A method and system for detecting false information in knowledge-enhanced heterogeneous graph attention.

Abstract

This invention relates to the field of information content security technology, and particularly to a method and system for detecting false information using knowledge-enhanced heterogeneous graph attention. It constructs a heterogeneous text graph to model the dependencies between different semantic units within a post and the association with external background knowledge. A node feature representation based on a two-layer graph attention mechanism is designed to capture the semantic contribution of different types of node neighborhoods to the target node and the importance of nodes within different types of node neighborhoods. A multi-objective optimization strategy is employed, using adaptive feature aggregation to automatically select and aggregate features to obtain the fused features of the post. The fused feature vector representation is input into a multilayer perceptron to judge the authenticity of the post content. This invention can mine high-value clues from multiple dimensions such as post presentation patterns and entity semantics, improving model detection performance and alleviating the semantic sparsity problem caused by the characteristics of short texts.

Description

Technical Field

[0001] This invention relates to the field of information content security technology, and in particular to a method and system for detecting false information using knowledge-enhanced heterogeneous graph attention. Background Technology

[0002] Due to the short-text nature of social media and the diverse forms of misinformation, it is difficult to extract effective clues for identifying the authenticity of information solely from the content of the posts themselves, making misinformation detection on social media extremely challenging. To alleviate the semantic sparsity problem caused by the short-text nature of social media, researchers have proposed detection methods based on content semantic enhancement, aiming to solve the problem of early misinformation detection in low-resource scenarios. The core idea is that key entities contained in social network posts can reveal the core theme, and introducing background knowledge of these entities can enrich the semantic content of the short text. Therefore, how to obtain the background knowledge of the post text content and integrate this knowledge into the text information becomes a crucial issue.

[0003] Based on different knowledge acquisition strategies, methods can be broadly categorized into two main directions: methods based on pre-trained language models and methods based on knowledge-driven graph convolution. Methods based on pre-trained language models enhance semantic understanding by leveraging the model's internal knowledge structure and employing techniques such as fine-tuning and cue learning to mine latent language patterns and semantic cues in posts, effectively improving the detection of false information. This method is only suitable for situations where relevant background knowledge has been sufficiently incorporated during model pre-training and fine-tuning. Knowledge-driven graph convolution detection methods model the intrinsic semantic relationships and extrinsic knowledge relationships within the text in the form of graph structures, deeply mining the inherent meanings of key information elements in posts (such as people, organizations, and specific nouns). Building upon this, multi-layer graph convolution operations are used to deeply aggregate high-order neighborhood information, enhancing the text's semantic representation capabilities and improving the model's detection performance.

[0004] While these methods have improved the performance of fake information detection to some extent, they still fall short in accurately characterizing the complex semantic features of text. Firstly, existing methods focus on modeling sequence relationships and co-occurrence relationships between words, while relatively neglecting deeper semantic dependencies, such as modal relationships, agent-patient relationships, causal relationships, and dependent relationships. These deep semantic relationships are crucial for models to accurately capture the inherent connections within information and uncover key suspicious clues in fake information. Figure 1-(b) provides an example to further illustrate this point. By performing semantic dependency analysis on the post and transforming it into an intuitive dependency graph, the potentially misleading logical connections in the misinformation can be clearly revealed: that is, "myoclonic jerks" are incorrectly associated with the cause of "decreased physical function." Secondly, existing methods typically focus on single-perspective analysis (either focusing on mining pattern features of the post or on semantic understanding of the text content), lacking the ability to capture fine-grained misinformation cues from multiple perspectives. For example... Figure 1 -(c) A post is typically constructed from a diverse set of words, including not only key entity words that reveal the core theme, such as "myoclonus" and "body function," but also auxiliary expressive words that reflect the writing style and emotional tone of the post, such as "certainly," "note," "very good," and "indicates." These different types of words provide key insights for identifying misinformation from different perspectives. Summary of the Invention

[0005] This invention aims to address the problem that existing detection methods are insufficient in terms of fine-grained semantic perception due to the semantic sparsity of short text posts. It proposes a knowledge-enhanced heterogeneous graph attention-based method and system for detecting false information. This method can mine high-value clues from multiple dimensions such as post display patterns and entity semantics, thereby improving the model's detection performance and alleviating to some extent the semantic sparsity problem caused by the characteristics of short texts and the challenges brought about by the diversity of false information manifestations.

