Rumor detection methods, devices, equipment and storage media

Abstract

This application relates to the field of text processing technology, and provides a method, apparatus, device, and storage medium for rumor detection. The method includes: acquiring text information to be detected; extracting text features and word frequency statistical features from the text information; determining the credibility of the text information based on the text features, and determining the fusion degree of the text features and the word frequency statistical features based on the credibility; fusing the text features and the word frequency statistical features based on the fusion degree to obtain fused features; and detecting whether the text information contains rumor information based on the fused features. The rumor detection method provided by this application determines the fusion degree of word frequency statistical features and text features through the credibility of the text information, achieving adaptive feature fusion. This effectively prevents overfitting caused by excessive use of auxiliary features, and adaptively provides necessary auxiliary features for rumor detection based on the credibility of the text information, thereby improving the accuracy of rumor detection.

Description

Technical Field

[0001] This application relates to the field of text processing technology, specifically to a rumor detection method, apparatus, device, and storage medium. Background Technology

[0002] In the internet age, social media platforms are popular due to their freedom of speech and information sharing. However, their massive information volume and rapid dissemination also lead to a large number of rumors and misinformation circulating on these platforms. Allowing rumors to spread unchecked can severely disrupt people's normal lives, affect social stability, and cause incalculable harm to society. Most people, due to limitations in professional knowledge, time, and location, are unable to accurately identify rumors. Therefore, quickly and effectively identifying rumors is crucial for purifying the online environment and maintaining social stability.

[0003] Currently, rumor detection primarily relies on extracting text features. While this method can automatically learn text features, its accuracy in identifying rumors is low due to the limited variety of text features. Alternatively, introducing other features on top of the extracted text features can improve detection performance, but this over-focus on extracting features outside the text itself neglects the inherent characteristics of the text, leading to overfitting and ultimately affecting the accuracy of rumor detection. Summary of the Invention

[0004] This application provides a rumor detection method, apparatus, device, and storage medium to address the technical problem that existing rumor detection methods are prone to overfitting due to excessive influx of auxiliary features, thereby affecting the accuracy of rumor detection.

[0005] In a first aspect, embodiments of this application provide a rumor detection method, including:

[0006] Obtain the text information to be detected, and extract the text features and word frequency statistical features of the text information;

[0007] The credibility of the text information is determined based on the text features, and the degree of fusion between the text features and the word frequency statistics features is determined based on the credibility.

[0008] Based on the fusion degree, the text features and the word frequency statistical features are fused to obtain the fused features;

[0009] The text information is detected based on the fusion features to determine whether it contains rumors.

[0010] In one embodiment, the step of fusing the text features and the word frequency statistical features based on the fusion degree to obtain the fused features includes:

[0011] The text features are projected onto a preset information space to obtain text projection features;

[0012] The word frequency statistical features are projected onto the information space to obtain statistical projection features;

[0013] The fusion weight value of the statistical projection feature is determined based on the fusion degree. Based on the fusion weight value, the statistical projection feature and the text projection feature are fused to obtain the fused feature.

[0014] In one embodiment, the step of determining the credibility of the text information based on the text features includes:

[0015] Determine the feature information of the text features and obtain the parameter matrix corresponding to the text features;

[0016] Based on the parameter matrix and the feature information, the text projection features are activated using a preset activation function to obtain the credibility variable value corresponding to the text features. The feature information includes dimension information, content information, and length information.

[0017] The credibility of the text information is evaluated using the credibility variable values.

[0018] In one embodiment, the step of detecting whether the text information contains rumor information based on the fusion features includes:

[0019] Based on a preset attention mechanism, attention fusion is performed on the fused features and the text features to obtain the target features;

[0020] The target features are projected into a preset label space to classify the text information as rumors.

[0021] Based on the rumor classification results of the text information, determine whether the text information contains rumor information.

[0022] In one embodiment, the step of extracting the text features and word frequency statistical features of the text information includes:

[0023] Obtain the word frequency statistical vector of each word in the text information;

[0024] Encoding is performed on each of the aforementioned word frequency statistical vectors to obtain the word frequency statistical features of the text information, wherein the word frequency statistical features contain global information of each of the aforementioned word frequency statistical vectors;

[0025] Extract the sentence vectors from the text information, where the sentence vectors contain global semantic information of the text information;

[0026] Extract the sequence features and semantic features of the text information based on the sentence vectors;

[0027] Based on a preset attention mechanism, attention scores are applied to the sequence features and the semantic features respectively to obtain the corresponding attention sequence features and attention semantic features;

[0028] The attention sequence features and the attention semantic features are concatenated to obtain the text features of the text information.

[0029] In one embodiment, the step of encoding each of the word frequency statistical vectors to obtain the word frequency statistical features of the text information includes:

[0030] Calculate the feature values ​​of each of the word frequency statistical vectors, wherein the feature values ​​include at least the mean and variance of the word frequency statistical vectors;

[0031] Based on the feature values, a variational posterior of the text information is constructed, and the variational posterior is sampled to obtain the latent variables of the text information.

[0032] The word frequency statistical features of the text information are generated based on the latent variables.

[0033] In one embodiment, the step of obtaining the text information to be detected includes:

[0034] Obtain the original text dataset, and use regular expressions to filter and clean the text information in the original text dataset to obtain the text information to be detected.

