A false news detection method, device, equipment and medium

Abstract

The present invention relates to a method, device, equipment and medium for detecting fake news, and belongs to the field of information detection technology. It solves the problems in the prior art that comments are not diverse enough and the understanding of news content is not comprehensive enough. The fake news detection method includes: obtaining news to be detected and generating text prompt words; generating diverse comments according to the text prompt words through a large language model; extracting features of the diverse comments to obtain semantic features and diversity features; fusing the semantic features and the diversity features to obtain final features; calculating the authenticity prediction probability of the news to be detected according to the final features, and optimizing parameters according to the authenticity prediction probability. The present invention realizes the generation of rich and diverse comments, and performs feature extraction, feature fusion and other processing, so that the model can more comprehensively understand and learn the content of the news and the diverse comments, thereby improving the accuracy of fake news detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of information detection, and in particular relates to a false news detection method, device, equipment and medium. Background Art

[0002] With the development of the Internet, the detection of fake news has become crucial; fake news refers to false information that is inconsistent with objective facts and is deliberately fabricated. Fake news detection refers to a binary classification task of judging whether the news is true or false based on the content of the news or other information related to the content of the news.

[0003] Existing fake news detection methods usually use comments posted by real users, which are limited in number and coverage, and ignore the acquisition and analysis of the positions, opinions, and emotions of users who have not posted comments. Existing detection methods mainly include two categories: news content-based methods and social context-based methods. Among them, news content-based methods judge the authenticity of news by mining features related to news content, such as style features, semantic features, and emotional features. However, since fake news often evades detection by imitating real news, using only news content will limit the detection effect of fake news; social context-based methods use richer information other than news content, such as user comments, propagation paths, etc., so they have better detection effects; user comments are a widely used and proven effective social context information, because the positions, opinions, emotions, etc. in user comments can be used as effective feedback signals to help understand the corresponding news content and identify its authenticity. However, due to factors such as exposure bias caused by the recommendation mechanism and the user's personal willingness to publish, these comments are often limited, especially in the early stage of news release. Therefore, existing fake news detection methods have technical problems such as insufficient diversity of comments and insufficient understanding of news content, which further affect the accuracy of detection. Summary of the invention

[0004] In view of the above analysis, the embodiments of the present invention are intended to provide a fake news detection method, apparatus, device and medium to solve one or more of the above problems existing in the prior art.

[0005] The object of the present invention is achieved in that:

[0006] A first aspect of the present invention provides a method for detecting fake news, comprising:

[0007] S1. Obtain the news to be detected and generate text prompt words;

[0008] S2, generating diverse comments according to the text prompt words through a large language model;

[0009] S3, extracting features from the diverse comments to obtain semantic features and diversity features;

[0010] S4, fusing the semantic feature and the diversity feature to obtain a final feature;

[0011] S5. Calculate the authenticity prediction probability of the news to be detected according to the final features, and optimize parameters according to the authenticity prediction probability.

[0012] Furthermore, the generating of text prompt words includes: selecting m kinds of attribute features according to demographic characteristics; setting respective attribute values ​​for the m kinds of attribute features, and performing permutations and combinations to obtain n kinds of attribute combinations representing different types of users; and generating n kinds of text prompt words according to the news to be detected and the attribute combinations.

[0013] Furthermore, generating diverse comments according to the text prompt words through a large language model includes: inputting n kinds of the text prompt words into the large language model respectively to generate n kinds of the diverse comments C = {c1, ..., c n}.

[0014] Furthermore, feature extraction is performed on the diverse comments to obtain semantic features, including: extracting n kinds of diverse comments C = {c1, ..., c n} Perform feature extraction to obtain a set of n feature vectors According to the m attribute features, the set E of n feature vectors c Divide into m views V i , i = 1, ..., m; each of the views V i There are multiple subgroups Each of the subgroups G p There are multiple feature vectors in Calculate each of the views V i Each of the subgroups G p The average feature of all feature vectors in is used as the semantic feature

[0015]

[0016] Further, feature extraction is performed on the diversity comments to obtain diversity features, including: calculating each view V i Different subgroups G p The feature distribution difference between them; stitching each of the views V i The distribution difference of all features in the view is obtained Concatenate the intra-view diversity features of all views V To obtain the diversity feature d.