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

[0007] A method for detecting false information in knowledge-enhanced heterogeneous graph attention includes the following steps:

[0008] Model the dependency relationships between different semantic units within a post and the association with external background knowledge by constructing a heterogeneous text graph;

[0009] 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 in different types of node neighborhoods;

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

[0011] The fused feature vector is input into a multilayer perceptron to determine the authenticity of the post content.

[0012] According to the knowledge-enhanced heterogeneous graph attention-based false information detection method of the present invention, the process of constructing a heterogeneous text graph further includes generating multiple types of nodes and initial vector representations of the nodes: firstly, forming an entity word set S for each post. te Pattern word set Stp and the set of conceptual knowledge S tk Initialization was performed using the pre-trained model BERT. in The initial feature representation of all entity type words. The initial feature representation of all pattern type words. Let the initial feature representation of all concept descriptions be given; This represents the initial node feature representation matrix consisting of all nodes in the graph. |V| = 2n + l, where 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 knowledge-enhanced heterogeneous graph attention-based false information detection method of the present invention, the process of constructing a heterogeneous text graph further includes constructing inter-node relationships based on a strategy of dual metrics of semantic dependency and co-occurrence association:

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

[0015]

[0016] In the formula, SD<u,v> =1 indicates a semantic dependency between lexical node u and lexical node v; mutual information PMI greater than 0 indicates a co-occurrence association edge between lexical node u and lexical 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 present invention's knowledge-enhanced heterogeneous graph attention-based false information detection method, the two-layer graph attention mechanism further includes type-level attention and node-level attention:

[0018] Type-level attention aims to capture the semantic contribution of different types of node neighborhoods to the target node. The formula for calculating the type-level attention weights for node v is as follows:

[0019]

[0020] In the formula, x v It is the feature representation of the current node, x vt It 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; It is a learnable parameter vector in type-level attention mechanisms;

[0021] Node-level attention aims to capture the importance of nodes in the neighborhood of different types of nodes. The formula for the node-level attention weights of node v is:

[0022]

[0023] In the formula, x v′ The feature representation of a neighbor node v′ of type t of node v, ν T This represents the learnable parameter vector in the node-level attention mechanism.

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

[0025] The representations of all nodes in a heterogeneous text graph are updated by aggregating features from their neighboring nodes of different types, which can be formally described as follows:

[0026]

[0027] In the formula, This represents the embedding representation of all nodes after the attention operation on the heterogeneous graph at layer l, where σ(·) is the activation function. Let be 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 v-th row and v′-th column represents the semantic influence of node v′ on node v. This represents the embedding representation of all nodes of type t after the (l-1)th layer graph convolution operation; The transformation matrix of type t is used when convolving the (l-1)th layer graph.

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

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

[0030] (1) Characteristic Utility Assessment

[0031] A single-feature utility evaluation task is constructed to obtain the effectiveness score of each feature; an entity semantic feature utility evaluator (MLP) based on a multilayer perceptron is then used. e (·), to obtain the effectiveness evaluation score S of entity semantic features. e ; using a pattern feature utility estimator (MLP) based on a multilayer perceptron p (·), to obtain the effective differentiation evaluation score S of the pattern features. p ;

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

[0033] By learning the weights of features from different perspectives through a multilayer perceptron, dynamic optimization and aggregation of features are achieved, as shown in the following formula:

[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 knowledge-enhanced heterogeneous graph attention-based false information detection method of the present invention, during the model training process, 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 total loss function is constructed by weighted summation of the above three loss functions.

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

[0038] The heterogeneous text graph construction module is used to model the dependency relationships between different semantic units within a post and the association with external background knowledge by constructing heterogeneous text graphs.

[0039] The dual-layer graph attention node representation module is used to design node feature representations based on the dual-layer graph attention mechanism, capturing the semantic contribution of different types of node neighborhoods to the target node and the importance of nodes in different types of node neighborhoods.

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

[0041] The prediction module is used to input the fused feature vector representation into the multilayer perceptron to determine the authenticity of the post content.

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

[0043] This invention proposes a knowledge-enhanced heterogeneous graph attention-based fake information detection method, KHGAT. This method constructs a heterogeneous text graph that integrates external knowledge to finely characterize the intrinsic semantic dependencies and extrinsic knowledge associations of different semantic units in a post. Based on the heterogeneous text graph structure, a two-layer graph attention mechanism is designed to refine the learning of post pattern features and knowledge-enhanced entity semantic features. An adaptive feature aggregation module is introduced, endowing the model with the ability to automatically select and optimize feature combinations, thereby enhancing the model's ability to perceive fine-grained semantics of post content and capture fake information cues from multiple perspectives. Experimental results show that KHGAT can effectively alleviate the challenges posed by the semantic sparsity of short texts to the task of detecting fake information based on post content. On two widely used social media datasets, Weibo and Twitter, compared with fine-tuning the large language model GLM3_6B, the model's detection accuracy is improved by an average of 3.3%, and compared with the graph structure CompareNet method, the detection accuracy is improved by an average of 4.4%. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.