[0035] Secondly, embodiments of this application provide a rumor detection device, comprising:

[0036] The feature extraction module is used to acquire the text information to be detected and extract the text features and word frequency statistical features of the text information;

[0037] An adaptive gating module is used to determine the credibility of the text information based on the text features, and to determine the fusion degree of the text features and the word frequency statistical features based on the credibility.

[0038] The feature fusion module is used to fuse the text features and the word frequency statistical features based on the fusion degree to obtain fused features;

[0039] The rumor detection module is used to detect whether the text information contains rumor information based on the fusion features.

[0040] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the program to implement the steps of the rumor detection method described in the first aspect.

[0041] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the rumor detection method described in the first aspect.

[0042] The rumor detection method, apparatus, device, and storage medium provided in this application extract text features and word frequency statistical features of the text information to be detected, evaluate the credibility of the text information based on the extracted text features, determine the fusion degree of word frequency statistical features and text features based on the credibility, thereby establishing an adaptive fusion mechanism of word frequency statistical features and text features, realizing adaptive feature fusion, which can effectively prevent overfitting caused by the overuse of auxiliary features, and adaptively provide necessary auxiliary features for rumor detection based on the credibility of text information, thereby improving the accuracy of rumor detection. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is one of the flowcharts illustrating the rumor detection method provided in the embodiments of this application;

[0045] Figure 2 This is one of the schematic diagrams of the feature extraction network structure of the rumor detection method provided in the embodiments of this application;

[0046] Figure 3 This is the second schematic diagram of the feature extraction network structure of the rumor detection method provided in the embodiments of this application;

[0047] Figure 4 This is the third schematic diagram of the feature extraction network structure of the rumor detection method provided in this application embodiment;

[0048] Figure 5 This is one of the structural schematic diagrams of the rumor detection device provided in the embodiments of this application;

[0049] Figure 6 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0051] Figure 1 This is one of the flowcharts illustrating the rumor detection method provided in this application. (Refer to...) Figure 1 The rumor detection method provided in this application includes:

[0052] Step 100: Obtain the text information to be detected, and extract the text features and word frequency statistical features of the text information;

[0053] In this embodiment, when detecting rumors, the text information to be detected is first obtained. This text information may be one or more tweets posted by various users on a social platform. When the tweets contain images, the text information may also include text content extracted from the images.

[0054] The process involves extracting relevant text features and word frequency statistics from the text information to be detected. Text features are inherent features of the text, including semantic and sequence information. Word frequency statistics are statistical features of the text, containing statistical information about the text. The extraction of text features and word frequency statistics can be performed simultaneously, for example, using a parallel network structure, or sequentially. Sequential extraction does not distinguish the order of extraction; no specific limitation is made here. The following explanation uses a parallel network structure to extract both text features and word frequency statistics simultaneously as an example.

[0055] Step 200: Determine the credibility of the text information based on the text features, and determine the fusion degree of the text features and the word frequency statistics features based on the credibility.

[0056] The credibility of the text information is determined based on the extracted text features, and the degree of fusion between the text features and the word frequency statistics features is determined based on the credibility of the text information. In this embodiment, an adaptive gating mechanism is introduced, using the word frequency statistics features of the extracted text information as auxiliary features of the text features. Based on the credibility of the text information, the degree of fusion between the text features and the word frequency statistics features is determined, thereby determining the inflow of auxiliary features and avoiding excessive inflow of auxiliary features that ignore the inherent text features of the text information itself, thus preventing overfitting.

[0057] Furthermore, the text features of the extracted text information can be used as auxiliary features of the word frequency statistical features, and the degree of fusion between the two can be determined based on the reliability of the statistical features. This will not be elaborated further here.

[0058] Step 300: Based on the fusion degree, fuse the text features and the word frequency statistical features to obtain the fused features;

[0059] Step 400: Detect whether the text information contains rumor information based on the fusion features.

[0060] Based on the fusion degree of text features and word frequency statistical features, the text features and word frequency statistical features are fused to obtain fused features of the text information. These fused features contain multi-dimensional features of the text information. Specifically, when performing feature fusion based on the fusion degree of text features and word frequency statistical features, this fusion degree can be used as the weight value of the word frequency statistical features. This weight value can be 0 or 1, i.e., the fusion degree is 0% or 100%. Whether to fuse all word frequency statistical features based on the text features is determined based on the credibility of the text information. Specifically, the credibility threshold is a probability value. When it is close to 0 or 1, the credibility is high; when it is close to 0.5, the credibility is low. When the credibility is within a certain threshold range, all word frequency statistical features are fused based on the text features to obtain the fused features, at which point the fusion degree is 100%. When the credibility is outside the set feature fusion threshold range, all word frequency statistical features are discarded, at which point the fusion degree is 0%.

[0061] Furthermore, the fusion degree can be adjusted based on the credibility, thereby adjusting the corresponding weight values ​​of the word frequency statistical features. For example, when the credibility is 0.5, the fusion degree is 100%, fusing all word frequency statistical features on the basis of text features. When the credibility is (0.4, 0.5) or (0.5, 0.6), the fusion degree is 80%, fusing 80% of the word frequency statistical features on the basis of text features. In this case, the weight value of the word frequency statistical features is 0.8. Thus, the fusion degree of text features and word frequency statistical features can be determined based on the credibility value, thereby determining the weight value of auxiliary features. Feature fusion is then performed based on this weight value to achieve adaptive fusion of features of different dimensions, which can effectively prevent overfitting caused by excessive use of auxiliary features.

