False comment detection method integrating text features and aspect features

By integrating text features and aspect features, combined with Gaussian hybrid algorithm and fuzzy clustering algorithm, the problem of degradation of false comment detection accuracy in the existing technology is solved, and more accurate false comment recognition and true comment presentation are achieved.

CN119067683AInactive Publication Date: 2024-12-03AGRI INFORMATION INST OF CHINESE ACAD OF AGRI SCI +1

Patent Information

Application Number
CN202411122778.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-12-03
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Existing false comment detection methods mainly rely on text features or behavioral features, making it difficult to accurately identify false comments. Especially when the level of false comment writers increases, the similarity between true comments and false comments increases, resulting in a decrease in detection accuracy.

Method used

Using the method of integrating text features and aspect features, the text vectorized features and aspect features are preprocessed by obtaining comment text data, and the text vectorized features and aspect features are extracted, combined with Gaussian hybrid algorithm and fuzzy clustering algorithm, semantic correlation analysis and discrete analysis are carried out to establish false comment detection. Model, realizes the organic combination of coarse and fine-grained information.

Benefits of technology

It improves the accuracy of false comment detection, enables more comprehensive analysis of review text, accurately identify false comments, and provides real product review information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067683A_ABST
    Figure CN119067683A_ABST
Patent Text Reader

Abstract

The invention discloses a false comment detection method integrating text features and aspect features, and the method comprises the steps: obtaining comment text data, and carrying out the preprocessing of the comment text data; extracting text features of the evaluation text data, and extracting aspect features of the evaluation text data; performing first clustering on the evaluation text data to obtain first classification data, and performing second clustering on the evaluation text data to obtain second classification data; performing semantic association analysis on the first classification data to obtain a false degree, and performing discrete analysis on the second classification data to obtain a deviation degree; and establishing a false comment detection model according to the false degree and the deviation degree, and inputting a comment text to be detected into the false comment detection model to obtain a detection result. The method not only can improve the accuracy of false comment detection integrating text features and aspect features, but also has good interpretability, and can be directly applied to a false comment detection system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of natural language processing, and in particular to a false comment detection method integrating text features and aspect features. Background Art

[0002] In the Internet age, people are increasingly inclined to shop online. Since online shopping limits people's personal experience of products, they usually can only rely on product reviews to make purchasing decisions. Positive product reviews can often bring huge profits to merchants, while negative reviews have the opposite effect. Since reviews have a direct impact on economic benefits, some merchants will hire people to write false reviews for their brands in order to attract more users. However, it is difficult for consumers to accurately distinguish between false reviews and real reviews when browsing product reviews, which may lead to mistakes in purchasing decisions and thus affect economic benefits. The current false review detection method mainly relies on text features or behavioral features, and it is difficult to accurately identify false reviews by relying solely on coarse-grained information. With the improvement of the level of false review writers, the similarity between false reviews and real reviews increases, but the review content of the two is different in terms of products. In order to improve the detection accuracy, the present invention proposes a false review detection method integrating text features and aspect features, which realizes the organic combination of coarse-grained and fine-grained information, thereby effectively detecting false reviews and presenting real product reviews to consumers. Summary of the invention

[0003] The purpose of the present invention is to provide a false review detection method integrating text features and aspect features.

[0004] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0005] The present invention comprises the following steps:

[0006] Acquire comment text data, and preprocess the comment text data;

[0007] Extracting text features of the comment text data, and extracting aspect features of the comment text data;

[0008] Performing a first clustering on the comment text data to obtain first classification data, and performing a second clustering on the comment text data to obtain second classification data;

[0009] Performing semantic association analysis on the first classification data to obtain a falsehood degree, and performing discrete analysis on the second classification data to obtain a deviation degree;

[0010] A false review detection model is established according to the false degree and the deviation degree, and the review text to be detected is input into the false review detection model to obtain a detection result.

