A social network information credibility evaluation method based on D-S evidence theory

Abstract

The application discloses a social network information credibility evaluation method based on D-S evidence theory, which comprises the following steps: 1, feature extraction, using BERT pre-training model, ResNet50 neural network, TF-IDF algorithm and cosine similarity method to extract text features, image-text consistency features and title-text consistency features; 2, evidence generation, using the basic probability assignment method and related formula to convert the original evidence into the corresponding BPA; 3, evidence fusion, calculating the evidence clarity, evidence credibility and evidence weight, and combining the evidence synthesis rule to realize the evidence fusion. The method firstly respectively utilizes the text features, image-text consistency features and title-text consistency features to evaluate the information credibility; then converts the credibility evaluation results based on the three single features into evidence, and uses the improved D-S evidence theory fusion rule to make fusion and decision, and finally outputs the final result of the comprehensive credibility evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of social network information credibility evaluation methods, in particular to a social network information credibility evaluation method based on D-S evidence theory. BACKGROUND

[0002] The social network information credibility evaluation method based on D-S evidence theory refers to converting the features of information content (text, image) in a social network into evidence, and then using D-S evidence theory to fuse the evidence to obtain the probability of determining the authenticity of information, i.e., information credibility.

[0003] At present, the existing social network information credibility evaluation method is based on the combination of "text features + image-text consistency features", and is realized by combining a deep learning feature fusion method. However, the feature combination used in the current information credibility evaluation method has low effectiveness, and fails to fully exploit the similarity relationship between article titles and article content. Moreover, the existing deep learning-based feature fusion method relies heavily on prior knowledge, and has low evaluation accuracy on specific small sample datasets. Therefore, the accuracy of the social network information credibility evaluation method has room for improvement. SUMMARY

[0004] In view of the problem that the existing deep learning-based feature fusion method relies heavily on prior knowledge and has low evaluation accuracy on specific small sample datasets, the information fusion method based on D-S evidence theory is introduced into information credibility evaluation, and a social network information credibility evaluation method based on D-S evidence theory is proposed by combining the new feature combination. The method first uses text features, image-text consistency features, and title-text consistency features to evaluate information credibility, respectively. Then, the credibility evaluation results based on single features are converted into evidence, and an improved D-S evidence theory fusion rule is used for fusion and decision-making. Finally, the final result of comprehensive credibility evaluation is output.

[0005] In order to achieve the above-mentioned application purposes and solve the technical problems, the technical solutions adopted are as follows:

[0006] A social network information credibility evaluation method based on D-S evidence theory, comprising the following steps:

[0007] Step 1: Feature extraction, using BERT pre-training model, ResNet50 neural network, TF-IDF algorithm, and cosine similarity method to extract text features, image-text consistency features, and title-text consistency features;

[0008] Step 2: Evidence generation, using basic probability assignment method and related formulas to convert original evidence into corresponding BPA;

[0009] Step 3: Evidence fusion, calculate evidence clarity, evidence credibility, and evidence weight, and combine evidence synthesis rules to realize evidence fusion.

[0010] Further, step 1 includes the following:

[0011] Step 11: Text feature extraction of the main text;

[0012] First, the original data is preprocessed, where the number of characters is set to 512, and characters less than 512 are supplemented, and characters exceeding 512 are truncated, and then the BERT pre-training model is used to extract text features:

[0013]

[0014] Where t i represents the i-th input sentence, is a 768-dimensional vector representing the text feature vector after BERT embedding, and the extraction of the text features of the main text is completed;

[0015] Step 12: Image-text consistency feature extraction;

[0016] The output of the convolutional neural network is used as the image bottom layer feature, and the input image is encoded using the ResNet50 pre-training model. In the pre-trained ResNet50 model, a 2048-dimensional fully connected layer is used to encode the image features, and a 2048-dimensional vector is used to represent the image:

[0017] h v = ResNet50(v) (2)

