A method and system for detecting fake news based on semantic decoupling and knowledge enhancement

Abstract

This invention relates to a method and system for detecting fake news based on semantic decoupling and knowledge enhancement. The method includes: acquiring news text data; preprocessing the news text data to obtain standardized news text data and corresponding news auxiliary information; extracting news features and auxiliary information features from the news text data and corresponding news auxiliary information using a pre-trained model; extracting topic features, content features, and auxiliary features using multiple autoencoders; fusing the topic features and auxiliary features based on a bidirectional attention mechanism; calculating a separation constraint signal based on the fused features; constructing a loss function based on the separation constraint signal and the fused features; jointly training the multiple autoencoders using the loss function; and detecting target news using the trained multiple autoencoders. This invention improves the accuracy and generalization ability of fake news detection by combining multiple encoders and a bidirectional attention mechanism.

Description

Technical Field

[0001] This invention belongs to the fields of natural language processing and information security technology, specifically relating to a method and system for detecting fake news based on semantic decoupling and knowledge enhancement. Background Technology

[0002] With the rapid development of social media and online news platforms, the spread of fake news has shown an exponential growth trend, posing a serious threat to social stability and public safety. Existing fake news detection technologies are mainly divided into two categories: methods based on dissemination patterns and methods based on content analysis.

[0003] Methods based on dissemination patterns primarily detect news by analyzing its dissemination path and user interaction patterns on social networks, such as the graph neural network-based dissemination pattern analysis method proposed in patent CN115269853A. While these methods have some effectiveness, they rely on complete dissemination data, making it difficult to obtain sufficient information in the early stages of news release, resulting in a detection lag.

[0004] Content-based analysis methods mainly focus on the linguistic features of news texts, such as the deep learning-based text feature extraction method disclosed in patent CN120277400A. However, these methods have the following technical shortcomings: (1) they fail to fully consider the semantic overlap between real and fake news under the same topic, resulting in limited detection accuracy; (2) they lack the ability to fine-grained decouple the real and fake components in news content; and (3) they fail to effectively utilize external knowledge to assist in judgment, resulting in insufficient generalization ability when facing emerging topics.

[0005] Furthermore, existing technologies, such as the autoencoder architecture proposed in patent CN120316253A, while capable of extracting latent feature representations of news, cannot explicitly distinguish between general theme features and content-specific features, making it difficult to capture the typical "mixed truth and falsehood" feature patterns in fake news. The large-scale language model direct detection scheme used in patent CN119782889A, while possessing strong semantic understanding capabilities, suffers from high computational resource consumption and poor interpretability.

[0006] Therefore, there is an urgent need for a new fake news detection technology that can effectively decouple the general features of the theme from the specific features of the content, and integrate external knowledge to assist in the judgment, so as to solve the technical problems of insufficient detection accuracy and limited generalization ability of existing methods. Summary of the Invention

[0007] To improve the accuracy and generalization ability of fake news detection, a fake news detection method based on semantic decoupling and knowledge enhancement is provided in the first aspect of the present invention, comprising:

[0008] The process involves: acquiring news text data; preprocessing the news text data to obtain standardized news text data and corresponding news auxiliary information; the auxiliary information including topic description, credibility, and background knowledge; extracting news features and auxiliary information features from the news text data and corresponding news auxiliary information using a pre-trained model; extracting topic features, content features, and auxiliary features from the news features and auxiliary information features using multiple autoencoders; fusing the topic features and auxiliary features based on a bidirectional attention mechanism; calculating a separation constraint signal based on the fused features; constructing a loss function based on the separation constraint signal and the fused features; jointly training the multiple autoencoders using the loss function; and detecting target news using the trained multiple autoencoders.

[0009] In some embodiments of the present invention, the step of extracting news features and auxiliary information features from the news text data and corresponding news auxiliary information using a pre-trained model includes: extracting news features and auxiliary information features from the news text data and corresponding news auxiliary information using two pre-trained language models.