[0011] Furthermore, the method of obtaining comment text data and preprocessing the comment text data includes:

[0012] The comment text data includes real comments and fake comments. The preprocessing of the comment text data includes deleting special characters, word segmentation, removing stop words, and stemming.

[0013] Furthermore, the method of extracting text features of the comment text data and extracting aspect features of the comment text data includes:

[0014] Text features include text vectorization features and text discretization features. The sentence-level and word-level vectors obtained by inputting the comment text data into the longformer model are used as text vectorization features. Text discretization features include sentiment polarity features and sentiment features.

[0015] A pre-trained aspect-level sentiment analysis model is used to obtain aspect words and aspect sentiment types of the comment text data. The pre-trained aspect-level sentiment analysis model is a Bi-LSTM classification model. The Bi-LSTM classification model is trained using a comment sample set containing aspect word labels and aspect sentiment type labels to obtain a Bi-LSTM classification model that can identify aspect words and aspect sentiment types in the comment text data.

[0016] Cluster the aspect words and aspect sentiment types to obtain aspect types, obtain the score of the aspect types by summing up the sentiment values ​​of the aspect types, and calculate the difference between the score of the aspect types and the average score;

[0017] Aspect features include aspect-based features and aspect type-based features. Aspect-based features include the number of aspect words and the number of corresponding sentiment polarities, and aspect type-based features include the number of aspect types, scores, sentiment values, and score differences.

[0018] Furthermore, the method of performing a first clustering on the comment text data to obtain first classification data includes:

[0019] Text features and aspect features are used to establish data points, and the Gaussian mixture algorithm is used for the first clustering:

[0020] Probability density function of Gaussian distribution in Gaussian mixture algorithm:

[0021]

[0022] Where x is the data point, Λ is the dimension of the data point, ρ is the mean vector, ψ is the covariance matrix, |ψ| is the determinant of the covariance matrix, exp represents the exponential function, and T represents the transpose of the matrix.

[0023] Probability density function of a data point:

[0024]

[0025] Where K is the number of Gaussian distributions in the Gaussian mixture algorithm, j is the index of the Gaussian distribution, and q is the mixing coefficient.

[0026] Posterior probability:

[0027]

[0028] where α ik is the posterior probability that the i-th data point belongs to the k-th Gaussian distribution in the Gaussian mixture algorithm, x i is the ith data point,

[0029] Mean vector and mixing coefficient:

[0030]

[0031] Where n is the number of data points and the covariance matrix is:

[0032]

[0033] where τ ik is the contribution of the covariance matrix of the kth Gaussian distribution. The contribution of the covariance matrix is ​​obtained based on the likelihood ratio test, and I is the unit matrix.

[0034] Furthermore, the method of performing a second clustering on the comment text data to obtain second classification data includes:

[0035] The number of clusters of data points is determined according to the elbow rule, where the cost function value is calculated:

[0036]

[0037] in, is the nth data point of cluster m, k is the number of clusters, O is the number of data points, q m is the cluster center of cluster m, express and q m The distance between

[0038] Randomly select cluster centers and use cosine similarity as the distance between data points and cluster centers:

[0039]

[0040] Where g(p a ,q b ) is the ath data point p aand the b-th cluster center q b The cosine similarity between ai is the data point p a The i-th data, q bi is the cluster center point q b The i-th data, n is the dimension of the data point,

[0041] Divide the data points into the cluster where the nearest cluster center point is located, and update the cluster center point:

[0042]

[0043] where q m (t+1) is the mth cluster Q of the t+1th iteration m The cluster center point, η(t) is the learning rate, g∈Q m (t) represents the cluster Q of the tth iteration m The distance between the data points within is the cluster center point q m The gradient of D(Q m (t)|q m ) is the clustering conditional probability,

[0044] Objective function:

[0045]

[0046] where p n ∈Q m Q m The cluster center point and the objective function are iteratively updated for a set of data points within the cluster until a preset number of iterations is reached, and the clustering result is output as the second classification data.

