False news detection method and device based on semantic intention alignment learning

Abstract

The invention provides a false news detection method based on semantic intention alignment learning, which comprises the following steps: splitting news into sentences and entities, and constructing a semantic graph and an intention graph; updating each node in the semantic graph and the intention graph through weighted information between the nodes and neighbor nodes to obtain a semantic intermediate update graph and an intention intermediate update graph; respectively embedding original root nodes in the semantic intermediate update graph and the intention intermediate update graph and combining the original root nodes with weighted global information of all respective graph nodes to obtain respective super root nodes of the semantic intermediate update graph and the intention intermediate update graph; node embedding of the global context updating semantic intermediate updating graph and the intention intermediate updating graph is used through the updated super root nodes, and a new semantic graph and a new intention graph are obtained; aligning the new semantic graph with the new intention graph by constructing pseudo nodes and bidirectional pseudo edges to obtain an aligned graph; and aggregating the embedding representation of the pseudo nodes in the alignment graph to obtain final classification features, and performing true and false news classification based on the final classification features.

Description

Technical Field

[0001] The present invention relates to the technical fields of computer technology, natural language processing (NLP) technology, graph neural network (GNN) technology, and fake news detection technology, and particularly relates to a fake news detection method, device, electronic device, computer-readable storage medium, and computer program product based on semantic intention alignment learning. Background Art

[0002] With the wide application of social media and news platforms, fake news has become a serious problem in society. Current fake news detection mainly relies on semantic features of news content, such as sentiment analysis, tone analysis, and writing style. Existing technologies, such as fake news detection methods based on sentiment and style, identify the falsity of news by analyzing semantic clues in the text. These methods often ignore the underlying intentions behind the news, resulting in insufficient detection capabilities for adversarial writing and forged content. In addition, some technologies attempt to introduce social context information into the detection process, such as through user comments or dissemination paths, but these methods still face challenges of information loss, especially in the early stages of fake news dissemination.

[0003] The disadvantages of the existing technologies are as follows: (1) Over-reliance on surface semantic features, being easily affected by malicious rewriting or context changes. (2) Failure to fully consider the motivation and intention of news creation, which makes the detection results of fake news vulnerable to interference from local semantic patterns. (3) The effectiveness of social context information in practical applications is relatively low, especially in the early dissemination stage, where the detection accuracy decreases in the case of scarce auxiliary information. Summary of the Invention

[0004] The purpose of the present invention is to solve the defect in the existing technology that fake news detection relies on surface semantic features, and proposes a fake news detection method based on semantic intention alignment learning, which improves the accuracy and robustness of fake news detection by combining semantic and intention signals.

[0005] Aiming at the deficiencies of the existing technology, as Figure 2 shown, the present invention proposes a fake news detection method based on semantic intention alignment learning, which includes:

[0006] Initial step, obtaining news articles with labeled true or false category labels, splitting them into sentences and entities, and constructing a semantic graph and an intention graph; the nodes in the semantic graph represent sentences or entities, and the edges between the nodes in the semantic graph represent the relationships between entities and sentences or the relationships between sentences; the nodes in the intention graph represent news intentions, and the edges between the nodes represent the relationships between news intentions at each level;

[0007] Update step: For each node in the semantic graph and the intention graph, update it through the weighted information with neighbor nodes to obtain the intermediate updated semantic graph and the intermediate updated intention graph. Combine the original root node embedding in the intermediate updated semantic graph and the intermediate updated intention graph with the weighted global information of all graph nodes respectively to obtain the super root nodes of the intermediate updated semantic graph and the intermediate updated intention graph respectively. Use the updated super root nodes to update the node embeddings of the intermediate updated semantic graph and the intermediate updated intention graph with the global context to obtain the new semantic graph and the new intention graph.

[0008] Alignment step: Align the new semantic graph and the new intention graph by constructing pseudo nodes and bidirectional pseudo edges to obtain the alignment graph. The semantic nodes in the alignment graph are connected to the intention nodes through pseudo nodes.

[0009] Training step: Aggregate the embedding representations of the pseudo nodes in the alignment graph to obtain the final classification features. The multi-layer perceptron classifies true and false news based on the final classification features, constructs a loss function according to the classification results and the true and false category labels, trains the multi-layer perceptron to obtain the news detection model. Input the final classification features of the news to be detected into the news detection model to obtain the detection result of whether the news to be detected is false.

[0010] The above-mentioned false news detection method based on semantic-intention alignment learning, wherein the generation process of the intention graph in the initial step includes:

[0011]

[0012] where f j is the feature of the intention node in the intention graph, initially initialized as a learnable embedding, c i is the news intention feature extracted by the generative language model, and H sen is the sentence representation in the news article.

[0013] The above-mentioned false news detection method based on semantic-intention alignment learning, wherein the update step includes:

[0014] The representation h v of each node v in the semantic graph and the intention graph is updated through weighted information transmission with neighbor node u, and the update formula is as follows:

[0015]

[0016] where, represents the representation of node v after the l-th layer update, w uv represents the weight of edge (u, v), and N(v) represents the set of neighbor nodes of node v;

[0017] Update the representation v of the hyper root node by combining the original root node embedding with the weighted global information from all graph nodes root :

[0018] v root = v root + softmax(W·H + b)·H

[0019] where W is the weight matrix, H is the representation of all nodes in the graph, and b is the bias term.

