An early fake news detection method based on hierarchical event domain semantic modeling
By using hierarchical event domain semantic modeling, and decoupling and integrating event perception and authenticity-oriented representation, the problem of insufficient generalization ability of fake news detection methods in the case of unknown events is solved, and higher detection accuracy and robustness are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-03-17
- Publication Date
- 2026-06-19
AI Technical Summary
Existing fake news detection methods lack generalization ability in the face of unknown events and cannot effectively utilize event-specific contextual information, leading to a decrease in detection accuracy.
We adopt a hierarchical event domain semantic modeling approach. We generate initial semantic embeddings through an encoder, use a generator network to decouple hierarchical semantics, and combine contrastive learning and mutual semantic injection mechanisms to dynamically and adaptively fuse event perception and authenticity-oriented representations to form perturbation features. Finally, we generate news authenticity prediction results through an adaptive fusion network.
It significantly improves the model's detection performance on unknown events and domain-specific unknown events, enhances cross-event generalization ability, structures semantic representation, and improves the model's flexibility and robustness.
Smart Images
Figure CN122240923A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of fake news detection technology, specifically involving an early fake news detection method based on hierarchical event domain semantic modeling. Background Technology
[0002] With the widespread use of social media, the rapid spread of fake news has posed a serious challenge, especially in the early stages of public emergencies where information is limited. Existing fake news detection methods are mainly divided into two categories: unimodal and multimodal. Unimodal methods typically rely on single features of text or images, such as the rhetorical style or sentiment of the text, or traces of image manipulation. Multimodal methods, on the other hand, attempt to fuse text and image information, improving detection performance by modeling cross-modal consistency or introducing attention mechanisms.
[0003] While existing methods perform well on known events, they generally suffer from insufficient generalization when dealing with unknown events. This is because different events vary significantly in narrative style, topic framing, and multimedia content. One existing strategy is to treat each event as an independent domain and apply domain generalization techniques. However, this approach oversimplifies the complex semantic landscape of real-world news. In practice, events are correlated, especially those within the same domain (such as natural disasters), often sharing discourse structures, stylistic features, or visual patterns. Completely ignoring these shared attributes limits the model's ability to generalize using effective contextual cues.
[0004] Existing techniques, such as the DEAR (Disentangled Event-Agnostic Representation) method, attempt to decouple news representation into two parts: truth-related and event-specific, retaining only the truth-related part for detection. However, this method treats all event-specific features as noise and discards them. Nevertheless, event-specific features are not always harmful noise; when the target event has semantic or structural similarity to the training event, these features can provide valuable supplementary information.
[0005] In conclusion, there is an urgent need for a new method for early detection of fake news. This method should have more sophisticated modeling techniques that can extract cross-event, transferable authenticity features and conditionally utilize event-specific contextual information to address the detection challenges of unknown events, thereby improving the accuracy of fake news detection. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention proposes an early fake news detection method based on hierarchical event domain semantic modeling. The method includes: acquiring early news data and inputting it into a trained fake news detection model for processing to obtain news authenticity detection results.
[0007] The training process for a fake news detection model includes:
[0008] S1: Obtain training data for fake news; training data for fake news includes text modal data or image modal data;
[0009] S2: An encoder is used to process the fake news training data, and an initial semantic embedding for each modality is generated through a feedforward network;
[0010] S3: A generator network is used to perform hierarchical semantic decoupling on the initial semantic embedding of each modality data to obtain the event-aware representation and authenticity-oriented representation of each modality data;
[0011] S4: Perform contrastive learning training based on event-aware representation and reality-oriented representation, and calculate the total contrastive loss;
[0012] S5: Use a mutual semantic injection mechanism to perform semantic injection on the event perception representation and the authenticity-oriented representation to obtain the perturbation event perception representation and the perturbation authenticity-oriented representation;
[0013] S6: Input the perturbation event perception representation into the event classifier for processing to obtain the event classification prediction result; input the perturbation authenticity guidance representation into the authenticity classifier for processing to obtain the initial prediction result of the news authenticity; calculate the classification loss based on the event classification prediction result and the news authenticity prediction result;
[0014] S7: Concatenate the perturbation event perception representation and perturbation authenticity guidance representation of all modalities to obtain a decoupled feature sequence; use an adaptive fusion network to process the decoupled feature sequence to obtain the final news authenticity prediction result; calculate the supervised loss;
[0015] S8: Use the sum of the total contrastive loss, classification loss, and supervision loss as the total basic training loss; perform iterative basic training based on the total basic training loss and adjust the model parameters;
[0016] S9: After several rounds of basic training, freeze the parameters of the event classifier and the realism classifier, calculate the adversarial loss, and iterate and adjust the generator network parameters again based on the adversarial loss to obtain the trained fake news detection model.
