Method and system for detecting fake reviews

Abstract

The application discloses a false comment detection method and system, to solve the technical problem of low accuracy of false comment detection. Among them, a false comment detection scheme, according to the false degree evaluation results of the four dimensions of comment text, input object, input group and output object, the comprehensive false degree value more accurately reflecting the false comment is calculated. Based on the comprehensive false degree value, the false comment is identified, and the false comment detection scheme provided by the application improves the detection accuracy of the false comment.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and system for detecting fake comments. Background Technology

[0002] With the continuous development of internet technology, shopping methods are also changing. Currently, more and more people are shopping online. When selecting products online, in addition to comparing product information, people usually read product reviews before deciding whether to buy from a store. Therefore, product reviews have a significant influence on consumers' purchasing behavior.

[0003] To attract users, merchants create a large number of fake reviews. Some users also maliciously create fake reviews to damage a product's reputation. Clearly, false or inaccurate reviews can mislead buyers, seriously harm user rights, and disrupt market balance.

[0004] In the process of developing the existing technology, the inventors discovered that:

[0005] Current technologies for detecting fake comments only focus on the semantic and sentiment features of the comment text. However, modern fake comments deliberately imitate or even plagiarize genuine comments, resulting in very similar text content. This leads to low accuracy of current technologies in detecting fake comments.

[0006] Therefore, a new fake review detection scheme is needed to solve the technical problem of low accuracy in fake review detection. Summary of the Invention

[0007] This application provides a new scheme for detecting fake reviews, which solves the technical problem of low accuracy in detecting fake reviews.

[0008] Specifically, a method for detecting fake reviews includes the following steps:

[0009] Get the comment text;

[0010] Determine the textual and semantic features of the comment text;

[0011] Based on the textual and semantic features of the comment text, the first dimension of the falsehood is evaluated, and the first dimension of the falsehood value of the comment text is obtained as the first evaluation factor;

[0012] Identify the input object for the corresponding comment text and the behavioral characteristics of the input object;

[0013] Based on the behavioral characteristics of the input object, a second-dimensional falseness evaluation is performed on the input object to obtain the second-dimensional falseness value of the input object, which is used as the second evaluation factor;

[0014] Determine the input group corresponding to the input object, and the group characteristics of the input group;

[0015] Based on the group characteristics of the input group, a third-dimensional falsehood evaluation is performed on the input group to obtain the third-dimensional falsehood value of the input group, which is used as the third evaluation factor;

[0016] The sum of the first evaluation factor, the second evaluation factor, and the third evaluation factor is calculated using preset weighting percentages to obtain the overall falsehood value.

[0017] When the overall falsehood value exceeds a preset threshold, the comment text is determined to be a false comment.

[0018] Furthermore, the method also includes:

[0019] Determine the output object for the corresponding comment text, and the output group for the corresponding output object;

[0020] Based on the textual and semantic features of the comment text, the behavioral features of the input objects, the behavioral features of the input groups, the output objects, and the output groups, construct a comment relationship graph that represents the relationship between the comment text, the output objects, and the input objects;

[0021] Based on the comment relationship diagram, update the first evaluation factor, the second evaluation factor, and the third evaluation factor;

[0022] The comment relationship graph is represented as G = (V, E);

[0023] In the formula, V = {v1, v2, ..., v} n} is a set of vertices;

[0024] E={e ij For i, j∈{1,...,n}} is the set of edges;

[0025] And vertex v i With vertex v j There is an edge e between them ij ;

[0026] v i It has type τ(v)∈A, A={RW,R,P};

[0027] In the formula, RW represents the input object, R represents the comment text, and P represents the output object;

[0028] Vertices of type RW input object have attributes The input group that represents the corresponding input object;

[0029] Vertices of type P in the output object have an attribute ψ, which represents the output group of the corresponding output object.

[0030] Furthermore, based on the aforementioned comment relationship graph, the first evaluation factor, the second evaluation factor, and the third evaluation factor are updated, specifically including:

[0031] With F R This represents the first evaluation factor before the update, denoted by F. RW This represents the second evaluation factor before the update, denoted by F. GRW This indicates the third evaluation factor before the update;

[0032] by This represents the updated first evaluation factor, with This represents the updated second evaluation factor, in order to This represents the updated third evaluation factor;

[0033] Then there is

[0034]

[0035]

[0036]

[0037] In the formula, F PR (i) represents the falsity value of the output object corresponding to comment text i;

[0038] F RWR (i) represents the falsity value of the input object corresponding to comment text i;

[0039] F PRW (i) represents the falsity value of the output object corresponding to input object i;

[0040] F RRW (i) represents the falsity value of the comment text corresponding to input object i;

[0041] The elements of the GR(i) set are the comment texts corresponding to the input group i;

[0042] The elements of the GRW(i) set are the input objects corresponding to input group i;

[0043] The elements of the GP(i) set are the output objects corresponding to the input group i;

[0044] F P (i) represents the spuriousness value of output object i;

[0045] and

[0046]

[0047] The elements of the PR(i) set are the output objects corresponding to comment text i;

[0048]

[0049] The elements of the RWR(i) set are the output objects corresponding to the input object i;

[0050]

[0051] The elements of the PRW(i) set are the output objects corresponding to the input object i;

[0052]

[0053] The elements of the RRW(i) set are the comment texts corresponding to the input object i.

[0054] Furthermore, the method also includes:

[0055] when and and When that happens, stop updating the first evaluation factor, the second evaluation factor, and the third evaluation factor.

[0056] Furthermore, the method also includes:

[0057] Based on the identity characteristics of the output object, the fourth dimension of the falseness evaluation is performed on the output object to obtain the fourth dimension falseness value of the output object, which is used as the fourth evaluation factor;

[0058] The sum of the first evaluation factor, the second evaluation factor, the third evaluation factor, and the fourth evaluation factor is calculated using preset weight ratios to obtain the overall falseness value.

[0059] Furthermore, the method also includes:

[0060] Update the fourth evaluation factor based on the aforementioned comment relationship diagram;

[0061] With F P This represents the fourth evaluation factor before the update. This represents the updated fourth evaluation factor;

[0062] Then there is

[0063]

[0064] F RP (i) represents the degree of falsity of the comment text corresponding to output object i;

[0065] F RWP (i) represents the falsity of the input object corresponding to output object i;

[0066] and

[0067]

[0068] The elements of the RP(i) set are the comment text corresponding to output object i;

[0069]

[0070] The elements of the RP(i) set are the input objects corresponding to the output object i.

[0071] This application also provides a fake review detection system.

[0072] Specifically, a fake review detection system includes:

[0073] The first-dimensional evaluation module is used to acquire the comment text; it is also used to determine the textual and semantic features of the comment text; and it is also used to evaluate the first-dimensional falsity of the comment text based on the textual and semantic features of the comment text, thereby obtaining the first-dimensional falsity value of the comment text as the first evaluation factor.

[0074] The second-dimensional evaluation module is used to determine the input object of the corresponding comment text and the behavioral characteristics of the input object; it is also used to evaluate the second-dimensional falsity of the input object based on the behavioral characteristics of the input object, and obtain the second-dimensional falsity value of the input object as the second evaluation factor.

[0075] The third-dimensional evaluation module is used to determine the input group of the corresponding input object and the group characteristics of the input group; it is also used to evaluate the falseness of the input group in the third dimension based on the group characteristics of the input group, and obtain the falseness value of the input group in the third dimension as the third evaluation factor.

[0076] The output module is used to calculate the sum of the first evaluation factor, the second evaluation factor, and the third evaluation factor with preset weight ratios to obtain the comprehensive falsehood value; it is also used to determine that the comment text is a false comment when the comprehensive falsehood value is greater than a preset threshold.

[0077] Furthermore, the system also includes:

[0078] The update module is used to determine the output object of the corresponding comment text and the output group of the corresponding output object; it is also used to construct a comment relationship graph representing the relationship between the comment text, the output object, and the input object based on the textual and semantic features of the comment text, the behavioral features of the input object, the behavioral features of the input group, the output object, and the output group; and it is also used to update the first evaluation factor, the second evaluation factor, and the third evaluation factor based on the comment relationship graph.

