Fake news detection method and system based on time interval and knowledge fusion

Abstract

This invention proposes a fake news detection method and system based on time interval and knowledge fusion, belonging to the fields of internet security and deep learning technology. The method involves extracting image features from target news images using a convolutional neural network and converting them into image feature vectors. A text encoder is used to semantically represent the target news text and convert it into a text feature vector. Entities are extracted from the target news text and linked to obtain external knowledge supplements, which are then converted into external knowledge feature vectors. The time interval is calculated by extracting the occurrence time of the target news event and the news release time, and converted into a time interval feature vector. The image feature vector, text feature vector, external knowledge feature vector, and time interval feature vector are fused, and the fused features are input into a classifier to determine the authenticity of the news data. This invention can significantly improve the accuracy and timeliness of fake news identification.

Description

Technical Field

[0001] The present invention belongs to the field of Internet security and deep learning technology, and in particular relates to a fake news detection method and system based on time interval and knowledge fusion. Background Art

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

[0003] With the rapid development of the internet, social media has revolutionized how people access information, fueling a proliferation of misinformation. This misinformation often includes fabricated multimedia content, particularly fake news. The spread of fake news has had a serious impact on society, with some criminals even exploiting this information to manipulate public opinion and disrupt social order. Therefore, detecting and regulating fake news, ensuring that users receive authentic information, and thus maintaining social harmony, is a top priority.

[0004] However, the identification and detection of fake news on social media is still in its infancy. Relying solely on manual identification and detection is difficult to fully address already published fake news. Consequently, researchers in related fields have begun developing algorithmic models based on deep learning to automatically detect fake news, significantly improving detection efficiency compared to manual methods.

[0005] Currently, researchers in related fields have been conducting increasingly in-depth research on fake news detection using deep learning methods. However, the inventors have discovered that existing fake news detection methods still have some technical problems, such as:

[0006] (1) To improve the accuracy of fake news detection, existing fake news detection methods have begun to analyze news data based on multiple modalities. However, in the process of model design and feature extraction, modal features are only focused on the direct fusion of news text and news images, while ignoring more complex semantic associations. This leads to the neglect of the rich knowledge information behind the news text; this information is useful for fake news detection, and external knowledge can provide rich association information. Therefore, the accuracy of existing fake news detection methods is still not ideal.

[0007] (2) Existing fake news detection methods lack the analysis and judgment of the time interval information of news events during detection, ignoring the characteristics of the time dimension. Without the time interval driving, it is difficult to accurately capture the evolution of information and the changes in relationships. Therefore, the existing technology lacks timeliness in judging the authenticity of news. Summary of the Invention

[0008] To overcome the above-mentioned shortcomings of the existing technology, the present invention provides a fake news detection method and system based on time interval and knowledge fusion, which can take into account the complex semantic associations and time dimension characteristics in news data, thereby improving the accuracy and timeliness of identifying fake news.

[0009] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0010] A first aspect of the present invention provides a fake news detection method based on time interval and knowledge fusion.

[0011] The fake news detection method based on time interval and knowledge fusion includes:

[0012] Acquire target news data, wherein the target news data includes a target news image and a target news text;

[0013] Image features of the target news image are extracted based on the VGG-19 convolutional neural network, and the obtained image features are converted into image feature vectors; semantic representation of the target news text is performed based on the BERT text encoder, and the obtained semantic representation is converted into a text feature vector; entities are extracted from the target news text and entity linking is performed to obtain external knowledge supplements, and the obtained external knowledge supplements are converted into external knowledge feature vectors; the time interval is calculated by extracting the occurrence time of the target news event and the news release time, and the obtained time interval is converted into a time interval feature vector;

[0014] The image feature vector, the text feature vector, the external knowledge feature vector and the time interval feature vector are subjected to feature fusion, and the fused features are input into a multi-layer perceptron classifier to judge the authenticity of the target news data.

[0015] Furthermore, after acquiring the target news data and before extracting the features, it is necessary to determine whether there is noise data in the target news text. If so, the target news text is denoised; wherein, the noise data includes special symbols and expressions in the news text.

[0016] Furthermore, the last layer of the VGG-19 convolutional neural network is a fully connected layer, which is used to adjust the dimension of the output image features.

[0017] Furthermore, the target news text is semantically represented based on the BERT text encoder, including: vectorizing the target news text and processing it into a word sequence, and inputting the word sequence into the BERT text encoder; encoding the word sequence based on the BERT text encoder, and using the aggregated sequence of the BERT text encoder as the semantic representation of the target news text.

[0018] Furthermore, entity linking tools are used to obtain external knowledge supplements and eliminate ambiguity between text and knowledge based on entity associations.