[0017] Further, the semantic feature and the diversity feature are integrated to obtain the final feature, including: extracting features of the news to be detected to obtain news features e o ; Calculate the semantic features With the news feature o The inner product of each view V i The weight w of the semantic feature in i ; According to the weight w i For each of the semantic features in the view Perform weighted calculation to obtain each view V i Overall characteristics According to the news features o and the diversity feature d to calculate the view V i The weight vector a of the overall feature A weighted calculation is performed to obtain the final feature r of all views V.

[0018] Further, the authenticity prediction probability of the news to be detected is calculated according to the final feature, and the parameters are optimized according to the authenticity prediction probability, including: according to the final feature r and the news feature e o Predict the authenticity of the news to be detected to obtain the authenticity prediction probability Predict the probability based on the truth The model parameters are optimized by minimizing the cross entropy loss function.

[0019] A second aspect of the present invention provides a false news detection device, comprising:

[0020] The acquisition generation module is used to acquire the news to be detected and generate text prompt words;

[0021] A comment generation module, used to generate diverse comments according to the text prompt words through a large language model;

[0022] A feature extraction module, used for extracting features from the diverse comments to obtain semantic features and diversity features;

[0023] A feature fusion module, used for fusing the semantic feature and the diversity feature to obtain a final feature;

[0024] The classification prediction module is used to calculate the authenticity prediction probability of the news to be detected according to the final features, and optimize parameters according to the authenticity prediction probability.

[0025] According to a third aspect of the present invention, an electronic device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the false news detection method described in any one of the embodiments is implemented.

[0026] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the false news detection method described in any one of the embodiments.

[0027] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0028] The fake news detection method provided by the present invention simulates different types of users according to text prompt words through a large language model, generates rich and diverse comments, and performs feature extraction, feature fusion and other processing on the diverse comments, so that the model can more comprehensively understand and learn the content of the news and the diverse comments, thereby improving the accuracy of fake news detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this specification. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0030] Figure 1 A flowchart of a false news detection method provided in Example 1 of the present invention;

[0031] Figure 2 A schematic diagram of a false news detection device provided in Example 2 of the present invention;

[0032] Figure 3 This is a schematic diagram of the electronic device architecture provided in Example 3 of the present invention. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. It should be noted that, in the absence of conflict, the embodiments and features in the embodiments disclosed in this disclosure can be combined, separated, interchanged and / or rearranged with each other. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0034] In the accompanying drawings, the size and relative size of components may be exaggerated for clarity and / or descriptive purposes. When the exemplary embodiments may be implemented differently, a specific process sequence may be performed in a different order than described. For example, two successively described processes may be performed substantially simultaneously or in an order opposite to the order described. In addition, the same reference numerals represent the same components.

[0035] Example 1

[0036] A specific embodiment of the present invention, as Figure 1 As shown, a false news detection method is disclosed, comprising the following steps:

[0037] It should be noted that the steps provided in this embodiment are all completed by a large model. The large model can be a Transformer or a combination of multiple models, which is not limited in this embodiment.

[0038] S1. Obtain news to be detected and generate text prompt words.

[0039] In this embodiment, generating a text prompt word specifically includes the following steps:

[0040] S101, select m attribute features according to demographic characteristics;

[0041] Specifically, the attribute characteristics include, but are not limited to, gender, age, income, education level, political inclination, personal preferences, etc.; in this embodiment, the three attributes of gender, age, and education level are taken as examples.

[0042] S102, setting respective attribute values ​​for the m attribute features, and performing permutations and combinations to obtain n attribute combinations representing different types of users;

[0043] Taking the three attribute features selected in step S101 as an example, the specific attribute values ​​of the three attributes are set as gender {male, female}, age {less than 17 years old, 18-29 years old, 30-49 years old, 50-64 years old, greater than 65 years old}, and education level {high school or below, college student, college graduate};

[0044] By permuting and combining these attribute values, 30 attribute combinations representing different types of users can be obtained.

[0045] S103: Generate n types of text prompt words according to the news to be detected and the attribute combination.

[0046] According to the news to be detected and the n attribute combinations obtained in step S102, n text prompt words are generated to assist in generating diverse comments. For example, assuming that you are a female social media user, your age is 18-29 years old, your education level is a college graduate, and you will be provided with a piece of news; please write a relevant comment based on the news, and please note that your comment must conform to the user identity assigned to you.

[0047] S2. Generate diverse comments based on the text prompt words through a large language model.