[0047] Furthermore, the method of performing semantic association analysis on the first classification data to obtain the falsehood degree includes:

[0048] Calculate the semantic similarity between data points of the first classification data:

[0049]

[0050] Where S(x,c) represents the semantic similarity between data point x and data point c, and the i-th text vectorization feature of data point x is x i , the i-th text vectorization feature of data point c is c i , the j-th text discretization feature of data point x is x j , the j-th text discretization feature of data point c is c j, the coefficient of the text vectorization feature is a, the coefficient of the text discretization feature is b, the number of text vectorization features is u, and the number of text discretization features is e.

[0051] The local density of data points in the first category of data:

[0052]

[0053] where d k (x p ,x o ) is the data point x p and data point x o The distance between data points is the semantic similarity between data points, N k (x p ) is the data point x p The k-distance neighborhood of

[0054] Calculate the degree of spuriousness based on the local density:

[0055]

[0056] Where L k (x p ) is the data point x p degree of falsehood.

[0057] Furthermore, a method for performing discrete analysis on the second classification data to obtain a degree of deviation includes:

[0058] The average intra-cluster distance of data points in the second classification data:

[0059]

[0060] where g avg,m is the average intra-cluster distance of cluster m, |Q m | is the number of data points in cluster m,

[0061] Calculation of the degree of deviation:

[0062]

[0063] The data point p n The degree of deviation is σ n .

[0064] Furthermore, a method for establishing a false review detection model according to the falseness degree and the deviation degree includes:

[0065] The degree of deviation and falsehood are standardized and preprocessed to obtain data vectors. Fuzzy clustering algorithm is used to establish a false review detection model. The objective function and constraints are as follows:

[0066]

[0067] Where min J is the objective function, st represents the constraint condition, c is the number of data clusters, m is the fuzzy factor, x j is the jth data vector, N is the number of data vectors, is the jth control weight coefficient, e is the number of reference data vectors, represents rounding up, e∈[1,N], r is the data reference rate, r∈(0,1], v h is the hth cluster center, is the membership of the jth data vector to the hth data cluster, is the fuzzy local information factor of the jth data vector to the hth data cluster, and the calculation formula is:

[0068]

[0069] where d kj is the data vector x j and data vector x k The spatial distance between j is the data vector x j Neighborhood of , with the cutoff value determined based on a reference data vector:

[0070]

[0071] in is the cutoff value, is the membership degree of the e-th data vector to the h-th data cluster, x e is the e-th data vector, is the fuzzy local information factor of the e-th data vector to the h-th data cluster,

[0072] Control weighting coefficient:

[0073]

[0074] The Lagrange multiplier method is used to solve the objective function:

[0075]

[0076] in is an auxiliary function with Lagrangian constraints, λ j is the Lagrange multiplier, and the membership update formula and cluster center update formula are obtained according to the auxiliary function:

[0077]

[0078] where v kis the kth cluster center, is the fuzzy local information factor of the j-th data vector to the k-th data cluster,

[0079] Calculate the false probability of comment text data based on fuzzy clustering results:

[0080]

[0081] Where P false,j is the false probability of data vector j, β h is the proportion of false comments in the data cluster h, and the comment text data with a false probability greater than the preset false threshold is detected as a false comment.

[0082] The beneficial effects of the present invention are:

[0083] The present invention is a false review detection method integrating text features and aspect features. Compared with the prior art, the present invention has the following technical effects:

[0084] The present invention provides a method for detecting fake reviews by integrating multi-granularity features, which can consider text features and aspect features at the same time, comprehensively analyze the review text, and thus more accurately identify fake reviews. By extracting multi-dimensional features such as text vectorization features, text discretization features, aspect-based features, and aspect type-based features, an organic combination of coarse-grained and fine-grained information is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 The present invention is a flowchart of the steps of a false review detection method integrating text features and aspect features. DETAILED DESCRIPTION