[0045] Figure 1 This is an example image of a Weibo post;

[0046] Figure 2 This is a flowchart illustrating the method for detecting false information using knowledge-enhanced heterogeneous graph attention according to an embodiment of the present invention.

[0047] Figure 3 This is a diagram illustrating the heterogeneous text graph construction process according to an embodiment of the present invention;

[0048] Figure 4 This is a visual analysis of the feature validity based on heatmaps, according to an embodiment of the present invention.

[0049] Figure 5 This is an ablation experiment of KHGAT on two datasets according to an embodiment of the present invention;

[0050] Figure 6 These are the semantic dependencies with the top 10 loss ratios in the test data labeled Fake according to embodiments of the present invention.

[0051] Figure 7 This is a comparison diagram of the semantic representation similarity between entities and pattern nodes in an embodiment of the present invention. Detailed Implementation

[0052] The exemplary solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art.

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

[0054] Step S101: Model the dependency relationships between different semantic units within a post and the association with external background knowledge by constructing a heterogeneous text graph.

[0055] Figure 3 The entire process of constructing a heterogeneous text graph is presented, which is divided into two parts: generating multiple types of nodes and constructing the relationships between nodes.

[0056] (1) Generate multiple types of nodes

[0057] A post is typically constructed from a diverse set of words, with different types of words carrying different semantic functions. For example, entity words embody the key elements of the text's content description, usually including people, organizations, places, and some proper nouns. A deep understanding of these entities is crucial for depicting the deeper semantic features of the information. In addition, posts also contain pattern words that reflect their expressive style and emotional inclination, which are often reflected in abstract expressive elements.

[0058] Therefore, for each social media post to be detected, the first step is to use open-source knowledge extraction tools to extract entity-related information such as people, places, organizations, and proper nouns contained in the post, forming a corresponding entity word set S. te This solution uses the DeepKE-LLM extraction tool, an open-source tool from Zhejiang University. Then, the remaining words in the post that are not included in the stop word list are considered as a set of pattern words S describing the post's writing pattern. tp Due to the short text nature of social media, models struggle to deeply understand the intrinsic meaning of core entities within limited context. Entity linking technology is used to extract relevant conceptual descriptions from high-quality external knowledge bases, forming a conceptual knowledge set S. tk .

[0059] For the initial feature representations of different types of nodes in heterogeneous text graphs, the pre-trained model BERT is used for initialization to obtain... in The initial feature representation of all entity type words. The initial feature representation for all pattern type words. Let the initial feature representation of all concept descriptions be given; This represents the initial node feature representation matrix consisting of all nodes in the graph. |V| = 2n + l, where n represents the number of entities contained in the post and l represents the number of pattern words contained in the post.

[0060] (2) Constructing relationships between nodes

[0061] To deeply analyze and accurately depict the inherent logical connections between various lexical units in a post, this scheme adopts a dual-measure strategy of semantic dependency and co-occurrence association to establish dependency relationships between different words. Specifically, semantic dependency analysis techniques are used to reveal deep semantic dependencies between words, denoted by SD. To better utilize the potential semantic connections between words and alleviate the sparsity problem of semantic dependency relationships in text graphs, this scheme simplifies the dependency relationships between nodes, transforming directed semantic dependencies into undirected ones. That is, if there is such a dependency relationship between word node u and word node v, then SD is defined.<u,v> =SD<v,u> =1, otherwise SD<u,v> =SD<v,u> =0. Next, for lexical node pairs without direct semantic dependencies, the Pointwise Mutual Information (PMI) score is used to quantify the tightness of their association. The specific calculation formula is as follows:

[0062]

[0063]

[0064]

[0065] Where C(w) represents the number of sliding windows, C(u) represents the number of sliding windows containing word node u, C(u,v) represents the number of sliding windows containing both words u and v, p(u,v) represents the probability of words u and v co-occurring, p(u) represents the percentage of sliding windows containing word u out of the total number of sliding windows, and p(v) represents the percentage of sliding windows containing word v out of the total number of sliding windows. Co-occurrence association edges are established between word pairs with PMI greater than 0. Simultaneously, to model the external knowledge associations of posts, the extracted concept descriptions are treated as nodes linked to corresponding entities, and KE represents 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 dependencies between different words in the post can be shown in detail, while also reflecting the knowledge connections outside the post.