[0062] Based on the fusion feature detection method, the method detects whether the text information contains rumors. Based on the adaptive gating mechanism, the fusion degree of auxiliary features is adaptively controlled according to the credibility value of the text information. Without ignoring the inherent features of the text information, the method avoids the overuse of auxiliary features and enhances the classification effect of rumors, thereby improving the detection accuracy of rumors.

[0063] In this embodiment, text features and word frequency statistical features of the text information to be detected are extracted. The credibility of the text information is evaluated based on the extracted text features. The degree of fusion between word frequency statistical features and text features is determined based on the credibility. Thus, an adaptive fusion mechanism between word frequency statistical features and text features is established to achieve adaptive feature fusion. This can effectively prevent overfitting caused by the overuse of auxiliary features. Based on the credibility of the text information, necessary auxiliary features are adaptively provided for rumor detection, thereby improving the accuracy of rumor detection.

[0064] Furthermore, based on an adaptive gating mechanism, the inflow of auxiliary features is automatically controlled, thereby adjusting the degree of use of auxiliary features by text information under different confidence values. This allows for a better balance between text features and word frequency statistical features, which can not only effectively avoid overfitting caused by excessive use of auxiliary features, but also improve the generalization ability of the rumor detection model.

[0065] In one embodiment, step 100, the step of obtaining the text information to be detected, includes:

[0066] Step 110: Obtain the original text dataset, and use regular expressions to filter and clean the text information in the original text dataset to obtain the text information to be detected.

[0067] When obtaining the text information to be detected, the original text dataset is first obtained. Regular expressions are then used to filter and clean the information in the original text dataset, removing useless or interfering information such as stop words and special symbols, to obtain the text information to be detected.

[0068] Furthermore, in step 100, extracting text features and word frequency statistical features from the text information specifically includes:

[0069] Step 101: Obtain the word frequency statistics vector of each word in the text information;

[0070] Step 102: Encode each of the word frequency statistical vectors to obtain the word frequency statistical features of the text information, wherein the word frequency statistical features contain global information of each of the word frequency statistical vectors;

[0071] Step 103: Extract the sentence vector of the text information, wherein the sentence vector contains the global semantic information of the text information;

[0072] Step 104: Extract the sequence features and semantic features of the text information based on the sentence vectors;

[0073] Step 105: Based on a preset attention mechanism, perform attention scoring on the sequence features and the semantic features respectively to obtain the corresponding attention sequence features and attention semantic features;

[0074] Step 106: Concatenate the attention sequence features and the attention semantic features to obtain the text features of the text information.

[0075] In this embodiment, the rumor detection method mainly includes three parts: text feature extraction, word frequency statistical feature extraction, and rumor prediction and classification. The text feature extraction part includes a preprocessing layer, sequence feature extraction, and semantic feature extraction. Furthermore, an attention mechanism is introduced, which can not only highlight the key features of word frequency statistical features, but also extract richer text features.

[0076] Specifically, when extracting word frequency statistical features, the word frequency statistical vector of each word in the text information is first obtained. Based on the distribution of each word on different labels in the text information, a preliminary word frequency statistical vector is generated. If w represents a word and c represents the label type, then the word frequency statistical vector Fc of word w is expressed in the form of the following Formula 1:

[0077] ρ w =[ρ1,...,ρ c (1)

[0078] In Formula 1, ρ i Let S represent the distribution of word w on tag i; if S represents a tweet and m represents the length of the tweet, then the word frequency statistics vector Fc of tweet S can be represented by the following formula 2:

[0079]

[0080] in Let S represent the word frequency statistics vector of the m-th word in tweet S. The word frequency statistics vectors corresponding to each word are encoded to obtain the word frequency statistics features of the text information. These features contain global information about the word frequency statistics vectors.

[0081] Furthermore, when extracting text features, we first obtain sentence vectors containing global semantic information of the text, and then extract sequence features and semantic features of the text information based on the sentence vectors. Based on a pre-defined attention mechanism, we perform attention scoring on the sequence features and semantic features respectively, assigning attention weight values ​​to each word in the text information, thereby obtaining attention sequence features and attention semantic features. We then concatenate the attention sequence features and attention semantic features to obtain multi-dimensional text features of the text information.

[0082] In this embodiment, the text feature extraction part includes preprocessing, sequence feature extraction, semantic feature extraction, and an attention mechanism. Taking the BERT (Bidirectional Encoder Representation from Transformers) model for preprocessing text information and a parallel network of Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) for extracting multi-dimensional text features as an example, when extracting text features, firstly, a tweet text of length m is fed into the BERT model for pre-training. After encoding, a sentence vector X that incorporates global semantic information is output. If x... j ∈R d Let d represent the d-dimensional word vector of the j-th word. The word vectors together form the sentence vector of the text information. Then:

[0083] X 1：m = [x1; x2; ...; x m (3)

[0084] Then, the sentence vector X is input into a parallel attention mechanism network of BiLSTM_Attention and CNN_Attention to extract the sequence features H and semantic features K of the sentence vector, respectively, where:

[0085]

[0086] K = CNN(x) (5)

