False information detection method and system for knowledge-enhanced heterogeneous graph attention

Abstract

The invention relates to the technical field of information content security, in particular to a false information detection method and system for knowledge-enhanced heterogeneous graph attention, and the method comprises the steps: constructing a dependency relationship between different semantic units in a heterogeneous text graph modeling post and the correlation of external background knowledge; designing node feature representation based on a double-layer graph attention mechanism, and capturing semantic contribution degrees of different types of node neighborhoods to a target node and importance of nodes under the different types of node neighborhoods; a multi-objective optimization strategy is adopted, and features are automatically selected and aggregated through adaptive feature aggregation to obtain features after post-text fusion; and inputting the fused feature vector representation into a multi-layer perceptron, and judging the authenticity of the post content. According to the method, high-value clues can be mined from multiple dimensions such as presentation patterns and entity semantics of the post, the model detection performance is improved, and the semantic sparse problem caused by short text characteristics is relieved.

Description

Technical Field

[0001] The present invention relates to the field of information content security technology, and in particular to a false information detection method and system for knowledge-enhanced heterogeneous graph attention. Background Art

[0002] Due to the short text characteristics of social media and the diversity of false information, it is difficult to mine effective clues to identify the authenticity of information based solely on the content of the post itself, making the false information detection task for social media extremely challenging. In order to alleviate the semantic sparsity problem caused by the short text characteristics of social media, researchers have proposed a detection method based on content semantic enhancement, aiming to solve the problem of early detection of false information in low-resource scenarios. The core idea is that the key entity words contained in social network posts can reveal the core topics, and the introduction of background knowledge of these entities can enrich the semantic content of short texts. Therefore, how to obtain the background knowledge of the post text content and integrate this knowledge into the text information has become a key issue.

[0003] According to the different strategies for acquiring model knowledge, it can be roughly divided into two directions: methods based on pre-trained language models and methods based on knowledge-driven graph convolution. The core idea of ​​the method based on pre-trained language models to enhance the semantic understanding of content is to use the knowledge structure contained in the model to mine the potential language patterns and semantic clues in the posts through fine-tuning, prompt learning and other techniques, effectively enhancing the detection effect of false information. This method is only applicable to situations where relevant background knowledge has been fully involved in the process of model pre-training and fine-tuning. The knowledge-driven graph convolution detection method models the intrinsic semantic associations and external knowledge associations of text in the form of graph structures, and deeply mines the intrinsic meaning of key information elements in posts (such as people, organizations, and specific nouns); on this basis, high-order neighborhood information is deeply aggregated through multi-layer graph convolution operations to enhance the semantic representation ability of text and improve the detection performance of the model.

[0004] Although these methods have improved the detection performance of fake information to a certain extent, they are still insufficient in accurately describing the complex semantic features of text. First, existing methods focus on modeling the sequence relationship and co-occurrence relationship between words, while relatively ignoring the deep semantic dependency structure between words, such as modality relationship, agent-patient relationship, causal relationship, and basis relationship. These deep semantic relationships are crucial for the model to accurately capture the intrinsic relationship of information and discover key suspicious clues in fake information. Figure 1-(b) provides an example to further illustrate this point. By performing semantic dependency analysis on the posts and converting them into an intuitive dependency graph, the potentially misleading logical connections in the false information can be clearly revealed: that is, "muscle twitching" is mistakenly associated with the cause of "decreased physical function." Secondly, existing methods usually focus on single-perspective analysis (either focusing on mining the pattern features of posts, or focusing on the semantic understanding of text content), and lack the ability to capture fine-grained false information clues from multiple perspectives. Figure 1 -(c) A post is usually constructed from a diverse vocabulary set, which includes not only key entity words that reveal the core theme, such as "muscle twitch" and "physical function", but also auxiliary expression words that reflect the writing pattern and emotional color of the article, such as "certain", "attention", "very good", "indicate", etc. These different vocabulary types provide key insights for identifying false information from different perspectives. Summary of the invention

[0005] The present invention aims to solve the problem that the semantic sparsity of short texts in posts leads to insufficient fine-grained semantic perception of existing detection methods. A false information detection method and system with knowledge-enhanced heterogeneous graph attention is proposed, which can mine high-value clues from multiple dimensions such as the presentation mode and entity semantics of posts, improve the model detection performance, and alleviate to a certain extent the semantic sparsity problem brought by the characteristics of short texts and the challenges brought by the diversity of false information presentation forms.

[0006] To achieve the above purpose, the technical solution adopted is:

[0007] A false information detection method based on knowledge-enhanced heterogeneous graph attention includes the following steps:

[0008] By constructing a heterogeneous text graph model, the dependencies between different semantic units within the post and the association with external background knowledge are modeled;

[0009] Design node feature representation based on a two-layer graph attention mechanism to capture the semantic contribution of different types of node neighborhoods to the target node and the importance of nodes under different types of node neighborhoods;

[0010] A multi-objective optimization strategy is adopted to automatically select and aggregate features through adaptive feature aggregation to obtain the features after post fusion;

[0011] The fused feature vector representation is input into the multi-layer perceptron to judge the authenticity of the post content.

