A false news detection method and system based on size model cooperation

Abstract

The application discloses a false news detection method and system based on size model cooperation, relates to the technical field of data mining, and comprises the following steps: acquiring to-be-detected news data, generating a news data set based on the to-be-detected news data, dividing part of data in the news data set to obtain a training set; inputting the training set into a previously established false news detection model based on size model cooperation for training, and outputting a trained false news detection model based on size model cooperation, wherein the size model comprises a large language model and a detection small model; inputting the news data set into the trained false news detection model based on size model cooperation for detection, obtaining a false news detection result, and realizing a more accurate detection result.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, specifically a method and system for detecting fake news based on the collaboration of large and small models. Background Technology

[0002] Short video platforms have become important channels for news dissemination, but their openness has led to a proliferation of fake news. The strong appeal of video content combined with the platform's reach causes fake news to spread explosively. Because fake news spreads rapidly in its early stages, developing targeted content-based fake news detection methods is crucial.

[0003] Existing methods primarily integrate diverse information from short video news by designing complex modules. By incorporating richer information, detection capabilities can be improved, demonstrating continuously optimizing detection performance. Although existing fake news detection methods have achieved initial success, they only detect descriptive errors in news content, ignoring factual errors, which leads to limited detection accuracy. Summary of the Invention

[0004] To address the shortcomings mentioned in the background section, the present invention aims to provide a method and system for detecting fake news based on size model collaboration.

[0005] Firstly, the objective of this invention can be achieved through the following technical solution: a method for detecting fake news based on size model collaboration, the method comprising the following steps:

[0006] Obtain the news data to be detected, generate a news dataset based on the news data to be detected, and divide the data within the news dataset to obtain the training set;

[0007] The training set is input into a pre-established fake news detection model based on the collaboration of large and small models for training, and the trained fake news detection model based on the collaboration of large and small models is output. The large and small models include a large language model and a small detection model.

[0008] The news dataset is input into a trained fake news detection model based on size-model collaboration for detection, and the fake news detection results are obtained.

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the news dataset contains One news item, among which... It is a positive integer;

[0010] The large language model uses prompting learning techniques to transfer common sense knowledge from the large language model to the detection small model.

[0011] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the pre-established fake news detection model based on size model collaboration includes: common sense knowledge extraction, fine-grained feature encoding, feature cleansing, and collaborative detection;

[0012] The common sense knowledge extraction process takes news text and video information as input, employs prompting learning technology, and outputs common sense information.

[0013] The fine-grained feature encoding takes the news text, video information, and common sense information as input, and uses a pre-trained feature encoder to output the global text features of the news. Local text features Global video features Local video features Common sense characteristics ,in, The features are all in tensor form. Dimensions representing features Corresponding to the number of characters in the input text, Corresponding to the number of frames in the input video;

[0014] The feature cleansing uses the global textual semantic features of the news. Local text semantic features Global video semantic features Local video semantic features Input: Cleaned text features Video features , The number of features to retain;

[0015] The collaborative detection uses cleaned text features Video features Common sense characteristics The input is the detection result, which is used as the fake news detection result.

[0016] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of extracting common sense knowledge using cue learning technology and outputting common sense information, including:

[0017] Using pre-built prompt word templates, the prompt words are combined with news text and video information to construct the request information;

[0018] The request information is input into the multimodal large language model through the application programming interface, and common sense information is output.

[0019] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the feature purification includes text feature purification and video feature purification, including:

[0020] The text feature cleansing is achieved through global text features. and global video features Guided by the algorithm, the system filters out news-related parts of local text features and outputs purified text features. ;

[0021] The video feature cleanup utilizes global text features. and global video features Guided by the algorithm, the system filters out news-related features from local video features and outputs the purified video features. .

[0022] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the text feature purification includes the following steps:

[0023] Calculate local text features With global text features The similarity between the texts is used to obtain the text-to-text similarity matrix. ;

[0024] Calculate local text features With global video features The similarity between the two is used to obtain the text-video similarity matrix. ;

[0025] Text-to-text similarity matrix Text-video similarity matrix Add them together to obtain the final text filtering matrix. ;

[0026] Based on the final text filtering matrix The bubble sort algorithm is used to retain the most similar results. By analyzing the features at each location, we can obtain the purified text features. .

