False news real-time detection method and system fusing background knowledge and user characteristics

Abstract

The invention provides a false news real-time detection method and system fusing background knowledge and user features, and the method comprises the steps: carrying out the feature extraction of a text of a to-be-detected post through a pre-training language model and a gating loop unit, and generating the text semantic features of the post; obtaining entity interpretation in the post text and entity concept information in the post text through knowledge distillation; performing secondary enhancement on the final user features to obtain re-enhanced user features; performing feature mining on the historical information of the user by using an emotion value calculation method to obtain final user features, and performing secondary enhancement on the final user features to obtain re-enhanced user features; inputting the double-gating fusion features into a multi-layer perceptron, and performing false news detection classification by connecting a normalized exponential function to obtain a classification probability; according to the method, fine-grained alignment of text-user features is realized through double-gating cross fusion, and cross-modal correlation errors are effectively reduced.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and natural language processing, and particularly to a real-time false news detection method and system that integrates background knowledge and user characteristics. Background Art

[0002] With the rapid development of social media, false news, characterized by its fast spread and wide coverage, has brought many negative impacts to society. Among existing false news detection methods, models based on propagation characteristics (such as forwarding path analysis and user interaction network modeling) require complete propagation chain data, resulting in significant detection delays and difficulty in meeting the real-time detection requirements. Detection methods based on text content, on the other hand, are limited in context semantic modeling and implicit intention capture due to the short, semantically sparse, and noisy content of social media posts (such as emojis and informal language), making it difficult to achieve efficient and accurate false news identification.

[0003] Existing methods have significant limitations in modeling user behavior characteristics: on the one hand, most studies only focus on the shallow social attributes of users (such as activity level and number of followers), while ignoring the deep behavior characteristics of users in the spread of false news (such as professionalism, rationality value, and historical credibility), resulting in a single-dimensional user portrait; on the other hand, the influence characteristics of the user social network (such as dynamic propagation weight and key node identification) have not been fully explored, making it difficult to locate highly influential users at the early stage of dissemination for precise interception. In addition, existing methods do not consider the dynamic evolution characteristics of entity relationships (such as the change of entity association strength over time windows and cross-modal entity alignment), resulting in semantic modeling lagging behind the rapid evolution of false news content.

[0004] Traditional methods have obvious deficiencies in multi-modal data fusion and classification robustness: firstly, external background knowledge (such as entity concepts and domain terms in knowledge graphs) has not been deeply integrated, and cannot alleviate the problem of text semantic sparsity; secondly, single-task learning frameworks are difficult to capture the complex associations between multi-modal features (text, user behavior, and temporal dynamics), resulting in the loss of high-order semantic information; finally, existing hierarchical fusion methods are mostly static weighted splicing, without achieving adaptive interaction and noise suppression of cross-modal features, and lacking a dynamic calibration mechanism for confidence distribution during classification decision-making, resulting in insufficient robustness of the model to fuzzy samples and noise interference, and high false alarm and miss rate. Summary of the Invention

[0005] In view of the above situation, the main purpose of the present invention is to propose a real-time false news detection method and system that integrates background knowledge and user characteristics to solve the above technical problems.

[0006] The present invention proposes a real-time false news detection method that integrates background knowledge and user characteristics, and the method includes the following steps: Step 1: Use a pre-trained language model and a gated recurrent unit to extract features from the text of the post to be detected, and generate the text semantic features of the post; Step 2: Perform entity recognition on the text of the post to be detected to obtain the recognized key entity information, and map the recognized key entity information to the knowledge base through entity linking technology for entity node correspondence operations, so as to obtain the entity interpretation in the post text and the entity concept information in the post text; Step 3: Use a pre-trained language model to fuse the entity interpretation in the post text, the entity concept information in the post text, and the text semantic features of the post through a cross-attention network and a weighted graph attention network to obtain the final knowledge-enhanced semantic features; Perform secondary enhancement on the final knowledge-enhanced semantic features to obtain the semantic features with enhanced knowledge again; Step 4: Use a sentiment value calculation method to mine features from the user's historical information to obtain the final user features; Perform secondary enhancement on the final user features to obtain the user features with enhanced again; Step 5: Perform dual-gated feature fusion on the knowledge semantic features with enhanced again and the user features with enhanced again to obtain the dual-gated fusion features; Step 6: Input the dual-gated fusion features into a multi-layer perceptron to generate a classification result.