[0018] Where v represents the original image input into ResNet50, and h v represents the visual feature vector extracted after ResNet50 network;

[0019] BiGRU is used to further extract the features extracted by BERT, i.e. BiGRU is used to further extract the temporal properties in the text features, so as to convert the text features into a text feature sequence:

[0020]

[0021] Where, represents the text feature vector after BERT embedding, represents the text feature sequence extracted by BiGRU;

[0022] The method using similarity measure calculates the similarity of text and image information, for using the features of two kinds of information in common representation space, using a fully connected layer in the last layer of each feature extraction network, and making the two networks share the weights of the last layer, to obtain semantic representation s of text and image t and s v , and using cosine similarity to measure the similarity between text and image:

[0023]

[0024] wherein s t and s v represent the text feature sequence and the image semantic sequence respectively, s tv represents the similarity between the two, s tv , the value range of s tv is [-1, 1], and the greater the value of s h , the higher the similarity between the two, thus, the extraction of the text-image consistency feature is completed;

[0025] Step 13: title-text consistency feature extraction;

[0026] The TF-IDF algorithm is used to calculate the text similarity of title t h and text t m , so as to extract the title-text consistency feature. First, the jieba word segmentation is used to respectively perform word segmentation processing on t h and t m in the data, find out the respective keywords of the two, and convert them into corresponding vectors t h1 , t h2 …t hi and t m1 , t m2 …t mj (i, j are positive integers), then all the keywords are combined into a set k:

[0027]

[0028] wherein k is not an empty set;

[0029] The TF-IDF values d h , d m , f h and f m of the title and the text relative to the keywords in the set k are respectively calculated;

[0030] The TF-IDF vectors of the title and the text are respectively calculated as and The calculation formula is as follows:

[0031]

[0032]

[0033] and cosine similarity of The calculation formula is as follows:

[0034]

[0035] wherein, and respectively represent the TF-IDF vectors of the title and the text, represent the similarity between the two, that is, the title-text consistency feature, and thus the extraction of the image-text consistency feature is completed.

[0036] Further, step 2 includes the following contents:

[0037] Step 21: Obtain the original evidence;

[0038] In the feature selection task, after extracting the three features respectively, the corresponding feature vectors are obtained: the text feature vector the image-text consistency feature vector s tv and the title-text consistency feature vector Then, the three feature vectors are used to evaluate the information credibility respectively, based on the credibility evaluation method of the three features;

[0039] The credibility evaluation method based on the text feature is to send into the classifier to obtain the classification result, and then convert the result into a percentage R a , R a is the result of using the text feature to evaluate the credibility;

[0040] The credibility evaluation method based on the image-text consistency feature is to send s tv through the Sigmoid function to map it between [0, 1] to get a percentage R b and as the final result of the evaluation method based on the image-text consistency feature;

[0041] The credibility evaluation method based on the title-text consistency feature is to send through the Sigmoid function to map it between [0, 1] to get a percentage R c and as the final result of the evaluation method based on the title-text feature;

[0042] Step 22: Generate the original BPA;

[0043] Considering that D-S evidence theory is an ideal method to deal with uncertainty, the basic probability assignment (BPA) function is calculated to transform the evaluation results into probability values in the recognition framework when using evidence theory to fuse the evaluation results of different features. In the problem of false information detection, the result of detection is true (Real, R) or false (False, F), and the recognition framework is Φ = {R, F}, where R and F are mutually exclusive, and Φ is a finite and complete set. The BPA function calculation formula under different attributes is as follows:

[0044] λ(x) = 1 - 2|x - 0.5|, x ∈ [0, 1] (9)

[0045]

[0046]

[0047] m a,φ =1-m a,R -m a,F (12)

[0048]

[0049]

[0050] m b,φ =1-m b,R -m b,F (15)

[0051]

[0052]

[0053] m c,φ =1-m c,R -m c,F (18)