[0020] The fake news detection method based on semantic intention alignment learning, wherein the alignment step includes:

[0021] Perform bidirectional message passing and aggregation on the new semantic graph and the new intention graph using the pseudo nodes, and the pseudo node set V pd is initialized as a learnable embedding H pd ∈ R p×d , where p is the number of pseudo nodes and d is the feature dimension. The newly added pseudo edges E com connect the nodes of the new semantic graph and the new intention graph to the pseudo nodes;

[0022] For nodes u, v in the common space V com formed by the pseudo nodes, the nodes in the new semantic graph, and the nodes in the new intention graph, the edge (u, v) from u to v has an edge attribute e uv , and combine the edge type and node type context to guide the message passing process:

[0023] e uv = MLP(Concat(OneHot(u), OneHot(v), OneHot((u, v))))

[0024] where OneHot(·) performs one-hot encoding on the source node type u, the target node type v, and the edge (u, v) between them;

[0025] Use the attention mechanism to assign weights α uv to the edges of the pseudo nodes, and dynamically update the node embedding h v ∈ h pd , and the message passing formula in the graph alignment process is as follows:

[0026]

[0027] where, represents the representation of node v updated by the l-th layer of the multi-layer perceptron, and α uvis the edge weight of edge (u, v) calculated by the attention mechanism. MLP is a multi-layer perceptron for feature update. Concat(·) represents the feature concatenation operation.

[0028] As Figure 3 shown, the present invention also proposes a fake news detection device based on semantic intention alignment learning, which includes:

[0029] An initial module that obtains news articles with labeled true / false category labels, splits them into sentences and entities, and constructs a semantic graph and an intention graph. Nodes in the semantic graph represent sentences or entities, and edges between nodes in the semantic graph represent the relationship between entities and sentences or the relationship between sentences. Nodes in the intention graph represent news intentions, and edges between nodes represent the relationship between news intentions at each level.

[0030] An update module that updates each node in the semantic graph and the intention graph respectively by weighted information with neighbor nodes to obtain a semantic intermediate update graph and an intention intermediate update graph. Combine the original root node embedding in the semantic intermediate update graph and the intention intermediate update graph with the weighted global information of all graph nodes respectively to obtain the super root nodes of the semantic intermediate update graph and the intention intermediate update graph respectively. Use the updated super root nodes to update the node embeddings of the semantic intermediate update graph and the intention intermediate update graph with global context to obtain a new semantic graph and a new intention graph.

[0031] An alignment module that aligns the new semantic graph and the new intention graph by constructing pseudo nodes and bidirectional pseudo edges to obtain an alignment graph. Semantic nodes in the alignment graph are connected to intention nodes through pseudo nodes.

[0032] A training module that aggregates the embedding representations of pseudo nodes in the alignment graph to obtain final classification features. A multi-layer perceptron classifies true / false news based on the final classification features, constructs a loss function according to the classification result and the true / false category label, trains the multi-layer perceptron to obtain a news detection model. Input the final classification features of the news to be detected into the news detection model to obtain the detection result of whether the news to be detected is fake.

[0033] In the above-mentioned fake news detection device based on semantic intention alignment learning, the generation process of the intention graph in the initial module includes:

[0034]

[0035] where f j is the feature of the intention node in the intention graph, initially initialized as a learnable embedding, c i is the news intention feature extracted by using a generative language model, and H sen is the sentence representation in the news article.

[0036] The described fake news detection device based on semantic intention alignment learning, wherein the update module includes:

[0037] The representations h of each node v in the semantic graph and the intention graph v will be updated through weighted information passing with neighbor nodes u, and the update formula is as follows:

[0038]

[0039] where, represents the representation of node v after the l-th layer update, w uv represents the weight of edge (u, v), and N(v) represents the set of neighbor nodes of node v;

[0040] By combining the original root node embedding with the weighted global information from all graph nodes, update the representation v of the super root node root :

[0041] v root = v root + softmax(W·H + b)·h

[0042] where W is the weight matrix, H is the representation of all nodes in the graph, and b is the bias term;

[0043] The alignment module includes:

[0044] Utilize the pseudo-node to perform bidirectional message passing and aggregation on the new semantic graph and the new intention graph. The pseudo-node set V pd is initialized as a learnable embedding H pd ∈ R p×d , where p is the number of pseudo-nodes and d is the feature dimension. The newly added pseudo-edge E com connects the nodes of the new semantic graph and the new intention graph to the pseudo-node;

[0045] For nodes u, v in the common space V com formed by the pseudo-node, the nodes in the new semantic graph, and the nodes in the new intention graph, the edge (u, v) from u to v has an edge attribute e uv , and combine the edge type and node type context to guide the message passing process:

[0046] e uv = MLP(Concat(OneHot(u), OneHot(v), OneHot((u, v))))

[0047] where OneHot(·) performs one-hot encoding on the source node type u, the target node type v, and the edge (u, v) between them;

[0048] Assign weights α to the edges of the pseudo-nodes using the attention mechanism uv , and dynamically update the node embedding h based on the similarity between the neighborhood node embeddings v ∈H pd , and the message passing formula in the graph alignment process is as follows:

[0049]

[0050] where represents the representation of node v updated by the l-th layer of the multi-layer perceptron, and α uv is the edge weight of the edge (u, v) calculated by the attention mechanism. MLP is the multi-layer perceptron for feature update, and Concat(·) represents the feature concatenation operation.