[0007] The present invention also proposes a real-time fake news detection system that fuses background knowledge and user features, and the system includes: A knowledge distillation module for: Use a pre-trained language model and a gated recurrent unit to extract features from the text of the post to be detected, and generate the text semantic features of the post; Perform entity recognition on the text of the post to be detected to obtain the recognized key entity information, and map the recognized key entity information to the knowledge base through entity linking technology for entity node correspondence operations, so as to obtain the entity interpretation in the post text and the entity concept information in the post text; A knowledge fusion module for: Use a pre-trained language model to fuse the entity interpretation in the post text, the entity concept information in the post text, and the text semantic features of the post through a cross-attention network and a weighted graph attention network to obtain the final knowledge-enhanced semantic features; Perform secondary enhancement on the final knowledge-enhanced semantic features to obtain the semantic features with enhanced knowledge again; A user feature learning module for: Use a sentiment value calculation method to mine features from the user's historical information to obtain the final user features; Perform secondary enhancement on the final user features to obtain further enhanced user features; Perform dual-gated feature fusion on the further enhanced knowledge semantic features and the further enhanced user features to obtain dual-gated fusion features; Result classification module, for: Input the dual-gated fusion features into a multi-layer perceptron to generate a classification result.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Through low-rank projection and cross-subspace attention mechanism, the present invention captures multi-granularity semantic associations, combines parametric entropy value gating to adaptively focus on key features, improves the metaphor and noise text modeling ability, and enhances the robustness in sparse scenarios; realizes high-order semantic retention through elastic residual fusion, and solves the problem of information loss in traditional hierarchical fusion; 2. Based on rational value extension vector and professional degree bilinear interpolation, the present invention constructs a dynamic mask matrix for exponential enhancement, quantifies the evolution trend of user behavior, and improves the recognition accuracy of key communication nodes; 3. Through dual-gated cross-fusion, the present invention realizes fine-grained alignment of text-user features, effectively reduces cross-modal association errors; uses a dynamic temperature classifier combined with multi-order non-linear projection to adaptively calibrate the confidence distribution, effectively reduces the false alarm rate while improving the classification accuracy of fuzzy samples.

[0009] The additional aspects and advantages of the present invention will be partially given in the following description, partially become apparent from the following description, or be understood through the embodiments of the present invention. Brief Description of the Drawings

[0010] Figure 1 It is a step flow chart of the multi-modal fake news detection method based on dual-evidence enhancement and text-image similarity perception proposed by the present invention.

[0011] Figure 2 It is an architecture diagram of the fake news real-time detection method that fuses background knowledge and user features proposed by the present invention.

[0012] Figure 3 It is a flow chart of the secondary enhancement of the final knowledge semantic features of the fake news real-time detection method that fuses background knowledge and user features proposed by the present invention.

[0013] Figure 4 It is a flow chart of the secondary enhancement of the final user features of the fake news real-time detection method that fuses background knowledge and user features proposed by the present invention.

[0014] Figure 5 It is a schematic diagram of the fake news real-time detection system that fuses background knowledge and user features proposed by the present invention. Detailed implementation manners

[0015] The embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present invention and should not be construed as limiting the present invention.

[0016] These and other aspects of the embodiments of the present invention will be clear with reference to the following description and drawings. In these descriptions and drawings, some specific implementation manners in the embodiments of the present invention are specifically disclosed as some ways to implement the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0017] Please refer to Figure 1 , an embodiment of the present invention proposes a real-time false news detection method that fuses background knowledge and user characteristics. The method includes the following steps: Step 1: Use a pre-trained language model and a gated recurrent unit to extract features from the text of the post to be detected, and generate the text semantic features of the post.

[0018] Please refer to Figure 2 , in the said Step 1, use a pre-trained language model and a gated recurrent unit to extract features from the text of the post to be detected, and generate the text semantic features of the post. The corresponding relationship in the process is as follows: ; Among them, represents the text of the post to be detected, represents the gated recurrent unit, represents the text of the post to be detected, represents the pre-trained language model.

[0019] Step 2: Perform entity recognition on the text of the post to be detected to obtain the identified key entity information, and map the identified key entity information to the knowledge base through entity linking technology for entity node corresponding operations, so as to obtain the entity interpretation in the post text and the entity concept information in the post text.

[0020] Step 3: Use the pre-trained language model to fuse the entity interpretation in the post text, the entity concept information in the post text, and the text semantic features of the post through a cross-attention network and a weighted graph attention network to obtain the final knowledge-enhanced semantic features; Perform secondary enhancement on the final knowledge-enhanced semantic features to obtain the semantic features with enhanced knowledge again.

[0021] In step 3, the pre-trained language model is used to fuse the entity explanations in the post text, the entity concept information in the post text, and the text semantic features of the post through a cross-attention network and a weighted graph attention network to obtain the final knowledge-enhanced semantic features. The specific steps are as follows: The pre-trained language model is used to perform semantic extraction on the entity explanations in the post text to obtain the semantic feature representation of the post entity explanations; The semantic feature representation of the post entity explanations and the text semantic features of the post are fused through a cross-attention mechanism to obtain the fused features of the post text semantics and entity explanation semantics; The point mutual information calculation operation and the text transformation operation are respectively performed on the post text to obtain the point mutual information between words and the transformed graph structure; A discrimination operation is performed on the point mutual information between words to obtain the adjacency matrix; The entity concept information in the post text is used to supplement the nodes of the transformed graph structure to obtain the graph structure with nodes supplemented; The weighted graph attention network is used to perform representation learning on the graph structure with nodes supplemented to obtain the updated features of the nodes; An aggregation operation is performed on the updated features of the nodes to obtain the final text graph feature representation; The final text graph feature representation, the fused features of the post text semantics and entity explanation semantics are concatenated to obtain the final knowledge-enhanced semantic features.