[0054] where λ(x) is a function for calculating the fuzzy rate of probability value x, and λ(x) ∈ [0, 1]; when the value of x is closer to 0.5, it means that it is difficult to distinguish whether the news is false information, and the corresponding fuzzy rate will be higher; m a,R represents the transformation of R a into the BPA value of true result, m a,F represents the transformation of R a into the BPA value of false result, m a,φ represents the transformation of R a into the BPA value of uncertain result; similarly, m b,R represents the transformation of R b into the BPA value of true result, m b,F represents the transformation of R bTransformed into BPA values of false results, m b,φ R represents the transformation of R b Transformed into BPA values of uncertain results, m c,R R represents the transformation of R c Transformed into BPA values of true results, m c,F R represents the transformation of R c Transformed into BPA values of false results, m c,φ R represents the transformation of R c Transformed into BPA values of uncertain results.

[0055] Further, step 3 includes the following contents:

[0056] Step 31: evidence credibility calculation;

[0057] The evidence credibility is measured using the similarity of the evidence, first calculating the evidence similarity as the initial credibility, and then correcting the initial credibility using the evidence sorting factor, and the similarity calculation method is as follows:

[0058]

[0059] Wherein, m i And m j Respectively represent the corresponding evidence vector m a , m b , m c Any two of them, m i -m j Is the vector difference between evidence m i And m j , indicating the distance between the input evidence, D is a matrix based on the focus element first norm, and D = |A∩B| / |A∪B|;

[0060] The credibility of the evidence is defined as follows:

[0061]

[0062] Wherein, Cred(m i ) represents the credibility of evidence m i , ms is the set of evidence participating in fusion, all evidence in ms must be in the same identification framework as m, and the similarity between the evidence is used to reflect the overall credibility of the evidence, the higher the sum of the similarity between the evidence, the greater the credibility;

[0063] The sorting factor is used to correct the credibility formula as shown in formula 21:

[0064]

[0065] Wherein, SortFactor(mi ) represents the ranking factor corresponding to the evidence mi;

[0066] Step 32: Evidence clarity calculation;

[0067] The clarity of evidence is represented by calculating the distance between each evidence and the uniform distribution, and the result is normalized. The Wasserstein distance between two evidences Ei and Ej is as follows:

[0068]

[0069] where W(E i ,E j ) represents the difference in probability distribution between evidence Ei and Ej, that is, the distance between the two evidences. The recognition framework contains n elements, inf represents the maximum lower bound of the expression, and the value range of γ is [0, (2n-1)2]. The clarity calculation formula of evidence m i is as follows:

[0070] Clar(m i ) = W(m i , U) (23)

[0071]

[0072] where m i represents the evidence to be calculated, U represents the uniform distribution, |m| represents the number of focal elements in any evidence body, and Clar(m i ) represents the clarity of evidence m i ;

[0073] Step 33: Evidence weight determination;

[0074] The evidence weight is determined by the evidence credibility and clarity. The evidence weight is calculated using the weight determination method, and the calculation formula is as follows:

[0075]

[0076] where N is the number of evidences participating in fusion, and w(m i ) is the weight of evidence m i ;

[0077] Step 34: Evidence synthesis according to evidence synthesis rule;

[0078] The normalized evidence MAE is defined according to the weight of the evidence. MAE is the evidence corrected according to the weight, and the definition of MAE is as follows:

[0079] Suppose m1, m2…m N are N evidences under the same recognition framework, and w(mi is the weight of the evidence m i The MAE calculation formula is as follows:

[0080]

[0081] The final fusion result R is obtained by using the basic formula to fuse n-1 times abc .