[0027] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the video feature purification includes the following steps:

[0028] Calculate local video features With global text features The similarity between the two is used to obtain the video-text similarity matrix. ;

[0029] Calculate local video features With global video features The similarity between the videos is used to obtain a video-video similarity matrix. ;

[0030] Video-text similarity matrix Video-to-video similarity matrix Add them together to obtain the final video filtering matrix. ;

[0031] Based on the final video selection matrix The bubble sort algorithm is used to retain the most similar results. By analyzing the features at each location, the purified video features are obtained. In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the cooperative detection process, comprising:

[0032] Cleaned text features Video features Common sense characteristics The fusion process is performed to obtain the fused news features. The fusion expression is as follows:

[0033]

[0034] in, For average pooling function, For concatenation functions, A learnable mapping matrix, For bias vectors, The characteristics of the merged news;

[0035] Based on the fused news features, a classifier is used for classification, and the classification expression is:

[0036] in, A learnable mapping matrix, For bias vectors, For the classification results, when When the news is true, it is true; otherwise, it is false.

[0037] Secondly, in order to achieve the above objectives, this invention discloses a fake news detection system based on size model collaboration, comprising:

[0038] The data processing module is used to acquire the news data to be detected, generate a news dataset based on the news data to be detected, and divide the data within the news dataset to obtain a training set.

[0039] The model training module is used to input the training set into a pre-established fake news detection model based on the collaboration of large and small models for training, and output the trained fake news detection model based on the collaboration of large and small models, wherein the large and small models include a large language model and a small detection model;

[0040] The news detection module is used to input news datasets into a trained fake news detection model based on size-model collaboration for detection, and obtain fake news detection results.

[0041] In another aspect of the present invention, in order to achieve the above-mentioned objective, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor, and when the processor loads and executes the computer program, it employs a fake news detection method based on size model collaboration as described above.

[0042] The beneficial effects of this invention are:

[0043] This invention constructs a fake news detection model based on the collaboration of large and small models, which can effectively mine users' core behavioral intentions and avoid the introduction of redundant behavioral features by the whole graph learning technique, thereby more accurately modeling user behavioral features. At the same time, by constraining the training of the multimedia recommendation model through self-supervised learning auxiliary tasks, the extracted user features are made more in line with user preferences, avoiding the contamination of user features by noisy behavior, and further enhancing the distinguishability of user features. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0046] Figure 2 This is a schematic diagram of the fake news detection model structure based on the collaboration of large and small models in this invention;

[0047] Figure 3 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Example 1:

[0050] like Figure 1 As shown, a fake news detection method based on size model collaboration is proposed, which includes the following steps:

[0051] S101: Obtain the news data to be detected, generate a news dataset based on the news data to be detected, divide the data within the news dataset to obtain the training set;

[0052] The news dataset contains One news item, among which... The value is a positive integer; in this embodiment of the invention, 6047 news items were collected. The value is 6047.

[0053] The large language model uses prompting learning techniques to transfer common sense knowledge from the large language model to the detection small model.

[0054] S102: Input the training set into the pre-established fake news detection model based on the collaboration of large and small models for training, and output the trained fake news detection model based on the collaboration of large and small models, wherein the large and small models include a large language model and a small detection model;

[0055] The pre-established fake news detection model based on big-small model collaboration includes: common sense knowledge extraction, fine-grained feature encoding, feature cleansing, and collaborative detection;

[0056] The common sense knowledge extraction process takes news text and video information as input, employs prompting learning technology, and outputs common sense information.

[0057] The fine-grained feature encoding takes the news text, video information, and common sense information as input, and uses a pre-trained feature encoder to output the global text features of the news. Local text features Global video features Local video features Common sense characteristics ,in, The features are all in tensor form. Dimensions representing features Corresponding to the number of characters in the input text, Corresponding to the number of frames in the input video; in this embodiment, It is 64;

[0058] The feature cleansing uses the global textual semantic features of the news. Local text semantic features Global video semantic features Local video semantic features Input: Cleaned text features Video features , To preserve the number of features; in this embodiment, It is 5;

[0059] The collaborative detection uses cleaned text features Video features Common sense characteristics The input is the detection result, which is used as the fake news detection result.

[0060] The common sense knowledge extraction process employs cue-based learning technology, and the output of common sense information includes:

[0061] Using pre-built prompt word templates, the prompt words are combined with news text and video information to construct the request information;

[0062] The request information is input into the multimodal large language model through the Application Programming Interface (API), and common sense information is output.

[0063] The feature cleanup includes text feature cleanup and video feature cleanup, including:

[0064] The text feature cleansing is achieved through global text features. and global video features Guided by the algorithm, the system filters out news-related parts of local text features and outputs purified text features. ;

[0065] The video feature cleanup utilizes global text features. and global video features Guided by the algorithm, the system filters out news-related features from local video features and outputs the purified video features. .