[0022] In the step of using the pre-trained language model to perform semantic extraction on the entity explanations in the post text to obtain the semantic feature representation of the post entity explanations, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the semantic feature representation of the post entity explanations, represents the entity explanations in the post text obtained through knowledge distillation.

[0023] In the step of fusing the semantic feature representation of the post entity explanations and the text semantic features of the post through a cross-attention mechanism to obtain the fused features of the post text semantics and entity explanation semantics, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the fused features of the post text semantics and entity explanation semantics, represents the normalized exponential function, represents being processed through the cross-attention mechanism, represents the query parameters of, Represents a key parameter, represents a value parameter, represents the feature dimension of the parameter.

[0024] In the steps of performing point mutual information calculation operations and text conversion operations on the text of the post respectively, to obtain the point mutual information between words and the converted graph structure, the relational expressions existing in the corresponding processes are as follows: ; where, represents the marginal probability of the th word , represents the marginal probability of the th word , represents the number of times the word appears in the corpus, represents the total number of times all words in the corpus appear, represents the word and the word joint probability, represents the word and the word the number of times they appear simultaneously in the context, represents the logarithmic function, represents the word and the word point mutual information, represents the converted graph structure, represents the set of nodes composed of words, represents the set of edges indicating the degree of association between nodes.

[0025] In the step of performing a discrimination operation on the point mutual information between words to obtain the adjacency matrix, the relational expressions existing in the corresponding process are as follows: ; where, represents the value of the word and the word in the adjacency matrix .

[0026] In the step of performing representation learning on the graph structure supplemented by nodes using a weighted graph attention network to obtain the updated features of the nodes, the relational expressions existing in the corresponding process are as follows: ; where, represents the th node and the The degree of association of a node indicating the th node and the th node The degree of association of indicating the activation function indicating the first learning parameter matrix indicating the feature representation of node indicating the feature representation of node indicating the feature representation of node and node The attention weight value between indicating the set of neighbor nodes of node indicating the exponential function indicating the updated feature of node indicating the activation function indicating the concatenation operation

[0027] In the step of aggregating the updated features of the nodes to obtain the final text graph feature representation, the relational expressions in the corresponding process are as follows: ; wherein indicates the final text graph feature representation

[0028] In the step of concatenating the final text graph feature representation, the fused feature of the post text semantics and the entity interpretation semantics to obtain the final knowledge-enhanced semantic feature, the relational expressions in the corresponding process are as follows: ; wherein indicates the final knowledge-enhanced semantic feature indicating being processed by concatenation

[0029] Please refer to Figure 3 for the secondary enhancement of the final knowledge-enhanced semantic feature to obtain the knowledge-reinforced semantic feature, which specifically includes the following steps: Perform low-rank projection matrix processing and Gaussian error linear unit activation processing on the final knowledge-enhanced semantic feature in sequence to obtain the semantic subspace feature. The relational expressions in the corresponding process are as follows: ; wherein indicates the ​​​​​A semantic subspace feature, represents the Gaussian error linear unit activation function, represents the th low-rank projection matrix ; Performing cross-subspace attention calculation on the semantic subspace feature to obtain a cross-subspace attention weight matrix, and the relational expressions existing in the corresponding process are as follows: ; Among them, represents the cross-subspace attention weight matrix, represents the first semantic subspace feature, represents the second semantic subspace feature, represents the transpose of the second semantic subspace feature , represents the dimension of the final knowledge-enhanced semantic feature; Using the feature reconstruction matrix and the cross-subspace attention weight to reconstruct the semantic subspace feature to obtain a reconstructed semantic focus feature, and the relational expressions existing in the corresponding process are as follows: ; Among them, represents the reconstructed semantic focus feature, represents the layer normalization operation, represents the feature reconstruction matrix; Introducing a parameterized entropy value mixing mechanism to generate a dynamic gating vector, and the relational expressions existing in the corresponding process are as follows: ; Among them, represents the dimension of the dynamic gating vector, represents the sigmoid activation function, represents the information entropy calculation of the input feature vector with dimension , and respectively represent different learnable gating parameters, represents the logarithmic function, represents the distribution probability of class calculated by a sliding window , represents the index of the class calculated by the window, represents the hyperbolic tangent function, represents the entropy value scaling coefficient; Using the dynamic gating vector to perform an element-wise multiplication operation on the final knowledge-enhanced semantic feature to obtain a gated-enhanced semantic feature, and the relational expression of the corresponding process is as follows: ; Among them, represents the gated enhanced semantic feature, represents element-wise multiplication; Perform a feature splicing operation on the reconstructed semantic focus feature and the gated enhanced semantic feature to obtain the spliced fusion feature, and perform feature fusion again on the spliced fusion feature and the final knowledge-enhanced semantic feature to obtain the knowledge-reinforced semantic feature. The relational expressions for the corresponding process are as follows: ; Among them, represents the knowledge-reinforced semantic feature, represents the fusion through the fully connected layer after feature splicing.