[0051] The present invention also proposes an electronic device, which includes the above-mentioned fake news detection device based on semantic intention alignment learning. The electronic device is connected to an information display device, and the information display device is used to display the detection result with the display parameters, attributes set by the user or through an artificial intelligence model.

[0052] The present invention also proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of any of the above-mentioned fake news detection methods based on semantic intention alignment learning.

[0053] The present invention also proposes a computer program product, including a computer program. When the computer program is executed by a processor, it implements the steps of any of the above-mentioned fake news detection methods based on semantic intention alignment learning.

[0054] As can be seen from the above solutions, the advantages of the present invention are as follows:

[0055] Compared with the prior art, the present invention has the following advantages: (1) Improve accuracy: By combining semantic and intention signals, the ability to identify fake news is effectively improved, especially when facing dynamically changing writing patterns. (2) Enhance robustness: By bidirectional message passing and graph alignment, the representation gap between semantics and intention is solved, enabling the model to handle complex writing styles and news creation intentions. (3) Optimize computational efficiency: By adopting a dynamic path alignment module, redundant node connections are avoided, and the computational efficiency of the model is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flowchart of an embodiment of the present invention;

[0057] Figure 2 is a flowchart of the method of the present invention;

[0058] Figure 3 is a module diagram of the device of the present invention;

[0059] Figure 4 This is a schematic structural diagram of the first electronic device of the present invention;

[0060] Figure 5 This is a schematic structural diagram of the application environment of the first electronic device of the present invention;

[0061] Figure 6 This is a schematic structural diagram of the second electronic device of the present invention.

[0062] Reference numerals:

[0063] A - The first electronic device;

[0064] B - Fake news detection device based on semantic intention alignment learning;

[0065] C - Data acquisition device;

[0066] D - Information display device;

[0067] 1000 - The second electronic device;

[0068] Ⅰ - Computing unit;

[0069] Ⅱ - ROM;

[0070] Ⅲ - RAM;

[0071] Ⅳ - Bus;

[0072] Ⅴ - Interface;

[0073] Ⅵ - Input unit;

[0074] Ⅶ - Output unit;

[0075] Ⅷ - Storage medium;

[0076] Ⅸ - Communication unit. Detailed implementation manners

[0077] It should be noted that in this application, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0078] Without further limitations, an element qualified by the statement "comprising a..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0079] The processor described in the present invention is the control center of an electronic device, which can be a single processor or a collective term for multiple processing elements. For example, it can be one or more central processing units (CPUs), or it can be an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention. For example: one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs).

[0080] Optionally, the processor can execute various functions of the electronic device by running or executing software programs stored in the memory and by invoking data stored in the memory.

[0081] In a specific implementation, as an embodiment, the processor can include one or more CPUs. Each of these processors can be a single-core processor (single-CPU) or a multi-core processor (multi-CPU). Here, the processor can refer to one or more devices, circuits, and / or processing cores for processing data (such as computer program instructions). The electronic device can include: servers, desktop computers, laptop computers, smartphones, tablet computers, embedded computers, etc., where the embedded computer includes vehicles and robots, etc.

[0082] The memory is used to store the software program for implementing the solution of the present invention and is controlled by the processor for execution. The specific implementation manner can refer to the above method embodiments and will not be elaborated here.

[0083] It should be noted that the structure of the electronic device shown in the drawings of the present invention does not constitute a limitation thereto. The actual knowledge structure recognition device may include more or fewer components than shown in the drawings, or combine certain components, or have a different component arrangement.

[0084] The above embodiments can be implemented in whole or in part by software, hardware (such as circuits), firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present invention are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (such as infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more collections of available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, or magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0085] It should also be understood that the term "and / or" in this document is merely a description of the association relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: the individual existence of A, the simultaneous existence of A and B, and the individual existence of B. Here, A and B can be singular or plural. Additionally, the character " / " in this document generally represents an "or" relationship between the associated objects before and after, but it may also represent an "and / or" relationship, which can be specifically understood with reference to the context.

[0086] In the present invention, "at least one" means one or more, and "a plurality" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, or c can represent: a, b, c, a - b, a - c, b - c, or a - b - c, where a, b, and c can be single or multiple.

[0087] It should also be understood that in various embodiments of the present invention, the magnitudes of the sequence numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0088] In several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical or other forms.

[0089] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0090] In addition, in each embodiment of the present invention, the functional units can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit.

[0091] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks or optical discs that can store program codes.

