False news detection method and system based on fact-sentiment dual uncertainty

Abstract

This invention belongs to the field of data detection and provides a method and system for detecting fake news based on dual uncertainty of fact and sentiment. The method involves acquiring the text and image of each news item from a social network; extracting features from the text and image to obtain text embeddings and image embeddings; calculating Gaussian uncertainty representations by Gaussian reweighting of the text and image embeddings respectively; filtering the Gaussian uncertainty representations based on a variational information bottleneck strategy to obtain multimodal uncertainty representations; processing the multimodal uncertainty representations to generate fact inconsistency representations; processing the multimodal uncertainty representations based on a sentiment manipulation graph convolutional network to generate sentiment inconsistency representations; fusing the fact inconsistency representations and sentiment inconsistency representations to generate fused features; and classifying based on the fused features to obtain the fake news detection classification result. This invention alleviates the limitations of cross-modal complementarity caused by multimodal data uncertainty.

Description

Technical Field

[0001] This invention belongs to the field of data detection technology, specifically relating to a method and system for detecting fake news based on both factual and sentiment uncertainties. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Fake news refers to reports that deliberately fabricate, exaggerate, or distort events and information with the aim of misleading the public. With the rapid development of multimodal technologies and online platforms, social information security has received increasing public attention, and fake news has evolved from simple unimodal representations (pure text or pure images) to increasingly complex multimodal representations. Therefore, multimodal fake news detection, due to its ability to effectively utilize the complementary relationship between image and text content to achieve more accurate detection results, has become an important research hotspot.

[0004] Currently, multimodal fake news detection methods can be categorized into modality representation methods and modality inconsistency methods. Modality representation methods extract text and image features through text and visual encoders, then integrate multimodal information through feature concatenation or interactive fusion strategies to facilitate fake news identification. In contrast, modality inconsistency methods emphasize modeling the semantic inconsistencies between text and images by explicitly capturing the contrastive differences between them. Furthermore, recent research has begun to explore the problem of semantic ambiguity from the perspective of multimodal data uncertainty, aiming to improve detection performance by reducing errors caused by ambiguity.

[0005] While the aforementioned methods have made some progress, they often struggle to capture the intrinsic characteristics of multimodal fake news. Fake news typically attracts user attention by deliberately fabricating facts that contradict common sense and further manipulates audience perception through emotion, thus exhibiting a dual nature: factual contradiction and emotional manipulation. Therefore, explicitly addressing factual contradiction and emotional manipulation during the detection process can significantly improve the accuracy and robustness of detection. However, two key challenges hinder the effective modeling of these characteristics, specifically:

[0006] (1) Data uncertainty: Multimodal fake news on social media contains heterogeneous data, in which different modalities inherently exist in unequal semantic subspaces; fake news samples often contain diverse and inconsistent noise, thereby reducing the discriminative power of visual and textual features. This difference introduces inherent randomness into the data, weakens the complementary information between modalities, and leads to the problem of multimodal data uncertainty, which severely restricts the effective complementarity of cross-modal information.

[0007] (2) Uncertainty in fact-emotion fusion: The fusion of fact semantics and emotional semantics is often hindered by modal inertia, that is, the image and text modalities exhibit a vague dominant relationship in the joint modeling process; one modality will have a disproportionate impact on the optimization process. In this case, the dominant modality will often drive the reduction of the overall loss, thereby inhibiting the gradient flow and backpropagation contribution of the other modality. This ambiguity will reduce the modality's adaptability and impair the joint optimization performance in cross-modal training. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes a method and system for detecting fake news based on both factual and sentiment uncertainties. This invention recalibrates modal contributions through multivariate Gaussian weighted averaging and utilizes a memory bank to generate multivariate Gaussian uncertainty representations, thereby helping to mitigate the uncertainty caused by the inherent randomness of multimodal data. It can dynamically select the text-image pair with the highest information content based on routing probabilities, thus effectively capturing factual discrepancies. Simultaneously, it employs sentiment manipulation graph convolution to reconstruct the edge weights between text and visual nodes, thereby explicitly extracting sentiment differences to solve the uncertainty problem of factual-sentiment fusion.

[0009] According to some embodiments, the first solution of the present invention provides a method for detecting fake news based on both factual and sentiment uncertainties, employing the following technical solution:

[0010] Fake news detection methods based on both factual and sentiment uncertainties include:

[0011] Get the text and images of every news item on social networks;

[0012] Feature extraction is performed on text and image to obtain text embeddings and image embeddings. Gaussian uncertainty representations are generated by Gaussian reweighting of the text embeddings and image embeddings respectively. Based on the variational information bottleneck strategy, the Gaussian uncertainty representations are filtered to obtain multimodal uncertainty representations.

[0013] The multi-head mechanism is used to perform adaptive routing in parallel to process multimodal uncertainty representations and generate factual inconsistency representations. Then, a sentiment manipulation graph convolutional network is used to process multimodal uncertainty representations and generate sentiment inconsistency representations. The factual inconsistency representations and sentiment inconsistency representations are fused to generate fused features. Classification is performed based on the fused features to obtain the fake news detection classification results.

[0014] Furthermore, Gaussian uncertainty representations are generated by Gaussian reweighting of both text embeddings and image embeddings, including:

[0015] Gaussian reweighting is applied to text embeddings using image embeddings to obtain reweighted text embeddings.

[0016] Gaussian reweighting is applied to the image embedding using text embedding to obtain the reweighted image embedding;

[0017] The mean and variance are calculated based on reweighted text embedding and reweighted image embedding. The image multivariate Gaussian distribution and text multivariate Gaussian distribution corresponding to each news item are calculated based on the mean and variance.