[0079] The comment relationship graph is represented as G = (V, E);

[0080] In the formula, V = {v1, v2, ..., v} n} is a set of vertices;

[0081] E={e ij For i, j∈{1,...,n}} is the set of edges;

[0082] And vertex v i With vertex v j There is an edge e between them ij ;

[0083] v i It has type τ(v)∈A, A={RW,R,P};

[0084] In the formula, RW represents the input object, R represents the comment text, and P represents the output object;

[0085] Vertices of type RW input object have attributes The input group that represents the corresponding input object;

[0086] Vertices of type P in the output object have an attribute ψ, which represents the output group of the corresponding output object.

[0087] Furthermore, the update module updates the first evaluation factor, the second evaluation factor, and the third evaluation factor based on the comment relationship graph, specifically including:

[0088] With F R This represents the first evaluation factor before the update, denoted by F. RW This represents the second evaluation factor before the update, denoted by F. GRW This indicates the third evaluation factor before the update;

[0089] by This represents the updated first evaluation factor, with This represents the updated second evaluation factor, in order to This represents the updated third evaluation factor;

[0090] Then there is

[0091]

[0092]

[0093]

[0094] In the formula, F PR (i) represents the falsity value of the output object corresponding to comment text i;

[0095] F RWR(i) represents the falsity value of the input object corresponding to comment text i;

[0096] F PRW (i) represents the falsity value of the output object corresponding to input object i;

[0097] F RRW (i) represents the falsity value of the comment text corresponding to input object i;

[0098] The elements of the GR(i) set are the comment texts corresponding to the input group i;

[0099] The elements of the GRW(i) set are the input objects corresponding to input group i;

[0100] The elements of the GP(i) set are the output objects corresponding to the input group i;

[0101] F P (i) represents the spuriousness value of output object i;

[0102] and

[0103]

[0104] The elements of the PR(i) set are the output objects corresponding to comment text i;

[0105]

[0106] The elements of the RWR(i) set are the output objects corresponding to the input object i;

[0107]

[0108] The elements of the PRW(i) set are the output objects corresponding to the input object i;

[0109]

[0110] The elements of the RRW(i) set are the comment texts corresponding to the input object i.

[0111] Furthermore, the update module is also used for:

[0112] when and and When that happens, stop updating the first evaluation factor, the second evaluation factor, and the third evaluation factor.

[0113] The technical solution provided in this application has at least the following beneficial effects:

[0114] Based on the evaluation results of the fakeness of the comment text, input object, input group, and output object, a comprehensive fakeness value that more accurately reflects fake comments is calculated. By identifying fake comments based on this comprehensive fakeness value, the fake comment detection scheme provided in this application improves the accuracy of fake comment detection. Attached Figure Description

[0115] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0116] Figure 1 This is a flowchart illustrating a method for detecting fake comments provided in an embodiment of this application.

[0117] Figure 2 This is a schematic diagram of the structure of a fake review detection system provided in an embodiment of this application.

[0118] 100 Fake Review Detection System

[0119] 11 First Dimension Evaluation Module

[0120] 12 Second Dimension Evaluation Module

[0121] 13 Third Dimension Evaluation Module

[0122] 14 Output Module

[0123] 15 Update Module Detailed Implementation

[0124] 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 in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0125] Please refer to Figure 1 To address the technical problem of low accuracy in detecting fake reviews, this application provides a method for detecting fake reviews, comprising the following steps:

[0126] S110: Get the comment text.

[0127] S120: Determine the textual and semantic features of the comment text.

[0128] S130: Based on the textual and semantic features of the comment text, evaluate the first dimension of the falsehood of the comment text and obtain the first dimension of the falsehood value of the comment text as the first evaluation factor.

[0129] It is understood that the comment text includes at least phrases, sentences, paragraphs, or documents composed of text. The text can be understood as at least one of the following: words, numbers, letters, and symbols. In the specific embodiments provided in this application, the comment text can be a review of a product, a comment on an opinion, or a message about media content. Of course, the comment text can be categorized as genuine or fake based on the veracity of its content. Fake comments are usually malicious, misleading users and influencing their judgment.

[0130] Comment texts typically possess textual and semantic features. This application uses these textual and semantic features to calculate the probability that a comment text is a fake comment, thereby identifying fake comments.

[0131] Furthermore, obtaining comment text specifically involves collecting comment texts from e-commerce platforms over a period of time, and identifying the comment texts by store name, product number, and commenter ID.

[0132] Furthermore, the text features are manifested in at least one of the following: comment text length (i.e., the number of characters contained in the comment text), text complexity (i.e., the complexity of the vocabulary used in the text), usefulness votes (i.e., the number of supporters who believe the comment is useful), early time window (i.e., whether any comment is an early comment on the product or service), similarity of comment content (i.e., the degree of similarity between any comment and other comments on the same product or service), comment reply status (i.e., the number of replies to any comment, reflecting its level of attention), coherence with other comments (i.e., the degree of consistency between any comment and other comments on the same product or service), and consistency between initial and follow-up comments (i.e., the time interval and the degree of content progression between the initial and follow-up comments). In the specific embodiments provided in this application, the process of determining the text features of the comment text is manifested as follows: using a trained algorithm model to identify the text features of the comment text, and converting the text features into text feature vector values.

[0133] Furthermore, the semantic features are used to characterize the positivity of the comment text. Considering that fake comments contain more sentiment words than real comments, i.e., fake comments are more positive or more negative in meaning than real comments, the semantic features of the comment text are an important aspect of distinguishing fake comments. In the specific embodiments provided in this application, the process of determining the semantic features of the comment text is as follows: the comment text is segmented into words to obtain a text set whose elements are word units; stop word units are filtered out from the text set to obtain a preprocessed text set; for the preprocessed text set, the Apriori algorithm is used to identify the proportion of sentiment word units; the sentiment word with the highest proportion of sentiment word units is taken as the sentiment tendency of the comment, i.e., the number of times the words "Chinese positive evaluation", "Chinese negative evaluation", "Chinese positive sentiment", and "Chinese negative sentiment" appear in the comment are calculated as the semantic feature vector value of the evaluation.

[0134] Furthermore, the first dimension of falsity evaluation can be understood as the probability that the comment text is a fake comment. Specifically, this application uses a logistic regression model to calculate the probability value of the comment text being a fake comment (i.e., the first dimension of falsity value of the comment text) based on the text feature vector value and semantic feature vector value of the comment text, and uses it as the first evaluation factor.

[0135] S140: Determine the input object for the corresponding comment text and the behavioral characteristics of the input object.

[0136] S150: Based on the behavioral characteristics of the input object, perform a second-dimensional falsity evaluation on the input object to obtain the second-dimensional falsity value of the input object, which serves as the second evaluation factor.

[0137] The input object for the corresponding comment text can be understood as the object that posts the comment text, such as an e-commerce platform user. In the application scenario provided in this application, the input object can be represented as a comment account, comment ID, or comment IP address.

[0138] Typically, the input object and the comment text are related; that is, the corresponding input object can be determined through the comment text, or the comment text published can be determined through the input object.

[0139] Of course, the input objects can be categorized as real users or fake users based on the authenticity of the comments posted. Fake users are usually malicious; they target specific output objects with specific information to quickly publish a large number of fake comments. The comment text posted by fake users is itself a fake comment, which confuses users and influences their judgment. In specific application scenarios, fake users manifest as order-boosting programs or online trolls.

[0140] The input object typically possesses behavioral characteristics. In specific application scenarios, these behavioral characteristics can manifest as at least one of the following: registration rate, registration window, rating rate, comment ratio, comment content similarity, and comment time window. This application uses the behavioral characteristics of the input object to calculate the probability that the input object is a fake user, thereby identifying fake users.

[0141] Furthermore, the registration rate of the input object can be expressed as:

[0142]

[0143] In the formula, RUR represents the registration rate of the input object, |u i | represents the number of registered accounts of the input object, and λ1 represents the threshold for a single IP address to register an account. Since there is no direct method to calculate the exact number of registered accounts of the input object, this application preferably uses IP addresses to distinguish the input objects.