[0030] Step 4: Use the sentiment value calculation method to perform feature mining on the user's historical information to obtain the final user feature; Perform secondary enhancement on the final user feature to obtain the further enhanced user feature.

[0031] Please refer to Figure 4 , in Step 4, use the sentiment value calculation method to perform feature mining on the user's historical information to obtain the final user feature, which specifically includes the following steps: Extract features from the user's historical information to obtain the user's basic social features; Based on the user feature learning module, perform sentiment calculation on the user's historical information to obtain the user's rational value; Perform topic distribution mining on the user's historical information through the Latent Dirichlet Allocation to obtain the topic distribution mining result, and perform professionalism calculation on the topic distribution mining result to obtain the user's professionalism feature; Perform splicing processing and embedding mapping operations on the user's rational value, the user's professionalism feature, and the user's basic social features in sequence to obtain the final user feature.

[0032] In the step of performing sentiment calculation on the user's historical information based on the user feature learning module to obtain the user's rational value, the relational expressions for the corresponding process are as follows: ; Among them, represents the th post 's sentiment value, represents the sentiment word library, represents the th word 's sentiment value, represents the word 's degree value of modifying the sentiment, represents the post The controversy and respectively represent the number of positive and negative sentiment comments, represents the user rational value, represents the number of historical posts published by the user.

[0033] In the step of mining the topic distribution of the user's historical information through the Latent Dirichlet Allocation to obtain the topic distribution mining result, and calculating the professionalism of the topic distribution mining result to obtain the professionalism characteristics of the user, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the user professionalism of the post , represents the cosine function, represents the topic probability distribution of the post , represents the topic probability distribution of the post .

[0034] In the step of sequentially performing splicing processing and embedding mapping operations on the user's rational value, the user's professionalism characteristics, and the user's basic social characteristics to obtain the final user characteristics, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the final user characteristics, represents the embedding representation function, represents the basic social characteristics of the user.

[0035] Perform secondary enhancement on the final user characteristics to obtain the user characteristics with enhanced again, which specifically includes the following steps: Perform expansion processing on the user's rational value to obtain the expanded vector, and the relational expressions existing in the corresponding process are as follows: ; Among them, represents the expanded vector; Perform bilinear interpolation operation on the user's professionalism characteristics to obtain the characteristics after bilinear interpolation, and the relational expressions existing in the corresponding process are as follows: ; Among them, represents the characteristics after bilinear interpolation, represents the bilinear interpolation operation; Activate the final user feature combination gating learning parameter to obtain the activated feature. Concatenate the extended vector and the bilinearly interpolated feature to obtain the concatenated feature of extension and bilinear interpolation. Perform bitwise multiplication on the activated feature and the concatenated feature of extension and bilinear interpolation to obtain the attribute-aware dynamic mask matrix. The corresponding relationship in the process is as follows: ; Among them, represents the attribute-aware dynamic mask matrix, represents the second learnable parameter; Use the attribute-aware dynamic mask matrix to perform exponential enhancement on the final user feature to obtain the exponentially enhanced user feature. The corresponding relationship in the process is as follows: ; Among them, represents the exponentially enhanced user feature, represents the global enhancement factor, represents element-wise exponential operation; Perform elastic residual mechanism processing on the final user feature, the exponentially enhanced user feature, the user's rational value, and the user's professionalism feature to generate the further enhanced user feature. The corresponding relationship in the process is as follows: ; Among them, represents the elastic mixing coefficient, represents the further enhanced user feature, represents the steepness coefficient, represents the normalized rational value, represents the mean of the professionalism feature.

[0036] Step 5: Perform dual-gating feature fusion on the further enhanced knowledge semantic feature and the further enhanced user feature to obtain the dual-gating fusion feature.

[0037] In Step 5, perform dual-gating feature fusion on the further enhanced knowledge semantic feature and the further enhanced user feature to obtain the dual-gating fusion feature, which specifically includes the following steps: Perform gating generation processing on the knowledge further enhanced semantic feature and the further enhanced user feature respectively to obtain the knowledge feature credibility contribution gating and the user feature personalized adaptation gating; Perform feature crossing between the credibility contribution degree gating of knowledge features and the semantic features after knowledge re-enhancement to obtain knowledge-enhanced crossing features. Perform feature crossing between the personalized adaptation gating of user features and the user features after re-enhancement to obtain user-enhanced crossing features. Perform an addition operation on the knowledge-enhanced crossing features and the user-enhanced crossing features to obtain double-gating crossing features; Perform residual gating aggregation processing on the semantic features after knowledge re-enhancement and the user features after re-enhancement in combination with the double-gating crossing features to obtain double-gating fusion features.