[0018] A multi-head attention mechanism is used to fuse the multivariate Gaussian distribution of images and the multivariate Gaussian distribution of text to generate a Gaussian uncertainty representation.

[0019] Furthermore, the text embedding is Gaussian reweighted using image embedding to obtain reweighted text embedding, including:

[0020]

[0021] in, Indicates the first in the text Weighted text embedding of each word This represents the activation function. Indicates the first in the text Text embedding of each word Indicates the first Image embedding of each word in the corresponding region block of the image. Indicates temperature parameter;

[0022] Then, Gaussian reweighting is applied to all words in the text embedding to obtain the reweighted text embedding. .

[0023] Furthermore, a multi-head mechanism is used to perform adaptive routing in parallel to process multimodal uncertainty representations, generating fact-inconsistent representations, including:

[0024]

[0025] in, For the first Adaptive routing representation of head self-attention mechanism , This indicates the number of heads in the multi-head self-attention mechanism. , , and They represent the representations of multimodal uncertainty, respectively. The transformed query, key, and value vectors For adaptive routing probability, The matching coefficient between each pair of images. This indicates that the use of a multi-head mechanism to perform adaptive routing in parallel generates inconsistent representations. Represents the learnable parameter weights. This represents the scaling factor.

[0026] Furthermore, based on the sentiment manipulation graph convolutional network for processing multimodal uncertainty representations, sentiment inconsistency representations are generated, including:

[0027] Constructing an emotion manipulation graph matrix based on multimodal uncertainty representation;

[0028] An emotion manipulation graph convolutional network is used to perform graph convolution operations on the emotion manipulation graph matrix to generate an emotion inconsistency representation.

[0029] Furthermore, the generation of the sentiment inconsistency representation specifically includes:

[0030]

[0031] in, Indicating inconsistency in emotions, Represents the emotion manipulation graph matrix. It is a non-linear activation function. This indicates the output hidden state of the previous GCN layer. Indicates the degree of the tree. Represents the learnable parameter weights. This represents the offset value.

[0032] According to some embodiments, a second aspect of the present invention provides a fake news detection system based on both factual and sentiment uncertainties, employing the following technical solution:

[0033] A fake news detection system based on both factual and sentiment uncertainties includes:

[0034] The news data acquisition module is configured to acquire the text and images of each news item from the social network;

[0035] The multimodal data uncertainty module is configured to extract features from text and images to obtain text embeddings and image embeddings, respectively, and then calculate Gaussian uncertainty representations by performing Gaussian reweighting on the text embeddings and image embeddings. Based on the variational information bottleneck strategy, the Gaussian uncertainty representations are filtered to obtain multimodal uncertainty representations.

[0036] The fact-sentiment fusion uncertainty module is configured to use a multi-head mechanism to perform adaptive routing in parallel to process multimodal uncertainty representations and generate fact-inconsistency representations. It then uses a sentiment manipulation graph convolutional network to process multimodal uncertainty representations and generate sentiment-inconsistency representations. The fact-inconsistency representations and sentiment-inconsistency representations are fused to generate fused features. Based on the fused features, classification is performed to obtain the fake news detection classification results.

[0037] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium.

[0038] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the fake news detection method based on fact-sentiment dual uncertainty as described in the first embodiment above.

[0039] According to some embodiments, a fourth aspect of the present invention provides a computer device.

[0040] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the fake news detection method based on fact-sentiment dual uncertainty as described in the first embodiment above.

[0041] According to some embodiments, a fifth aspect of the present invention provides a computer program product or computer program.

[0042] A computer program product or computer program includes computer instructions stored in a computer-readable storage medium, wherein a processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the fake news detection method based on fact-sentiment dual uncertainty as described in the first embodiment above.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] 1. This invention proposes a Fact-Sentiment Dual Uncertainty Framework (FSDUF) for accurate fake news detection; specifically, FSDUF comprises two main modules: Multimodal Data Uncertainty Module (MDUM) and Fact-Sentiment Fusion Uncertainty Module (FSFUM): MDUM aims to address two problems of multimodal data uncertainty. (1) The first problem stems from the inherent randomness of multimodal inputs. Fake news detection models rely on the mapping from the input space X to the label space Y, and random fluctuations in X may corrupt the predicted output of Y. To mitigate this problem, MDUM introduces a multivariate Gaussian reweighting mechanism, which recalibrates textual and visual features to ensure that each modality contributes optimally to the final decision-making process. This strategy helps to mitigate the uncertainty caused by the inherent randomness of multimodal data. (2) The second problem lies in the inherent noise in multimodal data. While a soft weighting strategy based on multivariate Gaussian distribution can reduce biases associated with hard decisions and reduce the weights of features with less information, it may inadvertently suppress some non-misleading signals. To address this issue, MDUM incorporates a Variational Information Bottleneck (VIB) strategy on top of Gaussian weighting. VIB aims to preserve task-relevant information beneficial for fake news detection while compressing redundant signals unrelated to category labels. By minimizing the KL divergence, the latent representation is regularized to a standard Gaussian prior, effectively filtering out modality combinations that do not contribute to improved classification performance. Through these two core techniques, the proposed FSDUF framework alleviates the limitations of cross-modal complementarity caused by the uncertainty of multimodal data.