[0144] Furthermore, the registration window for determining the input object can be displayed as follows:

[0145]

[0146] avg(|t i |)=avg(|max_create_timei-min_create_timei|) / user_cnt

[0147] In the formula, URW represents the time window for the registration of the input object, and avg(|t i |) represents the average time interval between two adjacent accounts of the input object, and γ1 represents the time interval threshold between two registered accounts.

[0148] Furthermore, the evaluation rate of the input object can be expressed as:

[0149]

[0150] In the formula, RR represents the evaluation rate of the input object, |u r | represents the number of comment texts published by the input object, and λ2 represents the preset first analysis threshold.

[0151] Furthermore, determining the comment ratio of the input object can be expressed as:

[0152]

[0153] In the formula, RRM represents the comment ratio of the input object, |u mi | represents the number of comment texts published by the input object for the output group, and λ3 represents the preset second analysis threshold.

[0154] It is understood that the comment text posted by the input object is directed at the output object. The output object possesses identity characteristics, and output objects with the same identity characteristics constitute an output group. In the specific application scenario provided in this application, the output object can represent a product on an e-commerce platform, and the output group can represent a shop on the e-commerce platform. The shop sells the product on the e-commerce platform.

[0155] Furthermore, determining the similarity of the comments from the input objects can be represented as follows:

[0156] RCS = avg(RMCS(u, m) i ))

[0157]

[0158] In the formula, RCS represents the similarity of the comment content of the input objects, and RMCS represents the similarity of the comment content of the input objects for the output group. avg(RMCS(u,m) i )) represents the mean of RMCS. mil This represents the l-th comment text posted by the input object for the i-th output group, r mik This represents the text of the k-th comment posted by the input object to the i-th output group. (avg(sim)) u (r mil r mik )) represents the average similarity of the comment texts published by the input object for the i-th output group. If the input object has only published one comment text, then RMCS is 0.

[0159] Furthermore, determining the comment time window of the input object can be represented as follows:

[0160] RTW = avg(RMTW(u, m) i ))

[0161]

[0162] In the formula, RTW represents the time window during which the input object posts a comment, and RMTW represents the time window during which the input object posts a comment for the output group. avg(RMTW(u, m) i )) represents the mean of RMTW. LT mil This represents the latest time that the input object has posted a comment text for the i-th output group. (FT) mik γi represents the earliest time when the input object posts a comment text for the i-th output group. γ2 represents the time interval threshold for the input object to post a comment text for the i-th output group.

[0163] Furthermore, in one specific embodiment provided in this application, based on the behavioral characteristics of the input object, a second-dimensional falsehood evaluation is performed on the input object to obtain a second-dimensional falsehood value of the input object, which serves as a second evaluation factor, and can be expressed as follows:

[0164]

[0165] In the formula, F RW This represents the probability that the input object is a fake user (i.e., the second dimension of the input object's fakeness value), and is used as the second evaluation factor. i w represents the behavioral characteristics of the input object. i This represents the weight value of the corresponding behavioral feature.

[0166] S160: Determine the input group corresponding to the input object and the group characteristics of the input group.

[0167] S170: Based on the group characteristics of the input group, evaluate the falsehood of the input group in the third dimension and obtain the falsehood value of the input group in the third dimension as the third evaluation factor.

[0168] The input group corresponding to the input object can be understood as an input group composed of input objects, that is, the group to which the input object belongs. Given that the input object is a fake user, a large number of fake users will typically post specific information targeting a specific output object to quickly publish a large number of fake comments. This application defines such user groups with similar behavioral characteristics or user groups with the same IP address as input groups.

[0169] Specifically, determining the input group corresponding to the input object can be represented as follows:

[0170] Based on the behavioral characteristics of the input objects, the input objects are clustered to generate several cluster sets; the clustering methods can be k-means algorithm, hierarchical clustering, spectral clustering, etc.

[0171] After clustering is completed, the centroids of the cluster sets are calculated (for example, using the same method as in the k-means algorithm).

[0172] For any cluster set, calculate the maximum distance from each point (i.e., the input object) in the cluster set to the center point;

[0173] Points with a distance less than half of the maximum distance are identified and grouped into an input group.

[0174] The input group possesses group characteristics. Specifically, these characteristics include group comment ratio, group comment content similarity, group comment time window, and group size ratio. This application uses these group characteristics to calculate the probability that the input group is a fake user group, thereby identifying fake user groups.

[0175] Furthermore, the group comment ratio of the input group can be represented as follows:

[0176]

[0177] In the formula, RGM represents the ratio of comments posted by the input group to the output group, |g mi | represents the number of comment texts published by the input group for the output group, and λ4 represents the preset third analysis threshold.

[0178] Furthermore, determining the similarity of group comments in the input group can be represented as follows:

[0179] GMCS(g j m i ) = avg(sim g (r mil r mik ))

[0180] In the formula, GMCS(g j m i ) represents the text similarity of comments posted by the input group to the output group, where avg(sim g (r mil r mik )) represents the average similarity of group comments among the input groups. mil This represents the l-th comment text posted by the input group to the i-th output group, r. mik This represents the text of the k-th comment posted by the input group to the i-th output group.

[0181] Furthermore, determining the time window for group comments in the input group can be represented as follows:

[0182]

[0183] In the formula, GMTW represents the time window during which the input group comments on the output group. LT g(mil) This represents the latest time that the input group can post a comment on the i-th output group. (FT) g(mik) γ represents the earliest time when the input group comments on the i-th output group. γ3 represents the time interval threshold for the input group to comment on the i-th output group.

[0184] Furthermore, determining the group size ratio of the input group can be expressed as:

[0185]

[0186] In the formula, GSR represents the group size ratio of the input group. |g i | represents the number of fake users in the i-th input group, and |max(g)| represents the maximum number of fake users across all input groups.

[0187] Furthermore, based on the group characteristics of the input group, the third-dimensional evaluation of the falsehood of the input group can be expressed as follows:

[0188] Obtain the group feature vector values ​​of the input group;

[0189] If the group feature vector value of the input group is greater than the false detection threshold, the candidate set is classified as a false user group.

[0190] Then there is

[0191]

[0192] In the formula, F RG This represents the probability that the input group is a fake user group (i.e., the third dimension of the fakeness score of the input group), and serves as the third evaluation factor. The elements of the GRW(i) set are the fake users corresponding to fake user group i; F RW (i) represents the second dimension, the falsity score, for fake user i. In other words, F RG This represents the average of the second dimension's deception score for all fake users within the fake user group.

[0193] S180: Calculate the sum of the first evaluation factor, the second evaluation factor, and the third evaluation factor using preset weight ratios to obtain the comprehensive falsehood value.

[0194] S190: When the overall falsehood value is greater than the preset threshold, the comment text is determined to be a false comment.

[0195] It is understood that the preset weight percentages respectively provide the preset weight values ​​for the first evaluation factor, the second evaluation factor, and the third evaluation factor. The preset weight values ​​of the evaluation factors are related to the credibility of the evaluation factors; generally, evaluation factors with higher credibility also have higher preset weight values.

[0196] In one specific embodiment provided in this application, the preset weight values ​​of the first evaluation factor, the second evaluation factor, and the third evaluation factor are different. For example, the preset weight value of the first evaluation factor is 0.4, the preset weight value of the second evaluation factor is 0.3, and the preset weight value of the third evaluation factor is 0.3.

[0197] Alternatively, in another specific embodiment provided in this application, the preset weight value of any evaluation factor is 1, and the preset weight value of the other evaluation factors is 0.

[0198] Of course, the preset weight values ​​of the first evaluation factor, the second evaluation factor, and the third evaluation factor can all be the same, for example, they can all be set to 1.

[0199] A comprehensive falsehood score can be calculated based on preset weighting percentages. This score integrates the falsehood evaluation results from three dimensions: the comment text itself, the commenting user, and the group the commenting user belongs to. Even if the text content of a false comment and a genuine comment are very similar, the commenting user or their group will still affect the comprehensive falsehood score. Therefore, the comprehensive falsehood score provided in this application more accurately reflects false comments. This application identifies false comments based on the comprehensive falsehood score, improving the accuracy of false comment detection.