[0082] Compared with the prior art, the present application has the following advantages and positive effects:

[0083] The social network information credibility evaluation method based on the D-S evidence theory effectively improves the accuracy of the social network information credibility evaluation result, solves the problem of low feature combination effectiveness in the current information credibility evaluation method, and solves the problem that the existing feature fusion method based on deep learning relies on prior knowledge and has low evaluation accuracy on a specific small sample data set. When measuring the credibility of social network information, the method has high measurement accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0084] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor. In the drawings:

[0085] Figure 1 is the data fusion method framework diagram based on the D-S evidence theory of the present application;

[0086] Figure 2 is the text feature extraction schematic diagram in the present application;

[0087] Figure 3 is the text feature extraction schematic diagram in the present application;

[0088] Figure 4 is the title-text consistency feature extraction schematic diagram in the present application. DETAILED DESCRIPTION

[0089] The technical solutions of the present application will be described in detail below with reference to the drawings. Obviously, the described embodiments are part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0090] The embodiment discloses a social network information credibility evaluation method based on D-S evidence theory, and comprises the following steps:

[0091] Step 1: feature extraction, using a BERT pre-training model, a ResNet50 neural network, a TF-IDF algorithm and a cosine similarity method to extract text features, image-text consistency features and title-text consistency features;

[0092] Step 2: evidence generation, using a basic probability assignment method and related formulas to convert original evidence into corresponding BPA;

[0093] Step 3: evidence fusion, calculating evidence clarity, evidence credibility and evidence weight, and combining evidence synthesis rules to realize evidence fusion.

[0094] Further, step 1 comprises the following contents:

[0095] Step 11: text feature extraction;

[0096] First, the original data is preprocessed, wherein the number of characters is set to 512, characters less than 512 are supplemented, and characters exceeding 512 are truncated, and then a BERT pre-training model is used to extract text features, as shown in formula 1:

[0097]

[0098] Wherein, t i represents the i-th input sentence, is a 768-dimensional vector, which represents the text feature vector after BERT embedding, and the extraction of the text feature is completed. The overall framework of the text feature extraction method is shown in Figure 2 In the text feature extraction method, first, the text content is preprocessed, the original text information is vectorized, and then the pre-trained BERT model is used to complete feature extraction.

[0099] Step 12: image-text consistency feature extraction;

[0100] The output of the convolutional neural network is used as the image bottom layer feature, and the input image is encoded by combining the ResNet50 pre-training model. In the pre-trained ResNet50 model, a 2048-dimensional fully connected layer is used to encode the image features, and the image is represented by a 2048-dimensional vector, as shown in formula 2:

[0101] h v = ResNet50(v) (2)

[0102] Wherein, v represents the original image input into the ResNet50, h vdenotes the visual feature vector extracted by the ResNet50 network;

[0103] To more thoroughly capture global feature information and better fuse image information, the BiGRU is used to extract the feature extracted by the BERT again, that is, the BiGRU is used to further extract the temporal attribute in the text feature, so as to convert the text feature into a text feature sequence, as shown in formula 3:

[0104]

[0105] wherein, denotes the text feature vector after BERT embedding, denotes the text feature sequence after BiGRU extraction.

[0106] The authenticity of the overall information is evaluated by calculating the correlation between the text feature and the corresponding visual feature. In the embodiment, the similarity measurement method is used to calculate the similarity of the text and image information. In order to use the features of the two kinds of information in a common representation space, a fully connected layer is used in the last layer of each feature extraction network, and the weights of the last layer are shared by the two networks, to obtain the semantic representation s t and s v of the text and image, and the cosine similarity is used to measure the similarity between the text and image, as shown in formula 4:

[0107]

[0108] wherein, s t and s v denote the text feature sequence and the image semantic sequence respectively, s tv denotes the similarity between the two, s tv has a value range of [-1, 1], s tv The greater the value, the higher the similarity between the two. Thus, the extraction of the text-image consistency feature is completed; the overall framework of the text-image consistency feature extraction method is shown in the following formula 5: Figure 3 In the text-image consistency feature extraction method, first, the BERT and the ResNet are used to vectorize the text and picture original information, then the weights of the vectors are shared, and finally the similarity is calculated, that is, the text-image consistency feature.