[0045] 2. The Fact-Emotion Fusion Uncertainty Module (FSFUM) in this invention aims to address two main problems of fact-emotion fusion uncertainty: (1) Fake news often deliberately contradicts common-sense facts to attract user attention and uses emotions to manipulate audience perception, exhibiting both factual discrepancies and emotional manipulation. To address this issue, FSFUM develops an adaptive routing mechanism that dynamically selects the most informative text-image pairs based on routing probabilities, thereby effectively capturing factual inconsistencies. Furthermore, the sentiment manipulation graph convolutional network reconstructs the edge weights between text and visual nodes to explicitly extract emotional inconsistencies. (2) In the joint modeling of image and text modalities, ambiguous dominance relationships may arise, where one modality has a disproportionate impact on the optimization process. In this case, the dominant modality tends to drive the reduction of the overall loss, thereby suppressing the gradient flow and backpropagation contribution of the other modality. To address this issue, FSFUM introduces a prediction risk minimization strategy and a gradient correction mechanism. In each iteration, the gradient correction matrix is ​​used to adjust the training gradient so that the direction of parameter updates is orthogonal to the hyperplane formed by the previously encoded modal features. This strategy can mitigate cross-modal interference and support independent and joint optimization of text and image modalities. Through these core technologies, FSDUF effectively alleviates modal inertia caused by fact-sentiment fusion uncertainty. Attached Figure Description

[0046] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0047] Figure 1 This is the overall framework of a fake news detection method based on both factual and sentiment uncertainties in an embodiment of the present invention;

[0048] Figure 2 This is a visualization result of the multimodal data uncertainty after removing Gaussian uncertainty representation and variational information in an embodiment of the present invention;

[0049] Figure 3 This is a visualization result of the uncertainty of multimodal data using this method in an embodiment of the present invention;

[0050] Figure 4 This is a visualization of the dynamic routing mechanism on the Twitter dataset using ARM and SMGCN in an embodiment of the present invention.

[0051] Figure 5 This is a visualization result of the emotion manipulation graph neural network on the Twitter dataset using ARM and SMGCN in an embodiment of the present invention. Detailed Implementation

[0052] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0053] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0054] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0055] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0056] Example 1

[0057] This embodiment provides a method for detecting fake news based on both factual and sentiment uncertainties. This embodiment uses the application of this method to a server as an example for illustration. It is understood that this method can also be applied to terminals, and can also be applied to systems including terminals, servers, and other components, and can be implemented through interaction between the terminal and the server. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network servers, cloud communication, middleware services, domain name services, CDN security services, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein. In this embodiment, the method includes the following steps:

[0058] Get the text and images of every news item on social networks;

[0059] Feature extraction is performed on text and image to obtain text embeddings and image embeddings. Gaussian uncertainty representations are generated by Gaussian reweighting of the text embeddings and image embeddings respectively. Based on the variational information bottleneck strategy, the Gaussian uncertainty representations are filtered to obtain multimodal uncertainty representations.

[0060] The multi-head mechanism is used to perform adaptive routing in parallel to process multimodal uncertainty representations and generate factual inconsistency representations. Then, a sentiment manipulation graph convolutional network is used to process multimodal uncertainty representations and generate sentiment inconsistency representations. The factual inconsistency representations and sentiment inconsistency representations are fused to generate fused features. Classification is performed based on the fused features to obtain the fake news detection classification results.

[0061] like Figure 1 As shown, this embodiment provides a Fact-Sentiment Dual Uncertainty Framework (FSDUF). Specifically, FSDEF recalibrates modal contributions through multivariate Gaussian reweighting and resolves semantic ambiguity by generating multivariate Gaussian uncertainty representations using a memory. By further integrating variational information bottlenecks, FSDEF effectively filters out redundant information while retaining task-relevant content, thereby reducing the adverse effects of inherent uncertainty noise in multimodal data. Furthermore, FSDEF includes a fact-sentiment inconsistency fusion component, which combines an adaptive routing mechanism with a sentiment manipulation graph convolutional network to jointly capture factual and sentiment semantics, revealing the intrinsic properties of fake news. To address fusion uncertainty caused by modal inertia, this component also employs a gradient correction strategy.

[0062] Step S1: Obtain the text and images of each news item on the social network.

[0063] Step S2: Extract features from the text and image to obtain text embeddings and image embeddings. Calculate Gaussian uncertainty representations by performing Gaussian reweighting on the text embeddings and image embeddings respectively. Filter the Gaussian uncertainty representations based on the variational information bottleneck strategy to obtain multimodal uncertainty representations.

[0064] Multimodal fake news on social media contains heterogeneous data, where different modalities inherently exist in unequal semantic subspaces. This disparity introduces inherent randomness into the data, leading to uncertainty in multimodal data and severely restricting the effective complementarity of cross-modal information. Specifically, the intertwined nature of factual discrepancies and sentiment manipulation in fake news further amplifies multimodal noise in the representation learning process, making the fusion process more complex. To address this challenge, a multivariate Gaussian reweighting method is proposed in the representation stage, combined with a variational information bottleneck strategy, to generate robust multimodal representations that can effectively overcome uncertainty and noise.

[0065] Step S2.1: Perform feature extraction on the text and image to obtain text embeddings and image embeddings.

[0066] First, BERT, with its vast amount of pre-trained knowledge, is used as the text encoder to extract semantic features. Simultaneously, a pre-trained ViT is used as the image encoder to extract image features.

[0067] The extraction of text and image features can be defined as:

[0068]

[0069] in, Indicates the first Text embeddings generated from encoded text in each sample Indicates the first in the text Text embedding of a word, Indicates the first Image embeddings generated by encoding images from each sample Indicates the first Image embedding of each word in the corresponding region block of the image. Understandably, the text encoder and image encoder here can choose other conventional models to perform this step as needed.

[0070] Step S2.2: Calculate the Gaussian uncertainty representation by performing Gaussian reweighting on both the text embedding and the image embedding, specifically as follows:

[0071] Gaussian reweighting is applied to text embeddings using image embeddings to obtain reweighted text embeddings.

[0072] Gaussian reweighting is applied to the image embedding using text embedding to obtain the reweighted image embedding;

[0073] The mean and variance are calculated based on reweighted text embedding and reweighted image embedding. The image multivariate Gaussian distribution and text multivariate Gaussian distribution corresponding to each news item are calculated based on the mean and variance.