[0048] Specifically, the n types of text prompt words generated in step S103 are respectively input into the large language model to generate n types of diverse comments C = {c1, ..., c n Large language models include GLM, GPT, etc., which are not limited in this embodiment. Different identities are given to large language models through text prompt words, so as to generate richer and more diverse comments that are closer to real users.

[0049] S3. Extract features from the diverse comments to obtain semantic features and diversity features.

[0050] In this embodiment, in order to enable the model to use the diverse comments to more comprehensively understand the news to be detected, thereby improving the detection effect of false news; therefore, feature extraction is performed on the diverse comments, specifically including the following steps:

[0051] S301, for n kinds of diversity comments C = {c1, ..., c n} Perform feature extraction to obtain a set of n feature vectors

[0052] Specifically, the diversity comments C generated in step S2 = {c1,...,c n}, use sentence embedding representation models, such as Sentence Transformer, to extract the features of all comments and represent all the extracted feature vectors as a set

[0053] S302: According to the m attribute features, a set E of n feature vectors is c Divide into m views V i ,i=1,...,m;

[0054] Exemplarily, taking the three attribute features selected in step S101 as an example, step S301 will obtain the corresponding 30 feature vectors. Using the three attribute features of gender, age, and education level as classification criteria, the set of 30 feature vectors is divided into gender view V1, age view V2, and education level view V3.

[0055] S303: Each of the views V i There are multiple subgroups Each of the subgroups G p There are multiple feature vectors in

[0056] Specifically, taking the three views obtained in step S302 as an example, the gender view V1 includes the male subgroup and female subgroups Age view V2 includes the under 17 subgroup 18-29 years old subgroup 30-49 years old subgroup 50-64 years old subgroup and the over 65 age group Education level view V3 includes subgroups with high school education or below University student subgroup and college graduate subgroup Each subgroup G p Including multiple feature vectors

[0057] All feature vectors in the collection are divided into corresponding views according to attribute features, so that the model can process them more efficiently.

[0058] S304: Calculate each view V i Each of the subgroups G p The average feature of all feature vectors in is used as the semantic feature

[0059] Specifically, for view V i , calculate each subgroup G in the view p The average feature of all feature vectors in is expressed as follows:

[0060]

[0061] in, represents the average feature, that is, the semantic feature, Represents view V i A subgroup G in p The set of eigenvectors in , Representing a collection The number of eigenvectors in , p∈[1,2,...j].

[0062] Considering the same view, there are differences between different subgroups. In order to make the model understand these differences more comprehensively, S3 also includes the following steps:

[0063] S305: Calculate each view V i Different subgroups G p Differences in feature distribution between

[0064] Specifically, for view V i , first, calculate each subgroup G p The probability distribution vector corresponding to all eigenvectors in is expressed as follows:

[0065]

[0066] in, represents the probability distribution vector, Softmax(·) represents the Softmax function, represents a certain feature vector.

[0067] Then, the difference values ​​between the two subgroups in the view are calculated, for example, by the KL divergence function, which is expressed as follows:

[0068]

[0069] Among them, KL(G p ,G q ) represents the difference between subpopulations Gp and Gq, and Respectively represent subgroups G p and G q The number of eigenvectors in , kl_div(·,·) represents the KL divergence function, and Respectively represent subgroups G p and G q Eigenvector and The corresponding probability distribution vector.

[0070] S306: stitching each of the views V i The distribution difference of all features in the view is obtained

[0071] Specifically, according to a view V obtained in step S305 i The difference values ​​between different subgroups within the view V i All the difference values ​​in the image are concatenated into a vector to obtain the intra-view diversity feature of each view, which is expressed as follows:

[0072]

[0073] in, represents the diversity feature in the view, ⊕ represents splicing, KL(Gp ,G q ) represents the subgroup G p and G q The difference between .

[0074] S307: stitching the intra-view diversity features of all views V To obtain the diversity feature d.

[0075] Specifically, in order to facilitate the model to learn features, the intra-view diversity features corresponding to all views obtained by the splicing operation processing step S306 are To obtain the diversity features of all views, it is expressed as follows:

[0076]

[0077] Among them, d represents the diversity feature, A view V representing a view i Internal diversity characteristics.

[0078] S4. Fusing the semantic feature and the diversity feature to obtain a final feature.