[0010] In some embodiments of the present invention, the step of extracting topic features, content features and auxiliary features from the news features and the auxiliary information features by means of multiple autoencoders includes: extracting topic features from the news features by means of a first autoencoder; extracting content features from the news features by means of a second autoencoder; and extracting auxiliary features from the auxiliary information features by means of a third autoencoder.

[0011] In some embodiments of the present invention, the fusion of topic features and auxiliary features based on a bidirectional attention mechanism includes: calculating a cross-modal attention matrix of topic features and auxiliary features, and generating a first attention weight using a preset activation function and learnable parameters; calculating a topic feature representation perceived by auxiliary features using the first attention weight; calculating a cross-modal attention matrix of topic features and auxiliary features, and generating a second attention weight using a preset activation function and learnable parameters; calculating an auxiliary feature representation perceived by topic features using the second attention weight; and fusing the topic feature representation perceived by auxiliary features and the auxiliary feature representation perceived by topic features.

[0012] Furthermore, the step of calculating the separation constraint signal based on the fused features includes: calculating the separation constraint signal based on the fused features using a multilayer perceptron and a preset activation function.

[0013] In some embodiments of the present invention, the step of constructing a loss function based on the separation constraint signal and the fused features includes: constructing a detection loss function based on the fused features; constructing a decoupling loss function based on the separation constraint signal; and constructing a consistency loss function and a reconstruction loss function based on the topic features and auxiliary features.

[0014] A second aspect of the present invention provides a fake news detection system based on semantic decoupling and knowledge enhancement, comprising: an acquisition module for acquiring news text data; preprocessing the news text data to obtain standardized news text data and corresponding news auxiliary information; the auxiliary information including topic description, credibility, and background knowledge; an extraction module for extracting news features and auxiliary information features from the news text data and corresponding news auxiliary information using a pre-trained model; extracting topic features, content features, and auxiliary features from the news features and the auxiliary information features using multiple autoencoders; a fusion module for fusing topic features and auxiliary features based on a bidirectional attention mechanism; calculating a separation constraint signal based on the fused features; a training module for constructing a loss function based on the separation constraint signal and the fused features; jointly training the multiple autoencoders using the loss function; and a detection module for detecting target news using the trained multiple autoencoders.

[0015] A third aspect of the present invention provides an electronic device comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the fake news detection method based on semantic decoupling and knowledge enhancement provided in the first aspect of the present invention.

[0016] In a fourth aspect, the present invention provides a computer-readable medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the fake news detection method based on semantic decoupling and knowledge enhancement provided in the first aspect of the present invention.

[0017] The beneficial effects of this invention are:

[0018] This invention employs a three-way feature decoupling architecture to decompose news content into general theme features and content-specific features. Combined with auxiliary information generated by a large language model, it achieves accurate separation of mixed semantics between true and false news. Utilizing auxiliary information as semantic anchors and dynamically fusing external knowledge through a bidirectional attention mechanism, it significantly enhances the ability to discriminate emerging topics. The designed multi-objective collaborative optimization framework, through joint training with four loss functions, strengthens feature independence and maintains theme consistency while ensuring classification accuracy, effectively overcoming the feature coupling defects in traditional methods. The sparse feature space construction mechanism, combined with dimensional constraints, not only improves model generalization but also generates interpretable decision-making criteria, intuitively displaying key discriminative features. In terms of computational efficiency, through feature compression and parallel processing architecture, it significantly reduces inference latency while maintaining high accuracy, providing a solution for complex fake news detection. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the basic process of a fake news detection method based on semantic decoupling and knowledge enhancement in some embodiments of the present invention;

[0020] Figure 2 This is a schematic diagram illustrating the specific process of a fake news detection method based on semantic decoupling and knowledge enhancement in some embodiments of the present invention;

[0021] Figure 3 This is a schematic diagram of the structure of a fake news detection system based on semantic decoupling and knowledge enhancement in some embodiments of the present invention;

[0022] Figure 4 This is a schematic diagram of the structure of an electronic device in some embodiments of the present invention. Detailed Implementation

[0023] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0024] refer to Figure 1 and Figure 2 In a first aspect, the present invention provides a method for detecting fake news based on semantic decoupling and knowledge enhancement, comprising:

[0025] S100. Acquire news text data; preprocess the news text data to obtain standardized news text data and corresponding news auxiliary information; the auxiliary information includes topic description, credibility and background knowledge;

[0026] S200. Extract news features and auxiliary information features from the news text data and corresponding news auxiliary information using a pre-trained model; extract topic features, content features and auxiliary features from the news features and the auxiliary information features using multiple autoencoders;

[0027] S300. Based on a bidirectional attention mechanism, the topic features and auxiliary features are fused; based on the fused features, the separation constraint signal is calculated.