[0092] The present invention overcomes these drawbacks and improves the robustness and accuracy of fake news detection by introducing the intention signal of news and performing alignment joint learning in combination with semantic signals. When conducting research on fake news detection, the inventors found that in the prior art, the variability of semantic clues is one of the main reasons for detection failure, especially in a news environment with diverse and rapidly changing writing patterns. After multiple rounds of experiments, the inventors proposed to jointly model semantic and intention signals through a graph neural network, thereby revealing hidden thoughts and motives in news content. Further research shows that by adopting the technical means of semantic-intention space alignment based on dynamic paths, the representational gap between semantics and intention can be effectively eliminated, making the detection of fake news more accurate and effective. To achieve the above technical effects, the present invention proposes the following key technical points:

[0093] Key point 1: Mining intention signals to enhance the ability to detect forgery motives. This method enhances the expressive ability of the fake news detection model by combining semantic and intention signals, enabling the model to not only rely on surface semantics but also understand the motives behind the news.

[0094] Key point 2: Deep semantic and intention modeling based on graph neural network. This method uses a graph neural network (GNN) to deeply model news content, effectively capturing the dependencies between news sentences and entities, and enhancing the ability to understand long-distance contexts.

[0095] Key point 3: Bidirectional alignment of semantic and intention signals. This method constructs a dynamic path by introducing pseudo nodes and pseudo edges, performs bidirectional alignment of semantic and intention graphs, realizes the precise fusion of signals, and reduces redundant calculations.

[0096] To make the above features and effects of the present invention more clearly and understandably described, specific embodiments are hereinafter given and detailed descriptions are made in conjunction with the accompanying drawings of the specification. This specification discloses one or more embodiments incorporating the features of the present invention. The disclosed embodiments are merely for illustrative purposes. The protection scope of the present invention is not limited to the disclosed embodiments, and the present invention is defined by the appended claims.

[0097] Figure 1 It is the overall flowchart of the present invention, showing the process of constructing semantic and intention graphs, bidirectional message passing, and alignment.

[0098] The technical solution of the present invention can be realized through the following steps:

[0099] Step 1: Preprocess the news content. The news source can be, for example, short - text news (Weibo news) or long - text news (webpage news). Split the news into sentences and entities, and construct a semantic graph and an intent graph. To extract semantic clues, different from previous work that regarded news narrative as a whole, we consider the local narrative structure of the news and analyze the news at the sentence level. In addition, we combine the global relationships between news sentences by using entities as references, thus achieving long - distance context interaction and facilitating joint learning with news intents. By integrating local inter - sentence relationships and global sentence - entity relationships, the semantic graph can achieve the semantic representation of the whole news, capture the narrative flow, and ensure the logical and temporal progression of the news article. To comprehensively extract intent signals, this method adopts a coarse - to - fine strategy. It uses a generative language model to extract coarse - grained news intents, capturing four high - level intent components including belief, desire, plan, and result. For example, asking an open - source or closed - source large language model "What are the belief, desire, plan, and expected result of this news respectively", and the 4 answer texts obtained will be input into a text encoder as the representation of the coarse - grained intent at different levels. At the same time, considering the fine - grained correlation between intent and narrative features, the constructed intent graph contains fine - grained intent nodes, and the edges between nodes represent the relationships between intents. The relationship between coarse - grained intents is a directed edge defined by the "news intent theory" of existing research, and the edge between a coarse - grained intent and its corresponding fine - grained intent is a fully - connected bidirectional edge, representing a subordinate relationship.

[0100] Model the intent realization by relying on the correlation between the coarse - grained intent and news narrative features.

[0101]

[0102] where f j is a fine - grained intent node, initially initialized as a learnable embedding, c i is the representation of the coarse - grained intent node, c i is the feature value output by the text encoder of the above - mentioned generative language model, H sen is the representation of the sentence in the news. Adaptive optimization from specific news details is achieved through training. This coarse - to - fine strategy enables the intent graph to combine high - level intent structures with fine - grained details, providing a comprehensive and flexible intent representation.

[0103] Step 2: Perform dual feature updates on semantic and intent signals through a graph neural network to extract the feature representation of the news. Specifically, this Step 2 is not an essential technical feature. Even if the feature updates on semantic and intent signals are not performed, news true / false detection can still be completed, but the recognition accuracy will decrease. From the experimental data, there is already a good news true / false detection effect without performing Step 2. However, the news true / false detection effect is better when Step 2 is adopted and trained for several more rounds. Generally, overfitting will start after about 10 rounds.

[0104] For the semantic graph and the intent graph, we perform dual graph updates, considering two types of message passing between graph nodes, including: (i) local message passing, which captures the dependencies between nodes according to learnable edge weights and aggregates information from neighboring nodes; (ii) global message passing, which updates node embeddings using the global context through a hyper root node to ensure that the global perspective affects all nodes. In this method, the hyper root node is represented as a learnable embedding, which is randomly initialized and optimized during training. This dynamic representation enables the hyper root node to adapt to the specific information and structural properties of the graph. For the node embeddings of the semantic graph or the intent graph, its contribution to the global information can be calculated through a linear transformation and the softmax function.

[0105] Specifically, for each node v, its representation h v is updated through weighted information passing with its neighbor node u. The update formula is as follows:

[0106]

[0107] where, represents the representation of node v after the l-th layer update of the graph neural network, w uv represents the weight of edge (u, v), N(v) represents the set of neighbor nodes of node v, and W1 and W2 are learnable weight parameters.

[0108] The representation of the hyper root node is updated by combining the original root node embedding with the weighted global information from all graph nodes.