[0074] A multi-head attention mechanism is used to fuse the multivariate Gaussian distribution of images and the multivariate Gaussian distribution of text to generate a Gaussian uncertainty representation.

[0075] It should be noted that after acquiring text and image representations, directly fusing text and visual features may adversely affect the final decision-making process if one modality contains excessive noise or insufficient discriminative information. Therefore, a Gaussian reweighting mechanism is introduced on the text and image features to ensure that the contribution of each modality is appropriately calibrated, thereby enhancing the robustness of multimodal decision-making.

[0076] Gaussian reweighting is applied to text embeddings using image embeddings to obtain reweighted text embeddings, based on the image embeddings. Taking a single word as an example, the specific details are as follows:

[0077]

[0078] in, Indicates the first in the text Weighted text embedding of each word This represents the activation function. Indicates the first in the text Text embedding of a word, Indicates the first Image embedding of each word in the corresponding region block of the image. Indicates temperature parameter;

[0079] Then, Gaussian reweighting is applied to all words in the text embedding to obtain the reweighted text embedding. ;

[0080] The construction method for reweighted image embedding is similar. It utilizes text embedding to apply Gaussian reweighting to the image embedding, resulting in a reweighted image embedding. Taking the region block corresponding to each word in the image as an example, the specific details are as follows:

[0081]

[0082] in, Indicates the first The reweighted text embedding of each word in the corresponding region block of the image;

[0083] Then, Gaussian reweighting is applied to all regions in the image embedding to obtain the reweighted image embedding. .

[0084] The mean and variance are calculated based on reweighted text embedding and reweighted image embedding. Then, a multivariate Gaussian distribution is applied to the image corresponding to each news item based on the mean and variance. and textual multivariate Gaussian distribution .

[0085] Textual multivariate Gaussian distribution:

[0086] The method primarily uses reweighted text embedding, interacting with the corresponding region blocks in the reweighted image embedding word by word to obtain an enhanced text representation. Then, through memory block aggregation, the text mean and text variance are calculated.

[0087] The text is calculated using the text mean and text variance, resulting in a multivariate Gaussian distribution.

[0088] Image multivariate Gaussian distribution:

[0089] The method primarily uses reweighted image embedding, interacting with the corresponding words in the reweighted text embedding of each image region block to obtain an enhanced image representation. Then, through memory block aggregation, the image mean and image variance are calculated.

[0090] The image is calculated based on the image mean and image variance, resulting in a multivariate Gaussian distribution.

[0091] Image multivariate Gaussian distribution Taking an example, the calculation process is as follows:

[0092]

[0093] in, This represents a memory block that stores the features of a text-image pair in discrete storage units. This memory block directly calculates the cosine similarity between instances in a non-parametric manner based on the features of the text-image pair, thus obtaining the final average value. Let be a conditional probability distribution, representing a given input state. , Follows a normal distribution , The average value of the image. For image variance, The image covariance matrix, It is an identity matrix.

[0094] This embodiment utilizes a memory bank to calculate the average value. and variance Each sample (each news item) Initial uncertainty representation in the latent space Defined as having The multivariate Gaussian distributions of the variables are obtained and maintained through a memory bank.

[0095] By fusing the multi-head attention mechanism of text and image multi-Gaussian distributions, a Gaussian uncertainty representation is generated. To better quantify the uncertainty of the data.

[0096] Step S2.3: Based on the variational information bottleneck strategy, the Gaussian uncertainty representation is filtered to obtain the multimodal uncertainty representation.

[0097] Based on the generated multivariate Gaussian uncertainty representation, a Variational Information Bottleneck (VIB) strategy is introduced to retain task-relevant information while suppressing irrelevant or redundant content. This method effectively alleviates the limitations of cross-modal information complementarity caused by the uncertainty of multimodal data.

[0098] The variational information bottleneck strategy is defined as follows:

[0099]

[0100] Among them, cross-entropy Encourage the Sample Initial uncertainty representation in the latent space Retain the target label corresponding to the prediction. Useful information, ensure It can effectively predict the corresponding target label. Meanwhile, KL divergence force Approximates the prior distribution, i.e., the standard Gaussian distribution. This reduces Zhongyu The removal of irrelevant redundant information makes it a compressed and robust representation. The loss function represents the bottleneck of variational information. This represents the use of latent variables and initial uncertainty. To predict the target label of the corresponding output .

[0101] By backpropagating parameter updates using this variational information bottleneck strategy, the model retains uncertainty modeling during iterations while filtering out irrelevant information through the information bottleneck. This results in a final multimodal representation that is both uncertainty-inclusive and more robust. Furthermore, a reparameterization technique is subsequently employed to sample the multimodal representation, ensuring that gradients can be effectively backpropagated during subsequent fusion. This helps optimize distribution parameters and generates compressed, noise-resistant multimodal uncertainty representations.

[0102]

[0103] in, This represents multimodal uncertainty representation, including text uncertainty representation and image uncertainty representation. and .

[0104] Step S3: Use a multi-head mechanism to perform adaptive routing in parallel to process the multimodal uncertainty representation and generate factual inconsistency representation. Use a sentiment manipulation graph convolutional network to process the multimodal uncertainty representation and generate sentiment inconsistency representation. Factual inconsistency representation and sentiment inconsistency representation are fused to generate fused features. Classify based on fused features to obtain fake news detection classification results.