[0109] Step 13: title-text consistency feature extraction;

[0110] The TF-IDF algorithm is used to calculate the text similarity of the title t h and the text t m , so as to extract the title-text consistency feature. First, the jieba word segmentation is used to respectively segment t h and t mThe content is segmented, and the respective keywords of both are found and converted into corresponding vectors t h1 ,t h2 …t hi and t m1 ,t m2 …t mj (i, j are positive integers), and then all keywords form a set k, as shown in equation 5:

[0111]

[0112] wherein k is not an empty set;

[0113] The TF-IDF values d h , d m , f h and f m of the title and the text with respect to the keywords in set k are calculated respectively, as shown in Table 1;

[0114]

[0115] Table 1: Keyword TF-IDF value table

[0116] The TF-IDF vectors of the title and the text are calculated respectively and The calculation process is shown in equations 6 and 7;

[0117]

[0118]

[0119] and the cosine similarity of t is calculated as shown in equation 8;

[0120]

[0121] wherein t and t represent the TF-IDF vectors of the title and the text respectively, represents the similarity between the two, i.e., the title-text consistency feature, and thus the extraction of the title-text consistency feature is completed.

[0122] The overall framework of the extraction method of the title-text consistency feature is shown in Figure 4The title-text consistency feature extraction process mainly uses jieba segmentation technology and TF-IDF algorithm. First, the jieba segmentation is used to segment the text of the title and the text of the body to obtain the corresponding segmentation vectors, and the public keyword vector set of the text of the title and the text of the body is determined; then the TF-IDF value of the text of the title and the text of the body relative to the keyword is calculated and converted into a TF-IDF vector, and the similarity between the two is calculated, that is, the title-text consistency feature.

[0123] Further, step 2 includes the following contents:

[0124] Step 21: obtaining original evidence;

[0125] In the feature selection task, after extracting the three features respectively, the corresponding feature vectors are obtained: the text feature vector of the body The figure-text consistency feature vector s tv And the title-text consistency feature vector Then, the three feature vectors are used to evaluate the information credibility respectively, and the credibility evaluation methods based on the three features are:

[0126] The credibility evaluation method based on the text feature of the body is to input Into the classifier to obtain the classification result, and then convert the result into a percentage R a , R a is the result of using the text feature to evaluate the credibility;

[0127] The credibility evaluation method based on the figure-text consistency feature is to input s tv Through the Sigmoid function mapping between [0,1] to obtain the percentage R b and as the final result of the evaluation method based on the figure-text consistency feature;

[0128] The credibility evaluation method based on the title-text consistency feature is to input Through the Sigmoid function mapping between [0,1] to obtain the percentage R c and as the final result of the evaluation method based on the title-text feature;

[0129] Step 22: generating original BPA;

[0130] Considering that D-S evidence theory is an ideal method to deal with uncertainty, the basic probability assignment (BPA) function is calculated to transform the evaluation results into probability values in the recognition framework when using evidence theory to fuse the evaluation results of different features. In the problem of false information detection, the result of detection is true (Real, R) or false (False, F), and the recognition framework is Φ = {R, F}, where R and F are mutually exclusive, and Φ is a finite complete set. The BPA function under different attributes is calculated according to the following formulas 9-18:

[0131] λ(x) = 1 - 2|x - 0.5|, x ∈ [0, 1] (9)

[0132]

[0133]

[0134] m a,φ = 1 - m a,R -m a,F (12)

[0135]

[0136]

[0137] m b,φ = 1 - m b,R -m b,F (15)

[0138]

[0139]

[0140] m c,φ = 1 - m c,R -m c,F (18)

[0141] where λ(x) is a function for calculating the fuzzy rate of probability value x, and λ(x) ∈ [0, 1]; when the value of x is closer to 0.5, it means that it is difficult to distinguish whether the news is false information, and the corresponding fuzzy rate will be higher; m a,R represents the transformation of R a into the BPA value of true result, m a,F represents the transformation of R a into the BPA value of false result, m a,φ represents the transformation of R a into the BPA value of uncertain result; similarly, m b,R represents the transformation of R b into the BPA value of true result, m b,F represents the transformation of Rb BPA value converted to false result, m b,φ R represents the conversion of R b BPA value converted to uncertain result c,R R represents the conversion of R c BPA value converted to true result, m c,F R represents the conversion of R c BPA value converted to false result, m c,φ R represents the conversion of R c BPA value converted to uncertain result.