[0028] S400. Based on the separated constraint signal and the fused features, construct a loss function; jointly train the multiple autoencoders using the loss function;

[0029] S500 detects target news by using multiple trained autoencoders.

[0030] In step S100 of some embodiments of the present invention, news text data is acquired; the news text data is preprocessed to obtain standardized news text data and corresponding news auxiliary information; the auxiliary information includes topic description, credibility and background knowledge.

[0031] Specifically, S101. Obtain a text dataset containing real news and fake news, the dataset including news text, publication time, source website and other metadata; preprocess the obtained news text, including: removing HTML tags, special characters and non-text content; unifying the encoding format to UTF-8; standardize the text, including case conversion and spell correction, and finally obtain a data set D={p_1,p_2,…,p_n} containing n news instances, where p_i represents the i-th original news instance.

[0032] S102. Use a large language model to generate auxiliary information related to the news content. Specifically, guide the large language model to generate auxiliary information by designing prompt templates. The auxiliary information includes: 1. News topic information: Extract the core theme and key facts of the news. 2. Credibility analysis: Analyze from three dimensions: reasonableness, verifiability, and source credibility. 3. Background knowledge: Provide background information and common sense information related to the news content.

[0033] The designed large language model prompt words were combined with news articles. The specific content and metadata are input into the large language model to obtain corresponding auxiliary information. :

[0034] ,

[0035] in, This indicates the large language models used (such as Chat-GPT, LLaMA, DeepSeek, etc.); ultimately, a news dataset is constructed. Corresponding auxiliary information set .

[0036] Here's an example of a prompt template: I'll provide you with the text content and related metadata of a news article. You need to generate a corresponding topic introduction and text description based on this content. Furthermore, you'll need to analyze the news's credibility, verifiability, and source credibility.

[0037] Reasonableness assessment: Determine whether the news content is reasonable and consistent with common sense. Does the news article contain logically consistent elements, or are there any contradictions?

[0038] Verifiability: This analyzes whether the content stated in the news article can be independently verified through reliable sources. Are there specific references, evidence, or data available for cross-checking to confirm the accuracy of the information?

[0039] News Source Credibility: Assess the credibility of the source from which the news originates. Is the source reputable and known for publishing reliable news, or does it have a history of publishing false or biased content?

[0040] News content: $News content$.

[0041] News metadata: $News metadata$.

[0042] S103. Associate and store the original news text with the generated auxiliary information to construct a structured dataset. Each dataset is saved as a dataset file in JSON format. Each data sample contains the following attributes: original news text ID, preprocessed news text content, generated topic auxiliary information, generated background knowledge auxiliary information, news release time, and news authenticity label (true / false); specifically in the following format:

[0043] {

[0044] "news_id": "1",

[0045] "content": "Preprocessed news text",

[0046] "theme_aux": "Generated topic summary",

[0047] "knowledge_aux": "Generated background knowledge",

[0048] "time": "2025-01-01",

[0049] "label": 1 / / 0 = false, 1 = true

[0050] };

[0051] S104. The processed dataset is trained using five-fold cross-validation. Therefore, the model automatically and randomly divides the dataset into five equal parts. During each fold of training, one fold is used as the test set data, and the remaining four folds are used as the training set data. All model parameters are initialized before the start of each fold of training to avoid data leakage.

[0052] It is understandable that step S100 realizes data preprocessing and auxiliary information generation: obtaining news text datasets and related metadata, performing text cleaning and standardization processing; and using large language models to generate news topic descriptions, credibility analysis and background knowledge as auxiliary information.

