information and source text of a large number of historical events, it is possible to simulate and generate the propagation structure features corresponding to the source text when there is only the source text in the early stage, without waiting for the source text to be widely diffused, effectively Improve the timeliness of detection; in addition, by synthesizing the source text and the corresponding propagation structure characteristics of the source text, rumors can be detected in a timely and accurate manner when the source text is released in the early stage, avoiding the adverse effects caused by the spread of rumor events.

Step 103, inputting the first source text and the first propagation structure feature into a classification model, and obtaining the rumor detection result of the to-be-detected event output by the classification model; the classification model according to the second source text information and the second propagation structure feature and the rumor detection label of the sample event are obtained by training.

Among them, the network structure of the classification model can be set according to actual needs, such as the number of layers of the network and the number of nodes in each layer. The classification model may be constructed and generated by various artificial intelligence models with classification functions, including but not limited to convolutional neural networks, recurrent neural networks, computer vision group networks, and residual network series, which are not specifically limited in this embodiment.

The rumor detection label is the true rumor detection category of the sample event, including rumor and non-rumor.

Optionally, before step 103 is performed, a trained classification model may be obtained by pre-training, and the specific training method is iterative training. The specific training steps may include: first, acquiring the second source text information and the second propagation structure feature of the sample event ; Take the second source text information of the sample event and the second propagation structure feature as the input information, and use the rumor detection label of the sample event as the real label to iteratively train the classification model to obtain the information that accurately distinguishes rumor events and non-rumor events. classification model.

After the training of the classification model is completed, the first source text of the event to be detected and the first propagation structure feature are spliced and input into the trained classification model to determine whether the event to be detected is a rumor event or a non-rumor event.

In this example, a propagation structure diagram is constructed according to the second source text of the sample event and the response tweet, and the feature extraction is performed on the propagation structure diagram based on the graph convolutional neural network model to obtain the second propagation structure feature of the sample event. The second source text and the second propagation structure feature are iterated to obtain a conditional variational autoencoder that can simulate and generate the propagation structure feature according to the source text of the event; when the first source text of the event to be detected is obtained, the Based on the conditional variational autoencoder, the first propagation structure feature of the event to be detected is automatically generated by simulation; then, the first propagation structure feature of the event to be detected and the first source text can be combined to perform rumor detection on the event to be detected; on the one hand, In the entire rumor detection process, only the first source text of the event to be detected can be obtained, and the corresponding propagation structure features can be quickly generated, reducing the delay caused by collecting a large number of response tweets before obtaining the propagation structure features. Rumors can be detected at the source of the release, which effectively improves the timeliness of rumor detection; on the other hand, the combination of the source text information of the event and the communication structure features enriches the features related to rumor detection, making the rumor detection results more accurate. precise.

On the basis of the above embodiment, in this embodiment, the first source text and the first propagation structure feature are input into the classification model, and the rumor detection result of the to-be-detected event output by the classification model is obtained, including : Input the first source text into the feature extraction layer of the classification model to obtain the text features of the first source text; after splicing the text features and the first propagation structure features, input the classification The classification layer of the model obtains the rumor detection result of the event to be detected.

Optionally, the classification model in this embodiment includes a feature extraction layer and a classification layer;

The feature extraction layer may be constructed and generated based on various neural network models, such as a cyclic neural network and a convolutional neural network, which are not specifically limited in this embodiment.

Optionally, according to the first source text and the first dissemination structure feature, the step of performing rumor detection on the event to be detected includes, first, inputting the first source text into a feature extraction layer of a classification model to obtain text features of the first source text;

Then, the text features of the first source text and the first propagation structure features are spliced to obtain fusion features; then the fusion features are input into the classification layer to obtain the probability P(rumor) of the event to be detected belonging to rumors, according to the event to be detected. The probability of belonging to a rumor is determined to determine whether the event to be detected is a rumor.

The specific determination method may be as follows: comparing the probability of the event to be detected belonging to a rumor with a preset value, and if it is greater than the preset value, it is determined that the event to be detected belongs to a rumor event, otherwise it belongs to a rumor event.

In this embodiment, the feature extraction layer of the classification model performs feature extraction on the first source text, which can extract the deep features that affect the rumor detection, so that the rumor detection result is more accurate; , rumor detection of the event to be detected can further improve the accuracy of rumor detection.