[0087] Then, H and K are used as inputs to the attention layer, and attention scores are applied to the extracted sequence features H and semantic features K respectively. Different attention weights are assigned to each word to obtain the corresponding features TE1 and TE2. The calculation process is shown in the following formulas 6 to 9:

[0088] u1 = Tanh(W H H+b H ), u2 = Tanh(W K H+b K (6)

[0089]

[0090] TE1=∑a i1 H (8)

[0091] TE2=∑a i2 K (9)

[0092] Where u1 and u2 represent the attention hidden layer representations of sequence feature H and semantic feature K, respectively, and W H It is the parameter matrix of the sequence feature H, b H It is the bias term of the input sequence feature H; similarly, W K It is the parameter matrix of semantic feature K, b K It is the bias term of the input semantic features K, where W H and W K It can be obtained through training. i1 and a i2 These are the normalized weights obtained by applying the Softmax function to the sequence feature H and semantic feature K, respectively. TE1 and TE2 are the attention sequence feature and attention semantic feature obtained after attention scoring. Softmax is the normalization exponential function, and Tanh is the hyperbolic tangent function. Here, both Softmax and Tanh are activation functions of the attention mechanism. and For the transpose of u1 and u2, u s Let S represent the parameter matrix of the tweet S.

[0093] By concatenating TE1 and TE2, we can obtain the text feature TE, which integrates multi-dimensional features. (Refer to...) Figure 2 The diagram shows a parallel text feature extraction network structure consisting of a convolutional neural network and a bidirectional long short-term memory network. This feature extraction network includes BiLSTM_Attention and CNN_Attention. The sentence vector X, containing six word vectors w1 to w6, is input into the parallel network structure of the convolutional neural network and the bidirectional long short-term memory network to extract the sequence features and semantic features of the sentence vector X. Then, the extracted sequence features and semantic features are input into a pre-trained attention model to assign attention weight values ​​to the word vector of each word, thereby obtaining the corresponding attention sequence features and attention semantic features. The obtained features are then concatenated to obtain the final text features containing multi-dimensional features.

[0094] Furthermore, step 102, which involves encoding the word frequency statistical vector to obtain the word frequency statistical features of the text information, may further include:

[0095] Step 1021: Calculate the feature values ​​of each of the word frequency statistical vectors, wherein the feature values ​​include at least the mean and variance of the word frequency statistical vectors;

[0096] Step 1022: Construct the variational posterior of the text information based on the feature values, and perform sampling processing on the variational posterior to obtain the latent variables of the text information;

[0097] Step 1023: Generate word frequency statistical features of the text information based on the latent variables.

[0098] In this embodiment, each word frequency statistical vector is encoded based on a preset variational encoder to obtain the word frequency statistical features of the text information. When encoding the word frequency statistical vectors, the feature values ​​of each vector are first calculated, including at least the mean and variance of the vectors. Then, a variational posterior, such as a normal distribution, is constructed based on the calculated feature values. The variational posterior is then sampled to obtain the corresponding latent variables. Based on the latent variables of each word frequency statistical vector, the word frequency statistical features of the text information are generated. Specifically, the variational encoder includes two encoders. Taking the extracted word frequency statistical vectors ρ (including ρ1 to ρ6) as an example, refer to... Figure 3 The diagram shows the network structure of the variational encoder. This variational encoder comprises a mean-variance calculation module and a generator. The mean-variance calculation module includes two encoders. When encoding the word frequency statistical vectors, the two encoders in the mean-variance calculation module are used to calculate the mean and variance of the word frequency statistical vectors ρ1 to ρ6, respectively. The variational encoders encode the word frequency statistical vectors ρ1 to ρ6, calculating the mean and variance of each vector. Based on the calculated feature values, a variational posterior conforming to a normal distribution is constructed. Variable sampling is performed based on the constructed variational posterior to obtain the latent variables corresponding to the word frequency statistical vectors. Based on the latent variables corresponding to each word frequency statistical vector, the word frequency statistical features of the text information are generated.

[0099] Furthermore, the variational encoder requires pre-training. Specifically, assume a random process p θ (ρ|z) Generates Fc based on all data in the dataset corresponding to the text information, obtaining the latent variables:

[0100]

[0101] In Formula 10, N represents the size of the dataset corresponding to the text information, and the learning parameters θ and φ used to construct the deep learning network. Optimize variational approximation function Make it infinitely close to p θ (ρ|z), the optimization process is shown in Equation 11 below:

[0102]

[0103] In Formula 11, D KL denoted by D, the KL divergence measures the similarity between two distributions, p and q. A smaller value indicates a closer similarity between the two distributions, while a larger value indicates a greater difference. Since D... KL Non-negative, therefore LV Let L be the variational lower bound, i.e.:

[0104]

[0105] During the sampling process, the technique of reparameterization is used to ensure that the variational posterior of the latent variable z with respect to each sample follows a standard normal distribution p. θ (ρ|z) uses two encoders: one to calculate the mean μ of the word frequency statistics vector, and the other to calculate the variance σ of the word frequency statistics vector. 2 Since the approximate prior is multivariate Gaussian, a diagonal covariance structure is used to construct the variational posterior:

[0106]