[0038] In the step of performing gating generation processing on the semantic features after knowledge re-enhancement and the user features after re-enhancement respectively to obtain the credibility contribution degree gating of knowledge features and the personalized adaptation gating of user features, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the credibility contribution degree gating of knowledge features, represents the personalized adaptation gating of user features, represents the first learnable weight matrix, represents the second learnable weight matrix, represents the first bias term, represents the second bias term; In the step of performing feature crossing between the credibility contribution degree gating of knowledge features and the semantic features after knowledge re-enhancement to obtain knowledge-enhanced crossing features, performing feature crossing between the personalized adaptation gating of user features and the user features after re-enhancement to obtain user-enhanced crossing features, and performing an addition operation on the knowledge-enhanced crossing features and the user-enhanced crossing features to obtain double-gating crossing features, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the double-gating crossing features; In the step of performing residual gating aggregation processing on the semantic features after knowledge re-enhancement and the user features after re-enhancement in combination with the double-gating crossing features to obtain double-gating fusion features, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the weight coefficient for balancing the crossing features and the residual features, represents the double-gating fusion features, represents the third learnable weight matrix.

[0039] Step 6: Input the double-gating fusion features into a multi-layer perceptron to generate a classification result.

[0040] In step 6, the dual-gated fusion features are input into a multi-layer perceptron to generate a classification result, which specifically includes the following steps: The weight value calculation, residual connection, and scaling mechanism processing are sequentially performed on the dual-gated fusion features to obtain the scaled enhanced features. The relationship in the corresponding process is as follows: ; Where, represents the weight value, represents the fourth learnable weight matrix, represents the third bias term, represents the scaled enhanced features; The scaled enhanced features are input into a multi-layer perceptron for first-order non-linear projection processing to obtain the first-order non-linear projection processing result. The relationship in the corresponding process is as follows: ; Where, represents the first-order non-linear projection processing result, represents the fifth learnable weight matrix, represents the fourth bias term; The first-order non-linear projection processing result is subjected to second-order non-linear projection processing to obtain the second-order non-linear projection result. The relationship in the corresponding process is as follows: ; Where, represents the second-order non-linear projection result, represents the sixth learnable weight matrix, represents the fifth bias term; The second-order non-linear projection result is fused with the scaled enhanced features to obtain the final classification features. The relationship in the corresponding process is as follows: ; Where, represents the final classification features; Based on the final classification features, a temperature parameter is generated, and a classification probability is generated based on the temperature parameter. The relationship in the corresponding process is as follows: ; Where, represents the temperature parameter, represents the transpose of the learnable vector , represents the classification probability, represents the classification weight matrix.

[0041] Please refer to Figure 5, the present invention also proposes a real-time fake news detection system that integrates background knowledge and user characteristics, and the system includes: A knowledge distillation module, which is used for: Using a pre-trained language model and a gated recurrent unit to extract features from the text of the post to be detected, and generating the text semantic features of the post; Performing entity recognition on the text of the post to be detected, obtaining the identified key entity information, and mapping the identified key entity information to the knowledge base through entity linking technology for entity node corresponding operations, so as to obtain the entity interpretation in the post text and the entity concept information in the post text; A knowledge fusion module, which is used for: Using a pre-trained language model to fuse the entity interpretation in the post text, the entity concept information in the post text, and the text semantic features of the post through a cross-attention network and a weighted graph attention network to obtain the final knowledge-enhanced semantic features; Performing secondary enhancement on the final knowledge-enhanced semantic features to obtain the knowledge-reinforced semantic features; A user feature learning module, which is used for: Using a sentiment value calculation method to mine features from the user's historical information to obtain the final user features; Performing secondary strengthening on the final user features to obtain the re-enhanced user features; Performing double-gated feature fusion on the re-enhanced knowledge semantic features and the re-enhanced user features to obtain double-gated fusion features; A result classification module, which is used for: Inputting the double-gated fusion features into a multi-layer perceptron to generate a classification result.

[0042] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following well-known technologies in the art can be used: discrete logic circuits with logic gate circuits for implementing logic functions on data signals, application-specific integrated circuits with suitable combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0043] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0044] The above-described embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several variations and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the appended claims.

Claims

1. A real-time false news detection method that integrates background knowledge and user characteristics, characterized in that The method includes the following steps: Step 1: Use a pre-trained language model and a gated recurrent unit to extract features from the text of the post to be detected, generating the text semantic features of the post; Step 2: Perform entity recognition on the text of the post to be detected to obtain the identified key entity information, and map the identified key entity information to the knowledge base through entity linking technology for entity node correspondence operations, so as to obtain the entity explanations in the post text and the entity concept information in the post text; Step 3: Use a pre-trained language model to fuse the entity explanations in the post text, the entity concept information in the post text, and the text semantic features of the post through a cross-attention network and a weighted graph attention network to obtain the final knowledge-enhanced semantic features; Perform secondary enhancement on the final knowledge-enhanced semantic features to obtain the semantic features with enhanced knowledge again; Step 4: Use a sentiment value calculation method to mine features from the user's historical information to obtain the final user features; Perform secondary enhancement on the final user features to obtain the user features with enhanced again; Step 5: Perform double-gated feature fusion on the knowledge semantic features with enhanced again and the user features with enhanced again to obtain double-gated fusion features; Step 6: Input the double-gated fusion features into a multi-layer perceptron to generate a classification result.