[0200] Furthermore, considering the corresponding relationship between comment text, input object, and input group, and that the output object of the corresponding comment text also has a certain impact on the judgment of fake comments, in order to further improve the accuracy of fake comment detection, the fake comment detection method provided in this application also includes:

[0201] Determine the output object for the corresponding comment text, and the output group for the corresponding output object;

[0202] Based on the textual and semantic features of the comment text, the behavioral features of the input objects, the behavioral features of the input groups, the output objects, and the output groups, construct a comment relationship graph that represents the relationship between the comment text, the output objects, and the input objects;

[0203] Based on the comment relationship diagram, update the first evaluation factor, the second evaluation factor, and the third evaluation factor;

[0204] The comment relationship graph is represented as G = (V, E);

[0205] In the formula, V = {v1, v2, ..., v} n} is a set of vertices;

[0206] E={e ij For i, j∈{1,...,n}} is the set of edges;

[0207] And vertex v i With vertex v j There is an edge e between them ij ;

[0208] v iIt has type τ(v)∈A, A={RW,R,P};

[0209] In the formula, RW represents the input object, R represents the comment text, and P represents the output object;

[0210] Vertices of type RW input object have attributes The input group that represents the corresponding input object;

[0211] Vertices of type P in the output object have an attribute ψ, which represents the output group of the corresponding output object.

[0212] It should be reiterated that the comment text posted by the input object is typically directed at the output object. The output object possesses identity characteristics, and output objects with the same identity characteristics constitute an output group. In the specific application scenario provided in this application, the output object can represent a product on an e-commerce platform, and the output group can represent a shop on the e-commerce platform. The shop sells the product on the e-commerce platform.

[0213] The product itself, as the source of the review text, also influences the identification of fake reviews. This is because a large number of fake reviews on a product could indicate that the store is using fake review programs or hiring online trolls to post a large number of positive reviews, thus confusing users and affecting their judgment. Therefore, identifying fake reviews from both product and store perspectives will improve detection accuracy.

[0214] To this end, this application determines the output object of the corresponding comment text and the output group of the corresponding output object; and constructs a comment relationship graph representing the relationship between the comment text, the output object, and the input object. The comment relationship graph G consists of a vertex set V(G) and an edge set E(G). The comment text, the output object, and the input object are the vertices in the comment relationship graph, and the relationship between the comment text, the output object, and the input object is the edge in the comment relationship graph.

[0215] For example, a comment relationship graph can be represented by a graph G, defined as G = (V, E). Where V = {v1, v2, ..., v...} n} is the set of vertices, E = {e ij Let $\mathbf{i,j}} be the set of edges. i It has a type τ(v)∈A, A={RW, R, P}; where RW represents the input object, R represents the comment text, and P represents the output object; the vertices of type RW of the input object have attributes The input group represents the corresponding input object; the vertices of the output object P type have the attribute ψ, which represents the corresponding output object's output group.

[0216] If there is an input object v i Published about output object v j Comment text v k Then vertex v i and v k There is an edge e between them ik v j and v k There is an edge e between them jk .

[0217] Intuitively, a comment relationship diagram can map the relationships between comment text, output objects, and input objects.

[0218] Furthermore, based on the aforementioned comment relationship graph, the first evaluation factor, the second evaluation factor, and the third evaluation factor are updated, specifically including:

[0219] With F R This represents the first evaluation factor before the update, denoted by F. RW This represents the second evaluation factor before the update, denoted by F. GRW This indicates the third evaluation factor before the update;

[0220] by This represents the updated first evaluation factor, with This represents the updated second evaluation factor, in order to This represents the updated third evaluation factor;

[0221] Then there is

[0222]

[0223]

[0224]

[0225] In the formula, F PR (i) represents the falsity value of the output object corresponding to comment text i;

[0226] F RWR (i) represents the falsity value of the input object corresponding to comment text i;

[0227] F PRW (i) represents the falsity value of the output object corresponding to input object i;

[0228] F RRW (i) represents the falsity value of the comment text corresponding to input object i;

[0229] The elements of the GR(i) set are the comment texts corresponding to the input group i;

[0230] The elements of the GRW(i) set are the input objects corresponding to input group i;

[0231] The elements of the GP(i) set are the output objects corresponding to the input group i;

[0232] F P (i) represents the spuriousness value of output object i;

[0233] and

[0234]

[0235] The elements of the PR(i) set are the output objects corresponding to comment text i;

[0236]

[0237] The elements of the RWR(i) set are the output objects corresponding to the input object i;

[0238]

[0239] The elements of the PRW(i) set are the output objects corresponding to the input object i;

[0240]

[0241] The elements of the RRW(i) set are the comment texts corresponding to the input object i.

[0242] Furthermore, the updating of the first evaluation factor, the second evaluation factor, and the third evaluation factor based on the comment relationship graph can be iteratively updated multiple times. In one embodiment provided in this application, the iterative updating of the first evaluation factor, the second evaluation factor, and the third evaluation factor is stopped after 10 iterations. In another embodiment provided in this application, when and and When that happens, stop updating the first evaluation factor, the second evaluation factor, and the third evaluation factor.

[0243] Furthermore, considering that the product itself is the output object of the review text, it also has a certain impact on the judgment of fake reviews. To further improve the accuracy of fake review detection, the fake review detection method provided in this application also includes:

[0244] Based on the identity characteristics of the output object, the fourth dimension of the falseness evaluation is performed on the output object to obtain the fourth dimension falseness value of the output object, which is used as the fourth evaluation factor;

[0245] The sum of the first evaluation factor, the second evaluation factor, the third evaluation factor, and the fourth evaluation factor is calculated using preset weight ratios to obtain the overall falseness value.

[0246] It needs to be emphasized again that the output object can be understood as the object to which the comment text is addressed. In the specific application scenario provided in this application, the output object can be represented as a product on an e-commerce platform. Typically, the output object and the comment text are related; that is, the corresponding output object can be determined through the comment text, and vice versa.

[0247] Of course, the output objects can be categorized as genuine or malicious products based on the veracity of the corresponding review content. The output objects also have a certain impact on the judgment of fake reviews. This is because if a product's review text contains a large number of fake reviews, it's possible that the store is using order-boosting programs or hiring online trolls to publish a large number of positive reviews, thereby confusing users and affecting their judgment. Therefore, identifying fake reviews from a product perspective can improve detection accuracy. Here, products corresponding to a large number of positive reviews published by order-boosting programs or hired online trolls are considered malicious products.

[0248] The following describes the specific implementation process of evaluating the output object based on its identity characteristics, obtaining the fourth-dimensional falsity value of the output object, and using it as the fourth evaluation factor:

[0249] The output object has identity characteristics, such as product number and product name.

[0250] Based on the identity characteristics of the output objects and the relationship between the output objects and the comment texts, the set of comment texts corresponding to output objects with the same identity characteristics can be determined.

[0251] Then there is

[0252]

[0253] In the formula, F P This represents the probability value that the output object is a malicious product (i.e., the fourth dimension of the output object's falsity value), serving as the fourth evaluation factor. The elements of the RP(i) set are the comment texts corresponding to output object i; F R (i) represents the first dimension of the falsity score of the comment text i corresponding to the output object. In other words, F P The average of the first dimension of the falsehood score of the comment text corresponding to malicious products.

[0254] Furthermore, based on the introduction of a fourth evaluation factor, preset weight values ​​are given for the first, second, third, and fourth evaluation factors, respectively. A comprehensive falsehood score can be calculated based on these preset weight values. This comprehensive falsehood score integrates the falsehood evaluation results from four dimensions: the review text itself, the reviewer's reputation, the reputation of the reviewer's group, and the product's own reputation. Even if the text content of a false review and a genuine review are very similar, the reviewer's reputation, the reviewer's group reputation, or the product's own reputation will still affect the comprehensive falsehood score. Therefore, the comprehensive falsehood score provided in this application more accurately reflects false reviews. This application identifies false reviews based on the comprehensive falsehood score, improving the accuracy of false review detection.