[0107] Based on the constructed variational posterior, an unsupervised training process is performed using a variational encoder, such as a VAE (Variational Auto-Encoder), to fit the variational posterior. The error loss during the fitting process is measured using KL divergence. Specifically, based on the collected latent variables, a new word frequency statistical vector (i.e., ...) is generated by a generator. Figure 3 In The word frequency statistics vector is then compared with the word frequency statistics vector (i.e., the statistical information input as the text information to be detected). Figure 3 The KL divergence is determined by comparing the newly generated word frequency statistical vectors (ρ1 to ρ6) with the input word frequency statistical vectors. The KL divergence is then used to measure the similarity between the newly generated word frequency statistical vectors and the input word frequency statistical vectors. A low similarity indicates a low variational posterior (i.e.,...) Figure 3 If the fit to the normal distribution (in the model) is not ideal, the learning parameters are adjusted, and training is repeated based on the adjusted learning parameters until a new word frequency statistical vector is generated. The similarity with the input word frequency statistical vectors, i.e., ρ1 to ρ6, is calculated until the condition is met. Following this approach, the fitted variational posterior can be directly used to extract word frequency statistical features from text information for rumor detection. Specifically, the variational posterior constructed based on the text information to be detected is sampled to obtain encoded latent variables. Then, the word frequency statistical features ρ of the text information are generated based on the latent variables obtained from sampling each word frequency statistical vector. z .

[0108] In this embodiment, by introducing an attention mechanism and extracting inherent features of text information from different dimensions such as sequence and semantics, the extracted text features from different dimensions are concatenated to obtain multi-dimensional text features, which is beneficial to improving the detection accuracy of rumors.

[0109] In one embodiment, step 300 specifically includes:

[0110] Step 301: Project the text features into a preset information space to obtain text projection features;

[0111] Step 302: Project the word frequency statistical features onto the information space to obtain statistical projection features;

[0112] Step 303: Determine the fusion weight value of the statistical projection feature based on the fusion degree, and fuse the statistical projection feature with the text projection feature based on the fusion weight value to obtain the fused feature.

[0113] Based on the extracted text features TE and word frequency statistical features ρ z For example, the fusion of text features and word frequency statistics is achieved based on a pre-defined information space, which is a shared space for both text features and word frequency statistics. When fusing text features and word frequency statistics, the text features TE are first projected onto the pre-defined information space to obtain the text projection features G. TE Then, the word frequency statistical feature ρ z Similarly, projecting these features into the information space shared with the text features yields the statistical projection features G. ρ Taking the projection of word frequency statistical features as an example, the specific projection method can be as shown in the following formula 14:

[0114] G ρ =W ρ (ρ z )+b ρ (14)

[0115] Among them, W ρ It is the projection parameter matrix corresponding to the word frequency statistical features, which can be obtained through pre-training, b ρ It is the bias term of the input word frequency statistical features, which is activated by G. TE To measure the credibility of textual information, an adaptive gating mechanism is employed. Based on the credibility of the textual information, the degree of fusion between textual features and word frequency statistical features is determined. This degree of fusion is then used as the statistical projection feature G obtained after projecting the word frequency statistical features. ρ The fusion weight value is used to fuse the statistical projection features and the text projection features, resulting in the fused feature G. 0 The fusion formula is shown below:

[0116]

[0117] In Equation 15, the function AdaGate is the adaptive gating function in this embodiment, ReLU is the activation function, and α is a hyperparameter used to adjust the confidence threshold. For example, when α = 0, all statistical features are discarded, and when α = 0.5, all statistical features are fused. The Valve function is a filter used to match useful statistical features G. ρ Given low-confidence text, specifically, based on an adaptive gating mechanism, the Valve function adjusts the confidence variable G′ based on the text information. TE To assess the credibility of textual information, statistical projection features G are used. ρ Filtering is performed to determine the statistical projection feature G. ρ The degree of fusion is used to provide necessary word frequency statistical features for text information with low credibility. Furthermore, the credibility variable value G′ of the text information... TE It is a probability value of (0, 1). When it is close to 0 or 1, the text credibility is high; when it is close to 0.5, the text credibility is low. The value of α is determined based on the credibility variable value of the text information. For example, when G′ TE When α = 0 or 1, α = 0.

[0118] Furthermore, in this embodiment, the credibility G′ of the text information is evaluated based on the text projection features obtained after text feature projection. TE Specifically, in this embodiment, a fully connected network is used to project the text features TE onto a preset information space to obtain G. TE The credibility of text features is evaluated by activating the text projection features using activation functions. For example, the activation of text projection features is shown in Equation 16 below:

[0119] G′ TE =Sigmoid(G TE ) = Sigmoid(W TE TE+b TE (16)

[0120] Basically the same as Formula 15, in Formula 16, W TE Let b be the parameter matrix of the text features. TE It is the bias term of the input text features, and Sigmoid is the activation function.

[0121] In step 200, determining the credibility of text information based on text features may include:

[0122] Step 201: Determine the feature information of the text features and obtain the parameter matrix corresponding to the text features;

[0123] Step 202: Based on the parameter matrix and the feature information, activate the text projection features using a preset activation function to obtain the credibility variable value corresponding to the text features, wherein the feature information includes dimension information, content information and length information;

[0124] Step 203: Evaluate the credibility of the text information using the credibility variable value.

[0125] When evaluating the credibility of the text information to be detected based on text features, the text projection feature G obtained by projecting the text features is used. TE First, determine the feature information of the input text, which includes content information, length information, and dimension information, and then obtain the parameter matrix W corresponding to the text features. TE Then, based on the feature information and parameter matrix W of the text features... TE The text projection features G of the text features are obtained by using a preset activation function. TE Activation is performed to obtain the confidence variable value G′ corresponding to the text features. TE , G′ TE It is a probability value between 0 and 1, according to G′ TE The probability value is used to assess the credibility of textual information.