[0142] Further, step 3 includes the following:

[0143] Step 31: evidence credibility calculation;

[0144] The evidence credibility is measured using the similarity of the evidence, first calculating the evidence similarity as the initial credibility, and then correcting the initial credibility using the evidence ranking factor, the similarity calculation method is shown in the following formula 19:

[0145]

[0146] Wherein, m i and m j respectively represent the corresponding evidence vector m a , m b , m c corresponding to any two of them, m i -m j is the vector difference between evidence m i and m j , indicating the distance between the input evidence, D is a matrix based on the focus element first norm, and D = |A∩B| / |A∪B|; 19

[0147] The credibility of the evidence is defined as follows 20:

[0148]

[0149] Wherein, Cred(m i ) represents the credibility of evidence m i , ms is the set of evidence participating in the fusion, all evidence in ms must be in the same recognition framework as m, the similarity between evidence is used to reflect the overall credibility of the evidence, the higher the sum of the similarity between the evidence, the greater the credibility;

[0150] The credibility formula is corrected using the ranking factor as shown in formula 21:

[0151]

[0152] where SortFactor(m i ) represents the sorting factor corresponding to the evidence m i ;

[0153] Step 32: Evidence clarity calculation;

[0154] The clarity of the evidence is calculated based on the Wasserstein distance formula. The idea of Wasserstein distance comes from the optimal transport theory. The clarity of the evidence is represented by calculating the distance between each evidence and the uniform distribution (representing extremely vague evidence). The result is normalized, and the Wasserstein distance between two evidences Ei and Ej is shown in equation 22:

[0155]

[0156] where W(E i , E j ) represents the difference in probability distribution between the evidence Ei and Ej, i.e., the distance between the two evidences. The recognition framework contains n elements, inf represents the maximum lower bound of the expression, and γ is in the range of [0, (2n-1)2]. The clarity calculation formula of the evidence m i is shown in equations 23 and 24:

[0157] Clar(m i ) = W(m i , U) (23)

[0158]

[0159] where m i represents the evidence to be calculated, U represents the uniform distribution, |m| represents the number of focal elements in any evidence body, and Clar(m i ) represents the clarity of the evidence m i ;

[0160] Step 33: Evidence weight determination;

[0161] The evidence weight is determined by the evidence credibility and clarity. The evidence weight is calculated using the weight determination method, and the calculation formula is shown in equation 25:

[0162]

[0163] where N is the number of evidences participating in fusion, and w(m i ) is the weight of the evidence m i ;

[0164] Step 34: Evidence synthesis according to evidence synthesis rules;

[0165] According to the weight of the evidence, the normalized evidence (Modified Average Evidence, MAE) is defined as the evidence modified according to the weight, and the definition of MAE is as follows:

[0166] Suppose m1, m2…m N are N evidences under the same recognition framework, and w(m i ) is the weight of the evidence m i , and the MAE calculation formula is as shown in the following formula 26:

[0167]

[0168] Using the basic formula, the MAE is fused n-1 times to obtain the final fusion result R abc .