[0086] The present invention is further described below by means of specific embodiments. The illustrative embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0087] The present invention provides a false review detection method integrating text features and aspect features, comprising the following steps:

[0088] like Figure 1 As shown, in this embodiment, the following steps are included:

[0089] Acquire comment text data, and preprocess the comment text data;

[0090] In the actual evaluation, the collected review text data includes:

[0091] Review 1 (false): This phone is really great, fast, long battery life, I am very satisfied;

[0092] Comment 2 (true): The phone is good, but the battery life can be improved;

[0093] Review 3 (False): I am very happy with this product, no issues at all, highly recommend;

[0094] Comment 4 (true): The performance is very good, but the price is a bit high;

[0095] Review 5 (False): This is the best product I have ever used, everything works perfectly;

[0096] Comment 6 (True): The screen display is great, but the camera is a bit noisy in low light conditions;

[0097] Extracting text features of the comment text data, and extracting aspect features of the comment text data;

[0098] In the actual evaluation, the text discretization features in the aspect features and text features include: Comment 1: sentiment polarity feature, positive, sentiment feature, 0.9, aspect words: [mobile phone, running speed, battery life], sentiment polarity: [positive, positive, positive], number of aspect types, 3, sum of sentiment values, 2.7, average rating, 0.9, rating difference: 0, 0, 0; Comment 2: sentiment polarity feature, neutral, sentiment feature, 0.6, aspect words: [mobile phone, battery life], sentiment polarity: [positive, neutral], number of aspect types, 2, sum of sentiment values, 1.4, average rating, 0.7, rating difference: 0.1, 0.1; Comment 3: sentiment polarity feature, positive, sentiment feature, 0.9, aspect words: [mobile phone], sentiment polarity: [positive], number of aspect types, 1, sum of sentiment values, 0.9, average Average rating, 0.9, rating difference: 0; Review 4: Sentiment polarity feature, neutral, sentiment feature, 0.6, aspect words: [performance, price], sentiment polarity: [positive, negative], number of aspect types, 2, sum of sentiment values, 1.3, average rating, 0.65, rating difference: 0.35, 0.35; Review 5: Sentiment polarity feature, positive, sentiment feature, 0.9, aspect words: [product], sentiment polarity: [positive], number of aspect types, 1, sum of sentiment values, 0.9, average rating, 0.9, rating difference: 0; Review 6: Sentiment polarity feature, positive, sentiment feature, 0.7, aspect words: [screen display effect, camera], sentiment polarity: [positive, neutral], number of aspect types, 2, sum of sentiment values, 1.3, average rating, 0.65, rating difference: 0.35, 0.35;

[0099] Performing a first clustering on the comment text data to obtain first classification data, and performing a second clustering on the comment text data to obtain second classification data;

[0100] In actual evaluation, the first category of data includes:

[0101] Category 1 includes Comment 1, Comment 4, and Comment 6, and Category 2 includes Comment 2, Comment 3, and Comment 5;

[0102] The second category of data includes:

[0103] Category 1 includes Comment 1, Comment 3, and Comment 5, and Category 2 includes Comment 2, Comment 4, and Comment 6;

[0104] Performing semantic association analysis on the first classification data to obtain a falsehood degree, and performing discrete analysis on the second classification data to obtain a deviation degree;

[0105] In actual evaluation, the first category of data includes:

[0106] The falsehood levels of category 1 comments 1, 4, and 6 are 0.7, 0.2, and 0.3, respectively; the falsehood levels of category 2 comments 2, 3, and 5 are 0.2, 0.6, and 0.8, respectively;

[0107] The second category of data includes:

[0108] The deviations of category 1 comments 1, 3, and 5 are 0.7, 0.8, and 0.9, respectively, and the deviations of category 2 comments 2, 4, and 6 are 0.1, 0.3, and 0.2, respectively;

[0109] Establishing a false review detection model according to the falseness degree and the deviation degree, inputting the review text to be detected into the false review detection model, and obtaining a detection result;

[0110] In the actual evaluation, the false probabilities of comment 1, comment 2, comment 3, comment 4, comment 5 and comment 6 are 0.9, 0.1, 0.85, 0.05, 0.95 and 0.15 respectively, and the preset false threshold is 0.5.