[0012] According to the false information detection method for enhancing heterogeneous graph attention based on the present invention, further, the process of constructing a heterogeneous text graph includes generating multiple types of nodes and initial vector representations of the nodes: first, an entity word set S is formed for each post te , pattern word set Stp and concept knowledge set S tk ; Initialized using the pre-trained model BERT in Represents the initial feature representation of all entity type words, represents the initial feature representation of all pattern type words, represents the initial feature representation of all concept descriptions; let Represents the initial node feature representation matrix composed of all nodes in the graph, |V|=2n+l, n represents the number of entities contained in the post, and l represents the number of pattern words contained in the post.

[0013] According to the false information detection method for enhancing heterogeneous graph attention based on the knowledge of the present invention, further, the process of constructing a heterogeneous text graph also includes constructing the association relationship between nodes based on the strategy of semantic dependency and co-occurrence association dual measurement:

[0014] Let A represent the adjacency matrix of the heterogeneous text graph. For any type of node u, the adjacency relationship between its corresponding different types of nodes v is expressed as follows:

[0015]

[0016] In the formula, SD<u,v> =1 indicates that there is a semantic dependency relationship between vocabulary node u and vocabulary node v. Mutual information PMI greater than 0 indicates that a co-occurrence edge is established between vocabulary node u and vocabulary node v. KE<u,v> =1 indicates that there is an entity concept description relationship between node u and node v.

[0017] According to the false information detection method of the present invention that enhances heterogeneous graph attention, further, the two-layer graph attention mechanism includes type-level attention and node-level attention:

[0018] Type-level attention is used to capture the semantic contribution of different types of node neighborhoods to the target node. The type-level attention weight calculation formula for node v is:

[0019]

[0020] In the formula, x v is the feature representation of the current node, x vt is the neighborhood representation of node v of type t, which is represented by the sum of the features of all nodes of type t adjacent to node v; is the learnable parameter vector in the type-level attention mechanism;

[0021] Node-level attention is to capture the importance of nodes in different types of node neighborhoods. The node-level attention weight formula of node v is:

[0022]

[0023] In the formula, x v′ Represents the feature representation of node v’s neighbor node v’ of type t, ν T Represents the learnable parameter vector in the node-level attention mechanism.

[0024] According to the false information detection method of enhancing heterogeneous graph attention according to the present invention, further, the node feature representation based on the double-layer graph attention mechanism includes:

[0025] The representation of all nodes in the heterogeneous text graph is updated by aggregating the features of their different types of neighboring nodes, which can be formally described as:

[0026]

[0027] In the formula, represents the embedded representation of all nodes after the attention operation of the l-th layer heterogeneous graph, σ(·) is the activation function, is the attention matrix of type t, where the rows represent all nodes in the graph and the columns represent all nodes of type t. The element value in the vth row and vth column represents the semantic influence of node v′ on node v. Represents the embedded representation of all nodes of type t after the l-1th layer of graph convolution operation; It is the transformation matrix of type t for the l-1th layer of graph convolution.

[0028] According to the false information detection method of knowledge-enhanced heterogeneous graph attention of the present invention, further, for entity type word nodes and pattern type word nodes, a learnable node weight layer is designed respectively to obtain knowledge-enhanced entity semantic feature representation and post pattern feature representation

[0029] According to the false information detection method for enhancing heterogeneous graph attention based on the knowledge of the present invention, further, the features obtained by automatically selecting and aggregating features through adaptive feature aggregation after post fusion include:

[0030] (1) Feature utility evaluation

[0031] By constructing a single feature utility evaluation task, the effectiveness score of each feature is obtained; through the entity semantic feature utility evaluator MLP based on multi-layer perceptron e (·), obtain the validity evaluation score S of the entity semantic feature e ; Through the multi-layer perceptron-based pattern feature utility evaluator MLP p (·), and obtain the effective classification evaluation score S of the pattern feature. p ;

[0032] (2) Adaptive aggregation of multi-view features

[0033] The weights of features from different perspectives are learned through a multi-layer perceptron to achieve dynamic optimization and aggregation of features. The process is as follows:

[0034] h=Sigmoid(MLP′ e (S e ))h e +Sigmoid(MLP′ p (S p ))h p

[0035] In the formula, h represents the fused features, MLP′ e (·) represents the entity semantic feature weight learning module, MLP′ p (·) represents the pattern feature weight learning module.

[0036] According to the false information detection method of knowledge-enhanced heterogeneous graph attention of the present invention, further, in the model training process, three cross entropy loss functions are used to quantify the deviation between the pattern feature prediction, the entity semantic feature prediction, and the final prediction after adaptive fusion of two perspective features and the actual label; the total loss function is constructed by the weighted sum of the above three loss functions.

[0037] Furthermore, the present invention also proposes a false information detection system for knowledge-enhanced heterogeneous graph attention, comprising:

[0038] Heterogeneous text graph construction module, which is used to model the dependency relationship between different semantic units in a post and the association of external background knowledge by constructing a heterogeneous text graph;

[0039] The two-layer graph attention node representation module is used to design node feature representation based on the two-layer graph attention mechanism to capture the semantic contribution of different types of node neighborhoods to the target node and the importance of nodes under different types of node neighborhoods;

[0040] The adaptive feature aggregation module is used to adopt a multi-objective optimization strategy to automatically select and aggregate features through adaptive feature aggregation to obtain the features after post fusion;

[0041] The prediction module is used to input the fused feature vector representation into the multi-layer perceptron to judge the authenticity of the post content.