[0017] Preferably, the process of hierarchically decoupling the initial semantic embeddings of each modality data using a generator network includes:
[0018] The initial semantic embedding of each modality data is input into the event-aware semantic generator to obtain a preliminary event-aware representation;
[0019] The residual is obtained by removing the projection component of the initial semantic embedding in the direction of the initial event-aware representation.
[0020] The residual is input into the authenticity-oriented semantic generator to obtain the authenticity-oriented representation;
[0021] A linear transformation is performed on the preliminary event perception representation to obtain the final event perception representation;
[0022] Furthermore, both the event-aware semantic generator and the authenticity-oriented semantic generator are two-layer, multi-layer perceptrons.
[0023] Preferably, the formula for calculating the total contrast loss is:
[0024]
[0025]
[0026]
[0027] in, Indicates the total comparative loss. Indicates the loss of the true triplet. Indicates the loss of the event triple. Represents the Euclidean distance function. This represents the authenticity-oriented representation of the i-th sample. Indicates from and The truth-oriented representation of positive sample j selected from samples with the same true / false labels but different domain labels. Indicates from and The authenticity-oriented representation of negative sample k selected from samples with different true / false labels but the same event label. Indicates interval, This represents the event-aware representation of the i-th sample. Indicates from and The event-aware representation of positive sample j selected from samples of different events but with the same domain label. Indicates from and Event-aware representation of negative sample k selected from samples with different domain labels.
[0028] Preferably, the disturbance event perception representation and the disturbance authenticity-oriented representation are obtained as follows:
[0029]
[0030]
[0031] in, This represents the perturbation-oriented representation of the i-th sample. Let represent the perturbation event perception representation of the i-th sample. This represents the event-aware representation of the i-th sample. This represents the event-aware representation of the j-th sample. This represents the authenticity-oriented representation of the j-th sample. This indicates the AdaIN operation.
[0032] Preferably, the formula for calculating the classification loss is:
[0033]
[0034] in, Represents classification loss, Represents the cross-entropy loss function. This represents a realism classifier used for text. This represents an event classifier used for text. Used for image realism classifiers This represents an event classifier used for images. This represents the true / false label of the i-th sample. This represents the event label of the i-th sample. This represents the perturbation-guided representation of the i-th sample image. Let represent the perturbation event perception representation of the i-th sample image. This represents the perturbation-oriented representation of the i-th sample text. Let represent the perturbation event perception representation of the i-th sample text.
[0035] Preferably, the process of processing the decoupled feature sequence using an adaptive fusion network includes:
[0036] The decoupled feature sequence is input into a 2-layer Transformer encoder to obtain contextualized features;
[0037] The contextualized features are subjected to average pooling and max pooling in the sequence dimension to obtain average pooling features and max pooling features respectively;
[0038] The average pooling feature and the max pooling feature are concatenated to obtain the global representation;
[0039] The global representation is input into the classification feedforward network to obtain the final fused representation; the final fused representation is then passed through a linear classifier. and The function processes the data to obtain the final prediction result of whether the news is true or false.
[0040] Preferably, the formula for calculating adversarial loss is:
[0041]
[0042] in, Indicating resistance to loss, This indicates a reality-oriented semantic generator. This represents an event-aware semantic generator. This represents the final fusion representation. This represents the authenticity-guided representation of the i-th sample image. Let i represent the event-aware representation of the i-th sample image. This represents the authenticity-oriented representation of the i-th sample text. This represents the event-aware representation of the i-th sample text. This represents a realism classifier used for text. This represents an event classifier used for text. Used for image realism classifiers This represents an event classifier used for images.
[0043] The beneficial effects of this invention are as follows:
[0044] 1. Enhanced cross-event generalization ability: By explicitly decoupling the semantics of "authenticity orientation" and "event awareness", this invention can learn discriminative patterns that are transferable across events more purely. At the same time, when the semantics of events are similar, it can conditionally use event-specific features as a supplement, which significantly improves the detection performance of the model on completely unknown events and unknown events in the domain.