[0109] The update formula for the super root node is as follows:

[0110] v root = v root + softmax(W·H + b)·H

[0111] where, v root is the representation of the super root node. The initial super root node is randomly initialized. W is a learnable weight matrix, H is the representation of all nodes in the graph, and b is a learnable bias term.

[0112] By leveraging the hyper root node and its dynamic interaction with graph nodes, this global message passing mechanism enables nodes representing sentences, entities, or intents to be enriched contextually not only by obtaining information from their direct neighbors but also through global embeddings, allowing semantic and intent representations to capture detailed and holistic narratives.

[0113] Step 3: Employ a dynamic path alignment module for bidirectional message passing to align semantic and intent signals; After obtaining the graph representations of semantics and intents, an intuitive joint learning method is to directly connect the nodes in the intent graph and the semantic graph. However, this simple operation ignores the gap between intent and semantics and may lead to quadratic redundant computational costs due to the increased number of nodes. To address this, this method proposes a graph alignment module based on dynamic paths, which uses pseudo-nodes representing the transformation weights (aggregation / contribution weights) between semantic nodes and intent nodes to perform bidirectional message passing and aggregation in a common space. The common space is the connection between the semantic graph and the intent graph, and the number of common space nodes = the number of semantic graph nodes + the number of intent graph nodes + the number of pseudo-nodes. For pseudo-node i, all nodes in the intent graph will be connected to pseudo-node i, and all nodes in the semantic graph will also be connected to i. The final edge connection weights will be adaptively assigned during the training process.

[0114] We decompose the relationship between semantic graph and intent graph nodes by introducing pseudo-nodes and pseudo-edges. Specifically, the pseudo-nodes act as conceptual bridges, promoting effective interaction between the semantic graph and the intent graph by constructing bidirectional (i.e., semantic-pseudo-node-intent and intent-pseudo-node-semantic) pseudo-edges. Thus, the pseudo-nodes and pseudo-edges can serve as dynamic paths, enabling this method to achieve alignment between the semantic graph and the intent graph while eliminating excessive and redundant node connections. Specifically, the pseudo-nodes are initialized as learnable embeddings that fully connect the nodes of the semantic graph and the intent graph to the pseudo-nodes, transforming the nodes in the semantic graph and the intent graph into the common space: the set of pseudo-nodes V pd is initialized as a learnable embedding H pd ∈R p×d , where p is the number of pseudo-nodes and d is the feature dimension. The newly added connecting edges E com fully connect the nodes of the semantic graph and the intent graph to the pseudo-nodes.

[0115] Based on the common space, this method further designs a dynamic message passing mechanism to ensure that this method can capture the relationships between heterogeneous graph nodes, support bidirectional message passing and aggregation, and thus obtain aligned and effective news representations.

[0116] For nodes u, v in the common space V com and the connecting edge (u, v) from u to v, we first introduce its edge attribute e uv, guiding the message passing process by combining the edge type and node type context, as follows:

[0117] e uv = MLP(Concat(OneHot(u), OneHot(v), OneHot((u, v))))

[0118] where OneHot(r) performs one-hot encoding on the source node type u, target node type v, and the edge (u, v) between them. Through the edge attribute e uv , the model can distinguish different relationships between nodes and the direction of their edges, ensuring that the interaction between nodes is affected by their semantic or intended roles.

[0119] Based on the edge attribute, this method uses the attention mechanism to assign weights α to the edges of the pseudo-nodes uv , indicating the key information path and reasoning about false clues. Dynamically update the node embedding h based on the similarity between neighborhood node embeddings v ∈H pd , the message passing formula in the graph alignment process is as follows:

[0120]

[0121] where α uv is the edge weight of the edge (u, v) calculated by the attention mechanism, MLP is a multi-layer perceptron for feature update, and Concat(·) represents the feature concatenation operation. Use the attention mechanism to assign weights to the pseudo-edges, indicating the key information path and reasoning about forged clues, and dynamically update the node embeddings based on the neighborhood node embeddings. During the information interaction process, the pseudo-nodes ensure effective information exchange by acting as a mediator between the semantic graph and the intention graph. By dynamically adjusting the edge attributes, key information is enhanced, while irrelevant or redundant information is gradually discarded, thereby enhancing the model's ability to extract complementary features from semantic and intention signals and ultimately improving its ability to detect fake news.

[0122] Step 4: Classify fake news through the multi-layer perceptron MLP. This method obtains the final classification feature by aggregating the embedding representations of the pseudo-nodes, and the feature is predicted through an MLP. Finally, the model is optimized using the cross-entropy loss.

[0123] The following is a system embodiment corresponding to the above method embodiment. This embodiment can be implemented in cooperation with the above embodiment. The relevant technical details mentioned in the above embodiment are still valid in this embodiment. To avoid repetition, they are not elaborated here. Correspondingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiment.