[0042] The beneficial effects achieved by adopting the above technical solution are:

[0043] The present invention proposes a false information detection method KHGAT with knowledge-enhanced heterogeneous graph attention. The method constructs a heterogeneous text graph that integrates external knowledge to fine-grainedly characterize the intrinsic semantic dependency and external knowledge association of different semantic units of posts; designs a double-layer graph attention mechanism based on the heterogeneous text graph structure to fine-tune the pattern features of posts and the entity semantic features of knowledge enhancement; introduces an adaptive feature aggregation module to give the model the ability to automatically select and optimize feature combinations, so as to enhance the model's fine-grained semantic perception of the content of posts and its ability to capture false information clues from multiple perspectives. Experimental results show that KHGAT can effectively alleviate the challenges brought by the semantic sparsity of short texts to the task of false information detection based on post content. On two widely used social media datasets, Weibo and Twitter, the detection accuracy of the model is improved by an average of 3.3% compared with the fine-tuning of the large language model GLM3_6B, and the detection accuracy is improved by an average of 4.4% compared with the graph structure CompareNet method. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention, wherein the drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0045] Figure 1 This is a Weibo post example picture;

[0046] Figure 2 It is a flow chart of a false information detection method of knowledge-enhanced heterogeneous graph attention according to an embodiment of the present invention;

[0047] Figure 3 It is a description diagram of the heterogeneous text graph construction process of an embodiment of the present invention;

[0048] Figure 4 It is a heat map-based visual analysis of feature effectiveness according to an embodiment of the present invention;

[0049] Figure 5 This is an ablation experiment of KHGAT of an embodiment of the present invention on two data sets;

[0050] Figure 6 are the semantic dependencies with the top 10 loss ratios in the test data labeled Fake in the embodiment of the present invention;

[0051] Figure 7 It is a comparison diagram of the similarity between the semantic representations of entities and pattern nodes in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings of specific embodiments of the present invention to clearly and completely describe the exemplary scheme of the embodiment of the present invention. Unless otherwise defined, the technical terms or scientific terms used in the present invention should be the common meanings understood by people with ordinary skills in the field.

[0053] like Figure 2 As shown, this embodiment discloses a false information detection method for knowledge-enhanced heterogeneous graph attention, which includes the following contents:

[0054] Step S101: construct a heterogeneous text graph model to model the dependencies between different semantic units within the post and the association with external background knowledge.

[0055] Figure 3 The whole process of constructing heterogeneous text graph is given, which is divided into two parts: generating multi-type nodes and building relationships between nodes.

[0056] (1) Generate multiple types of nodes

[0057] A post is usually constructed from a diverse set of words, and different types of words carry different semantic functions. For example, entity words can reflect the key elements of the text content description, usually including people, organizations, places, and some proper nouns. In-depth understanding of these entities is crucial to portraying the deep semantic features of information. In addition, posts also contain pattern words that reflect their expression style and emotional tendencies, which are often reflected in abstract expression elements.

[0058] In view of this, for each social media post to be detected, we first use open source knowledge extraction tools to extract entity-related information such as people, places, organizations, proper nouns, etc. contained in the post to form a corresponding entity word set S te This scheme uses the open source DeepKE-LLM extraction tool from Zhejiang University; then, the remaining words in the post that are not included in the stop word list are regarded as the pattern word set S that describes the post writing mode. tp Due to the short text characteristics of social media, it is difficult for the model to deeply understand the intrinsic meaning of the core entity in a limited context. Through entity linking technology, the relevant concept description information is extracted from the external high-quality knowledge base to form a concept knowledge set S tk .

[0059] For the initial feature representation of different types of nodes in heterogeneous text graphs, the pre-trained model BERT is used to initialize the in Represents the initial feature representation of all entity type words, represents the initial feature representation of all pattern type words, represents the initial feature representation of all concept descriptions; let Represents the initial node feature representation matrix composed of all nodes in the graph, |V|=2n+l, n represents the number of entities contained in the post, and l represents the number of pattern words contained in the post.

[0060] (2) Building relationships between nodes

[0061] In order to deeply analyze and accurately describe the internal logical connection between various vocabulary units in the post, this scheme adopts a dual measurement strategy of semantic dependency and co-occurrence association to establish dependency relationships between different words. Specifically, the semantic dependency analysis technology is used to reveal the deep semantic dependency relationship between words, which is represented by SD. In order to better utilize the potential semantic connection between words and alleviate the sparsity problem of semantic dependency in text graphs, this scheme simplifies the dependency relationship between nodes and converts directed semantic dependency into undirected, that is: if there is such a dependency relationship between vocabulary node u and vocabulary node v, then define SD<u,v> =SD<v,u> =1, otherwise SD<u,v> =SD<v,u> = 0. Next, for the vocabulary node pairs without direct semantic dependency, the Pointwise Mutual Information (PMI) score between vocabulary node pairs is used to quantify the closeness of the association between them. The specific calculation formula is as follows:

[0062]

[0063]

[0064]

[0065] Among them, C(w) represents the number of sliding windows, C(u) represents the number of sliding windows containing vocabulary node u, C(u,v) represents the number of sliding windows containing vocabulary u and vocabulary v at the same time, p(u,v) represents the probability of co-occurrence of vocabulary u and vocabulary v, p(u) represents the percentage of sliding windows containing vocabulary u to the total number of sliding windows, and p(v) represents the percentage of sliding windows containing vocabulary v to the total number of sliding windows; co-occurrence association edges are established between vocabulary pairs with PMI greater than 0. At the same time, in order to model the external knowledge association of the post, the extracted concept description is regarded as a node link to the corresponding entity, and KE is used to represent the relationship between entity concept descriptions. If there is an entity concept description relationship between node u and node v, then KE<u,v> =KE<v,u> =1, otherwise both are 0. In this way, the complex dependency relationship between different words in the post can be carefully displayed, and the external knowledge association of the post can also be reflected.