[0126] Furthermore, step 400 may specifically include:

[0127] Step 401: Based on a preset attention mechanism, perform attention fusion on the fused features and the text features to obtain the target features;

[0128] Step 402: Project the target features into a preset label space to classify the text information as rumors;

[0129] Step 403: Based on the rumor classification results of the text information, determine whether the text information contains rumor information.

[0130] When detecting whether text information contains rumors, the system first performs attention fusion on the fused features and text features based on a pre-defined attention mechanism, namely:

[0131] Attention(G o ,TE)=Softmax(G o TE T )TE (17)

[0132] The target features for detection are obtained through the attention fusion mechanism shown in Equation 17. These target features are then projected into a predefined label space to classify the text information as rumors. Based on the rumor classification results, it is determined whether the text information contains rumors and the type of rumors it contains.

[0133] Specifically, the attention mechanism used in rumor detection can be the same as or different from the attention mechanism in text feature extraction networks; no specific limitation is made here. Based on Equation 17, the attention mechanism is used to enhance the text features, i.e., the fused feature G, after feature enhancement through fusion of word frequency statistics. 0 The target features are combined with the extracted original text features (TE), then passed through a fully connected layer and a softmax activation function layer. The resulting vector representation of the target features is projected onto the label space, and a cross-entropy loss function is used as the optimizer for rumor classification, thereby determining whether the text information to be detected contains rumor information. It is important to note that when word frequency statistical features are completely discarded, for example, when α = 0, G... 0 =TE, at this time, G 0 Without incorporating word frequency statistics, the attention fusion mechanism shown in Formula 17 is a self-attention mechanism when classifying rumors.

[0134] In this embodiment, by fusing word frequency statistics features and text features through feature projection, and by introducing an attention mechanism again during rumor classification and detection, key features can be further highlighted and the accuracy of rumor detection can be improved.

[0135] Reference Figure 4 , Figure 4 This is another schematic diagram of the feature extraction network for the rumor detection method provided in this application embodiment. In this embodiment, based on Figure 4 The feature extraction network structure shown in this paper provides a detailed explanation of the rumor detection method provided in this application.

[0136] Figure 4 The feature extraction network structure shown mainly includes a text feature extraction network, a word frequency statistics feature extraction network, and a rumor classification and detection network. When detecting rumors, the original text dataset is first obtained. Regular expressions are then used to filter and clean the original text dataset, removing useless or interfering information to obtain the text information to be detected. The word frequency statistics vector ρ and sentence vector X of the text information are then obtained respectively. Figure 4 In the text, the word frequency statistics vector ρ has a dimension of 5×4, that is, (w1~w5)×(c1~c4), and the sentence vector X=[w1,w2,w3,w4,w5].

[0137] For text feature extraction networks, a network structure is used that combines convolutional neural network (CNN) attention mechanisms with bidirectional long short-term memory (BiLSTM) attention mechanisms, namely CNN_Attention and BiLSTM_Attention. The sentence vector X is taken as input and passed through both the CNN and BiLSTM networks to extract sequence and semantic features. These extracted features are then passed through an attention layer for attention scoring, assigning a corresponding attention weight to each word to obtain the corresponding attention sequence features and attention semantic features. These features are concatenated to obtain the text feature TE represented by a vector. The text feature TE is then projected onto the information space to obtain the corresponding projected text feature G. TE .

[0138] The word frequency feature extraction network includes normalization and a variational autoencoder (VAE). Taking the word frequency statistics vector ρ as input, the input vector is first normalized. Then, the VAE is used to calculate the feature values ​​of the normalized vector ρ, which include the mean μ and variance σ. 2 Based on the calculated feature values, a variational posterior conforming to a standard normal distribution is constructed. This variational posterior is then sampled to obtain the corresponding latent variable z. Based on the latent variable z corresponding to each word frequency statistical vector ρ, the word frequency statistical features ρ of the text information are generated. z And the word frequency statistical feature ρ z Projecting these features onto an information space shared with the text features yields the corresponding statistical projection features G. ρ .

[0139] Features G obtained from a text feature extraction network TE And the feature G obtained based on the word frequency statistical feature extraction network. ρ An adaptive fusion mechanism for word frequency statistical features is established using a hyperparameter α. Specifically, the credibility of text information is evaluated using text features TE, yielding the corresponding credibility variable value G′. TE The confidence variable value is a probability value of (0, 1), and the value of α is based on G′. TE The value of is determined to adjust the confidence threshold, thereby determining whether word frequency statistics features need to be fused, and the degree of fusion of word frequency statistics features. According to G... TE G ρ G′ TE And α to construct the word frequency statistical feature ρ z The adaptive fusion mechanism uses an adaptive gating component, which acts as a filter, to control the fusion degree of word frequency statistical features based on the credibility of the text information.

[0140] Finally, based on the fused text enhancement features G 0 The extracted text features (TE) are fused using an attention mechanism to further highlight key features, resulting in a vector representation of the target features. This target feature is then projected onto a predefined label space using an activation function to classify the text information as rumors. Based on the classification results, it is determined whether the text to be detected contains rumors.