[0124] Such as Figure 3As shown in the figure, the present invention also proposes a fake news detection device based on semantic intention alignment learning, which includes:

[0125] An initial module that obtains news articles with labeled true or false category labels, splits them into sentences and entities, and constructs a semantic graph and an intention graph; the nodes in the semantic graph represent sentences or entities, and the edges between the nodes in the semantic graph represent the relationships between entities and sentences or the relationships between sentences; the nodes in the intention graph represent news intentions, and the edges between the nodes represent the relationships between news intentions at various levels;

[0126] An update module that updates each node in the semantic graph and the intention graph respectively by weighted information with neighboring nodes to obtain a semantic intermediate update graph and an intention intermediate update graph; combines the original root node embeddings in the semantic intermediate update graph and the intention intermediate update graph with the weighted global information of all graph nodes respectively to obtain the super root nodes of the semantic intermediate update graph and the intention intermediate update graph respectively; uses the updated super root nodes to update the node embeddings of the semantic intermediate update graph and the intention intermediate update graph with global context to obtain a new semantic graph and a new intention graph;

[0127] An alignment module that aligns the new semantic graph and the new intention graph by constructing pseudo nodes and bidirectional pseudo edges to obtain an alignment graph; the semantic nodes in the alignment graph are connected to the intention nodes through pseudo nodes;

[0128] A training module that aggregates the embedding representations of the pseudo nodes in the alignment graph to obtain final classification features, and a multi-layer perceptron classifies true and false news based on the final classification features, constructs a loss function according to the classification results and the true and false category labels, trains the multi-layer perceptron to obtain a news detection model; inputs the final classification features of the news to be detected into the news detection model to obtain the detection result of whether the news to be detected is false.

[0129] In the fake news detection device based on semantic intention alignment learning, the generation process of the intention graph in the initial module includes:

[0130]

[0131] where f j is the feature of the intention node in the intention graph, which is initially initialized as a learnable embedding, c i is the news intention feature extracted by using a generative language model, and H sen is the sentence representation in the news article.

[0132] In the fake news detection device based on semantic intention alignment learning, the update module includes:

[0133] The representation h of each node v in the semantic graph and the intention graph vIt will be updated through weighted information transmission with the neighbor node u, and the update formula is as follows:

[0134]

[0135] Among them, represents the representation of node v after the l-th layer update, w uv represents the weight of the edge (u, v), and N(v) represents the set of neighbor nodes of node v;

[0136] By combining the original root node embedding with the weighted global information from all graph nodes, update the representation of the hyper root node v root :

[0137] v root = v root + softmax(W·H + b)·H

[0138] Among them, W is the weight matrix, H is the representation of all nodes in the graph, and b is the bias term;

[0139] This alignment module includes:

[0140] Use this pseudo-node to perform bidirectional message passing and aggregation of this new semantic graph and this new intent graph. The pseudo-node set V pd is initialized as a learnable embedding H pd ∈ R p×d , where p is the number of pseudo-nodes and d is the feature dimension. The newly added pseudo-edge E com connects the nodes of the new semantic graph and the new intent graph with this pseudo-node;

[0141] For nodes u, v in the common space V com formed by the pseudo-node, the nodes in this new semantic graph, and the nodes in this new intent graph, the connecting edge (u, v) from u to v has an edge attribute e uv , and combine the edge type and node type context to guide the message passing process:

[0142] e uv = MLP(Concat(OneHot(u), OneHot(v), OneHot((u, v))))

[0143] Among them, OneHot(·) performs one-hot encoding on the source node type u, the target node type v, and the edge (u, v) between them;

[0144] Use the attention mechanism to assign weights α uv to the connecting edges of the pseudo-node, and dynamically update the node embedding h v ∈ H pd , and the message passing formula in the graph alignment process is as follows:

[0145]

[0146] Among them, represents the representation of node v after the update of the l-th layer of the multi-layer perceptron, and α uv is the edge weight of the edge (u, v) calculated by the attention mechanism. MLP is a multi-layer perceptron for feature update, and Concat(·) represents the feature concatenation operation.

[0147] As Figure 4 shown, in another embodiment of the present invention, a first electronic device A is further proposed, which includes the above-mentioned fake news detection device based on semantic intention alignment learning.

[0148] As Figure 5 shown, the first electronic device A can also be connected to the data acquisition device C and the information display device D through a wired or wireless information transmission scheme. The data acquisition device C is used to collect news articles to be detected, and the information display device D is used to display the detection results analyzed by the present invention.

[0149] Among them, the information display device D can process and organize the data output by the first electronic device A based on the information display mechanism to improve the readability of the data output by the first electronic device A. The information display mechanism can be preset manually. For example, the data output by the first electronic device A is visually displayed, and it can display according to the display parameters and / or attributes set by the user. The display parameters can be, for example, the display data range, and the display attributes can be, for example, the display font, color, whether to scroll and play, etc. The key information specified by the user is presented to the user, and the user can understand this information more timely without having to access the secondary page or scroll the page, saving the user's operation. Or the information display mechanism can be an artificial intelligence AI display model, which can learn the key information of the user according to the user's previous usage habits, such as viewing duration, click times, editing times, etc., and then automatically present rich and necessary key information to the user.

[0150] The present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a readable storage medium. When the computer program is executed by a processor, the computer can execute the fake news detection method based on semantic intention alignment learning provided by the above-mentioned various methods.

[0151] In another embodiment, the present invention also provides a storage medium VIII for storing a computer program for executing the fake news detection method based on semantic intention alignment learning. It should be understood that the storage medium in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of random access memory (RAM) are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchlink dynamic random access memory (SLDRAM), and direct rambus random access memory (DR RAM).