[0105] The dual semantic inconsistencies arising from factual discrepancies and sentiment manipulation are crucial for effectively identifying potential differences in multimodal fake news. Specifically, factual semantics highlights information that differs significantly from general audience perception. Therefore, detecting factual inconsistencies is essential for uncovering deceptive content. Simultaneously, sentiment semantics conveys implicit emotional cues, where visual elements intuitively represent emotions, contrasting with textual guidance to reveal sentiment inconsistencies. This contrast further reveals the intent behind potential sentiment manipulation and information dissemination. To effectively capture these intrinsic relationships, an adaptive routing mechanism and a fact-sentiment interaction-based sentiment manipulation graph convolutional network are designed to jointly model factual and sentiment inconsistencies, thereby improving detection accuracy.

[0106] Step S3.1: Use a multi-head mechanism to perform adaptive routing in parallel to process multimodal uncertainty representations and generate fact-inconsistent representations, specifically:

[0107]

[0108] in, For the first Adaptive routing representation using a head-based self-attention mechanism , This indicates the number of heads in the multi-head self-attention mechanism. , , and They represent the representations of multimodal uncertainty, respectively. The transformed query, key, and value vectors For adaptive routing probability, This is the matching coefficient between each pair of images. If the image patch is within the attention range of the text target, the coefficient is set to 1; otherwise, it is set to 0. This indicates that the use of a multi-head mechanism to perform adaptive routing in parallel generates inconsistent representations. Represents the learnable parameter weights. This represents the scaling factor.

[0109] Inconsistencies arise when the objects referenced in a textual description do not match the actual visual content, or when the textual description contrasts sharply with the visual presentation. To effectively capture these factual discrepancies, an adaptive routing mechanism is introduced. This mechanism dynamically selects the text-image pair with the highest information content based on the calculated routing probabilities. This method enables the model to learn richer feature representations from different subspaces and routing paths, thereby more effectively capturing factual discrepancies.

[0110] Step S3.2: Process multimodal uncertainty representations based on sentiment manipulation graph convolutional networks to generate sentiment inconsistency representations, specifically as follows:

[0111] Constructing an emotion manipulation graph matrix based on multimodal uncertainty representation;

[0112] Emotional inconsistency refers to the discrepancy between the emotions expressed or implied in textual content and the emotional atmosphere of visual presentation. By explicitly modeling emotional inconsistency in multimodal data, FSDUF can effectively capture the characteristics of emotional manipulation and significantly improve the detection capability of deceptive content. To this end, an emotion manipulation graph convolutional network is constructed to generate accurate representations of emotional inconsistency. The construction of the emotion manipulation graph matrix is ​​as follows:

[0113]

[0114]

[0115] in, This represents the emotion manipulation graph matrix. Represents the cosine similarity between text and images. Indicates the first Textual uncertainty representation of a single word, Indicates the first Image uncertainty representation of a region block. The sentiment score is used to combine external sentiment knowledge to optimize graph edge weights, thereby capturing sentiment manipulation features. This indicates the sentiment value retrieved from the SenticNet database. If the entity is not found in SenticNet, its value is set to 0. Because... It is a graph matrix, so Indicates the first OK, Representing the Column, that is, the first Line number Column elements, This represents the number of words in the text and the number of regions corresponding to those words in the image; both are equal in number. Each line has Words and There are [number] blocks, therefore, the maximum number of rows and columns is [number]. In other words, this is a The matrix.

[0116] An emotion manipulation graph convolutional network is used to perform graph convolution operations on the emotion manipulation graph matrix to generate an emotion inconsistency representation, specifically:

[0117]

[0118] in, Indicating inconsistency in emotions, Represents the emotion manipulation graph matrix. It is a non-linear activation function. This indicates the output hidden state of the previous GCN layer. Indicates the degree of the tree. Represents the learnable parameter weights. This represents the offset value.

[0119] Step S3.3: Fuse the factual inconsistency representation and the sentiment inconsistency representation to generate a fused feature. Classify the information based on the fused feature to obtain the fake news detection classification result.

[0120] Based on the extracted factual inconsistency and sentiment inconsistency representations, a fused feature is generated. Factual inconsistency and sentiment inconsistency representations can be effectively fused using, but are not limited to, attention-based methods.

[0121] Subsequently, the softmax activation function is used for the final fake news detection and classification:

[0122]

[0123] in, and Represents the learnable parameters. and Represents the offset value.

[0124] However, determining the dominant modality between textual and visual information remains challenging in the fusion of fact and sentiment inconsistencies. This uncertainty in fact-sentiment fusion limits the adaptability of multimodal joint optimization, negatively impacting cross-modal training and fusion performance. To address this issue, a prediction risk minimization strategy and gradient correction mechanism are introduced to effectively mitigate the interference caused by fusion uncertainty, thereby improving the performance of cross-modal training.

[0125]

[0126]

[0127] in, It is the gradient correction loss function. This represents the expectation operation. Represents cross-entropy loss, Represents the prediction function. It is the input sample. These are model parameters. It is a gradient correction factor Input sample The true label. This represents the partial derivative of the loss calculated during training with respect to the correction parameters. Representative in the sample The loss value on. It is a fusion of fact and emotion. It is a real fake news label. Indicates the optimization parameters. This represents the gradient correction matrix generated by the recursive least squares algorithm. In each iteration... At this time, a gradient correction matrix is ​​introduced. This method corrects the training gradient to ensure that the parameter update direction is orthogonal to the plane formed by the encoded features of the previous modality. This avoids mutual interference between different modalities and can achieve both independent optimization of text and image modalities, as well as joint optimization.

[0128] It is understandable that the above process is only used during the model training phase to update model parameters. In the actual application phase (when reasoning / detecting the authenticity of news), the model has fixed parameters and directly outputs the prediction results.

[0129] Datasets and Evaluation Metrics

[0130] The proposed FSDUF model was evaluated on three publicly available social media benchmark datasets: Weibo, Twitter, and Pheme. All datasets were randomly split into training, validation, and test sets in a 7:1:2 ratio.