On the basis of the above embodiment, the feature extraction layer of the classification model in this embodiment is constructed and generated based on the cyclic neural network model; the classification layer of the classification model is constructed and generated based on the softmax function.

Optionally, the classification model is constructed and generated based on a recurrent neural network and a classification layer;

Among them, the cyclic neural network may be a standard cyclic neural network (Recurrent Neural Network, RNN), which is also one of its derivative algorithms, such as an LSTM (Long short-term memory, long short-term memory) network, which is not made in this embodiment. Specifically defined.

As shown in Figure 2, the classification model is constructed by LSTM network and softmax function.

The rumor detection method in this embodiment is described below by taking the classification model constructed by the LSTM network and the softmax function as an example.

Before performing rumor detection on the event to be detected based on the classification model, the trained classification model can be pre-trained, and the specific training steps can include:

n_i是源文本向量c_i中包含的单词数量，x_t代表源文本向量c_i中的第t个词的词向量。First, vectorize the second source text of the sample event to obtain the second source text vector of the sample event i

n _i is the number of words contained in the source text vector c _i , and x _t represents the word vector of the t-th word in the source text vector c _i .

Then, ci is fed into an _LSTM network to extract high-order feature representations of the source text vectors

Among them, the LSTM network contains three gates, namely forget gate, input gate and output gate. There are three inputs to each gate unit, which are the input value x _t of the LSTM network at the current moment, the output value a _t-1 of the LSTM network at the previous moment, and the unit state value c _t-1 at the previous moment.

The calculation formula of the forget gate in the LSTM network is:

F _f =σ(W _f [a _t-1 , x _t ]+b _f );

F _f is the output of the forget gate, which is used to read the input information x _t of the current network and the output value a _t-1 of the LSTM unit at the previous moment; σ( ) is the activation function, and the control output is between 0 and 1 ; W _f is the weight coefficient, b _f bias term, at _-1 is the output value of the LSTM network at the previous moment. F _f determines how much information of the state value of the LSTM network at the last moment will be abandoned, and the value is between 0 and 1; when F _f is 0, it means forgetting the state value c _t-1 of the unit at the previous moment. All information, if the value is 1, all the information in c _t-1 is retained.

The input gate in LSTM is used to control how much information of the input data x _t is input into the unit. The calculation formula is:

F _i =σ(W _i [a _t-1 , x _t ]+ _bi );

是LSTM网络状态值的中间过渡量，c_t是当前单元更新后的状态值；W_i和W_c是权重系数，b_i与b_c是偏置项。F _i is the output of the input gate, which controls how much input information can be fed into the network unit

is the intermediate transition amount of the state value of the LSTM network, c _t is the updated state value of the current unit; _Wi and W _c are the weight coefficients, and _bi and b _c are the bias terms.

The output gate of LSTM controls how much information is output from the current cell state value c _t-1 . The calculation formula is:

F _o =σ(W _o [a _t-1 , x _t ]+b _o );

a _t =F _o *c _t ;

F _o is the output control gate, which controls the information output ratio of the current unit state value _ct . W _o and b _o are weight coefficients and bias terms. at is the high-order feature vector of the source text finally output by the _LSTM network.

Then, the conditional variational autoencoder is extracted to generate the second propagation structure feature of the sample event, and the second propagation structure feature is spliced with the corresponding source text feature vector a _i , and then passes through a fully connected layer and uses the softmax function for classification and detection, To get the probability of belonging to the rumor label, the calculation formula is:

是分类模型输出的第i个样本事件cⁱ的谣言检测结果，a_i是样本事件cⁱ的源文本的高阶特征表示；H_i为样本事件cⁱ的传播结构特征。in,

is the rumor detection result of the ith sample event c ⁱ output by the classification model, a _i is the high-order feature representation of the source text of the sample event c ⁱ ; H _i is the propagation structure feature of the sample event c ⁱ .