[0152] Figure 6 FIG. shows a schematic block diagram of a second electronic device 1000 that can be used to implement embodiments of the present invention. The second electronic device 1000 is intended to represent various forms of digital computers, such as, for example, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The second electronic device 1000 may also represent various forms of mobile devices, such as, for example, personal digital processors, cellular telephones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely exemplary and are not intended to limit the implementation of the present invention described and / or claimed herein. The second electronic device 1000 may be the same as or different from the first electronic device A.

[0153] The second electronic device 1000 includes a computing unit Ⅰ, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory Ⅱ (ROM) or a computer program loaded from a storage medium Ⅷ into a random access memory (RAM) Ⅲ. In the RAM Ⅲ, various programs and data required for the operation of the device 1000 can also be stored. The computing unit Ⅰ, the ROM Ⅱ, and the RAM Ⅲ are connected to each other through a bus Ⅳ. An input / output (I / O) interface Ⅴ is also connected to the bus Ⅳ.

[0154] Multiple components in the second electronic device 1000 are connected to the I / O interface Ⅴ, including: an input unit Ⅵ, such as a keyboard, a mouse, etc.; an output unit Ⅶ, such as various types of displays, speakers, etc.; a storage medium Ⅷ, such as a magnetic disk, an optical disc, etc.; and a communication unit Ⅸ, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit Ⅸ allows the second electronic device 1000 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0155] The computing unit Ⅰ can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit Ⅰ include but are not limited to a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit Ⅰ executes the various methods and processes described above, such as method steps S1 - S4. For example, in some embodiments, the method can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as the storage medium Ⅷ. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device 1000 via the ROM Ⅱ and / or the communication unit Ⅸ. When the computer program is loaded into the RAM Ⅲ and executed by the computing unit Ⅰ, one or more steps of the method described above can be executed. Alternatively, in other embodiments, the computing unit Ⅰ can be configured to execute the method in any other appropriate manner (e.g., by means of firmware).

[0156] Although the embodiments of the present invention have been disclosed as above, they are not limited to only the applications listed in the specification and embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the field, additional modifications can be easily made. Therefore, without departing from the general concept defined by the claims and the equivalent scope, the present invention is not limited to specific details and the illustrations shown and described herein.

Claims

1. A fake news detection method based on semantic intention alignment learning, characterized in that, It includes: Initial step: Obtain news articles with labeled true / false category labels, split them into sentences and entities, and construct a semantic graph and an intention graph. In this semantic graph, nodes represent sentences or entities, and the edges between nodes represent the relationships between entities and sentences or the relationships between sentences; in this intention graph, nodes represent news intentions, and the edges between nodes represent the relationships between news intentions at various levels. Update step: For each node in the semantic graph and the intention graph respectively, update it through weighted information with neighboring nodes to obtain a semantic intermediate updated graph and an intention intermediate updated graph; respectively combine the original root node embeddings in the semantic intermediate updated graph and the intention intermediate updated graph with the weighted global information of all graph nodes in each of them to obtain the super root nodes of the semantic intermediate updated graph and the intention intermediate updated graph respectively; use the global context to update the node embeddings of the semantic intermediate updated graph and the intention intermediate updated graph through the updated super root nodes to obtain a new semantic graph and a new intention graph. Alignment step: Align the new semantic graph and the new intention graph by constructing pseudo nodes and bidirectional pseudo edges to obtain an alignment graph. In this alignment graph, semantic nodes are connected to intention nodes through pseudo nodes. Training step: Aggregate the embedding representations of pseudo nodes in the alignment graph to obtain final classification features. A multi-layer perceptron classifies true / false news based on these final classification features, constructs a loss function according to the classification results and the true / false category labels, trains the multi-layer perceptron to obtain a news detection model; input the final classification features of the news to be detected into the news detection model to obtain the detection result of whether the news to be detected is false.

2. The fake news detection method based on semantic intention alignment learning according to claim 1, characterized in that The generation process of the intention graph in the initial step includes: where f j is the intent node feature in the intent graph, initially initialized as a learnable embedding, c i is the news intent feature extracted using a generative language model, H sen is the sentence representation in the news article.

3. The fake news detection method based on semantic intention alignment learning according to claim 1, characterized in that The update step includes: The representation h of each node v in this semantic graph and this intention graph v will be updated through weighted information passing with neighbor nodes u, and the update formula is as follows: Among them, represents the representation of the updated node v at the l-th layer, and w uv represents the weight of the edge (u, v), and N(v) represents the set of neighbor nodes of node v; Update the representation \(v\) of the hyper-root node by combining the embedding of the original root node with the weighted global information from all graph nodes root : v root = v root + softmax(W·H + b)·H Where W is the weight matrix, H is the representation of all nodes in the graph, and b is the bias term.