[0255] Furthermore, based on the introduction of a fourth evaluation factor, the fake review detection method provided in this application also includes:

[0256] Update the fourth evaluation factor based on the aforementioned comment relationship diagram;

[0257] With F P This represents the fourth evaluation factor before the update. This represents the updated fourth evaluation factor;

[0258] Then there is

[0259]

[0260] F RP (i) represents the degree of falsity of the comment text corresponding to output object i;

[0261] F RWP (i) represents the falsity of the input object corresponding to output object i;

[0262] F R (i) represents the first dimension of the falsity value of comment text i;

[0263] and

[0264]

[0265] The elements of the RP(i) set are the comment text corresponding to output object i;

[0266]

[0267] The elements of the RP(i) set are the input objects corresponding to the output object i.

[0268] In summary, the fake review detection method provided in this application calculates a more accurate comprehensive fake review value based on the fakeness evaluation results across four dimensions: review text, input object, input group, and output object. By identifying fake reviews based on this comprehensive fake review value, the fake review detection method provided in this application improves the accuracy of fake review detection.

[0269] Please refer to Figure 2 To support methods for detecting fake reviews, this application also provides a fake review detection system 100, comprising:

[0270] The first dimension evaluation module 11 is used to acquire the comment text; it is also used to determine the text features and semantic features of the comment text; and it is also used to evaluate the first dimension of the comment text based on the text features and semantic features of the comment text, so as to obtain the first dimension of the comment text as the first evaluation factor.

[0271] The second-dimensional evaluation module 12 is used to determine the input object of the corresponding comment text and the behavioral characteristics of the input object; it is also used to evaluate the second-dimensional falseness of the input object based on the behavioral characteristics of the input object, and obtain the second-dimensional falseness value of the input object as the second evaluation factor.

[0272] The third-dimensional evaluation module 13 is used to determine the input group of the corresponding input object and the group characteristics of the input group; it is also used to evaluate the falseness of the input group in the third dimension based on the group characteristics of the input group, and obtain the falseness value of the input group in the third dimension as the third evaluation factor.

[0273] Output module 14 is used to calculate the sum of the first evaluation factor, the second evaluation factor, and the third evaluation factor with preset weight ratios to obtain the comprehensive falsehood value; it is also used to determine that the comment text is a false comment when the comprehensive falsehood value is greater than a preset threshold.

[0274] The first dimension evaluation module 11 is used to acquire the comment text; it is also used to determine the textual and semantic features of the comment text; and it is also used to evaluate the first dimension of the comment text based on the textual and semantic features of the comment text, so as to obtain the first dimension of the comment text as the first evaluation factor.

[0275] It is understood that the comment text includes at least phrases, sentences, paragraphs, or documents composed of text. The text can be understood as at least one of the following: words, numbers, letters, and symbols. In the specific embodiments provided in this application, the comment text can be a review of a product, a comment on an opinion, or a message about media content. Of course, the comment text can be categorized as genuine or fake based on the veracity of its content. Fake comments are usually malicious, misleading users and influencing their judgment.

[0276] Comment texts typically possess textual and semantic features. The first-dimensional evaluation module 11 calculates the probability that a comment text is a fake comment based on its textual and semantic features, thereby identifying fake comments.

[0277] Furthermore, the first dimension evaluation module 11 specifically acquires comment text by collecting comment texts from e-commerce platforms over a period of time, and identifies the comment texts by store name, product number, and commenter ID.

[0278] Furthermore, the text features are manifested in at least one of the following: comment text length (i.e., the number of characters contained in the comment text), text complexity (i.e., the complexity of the vocabulary used in the text), usefulness votes (i.e., the number of supporters who believe the comment is useful), early time window (i.e., whether any comment is an early comment of the product or service), similarity of comment content (i.e., the degree of similarity between any comment and other comments of the same product or service), comment reply status (i.e., the number of replies to any comment, reflecting its level of attention), coherence with other comments (i.e., the degree of consistency between any comment and other comments of the same product or service), and consistency between initial and follow-up comments (i.e., the time interval and content progression between initial and follow-up comments). In the specific embodiments provided in this application, the process of the first dimension evaluation module 11 determining the text features of the comment text is manifested as follows: using a trained algorithm model to identify the text features of the comment text, and converting the text features into text feature vector values.

[0279] Furthermore, the semantic features are used to characterize the positivity of the comment text. Considering that fake comments contain more sentiment words than real comments, i.e., the semantics of fake comments are more positive or more negative than real comments, the semantic features of the comment text are an important aspect of distinguishing fake comments. In the specific embodiments provided in this application, the process of the first dimension evaluation module 11 determining the semantic features of the comment text is as follows: the comment text is segmented into words to obtain a text set whose elements are word units; stop word units in the text set are filtered out to obtain a preprocessed text set; for the preprocessed text set, the Apriori algorithm is used to identify the proportion of sentiment word units; the sentiment word with the highest proportion of sentiment word units is taken as the sentiment tendency of the comment, i.e., the number of times the words "Chinese positive evaluation", "Chinese negative evaluation", "Chinese positive sentiment", and "Chinese negative sentiment" appear in the comment are calculated as the semantic feature vector value of the evaluation.

[0280] Furthermore, the first dimension of falsehood assessment can be understood as the probability that the comment text is a false comment. Specifically, the first dimension assessment module 11 calculates the probability value (i.e., the first dimension of falsehood value) of the comment text as a false comment using a logistic regression model based on the text feature vector value and semantic feature vector value of the comment text, and uses it as the first assessment factor.

[0281] The second-dimensional evaluation module 12 is used to determine the input object of the corresponding comment text and the behavioral characteristics of the input object; it is also used to evaluate the second-dimensional falsity of the input object based on the behavioral characteristics of the input object, and obtain the second-dimensional falsity value of the input object as the second evaluation factor.

[0282] The input object for the corresponding comment text can be understood as the object that posts the comment text, such as an e-commerce platform user. In the application scenario provided in this application, the input object can be represented as a comment account, comment ID, or comment IP address.

[0283] Typically, the input object and the comment text are related. The second-dimensional evaluation module 12 can determine the corresponding input object through the comment text, or it can determine the comment text published by the input object.

[0284] Of course, the input objects can be categorized as real users or fake users based on the authenticity of the comments posted. Fake users are usually malicious; they target specific output objects with specific information to quickly publish a large number of fake comments. The comment text posted by fake users is itself a fake comment, which confuses users and influences their judgment. In specific application scenarios, fake users manifest as order-boosting programs or online trolls.

[0285] The input object typically possesses behavioral characteristics. In specific application scenarios, these behavioral characteristics can manifest as at least one of the following: registration rate, registration window, rating rate, comment ratio, comment content similarity, and comment time window. The second-dimensional evaluation module 12 calculates the probability that the input object is a fake user based on its behavioral characteristics, thereby identifying fake users.

[0286] Furthermore, the second-dimensional evaluation module 12 determines that the registration rate of the input object can be expressed as follows:

[0287]

[0288] In the formula, RUR represents the registration rate of the input object, |u i | represents the number of registered accounts of the input object, and λ1 represents the threshold for a single IP address to register an account. Since there is no direct method to calculate the exact number of registered accounts of the input object, this application preferably uses IP addresses to distinguish the input objects.

[0289] Furthermore, the registration window of the input object determined by the second-dimensional evaluation module 12 can be represented as follows:

[0290]

[0291] avg(|t i |)=avg(|max_create_timei-min_create_timei|) / user_cnt

[0292] In the formula, URW represents the time window for the registration of the input object, and avg(|t i |) represents the average time interval between two adjacent accounts of the input object, and γ1 represents the time interval threshold between two registered accounts.

[0293] Furthermore, the second-dimensional evaluation module 12 determines that the evaluation rate of the input object can be expressed as follows:

[0294]

[0295] In the formula, RR represents the evaluation rate of the input object, |u r | represents the number of comment texts published by the input object, and λ2 represents the preset first analysis threshold.

[0296] Furthermore, the second-dimensional evaluation module 12 determines that the comment ratio of the input object can be expressed as follows:

[0297]

[0298] In the formula, RRM represents the comment ratio of the input object, |u mi| represents the number of comment texts published by the input object for the output group, and λ3 represents the preset second analysis threshold.