最终预测的标签概率表达式可转换为下式：Since the integral over Hi is not obtainable, Monte Carlo is used to sample the propagation structure features generated by the conditional variational _autoencoder M times and take the mean value as the expectation

The final predicted label probability expression can be transformed into the following formula:

是预测的谣言检测标签；a_i是源文本的高阶特征，z是隐变量。可使用负对数似然函数作为分类模型的损失函数，具体计算公式：

is the predicted rumor detection label; _ai is the high-order feature of the source text, and z is the latent variable. The negative log-likelihood function can be used as the loss function of the classification model. The specific calculation formula is:

是第i个样本事件的预测谣言检测标签。N is the number of sample events, a _i is the high-order feature representation of the source text of sample events c ⁱ ;

is the predicted rumor detection label of the ith sample event.

Finally, the classification model is iteratively trained according to the loss function of the classification model until the training termination condition of the classification model is satisfied, so as to obtain a classification model that can accurately distinguish rumor events and non-rumor events.

In this embodiment, by using a recurrent neural network, the conditional generative adversarial network can better capture the text used to characterize the rumor event and the non-rumor event in the time series, thereby making the detection result more accurate; in addition, using Monte Carlo The method of sampling the propagation structure features generated by the conditional variational autoencoder for many times and then calculating the mean value can transform the rumor detection probability function, and then the optimal classification model can be obtained quickly and effectively.

On the basis of the above embodiments, the propagation structure diagram in this embodiment is constructed based on the following steps: taking the second source text and each response tweet in the sample event as nodes, according to the relationship between any two nodes The response relationship constructs a propagation structure diagram of the sample event.

Among them, the dissemination structure diagram represents the interaction process among the posts in the event and the dissemination structure of the information, that is, the response relationship between the posts.

Optionally, take the event as a unit to construct a propagation structure diagram for each sample event; for the i-th event, its propagation structure diagram is G _i =(V _i ,E _i ); wherein, V _i is a node of the propagation diagram set, each node represents a post, E _i is the edge set of the ith event propagation structure graph, each edge represents the response relationship between any two post nodes, A _i represents the ith event propagation structure graph adjacency matrix.

Among them, the adjacency matrix can be constructed and generated according to the response relationship between any two post nodes.

The response relationship is the response relationship between any two posts, which can be set according to actual needs.

In this embodiment, by constructing a propagation structure diagram, it is possible to mine the propagation structure information among the posts in the event, capture richer contextual relationships, and utilize rumor detection.

On the basis of the foregoing embodiment, the response relationship in this embodiment includes a forwarding relationship, a time dimension relationship, and a commenting relationship.

Optionally, the forwarding relationship is used to represent the forwarding hierarchical relationship between two posts. There may or may not be a forwarding relationship between any two posts; for example, there is a forwarding relationship between the source post and the forwarding post; There is no forwarding relationship between the two forwarded posts of the forwarded text.

The comment relationship is used to represent the comment level relationship between two posts. There may or may not be a comment relationship between any two posts; for example, there is a forwarding relationship between the source post and the comment post; There is no comment relationship between the two comment posts.

The time dimension relationship is the time dimension that exists between any two posts that forward or comment on the same post; or the time dimension that exists when different posts are published.

The comprehensive response relationship in this embodiment includes a forwarding relationship, a time dimension relationship, and a comment relationship to construct a communication structure diagram, so that the communication structure diagram contains rich relationship features between posts, thereby making the rumor detection result more accurate.

On the basis of the above embodiments, the second propagation structure feature in this embodiment is obtained based on the following steps: inputting the propagation structure diagram of the sample event into the graph convolutional neural network model to obtain the graph volume The second propagation structure feature of the sample event output by the feature extraction layer of the convolutional neural network model; wherein, the graph convolutional neural network model is based on the propagation structure diagram of the sample event and the rumor detection label of the sample event. training acquisition.

It should be noted that, before the conditional variational autoencoder is trained, the second propagation structure feature of the sample event needs to be acquired in advance; the obtaining step of the second propagation structure feature includes: first, obtaining the propagation structure diagram of the sample event; then, Each node and adjacency matrix in the propagation structure graph are used as the input of the graph convolutional neural network model, and each node in the propagation structure graph is updated to obtain the corresponding propagation structural features. The update of the information of each node is formed by the aggregation of the adjacent nodes connected to it and its own information.