[0053] In step S200 of some embodiments of the present invention, extracting news features and auxiliary information features from the news text data and corresponding news auxiliary information using a pre-trained model includes:

[0054] S201. Extract news features and auxiliary information features from the news text data and corresponding news auxiliary information using two pre-trained language models.

[0055] Specifically, a pre-trained language model (preferably the BERT-based-Chinese model for Chinese data and the BERT-based-uncased model for English data) is used to encode the pre-processed news text. First, the news text content is input into the BERT model, and the hidden state at the [CLS] position is obtained as the initial representation vector of the news text. The same BERT model (without parameter sharing) is used to obtain the initial representation vector for the auxiliary information. .

[0056] In step S202 of some embodiments of the present invention, the extraction of topic features, content features, and auxiliary features from the news features and the auxiliary information features by means of multiple autoencoders includes:

[0057] S2021. Extract topic features from the news features using a first autoencoder;

[0058] S2022. Extract content features from the news features using a second autoencoder;

[0059] S2023. Extract auxiliary features from the auxiliary information features using a third autoencoder.

[0060] Specifically, a single-hidden-layer multilayer perceptron (MLP) is constructed as an autoencoder: the input is an initial representation vector of the text content (feature dimension 768 layers), which undergoes feature transformation through two fully connected layers. A non-linear mapping is achieved using the GeLU activation function, compressing the output dimension to a fraction of the original dimension. The output features are subjected to L1 norm constraints to force feature vector sparsity, retaining the most representative key content features. Dropout (with a probability of 0.1) and Layer Normalization are applied to the encoding layer to enhance the model's generalization ability.

[0061] By setting two independent parameters for the autoencoder network and Extract the common module features of the same topic from news articles of the same batch size. and content-specific module features Simultaneously set the third parameter to a non-shared autoencoder. Extract features of the same dimension from auxiliary information and align the output dimension with specific dimensions of topic features and content features.

[0062] ,

[0063] ,

[0064] ,

[0065] in, , , .

[0066] It is understandable that step S200 first uses a pre-trained model (BERT model) to extract news and auxiliary information features. Then, it extracts semantic features shared across news items through a topic-general encoder; captures authenticity discrimination features through a content feature encoder; and processes auxiliary information generated by the large language model through an auxiliary encoder.

[0067] In step S301 of some embodiments of the present invention, the fusion of topic features and auxiliary features based on the bidirectional attention mechanism includes:

[0068] S3011. Calculate the cross-modal attention matrix of topic features and auxiliary features, and generate the first attention weights through a preset activation function and learnable parameters; calculate the topic feature representation perceived by the auxiliary features through the first attention weights;

[0069] Specifically, compute the general features of the topic. Features of auxiliary information The cross-modal attention matrix is ​​obtained and processed using the SoftMax activation function and learnable parameters. Generate attention weights Therefore, the general feature representation of the topic for auxiliary information perception is calculated. :

[0070] ,

[0071] ，

[0072] 。

[0073] in Q, K, V The matrix parameters represent the cross-modal attention matrix;

[0074] S3012. Calculate the cross-modal attention matrix of topic features and auxiliary features, and generate a second attention weight through a preset activation function and learnable parameters; calculate the auxiliary feature representation of topic features perception through the second attention weight;

[0075] Specifically, the auxiliary information features are calculated using the same method as in step S2021. and common features of the theme Reverse attention weights And thereby calculate the auxiliary information feature representation of topic general perception. :

[0076] ，

[0077] ，

[0078] in, These are learnable parameters.

[0079] S3013. Integrate the topic feature representation perceived by auxiliary features and the auxiliary feature representation perceived by topic features. Perform feature fusion on the above-mentioned general topic features and the interaction attention calculation results of auxiliary information features, and concatenate the bidirectional interaction results: By generating separation constraint signals:

[0080] ,

[0081] in, These are learnable parameters.

[0082] It is understandable that step S300 achieves the separation of general theme features and content-specific features: establishing a two-way attention mechanism to realize deep interaction between content features and auxiliary features; and constructing separation constraint signals to ensure the independence of general theme features and content-specific features.