[0299] It is understood that the comment text posted by the input object is directed at the output object. The output object possesses identity characteristics, and output objects with the same identity characteristics constitute an output group. In the specific application scenario provided in this application, the output object can represent a product on an e-commerce platform, and the output group can represent a shop on the e-commerce platform. The shop sells the product on the e-commerce platform.

[0300] Furthermore, the second-dimensional evaluation module 12 determines that the similarity of the comment content of the input objects can be expressed as follows:

[0301] RCS = avg(RMCS(u, m) i ))

[0302]

[0303] In the formula, RCS represents the similarity of the comment content of the input objects, and RMCS represents the similarity of the comment content of the input objects for the output group. avg(RMCS(u,m) i )) represents the mean of RMCS. mil This represents the l-th comment text posted by the input object for the i-th output group, r mik This represents the text of the k-th comment posted by the input object to the i-th output group. (avg(sim)) u (r mil r mik )) represents the average similarity of the comment texts published by the input object for the i-th output group. If the input object has only published one comment text, then RMCS is 0.

[0304] Furthermore, the second-dimensional evaluation module 12 determines that the comment time window of the input object can be represented as follows:

[0305] RTW = avg(RMTW(u, m) i ))

[0306]

[0307] In the formula, RTW represents the time window during which the input object posts a comment, and RMTW represents the time window during which the input object posts a comment for the output group. avg(RMTW(u, m) i )) represents the mean of RMTW. LT mil This represents the latest time that the input object has posted a comment text for the i-th output group. (FT) mikγi represents the earliest time when the input object posts a comment text for the i-th output group. γ2 represents the time interval threshold for the input object to post a comment text for the i-th output group.

[0308] Furthermore, in one specific embodiment provided in this application, the second-dimensional evaluation module 12 evaluates the input object's falseness in the second dimension based on the input object's behavioral characteristics, obtaining the input object's second-dimensional falseness value, which serves as the second evaluation factor, and can be expressed as follows:

[0309]

[0310] In the formula, F RW This represents the probability that the input object is a fake user (i.e., the second dimension of the input object's fakeness value), and is used as the second evaluation factor. i w represents the behavioral characteristics of the input object. i This represents the weight value of the corresponding behavioral feature.

[0311] The third-dimensional evaluation module 13 is used to determine the input group corresponding to the input object and the group characteristics of the input group; it is also used to evaluate the falseness of the input group in the third dimension based on the group characteristics of the input group, and obtain the falseness value of the input group in the third dimension as the third evaluation factor.

[0312] The input group corresponding to the input object can be understood as an input group composed of input objects, that is, the group to which the input object belongs. Given that the input object is a fake user, a large number of fake users will typically post specific information targeting a specific output object to quickly publish a large number of fake comments. This application defines such user groups with similar behavioral characteristics or user groups with the same IP address as input groups.

[0313] Specifically, the third-dimensional evaluation module 13 determines the input group corresponding to the input object, which can be represented as follows:

[0314] Based on the behavioral characteristics of the input objects, the input objects are clustered to generate several cluster sets; the clustering methods can be k-means algorithm, hierarchical clustering, spectral clustering, etc.

[0315] After clustering is completed, the centroids of the cluster sets are calculated (for example, using the same method as in the k-means algorithm).

[0316] For any cluster set, calculate the maximum distance from each point (i.e., the input object) in the cluster set to the center point;

[0317] Points with a distance less than half of the maximum distance are identified and grouped into an input group.

[0318] The input group possesses group characteristics. Specifically, these characteristics include group comment ratio, group comment content similarity, group comment time window, and group size ratio. The third-dimensional evaluation module 13 calculates the probability that the input group is a fake user group based on these group characteristics, thereby identifying fake user groups.

[0319] Furthermore, the third-dimensional evaluation module 13 determines that the group comment ratio of the input group can be expressed as follows:

[0320]

[0321] In the formula, RGM represents the ratio of comments posted by the input group to the output group, |g mi | represents the number of comment texts published by the input group for the output group, and λ4 represents the preset third analysis threshold.

[0322] Furthermore, the third-dimensional evaluation module 13 determines that the similarity of the group comments content of the input group can be expressed as follows:

[0323] GMCS(g j m i ) = avg(sim g (r mil r mik ))

[0324] In the formula, GMCS(g j m i ) represents the text similarity of comments posted by the input group to the output group, where avg(sim g (r mil r mik )) represents the average similarity of group comments among the input groups. mil This represents the l-th comment text posted by the input group to the i-th output group, r. mik This represents the text of the k-th comment posted by the input group to the i-th output group.

[0325] Furthermore, the third-dimensional evaluation module 13 determines that the group comment time window of the input group can be represented as follows:

[0326]

[0327] In the formula, GMTW represents the time window during which the input group comments on the output group. LT g(mil) This represents the latest time that the input group can post a comment on the i-th output group. (FT) g(mik) γ represents the earliest time when the input group comments on the i-th output group. γ3 represents the time interval threshold for the input group to comment on the i-th output group.

[0328] Furthermore, the third-dimensional evaluation module 13 determines that the group size ratio of the input group can be expressed as follows:

[0329]

[0330] In the formula, GSR represents the group size ratio of the input group. |g i | represents the number of fake users in the i-th input group, and |max(g)| represents the maximum number of fake users across all input groups.

[0331] Furthermore, the third-dimensional evaluation module 13, based on the group characteristics of the input group, performs a third-dimensional falsehood evaluation of the input group, which can be expressed as follows:

[0332] Obtain the group feature vector values ​​of the input group;

[0333] If the group feature vector value of the input group is greater than the false detection threshold, the candidate set is classified as a false user group.

[0334] Then there is

[0335]

[0336] In the formula, F RG This represents the probability that the input group is a fake user group (i.e., the third dimension of the fakeness score of the input group), and serves as the third evaluation factor. The elements of the GRW(i) set are the fake users corresponding to fake user group i; F RW (i) represents the second dimension, the falsity score, for fake user i. In other words, F RG This represents the average of the second dimension's deception score for all fake users within the fake user group.

[0337] Output module 14 is used to calculate the sum of the first evaluation factor, the second evaluation factor, and the third evaluation factor with preset weight ratios to obtain the comprehensive falsehood value; it is also used to determine that the comment text is a false comment when the comprehensive falsehood value is greater than a preset threshold.

[0338] It is understood that the preset weight percentages respectively provide the preset weight values ​​for the first evaluation factor, the second evaluation factor, and the third evaluation factor. The preset weight values ​​of the evaluation factors are related to the credibility of the evaluation factors; generally, evaluation factors with higher credibility also have higher preset weight values.

[0339] In one specific embodiment provided in this application, the preset weight values ​​of the first evaluation factor, the second evaluation factor, and the third evaluation factor are different. For example, the preset weight value of the first evaluation factor is 0.4, the preset weight value of the second evaluation factor is 0.3, and the preset weight value of the third evaluation factor is 0.3.

[0340] Alternatively, in another specific embodiment provided in this application, the preset weight value of any evaluation factor is 1, and the preset weight value of the other evaluation factors is 0.

[0341] Of course, the preset weight values ​​of the first evaluation factor, the second evaluation factor, and the third evaluation factor can all be the same, for example, they can all be set to 1.

[0342] The output module 14 calculates a comprehensive falsehood score based on preset weighting ratios. This comprehensive falsehood score integrates the falsehood evaluation results from three dimensions: the comment text itself, the commenting user, and the group to which the commenting user belongs. Even if the text content of a false comment and a genuine comment are very similar, the commenting user or the group to which the commenting user belongs will still affect the comprehensive falsehood score. Therefore, the comprehensive falsehood score provided in this application more accurately reflects false comments. The false comment detection system 100 identifies false comments based on the comprehensive falsehood score, improving the accuracy of false comment detection.

[0343] Furthermore, considering the correspondence between comment text, input object, and input group, and that the output object corresponding to the comment text also has a certain impact on the judgment of fake comments, in order to further improve the accuracy of fake comment detection, the fake comment detection system 100 provided in this application also includes:

[0344] The update module 15 is used to determine the output object of the corresponding comment text and the output group of the corresponding output object; it is also used to construct a comment relationship graph representing the relationship between the comment text, the output object, and the input object based on the textual and semantic features of the comment text, the behavioral features of the input object, the behavioral features of the input group, the output object, and the output group; it is also used to update the first evaluation factor, the second evaluation factor, and the third evaluation factor based on the comment relationship graph.