4. The fake news detection method based on semantic intention alignment learning according to claim 3, wherein The alignment step includes: Use this pseudo-node to perform bidirectional message passing and aggregation of the new semantic graph and the new intention graph. The set of pseudo-nodes V pd is initialized as a learnable embedding H pd ∈R p×d , where p is the number of pseudo-nodes and d is the feature dimension. The newly added pseudo-edges E com connect the nodes of the new semantic graph and the new intention graph to this pseudo-node; For the common space V composed of pseudo nodes, nodes in the new semantic graph, and nodes in the new intention graph com For nodes u, v in it, the edge (u, v) from u to v has the edge attribute e uv , combining the edge type and node type context to guide the message passing process: e uv = MLP(Concat(OneHot(u), OneHot(v), OneHot((u, v)))) Where OneHot(·) performs one-hot encoding on the source node type u, the target node type v, and the edge (u, v) between them. Assign weights α to the edges connecting the pseudo-nodes using the attention mechanism uv and dynamically update the node embedding h based on the similarity between the neighborhood node embeddings v ∈ H pd During the graph alignment process, the message passing formula is as follows: Among them, represents the representation of the updated node v in the l-th layer of the multi-layer perceptron, and α uv is the edge weight of the edge (u, v) calculated through the attention mechanism. MLP is the multi-layer perceptron for feature update, and Concat(·) represents the feature concatenation operation.

5. A fake news detection device based on semantic intention alignment learning, characterized in that, It includes: Initial module: Obtain news articles with labeled true / false category labels, split them into sentences and entities, and construct a semantic graph and an intention graph. In this semantic graph, nodes represent sentences or entities, and the edges between nodes represent the relationships between entities and sentences or the relationships between sentences; in this intention graph, nodes represent news intentions, and the edges between nodes represent the relationships between news intentions at various levels. Update module: For each node in the semantic graph and the intention graph respectively, update it through weighted information with neighboring nodes to obtain a semantic intermediate updated graph and an intention intermediate updated graph; respectively combine the original root node embeddings in the semantic intermediate updated graph and the intention intermediate updated graph with the weighted global information of all graph nodes in each of them to obtain the super root nodes of the semantic intermediate updated graph and the intention intermediate updated graph respectively; use the global context to update the node embeddings of the semantic intermediate updated graph and the intention intermediate updated graph through the updated super root nodes to obtain a new semantic graph and a new intention graph. Alignment module: Align the new semantic graph and the new intention graph by constructing pseudo nodes and bidirectional pseudo edges to obtain an alignment graph. In this alignment graph, semantic nodes are connected to intention nodes through pseudo nodes. The training module aggregates the embedded representations of the pseudo nodes in the alignment graph to obtain the final classification features. The multi-layer perceptron classifies true and false news based on the final classification features, constructs a loss function according to the classification results and the true and false category labels, and trains the multi-layer perceptron to obtain a news detection model. The final classification features of the news to be detected are input into the news detection model to obtain the detection result of whether the news to be detected is false.

6. The fake news detection device based on semantic intention alignment learning according to claim 5, wherein, The generation process of the intent graph in the initial module includes: where f j is the intent node feature in the intent graph, initially initialized as a learnable embedding, c i is the news intent feature extracted using a generative language model, H sen is the sentence representation in the news article.

7. The fake news detection device based on semantic intention alignment learning according to claim 1, characterized in that The update module includes: The representation h of each node v in the semantic graph and the intention graph v is updated through weighted information passing with neighbor nodes u, and the update formula is as follows: Among them, represents the representation of node v after the l-th layer update, and w uv represents the weight of edge (u, v), and N(v) represents the set of neighbor nodes of node v; Update the representation \(v\) of the hyper root node by combining the embedding of the original root node with the weighted global information from all graph nodes root : v root = v root + softmax(W·H + b)·H where W is the weight matrix, H is the representation of all nodes in the graph, and b is the bias term; The alignment module includes: Use this pseudo-node to perform bidirectional message passing and aggregation of the new semantic graph and the new intention graph. The set of pseudo-nodes V pd is initialized as a learnable embedding H pd ∈R p×d , where p is the number of pseudo-nodes and d is the feature dimension. The newly added pseudo-edges E com connect the nodes of the new semantic graph and the new intention graph to this pseudo-node; For the common space V composed of pseudo nodes, nodes in the new semantic graph, and nodes in the new intention graph com For nodes u, v in it, the edge (u, v) from u to v has the edge attribute e uv , combined with the edge type and node type context to guide the message passing process: e uv = MLP(Concat(OneHot(u), OneHot(v), OneHot((u, v)))) where OneHot(·) performs one-hot encoding on the source node type u, the target node type v, and the edge (u, v) between them; Assign weights α to the edges connecting the pseudo-nodes using the attention mechanism uv and dynamically update the node embedding h v ∈H pd based on the similarity between the neighborhood node embeddings. The message passing formula in the graph alignment process is as follows: Among them, represents the representation of the updated node v in the l-th layer of the multi-layer perceptron, and α uv is the edge weight of the edge (u, v) calculated through the attention mechanism. MLP is the multi-layer perceptron for feature update, and Concat(·) represents the feature concatenation operation.

8. An electronic device, characterized in that, It includes a fake news detection device based on semantic intent alignment learning as described in claims 5-7. The electronic device is connected to an information display device, and the information display device is used to display the detection result with the display parameters, attributes set by the user, or through an artificial intelligence model.

9. A computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the fake news detection method based on semantic intent alignment learning as described in any one of claims 1-4.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the fake news detection method based on semantic intent alignment learning as described in any one of claims 1-4.

Images