[0045] 2. Structured semantic representation learning: The proposed cross-relational contrastive learning mechanism organizes the feature space not only at the true / false level, but also at the event-domain structural level, making the learned features more structured and interpretable, which helps the model understand the information associations at different levels.
[0046] 3. Dynamic Adaptive Fusion: The contextual conditional fusion module can dynamically adjust the dependency weights on different semantic components and different modalities based on the modal composition of the input news (plain text, text and images, etc.) and the semantic similarity between events, thereby enhancing the flexibility and adaptability of the method.
[0047] 4. Robustness Enhancement: Mutual semantic injection mechanism, as a data augmentation and regularization method, simulates the semantic mixture between events, forcing the model to learn robust real features that are invariant under complex changes, thereby improving the stability of the model when facing changes in the distribution of unknown events. Attached Figure Description
[0048] Figure 1 This is a flowchart illustrating the training process of the fake news detection model in this invention. Detailed Implementation
[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0050] This invention proposes an early fake news detection method based on hierarchical event domain semantic modeling. This method aims to enable the model to simultaneously capture universally applicable authenticity features and specific event contextual features through semantic decoupling, structured contrastive learning, and dynamic fusion. It also adaptively integrates information based on the semantic proximity between events, thereby achieving robust detection of new events. Figure 1 As shown, the method includes the following:
[0051] Early news data is acquired and fed into a trained fake news detection model for processing to obtain the results of news authenticity detection.
[0052] The training process for a fake news detection model includes:
[0053] S1: Obtain training data for fake news; training data for fake news includes text modal data or image modal data.
[0054] Obtain training data for fake news, with each news sample in the data. It can include text modalities, image modalities, or both.
[0055] S2: An encoder is used to process the fake news training data, and an initial semantic embedding for each modality is generated through a feedforward network.
[0056] For text modal data, the CLIP text encoder is used to process the input text sequence into a word sequence with a maximum length of 77, thus obtaining the text feature vector.
[0057] For image modal data, the CLIP image encoder is used to process the input image and scale it to 224×224 pixels to obtain the image feature vector.
[0058] Subsequently, in order to adapt the general features to the fake news detection task, projection transformations were performed using two independent feedforward neural networks (FFN_t and FFN_v) to obtain task-specific initial semantic embeddings.
[0059] S3: A generator network is used to perform hierarchical semantic decoupling on the initial semantic embedding of each modality data to obtain the event-aware representation and authenticity-oriented representation of each modality data.
[0060] Embed the initial semantics of each modality Decoupling into two orthogonal and complementary subspace representations: Reality-oriented semantic representation and event-aware semantic representation This is achieved through a generator network containing two generators, and the specific implementation process is as follows:
[0061] Event-aware semantic generation: embedding initial semantics into each modality of data Input to event-aware semantic generator In the process, a preliminary event perception representation is obtained:
[0062]
[0063] Orthogonalization and Realistic Semantic Extraction: To ensure Orthogonal to the event information, from Remove it from The projection components in the direction are used to obtain residuals independent of the event. Specifically, orthogonal projection is calculated: .in This represents the inner product. This residual... It is believed to primarily contain information that remains unchanged across events and is related to the determination of authenticity.
[0064] residual Input to the authenticity-oriented semantic generator In this process, we obtain a representation of authenticity:
[0065]
[0066] Preliminary event perception representation Perform a linear transformation to obtain the final event-aware representation:
[0067]
[0068] in, and These are learnable parameters.
[0069] S4: Perform contrastive learning training based on event-aware representation and reality-oriented representation, and calculate the total contrastive loss.
[0070] To construct structured semantic relationships within the feature space, comparative learning is performed on the reality-oriented space and the event-aware space, specifically:
[0071] Realism-oriented spatial comparison: The goal is to bring samples with the same true / false labels closer together, and to push samples with different labels further apart, while encouraging the model to focus on realism patterns across events. For the realism features of an anchor sample... Construct triples Positive samples Negative samples are selected from samples with the same true / false labels as the anchor samples, but from different events or domains. Prioritize selecting samples from those that are in the same event as the anchor sample but have different true / false labels.
[0072] Event-aware spatial contrast: The goal is to model the hierarchical structure of events and domains. This involves defining the event features for a given anchor sample. Construct triples Positive samples Negative samples are selected from samples within the same domain as the anchor sample but from different events. Samples were selected from different fields.