[0131] When experimenting with the proposed FSDUF, we used three datasets to validate it, including accuracy, precision, recall, and F1 score.

[0132] Experimental setup

[0133] To ensure fair comparison, the same processing methods as previous studies were used to process the three datasets mentioned above, and 768-dimensional BERT and ViT were used to generate text and image vectors. For hyperparameter settings, the batch size was set to 32, and... The learning rate was used to train the model for 20 epochs. The multimodal weight decay was set to... It also employs an early stopping strategy, with the stop value set to 10.

[0134] Baseline model

[0135] The performance of FSDUF was evaluated by comparing it with a range of state-of-the-art fake news detection models.

[0136] Single-modal method: BERT+VGG-19.

[0137] Representation methods: EANN, MVAE, SpotFake.

[0138] Inconsistent methods: SAFE, MCAN, CAFÉ, BMR, CSFN, Event-Radar, MDPF, KAMP, MFCL.

[0139] Uncertainty methods: MM-UL, T-FND, MDF-FND.

[0140] Results and Analysis

[0141] The experimental results of the current state-of-the-art baseline model on three datasets are reported. As shown in Table 1, the proposed FSDUF model outperforms the current state-of-the-art baseline model on multiple datasets. On the Weibo dataset, FSDUF improves accuracy by 1.6% and F1 score by 4.3%. On the Pheme dataset, accuracy and F1 score improve by 2.9% and 1.3%, respectively. However, on the Twitter dataset, the model's F1 score only improves slightly by 0.3%, and accuracy does not improve significantly. The following conclusions can be drawn from the results:

[0142] First, unimodal models such as BERT and VGG-19 perform poorly on fake news detection tasks, suggesting that the complementarity of multimodal models can help improve detection accuracy. Notably, text-based BERT outperforms image-based VGG-19, indicating that the text modality plays a more dominant role than the visual modality during training.

[0143] Table 1 Performance comparison of FSDUF and baseline

[0144]

[0145] Evaluation metrics include accuracy, precision, recall, and F1 score.

[0146] Secondly, the FSDUF model proposed in this embodiment outperforms all existing baseline models that focus on cross-modal inconsistency modeling. The main reasons are as follows: Previous methods (e.g., MVAE and SpotFake) only use BERT and VGG-19 to fuse textual and visual features, which limits their ability to effectively capture cross-modal complementarity. Although models such as CAFE, BMR, and MFCL employ joint representation learning and contrastive learning paradigms to model cross-modal semantic interactions, they fail to explicitly consider the interrelated features of fake news, such as factual inconsistencies and sentiment manipulation. Event-Radar and MDPF incorporate semantic and sentiment information into fake news detection. Specifically, Event-Radar employs a multi-view feature learning mechanism, incorporating textual sentiment as an additional semantic cue into the overall feature representation. However, this method only models sentiment from the textual modality, failing to consider potential inconsistencies or manipulation relationships between sentiment and visual content. This single-modal sentiment modeling strategy ignores the interaction and alignment between image and text modalities, making it difficult to capture cross-modal sentiment conflicts or synergies. MDPF proposes a unified multimodal classification framework that integrates external knowledge and cross-modal graph convolutional networks for semantic and sentiment modeling. However, it neglects the modal laziness problem in the fusion of factual and sentiment information, thus hindering the joint optimization of fact-sentiment cross-modal learning. Furthermore, although MDPF introduces a graph attention mechanism, it may amplify inherent noise in factual and sentiment semantics during node information aggregation, thereby limiting the effectiveness of cross-modal interactive learning.

[0147] Third, compared to existing methods that incorporate uncertainty modeling, the model in this embodiment also demonstrates superior performance. Methods such as MM-ULN and MDF-FND attempt to address the inherent noise caused by multimodal semantic ambiguity through uncertainty estimation. However, these general methods often struggle to effectively handle the manipulative and implicit nature of complex fake news content, limiting their applicability in real-world scenarios. In contrast, the method in this embodiment introduces a VIB mechanism during uncertainty modeling to regulate the information flow in the latent space. This ensures that only task-relevant and information-rich features are retained, while suppressing redundant or noisy information. Furthermore, previous methods typically neglect the interrelationship between factual inconsistencies and sentiment manipulation, making it difficult to fundamentally address the challenges of fake news detection. By explicitly modeling these two aspects, the method in this embodiment exhibits better performance.

[0148] Finally, it was observed that FSDUF only improved its F1 score on the Twitter dataset, while its accuracy remained poor. The reasons are believed to be as follows: The Twitter dataset was collected through an official API, and some image resources are no longer available due to broken links or deleted content. Despite efforts to retrieve and supplement missing images, some gaps remain, resulting in incomplete visual information during training and inference. This affects the effectiveness of multimodal feature learning and fusion. Secondly, the text content on Twitter is highly informal and noisy, often containing abbreviations, slang, emojis, and other non-standard expressions. These characteristics significantly increase the difficulty of semantic modeling, especially when capturing fine-grained patterns such as factual inconsistencies and sentiment manipulation. Nevertheless, FSDUF still demonstrates overall performance superior to existing baseline models.

[0149] Considering the potential class imbalance in the datasets, the performance of the FSDUF model on the GossipCop and PolitiFact datasets was further evaluated, as shown in Table 2. To comprehensively and robustly evaluate performance, both datasets were split and evaluated, and a 5-fold cross-validation strategy was employed. The proposed FSDUF model performs excellently on both the GossipCop and PolitiFact datasets. On the GossipCop dataset, the FSDUF model improves accuracy by 1.7% and F1 score by 2.5%. On the PolitiFact dataset, accuracy and F1 score are improved by 1.0% and 1.2%, respectively. Experimental results on the GossipCop and PolitiFact datasets demonstrate that the proposed method possesses strong generalization ability and robustness.