[0066] The text feature cleansing includes the following steps:

[0067] Calculate local text features With global text features The similarity between the texts is used to obtain the text-to-text similarity matrix. ;

[0068] Calculate local text features With global video features The similarity between the two is used to obtain the text-video similarity matrix. ;

[0069] Text-to-text similarity matrix Text-video similarity matrix Add them together to obtain the final text filtering matrix. ;

[0070] Based on the final text filtering matrix The bubble sort algorithm is used to retain the most similar results. By analyzing the features at each location, we can obtain the purified text features. .

[0071] The video feature cleansing includes the following steps:

[0072] Calculate local video features With global text features The similarity between the two is used to obtain the video-text similarity matrix. ;

[0073] Calculate local video features With global video features The similarity between the videos is used to obtain a video-video similarity matrix. ;

[0074] Video-text similarity matrix Video-to-video similarity matrix Add them together to obtain the final video filtering matrix. ;

[0075] Based on the final video selection matrix The bubble sort algorithm is used to retain the most similar results. By analyzing the features at each location, the purified video features are obtained. .

[0076] The collaborative detection process includes:

[0077] Cleaned text features Video features Common sense characteristics The fusion process is performed to obtain the fused news features. The fusion expression is as follows:

[0078] in, For average pooling function, For concatenation functions, A learnable mapping matrix, For bias vectors, The characteristics of the merged news;

[0079] Based on the fused news features, a classifier is used for classification, and the classification expression is:

[0080] in, A learnable mapping matrix, For bias vectors, For the classification results, when When the news is true, it is true; otherwise, it is false.

[0081] S103: Input the news dataset into the trained fake news detection model based on size-model collaboration for detection, and obtain the fake news detection results.

[0082] Specifically, the present invention will be further illustrated below through embodiments:

[0083] The performance of the recommended method of this invention is compared with that of existing fake news detection methods. The comparison experiment is as follows:

[0084] This embodiment reproduces existing mainstream fake news detection methods based on the PyTorch framework and develops the fake news detection method based on large and small model collaboration proposed in this invention. The parameters in all comparison methods follow the optimal settings reported in the article, and the Adam optimizer is used to train the fake news detection model. The testing process of the fake news detection model proposed in this invention is as described in Embodiment 1.

[0085] The experimental datasets used for testing were two mainstream datasets: FakeSV and FakeTT.

[0086] All experiments were conducted on an NVIDIA Tesla V100 GPU.

[0087] The results of the comparative experiment are shown in Table 1. It can be seen that:

[0088] This method significantly outperforms existing fake news detection methods on multiple public datasets. By constructing a fake news detection model based on size-model collaboration, it can effectively mine users' core behavioral intentions, avoiding the introduction of redundant behavioral features by full graph learning techniques, thus more accurately modeling user behavioral features. At the same time, by constraining the training of the multimedia recommendation model through self-supervised learning auxiliary tasks, the extracted user features are made more consistent with user preferences, avoiding contamination of user features by noisy behaviors, and further enhancing the discriminability of user features.

[0089] Table 1 compares the performance of the fake news detection method based on size model collaboration with other existing fake news methods.

[0090]

[0091] Figure 2 This provides a more intuitive demonstration of the detection method. For short video news, video frames are first extracted, including video information and related video text information for each frame. Then, using a common sense knowledge extraction module, the news text and video information are used as input, and cue learning technology is employed to output common sense information. Subsequently, the extracted common sense information and the news text and video information are input into a fine-grained feature encoding module to output the global text features of the news. Local text features Global video features Local video features Common sense characteristics Secondly, the global textual semantic features of the news will be analyzed. Local text semantic features Global video semantic features Local video semantic features are input into the feature cleansing module, which outputs cleaned text features. Video features Finally, the collaborative detection module uses the cleaned text features... Video features Common sense characteristics The input is the detection result, which is used as the fake news detection result.

[0092] Example 2: To achieve the above objective, such as Figure 3 Based on Embodiment 1, this invention discloses a fake news detection system based on size model collaboration, comprising:

[0093] Data processing module 11 is used to acquire news data to be detected, generate a news dataset based on the news data to be detected, and divide some data within the news dataset to obtain a training set;

[0094] The model training module 12 is used to input the training set into a pre-established fake news detection model based on the collaboration of large and small models for training, and output the trained fake news detection model based on the collaboration of large and small models, wherein the large and small models include a large language model and a small detection model.

[0095] The news detection module 13 is used to input the news dataset into the trained fake news detection model based on size model collaboration for detection, and obtain fake news detection results.