[0079] In this embodiment, in order to further enhance the ability of the model to detect fake news, the final feature is obtained by fusing the semantic feature, the diversity feature, and the feature of the news to be detected, which specifically includes the following steps:

[0080] S401, extracting features of the news to be detected to obtain news features e o ;

[0081] S402: Calculate the semantic features With the news feature o The inner product of each view V i The weight w of the semantic feature in i ;

[0082] Specifically, considering that different views specialize in different analysis angles, and different subgroups within the same view represent different understandings from the analysis angle, it is necessary to calculate the weights of different semantic features within each view, which are expressed as follows:

[0083]

[0084] Among them, w i The weight of the semantic feature, Softmax(·) represents the Softmax function, Represents view V i All semantic features in e orepresents news features, and dim represents the dimension of semantic features and news features.

[0085] S403, according to the weight w i For each of the semantic features in the view Perform weighted calculation to obtain each view V i Overall characteristics

[0086] Specifically, after the weights of different semantic features in each view are obtained through step S402, the overall feature is obtained by weighted calculation, which is expressed as follows:

[0087]

[0088] in, represents the overall characteristics of each view, w i represents the weight of the semantic feature, Represents view V i All semantic features within.

[0089] S404, based on the news feature o and the diversity feature d to calculate the view V i The weight vector a of

[0090] Specifically, the diversity of views and the news to be detected both determine the importance of the view. Therefore, this embodiment provides a view gate to fuse the features between different views so that the model can adapt to the importance of different views. The input of the view gate is the diversity feature and the news feature, and the output is the weights of different views, which are expressed as follows:

[0091] a=Softmax(G(eo⊕d;θ))

[0092] Where a represents the weight vector of all views, Softmax(·) represents the Softmax function, G represents the view gate, ⊕ represents splicing, and e o represents news features, d represents diversity features, and θ represents the parameters of the view gate.

[0093] S405: The overall feature s is analyzed according to the weight vector a. Vi A weighted calculation is performed to obtain the final feature r of all views V.

[0094] Specifically, after step S404 calculates the weight vector a of all views, the overall feature Perform weighted calculation to obtain the final features of all views V, which are expressed as follows:

[0095]

[0096] Among them, r represents the final feature, a i represents the weight vector, Represents overall characteristics.

[0097] S5. Calculate the authenticity prediction probability of the news to be detected according to the final features, and optimize parameters according to the authenticity prediction probability.

[0098] In this embodiment, the authenticity prediction probability of the news to be detected is calculated according to the final feature, and the parameters are optimized according to the authenticity prediction probability, which specifically includes the following steps:

[0099] S501: Based on the final feature r and the news feature e o Predict the authenticity of the news to be detected to obtain the authenticity prediction probability The size of the authenticity prediction probability indicates the probability that the news to be detected is false news. The calculation process is as follows:

[0100]

[0101] in, represents the probability of authenticity prediction, Sigmoid(·) represents the Sigmoid activation function, MLP(·) represents the multi-layer perceptron, r represents the final feature, ⊕ represents concatenation, e o Indicates news features.

[0102] S502: Predicting probability based on the authenticity The model parameters are optimized by minimizing the cross entropy loss function, which is expressed as follows:

[0103]

[0104] Among them, L represents the cross entropy loss function, y represents the true value of the news to be predicted (0 is true, 1 is false), represents the probability of true prediction.

[0105] In some embodiments, the parameters of the model may be optimized by gradient descent method and the like to make the prediction more accurate.

[0106] Compared with the prior art, the fake news detection method provided in this embodiment simulates different types of users according to text prompt words through a large language model, generates rich and diverse comments, and performs feature extraction and feature fusion on the diverse comments, so that the model can understand and learn the content of the news and diverse comments more comprehensively, thereby improving the accuracy of fake news detection.

[0107] Example 2

[0108] This embodiment provides a false news detection device. Figure 2 As shown, including:

[0109] The acquisition generation module is used to acquire the news to be detected and generate text prompt words;

[0110] A comment generation module, used to generate diverse comments according to the text prompt words through a large language model;

[0111] A feature extraction module, used for extracting features from the diverse comments to obtain semantic features and diversity features;

[0112] A feature fusion module, used for fusing the semantic feature and the diversity feature to obtain a final feature;

[0113] The classification prediction module is used to calculate the authenticity prediction probability of the news to be detected according to the final features, and optimize parameters according to the authenticity prediction probability.