[0150] Table 2 compares the baseline performance of FSDUF with that of the GossipCop and PolitiFact datasets.

[0151]

[0152] Ablation Research

[0153] To further evaluate the contribution of each component, ablation studies were conducted by comparing FSDUF with several variants, as shown in Table 3.

[0154] Table 3 shows the ablation study results for the three datasets.

[0155]

[0156] GUR, VIB, FSIF, and GM represent Gaussian uncertainty representation, variational information bottleneck, fact-sentiment inconsistency fusion, and gradient correction, respectively.

[0157] First, removing the MDUM component (i.e., (a)) leads to a significant performance drop across all datasets, indicating that multimodal data uncertainty modeling for fake news can effectively address the inherent randomness hindering cross-modal information complementarity. To further investigate this phenomenon, we compare the effects of removing GUR (i.e., (b)) and VIB (i.e., (c)) separately, finding that the performance drop in case (b) is slightly greater than that in case (c). The performance drop observed after removing (b) suggests that the inherent randomness of multimodal inputs can perturb the predicted output in the latent space. The multivariate Gaussian reweighting mechanism helps to offset this uncertainty by mitigating the impact of randomness introduced by multimodal data. Similarly, the performance drop after removing setting (c) indicates that inherent noise in multimodal data also affects the effectiveness of fake news detection. VIB effectively addresses this issue by preserving task-relevant information that contributes to fake news detection while compressing redundant signals irrelevant to the classification objective. These results demonstrate that the combination of GUR and VIB can effectively alleviate the limitations of cross-modal complementarity caused by multimodal data uncertainty.

[0158] Secondly, removing the FSFUM component (d) also leads to a significant performance drop. Since fake news is typically designed to distort facts and evoke strong emotional responses, removing setting (e) prevents the model from capturing the intrinsic relationship between factual discrepancies and emotional manipulation, thus reducing detection performance. The GM component (f) is specifically designed to mitigate modal laziness, preventing a single modality from dominating the joint optimization process, thereby ensuring more balanced and efficient multimodal fusion for multimodal fake news detection. Overall, the above analysis confirms that both factual and sentiment information are crucial for accurately identifying fake news, and that mitigating modal laziness requires both independent optimization and cross-modal joint optimization.

[0159] Visualization

[0160] To further validate the effectiveness of the proposed FSDUF method, a visualization is provided to demonstrate how the dual uncertainty module (multimodal data uncertainty and fact-sentiment fusion uncertainty) significantly improves the accuracy and interpretability of fake news detection.

[0161] First, in the sample "Hurricane Sandy really had a positive impact on my life; it was absolutely beautiful and breathtaking," t-SNE is used to visualize the Gaussian representation learned by the uncertainty module of multimodal data. Specifically, a multivariate Gaussian reweighting method is introduced to quantify the uncertainty of multimodal data, and a variational information bottleneck strategy is combined to reduce the impact of inherent noise. This design enhances cross-modal complementarity, resulting in a more compact and discriminative feature distribution. For example... Figure 2As shown, simultaneously removing the Gaussian uncertainty representation and variational information bottleneck leads to a more dispersed feature space and greater noise, with a significant overlap in the distributions of real and fake news. In contrast, Figure 3 This indicates that the characteristics of fake news are mainly concentrated in the upper region, while the characteristics of real news are mainly located in the lower region. Although there is still slight overlap between the two categories, it is consistent with... Figure 2 compared to, Figure 3 The overall distribution is more compact, and the discrimination ability is stronger.

[0162] Secondly, Figure 4 shows the visualization effect of the adaptive routing mechanism. It can be seen that the model can accurately focus on the area in the image that corresponds to the text "Hurricane Sandy", thereby better capturing factual semantics to distinguish factual differences. Figure 5 An example of cross-modal graph construction is shown. Positive emotions expressed by words such as "positive" and "beautiful" are correlated with image regions. and The dark tones of the storm clouds create a stark contrast, revealing a strong emotional inconsistency. This suggests that emotion manipulation graph convolutional networks can effectively capture emotional manipulation by incorporating emotion knowledge to optimize the edge weights between text and visual nodes.

[0163] Example 2

[0164] This embodiment provides a fake news detection system based on both factual and sentiment uncertainties, including:

[0165] The news data acquisition module is configured to acquire the text and images of each news item from the social network;

[0166] The multimodal data uncertainty module is configured to extract features from text and images to obtain text embeddings and image embeddings, respectively, and then calculate Gaussian uncertainty representations by performing Gaussian reweighting on the text embeddings and image embeddings. Based on the variational information bottleneck strategy, the Gaussian uncertainty representations are filtered to obtain multimodal uncertainty representations.

[0167] The fact-sentiment fusion uncertainty module is configured to use a multi-head mechanism to perform adaptive routing in parallel to process multimodal uncertainty representations and generate fact-inconsistency representations. It then uses a sentiment manipulation graph convolutional network to process multimodal uncertainty representations and generate sentiment-inconsistency representations. The fact-inconsistency representations and sentiment-inconsistency representations are fused to generate fused features. Based on the fused features, classification is performed to obtain the fake news detection classification results.

[0168] The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1 above. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0169] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0170] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0171] Example 3

[0172] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the fake news detection method based on fact-sentiment dual uncertainty as described in Embodiment 1 above.

[0173] Example 4

[0174] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the fake news detection method based on fact-sentiment dual uncertainty as described in Embodiment 1 above.

[0175] Example 5

[0176] This embodiment provides a computer program product or computer program, including computer instructions stored in a computer-readable storage medium. The processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the fake news detection method based on fact-sentiment dual uncertainty described in Embodiment 1 above.