2. The real-time false news detection method that integrates background knowledge and user characteristics according to claim 1, wherein In the said Step 3, using a pre-trained language model to fuse the entity explanations in the post text, the entity concept information in the post text, and the text semantic features of the post through a cross-attention network and a weighted graph attention network to obtain the final knowledge-enhanced semantic features, specifically includes the following steps: Use a pre-trained language model to perform semantic extraction on the entity explanations in the post text to obtain the semantic feature representation of the post entity explanations; Fuse the semantic feature representation of the post entity explanations and the text semantic features of the post through a cross-attention mechanism to obtain the fusion feature of the post text semantics and entity explanation semantics; Perform pointwise mutual information calculation operations and text transformation operations on the text of the post respectively to obtain the pointwise mutual information between words and the transformed graph structure; Perform a discrimination operation on the pointwise mutual information between words to obtain an adjacency matrix; Supplement the nodes of the transformed graph structure with the entity concept information in the post text to obtain the graph structure with supplemented nodes; Use a weighted graph attention network to perform representation learning on the graph structure with supplemented nodes to obtain the updated features of the nodes; Perform an aggregation operation on the updated features of the nodes to obtain the final text graph feature representation; Perform a concatenation process on the final text graph feature representation, the post text semantics, and the fusion feature of the entity explanation semantics to obtain the final knowledge-enhanced semantic features.

3. The real-time false news detection method that integrates background knowledge and user characteristics according to claim 2, wherein, In the step of using a pre-trained language model to perform semantic extraction on the entity explanations in the post text to obtain the semantic feature representation of the post entity explanations, the relational expressions existing in the corresponding process are as follows: ； Among them, represents the semantic feature representation of the post entity explanation, represents the entity explanation in the post text obtained through knowledge distillation, represents the pre-trained language model; In the step of fusing the semantic feature representation of the post entity explanations and the text semantic features of the post through a cross-attention mechanism to obtain the fusion feature of the post text semantics and entity explanation semantics, the relational expressions existing in the corresponding process are as follows: ； Among them, represents the fusion feature of the post text semantics and the entity interpretation semantics, represents the normalization exponential function, represents being processed by the cross-attention mechanism, represents the query parameters, represents the key parameters, represents the value parameters, represents the feature dimension of the parameters; In the steps of performing point mutual information calculation operations and text transformation operations on the text of the post respectively to obtain the point mutual information between words and the transformed graph structure, the relational expressions existing in the corresponding processes are as follows: ； Among them, represents the marginal probability of the n-th word, represents the marginal probability of the m-th word, represents the number of times the word appears in the corpus, represents the total number of times all words appear in the corpus, represents the joint probability of the word and the word represents the number of times the word and the word appear simultaneously in the context, represents the logarithmic function, represents the pointwise mutual information of the word and the word represents the transformed graph structure, represents the set of nodes composed of words, represents the set of edges indicating the degree of association between nodes; In the step of performing a discrimination operation on the point mutual information between words to obtain an adjacency matrix, the relational expressions existing in the corresponding process are as follows: ； Among them, represents the word and the word in the adjacency matrix value; In the step of performing representation learning on the graph structure supplemented by nodes using a weighted graph attention network to obtain the updated features of the nodes, the relational expressions existing in the corresponding process are as follows: ； Among them, represents the th node and the th node correlation degree, represents the th node and the th node correlation degree, represents the activation function, represents the first learning parameter matrix, represents the node feature representation, represents the node feature representation, represents the node and the node attention weight value between them, represents the neighbor node set of the node , represents the exponential function, represents the node updated feature, represents the activation function, represents the concatenation operation; In the step of performing an aggregation operation on the updated features of the nodes to obtain the final text graph feature representation, the relational expressions existing in the corresponding process are as follows: ； Among them, represents the final text graph feature representation; In the step of concatenating the final text graph feature representation, the fusion feature of the post text semantics and the entity interpretation semantics to obtain the final knowledge-enhanced semantic feature, the relational expressions existing in the corresponding process are as follows: ； Among them, represents the final knowledge-enhanced semantic feature, indicating that splicing processing has been performed.