[0141] In this embodiment, multi-dimensional text features are extracted through an attention mechanism, and the credibility of text information is evaluated based on the extracted text features. This establishes an adaptive gating mechanism based on text credibility, enabling adaptive fusion of word frequency statistical features. This provides necessary statistical features for low-credibility text information and improves the detection accuracy of rumors.

[0142] Furthermore, when classifying and detecting rumors, the attention mechanism is used again to highlight the key features of the enhanced text features, thereby further improving the accuracy of rumor detection.

[0143] It should be noted that the activation functions, parameters, and specific parameter values ​​involved in the rumor detection process in the various embodiments provided in this application are only for illustrative purposes and do not constitute a limitation. In practical applications, the activation functions, parameters, and parameter values ​​can be adjusted and selected as needed, and will not be elaborated here.

[0144] The rumor detection device provided in the embodiments of this application is described below. The rumor detection device described below can be referred to in correspondence with the rumor detection method described above.

[0145] Reference Figure 5 The rumor detection device provided in this application includes:

[0146] The feature extraction module 10 is used to acquire the text information to be detected and extract the text features and word frequency statistical features of the text information;

[0147] An adaptive gating module 20 is used to determine the credibility of the text information based on the text features, and to determine the fusion degree of the text features and the word frequency statistical features based on the credibility.

[0148] Feature fusion module 30 is used to fuse the text features and the word frequency statistical features based on the fusion degree to obtain fused features;

[0149] The rumor detection module 40 is used to detect whether the text information contains rumor information based on the fusion features.

[0150] In one embodiment, the feature extraction module 10 is further configured to:

[0151] Obtain the word frequency statistical vector of each word in the text information;

[0152] Encoding is performed on each of the aforementioned word frequency statistical vectors to obtain the word frequency statistical features of the text information, wherein the word frequency statistical features contain global information of each of the aforementioned word frequency statistical vectors;

[0153] Extract the sentence vectors from the text information, where the sentence vectors contain global semantic information of the text information;

[0154] Extract the sequence features and semantic features of the text information based on the sentence vectors;

[0155] Based on a preset attention mechanism, attention scores are applied to the sequence features and the semantic features respectively to obtain the corresponding attention sequence features and attention semantic features;

[0156] The attention sequence features and the attention semantic features are concatenated to obtain the text features of the text information.

[0157] In one embodiment, the feature extraction module 10 is further configured to:

[0158] Calculate the feature values ​​of each of the word frequency statistical vectors, wherein the feature values ​​include at least the mean and variance of the word frequency statistical vectors;

[0159] Based on the feature values, a variational posterior of the text information is constructed, and the variational posterior is sampled to obtain the latent variables of the text information.

[0160] The word frequency statistical features of the text information are generated based on the latent variables.

[0161] In one embodiment, the feature fusion module 30 is further configured to:

[0162] The text features are projected onto a preset information space to obtain text projection features;

[0163] The word frequency statistical features are projected onto the information space to obtain statistical projection features;

[0164] The fusion weight value of the statistical projection feature is determined based on the fusion degree. Based on the fusion weight value, the statistical projection feature and the text projection feature are fused to obtain the fused feature.

[0165] In one embodiment, the adaptive gating module 20 is further configured to:

[0166] Determine the feature information of the text features and obtain the parameter matrix corresponding to the text features;

[0167] Based on the parameter matrix and the feature information, the text projection features are activated using a preset activation function to obtain the credibility variable value corresponding to the text features. The feature information includes dimension information, content information, and length information.

[0168] The credibility of the text information is evaluated using the credibility variable values.

[0169] In one embodiment, the rumor detection module 40 is further configured to:

[0170] Based on a preset attention mechanism, attention fusion is performed on the fused features and the text features to obtain the target features;

[0171] The target features are projected into a preset label space to classify the text information as rumors.

[0172] Based on the rumor classification results of the text information, determine whether the text information contains rumor information.

[0173] Figure 6 Example: A schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown, the electronic device may include: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other via the communication bus 640. The processor 610 can call a computer program in the memory 630 to execute the steps of the rumor detection method, such as including:

[0174] Obtain the text information to be detected, and extract the text features and word frequency statistical features of the text information;

[0175] The credibility of the text information is determined based on the text features, and the degree of fusion between the text features and the word frequency statistics features is determined based on the credibility.

[0176] Based on the fusion degree, the text features and the word frequency statistical features are fused to obtain the fused features;

[0177] The text information is detected based on the fusion features to determine whether it contains rumors.

[0178] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0179] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the steps of the rumor detection method provided in the above embodiments, such as including:

[0180] Obtain the text information to be detected, and extract the text features and word frequency statistical features of the text information;

[0181] The credibility of the text information is determined based on the text features, and the degree of fusion between the text features and the word frequency statistics features is determined based on the credibility.

[0182] Based on the fusion degree, the text features and the word frequency statistical features are fused to obtain the fused features;

[0183] The text information is detected based on the fusion features to determine whether it contains rumors.

[0184] On the other hand, embodiments of this application also provide a processor-readable storage medium storing a computer program for causing a processor to perform the steps of the methods provided in the above embodiments, such as including:

[0185] Obtain the text information to be detected, and extract the text features and word frequency statistical features of the text information;

[0186] The credibility of the text information is determined based on the text features, and the degree of fusion between the text features and the word frequency statistics features is determined based on the credibility.