[0066] Let A denote the adjacency matrix of a heterogeneous text graph. For any type of node u, the adjacency relationship between its corresponding different types of nodes v is represented 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 in different types of node neighborhoods, so as to enhance the representation of post patterns and entity semantic features.

[0069] For a node v in a text graph, from a semantic influence perspective, neighboring nodes of different types may have different impacts. For example, if node v is an entity type, intuitively, its neighboring concept description nodes and entity nodes of the same type contribute more to its semantic understanding than its pattern-type neighboring nodes. Based on this, this solution 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 in different types of node neighborhoods, thereby enhancing the model's ability to distinguish subtle differences between pattern and entity semantics.

[0070] (1) Type-level attention

[0071] Type-level attention aims 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 the features of all neighboring nodes of type t. The calculation process can be formally described as follows:

[0072]

[0073] Where, x vt Let D represent the neighborhood of node v of type t, where t∈T={te,tp,tk}. Let D denote the degree matrix of the text graph, and I denote the identity matrix. This represents the standardized adjacency matrix. X represents the proximity relationship between node v and all nodes of type t in the graph. t This represents the feature representation matrix for all types t.

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

[0075]

[0076] in, It is a learnable parameter vector in type-level attention mechanisms.

[0077] (2) Node-level attention

[0078] Node-level attention aims to capture the importance of nodes within the neighborhood of different node types. For a neighbor node v′ of type t of node v, x can be represented based on the node's characteristic representation. v and x v′ Calculate node-level attention weights, which can be specifically described as:

[0079]

[0080] Where, ν T This represents the learnable parameter vector in the node-level attention mechanism.

[0081] (3) Aggregation of neighborhood nodes in heterogeneous text graphs

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

[0083]

[0084] In the formula, This represents the embedding representation of all nodes after the attention operation on the heterogeneous graph at layer l, where σ(·) is the activation function. Let be 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 v-th row and v′-th column represents the semantic influence of node v′ on node v. This represents the embedding representation of all nodes of type t after the (l-1)th layer graph convolution operation; is a transformation matrix of type t for convolution of the (l-1)th layer graph, which transforms the representations of different types of nodes into a common dense low-dimensional space.

[0085] After aggregating features of different types of high-order neighborhood nodes through a heterogeneous graph attention layer (l-layer), the final representations of all nodes in the heterogeneous text graph (including entity, pattern, and concept descriptions) are obtained. Based on this, a learnable node weight layer is designed for entity-type word nodes and pattern-type word nodes respectively, to aggregate the features of all entity word nodes and obtain knowledge-enhanced entity semantic feature representations. Pattern characteristics of posts The specific process can be formally described as follows:

[0086]

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

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

[0089] Since the effectiveness of features from different perspectives varies depending on the type of misinformation, 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 endowed with the ability to automatically select and optimize feature combinations. This specifically includes two stages:

[0090] (1) Characteristic Utility Assessment

[0091] To enable the model to quantify the utility of features from different perspectives in false information detection, a multi-task joint learning strategy is adopted. This involves constructing a single-feature utility evaluation task to obtain the effectiveness score for each feature. Specifically, features from different perspectives are input into an evaluation model composed of a multilayer perceptron.

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

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

[0094] Among them, MLP e (·) denotes an entity semantic feature utility estimator based on a multilayer perceptron, S e This represents the effectiveness evaluation score of entity semantic features. Similarly, the pattern feature utility evaluator MLP based on a multilayer perceptron is used. p (·), to obtain the effective scoring score of pattern features: S p =Sigmoid(MLP) p (h p )).

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

[0096] Based on the assumption that features that can be correctly identified are more useful, the score of a feature predicting the correct label (i.e., the feature utility score) is used as the initial weight. Building upon this, a multilayer perceptron learns the weights of features from different perspectives 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 the multilayer perceptron to judge the authenticity of the post content.

[0100] The fused feature h is fed into a multilayer perceptron to predict the authenticity of the post, as shown in the following formula:

[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 formulas are as follows:

[0103]

[0104]

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

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

[0107] l = l cls +αl e +βl p

[0108] Here, α and β are hyperparameters with values ​​between 0 and 1. The average loss for each batch is calculated before backpropagation.