[0066] Let A represent the adjacency matrix of the heterogeneous text graph. For any type of node u, the adjacency relationship between its corresponding different types of nodes v is expressed as follows.

[0067]

[0068] Step S102: Design a node feature representation based on a two-layer graph attention mechanism to capture the semantic contribution of different types of node neighborhoods to the target node and the importance of nodes under different types of node neighborhoods, so as to enhance the representation of post patterns and entity semantic features.

[0069] For a node v in the text graph, from the perspective of semantic influence, adjacent nodes of different types may have different influences. For example, if the type of node v is entity type, intuitively, the adjacent concept description nodes and entity nodes of the same type contribute more to its semantic understanding than pattern neighbor nodes. Based on this, this scheme designs a node feature representation based on a two-layer graph attention mechanism to finely capture the semantic contribution of different types of node neighborhoods to the target node and the importance of nodes under different types of node neighborhoods, so as to enhance the model's ability to discern subtle differences in pattern and entity semantics.

[0070] (1) Type-level attention

[0071] Type-level attention is used to capture the semantic contribution of different types of node neighborhoods to the target node. For the target node v, its neighborhood features of type t can be represented by the sum of all node features of type t adjacent to node v. The calculation process can be formally described as:

[0072]

[0073] Among them, x vt represents the neighborhood representation of node v of type t, t∈T={te,tp,tk}. Let D represent the degree matrix of the text graph and I represent the identity matrix, then represents the normalized adjacency matrix. represents the neighbor relationship between node v and all nodes of type t in the graph, X t Represents the feature representation matrix of all types t.

[0074] Based on the feature representation x of the current node v The feature representation x with neighborhood type t vt , design a feedforward network with LeakyReLU activation function as the attention function, and normalize the type-level attention score through the softmax function to obtain the semantic importance of the neighborhood type t to the node v, that is, the type-level attention weight of the node v The formula is as follows:

[0075]

[0076] in, is the learnable parameter vector in the type-level attention mechanism.

[0077] (2) Node-level attention

[0078] Node-level attention is to capture the importance of nodes in different types of node neighborhoods. For a node v’s neighbor node of type t, we can represent the node x based on its feature representation. v and x v′ , calculate the node-level attention weight, which can be described as:

[0079]

[0080] Among them, ν T Represents the learnable parameter vector in the node-level attention mechanism.

[0081] (3) Neighborhood node aggregation of heterogeneous text graphs

[0082] The representation of all nodes in the heterogeneous text graph can be updated by aggregating the features of their different types of neighboring nodes, which can be formally described as:

[0083]

[0084] In the formula, represents the embedded representation of all nodes after the attention operation of the l-th layer heterogeneous graph, σ(·) is the activation function, is the attention matrix of type t, where the rows represent all nodes in the graph and the columns represent all nodes of type t. The element value in the vth row and vth column represents the semantic influence of node v′ on node v. Represents the embedded representation of all nodes of type t after the l-1th layer of graph convolution operation; is the transformation matrix of type t in the l-1th layer of graph convolution, which is used to transform the representations of different types of nodes into a common dense low-dimensional space.

[0085] After l layers of heterogeneous graph attention aggregation of different types of high-order neighborhood node features, we finally obtain all node representations in the heterogeneous text graph (including entity, pattern and concept descriptions). On this basis, we design a learnable node weight layer for entity type word nodes and pattern type word nodes respectively, which is used to aggregate all entity word node features to obtain knowledge-enhanced entity semantic feature representations. and post pattern feature representation The specific process can be formally described as:

[0086]

[0087] This process can automatically identify and emphasize entity words and pattern words that are crucial for false information detection, while weakening the influence of irrelevant words. Where σ(·) represents the Sigmoid activation function, represents the learnable parameter vector shared by all entity type node words, Represents the learnable parameter vector shared by all pattern type word nodes.

[0088] Step S103: adopt a multi-objective optimization strategy to automatically select and aggregate features through adaptive feature aggregation to obtain post-text fusion features.

[0089] Since the effectiveness of features from different perspectives varies depending on the type of false information, inappropriate combinations may weaken detection performance. Therefore, by designing adaptive feature aggregation and using feature judgment accuracy as the core criterion for measuring feature utility, the model is given the ability to automatically select and optimize feature combinations. Specifically, it includes two stages:

[0090] (1) Feature utility evaluation

[0091] In order to enable the model to quantify the effectiveness of different perspective features in false information discrimination, a multi-task joint learning strategy is adopted to obtain the effectiveness score of each feature by constructing a single feature utility evaluation task. Specifically, different perspective features are input into the evaluation model composed of a multi-layer perceptron.

[0092] Taking entity semantic feature utility evaluation as an example, the following formula shows the entire process of feature evaluation.

[0093] S e =Sigmoid(MLP e (h e ))

[0094] Among them, MLP e (·) represents the entity semantic feature utility evaluator based on multi-layer perceptron, S e Represents the effectiveness evaluation score of the entity semantic feature. Similarly, through the multi-layer perceptron-based pattern feature utility evaluator MLP p (·), obtain the effective classification evaluation score of the pattern feature: S p =Sigmoid(MLP p (h p )).