[0187] Based on the fusion degree, the text features and the word frequency statistical features are fused to obtain the fused features;

[0188] The text information is detected based on the fusion features to determine whether it contains rumors.

[0189] The processor-readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic memory (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO)), optical memory (e.g., CD, DVD, BD, HVD), and semiconductor memory (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)).

[0190] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0191] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A rumor detection method characterized by, The rumor detection method comprises the following steps: Obtain the text information to be detected, extract the text features and word frequency statistical features of the text information; Determine the credibility of the text information according to the text features, and determine the fusion degree of the text features and the word frequency statistical features according to the credibility; Fuse the text features and the word frequency statistical features based on the fusion degree to obtain fusion features; Detect whether the text information contains rumor information according to the fusion features; The step of fusing the text features and the word frequency statistical features based on the fusion degree to obtain fusion features comprises: Project the text features into a preset information space to obtain text projection features; Project the word frequency statistical features into the information space to obtain statistical projection features; The projection mode is as follows: ； wherein, is a statistical projection feature, is a term frequency statistical feature, is a projection parameter matrix corresponding to the term frequency statistical feature, is a bias term for the input term frequency statistical feature; Determine the fusion weight value of the statistical projection features according to the fusion degree, fuse the statistical projection features and the text projection features based on the fusion weight value to obtain fusion features; The fusion formula is as follows: ； wherein, is the fusion feature, function AdaGate is an adaptive gating function, ReLU is an activation function, function is a filter, is a hyper-parameter, used to adjust the threshold of the credibility, is the text projection feature, is the credibility variable value of the text information. 2.The rumor detection method of claim 1, wherein, The step of determining the credibility of the text information according to the text features comprises: Determine the feature information of the text features, obtain the parameter matrix corresponding to the text features; Based on the parameter matrix and the feature information, activate the text projection features by using a preset activation function to obtain a credibility variable value corresponding to the text features, wherein the feature information includes dimension information, content information and length information; Evaluate the credibility of the text information through the credibility variable value. 3.The rumor detection method of claim 1, wherein, The step of detecting whether the text information contains rumor information according to the fusion features comprises: Fuse the fusion features and the text features based on a preset attention mechanism to obtain target features; Project the target features into a preset label space to classify the text information as a rumor; Determine whether the text information contains rumor information according to the rumor classification result of the text information. 4.The rumor detection method of claim 1, wherein, The step of extracting the text features and word frequency statistical features of the text information comprises: Obtain the word frequency statistical vector of each word in the text information; Encode each word frequency statistical vector to obtain the word frequency statistical features of the text information, which contain global information of each word frequency statistical vector; Extract the sentence vector of the text information, which contains global semantic information of the text information; Extract the sequence features and semantic features of the text information according to the sentence vector respectively; Score the sequence features and semantic features based on a preset attention mechanism to obtain corresponding attention sequence features and attention semantic features; Concatenate the attention sequence features and the attention semantic features to obtain the text features of the text information. 5.The rumor detection method of claim 4, wherein, The step of encoding each word frequency statistical vector to obtain the word frequency statistical features of the text information comprises: Calculate the eigenvalues of each word frequency statistical vector, wherein the eigenvalues at least include the mean and variance of the word frequency statistical vector; According to the characteristic value, a variational posterior of the text information is constructed, and sampling processing is performed on the variational posterior to obtain a latent variable of the text information; According to the latent variable, a word frequency statistical feature of the text information is generated. 6.The rumor detection method of claim 1, wherein, The step of obtaining the text information to be detected comprises: An original text data set is obtained, and text information in the original text data set is filtered and cleaned by using a regular expression to obtain the text information to be detected.

7. A rumor detection apparatus characterized by comprising: The rumor detection device comprises: A feature extraction module is configured to obtain text information to be detected, and extract a text feature and a word frequency statistical feature of the text information; An adaptive gating module is configured to determine a credibility of the text information according to the text feature, and determine a fusion degree of the text feature and the word frequency statistical feature according to the credibility; A feature fusion module is configured to fuse the text feature and the word frequency statistical feature based on the fusion degree to obtain a fusion feature; A rumor detection module is configured to detect whether rumor information is contained in the text information according to the fusion feature; The step of fusing the text feature and the word frequency statistical feature based on the fusion degree to obtain a fusion feature comprises: The text feature is projected into a preset information space to obtain a text projection feature; The word frequency statistical feature is projected into the information space to obtain a statistical projection feature; The projection manner is as follows: ； wherein, is a statistical projection feature, is a term frequency statistical feature, is a projection parameter matrix corresponding to the term frequency statistical feature, is a bias term for the input term frequency statistical feature; According to the fusion degree, a fusion weight value of the statistical projection feature is determined, and the statistical projection feature and the text projection feature are fused based on the fusion weight value to obtain a fusion feature; The fusion formula is as follows: ； wherein, is the fusion feature, function AdaGate is an adaptive gating function, ReLU is an activation function, function is a filter, is a hyper-parameter used to adjust the threshold of the credibility, is the text projection feature, is the credibility variable value of the text information.

8. An electronic device comprising a processor and a memory having a computer program stored therein, characterized in that The processor executes the computer program to implement the steps of the rumor detection method in any one of claims 1 to 6. 9.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the rumor detection method in any one of claims 1 to 6.

Images