[0169] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for evaluating the credibility of social network information based on DS evidence theory, characterized in that, Includes the following steps: Step 1: Feature extraction. Use the BERT pre-trained model, ResNet50 neural network, TF-IDF algorithm and cosine similarity method to extract text features, image-text consistency features and title-text consistency features. Step 2: Evidence generation, using basic probability allocation methods and related formulas to convert the original evidence into the corresponding BPA; Step 3: Evidence fusion. Calculate the clarity, credibility, and weight of evidence, and combine them with the rules of evidence synthesis to achieve evidence fusion. Step 3 includes the following: Step 31: Calculation of the credibility of evidence; Evidence similarity is used to measure the credibility of evidence. First, the evidence similarity is calculated as the initial credibility. Then, the initial credibility is corrected using an evidence ranking factor. The similarity calculation method is as follows: Where, m i and m j These represent the evidence vectors m corresponding to the results of credibility evaluation using different single features. a m b m c For any two vectors in the vector, m i -m j It is evidence m i and m j The vector difference between them represents the distance between the input evidences. D is a matrix based on the first norm of the focal element, and D = |A∩B| / |A∪B|. The credibility of evidence is defined as follows: Among them, Cred(m i ) indicates evidence m i The credibility of evidence is determined by the following: ms is the set of evidence involved in the fusion. All evidence in ms must be within the same identification framework as m. The similarity between evidence is used to reflect the overall credibility of the evidence. The higher the sum of the similarities between evidence, the greater its credibility. The confidence formula for adjusting the ranking factor is shown in equation 21: Among them, SortFactor(m i ) represents the ranking factor corresponding to evidence mi; Step 32: Calculation of the clarity of evidence; The clarity of evidence is represented by calculating the distance between each piece of evidence and the uniform distribution, and the results are normalized. The Wasserstein distance between two pieces of evidence Ei and Ej is as follows: Among them, W(E) i E j ) represents the difference in probability distribution between evidence Ei and Ej, i.e., the distance between the two pieces of evidence. The identification frame contains n elements, inf represents taking the maximum lower bound of the expression, and the range of γ is [0, (2n-1)²]. Evidence m i The formula for calculating sharpness is as follows: Clar(m i )=W(m i ,U) (23) Where, m i Let U represent the evidence to be calculated, U represent a uniform distribution, |m| represent the number of focal elements within any piece of evidence, and Clar(m) represent the number of focal elements within any piece of evidence. i ) indicates evidence m i Clarity; Step 33: Determining the weight of evidence; The weight of evidence is determined by both the credibility and clarity of the evidence. The weight of evidence is calculated using the following formula: Where N is the number of pieces of evidence involved in the fusion, w(m i ) is evidence m i The weights; Step 34: Combine evidence according to the rules of evidence combination; The normalized evidence evaluation (MAE) is defined according to the weight of evidence. MAE is the evidence adjusted according to the weights. The definition of MAE is as follows: Assume m1, m2…m N There are N pieces of evidence that are within the same identification framework, and w(m) i ) is evidence m i If the weights are determined, then the MAE calculation formula is as follows: The MAE is fused n-1 times using the basic formula to obtain the final fusion result R. abc .

2. The social network information credibility evaluation method based on DS evidence theory according to claim 1, characterized in that, Step 1 includes the following: Step 11: Extract features from the main text; First, the raw data undergoes data preprocessing, where the character count is set to 512. Characters below 512 are padded, and those exceeding 512 are truncated. Then, a BERT pre-trained model is used to extract text features. Among them, t i This represents the i-th sentence in the input. It is a 768-dimensional vector representing the text feature vector after BERT embedding, which completes the extraction of text features; Step 12: Extraction of image-text consistency features; The output of the convolutional neural network is used as the low-level feature of the image, and combined with a ResNet50 pre-trained model to encode the input image. In the pre-trained ResNet50 model, a 2048-dimensional fully connected layer is used to encode the image features, and the image is represented by a 2048-dimensional vector. h v =ResNet50(v) (2) Where v represents the original image input to ResNet50, h v This represents the visual feature vector extracted by the ResNet50 network; BiGRU is used to further extract the features extracted by BERT, that is, BiGRU is used to further extract the temporal attributes in the text features, in order to transform the text features into a text feature sequence: in, This represents the text feature vector after BERT embedding. This represents the text feature sequence extracted by BiGRU; A similarity metric is used to calculate the similarity between text and image information. To utilize the features of both types of information in a common representation space, a fully connected layer is added to the last layer of each feature extraction network, and the weights of the last layer are shared between the two networks, resulting in semantic representations of the text and image. t and s v And use cosine similarity to measure the similarity between text and images: Among them, s t and s v Let s represent the text feature sequence and the image semantic sequence, respectively. tv s represents the similarity between the two. tv The value range of s is [-1, 1]. tv The larger the value, the higher the similarity between the two. At this point, the extraction of the image-text consistency feature is complete. Step 13: Extracting title-body consistency features; Calculate the title t using the TF-IDF algorithm. h and the main text t m To extract title-body consistency features, text similarity was calculated using jieba word segmentation. First, the t-values ​​in the data were analyzed. h and t m The content is segmented into words, and the keywords for each word are identified and converted into corresponding vectors t. h1 ,t h2 …t hi and t m1 ,t m2 …t mj (where i and j are both positive integers), and then all keywords are grouped into a set k: Where k is not an empty set; Calculate the TF-IDF values ​​d of the title and body text relative to the keywords in set k, respectively. h d m f h and f m ; Calculate the TF-IDF vectors of the title and text separately. and The calculation formula is as follows; and cosine similarity The calculation formula is as follows; in, and These represent the TF-IDF vectors of the title and text, respectively. This indicates the similarity between the two, i.e., the title-body consistency feature. At this point, the extraction of the image-text consistency feature is complete.