[0095] (2) Adaptive aggregation of multi-view features

[0096] Based on the assumption that features that can obtain correct judgments are more useful, the score of the feature predicting the correct label (i.e., the feature utility score) is used as the initial weight. On this basis, the weights of features from different perspectives are learned through a multi-layer perceptron to achieve dynamic optimization and aggregation of features. The specific process is as follows:

[0097] h=Sigmoid(MLP′ e (S e ))h e +Sigmoid(MLP′ p (S p ))h p

[0098] In the formula, h represents the fused features, MLP′ e (·) represents the entity semantic feature weight learning module, MLP′ p (·) represents the pattern feature weight learning module.

[0099] Step S104: Input the fused feature vector representation into a multi-layer perceptron to determine the authenticity of the post content.

[0100] The fused feature h is sent to the multi-layer perceptron to predict the authenticity of the post, as follows:

[0101]

[0102] During model training, three cross entropy loss functions are used to quantify the deviation between the pattern feature prediction, entity semantic feature prediction, and the final prediction after adaptive fusion of the two perspective features and the actual label. The specific formula is:

[0103]

[0104]

[0105] Among them, y i ∈{0,1} represents the true label of the i-th post, N represents the number of samples in each batch, represents the predicted probability of the true label of sample i based on the mixed features, represents the probability of predicting the true label of sample i based on entity semantic features, Represents the probability of predicting the true label of sample i based on the pattern features.

[0106] The overall loss function of the training process is obtained by the weighted summation of the above-mentioned loss functions:

[0107] l=l cls +αl e +βl p

[0108] Among them, α and β are hyperparameters, and their values ​​are between 0 and 1. Before backpropagation, the loss of each batch is averaged.

[0109] Corresponding to the above method, this embodiment also proposes a false information detection system for knowledge-enhanced heterogeneous graph attention, comprising:

[0110] The heterogeneous text graph construction module is used to model the dependencies between different semantic units within a post and the association with external background knowledge by constructing a heterogeneous text graph.

[0111] The two-layer graph attention node representation module is used to design node feature representation based on the two-layer graph attention mechanism to capture the semantic contribution of different types of node neighborhoods to the target node and the importance of nodes under different types of node neighborhoods.

[0112] The adaptive feature aggregation module is used to adopt a multi-objective optimization strategy to automatically select and aggregate features through adaptive feature aggregation to obtain the features after post fusion.

[0113] The prediction module is used to input the fused feature vector representation into the multi-layer perceptron to judge the authenticity of the post content.

[0114] In order to verify the effectiveness and accuracy of the model in a real social media environment, extensive experiments were conducted on two representative social media datasets, Weibo and Twitter. Both datasets contain two label categories, where Real means that the content of the information is true, and Fake means that the information is fake. Before formally training and testing the model, the two datasets were cleaned and preprocessed: first, the data was deduplicated according to the content of the post and the length of the data, and duplicate samples and samples with a length of less than 5 were deleted; secondly, the open source entity extraction tool was used to extract the names of people, places, institutions, events, proper nouns, etc. contained in the text. The specific statistical description of the preprocessed dataset is shown in Table 1.

[0115] Table 1 Statistical description of the two datasets Weibo and Twitter

[0116]

[0117] 1. Experimental setup

[0118] (1) Experimental environment

[0119] The machine configuration and environment for all experiments are: Intel(R) Xeon(R) Platinum 8474C (processor), 80GB (memory), 32G (GPU), and all codes are implemented in Python 3.8, Pytorch (2.0.1), and Cuda 11.8.

[0120] (2) Baseline comparison model

[0121] In order to verify the effectiveness of the proposed model, the proposed method is compared with some advanced baseline models. These methods can be divided into baseline methods based on pre-trained language models and false information detection methods based on graph neural networks.

[0122] ① Detection method based on pre-trained language model: With the help of the powerful context perception ability of the pre-trained model and the knowledge embedded in the model parameters, the language patterns and semantic details of the post are captured to improve the detection efficiency of the model.

[0123] BERT-Emo: Use the pre-trained model BERT to encode text semantics and characterize the emotional characteristics of the information publisher. Since the focus is on the content characteristics of the information itself, a simplified version of this method is used in the comparative experiment, that is, the emotional tendency of the information commentator is not considered.

[0124] GLM-6B-LORA: GLM-6B is an open source conversational language model that supports Chinese and English question-answering, with 6.2 billion parameters. The false information detection task is converted into a question-answering format, and the model is fine-tuned using the LORA method. The scaling factor is set to 3.12%, using 16-bit floating point operations.

[0125] LLM-Hiss Prompting: Based on the basic idea of ​​prompt learning, a hierarchical step-by-step prompting template is designed to guide LLMs to divide a piece of information into several sub-information, and then gradually verify each sub-information through multiple questions. Due to the short text characteristics of post data, in the comparative experiment, we used core entities instead of self-information for step-by-step verification. The large model mainly uses the API interface of Zhipu Qingyan Large Model (GLM-4-Flash).

[0126] ② Detection method based on graph neural network: This method converts the text into a graph structure and uses graph neural network to model the relationship between different semantic units of the text in a more fine-grained manner, thereby obtaining text semantic representation.

[0127] KMGCN: For each piece of multimodal information input, the model first obtains the related concept description based on the entity linking technology, and splices it with the original post as a supplement to the text content. Based on the co-occurrence relationship between words, a text graph is constructed, and on this basis, a two-layer graph convolutional neural network is used to obtain the semantic representation of the post.