[0109] Corresponding to the above method, this embodiment also proposes a fake 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 associations with external background knowledge by constructing heterogeneous text graphs.

[0111] The dual-layer graph attention node representation module is used to design node feature representations based on the dual-layer graph attention mechanism, capturing the semantic contribution of different types of node neighborhoods to the target node and the importance of nodes in different types of node neighborhoods.

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

[0113] The prediction module is used to input the fused feature vector representation into the multilayer perceptron to determine the authenticity of the post content.

[0114] To verify the effectiveness and accuracy of the model in real-world social media environments, extensive experiments were conducted on two representative social media datasets—Weibo and Twitter. Both datasets contain two label categories: Real indicates the content of the message is true, and Fake indicates the message is fake. Before formally training and testing the model, both datasets underwent cleaning and preprocessing: First, the data was deduplicated based on post content and data length, removing duplicate samples and samples shorter than 5 characters; second, open-source entity extraction tools were used to extract names of people, places, organizations, events, proper nouns, etc., from the text. Specific statistical descriptions of the preprocessed datasets are shown in Table 1.

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

[0116]

[0117] (I) Experimental Setup

[0118] (1) Experimental Environment

[0119] The machine configuration and environment for all experiments were as follows: Intel(R) Xeon(R) Platinum 8474C (processor), 80GB (memory), 32G (GPU). All code was implemented using Python 3.8, PyTorch (2.0.1), and CUDA 11.8.

[0120] (2) Baseline Comparison Model

[0121] To verify the effectiveness of the proposed model, the proposed method was compared with some state-of-the-art baseline models. These methods can be categorized into baseline methods based on pre-trained language models and fake information detection methods based on graph neural networks.

[0122] ① Detection method based on pre-trained language model: By leveraging the powerful context awareness capability of the pre-trained model and the knowledge embedded in the model parameters, the language patterns and semantic details of the posts are captured, thereby improving the detection efficiency of the model.

[0123] • BERT-Emo: This method uses the pre-trained BERT model to encode the semantics of text while simultaneously characterizing the sentiment of the information publisher. Since it focuses on the content features of the information itself, a simplified version of this method was used in the comparative experiments, which does not consider the sentiment of the information commenter.

[0124] • GLM-6B-LORA: GLM-6B is an open-source dialogue language model that supports both Chinese and English question-answering, with 6.2 billion parameters. It transforms the fake information detection task into a question-answering format and uses the LORA method to fine-tune the model. The scaling factor is set to 3.12%, and 16-bit floating-point operations are used.

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

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

[0127] • KMGCN: For each input multimodal information, the model first obtains related concept descriptions using entity linking techniques, and then concatenates these descriptions with the original post as supplementary text content. Based on the co-occurrence relationships 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 heterogeneous graphs containing nodes of various types, including topics, sentences, and entities. Second, it learns topic-rich sentence semantic representations and context-based entity semantic representations through a heterogeneous graph attention network. Finally, it designs a semantic contrast network to compare the consistency between context-based entity semantic representations and external knowledge base semantic representations, thereby improving the performance of fake information detection.

[0129] • KDRD: Transforms social media text into a semantic-entity graph, where nodes not only contain words from the original text, but also expand semantic words using world knowledge and language knowledge bases, thereby enriching the content of social media text; on this basis, a 2-layer graph convolution operation is used to obtain the overall semantic representation of the post.

[0130] (3) Model parameter configuration

[0131] The parameters in the proposed model are updated using the Adam algorithm, with the learning rate initialized to 0.00025. A pre-trained BERT model (bert-base-Chinese for Chinese and bert-base-uncased for English) is used to initialize semantic representations for different types of word nodes, with an initial dimension of 768. The dimension d of both the pattern word node representation and entity word node representation based on two-layer graph attention is set to 200. The batch size is 10, and the epochs are 25. To prevent gradient vanishing and overfitting during training, a dropout mechanism is used with a ratio of 0.5. The number of layers l in the heterogeneous graph attention is 2. The overall loss hyperparameters α and β are 0.5 and 0.5, respectively.