[0177] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0178] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0179] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0180] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0181] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0182] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for fake news detection based on fact-sentiment dual uncertainty, characterized in that, The method comprises the following steps: obtaining text and images of each news in a social network; performing feature extraction on the text and the images to obtain text embedding and image embedding, calculating Gaussian uncertainty representation after Gaussian reweighting of the text embedding and the image embedding respectively, and screening the Gaussian uncertainty representation based on a variational information bottleneck strategy to obtain a multi-modal uncertainty representation; wherein the Gaussian uncertainty representation is calculated after Gaussian reweighting of the text embedding and the image embedding respectively, and comprises the following steps: performing Gaussian reweighting on the text embedding using the image embedding to obtain reweighted text embedding; performing Gaussian reweighting on the image embedding using the text embedding to obtain reweighted image embedding; calculating the mean and variance based on the reweighted text embedding and the reweighted image embedding, and calculating the image multivariate Gaussian distribution and the text multivariate Gaussian distribution corresponding to each news according to the mean and variance; fusing the image multivariate Gaussian distribution and the text multivariate Gaussian distribution using a multi-head attention mechanism to generate the Gaussian uncertainty representation; wherein the reweighted text embedding is obtained by performing Gaussian reweighting on the text embedding using the image embedding, and comprises the following steps: ； wherein, represents a re-weighted text embedding of the word in the text, represents an activation function, represents a text embedding of the word in the text, represents an image embedding of the word in the corresponding region block in the image, represents a temperature parameter; then reweight all the words in the text embedding with a Gaussian, resulting in a reweighted text embedding ; wherein the variational information bottleneck strategy is defined as follows: ； wherein, represents a prediction of a target label for the corresponding output using a latent variable- initial uncertainty representation ;​ performing adaptive routing in parallel using a multi-head mechanism to process the multi-modal uncertainty representation to generate a fact inconsistency representation, performing processing on the multi-modal uncertainty representation based on a sentiment manipulation graph convolution network to generate a sentiment inconsistency representation, fusing the fact inconsistency representation and the sentiment inconsistency representation to generate fusion features, and classifying based on the fusion features to obtain a fake news detection classification result; performing adaptive routing in parallel using a multi-head mechanism to process the multi-modal uncertainty representation to generate a fact inconsistency representation, and comprises the following steps: ； wherein, is the number of heads for the multi-head self-attention mechanism, is the adaptive routing representation of the head self-attention mechanism, , is the number of heads for the multi-head self-attention mechanism, , and represent the multi-modal uncertainty representation based on the transformed query, key and value vectors, transformed query, key and value vectors, is the adaptive routing probability, is the matching coefficient between each pair of images, represents the factually inconsistent representation generated using the multi-head mechanism in parallel with the adaptive routing representation, represents the learnable parameter weight, represents the scaling factor; Construction of the sentiment manipulation graph matrix is as follows: ； ； wherein, denotes the sentiment manipulation graph matrix, denotes the cosine similarity between text and image, denotes the text uncertainty representation of the word, denotes the image uncertainty representation of the m-th region block, denotes the sentiment score, used to optimize the graph edge weight in combination with external sentiment knowledge, denotes the sentiment value of the entity retrieved from the SenticNet database, if the entity is not found in SenticNet, its value is set to 0, is a graph matrix, denotes the text uncertainty representation of the word, denotes the m-th column, that is, the element of the i-th row and the m-th column, denotes the number of words in the text and the number of region blocks corresponding to the words in the image, both of which are , each row has words and region blocks. 2.The fact-sentiment dual uncertainty based fake news detection method of claim 1, wherein, performing processing on the multi-modal uncertainty representation based on a sentiment manipulation graph convolution network to generate a sentiment inconsistency representation, and comprises the following steps: constructing a sentiment manipulation graph matrix based on the multi-modal uncertainty representation; performing graph convolution operation on the sentiment manipulation graph matrix using a sentiment manipulation graph convolution network to generate the sentiment inconsistency representation. 3.The fact-sentiment dual uncertainty based fake news detection method of claim 2, wherein, The generated sentiment inconsistency representation is specifically as follows: ； wherein, represents an affect inconsistency representation, represents an affect manipulation graph matrix, is a nonlinear activation function, represents the output hidden state of the previous GCN layer, represents the degree of a tree, represents a learnable parameter weight, represents an offset value.

4. The fact-sentiment dual uncertainty based fake news detection system for implementing the fact-sentiment dual uncertainty based fake news detection method as claimed in claim 1, wherein, The method comprises the following steps: a news data acquisition module configured to obtain text and images of each news in a social network; a multi-modal data uncertainty module configured to perform feature extraction on the text and the images to obtain text embedding and image embedding, calculate Gaussian uncertainty representation after Gaussian reweighting of the text embedding and the image embedding respectively, screen the Gaussian uncertainty representation based on a variational information bottleneck strategy, and obtain a multi-modal uncertainty representation; a fact-sentiment fusion uncertainty module configured to perform adaptive routing in parallel using a multi-head mechanism to process the multi-modal uncertainty representation to generate a fact inconsistency representation, perform processing on the multi-modal uncertainty representation based on a sentiment manipulation graph convolution network to generate a sentiment inconsistency representation, fuse the fact inconsistency representation and the sentiment inconsistency representation to generate fusion features, and classify based on the fusion features to obtain a fake news detection classification result.

5. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps in the fake news detection method based on fact-sentiment dual uncertainty according to any one of claims 1-3.

6. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the steps in the fact-sentiment dual uncertainty based fake news detection method in any one of claims 1-3 when executing the program.

7. A computer program product, characterised in that, The computer program product comprises a computer program, and the computer program, when executed by a processor, implements the steps in the fact-sentiment dual uncertainty based fake news detection method in any one of claims 1-3.

Images