[0083] In step S400 of some embodiments of the present invention, constructing the loss function based on the separation constraint signal and the fused features includes:

[0084] S401. Construct a detection loss function based on the fused features;

[0085] Specifically, the detection loss function is calculated to optimize the realism classification. Similar to the calculation process in step S300, two types of bidirectional features are calculated: "specific content-auxiliary feature interaction" and "auxiliary feature-specific content feature interaction". and :

[0086] ，

[0087] ，

[0088] ，

[0089] ，

[0090] in, and These are learnable parameters.

[0091] The two bidirectional features are concatenated: News prediction results are generated through a fully connected layer:

[0092] ,

[0093] in, These are learnable parameters.

[0094] After calculating the prediction result of the current news sample, it is compared with the label of that news sample. Calculate the cross-entropy loss function, perform back-tuning of the model, and identify the learnable parameters during the optimization process:

[0095] .

[0096] S402. Construct a decoupling loss function based on the separation constraint signal;

[0097] Specifically, the separation constraint signal obtained from the final calculation in step S300 is combined with... By matching the news sample tags Calculate the cross-entropy loss function, perform inverse tuning of the model, and minimize it. and The overlap of features between the two modules allows the model to more effectively decouple topic-general features from content-specific features:

[0098] .

[0099] S403. Based on topic features and auxiliary features, construct a consistency loss function and a reconstruction loss function.

[0100] Specifically, consistency loss prompts the development of general features of the topic. Content-specific features The previous representations were aligned. This process is achieved through contrastive learning, ensuring that the news representations remain consistent with relevant topics while differentiating themselves from irrelevant news. Unlike the detection loss in step S41, this loss acts as a self-supervised signal, requiring no detection labels. .

[0101] First, auxiliary features As anchor points (providing auxiliary information for reference), positive and negative sample pairs are constructed using samples from the same batch, and then... Common features of the corresponding theme The positive news items provide valid instances, while other news items in the same batch provide invalid instances as negative samples. A cosine similarity function is used. And set the temperature over-parameter. Calculate the consistency loss:

[0102] ,

[0103] in, Indicates the current batch.

[0104] Specifically, the reconstruction loss function preserves key semantics. The autoencoder's encoding process must both retain key information and reduce noise in the features, thus requiring the simultaneous use of... and To reconstruct the characteristics of news Similarly, it should be applied to the use of Reconstructing auxiliary information features To further optimize the extracted features, an additional sparse hyperparameter is introduced. This controls the maximum hidden layer size, thereby achieving sparsity and improving the interpretability of the learned representations.

[0105] The reconstruction loss formula is:

[0106] ,

[0107] Sparse constraints imposed on features , T The preset threshold is used; where, and This represents the decoder in an autoencoder. This indicates the size of the hidden layer for that feature.

[0108] Based on the above four types of loss functions, the end-to-end model training process is as follows: The total loss function of the method is:

[0109] ,

[0110] in, To balance the hyperparameters of various losses, The parameter setting range is [0.1, 0.9]; the preferred parameter configuration is: .

[0111] The model training process uses a gradient descent optimization algorithm (such as Adam) for optimization, with an initial learning rate set to... A hierarchical learning rate strategy is employed: the learning rate for the encoder layer is set to... The attention head learning rate is set to The learning rate for the classification head is set to To prevent overfitting during model training, the patience value is set to 5, and the F1 score on the test set is monitored to determine whether to continue training or terminate the process.

[0112] It is understandable that step S400 achieves joint optimization of the objectives: calculating the detection loss function to optimize realism classification; designing a separation loss function to strengthen feature decoupling; constructing a consistency loss function to maintain topic stability; and implementing a reconstruction loss function to preserve key semantics. This involves jointly optimizing four loss functions; employing a hierarchical learning rate strategy; and implementing an early stopping mechanism to prevent overfitting.