[0132] (II) Model Detection Performance

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

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

[0135] To analyze the ability of the multidimensional latent features extracted by the model to determine the authenticity of information, five real and five fake pieces of information were randomly selected as samples from the test set. First, the semantic feature h of each sample was calculated. e Pattern features hp The pairwise similarity of the mixed feature h is used to form a similarity matrix; then, it is displayed in the form of a heatmap, 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 acquired by the model all exhibit significant inter-class boundaries. This indicates that these features all possess a certain ability to distinguish the authenticity of information. Specifically, the heatmap of entity semantic features is the most chaotic compared to the other three features. This suggests that using entity semantics alone for false information detection is unstable. Pattern features exhibit strong intra-class similarity but insufficient inter-class differentiation. This is because true and false information overlap at the pattern feature level, resulting in insufficient discriminative power. When two different types of features are adaptively weighted and mixed, the mixed semantic features exhibit significant inter-class differentiation, effectively distinguishing true and false information in the feature space.

[0137] 2) Results Comparison and Analysis: Revealing the strengths and weaknesses of the model in the task of detecting false information.

[0138] The method of this invention was compared with six representative models, and four evaluation metrics, namely precision (Pre), recall (Re), F1 score and overall accuracy (Acc), were used to measure the performance of each model on the real dataset. Detailed data are shown in Table 2.

[0139] Table 2 shows the detection results (%) of KHGAT and baseline models on two standard datasets.

[0140]

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

[0142] A comparative analysis of the detection results of each model in Table 2 reveals the following:

[0143] ① Among the three detection methods based on pre-trained language models, GLM_6B_LORA outperforms the other two methods with a higher overall accuracy (Acc). However, it performs worse than the BERT_Emo model in terms of recall (Re) and F1 score for false information detection. Particularly in the Weibo dataset, GLM_6B_LORA's false information recall is only 0.583, indicating that it tends to consider the detected information as real. In contrast, although BERT_Emo outperforms GLM_6B_LORA and LLM_Hiss Prompting in false information recall and F1 score, it performs the worst in both recall and F1 score for judging real information. This result suggests that BERT_Emo tends to judge the detected information as false, and also indirectly reflects that there may be overlap between real and false information at the sentiment pattern feature level.

[0144] ②KDRD's overall detection performance is on average about 1.65 percentage points higher than KMGN, indicating that enhancing the background knowledge of core words in posts from multiple perspectives can improve the model's semantic understanding of post content, thereby improving the model's overall detection performance. The CompareNet model outperforms both KMGN and KDRD models in overall detection performance on two different datasets. This suggests that heterogeneous text graphs, compared to co-occurring text graphs, can model the semantic relationships within posts at a finer granularity, thus enhancing the model's semantic perception capabilities.

[0145] ③ The proposed method KHGAT outperforms the GLM_6B_LORA model by an average of 3.3 percentage points on both datasets, without showing any significant bias; compared to CompareNet, its detection performance is on average 4.4 percentage points higher. This is because the method combines the advantages of graph neural networks and heterogeneous graphs. By constructing a text heterogeneous graph, it integrates the complex relationships between entities, patterns, and external knowledge resources. Based on knowledge-enhanced entity semantics, it deeply understands information from both entity semantics and representational patterns, extracting key clues for identifying false information. Simultaneously, through an adaptive feature aggregation module, the model is endowed with the ability to automatically select and optimize feature combinations based on the characteristics of the posts to be detected, thereby improving the model's detection performance.

[0146] (III) Ablation Test

[0147] To demonstrate the effectiveness of each module of the model, a series of ablation experiments were conducted. These experiments mainly consisted of three parts:

[0148] 1) w / o SD: Remove semantic dependencies in heterogeneous text graphs to explore whether semantic dependencies can enhance the model’s ability to discover suspicious clues about false information.

[0149] 2) w / o HA: Remove node type level attention, convert the two-layer attention into a single-layer attention, evaluate the ability of the two-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 adaptive weighted aggregation module and concatenate and fuse the obtained entity semantic features and pattern features to verify the ability of the adaptive weighting module to improve the model’s ability to automatically select and optimize feature combinations.

[0151] Figure 5 The ablation experiment results of the KHGAT model on two datasets are shown in the figure:

[0152] First, semantic dependencies in the text graph were removed, retaining only word co-occurrence and concept description relationships. This resulted in an average decrease of 2.3 percentage points (Acc) in the overall detection performance across the two datasets, and an average decrease of 2.85 percentage points (F_f1) in the detection performance of fake information. To further explore the reasons behind the performance decline, the loss ratios of various semantic dependencies were statistically analyzed for samples labeled "Fake" on the test set. The top 10 relationships with the highest loss ratios in both datasets were selected for demonstration, such as... Figure 6 As shown.