[0111] In this embodiment, the method of obtaining comment text data and preprocessing the comment text data includes:

[0112] The comment text data includes real comments and fake comments. The preprocessing of the comment text data includes deleting special characters, word segmentation, removing stop words, and stemming.

[0113] In this embodiment, the method of extracting text features of the comment text data and extracting aspect features of the comment text data includes:

[0114] Text features include text vectorization features and text discretization features. The sentence-level and word-level vectors obtained by inputting the comment text data into the longformer model are used as text vectorization features. Text discretization features include sentiment polarity features and sentiment features.

[0115] The comment text data is matched by using the OpinionFinder opinion dictionary and the BingLiu opinion dictionary to obtain the sentiment value of the word and the sentiment polarity feature. The sentiment polarity feature includes positive sentiment value, negative sentiment value and neutral sentiment value.

[0116] The emotional features include the emotional values ​​of eight emotional types: anger, disgust, fear, expectation, happiness, sadness, surprise, and trust. The emotional value of each emotional type is obtained by using the NRC emotional dictionary;

[0117] A pre-trained aspect-level sentiment analysis model is used to obtain aspect words and aspect sentiment types of the comment text data. The pre-trained aspect-level sentiment analysis model is a Bi-LSTM classification model. The Bi-LSTM classification model is trained using a comment sample set containing aspect word labels and aspect sentiment type labels to obtain a Bi-LSTM classification model that can identify aspect words and aspect sentiment types in the comment text data.

[0118] Cluster the aspect words and aspect sentiment types to obtain aspect types, obtain the score of the aspect types by summing up the sentiment values ​​of the aspect types, and calculate the difference between the score of the aspect types and the average score;

[0119] Aspect features include aspect-based features and aspect type-based features. Aspect-based features include the number of aspect words and the number of corresponding sentiment polarities, and aspect type-based features include the number of aspect types, scores, sentiment values, and score differences.

[0120] In this embodiment, the method of performing a first clustering on the comment text data to obtain first classification data includes:

[0121] Text features and aspect features are used to establish data points, and the Gaussian mixture algorithm is used for the first clustering:

[0122] Probability density function of Gaussian distribution in Gaussian mixture algorithm:

[0123]

[0124] Where x is the data point, Λ is the dimension of the data point, ρ is the mean vector, ψ is the covariance matrix, |ψ| is the determinant of the covariance matrix, exp represents the exponential function, and T represents the transpose of the matrix.

[0125] Probability density function of a data point:

[0126]

[0127] Where K is the number of Gaussian distributions in the Gaussian mixture algorithm, j is the index of the Gaussian distribution, and q is the mixing coefficient.

[0128] Posterior probability:

[0129]

[0130] where α ik is the posterior probability that the i-th data point belongs to the k-th Gaussian distribution in the Gaussian mixture algorithm, x i is the ith data point,

[0131] Mean vector and mixing coefficient:

[0132]

[0133] Where n is the number of data points and the covariance matrix is:

[0134]

[0135] where τ ik is the contribution of the covariance matrix of the kth Gaussian distribution. The contribution of the covariance matrix is ​​obtained based on the likelihood ratio test, and I is the unit matrix.