[0096] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0097] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0098] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0099] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. A method for detecting fake news based on size-model collaboration, characterized in that, The method includes the following steps: Obtain the news data to be detected, generate a news dataset based on the news data to be detected, and divide the data within the news dataset to obtain the training set; The training set is input into a pre-established fake news detection model based on the collaboration of large and small models for training, and the trained fake news detection model based on the collaboration of large and small models is output. The large and small models include a large language model and a small detection model. The pre-established fake news detection model based on big-small model collaboration includes: common sense knowledge extraction, fine-grained feature encoding, feature cleansing, and collaborative detection; The common sense knowledge extraction process takes news text and video information as input, employs prompting learning technology, and outputs common sense information. The fine-grained feature encoding takes the news text, video information, and common sense information as input, and uses a pre-trained feature encoder to output the global text features of the news. Local text features Global video features Local video features Common sense characteristics ,in, The features are all in tensor form. Dimensions representing features Corresponding to the number of characters in the input text, Corresponding to the number of frames in the input video; The feature cleansing uses global text features of the news. Local text features Global video features Local video features Input: Cleaned text features Video features , The number of features to retain; The collaborative detection uses cleaned text features Video features Common sense characteristics The input is the detection result, which is used as the fake news detection result. The feature cleanup includes text feature cleanup and video feature cleanup, including: The text feature cleansing is achieved through global text features. and global video features Guided by the algorithm, the system filters out news-related parts of local text features and outputs purified text features. ; The video feature cleanup utilizes global text features. and global video features Guided by the algorithm, the system filters out news-related features from local video features and outputs the purified video features. ; The text feature cleansing includes the following steps: Calculate local text features With global text features The similarity between the texts is used to obtain the text-to-text similarity matrix. ; Calculate local text features With global video features The similarity between the two is used to obtain the text-video similarity matrix. ; Text-to-text similarity matrix Text-video similarity matrix Add them together to obtain the final text filtering matrix. ; Based on the final text filtering matrix The bubble sort algorithm is used to retain the most similar results. By analyzing the features at each location, we can obtain the purified text features. ; The video feature cleansing includes the following steps: Calculate local video features With global text features The similarity between the two is used to obtain the video-text similarity matrix. ; Calculate local video features With global video features The similarity between the videos is used to obtain a video-video similarity matrix. ; Video-text similarity matrix Video-to-video similarity matrix Add them together to obtain the final video filtering matrix. ; Based on the final video selection matrix The bubble sort algorithm is used to retain the most similar results. By analyzing the features at each location, we can obtain the purified video features. ; The news dataset is input into a trained fake news detection model based on size-model collaboration for detection, and the fake news detection results are obtained.

2. The method for detecting fake news based on size model collaboration according to claim 1, characterized in that, The news dataset contains One news item, among which... It is a positive integer; The large language model uses prompting learning techniques to transfer common sense knowledge from the large language model to the detection small model.

3. The method for detecting fake news based on size model collaboration according to claim 1, characterized in that, The common sense knowledge extraction process employs cue-based learning technology, and the output of common sense information includes: Using pre-built prompt word templates, the prompt words are combined with news text and video information to construct the request information; The request information is input into the multimodal large language model through the application programming interface, and common sense information is output.

4. The method for detecting fake news based on size model collaboration according to claim 3, characterized in that, The collaborative detection process includes: Cleaned text features Video features Common sense characteristics The fusion process is performed to obtain the fused news features. The fusion expression is as follows: in, For average pooling function, For concatenation functions, A learnable mapping matrix, For bias vectors, The characteristics of the merged news; Based on the fused news features, a classifier is used for classification, and the classification expression is: in, A learnable mapping matrix, For bias vectors, For the classification results, when When the news is true, it is true; otherwise, it is false.

5. A fake news detection system based on size-model collaboration, employing the fake news detection method based on size-model collaboration as described in any one of claims 1 to 4, characterized in that, include: The data processing module is used to acquire the news data to be detected, generate a news dataset based on the news data to be detected, and divide the data within the news dataset to obtain a training set. The model training module is used to input the training set into a pre-established fake news detection model based on the collaboration of large and small models for training, and output the trained fake news detection model based on the collaboration of large and small models, wherein the large and small models include a large language model and a small detection model; The news detection module is used to input news datasets into a trained fake news detection model based on size-model collaboration for detection, and obtain fake news detection results.

6. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on a processor. When the processor loads and executes the computer program, it employs a fake news detection method based on size model collaboration as described in any one of claims 1 to 4.

Images