Optionally, a Graph Convolution Network (GCN) model is used to construct event propagation structural features. The node update formula between each layer of GCN and the next layer of GCN is:

为第i个事件传播结构图的经过l层卷积更新后所有节点的特征矩阵；其中h_ij ^(l)为事件的第j个节点经过GCN更新后的特征向量；

为归一化对称邻接矩阵，D_i为邻接矩阵A_i的度矩阵；W^(l)为权重矩阵；B^(l)为偏置项；σ(·)为激活函数。当l＝0时，H_i ⁽⁰⁾＝X_i；X_i为第i个事件的文本向量；将X_i输入到l层GCN中，经过l层卷积更新后得到传播图特征矩阵H_i ^(l)，则所有事件传播图特征矩阵H^(l)＝{H₁ ^(l)，H₂ ^(l)，…，H_i ^(l)，…，H_n ^(l)}。in,

is the feature matrix of all nodes after the l-layer convolution update of the i-th event propagation structure diagram; wherein h _ij ^(l) is the feature vector of the j-th node of the event after GCN update;

is the normalized symmetric adjacency matrix, D _i is the degree matrix of the adjacency matrix A _i ; W ^(l) is the weight matrix; B ^(l) is the bias term; σ(·) is the activation function. When l=0, H _i ⁽⁰⁾ =X _i ; X _i is the text vector of the i-th event; X _i is input into the 1-layer GCN, and the propagation map feature matrix H _i is obtained after the 1-layer convolution update ^(l) , then all event propagation graph feature matrices H ^(l) = {H ₁ ^(l) , H ₂ ^(l) , ..., H _i ^(l) , ..., H _n ^(l) }.

The obtained propagation map feature H _i ^(l) is pooled to obtain H _i , and the softmax classifier is used to predict the rumor label of the event. The formula is:

是第i个事件的标签，W是权重参数，b是偏置项。in,

is the label of the ith event, W is the weight parameter, and b is the bias term.

Then, the loss function is obtained according to the predicted rumor label and the real rumor label; the graph convolutional neural network model is trained based on the loss function. Among them, the loss function uses the cross entropy function to optimize the model as:

和y_ic分别为事件的预测谣言标签和真实谣言标签；N为样本事件的数量；C是标签类别数量，本实施例中标签类别数量为2，即谣言事件与非谣言事件。

and y _ic are the predicted rumor label and the real rumor label of the event, respectively; N is the number of sample events; C is the number of label categories. In this embodiment, the number of label categories is 2, that is, rumor events and non-rumor events.

It should be noted that the rumor detection here is auxiliary training. After the training is completed, the propagation structure feature H _i of the sample event is obtained, which is used as the training data of the conditional variational autoencoder.

In this embodiment, the graph convolutional neural network model can quickly and accurately acquire the second propagation structure feature of the sample event, and use it as training data to train a conditional variational autoencoder that simulates and generates the propagation structure feature according to the source text.

On the basis of the above embodiments, in this embodiment, the first source text is input into the conditional variational autoencoder, and the first source of the to-be-detected event generated by the conditional variational autoencoder is obtained. Propagating structural features includes: preprocessing the first source text; wherein, the preprocessing includes removing garbled data and text vectorization; inputting the preprocessed first source text into the conditional variational auto-encoding In the device, the first propagation structure feature of the to-be-detected event is obtained.

Optionally, before performing step 102, data processing may be performed on the first source text; the preprocessing step includes: first, removing garbled data and/or blank line data in the first source text information, so as to obtain a data that does not contain Valid first source text information of garbled data and/or blank line data; then, use the Doc2vec model to encode the first source text into a text vector to obtain a text vector of the first source text information

In this embodiment, by converting the first source text into a text vector, each word in the text can be represented as a low-dimensional real number domain value vector, and the semantic feature of each word can also be encoded into the text vector.

After preprocessing the first source text information, the text vector of the first source text information can be obtained; the first source text information text vector can be input into the conditional variational autoencoder to obtain the first propagation of the event to be detected. Structure.

It should be noted that, in the training process, the second source text and response tweets of the sample events can also be preprocessed in the above-mentioned manner, and the preprocessed second source text and response tweet training samples can be used to evaluate the conditions. Variational autoencoders, graph convolutional neural network models, and classification models are trained.

In this embodiment, by preprocessing the first source text, a word vector containing the semantic features of each word can be generated, and at the same time, a foundation is laid for subsequent rumor detection.