[0136] In this embodiment, the method of performing a second clustering on the comment text data to obtain second classification data includes:

[0137] The number of clusters of data points is determined according to the elbow rule, where the cost function value is calculated:

[0138]

[0139] in, is the nth data point of cluster m, k is the number of clusters, O is the number of data points, q m is the cluster center of cluster m, express and q m The distance between

[0140] Randomly select cluster centers and use cosine similarity as the distance between data points and cluster centers:

[0141]

[0142] Where g(p a ,q b ) is the ath data point p a and the b-th cluster center q b The cosine similarity between ai is the data point p a The i-th data, q bi is the cluster center point q b The i-th data, n is the dimension of the data point,

[0143] Divide the data points into the cluster where the nearest cluster center point is located, and update the cluster center point:

[0144]

[0145] where q m (t+1) is the mth cluster Q of the t+1th iteration m The cluster center point, η(t) is the learning rate, g∈Q m (t) represents the cluster Q of the tth iteration m The distance between the data points within is the cluster center point q m The gradient of D(Q m (t)|q m ) is the clustering conditional probability,

[0146] Objective function:

[0147]

[0148] where p n ∈Q m Q m The cluster center point and the objective function are iteratively updated for a set of data points within the cluster until a preset number of iterations is reached, and the clustering result is output as the second classification data.

[0149] In this embodiment, the method of performing semantic association analysis on the first classification data to obtain the falsehood degree includes:

[0150] Calculate the semantic similarity between data points of the first classification data:

[0151]

[0152] Where S(x,c) represents the semantic similarity between data point x and data point c, and the i-th text vectorization feature of data point x is x i , the i-th text vectorization feature of data point c is c i , the j-th text discretization feature of data point x is x j , the j-th text discretization feature of data point c is c j , the coefficient of the text vectorization feature is a, the coefficient of the text discretization feature is b, the number of text vectorization features is u, and the number of text discretization features is e.

[0153] The local density of data points in the first category of data:

[0154]

[0155] where d k (xp ,x o ) is the data point x p and data point x o The distance between data points is the semantic similarity between data points, N k (x p ) is the data point x p The k-distance neighborhood of

[0156] Calculate the degree of spuriousness based on the local density:

[0157]

[0158] Where L k (x p ) is the data point x p degree of falsehood.

[0159] In this embodiment, the method of performing discrete analysis on the second classification data to obtain the degree of deviation includes:

[0160] The average intra-cluster distance of data points in the second classification data:

[0161]

[0162] where g avg,m is the average intra-cluster distance of cluster m, |Q m | is the number of data points in cluster m,

[0163] Calculation of the degree of deviation:

[0164]

[0165] The data point p n The degree of deviation is σ n .

[0166] In this embodiment, the method for establishing a false review detection model according to the falseness degree and the deviation degree includes:

[0167] The degree of deviation and falsehood are standardized and preprocessed to obtain data vectors. Fuzzy clustering algorithm is used to establish a false review detection model. The objective function and constraints are as follows:

[0168]

[0169] Where min J is the objective function, st represents the constraint condition, c is the number of data clusters, m is the fuzzy factor, x j is the jth data vector, N is the number of data vectors, is the jth control weight coefficient, e is the number of reference data vectors, represents rounding up, e∈[1,N], r is the data reference rate, r∈(0,1], v h is the hth cluster center, is the membership of the jth data vector to the hth data cluster, is the fuzzy local information factor of the jth data vector to the hth data cluster, and the calculation formula is:

[0170]

[0171] where d kj is the data vector x j and data vector x k The spatial distance between j is the data vector x j Neighborhood of , with the cutoff value determined based on a reference data vector:

[0172]

[0173] in is the cutoff value, is the membership degree of the e-th data vector to the h-th data cluster, x e is the e-th data vector, is the fuzzy local information factor of the e-th data vector to the h-th data cluster,

[0174] Control weighting coefficient:

[0175]

[0176] The Lagrange multiplier method is used to solve the objective function:

[0177]

[0178] in is an auxiliary function with Lagrangian constraints, λ j is the Lagrange multiplier, and the membership update formula and cluster center update formula are obtained according to the auxiliary function:

[0179]

[0180] where v k is the kth cluster center, is the fuzzy local information factor of the j-th data vector to the k-th data cluster,

[0181] Calculate the false probability of comment text data based on fuzzy clustering results:

[0182]

[0183] Where P false,jis the false probability of data vector j, β h is the proportion of false comments in the data cluster h, and the comment text data with a false probability greater than the preset false threshold is detected as a false comment.