[0128] CompareNet: This model first models posts as a heterogeneous graph containing multiple types of nodes such as topics, sentences, and entities. Secondly, it learns topic-rich sentence semantic representations and context-based entity semantic representations through a heterogeneous graph attention network. Finally, it designs a semantic comparison network to compare the consistency between the contextual entity semantic representation and the semantic representation of the external knowledge base, thereby improving the performance of false information detection.

[0129] KDRD: Converts social media text into a semantic-entity graph, where nodes not only contain words in the original text, but also expand semantic words by using world knowledge and language knowledge base, thus enriching the content of social media text; on this basis, through a two-layer graph convolution operation, the overall semantic representation of the post is obtained.

[0130] (3) Model parameter configuration

[0131] The parameters in the model proposed in the present invention are updated by the Adam algorithm, and the learning rate is initialized to 0.00025. The pre-trained BERT model (bert-base-Chinese for Chinese and bert-base-uncased for English) is used to initialize the semantic representation of different types of vocabulary nodes, and the initial dimension is 768. The dimensions d of the pattern word node representation and entity word node representation based on the double-layer graph attention in the model are both set to 200, the batchsize of the experiment is 10, and the epoch is 25. In order to prevent gradient vanishing and overfitting during training, the dropout mechanism is used in the model with a ratio of 0.5. The number of layers l of heterogeneous graph attention is 2. The values ​​of the overall loss hyperparameters α and β are 0.5 and 0.5, respectively.

[0132] 2. Model Detection Performance

[0133] Experiments are carried out from two perspectives: feature visualization analysis and result comparison analysis to demonstrate the effectiveness of the method proposed in this invention.

[0134] 1) Visual analysis: Explore the ability of implicit features learned by the model to judge the authenticity of information.

[0135] In order to analyze the ability of the multi-dimensional implicit features extracted by the model to determine the authenticity of information, 5 real information and 5 false information were randomly selected as samples in the test set. First, the semantic feature h of each sample was calculated. e , pattern feature hp and the pairwise similarities of the mixed feature h to form a similarity matrix; then, it is displayed in the form of a heat map, where each cell represents the cosine similarity value between the corresponding sample pairs, as shown in the following figure. Figure 4 .

[0136] It can be clearly observed that the entity semantic features, pattern features, and their mixed features obtained by the model all show relatively significant inter-class boundaries. This shows that these features all have a certain ability to discriminate the authenticity of information. Specifically, the heat map of the entity semantic feature is the most chaotic compared to the other three features. This shows that it is unstable to use only entity semantics for false information detection. The pattern feature shows strong intra-class similarity, but lacks in inter-class difference. This is because there is overlap between true and false information at the pattern feature level, resulting in insufficient discrimination. When the two different types of features are adaptively weighted and mixed, the mixed semantic feature shows a more significant inter-class difference, which can well distinguish true and false information from the feature space.

[0137] 2) Comparative analysis of results: revealing the pros and cons of the model’s performance in the task of false information detection.

[0138] The proposed method is compared with six representative models, and the performance of each model on the real data set is measured using four evaluation indicators: precision (Pre), recall (Re), F1 score, and overall accuracy (Acc). The specific data are shown in Table 2.

[0139] Table 2 Detection results of KHGAT and baseline models on two standard datasets (%)

[0140]

[0141] Note: The bold values ​​represent the best performance of each indicator.

[0142] By comparing and analyzing the test results of each model in Table 2, we can find that:

[0143] ①GLM_6B_LORA outperforms the other two methods with a higher overall accuracy (Acc) among the three detection methods based on pre-trained language models. However, its performance is not as good as the BERT_Emo model in terms of recall rate (Re) and F1 value evaluation indicators for false information detection. In particular, in the Weibo dataset, the false information recall rate of GLM_6B_LORA is only 0.583, which indicates that it tends to think that the information to be detected is true. In contrast, although BERT_Emo outperforms GLM_6B_LORA and LLM_Hiss Prompting in false information recall rate and F1 value indicators, it performs worst in terms of recall rate and F1 value in terms of true information judgment. This result shows that BERT_Emo tends to judge the information to be detected as false, and it also indirectly reflects that there may be overlap between true and false information at the level of emotional pattern features.

[0144] ② The overall detection performance of KDRD is about 1.65 percentage points higher than that of KMGN on average, which shows that enhancing the background knowledge of the core words of the post from multiple perspectives can improve the model's semantic understanding of the post content, thereby improving the overall detection performance of the model. The overall detection performance of the CompareNet model on two different datasets is better than that of the KMGN and KDRD models. This shows that heterogeneous text graphs can model the internal semantic associations of posts at a finer granularity than co-occurrence text graphs, thereby enhancing the semantic perception ability of the model.

[0145] ③ The overall performance of the proposed method KHGAT on the two data sets is 3.3 percentage points better than the GLM_6B_LORA model on average, and there is no significant tendency; compared with CompareNet, the detection performance is 4.4 percentage points higher on average. This is because the method of the present invention combines the advantages of graph neural networks and heterogeneous graphs, integrates the complex relationship between entities, patterns and external knowledge resources by constructing text heterogeneous graphs, and deeply understands information and extracts key clues for false information judgment from the two dimensions of entity semantics and performance patterns on the basis of knowledge-enhanced entity semantics; at the same time, through the adaptive feature aggregation module, the model is given the ability to automatically select and optimize feature combinations according to the characteristics of the post to be detected, thereby improving the detection performance of the model.