[0345] The comment relationship graph is represented as G = (V, E);

[0346] In the formula, V = {v1, v2, ..., v} n} is a set of vertices;

[0347] E={e ij For i, j∈{1,...,n}} is the set of edges;

[0348] And vertex vi With vertex v j There is an edge e between them ij ;

[0349] v i It has type τ(v)∈A, A={RW,R,P};

[0350] In the formula, RW represents the input object, R represents the comment text, and P represents the output object;

[0351] Vertices of type RW input object have attributes The input group that represents the corresponding input object;

[0352] Vertices of type P in the output object have an attribute ψ, which represents the output group of the corresponding output object.

[0353] It should be reiterated that the comment text posted by the input object is typically directed at the output object. The output object possesses identity characteristics, and output objects with the same identity characteristics constitute an output group. In the specific application scenario provided in this application, the output object can represent a product on an e-commerce platform, and the output group can represent a shop on the e-commerce platform. The shop sells the product on the e-commerce platform.

[0354] The product itself, as the source of the review text, also influences the identification of fake reviews. This is because a large number of fake reviews on a product could indicate that the store is using fake review programs or hiring online trolls to post a large number of positive reviews, thus confusing users and affecting their judgment. Therefore, identifying fake reviews from both product and store perspectives will improve detection accuracy.

[0355] To this end, update module 15 determines the output object of the corresponding comment text and the output group of the corresponding output object; and constructs a comment relationship graph representing the relationship between the comment text, the output object, and the input object. The comment relationship graph (Graph) G consists of a vertex set V(G) and an edge set E(G). The comment text, the output object, and the input object are the vertices in the comment relationship graph, and the relationship between the comment text, the output object, and the input object is the edge in the comment relationship graph.

[0356] For example, a comment relationship graph can be represented by a graph G, defined as G = (V, E). Where V = {v1, v2, ..., v...} n} is the set of vertices, E = {e ij Let $\mathbf{i,j}} be the set of edges. iIt has a type τ(v)∈A, A={RW, R, P}; where RW represents the input object, R represents the comment text, and P represents the output object; the vertices of type RW of the input object have attributes The input group represents the corresponding input object; the vertices of the output object P type have the attribute ψ, which represents the corresponding output object's output group.

[0357] If there is an input object v i Published about output object v j Comment text v k Then vertex v i and v k There is an edge e between them ik v j and v k There is an edge e between them jk .

[0358] Intuitively, a comment relationship diagram can map the relationships between comment text, output objects, and input objects.

[0359] Furthermore, the update module 15 updates the first evaluation factor, the second evaluation factor, and the third evaluation factor according to the comment relationship graph, specifically including:

[0360] With F R This represents the first evaluation factor before the update, denoted by F. RW This represents the second evaluation factor before the update, denoted by F. GRW This indicates the third evaluation factor before the update;

[0361] by This represents the updated first evaluation factor, with This represents the updated second evaluation factor, in order to This represents the updated third evaluation factor;

[0362] Then there is

[0363]

[0364]

[0365]

[0366] In the formula, F PR (i) represents the falsity value of the output object corresponding to comment text i;

[0367] F RWR (i) represents the falsity value of the input object corresponding to comment text i;

[0368] F PRW (i) represents the falsity value of the output object corresponding to input object i;

[0369] F RRW (i) represents the falsity value of the comment text corresponding to input object i;

[0370] The elements of the GR(i) set are the comment texts corresponding to the input group i;

[0371] The elements of the GRW(i) set are the input objects corresponding to input group i;

[0372] The elements of the GP(i) set are the output objects corresponding to the input group i;

[0373] F P (i) represents the spuriousness value of output object i;

[0374] and

[0375]

[0376] The elements of the PR(i) set are the output objects corresponding to comment text i;

[0377]

[0378] The elements of the RWR(i) set are the output objects corresponding to the input object i;

[0379]

[0380] The elements of the PRW(i) set are the output objects corresponding to the input object i;

[0381]

[0382] The elements of the RRW(i) set are the comment texts corresponding to the input object i.

[0383] Furthermore, the update module 15 can iteratively update the first evaluation factor, the second evaluation factor, and the third evaluation factor multiple times based on the comment relationship graph. In one embodiment provided in this application, the update module 15 stops iteratively updating after the first evaluation factor, the second evaluation factor, and the third evaluation factor have been iteratively updated 10 times. In another embodiment provided in this application, when and and When this happens, update module 15 stops updating the first evaluation factor, the second evaluation factor, and the third evaluation factor.

[0384] Furthermore, considering that the product is the output object of the review text, it also has a certain impact on the judgment of fake reviews. To further improve the accuracy of detecting fake reviews, the fake review detection system 100 provided in this application also includes:

[0385] The fourth-dimensional evaluation module is used to evaluate the falseness of the output object in the fourth dimension based on the identity characteristics of the output object, and obtain the fourth-dimensional falseness value of the output object as the fourth evaluation factor.

[0386] Output module 14 is also used to calculate the sum of the first evaluation factor, the second evaluation factor, the third evaluation factor, and the fourth evaluation factor with preset weight ratios to obtain the comprehensive falseness value.

[0387] It needs to be emphasized again that the output object can be understood as the object to which the comment text is addressed. In the specific application scenario provided in this application, the output object can be represented as a product on an e-commerce platform. Typically, the output object and the comment text are related; the fourth-dimensional evaluation module can determine the corresponding output object through the comment text, and vice versa.

[0388] Of course, the output objects can be categorized as genuine or malicious products based on the veracity of the corresponding review content. The output objects also have a certain impact on the judgment of fake reviews. This is because if a product's review text contains a large number of fake reviews, it's possible that the store is using order-boosting programs or hiring online trolls to publish a large number of positive reviews, thereby confusing users and affecting their judgment. Therefore, identifying fake reviews from a product perspective can improve detection accuracy. Here, products corresponding to a large number of positive reviews published by order-boosting programs or hired online trolls are considered malicious products.

[0389] The following describes the specific implementation process of the fourth-dimensional evaluation module, which evaluates the output object's falsity based on its identity characteristics, and obtains the fourth-dimensional falsity value of the output object as the fourth evaluation factor:

[0390] The output object has identity characteristics, such as product number and product name.

[0391] The fourth-dimensional evaluation module can determine the set of comment texts corresponding to output objects with the same identity characteristics based on the identity characteristics of the output objects and the relationship between the output objects and the comment texts.

[0392] Then there is

[0393]

[0394] In the formula, F PThis represents the probability value that the output object is a malicious product (i.e., the fourth dimension of the output object's falsity value), serving as the fourth evaluation factor. The elements of the RP(i) set are the comment texts corresponding to output object i; F R (i) represents the first dimension of the falsity score of the comment text i corresponding to the output object. In other words, F P The average of the first dimension of the falsehood score of the comment text corresponding to malicious products.

[0395] Furthermore, based on the introduction of a fourth evaluation factor, preset weight ratios are given for the first evaluation factor, the second evaluation factor, the third evaluation factor, and the fourth evaluation factor, respectively. The output module 14 can calculate the comprehensive falsehood score based on the preset weight ratios. The comprehensive falsehood score integrates the falsehood evaluation results from four dimensions: the review text itself, the reviewer's reputation, the reputation of the reviewer's group, and the product's own reputation. Even if the text content of a false review and a genuine review are very similar, the reviewer's reputation, the reviewer's group reputation, or the product's own reputation will still affect the comprehensive falsehood score. Therefore, the comprehensive falsehood score provided by this application more accurately reflects false reviews. This application identifies false reviews based on the comprehensive falsehood score, improving the accuracy of false review detection.