[0184] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A false review detection method integrating text features and aspect features, characterized in that: The following steps are involved: Acquire comment text data, and preprocess the comment text data; Extracting text features of the comment text data, and extracting aspect features of the comment text data; Performing a first clustering on the comment text data to obtain first classification data, and performing a second clustering on the comment text data to obtain second classification data; Performing semantic association analysis on the first classification data to obtain a falsehood degree, and performing discrete analysis on the second classification data to obtain a deviation degree; A false review detection model is established according to the false degree and the deviation degree, and the review text to be detected is input into the false review detection model to obtain a detection result.

2. According to claim 1, a false review detection method integrating text features and aspect features is characterized in that: The method of obtaining comment text data and preprocessing the comment text data comprises: The comment text data includes real comments and fake comments. The preprocessing of the comment text data includes deleting special characters, word segmentation, removing stop words, and stemming.

3. According to claim 1, a false review detection method integrating text features and aspect features is characterized in that: The method of extracting text features of the comment text data and extracting aspect features of the comment text data comprises: Text features include text vectorization features and text discretization features. The sentence-level and word-level vectors obtained by inputting the comment text data into the longformer model are used as text vectorization features. Text discretization features include sentiment polarity features and sentiment features. A pre-trained aspect-level sentiment analysis model is used to obtain aspect words and aspect sentiment types of the comment text data. The pre-trained aspect-level sentiment analysis model is a Bi-LSTM classification model. The Bi-LSTM classification model is trained using a comment sample set containing aspect word labels and aspect sentiment type labels to obtain a Bi-LSTM classification model that can identify aspect words and aspect sentiment types in the comment text data. Cluster the aspect words and aspect sentiment types to obtain aspect types, obtain the score of the aspect types by summing up the sentiment values ​​of the aspect types, and calculate the difference between the score of the aspect types and the average score; Aspect features include aspect-based features and aspect type-based features. Aspect-based features include the number of aspect words and the number of corresponding sentiment polarities, and aspect type-based features include the number of aspect types, scores, sentiment values, and score differences.

4. According to claim 1, a false review detection method integrating text features and aspect features is characterized in that: The method of performing a first clustering on the comment text data to obtain first classification data comprises: Text features and aspect features are used to establish data points, and the Gaussian mixture algorithm is used for the first clustering: Probability density function of Gaussian distribution in Gaussian mixture algorithm: Where x is the data point, Λ is the dimension of the data point, ρ is the mean vector, ψ is the covariance matrix, |ψ| is the determinant of the covariance matrix, exp represents the exponential function, and T represents the transpose of the matrix. Probability density function of a data point: Where K is the number of Gaussian distributions in the Gaussian mixture algorithm, j is the index of the Gaussian distribution, and q is the mixing coefficient. Posterior probability: where α ik is the posterior probability that the i-th data point belongs to the k-th Gaussian distribution in the Gaussian mixture algorithm, x i is the ith data point, Mean vector and mixing coefficient: Where n is the number of data points and the covariance matrix is: where τ ik is the contribution of the covariance matrix of the kth Gaussian distribution. The contribution of the covariance matrix is ​​obtained based on the likelihood ratio test, and I is the unit matrix.