[0073] Calculate the total contrast loss:
[0074]
[0075]
[0076]
[0077] in, Indicates the total comparative loss. Indicates the loss of the true triplet. Indicates the loss of the event triple. Represents the Euclidean distance function. This represents the authenticity-oriented representation of the i-th sample. Indicates from and The truth-oriented representation of positive sample j selected from samples with the same true / false labels but different domain labels. Indicates from and The authenticity-oriented representation of negative sample k selected from samples with different true / false labels but the same event label. Indicates interval, This represents the event-aware representation of the i-th sample. Indicates from and The event-aware representation of positive sample j selected from samples of different events but with the same domain label. Indicates from and Event-aware representation of negative sample k selected from samples with different domain labels.
[0078] S5: Use a mutual semantic injection mechanism to perform semantic injection on the event perception representation and the authenticity-oriented representation to obtain the perturbation event perception representation and the perturbation authenticity-oriented representation.
[0079] To enhance the model's robustness to unknown event distributions, this invention designs a mutual semantic injection mechanism to simulate the mixing of semantics between events. For the i-th sample in the batch:
[0080] Feature fusion: Randomly select another sample j from the batch. Use Adaptive Instance Normalization (AdaIN) to inject the semantic features of sample j into the features of sample i.
[0081] The authenticity features of injected event semantics are represented by perturbation authenticity-oriented representations:
[0082]
[0083] Event features infused with real semantics are thus represented as perturbation-based event perception representations:
[0084]
[0085] in, This represents the perturbation-oriented representation of the i-th sample. Let represent the perturbation event perception representation of the i-th sample. This indicates the AdaIN operation.
[0086] Both the text modality and the image modality are perturbed according to the above formula.
[0087] S6: Input the perturbation event perception representation into the event classifier for processing to obtain the event classification prediction result; input the perturbation authenticity guidance representation into the authenticity classifier for processing to obtain the initial prediction result of the news authenticity; calculate the classification loss based on the event classification prediction result and the news authenticity prediction result.
[0088] Features after mixing and The real classifiers for each input and event classifier Output the initial prediction results for the news's authenticity and the event classification prediction results, respectively, and calculate the classification loss:
[0089] in, Represents classification loss, Represents the cross-entropy loss function. This represents a realism classifier used for text. This represents an event classifier used for text. Used for image realism classifiers This represents an event classifier used for images. This represents the true / false label of the i-th sample. This represents the event label of the i-th sample. This represents the perturbation-guided representation of the i-th sample image. Let represent the perturbation event perception representation of the i-th sample image. This represents the perturbation-oriented representation of the i-th sample text. Let represent the perturbation event perception representation of the i-th sample text.
[0090] If only text modalities exist, the classification loss will only retain the text modal-related function terms.
[0091] S7: Concatenate the perturbation event perception representation and perturbation authenticity guidance representation of all modalities to obtain a decoupled feature sequence; use an adaptive fusion network to process the decoupled feature sequence to obtain the final news authenticity prediction result; calculate the supervision loss.
[0092] Feature concatenation and contextualization: For a sample, collect all available decoupling features and concatenate them to form a decoupling feature sequence. Where n is the number of features (n=2 for single-modality, n=4 for multi-modality). An adaptive fusion network is used to process the decoupled feature sequences, specifically:
[0093] Input H into a lightweight Transformer encoder (e.g., 2 layers). The encoder captures the interactions and dependencies between these features through a self-attention mechanism, and outputs contextualized features U.
[0094] Contextualized features are subjected to average pooling and max pooling along the sequence dimension to obtain average pooled features. and max pooling features This pooling method can simultaneously preserve the overall information of the sequence and its most salient features.
[0095] By concatenating the average pooling features and the max pooling features, we obtain the global representation:
[0096] The global representation is input into the classification feedforward network. The final fusion representation is obtained. The final fused representation is then processed by a linear classifier. and The function processes the data to obtain the final prediction result of whether the news is "false" or "true," outputting the probability distribution of whether the news is "false" or "true." .
[0097] The supervision loss for this part is the standard cross-entropy loss: .
[0098] S8: Use the sum of the total contrastive loss, classification loss, and supervision loss as the base training total loss; perform iterative training and adjust the model parameters based on the base training total loss.
[0099]
[0100] Iterative training is performed based on the total loss of the basic training, and the model parameters are continuously adjusted for a predetermined number of training rounds.