[0396] Furthermore, based on the introduction of the fourth evaluation factor, the update module 15 is also used for:

[0397] Update the fourth evaluation factor based on the aforementioned comment relationship diagram;

[0398] With F P This represents the fourth evaluation factor before the update. This represents the updated fourth evaluation factor;

[0399] Then there is

[0400]

[0401] F RP (i) represents the degree of falsity of the comment text corresponding to output object i;

[0402] F RWP (i) represents the falsity of the input object corresponding to output object i;

[0403] and

[0404]

[0405] The elements of the RP(i) set are the comment text corresponding to output object i;

[0406]

[0407] The elements of the RP(i) set are the input objects corresponding to the output object i.

[0408] In summary, the fake review detection system 100 provided in this application calculates a more accurate comprehensive fake review value based on the fakeness evaluation results across four dimensions: review text, input object, input group, and output object. By identifying fake reviews based on this comprehensive fake review value, the fake review detection system 100 provided in this application improves the accuracy of fake review detection.

[0409] It should be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0410] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0411] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for detecting fake reviews, characterized in that, Includes the following steps: Get the comment text; Determine the textual and semantic features of the comment text; Based on the textual and semantic features of the comment text, the first dimension of the falsehood is evaluated, and the first dimension of the falsehood value of the comment text is obtained as the first evaluation factor; Identify the input object for the corresponding comment text and the behavioral characteristics of the input object; Based on the behavioral characteristics of the input object, a second-dimensional falseness evaluation is performed on the input object to obtain the second-dimensional falseness value of the input object, which is used as the second evaluation factor; Determine the input group corresponding to the input object, and the group characteristics of the input group; Based on the group characteristics of the input group, a third-dimensional falsehood evaluation is performed on the input group to obtain the third-dimensional falsehood value of the input group, which is used as the third evaluation factor; The sum of the first evaluation factor, the second evaluation factor, and the third evaluation factor is calculated using preset weighting percentages to obtain the overall falsehood value. When the overall falsehood score exceeds a preset threshold, the comment text is determined to be a false comment; This also includes updating the first evaluation factor, the second evaluation factor, and the third evaluation factor, specifically including: by This represents the first evaluation factor before the update, with This represents the second evaluation factor before the update, in order to This indicates the third evaluation factor before the update; by This represents the updated first evaluation factor, with This represents the updated second evaluation factor, in order to This represents the updated third evaluation factor; Then there is ； ； ； In the formula, Indicates the corresponding comment text The output object's falsity value; Indicates the corresponding comment text The falsity value of the input object; Indicates the corresponding input object The output object's falsity value; Indicates the corresponding input object The falsity value of the comment text; The elements of the set are the input group The corresponding comment text; The elements of the set are the input group The corresponding input object; The elements of the set are the input group The corresponding output object; Indicates the output object The degree of falsity; and ； The elements of the set are comment text. The corresponding output object; ； The elements of the set are the input objects. The corresponding output object; ； The elements of the set are the corresponding input objects. The corresponding output object; ； The elements of the set are the corresponding input objects. The corresponding comment text; The method further includes: Based on the identity characteristics of the output object, the fourth dimension of the falseness evaluation is performed on the output object to obtain the fourth dimension falseness value of the output object, which is used as the fourth evaluation factor; The method further includes: Update the fourth evaluation factor; by This represents the fourth evaluation factor before the update, in... This represents the updated fourth evaluation factor; Then there is ； Indicates the corresponding output object The degree of falsity in the comment text; Indicates the corresponding output object The degree of falsity of the input object; and ； The elements of the collection are the output objects. The corresponding comment text; ； The elements of the collection are the output objects. The corresponding input object.

2. The method for detecting fake reviews as described in claim 1, characterized in that, The method further includes: Determine the output object for the corresponding comment text, and the output group for the corresponding output object; Based on the textual and semantic features of the comment text, the behavioral features of the input objects, the behavioral features of the input groups, the output objects, and the output groups, construct a comment relationship graph that represents the relationship between the comment text, the output objects, and the input objects; Based on the comment relationship diagram, update the first evaluation factor, the second evaluation factor, and the third evaluation factor; The comment relationship graph is represented as follows: ; In the formula It is a set of vertices; It is a set of edges; And the vertex With vertex There are edges ; Has type , ; In the formula Indicates the input object, This represents the comment text. Indicates the output object; Input object Vertices of type have attributes , representing the input group of the corresponding input object; Output object Vertices of type have attributes , which represents the output group of the corresponding output object.

3. The method for detecting fake reviews as described in claim 2, characterized in that, The method further includes: when ,and ,and When that happens, stop updating the first evaluation factor, the second evaluation factor, and the third evaluation factor.

4. A fake review detection system, characterized in that, include: The first-dimensional evaluation module is used to obtain the comment text; It is also used to determine the textual and semantic features of comment text; It is also used to evaluate the first dimension of the falsehood of the comment text based on its textual and semantic features, and to obtain the first dimension falsehood value of the comment text as the first evaluation factor; The second-dimensional evaluation module is used to determine the input object of the corresponding comment text and the behavioral characteristics of the input object; It is also used to evaluate the falseness of the input object in the second dimension based on the behavioral characteristics of the input object, and obtain the falseness value of the input object in the second dimension as the second evaluation factor; The third-dimensional evaluation module is used to determine the input group of the corresponding input object and the group characteristics of the input group; It is also used to evaluate the falseness of the input group in the third dimension based on the group characteristics of the input group, and obtain the falseness value of the input group in the third dimension as the third evaluation factor; The output module is used to calculate the sum of the first evaluation factor, the second evaluation factor, and the third evaluation factor with preset weight ratios to obtain the comprehensive falseness value; It is also used to determine that a comment text is a fake comment when the overall falsehood value is greater than a preset threshold; This also includes updating the first evaluation factor, the second evaluation factor, and the third evaluation factor, specifically including: by This represents the first evaluation factor before the update, with This represents the second evaluation factor before the update, in order to This indicates the third evaluation factor before the update; by This represents the updated first evaluation factor, with This represents the updated second evaluation factor, in order to This represents the updated third evaluation factor; Then there is ； ； ； In the formula, Indicates the corresponding comment text The output object's falsity value; Indicates the corresponding comment text The falsity value of the input object; Indicates the corresponding input object The output object's falsity value; Indicates the corresponding input object The falsity value of the comment text; The elements of the set are the input group The corresponding comment text; The elements of the set are the input group The corresponding input object; The elements of the set are the input group The corresponding output object; Indicates the output object The degree of falsity; and ； The elements of the set are comment text. The corresponding output object; ； The elements of the set are the input objects. The corresponding output object; ； The elements of the set are the corresponding input objects. The corresponding output object; ； The elements of the set are the corresponding input objects. The corresponding comment text; The system also includes: Based on the identity characteristics of the output object, the fourth dimension of the falseness evaluation is performed on the output object to obtain the fourth dimension falseness value of the output object, which is used as the fourth evaluation factor; The system also includes: Update the fourth evaluation factor; by This represents the fourth evaluation factor before the update, in... This represents the updated fourth evaluation factor; Then there is ； Indicates the corresponding output object The degree of falsity in the comment text; Indicates the corresponding output object The degree of falsity of the input object; and ； The elements of the collection are the output objects. The corresponding comment text; ； The elements of the collection are the output objects. The corresponding input object.

5. The fake review detection system as described in claim 4, characterized in that, The system also includes: The update module is used to determine the output object of the corresponding comment text and the output group of the corresponding output object; it is also used to construct a comment relationship graph representing the relationship between the comment text, the output object, and the input object based on the textual and semantic features of the comment text, the behavioral features of the input object, the behavioral features of the input group, the output object, and the output group; and it is also used to update the first evaluation factor, the second evaluation factor, and the third evaluation factor based on the comment relationship graph. The comment relationship graph is represented as follows: ; In the formula It is a set of vertices; It is a set of edges; And the vertex With vertex There are edges ; Has type , ; In the formula Indicates the input object, This represents the comment text. Indicates the output object; Input object Vertices of type have attributes , representing the input group of the corresponding input object; Output object Vertices of type have attributes , which represents the output group of the corresponding output object.

6. The fake review detection system as described in claim 5, characterized in that, The update module is also used for: when ,and ,and When that happens, stop updating the first evaluation factor, the second evaluation factor, and the third evaluation factor.

Images