As shown in FIG. 3 and FIG. 4 , an early rumor detection method based on generating propagation structure features proposed in this embodiment mainly includes the following steps: collecting samples; extracting data and vectorizing the text; constructing the propagation of sample events Structural graph, using graph convolutional network to construct propagation structural features; training graph variational autoencoder according to the source post and propagation structural features of sample events, and obtaining a generator that can accurately generate propagation structural features after training; Source post and propagation structure features, as well as real rumor labels, train the classification model to obtain a classification model that can accurately classify events; take the source post of the event to be detected as a condition, and the generator generates a simulated propagation corresponding to the condition. Structural features; combine the source post of the event to be detected with its corresponding simulated propagation structural features, and then use a softmax function to detect rumors.

The above steps are described in detail below, including:

The specific logic flow is as follows:

Step 1, collect sample data;

Step 2, text vectorization; this step specifically includes:

Step 2.1, process the sample data set, and aggregate the source posts and their repost information of the same event in units of events;

Step 2.2, use Doc2vec to vectorize all the posts of the sample event to get the text vector representation of each post;

Step 3, train the graph convolutional network GCN model to construct the propagation structure features of the event; this step specifically includes:

Step 3.1, construct the sample event propagation structure graph G=(V, E), and obtain the adjacency matrix A. V is the node set, E is the edge set, each node represents a post, and each edge represents the forwarding relationship, time dimension relationship and comment relationship between two post nodes;

Step 3.2, using the graph convolutional network model GCN to construct the propagation structure features of each sample event.

Step 3.3, input the obtained propagation structure features into the softmax function, obtain the rumor prediction label of the sample event, and train the graph convolution network model according to the loss function obtained from the rumor prediction label and the true label of the rumor, until the training of the model is terminated. condition;

The propagation structure diagram of the sample event is input into the trained graph convolutional network model, and the propagation structure feature H _i of the sample event is obtained, which is used as the training data set of the adversarial network model CGAN for subsequent conditions;

Step 4: Train the CGAN model according to the source post and propagation structure features of the sample event; this step specifically includes:

Step 4.1, extract the text vector of the source post of the sample event as the input condition in the conditional variational autoencoder CVAE model;

Step 4.2, extract the propagation structure feature H _i output by the graph convolutional network model in step 3 as the real sample data in the conditional variational autoencoder CVAE model.

Step 4.3, train CVAE model; wherein, CVAE consists of encoder E(·) and decoder D(·); the input of E(·) is source frame c _i (as a condition) and real sample data H _i , output is the mean μ and variance δ ² of the latent variable z; the input of D(·) is the source post c _i and the sampled latent variable z, and the output is the reconstructed data of the real sample data _Hi ;

Step 5: Train the classification model according to the source post and dissemination structure features of the sample event, as well as the true rumor label; this step includes:

Step 5.1, obtain the text vector and propagation structure features of the source post of the sample event;

Step 5.2, using the text vector and the propagation structure feature of the source post of the sample event as input information, and using the rumor detection label of the sample event as the real label, to train the classification model;

Step 6, generating corresponding simulated propagation structure features according to the source post of the event to be detected; this step includes:

Step 6.1, extract the text vector of the source post of the event to be detected;

Step 6.2, input the text vector of the source post into the conditional variational autoencoder, and generate the corresponding simulated propagation structure features through the conditional variational autoencoder CVAE;

Step 7: Perform early rumor detection according to the source post of the event to be detected and the generated simulated propagation structure features; this step includes:

Step 7.1, extract the text information of the source post of the event to be detected, and simulate the propagation structure characteristics;

Step 7.2, input the source text into the long-term and short-term LSTM after word2vec vectorization, and obtain the high-order text feature representation a _i of the source text;

Step 7.3, combine the text feature a _i with the simulated propagation structure feature and input it into the softmax function for rumor detection.

To sum up, in the rumor detection method in this embodiment, first, an event propagation structure graph is constructed according to the source post and repost of the sample event, and then the graph convolutional neural network is used to extract the propagation structure feature of the propagation structure graph of the sample event; then The source post of the sample event and the extracted propagation structure features are input into the conditional variational autoencoder, and the conditional variational autoencoder is trained, and finally a generation of simulated propagation structure features that can generate similar to the real propagation structure features is obtained. Then the source post of the event to be detected is input into the trained generator to generate the simulated propagation structure features corresponding to the source post; finally, the source post and the generated simulated propagation structure feature are combined for early rumor detection, which can be achieved in Rumor events can be detected quickly and accurately when the source post is published.