[0113] It should be noted that this invention decomposes news content into three complementary representations: shared topic features, authenticity discrimination features, and external knowledge features. It utilizes auxiliary information generated by a large language model as semantic anchors and achieves deep interaction between content features and external knowledge through a bidirectional attention mechanism. A four-fold loss function, including detection, separation, consistency, and reconstruction, is designed to maintain semantic integrity and model stability while ensuring feature decoupling. This invention can effectively identify the typical "mixed truth and falsehood" feature patterns in fake news, significantly improving the detection capability for emerging topics. Simultaneously, it enhances the interpretability of the model through sparse feature space construction, providing an efficient and reliable technical solution for content security governance.

[0114] Example 2

[0115] refer to Figure 3 In a second aspect, the present invention provides a fake news detection system 1 based on semantic decoupling and knowledge enhancement, comprising:

[0116] The acquisition module 11 is used to acquire news text data; preprocess the news text data to obtain standardized news text data and corresponding news auxiliary information; the auxiliary information includes topic description, credibility and background knowledge.

[0117] Extraction module 12 is used to extract news features and auxiliary information features from the news text data and corresponding news auxiliary information through a pre-trained model; and to extract topic features, content features and auxiliary features from the news features and the auxiliary information features through multiple autoencoders.

[0118] Fusion module 13 is used to fuse topic features and auxiliary features based on a bidirectional attention mechanism; and to calculate separation constraint signals based on the fused features.

[0119] Training module 14 is used to construct a loss function based on the separation constraint signal and the fused features; and to jointly train the multiple autoencoders through the loss function; detection module 15 is used to detect the target news through the trained multiple autoencoders.

[0120] Furthermore, the fusion module 13 includes: a first calculation unit, used to calculate the cross-modal attention matrix of topic features and auxiliary features, and generate a first attention weight through a preset activation function and learnable parameters; calculate the topic feature representation perceived by auxiliary features through the first attention weight; a second calculation unit, used to calculate the cross-modal attention matrix of topic features and auxiliary features, and generate a second attention weight through a preset activation function and learnable parameters; calculate the auxiliary feature representation perceived by topic features through the second attention weight; and a fusion unit, used to fuse the topic feature representation perceived by auxiliary features and the auxiliary feature representation perceived by topic features.

[0121] Example 3

[0122] refer to Figure 4 A third aspect of the present invention provides an electronic device comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the fake news detection method based on semantic decoupling and knowledge enhancement of the first aspect of the present invention.

[0123] Electronic device 500 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 501, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 502 or a program loaded from storage device 508 into random access memory (RAM) 503. The RAM 503 also stores various programs and data required for the operation of electronic device 500. The processing unit 501, ROM 502, and RAM 503 are interconnected via bus 504. An input / output (I / O) interface 505 is also connected to bus 504.

[0124] Typically, the following devices can be connected to I / O interface 505: input devices 506 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 507 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 508 including, for example, hard disks; and communication devices 509. Communication device 509 allows electronic device 500 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 An electronic device 500 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 4 Each box shown can represent a device or multiple devices as needed.

[0125] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 509, or installed from a storage device 508, or installed from a ROM 502. When the computer program is executed by a processing device 501, it performs the functions defined in the methods of embodiments of this disclosure. It should be noted that the computer-readable medium described in embodiments of this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A 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 may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, 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 devices, magnetic storage devices, or any suitable combination thereof. In embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0126] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more computer programs, which, when executed by the electronic device, cause the electronic device to:

[0127] Computer program code for performing the operations of embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, C++, and Python—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0128] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0129] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting fake news based on semantic decoupling and knowledge enhancement, characterized in that, The method comprises the following steps: obtaining news text data; preprocessing the news text data to obtain standardized news text data and corresponding news auxiliary information; the auxiliary information includes theme description, credibility and background knowledge; extracting news features and auxiliary information features from the news text data and corresponding news auxiliary information through a pre-trained model; extracting theme features, content features and auxiliary features from the news features and the auxiliary information features through multiple autoencoders; fusing the theme features and the auxiliary features based on a bidirectional attention mechanism; calculating a separation constraint signal based on the fused features; constructing a loss function based on the separation constraint signal and the fused features: constructing a detection loss function based on the fused features; constructing a decoupling loss function based on the separation constraint signal; constructing a consistency loss function and a reconstruction loss function based on the theme features and the auxiliary features; training the multiple autoencoders through the loss function; the decoupling loss function is represented as: ， ， ， wherein, denotes a separation constraint signal, denotes a news sample label denote a topic general feature representation and an auxiliary information feature representation, respectively, are learnable parameters; detecting target news through the trained multiple autoencoders.