4. The real-time false news detection method integrating background knowledge and user characteristics according to claim 3, characterized in that Perform secondary enhancement on the final knowledge-enhanced semantic feature to obtain the semantically re-enhanced knowledge feature, which specifically includes the following steps: Perform low-rank projection matrix processing and Gaussian error linear unit activation processing on the final knowledge-enhanced semantic feature in sequence to obtain the semantic subspace feature, and the relational expressions existing in the corresponding process are as follows: ； Among them, represents the th semantic subspace feature, represents the Gaussian error linear unit activation function, represents the th low-rank projection matrix ; Perform cross-subspace attention calculation on the semantic subspace feature to obtain the cross-subspace attention weight matrix, and the relational expressions existing in the corresponding process are as follows: ； Among them, represents the cross-subspace attention weight matrix, represents the first semantic subspace feature, represents the second semantic subspace feature, represents the transpose of the second semantic subspace feature , represents the dimension of the final knowledge-enhanced semantic feature; Use the feature reconstruction matrix combined with the cross-subspace attention weight to perform feature reconstruction on the semantic subspace feature to obtain the reconstructed semantic focus feature, and the relational expressions existing in the corresponding process are as follows: ； Among them, represents the reconstructed semantic focus feature, represents the layer normalization operation, represents the feature reconstruction matrix; Introduce a parameterized entropy value mixing mechanism to generate a dynamic gating vector, and the relational expressions existing in the corresponding process are as follows: ； Among them, represents the dynamic gating vector of dimension , represents the sigmoid activation function, represents the dimension for calculating the information entropy of the input feature vector, and respectively represent different learnable gating parameters, represents the logarithmic function, represents the class calculated by sliding window of the distribution probability represents the index of the class calculated by the window, represents the hyperbolic tangent function, represents the entropy value scaling coefficient; Use the dynamic gating vector to perform an element-wise multiplication operation on the final knowledge-enhanced semantic feature to obtain the gated enhanced semantic feature, and the relational expressions in the corresponding process are as follows: ； Among them, represents the gated enhanced semantic feature, represents element-wise multiplication; Perform a feature concatenation operation on the reconstructed semantic focus feature and the gated enhanced semantic feature to obtain the concatenated fusion feature, and perform feature fusion on the concatenated fusion feature and the final knowledge-enhanced semantic feature again to obtain the semantically re-enhanced knowledge feature, and the relational expressions existing in the corresponding process are as follows: ； Among them, represents the semantic feature of knowledge re-enhancement, represents the fusion through the fully connected layer after feature splicing.

5. The real-time false news detection method integrating background knowledge and user characteristics according to claim 4, characterized in that, In step 4, use the sentiment value calculation method to perform feature mining on the user historical information to obtain the final user feature, which specifically includes the following steps: Extract features from the user historical information to obtain the basic social features of the user; Based on the user feature learning module, perform sentiment calculation on the user historical information to obtain the rational value of the user, and the relational expressions existing in the corresponding process are as follows: ； Among them, represents the sentiment value of the nth post, represents the sentiment lexicon, represents the sentiment value of the nth word, represents the degree value of the word modifying the sentiment, represents the controversiality of the post, and respectively represent the number of positive and negative sentiment comments, represents the rationality value of the user, represents the number of historical posts published by the user; Perform topic distribution mining on the user historical information through the latent Dirichlet distribution to obtain the topic distribution mining result, and perform professionalism calculation on the topic distribution mining result to obtain the professionalism feature of the user, and the relational expressions existing in the corresponding process are as follows: ； Among them, represents the user 's professionalism towards the post ; represents the cosine function, represents the topic probability distribution of the post ; represents the topic probability distribution of the post ; Perform concatenation processing and embedding mapping operations on the rational value of the user, the professionalism feature of the user, and the basic social features of the user in sequence to obtain the final user feature, and the relational expressions existing in the corresponding process are as follows: ； Among them, represents the final user characteristics, represents the embedding representation function, represents the basic social characteristics of the user.

6. The real-time false news detection method integrating background knowledge and user characteristics according to claim 5, characterized in that, Perform secondary strengthening on the final user feature to obtain the re-enhanced user feature, which specifically includes the following steps: Perform an expansion process on the rational value of the user to obtain an expanded vector. The relational expressions in the corresponding process are as follows: ； Among them, represents the extended vector; Perform a bilinear interpolation operation on the professionalism feature of the user to obtain the bilinearly interpolated feature. The relational expressions in the corresponding process are as follows: ； Among them, represents the feature after bilinear interpolation, represents the bilinear interpolation operation; Perform an activation process on the final user feature combined with the gating learning parameter to obtain the activated feature. Concatenate the expanded vector and the bilinearly interpolated feature to obtain the concatenated feature of expansion and bilinear interpolation. Perform a bitwise multiplication process on the activated feature and the concatenated feature of expansion and bilinear interpolation to obtain the attribute-aware dynamic mask matrix. The relational expressions in the corresponding process are as follows: ； Among them, represents the attribute-aware dynamic mask matrix, represents the second learnable parameter; Use the attribute-aware dynamic mask matrix to perform exponential enhancement on the final user feature to obtain the exponentially enhanced user feature. The relational expressions in the corresponding process are as follows: ； Among them, represents the user feature with exponential enhancement, represents the global enhancement factor, represents the element-wise exponential operation; Perform an elastic residual mechanism process on the final user feature, the exponentially enhanced user feature, the rational value of the user, and the professionalism feature of the user to generate the further enhanced user feature. The relational expressions in the corresponding process are as follows: ； Among them, represents the elastic mixing coefficient, represents the re-enhanced user characteristics, represents the steepness coefficient, represents the normalized rational value, represents the mean of the professionalism characteristics.