[0101] S9: After several rounds of training, freeze the parameters of the event classifier and the realism classifier, calculate the adversarial loss, and adjust the generator network parameters according to the adversarial loss to obtain the trained fake news detection model.
[0102] To further enhance the discriminative power of the decoupled features, an adversarial training phase is introduced. After several rounds of basic training, the ground truth classifier is fixed. and event classifier The parameters are updated only for the semantically decoupled generator. and The parameters are optimized to minimize the adversarial loss. :
[0103]
[0104] in, Indicating resistance to loss, This indicates a reality-oriented semantic generator. This represents an event-aware semantic generator. This represents the authenticity-guided representation of the i-th sample image. Let represent the perturbation event perception representation of the i-th sample image. This represents the perturbation-oriented representation of the i-th sample text. Let represent the perturbation event perception representation of the i-th sample text.
[0105] If only text modalities exist, the classification loss will only retain the text modal-related function terms.
[0106] This loss drives the generator and The generated features can "deceive" existing, well-performing classifiers to the greatest extent possible. and This forces the generated features to be closer to the decision boundary of the classifier, thus making them more discriminative.
[0107] By iteratively training and continuously adjusting the generator network parameters using adversarial loss, a well-trained fake news detection model is obtained.
[0108] By acquiring early news data to be detected and inputting it into a trained fake news detection model, the authenticity of the news can be determined.
[0109] Simulation verification of the present invention:
[0110] The learning rate was set to 2e-5, using the Adam optimizer, with a batch size of 32, and training for approximately 50 epochs. During evaluation, a leave-one-out-of-event or leave-one-out-of-domain strategy was employed to simulate unknown event scenarios, with accuracy and F1 score as the primary evaluation metrics. The results of comparing this invention with comparative methods are shown in Table 1.
[0111] Table 1 Performance Comparison of the Invention and the Comparative Method
[0112]
[0113] Where T represents terrorist news, N represents natural news, M represents medical news, and P represents political news; TNM--->P means that the training set is (terrorist news + natural news + medical news) and the test set is political news, and the other representations are similar.
[0114] As can be seen from the experimental results in Table 1, the performance of the present invention is significantly better than that of existing mainstream baseline methods under various cross-event generalization settings.
[0115] In summary, this invention first separates invariant misleading clues (such as fixed deception patterns) from context-dependent event semantics (such as specific event background information) in fake news through a semantic untangling module; then, it fuses the two through a dynamic integration module, capturing transferable cross-event deception features while enhancing discriminative power by utilizing contextual signals from semantically related events. This method supports both unimodal (text / image) and multimodal data input, achieving efficient detection in the early stages (before events have fully propagated). On multiple publicly available benchmark datasets, this invention significantly outperforms existing models in both unimodal and multimodal detection accuracy, effectively improving generalization ability across scenarios (such as topic, style, and modality changes), providing a new approach to structured semantic utilization for event-level fake news detection, and demonstrating promising application prospects.
[0116] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An early fake news detection method based on hierarchical event domain semantic modeling, characterized in that, include: Early news data is acquired and fed into a trained fake news detection model for processing to obtain the news authenticity detection results. The training process for a fake news detection model includes: S1: Obtain training data for fake news; training data for fake news includes text modal data or image modal data; S2: An encoder is used to process the fake news training data, and an initial semantic embedding for each modality is generated through a feedforward network; S3: A generator network is used to perform hierarchical semantic decoupling on the initial semantic embedding of each modality data to obtain the event-aware representation and authenticity-oriented representation of each modality data; S4: Perform contrastive learning training based on event-aware representation and reality-oriented representation, and calculate the total contrastive loss; S5: Use a mutual semantic injection mechanism to perform semantic injection on the event perception representation and the authenticity-oriented representation to obtain the perturbation event perception representation and the perturbation authenticity-oriented representation; S6: Input the perturbation event perception representation into the event classifier for processing to obtain the event classification prediction result; input the perturbation authenticity guidance representation into the authenticity classifier for processing to obtain the initial prediction result of the news authenticity; calculate the classification loss based on the event classification prediction result and the news authenticity prediction result; S7: Concatenate the perturbation event perception representation and perturbation authenticity guidance representation of all modalities to obtain a decoupled feature sequence; use an adaptive fusion network to process the decoupled feature sequence to obtain the final news authenticity prediction result; calculate the supervised loss; S8: Use the sum of the total contrastive loss, classification loss, and supervision loss as the total basic training loss; perform iterative basic training based on the total basic training loss and adjust the model parameters; S9: After several rounds of basic training, freeze the parameters of the event classifier and the realism classifier, calculate the adversarial loss, and iterate and adjust the generator network parameters again based on the adversarial loss to obtain the trained fake news detection model.