To sum up, this embodiment uses the graph convolutional network to construct the propagation structure feature, which can better capture the close connection between the source post and the reposted post, and can better learn the hidden information features in the propagation structure; Encoder, which can train a generator with good performance, which can automatically generate the propagation structure features given only the text information of the source post, that is, it can predict the rumors in the early stage of the rumor when only the source post information is available. Propagation structure for early rumor detection. In addition, the generative propagation structure feature model augments the data of early rumor detection in the absence of early propagation information.

The following describes the early rumor detection device based on the generated propagation structure feature provided by the present invention. The early rumor detection device based on the generated propagation structure feature described below and the early rumor detection method based on the generated propagation structure feature described above may refer to each other correspondingly. .

As shown in FIG. 5 , this embodiment provides an early rumor detection device based on generating propagation structure features. The device includes: a text acquisition module 501, a propagation structure feature generation module 502, and a rumor detection module 503, wherein:

The text acquisition module 501 is used to acquire the first source text of the event to be detected;

The propagation structure feature generation module 502 is configured to input the first source text into a conditional variational auto-encoder, and obtain the first propagation structure feature of the to-be-detected event generated by the conditional variational auto-encoder; wherein, the The conditional variational autoencoder is obtained by training according to the second source text of the sample event and the second propagation structure feature; the second propagation structure feature is obtained by feature extraction on the propagation structure graph of the sample event based on the graph convolutional neural network model. ; The propagation structure diagram is constructed and obtained according to the second source text of the sample event and the response tweet corresponding to the second source text;

The rumor detection module 503 is configured to input the first source text and the first propagation structure feature into the classification model, and obtain the rumor detection result of the to-be-detected event output by the classification model; the classification model is based on the second The source text information, the second propagation structure feature, and the rumor detection label of the sample event are acquired through training.

In this embodiment, a propagation structure diagram is constructed according to the second source text of the sample event and the response tweet, and feature extraction is performed on the propagation structure diagram based on the graph convolutional neural network model to obtain the second propagation structure feature of the sample event. The second source text and the second propagation structure feature are iterated to obtain a conditional variational autoencoder that can simulate and generate the propagation structure feature according to the source text of the event; when the first source text of the event to be detected is obtained, that is, The first propagation structure feature of the event to be detected can be automatically simulated and generated based on the conditional variational autoencoder; then, the first propagation structure feature of the event to be detected and the first source text can be combined to perform rumor detection on the event to be detected; on the one hand , in the entire rumor detection process, only the first source text of the event to be detected can be obtained, and the corresponding propagation structure features can be quickly generated, reducing the delay caused by collecting a large number of response tweets before obtaining the propagation structure features. Rumors can be detected at the source of the release of rumors, which can effectively improve the timeliness of rumor detection; on the other hand, the combination of the source text information of the event and the characteristics of the communication structure enriches the features related to the rumor detection, making the rumor detection results. more precise.

On the basis of the above embodiment, the rumor detection module in this embodiment is specifically configured to: input the first source text into the feature extraction layer of the classification model to obtain text features of the first source text; After splicing the text feature and the first propagation structure feature, input the classification layer of the classification model to obtain the rumor detection result of the to-be-detected event.

On the basis of the above embodiment, the feature extraction layer of the classification model in this embodiment is constructed and generated based on the cyclic neural network model; the classification layer of the classification model is constructed and generated based on the softmax function.

On the basis of the above embodiments, the propagation structure diagram in this embodiment is constructed based on the following steps: taking the second source text and each response tweet in the sample event as nodes, according to the relationship between any two nodes The response relationship constructs a propagation structure diagram of the sample event.

On the basis of the foregoing embodiment, the response relationship in this embodiment includes a forwarding relationship, a time dimension relationship, and a commenting relationship.

On the basis of the above embodiments, the second propagation structure feature in this embodiment is obtained based on the following steps: inputting the propagation structure diagram of the sample event into the graph convolutional neural network model to obtain the graph volume The second propagation structure feature of the sample event output by the feature extraction layer of the convolutional neural network model; wherein, the graph convolutional neural network model is based on the propagation structure diagram of the sample event and the rumor detection label of the sample event. training acquisition.