3. The social network information credibility evaluation method based on DS evidence theory according to claim 1, characterized in that, Step 2 includes the following: Step 21: Obtain the original evidence; In the feature selection task, three types of features are extracted to obtain corresponding feature vectors: text feature vectors. Image-text consistency feature vector s tv Consistency feature vector between title and body text Then, the credibility of the information is evaluated using three feature vectors respectively, and a credibility evaluation method based on the three features is proposed. The credibility evaluation method based on the features of the main text is to... After being fed into the classifier, the classification result is obtained, and then the result is converted into a percentage R. a R a This refers to the result of credibility evaluation using text features; The credibility evaluation method based on the consistency feature of images and text is to use s tv The percentage R is obtained by mapping the Sigmoid function to the interval [0,1]. b And this serves as the final result of the evaluation method based on the consistency features of images and text; The credibility evaluation method based on the consistency feature of title and body text is to... The percentage R is obtained by mapping the Sigmoid function to the interval [0,1]. c And this serves as the final result of the evaluation method based on title-body features; Step 22: Generate the raw BPA; Considering that DS evidence theory is an ideal method for handling uncertainty problems, when using evidence theory to fuse evaluation results with different features, a basic probability allocation (BPA) function is calculated to transform the evaluation results into probability values ​​in the identification framework. In the false information detection problem, the detection result is either true (Real, R) or false (False, F), then the identification framework is Φ = {R, F}, where R and F are mutually exclusive, and Φ is a finite complete set. The formulas for calculating the BPA function under different attributes are as follows: λ(x)=1-2|x-0.5|,x∈[0,1] (9) m a,φ =1-m a,R -m a,F (12) m b,φ =1-m b,R -m b,F (15) m c,φ =1-m c,R -m c,F (18) Where λ(z) is a function for calculating the ambiguity rate of the probability value x, and λ(x)∈[0,1]; the closer the value of x is to 0.5, the more difficult it is for this feature to distinguish whether the news is fake information, and the higher its corresponding ambiguity rate will be; m a,R Indicates that R a Convert to a true BPA value, m a,F Indicates that R a Convert to a BPA value that results in a false result, m a,φ Indicates that R a This is converted into a BPA value with an uncertain outcome; similarly, m b,R Indicates that R b Convert to a true BPA value, m b,F Indicates that R b Convert to a BPA value that results in a false result, m b,φ Indicates that R b Convert to a BPA value with an uncertain outcome; m c,R Indicates that R c Convert to a true BPA value, m c,F Indicates that R c Convert to a BPA value that results in a false result, m c,φ Indicates that R c Converted into a BPA value with an uncertain outcome.

Images