2. The method for early fake news detection based on hierarchical event domain semantic modeling according to claim 1, characterized in that, The process of hierarchical semantic decoupling of the initial semantic embeddings for each modality data using a generator network includes: The initial semantic embedding of each modality data is input into the event-aware semantic generator to obtain a preliminary event-aware representation; The residual is obtained by removing the projection component of the initial semantic embedding in the direction of the initial event-aware representation. The residual is input into the authenticity-oriented semantic generator to obtain the authenticity-oriented representation; A linear transformation is performed on the initial event perception representation to obtain the final event perception representation.
3. The method for early fake news detection based on hierarchical event domain semantic modeling according to claim 2, characterized in that, Both the event-aware semantic generator and the authenticity-oriented semantic generator are two-layer, multi-layer perceptrons.
4. The method for early fake news detection based on hierarchical event domain semantic modeling according to claim 1, characterized in that, The formula for calculating the total contrast loss is: ； ； ； in, Indicates the total comparative loss. Indicates the loss of the true triplet. Indicates the loss of the event triple. Represents the Euclidean distance function. This represents the authenticity-oriented representation of the i-th sample. Indicates from and The truth-oriented representation of positive sample j selected from samples with the same true / false labels but different domain labels. Indicates from and The authenticity-oriented representation of negative sample k selected from samples with different true / false labels but the same event label. Indicates interval, This represents the event-aware representation of the i-th sample. Indicates from and The event-aware representation of positive sample j selected from samples of different events but with the same domain label. Indicates from and Event-aware representation of negative sample k selected from samples with different domain labels.
5. The method for early fake news detection based on hierarchical event domain semantic modeling according to claim 1, characterized in that, The disturbance event perception representation and the disturbance authenticity guidance representation are obtained as follows: ； ； in, This represents the perturbation-oriented representation of the i-th sample. Let represent the perturbation event perception representation of the i-th sample. This represents the event-aware representation of the i-th sample. This represents the event-aware representation of the j-th sample. This represents the authenticity-oriented representation of the j-th sample. This indicates the AdaIN operation.
6. The method for early fake news detection based on hierarchical event domain semantic modeling according to claim 1, characterized in that, The formula for calculating classification loss is: ； in, Represents classification loss, Represents the cross-entropy loss function. This represents a realism classifier used for text. This represents an event classifier used for text. Used for image realism classifiers This represents an event classifier used for images. This represents the true / false label of the i-th sample. This represents the event label of the i-th sample. This represents the perturbation-guided representation of the i-th sample image. Let represent the perturbation event perception representation of the i-th sample image. This represents the perturbation-oriented representation of the i-th sample text. Let represent the perturbation event perception representation of the i-th sample text.
7. The method for early fake news detection based on hierarchical event domain semantic modeling according to claim 1, characterized in that, The process of using an adaptive fusion network to process decoupled feature sequences includes: The decoupled feature sequence is input into a 2-layer Transformer encoder to obtain contextualized features; The contextualized features are subjected to average pooling and max pooling in the sequence dimension to obtain average pooling features and max pooling features respectively; The average pooling feature and the max pooling feature are concatenated to obtain the global representation; The global representation is input into the classification feedforward network to obtain the final fused representation; the final fused representation is then passed through a linear classifier. and The function processes the data to obtain the final prediction result of whether the news is true or false.
8. The method for early fake news detection based on hierarchical event domain semantic modeling according to claim 1, characterized in that, The formula for calculating adversarial loss is: ； in, Indicating resistance to loss, This indicates a reality-oriented semantic generator. This represents an event-aware semantic generator. This represents the final fusion representation. This represents the authenticity-guided representation of the i-th sample image. Let i represent the event-aware representation of the i-th sample image. This represents the authenticity-oriented representation of the i-th sample text. This represents the event-aware representation of the i-th sample text. This represents a realism classifier used for text. This represents an event classifier used for text. Used for image realism classifiers This represents an event classifier used for images.