On the basis of the above embodiments, the propagation structure feature generation module in this embodiment is specifically configured to: preprocess the first source text; wherein, the preprocessing includes removing garbled data and text vectorization; Inputting the preprocessed first source text into the conditional variational autoencoder to obtain the first propagation structure feature of the event to be detected.

FIG. 6 illustrates a schematic diagram of the physical structure of an electronic device. As shown in FIG. 6 , the electronic device may include: a processor (processor) 601, a communication interface (Communications Interface) 602, a memory (memory) 603 and a communication bus 604, The processor 601 , the communication interface 602 , and the memory 603 communicate with each other through the communication bus 604 . The processor 601 can invoke the logic instructions in the memory 603 to execute an early rumor detection method based on generating a propagation structure feature, the method comprising: acquiring a first source text of an event to be detected; inputting the first source text into a conditional variation In the autoencoder, the first propagation structure feature of the to-be-detected event generated by the conditional variational autoencoder is obtained; wherein, the conditional variational autoencoder is based on the second source text of the sample event and the second propagation structure The structural feature is obtained by training; the second propagation structural feature is obtained by extracting the propagation structure graph of the sample event based on the graph convolutional neural network model; the propagation structural graph is based on the second source text of the sample event and the The response tweet corresponding to the second source text is constructed and obtained; the first source text and the first propagation structure feature are input into the classification model, and the rumor detection result of the to-be-detected event output by the classification model is obtained; the classification The model is trained and acquired according to the second source text information and the second propagation structure feature, and the rumor detection label of the sample event.

In addition, the above-mentioned logic instructions in the memory 603 may be implemented in the form of software functional units and may be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention can be embodied in the form of a software product in essence, or the part that contributes to the prior art or the part of the technical solution. The computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, Read-Only Memory (ROM, Read-Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes .

In another aspect, the present invention also provides a computer program product, the computer program product includes a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer can Execute the early rumor detection method based on the generated propagation structure feature provided by the above methods, the method includes: obtaining the first source text of the event to be detected; inputting the first source text into the conditional variational encoder, and obtaining the the first propagation structure feature of the event to be detected generated by the conditional variational autoencoder; wherein, the conditional variational autoencoder performs training and acquisition according to the second source text and the second propagation structure feature of the sample event; The second propagation structure feature is obtained by feature extraction on the propagation structure diagram of the sample event based on the graph convolutional neural network model; the propagation structure diagram is obtained according to the second source text of the sample event and the response corresponding to the second source text Tweets are constructed and acquired; the first source text and the first propagation structure feature are input into the classification model, and the rumor detection result of the to-be-detected event output by the classification model is obtained; the classification model is based on the second source The text information and the second propagation structure feature, and the rumor detection label of the sample event are acquired through training.

In yet another aspect, the present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the computer program is implemented to execute the early rumors based on the generated propagation structure features provided by the above methods A detection method, the method comprising: acquiring a first source text of an event to be detected; inputting the first source text into a conditional variational auto-encoder, and obtaining a conditional variational auto-encoder of the to-be-detected event generated by the encoder The first propagation structural feature; wherein, the conditional variational autoencoder is obtained by training according to the second source text of the sample event and the second propagation structural feature; the second propagation structural feature is based on the graph convolutional neural network model. The propagation structure diagram of the event is obtained by feature extraction; the propagation structure diagram is constructed and obtained according to the second source text of the sample event and the response tweet corresponding to the second source text; the first source text and the first source text are obtained. The propagation structure feature is input into the classification model, and the rumor detection result of the to-be-detected event output by the classification model is obtained; the classification model is based on the second source text information and the second propagation structure feature, and the sample event. The rumor detection label is acquired for training.

The device embodiments described above are only illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in One place, or it can be distributed over multiple network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment. Those of ordinary skill in the art can understand and implement it without creative effort.