[0146] (III) Ablation experiment

[0147] In order to prove the effectiveness of each module of the model, a series of ablation experiments were conducted. It mainly includes three parts:

[0148] 1) w / o SD: remove semantic dependencies in heterogeneous text graphs and explore whether semantic dependencies can enhance the model's ability to mine suspicious clues of false information.

[0149] 2) w / o HA: Remove the node type level attention and convert the double-layer attention into a single-layer attention. We evaluate the ability of the double-layer graph attention mechanism to capture the differences in features of different types of words and refine the learning of post features.

[0150] 3) w / o AA: Remove the feature adaptive weighted aggregation module, concatenate and fuse the acquired entity semantic features with the pattern features, and verify the ability of the adaptive weighted module in improving the model's automatic selection and optimizing feature combinations.

[0151] Figure 5 The ablation experiment results of the KHGAT model on two datasets are shown. It can be seen from the figure:

[0152] First, the semantic dependencies in the text graph were deleted, and only the remaining word co-occurrence relationships and concept description relationships were retained, which caused the overall detection performance of the model on the two datasets to drop by an average of 2.3 percentage points (Acc), and the detection performance of fake information to drop by an average of 2.85 percentage points (F_f1). In order to further explore the reasons behind the performance degradation, the loss ratio of various semantic dependency relationships was further counted in detail for samples labeled Fake on the test set, and the top 10 relationships with the highest loss ratio in the two datasets were selected for display, as shown in the figure below. Figure 6 shown.

[0153] observe Figure 6 , we can find that many important semantic dependencies such as guest, cause, condition, modality, basis, etc. are lost in the text graph. The loss of these deep semantic dependency structures leads to a decline in model detection performance, which proves that these semantic relationships are crucial to enhancing the model's ability to identify potential clues of false information.

[0154] Secondly, after simplifying the dual-layer attention mechanism to a single-layer graph attention (removing the type-level attention), the overall detection accuracy (Acc) of the model decreased by an average of about 3.3 percentage points on both datasets, and the model's ability to identify false information decreased by an average of nearly 4.4 percentage points (F_f1). In order to explore the reasons for this experimental phenomenon, we further compared the semantic features of post entities h under the dual-layer graph attention and single-layer graph attention. e and pattern feature h p For intuitive display, 50 sample data are randomly selected from the test set to calculate the similarity between the entity semantics and the presentation mode features of each sample to verify the ability of the dual-layer attention to capture the differences between different types of features. Figure 7As shown. Generally, the smaller the similarity between two feature representations, the more the model can capture their differences at a more subtle level, and the stronger the model's ability to learn and identify refined features. The red nodes in the figure represent the similarity between the entity semantics and the pattern features of each sample after two layers of graph attention, and the green nodes represent the similarity between the entity semantics and the pattern features after only a single layer of graph attention.

[0155] according to Figure 7 It can be found that the similarity between entity semantics and pattern features after double-layer attention is mostly below 0.4; when it is replaced with single-layer graph attention, the similarity between the two is mostly above 0.4. Figure 7 , indicating that removing the two-layer attention module causes the model to lose the ability to finely capture the semantic contributions of lexical nodes of different categories, thereby weakening the model's ability to discern subtle differences between pattern features and entity semantics, ultimately affecting the overall effect of false information detection.

[0156] Finally, removing the feature adaptive aggregation module, the overall detection accuracy (Acc) of the model decreased by an average of 5.9 percentage points, and the F_f1 value decreased by an average of nearly 5.8 percentage points in identifying false information. This strongly proves that the adaptive feature aggregation module can significantly improve the model detection performance compared to feature splicing. Furthermore, in order to prove the contribution of a single loss in the multi-objective optimization paradigm proposed by the adaptive aggregation module in improving the automatic selection of the model and optimizing the feature combination, the loss of evaluating the utility of different features is gradually deleted for verification, as shown in Table 3.

[0157] Table 3 Analysis of individual loss contribution in multi-objective optimization

[0158]

[0159] Note: "√" indicates the introduction of the corresponding loss function, and "-" indicates that the corresponding loss function is not introduced. cls When , the feature fusion module directly degenerates into the addition of two features.

[0160] It can be observed that as more losses are deleted, the model performance gradually decreases. p and l e When there are two losses, the feature fusion module directly degenerates into the addition of two features, and the detection performance drops most significantly. This is because the removal of l p or e Loss, the model cannot adaptively capture the detection utility of different features according to the samples to be detected, and thus cannot optimize the combination of features. Therefore, the two losses are necessary to improve the model's ability to optimize features and automatically combine features. Figure 5,It can also be found that whether it is splicing or addition, compared with adaptive weighted aggregation, the detection performance of the model shows a significant decline.

[0161] Unless otherwise specifically stated, the components, steps, numerical expressions and values ​​set forth in these embodiments do not limit the scope of the present invention.

[0162] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0163] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.

[0164] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of combination of hardware and software.

[0165] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A false information detection method based on knowledge-enhanced heterogeneous graph attention, characterized in that: The following steps are involved: By constructing a heterogeneous text graph model, the dependencies between different semantic units within the post and the association with external background knowledge are modeled; Design node feature representation based on a two-layer graph attention mechanism to capture the semantic contribution of different types of node neighborhoods to the target node and the importance of nodes under different types of node neighborhoods; A multi-objective optimization strategy is adopted to automatically select and aggregate features through adaptive feature aggregation to obtain the features after post fusion; The fused feature vector representation is input into the multi-layer perceptron to judge the authenticity of the post content.