5. According to claim 1, a false review detection method integrating text features and aspect features is characterized in that: The method of performing a second clustering on the comment text data to obtain second classification data comprises: The number of clusters of data points is determined according to the elbow rule, where the cost function value is calculated: in, is the nth data point of cluster m, k is the number of clusters, O is the number of data points, and q m is the cluster center of cluster m, express and q m The distance between Randomly select cluster centers and use cosine similarity as the distance between data points and cluster centers: Where g(p a ,q b ) is the ath data point p a and the b-th cluster center q b The cosine similarity between ai is the data point p a The i-th data, q bi is the cluster center point q b The i-th data, n is the dimension of the data point, Divide the data points into the cluster where the nearest cluster center point is located, and update the cluster center point: where q m (t+1) is the mth cluster Q of the t+1th iteration m The cluster center point, η(t) is the learning rate, g∈Q m (t) represents the cluster Q of the tth iteration m The distance between the data points within is the cluster center point q m The gradient of D(Q m (t)|q m ) is the clustering conditional probability, Objective function: where p n ∈Q m Q m The cluster center point and the objective function are iteratively updated for a set of data points within the cluster until a preset number of iterations is reached, and the clustering result is output as the second classification data.

6. According to claim 1, a false review detection method integrating text features and aspect features is characterized in that: The method of performing semantic association analysis on the first classification data to obtain the falsehood degree includes: Calculate the semantic similarity between data points of the first classification data: Where S(x,c) represents the semantic similarity between data point x and data point c, and the i-th text vectorization feature of data point x is x i , the i-th text vectorization feature of data point c is c i , the j-th text discretization feature of data point x is x j , the j-th text discretization feature of data point c is c j , the coefficient of the text vectorization feature is a, the coefficient of the text discretization feature is b, the number of text vectorization features is u, and the number of text discretization features is e. The local density of data points in the first category of data: where d k (x p ,x o ) is the data point x p and data point x o The distance between data points is the semantic similarity between data points, N k (x p ) is the data point x p The k-distance neighborhood of Calculate the degree of spuriousness based on the local density: Where L k (x p ) is the data point x p degree of falsehood.

7. The method for detecting false comments by integrating text features and aspect features according to claim 1, characterized in that: The method of performing discrete analysis on the second classification data to obtain the degree of deviation includes: The average intra-cluster distance of data points in the second classification data: where g avg,m is the average intra-cluster distance of cluster m, |Q m | is the number of data points in cluster m, Calculation of the degree of deviation: The data point p n The degree of deviation is σ n .

8. The method for detecting false comments by integrating text features and aspect features according to claim 1, characterized in that: The method for establishing a false review detection model according to the falseness degree and the deviation degree includes: The degree of deviation and falsehood are standardized and preprocessed to obtain data vectors. Fuzzy clustering algorithm is used to establish a false review detection model. The objective function and constraints are as follows: Where minJ is the objective function, st represents the constraint condition, c is the number of data clusters, m is the fuzzy factor, and x j is the jth data vector, N is the number of data vectors, is the jth control weight coefficient, e is the number of reference data vectors, represents rounding up, e∈[1,N], r is the data reference rate, r∈(0,1], v h is the hth cluster center, is the membership of the jth data vector to the hth data cluster, is the fuzzy local information factor of the jth data vector to the hth data cluster, and the calculation formula is: where d kj is the data vector x j and data vector x k The spatial distance between j is the data vector x j Neighborhood of , with the cutoff value determined based on a reference data vector: in is the cutoff value, is the membership degree of the e-th data vector to the h-th data cluster, x e is the e-th data vector, is the fuzzy local information factor of the e-th data vector to the h-th data cluster, Control weighting coefficient: The Lagrange multiplier method is used to solve the objective function: in is an auxiliary function with Lagrangian constraints, λ j is the Lagrange multiplier, and the membership update formula and cluster center update formula are obtained according to the auxiliary function: where v k is the kth cluster center, is the fuzzy local information factor of the j-th data vector to the k-th data cluster, Calculate the false probability of comment text data based on fuzzy clustering results: Where P false,j is the false probability of data vector j, β h is the proportion of false comments in the data cluster h, and the comment text data with a false probability greater than the preset false threshold is detected as a false comment.

Citation Information

Patent Citations

  • False comment identification method based on rolling type cooperative training

    CN111666480A

  • False comment detection model training method, detection method and device

    CN112463966A

Cited By

  • Internet information intelligent early warning system based on big data

    CN119377620A