2. The method for fake news detection based on semantic decoupling and knowledge enhancement according to claim 1, characterized in that, The method comprises the following steps: extracting news features and auxiliary information features from the news text data and corresponding news auxiliary information through two pre-trained language models.

3. The method for fake news detection based on semantic decoupling and knowledge enhancement according to claim 1, characterized in that, The method comprises the following steps: extracting theme features from the news features through a first autoencoder; extracting content features from the news features through a second autoencoder; extracting auxiliary features from the auxiliary information features through a third autoencoder.

4. The method for fake news detection based on semantic decoupling and knowledge enhancement according to claim 1, characterized in that, The method comprises the following steps: calculating a cross-modal attention matrix of the theme features and the auxiliary features, and generating a first attention weight through a preset activation function and a learnable parameter; calculating a theme feature representation perceived by the auxiliary features through the first attention weight; calculating a cross-modal attention matrix of the theme features and the auxiliary features, and generating a second attention weight through a preset activation function and a learnable parameter; calculating an auxiliary feature representation perceived by the theme features through the second attention weight; fusing the theme feature representation perceived by the auxiliary features and the auxiliary feature representation perceived by the theme features.

5. The method for fake news detection based on semantic decoupling and knowledge enhancement according to claim 4, characterized in that, The method comprises the following steps: calculating a separation constraint signal through a multilayer perceptron and a preset activation function based on the fused features. 6.A false news detection system based on semantic decoupling and knowledge enhancement, characterized in that, The method comprises the following steps: an acquisition module is configured to obtain news text data; preprocessing the news text data to obtain standardized news text data and corresponding news auxiliary information; the auxiliary information includes theme description, credibility and background knowledge; an extraction module is configured to extract news features and auxiliary information features from the news text data and corresponding news auxiliary information through a pre-trained model; extracting theme features, content features and auxiliary features from the news features and the auxiliary information features through multiple autoencoders; a fusion module configured to fuse the topic feature and the auxiliary feature based on a bidirectional attention mechanism, and calculate a separation constraint signal based on the fused feature; a training module configured to construct a loss function based on the separation constraint signal and the fused feature, construct a detection loss function based on the fused feature, construct a decoupling loss function based on the separation constraint signal, construct a consistency loss function and a reconstruction loss function based on the topic feature and the auxiliary feature, and train the multiple autoencoders based on the loss functions; the decoupling loss function is represented as: a detection module configured to detect target news based on the trained multiple autoencoders. ， ， ， wherein, denotes a separation constraint signal, denotes a news sample label denotes a topic general feature representation and an auxiliary information feature representation, respectively, are learnable parameters; The fusion module comprises:

7. The system for fake news detection based on semantic decoupling and knowledge enhancement according to claim 6, characterized in that, a first calculation unit configured to calculate a cross-modal attention matrix of the topic feature and the auxiliary feature, generate a first attention weight based on a preset activation function and a learnable parameter, and calculate a topic feature representation perceived by the auxiliary feature based on the first attention weight; a second calculation unit configured to calculate a cross-modal attention matrix of the topic feature and the auxiliary feature, generate a second attention weight based on a preset activation function and a learnable parameter, and calculate an auxiliary feature representation perceived by the topic feature based on the second attention weight; a fusion unit configured to fuse the topic feature representation perceived by the auxiliary feature and the auxiliary feature representation perceived by the topic feature. one or more processors; 8. An electronic device, comprising: a storage device configured to store one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the method for detecting fake news based on semantic decoupling and knowledge enhancement according to any one of claims 1 to 5. The computer program, when executed by the processor, implements the method for detecting fake news based on semantic decoupling and knowledge enhancement according to any one of claims 1 to 5.

9. A computer readable medium having stored thereon a computer program, wherein, ​

Images