2. The false information detection method based on knowledge-enhanced heterogeneous graph attention according to claim 1 is characterized in that: The process of constructing a heterogeneous text graph includes generating multiple types of nodes and their initial vector representations: First, form an entity word set S for each post te , pattern word set S tp and concept knowledge set S tk ; Initialized using the pre-trained model BERT in Represents the initial feature representation of all entity type words, represents the initial feature representation of all pattern type words, represents the initial feature representation of all concept descriptions; let Represents the initial node feature representation matrix composed of all nodes in the graph, V|=2n+l, n represents the number of entities contained in the post, and l represents the number of pattern words contained in the post.

3. The false information detection method based on knowledge-enhanced heterogeneous graph attention according to claim 2 is characterized in that: The process of constructing a heterogeneous text graph also includes building node association relationships based on a dual measurement strategy of semantic dependency and co-occurrence association: Let A represent the adjacency matrix of the heterogeneous text graph. For any type of node u, the adjacency relationship between its corresponding different types of nodes v is expressed as follows: In the formula, SD<u,v> =1 indicates that there is a semantic dependency relationship between vocabulary node u and vocabulary node v. Mutual information PMI greater than 0 indicates that a co-occurrence edge is established between vocabulary node u and vocabulary node v. KE<u,v> =1 indicates that there is an entity concept description relationship between node u and node v.

4. The false information detection method based on knowledge-enhanced heterogeneous graph attention according to claim 1 is characterized in that: The two-layer graph attention mechanism includes type-level attention and node-level attention: Type-level attention is used to capture the semantic contribution of different types of node neighborhoods to the target node. The type-level attention weight calculation formula for node v is: In the formula, x v is the feature representation of the current node, x vt is the neighborhood representation of node v of type t, which is represented by the sum of the features of all nodes of type t adjacent to node v; is the learnable parameter vector in the type-level attention mechanism; Node-level attention is to capture the importance of nodes in different types of node neighborhoods. The node-level attention weight formula of node v is: In the formula, x v′ Represents the feature representation of node v’s neighbor node v’ of type t, ν T Represents the learnable parameter vector in the node-level attention mechanism.

5. The false information detection method of knowledge-enhanced heterogeneous graph attention according to claim 4 is characterized in that: The node feature representation based on the two-layer graph attention mechanism includes: The representation of all nodes in the heterogeneous text graph is updated by aggregating the features of their different types of neighboring nodes, which can be formally described as: In the formula, represents the embedded representation of all nodes after the attention operation of the l-th layer heterogeneous graph, σ(·) is the activation function, is the attention matrix of type t, where the rows represent all nodes in the graph and the columns represent all nodes of type t. The element value in the vth row and vth column represents the semantic influence of node v′ on node v. represents the embedded representation of all nodes of type t after the l-1th layer of graph convolution operation; W t (l-1) It is the transformation matrix of type t for the l-1th layer of graph convolution.

6. The false information detection method of knowledge-enhanced heterogeneous graph attention according to claim 5 is characterized in that: For entity type word nodes and pattern type word nodes, a learnable node weight layer is designed to obtain knowledge-enhanced entity semantic feature representation. and post pattern feature representation 7. The false information detection method of knowledge-enhanced heterogeneous graph attention according to claim 6 is characterized in that: The features obtained by automatically selecting and aggregating features through adaptive feature aggregation include: (1) Feature utility evaluation By constructing a single feature utility evaluation task, the effectiveness score of each feature is obtained; through the entity semantic feature utility evaluator MLP based on multi-layer perceptron e (·), obtain the validity evaluation score S of the entity semantic feature e ; Through the multi-layer perceptron-based pattern feature utility evaluator MLP p (·), and obtain the effective classification evaluation score S of the pattern feature. p ; (2) Adaptive aggregation of multi-view features The weights of features from different perspectives are learned through a multi-layer perceptron to achieve dynamic optimization and aggregation of features. The process is as follows: h=Sigmoid(MLP e '(S e ))h e +Sigmoid(MLP p '(S p ))h p In the formula, h represents the fused feature, MLP' e (·) represents the entity semantic feature weight learning module, MLP' p (·) represents the pattern feature weight learning module.

8. The false information detection method of knowledge-enhanced heterogeneous graph attention according to claim 1 is characterized in that: During the model training process, three cross entropy loss functions are used to quantify the deviation between pattern feature prediction, entity semantic feature prediction, and the final prediction after adaptive fusion of two perspective features and the actual label; the total loss function is constructed by the weighted sum of the above three loss functions.

9. A knowledge-enhanced heterogeneous graph attention false information detection system, characterized in that: Include: Heterogeneous text graph construction module, which is used to model the dependency relationship between different semantic units in a post and the association of external background knowledge by constructing a heterogeneous text graph; The two-layer graph attention node representation module is used to design node feature representation based on the two-layer graph attention mechanism to capture the semantic contribution of different types of node neighborhoods to the target node and the importance of nodes under different types of node neighborhoods; The adaptive feature aggregation module is used to adopt a multi-objective optimization strategy to automatically select and aggregate features through adaptive feature aggregation to obtain the features after post fusion; The prediction module is used to input the fused feature vector representation into the multi-layer perceptron to judge the authenticity of the post content.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images