[0153] observe Figure 6 It can be observed that many important semantic dependencies, such as subject, cause, condition, modality, and basis, are missing from the text graph. The loss of these deep semantic dependency structures leads to a decline in model detection performance, thus confirming that these semantic relationships are crucial for enhancing the model's ability to identify potential clues to false information.

[0154] Secondly, after simplifying the two-layer attention mechanism to a single-layer graph attention (removing type-level attention), the model's overall detection accuracy (Acc) decreased by an average of approximately 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). To explore the reasons for this experimental phenomenon, the semantic features h of post entities under the effects of two-layer graph attention and single-layer graph attention were further compared. e and pattern features h p To visually demonstrate the differences in similarity, 50 sample data points were randomly selected from the test set. The similarity between the semantics of each sample entity and the features representing the pattern was calculated to verify the ability of the dual-layer attention to capture the differences in different types of features. Specifically, as shown below... Figure 7As shown in the figure. Generally, the lower the similarity between two feature representations, the more granular the model can capture their differences, and the stronger the model's refined feature learning and recognition capabilities. In the figure, the red nodes represent the entity semantic and pattern feature similarity values ​​of each sample after two-layer graph attention, and the green nodes represent the entity semantic and pattern feature similarity values ​​after only one-layer graph attention.

[0155] according to Figure 7 It can be observed that the similarity values ​​between entity semantics and pattern features after double-layer attention are mostly below 0.4; when replaced with single-layer graph attention, the similarity values ​​are mostly above 0.4. Combined with... Figure 7 This indicates that removing the dual-layer attention module causes the model to lose its ability to capture the semantic contributions of different categories of word nodes in detail, thereby weakening the model's ability to distinguish pattern features and subtle differences in entity semantics, and ultimately affecting the overall performance of false information detection.

[0156] Finally, removing the adaptive feature aggregation module reduced the overall detection accuracy (Acc) 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 demonstrates that the adaptive feature aggregation module can significantly improve the model's detection performance compared to feature concatenation. Furthermore, to demonstrate the contribution of a single loss in the multi-objective optimization paradigm proposed by the adaptive aggregation module to improving the model's automated selection and optimization of feature combinations, verification was performed by progressively removing losses that evaluate the utility of different features, as shown in Table 3.

[0157] Table 3 Analysis of the contribution of a single loss in multi-objective optimization.

[0158]

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

[0160] It can be observed that as more loss functions are removed, the model performance gradually decreases. When l is removed simultaneously... p and l e With two losses, the feature fusion module degenerates directly into the sum of the two features, resulting in the most significant performance degradation. This is due to the removal of l p or l e The model cannot adaptively capture the detection utility of different features based on the samples to be detected, thus failing to optimize feature combinations. Therefore, both losses are necessary to improve the model's ability to optimize features and automatically combine them. Figure 5It can also be observed that, whether it is splicing or addition, the detection performance of the model shows a significant decrease compared to adaptive weighted aggregation.

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

[0162] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0163] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.

[0164] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, 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 embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.

[0165] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features within the technical scope disclosed in the present invention; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting false information in knowledge-enhanced heterogeneous graph attention, characterized in that, Includes the following steps: Step 1: Model the dependency relationships between different semantic units within a post and the associations with external background knowledge by constructing a heterogeneous text graph; Step 2: 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 in different types of node neighborhoods. Specifically, this includes: After aggregating high-order neighborhood features of three different types of nodes—entities, patterns, and concept descriptions—through a layer of heterogeneous graph attention, the representations of all nodes in the heterogeneous text graph are finally obtained. For entity-type word nodes and pattern-type word nodes, a learnable node weight layer is designed to obtain knowledge-enhanced semantic feature representations of entities. Pattern characteristics of posts The specific process can be formally described as follows: Where σ(·) represents the Sigmoid activation function, This represents a learnable parameter vector shared by node words of all entity types. This represents a learnable parameter vector shared by word nodes of all pattern types. This represents the final representation of all entity type nodes in the graph after passing through an l-layer graph attention network. This represents the final representation of all pattern type nodes in the graph after passing through an l-layer graph attention network. Step 3: Employ a multi-objective optimization strategy, automatically selecting and aggregating features through adaptive feature aggregation to obtain the fused post features, specifically including: (1) Characteristic Utility Assessment A single-feature utility evaluation task is constructed to obtain the effectiveness score of each feature; an entity semantic feature utility evaluator (MLP) based on a multilayer perceptron is then used. e (·), to obtain the effectiveness evaluation score S of entity semantic features. e ; using a pattern feature utility estimator (MLP) based on a multilayer perceptron p (·), to obtain the effective differentiation evaluation score S of the pattern features. p ; (2) Adaptive aggregation of multi-view features By learning the weights of features from different perspectives through a multilayer perceptron, dynamic optimization and aggregation of features are achieved, as shown in the following formula: h=Sigmoid(MLP e '(S e ))h e +Sigmoid(MLP′ p (S p ))h p 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; Step 4: Input the fused feature vector representation into the multilayer perceptron to judge the authenticity of the post content.