From the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and certainly can also be implemented by hardware. Based on this understanding, the above-mentioned technical solutions can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products can be stored in computer-readable storage media, such as ROM/RAM, magnetic A disc, an optical disc, etc., includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform the methods described in various embodiments or some parts of the embodiments.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be The technical solutions described in the foregoing embodiments are modified, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. a kind of early rumor detection method based on generating propagation structure feature, is characterized in that, comprises: Obtain the first source text of the event to be detected; Inputting the first source text into a conditional variational self-encoder, and obtaining the first propagation structure feature of the to-be-detected event generated by the conditional variational self-encoder; wherein, the conditional variational self-encoder is based on The second source text of the sample event and the second propagation structure feature are obtained by training; the second propagation structure feature is obtained by feature extraction on the propagation structure graph of the sample event based on the graph convolutional neural network model; the propagation structure graph Constructing and obtaining according to the second source text of the sample event and the response tweet corresponding to the second source text; The first source text and the first propagation structure feature are input into the classification model, and the rumor detection result of the to-be-detected event output by the classification model is obtained; the classification model is based on the second source text information and the second Propagation structural features, and rumor detection labels of the sample events are obtained by training. 2. The early-stage rumor detection method based on the feature of generating propagation structure according to claim 1, wherein the inputting the first source text and the feature of the first propagation structure into a classification model to obtain the output of the classification model The rumor detection results of the event to be detected, including: Inputting the first source text into the feature extraction layer of the classification model to obtain text features of the first source text; After splicing the text feature and the first propagation structure feature, input the classification layer of the classification model to obtain the rumor detection result of the to-be-detected event. 3. The early-stage rumor detection method based on generation propagation structure feature according to claim 1, is characterized in that, the feature extraction layer of described classification model is constructed and generated based on cyclic neural network model; The classification layer of the classification model is constructed and generated based on the softmax function. 4. the early rumor detection method based on the arbitrary described generation propagation structure feature according to claim 1-3, is characterized in that, described propagation structure diagram is constructed based on following steps: Taking the second source text and each response tweet in the sample event as nodes, the propagation structure diagram of the sample event is constructed according to the response relationship between any two nodes. 5 . The early rumor detection method based on generating a propagation structure feature according to claim 4 , wherein the response relationship includes a forwarding relationship, a time dimension relationship and a commenting relationship. 6 . 6. The early rumor detection method based on any one of claims 1-3, wherein the second propagation structure feature is obtained based on the following steps: Input the propagation structure diagram of the sample event into the graph convolutional neural network model, and obtain the second propagation structure feature of the sample event output by the feature extraction layer of the graph convolutional neural network model; Wherein, the graph convolutional neural network model is obtained by training according to the propagation structure diagram of the sample event and the rumor detection label of the sample event. 7. The early rumor detection method based on generating a propagation structure feature according to any one of claims 1-3, wherein the first source text is input into a conditional variational autoencoder to obtain the condition The first propagation structure feature of the to-be-detected event generated by the variational autoencoder includes: preprocessing the first source text; Wherein, the preprocessing includes removing garbled data and text vectorization; Inputting the preprocessed first source text into the conditional variational autoencoder to obtain the first propagation structure feature of the event to be detected. 8. An early-stage rumor detection device based on generating a propagation structure feature, characterized in that, comprising: A text acquisition module for acquiring the first source text of the event to be detected; A propagation structure feature generation module, configured to input the first source text into a conditional variational autoencoder, and obtain the first propagation structure feature of the to-be-detected event generated by the conditional variational autoencoder; wherein, the The conditional variational autoencoder is trained and acquired according to the second source text of the sample event and the second propagation structure feature; the second propagation structure feature is based on the graph convolutional neural network model. The propagation structure graph of the sample event is characterized. Extraction and acquisition; the propagation structure diagram is constructed and acquired according to the second source text of the sample event and the response tweet corresponding to the second source text; A rumor detection module, configured to input the first source text and the first propagation structure feature into a classification model, and obtain the rumor detection result of the to-be-detected event output by the classification model; the classification model is based on the second The source text information, the second propagation structure feature, and the rumor detection label of the sample event are acquired through training. 9. An electronic device, comprising a memory, a processor and a computer program stored on the memory and running on the processor, wherein the processor implements the program as claimed in claim 1 when executing the program Steps of the early rumor detection method based on generating propagation structure features described in any one of to 7. 10. A non-transitory computer-readable storage medium on which a computer program is stored, characterized in that, when the computer program is executed by a processor, the generation-based propagation structure feature according to any one of claims 1 to 7 is implemented The steps of the early rumor detection method.

Images