[0114] Example 3

[0115] This embodiment provides an electronic device, such as Figure 3 As shown, it includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the fake news detection method as described in any of the above embodiments is implemented.

[0116] Example 4

[0117] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the false news detection method as described in any of the above embodiments is implemented.

[0118] Computer readable storage media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0119] The professionals should further realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0120] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0121] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting fake news, characterized in that: include: S1. Obtain the news to be detected and generate text prompt words; S2, generating diverse comments according to the text prompt words through a large language model; S3, extracting features from the diverse comments to obtain semantic features and diversity features, including: extracting features from n diverse comments C = {c1, ..., c n } Perform feature extraction to obtain a set of n feature vectors According to m attribute features, the set E of n feature vectors c Divide into m views V i , i = 1, ... m; each of the views V i There are multiple subgroups Each of the subgroups G p There are multiple feature vectors in Calculate each of the views V i Each of the subgroups G p The average feature of all feature vectors in is used as the semantic feature Calculate each of the views V i Different subgroups G p The feature distribution difference between them; stitching each of the views V i The distribution difference of all features in the view is obtained Concatenate the intra-view diversity features of all views V To obtain the diversity feature d; S4, fusing the semantic features and the diversity features to obtain the final features, including: extracting features of the news to be detected to obtain news features e o ; Calculate the semantic features With the news feature o The inner product of each view V i The weight w of the semantic feature in i ; According to the weight w i For each of the semantic features in the view Perform weighted calculation to obtain each view V i Overall characteristics According to the news features o and the diversity feature d to calculate the view V i The weight vector a of the overall feature Perform weighted calculation to obtain the final feature r of all views V; S5. Calculate the authenticity prediction probability of the news to be detected according to the final features, and optimize parameters according to the authenticity prediction probability.

2. The false news detection method according to claim 1, characterized in that: The generating text prompt word comprises: Select m attribute features based on demographic characteristics; Setting respective attribute values ​​for the m attribute features, and performing permutations and combinations to obtain n attribute combinations representing different types of users; Generate n types of text prompt words according to the news to be detected and the attribute combination.

3. The false news detection method according to claim 2, characterized in that: Generate diverse comments based on the text prompt words through a large language model, including: Input the n text prompt words into the large language model respectively to generate n diverse comments C = {c1, ..., c n }.

4. The false news detection method according to claim 1, characterized in that: Calculating the authenticity prediction probability of the news to be detected according to the final feature, and optimizing parameters according to the authenticity prediction probability, including: According to the final feature r and the news feature e o Predict the authenticity of the news to be detected to obtain the authenticity prediction probability Predict the probability based on the truth The model parameters are optimized by minimizing the cross entropy loss function.

5. A false news detection device, characterized in that: The device comprises: The acquisition generation module is used to acquire the news to be detected and generate text prompt words; A comment generation module, used to generate diverse comments according to the text prompt words through a large language model; The feature extraction module is used to extract features from the diversity comments to obtain semantic features and diversity features, including: extracting n types of diversity comments C = {c1, ..., c n } Perform feature extraction to obtain a set of n feature vectors According to m attribute features, the set E of n feature vectors c Divide into m views V i , i = 1, ... m; each of the views V i There are multiple subgroups Each of the subgroups G p There are multiple feature vectors in Calculate each of the views V i Each of the subgroups G p The average feature of all feature vectors in is used as the semantic feature Calculate each of the views V i Different subgroups G p The feature distribution difference between them; stitching each of the views V i The distribution difference of all features in the view is obtained Concatenate the intra-view diversity features of all views V To obtain the diversity feature d; A feature fusion module is used to fuse the semantic features and the diversity features to obtain the final features, including: extracting features of the news to be detected to obtain news features e o ; Calculate the semantic features With the news feature o The inner product of each view V i The weight w of the semantic feature in i ; According to the weight w i For each of the semantic features in the view Perform weighted calculation to obtain each view V i Overall characteristics According to the news features o and the diversity feature d to calculate the view V i The weight vector a of the overall feature Perform weighted calculation to obtain the final feature r of all views V; The classification prediction module is used to calculate the authenticity prediction probability of the news to be detected according to the final features, and optimize parameters according to the authenticity prediction probability.

6. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the false news detection method according to any one of claims 1 to 4 is implemented.

7. A storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the false news detection method according to any one of claims 1 to 4 is implemented.

Images