7. The real-time false news detection method that integrates background knowledge and user characteristics according to claim 6, characterized in that In step 5, perform a double-gating feature fusion on the further enhanced knowledge semantic feature and the further enhanced user feature to obtain the double-gating fusion feature. The specific steps are as follows: Perform a gating generation process on the knowledge further enhanced semantic feature and the further enhanced user feature respectively to obtain the knowledge feature credibility contribution gating and the user feature personalized adaptation gating; Perform a feature cross process between the knowledge feature credibility contribution gating and the knowledge further enhanced semantic feature to obtain the knowledge enhanced cross feature. Perform a feature cross process between the user feature personalized adaptation gating and the further enhanced user feature to obtain the user enhanced cross feature. Perform an addition operation on the knowledge enhanced cross feature and the user enhanced cross feature to obtain the double-gating cross feature; Perform a residual gating aggregation process on the knowledge further enhanced semantic feature and the further enhanced user feature combined with the double-gating cross feature to obtain the double-gating fusion feature.

8. The real-time false news detection method that integrates background knowledge and user characteristics according to claim 7, characterized in that In the step of performing a gating generation process on the knowledge further enhanced semantic feature and the further enhanced user feature respectively to obtain the knowledge feature credibility contribution gating and the user feature personalized adaptation gating, the relational expressions in the corresponding process are as follows: ； Among them, represents the knowledge feature credible contribution degree gating, represents the user feature personalized adaptation gating, represents the first learnable weight matrix, represents the second learnable weight matrix, represents the first bias term, represents the second bias term; In the step of performing a feature cross process between the knowledge feature credibility contribution gating and the knowledge further enhanced semantic feature to obtain the knowledge enhanced cross feature, performing a feature cross process between the user feature personalized adaptation gating and the further enhanced user feature to obtain the user enhanced cross feature, and performing an addition operation on the knowledge enhanced cross feature and the user enhanced cross feature to obtain the double-gating cross feature, the relational expressions in the corresponding process are as follows: ； Among them, represents double-gated cross features; In the step of performing a residual gating aggregation process on the knowledge further enhanced semantic feature and the further enhanced user feature combined with the double-gating cross feature to obtain the double-gating fusion feature, the relational expressions in the corresponding process are as follows: ； Among them, represents the weight coefficient for balancing the cross feature and the residual feature, represents the double-gated fusion feature, represents the third learnable weight matrix.

9. The real-time false news detection method for integrating background knowledge and user characteristics according to claim 8, wherein In step 6, input the double-gating fusion feature into a multi-layer perceptron to generate a classification result. The specific steps are as follows: The weight value calculation, residual connection, and scaling mechanism processing are sequentially performed on the dual-gated fusion features to obtain the scaled enhanced features. The relational expressions existing in the corresponding process are as follows: ； Among them, represents the weight value, represents the fourth learnable weight matrix, represents the third bias term, represents the scaled enhanced feature; The scaled enhanced features are input into a multi-layer perceptron for the first-order non-linear projection processing to obtain the result of the first-order non-linear projection processing. The relational expressions existing in the corresponding process are as follows: ； Among them, represents the first-order non-linear projection processing result, represents the fifth learnable weight matrix, represents the fourth bias term; The result of the first-order non-linear projection processing is subjected to the second-order non-linear projection processing to obtain the result of the second-order non-linear projection. The relational expressions existing in the corresponding process are as follows: ； Among them, represents the second-order non-linear projection result, represents the sixth learnable weight matrix, represents the fifth bias term; The result of the second-order non-linear projection is fused with the scaled enhanced features to obtain the final classification features. The relational expressions existing in the corresponding process are as follows: ； Among them, represents the final classification feature; Based on the final classification features, a temperature parameter is generated, and a classification probability is generated based on the temperature parameter. The relational expressions existing in the corresponding process are as follows: ； Among them, represents the temperature parameter, represents the transpose of the learnable vector , represents the classification probability, represents the classification weight matrix.

10. A real-time fake news detection system that integrates background knowledge and user characteristics, characterized in that, The system applies the false news real-time detection method for fusing background knowledge and user features according to any one of claims 1 to 9 above. The system includes: A knowledge distillation module for: Using a pre-trained language model and a gated recurrent unit to extract features from the text of the post to be detected, and generating the text semantic features of the post; Performing entity recognition on the text of the post to be detected to obtain the identified key entity information, and mapping the identified key entity information to the knowledge base through entity linking technology for entity node corresponding operations, so as to obtain the entity interpretation in the post text and the entity concept information in the post text; A knowledge fusion module for: Using a pre-trained language model to fuse the entity interpretation in the post text, the entity concept information in the post text, and the text semantic features of the post through a cross-attention network and a weighted graph attention network to obtain the final knowledge-enhanced semantic features; Performing secondary enhancement on the final knowledge-enhanced semantic features to obtain the semantic features with knowledge re-enhanced; A user feature learning module for: Using a sentiment value calculation method to mine features from the user historical information to obtain the final user features; Performing secondary strengthening on the final user features to obtain the user features with re-enhanced; Performing dual-gated feature fusion on the re-enhanced knowledge semantic features and the re-enhanced user features to obtain dual-gated fusion features; A result classification module for: Inputting the dual-gated fusion features into a multi-layer perceptron to generate a classification result.

Images