2. The method for detecting false information in knowledge-enhanced heterogeneous graph attention according to claim 1, characterized in that, The process of constructing a heterogeneous text graph involves generating multiple types of nodes and their initial vector representations: first, forming an entity word set S for each post. te Pattern word set S tp and the set of conceptual knowledge S tk Initialization was performed using the pre-trained model BERT. in The initial feature representation of all entity type words. The initial feature representation for all pattern type words. Let the initial feature representation of all concept descriptions be given; This represents the initial node feature representation matrix consisting of all nodes in the graph. |V| = 2n + m, where n represents the number of entities contained in the post and m represents the number of pattern words contained in the post.

3. The method for detecting false information in knowledge-enhanced heterogeneous graph attention according to claim 2, characterized in that, The process of constructing a heterogeneous text graph also involves building relationships between nodes based on a strategy that uses both semantic dependency and co-occurrence correlation as dual measures. Let A denote the adjacency matrix of a heterogeneous text graph. For any type of node u, the adjacency relationship between its corresponding nodes v of different types is represented as follows: In the formula, SD<u,v> =1 indicates a semantic dependency between lexical node u and lexical node v; mutual information PMI greater than 0 indicates a co-occurrence association edge between lexical node u and lexical node v; KE<u,v> =1 indicates that there is an entity concept description relationship between node u and node v.

4. The method for detecting false information in knowledge-enhanced heterogeneous graph attention according to claim 1, characterized in that, Two-layer graph attention mechanisms include type-level attention and node-level attention: Type-level attention aims to capture the semantic contribution of different types of node neighborhoods to the target node. The formula for calculating the type-level attention weights for node v is as follows: In the formula, x v It is the feature representation of the current node, x vt It 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; It is a learnable parameter vector in type-level attention mechanisms; Node-level attention aims to capture the importance of nodes in the neighborhood of different types of nodes. The formula for the node-level attention weights of node v is: In the formula, x v′ The feature representation of a neighbor node v′ of type t of node v, ν T This represents the learnable parameter vector in the node-level attention mechanism.

5. The method for detecting false information in knowledge-enhanced heterogeneous graph attention according to claim 4, characterized in that, Node feature representation based on the two-layer graph attention mechanism includes: The representations of all nodes in a heterogeneous text graph are updated by aggregating features from their neighboring nodes of different types, which can be formally described as follows: In the formula, This represents the embedding representation of all nodes after the attention operation on the heterogeneous graph at layer l, where σ(·) is the activation function. Let be 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 v-th row and v′-th column represents the semantic influence of node v′ on node v. W represents the embedding representation of all nodes of type t after the (l-1)th layer graph convolution operation; t (l-1) The transformation matrix of type t is used when convolving the (l-1)th layer graph.

6. The method for detecting false information in knowledge-enhanced heterogeneous graph attention according to claim 1, characterized in that, 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, respectively; the total loss function is obtained by weighted summation of the above three cross-entropy loss functions.

7. A false information detection system for knowledge-enhanced heterogeneous graph attention, characterized in that, For implementing the method for detecting false information in knowledge-enhanced heterogeneous graph attention as described in claim 1, the system comprises: The heterogeneous text graph construction module is used to model the dependency relationships between different semantic units within a post and the association with external background knowledge by constructing heterogeneous text graphs. The dual-layer graph attention node representation module is used to design node feature representations based on the dual-layer graph attention mechanism, capturing the semantic contribution of different types of node neighborhoods to the target node and the importance of nodes in different types of node neighborhoods. The adaptive feature aggregation module is used to automatically select and aggregate features to obtain the features after post fusion by adopting a multi-objective optimization strategy. The prediction module is used to input the fused feature vector representation into the multilayer perceptron to determine the authenticity of the post content.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-6.

Images