[0019] Furthermore, normalization and encoding operations are performed on the time interval to convert the obtained time interval into a time interval feature vector.

[0020] Furthermore, when the fused features are input into the multi-layer perceptron classifier to judge the authenticity of the target news data, the softmax logistic regression function is used to obtain the binary classification result of whether the target news is fake news.

[0021] A second aspect of the present invention provides a fake news detection system based on time interval and knowledge fusion.

[0022] The fake news detection system based on time interval and knowledge fusion includes:

[0023] The data acquisition module is configured to: acquire target news data, wherein the target news data includes a target news image and a target news text;

[0024] The feature vector acquisition module is configured to: extract image features of the target news image based on the VGG-19 convolutional neural network and convert the obtained image features into an image feature vector; perform semantic representation of the target news text based on the BERT text encoder and convert the obtained semantic representation into a text feature vector; extract entities from the target news text and perform entity linking to obtain external knowledge supplements, and convert the obtained external knowledge supplements into an external knowledge feature vector; calculate the time interval by extracting the occurrence time of the target news event and the news release time, and convert the obtained time interval into a time interval feature vector;

[0025] The fake news detection module is configured to: perform feature fusion on the image feature vector, text feature vector, external knowledge feature vector and time interval feature vector, and input the fused features into a multi-layer perceptron classifier to judge the authenticity of the target news data.

[0026] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the fake news detection method based on time interval and knowledge fusion as described in the first aspect of the present invention.

[0027] A fourth aspect of the present invention provides an electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the fake news detection method based on time interval and knowledge fusion as described in the first aspect of the present invention are implemented.

[0028] One or more of the above technical solutions have the following beneficial effects:

[0029] (1) The present invention simultaneously combines four types of features in news data: text features, image features, temporal information features, and external knowledge. These four types of modal features are then extracted through a deep learning model and feature splicing is employed to achieve efficient fusion of multimodal information. Based on image and text features, the present invention utilizes entity extraction and entity linking techniques to associate target news text with an external knowledge base to obtain supplementary semantic information. Therefore, the present invention can significantly enhance the model's ability to understand news semantics, thereby improving the accuracy of identifying false news.

[0030] (2) When acquiring features, the present invention extracts the occurrence and release times of news events and uses time normalization to generate time interval features. This time interval-based driving force more accurately captures the evolution of information and changes in relationships. Therefore, the present invention makes the judgment of news authenticity more timely.

[0031] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0033] Figure 1 This is a flowchart of a fake news detection method based on time interval and knowledge fusion in Example 1 of the present invention.

[0034] Figure 2 Schematic diagram of the overall framework structure adopted by the fake news detection method based on time interval and knowledge fusion in Example 1 of the present invention. DETAILED DESCRIPTION

[0035] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0036] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to limit the exemplary embodiments according to the present invention.

[0037] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0038] The overall idea proposed by the present invention: The present invention provides a fake news detection method based on the fusion of time intervals and knowledge, aiming to solve the problems of insufficient feature extraction, low external knowledge utilization and unsatisfactory model performance in existing fake news detection. Specifically: first, relevant data is obtained from the target news, and the news images and texts are processed separately to extract image features and text features, and then converted into vectorized representations. Subsequently, entities are extracted from the news text and linked, external knowledge information is supplemented, and external knowledge feature vectors are obtained through further processing; at the same time, the occurrence time and release time of the target news event are extracted, and the time interval features are calculated and vectorized. Finally, the text features, image features, external knowledge features and time interval features are spliced, and the target news is classified as fake news. Based on the mining of multimodal features and external knowledge, this invention can effectively improve the accuracy and timeliness of fake news detection, and provide a comprehensive and efficient solution for fake news detection.

[0039] In addition, to facilitate understanding of the technical solutions of the present invention, the following terminology is provided:

[0040] 1) Multimodal fake news: news content that uses multiple different types of media elements (usually text and images) to spread false or inaccurate information.

[0041] 2) Multimodal fake news identification: Identify whether news is fake or not based on multimodal news data.

[0042] 3) Vectorized representation: The process of converting text or other unstructured data into numerical vectors that can be processed by computers. It can map high-dimensional, unstructured data into a low-dimensional or fixed-dimensional feature space, facilitating model learning and calculation.

[0043] 4) BERT (Bidirectional Encoder Representations from Transformers): A pre-trained natural language processing model based on the Transformer architecture that can effectively generate text feature vector representations for various NLP tasks.

[0044] 5) VGG-19: A deep convolutional neural network (CNN) that is widely used in various computer vision tasks such as image classification and feature extraction, with a depth of 19 layers.

[0045] 6) Normalization Layers: Use activation functions (such as Sigmoid and Tanh) to convert input information into a feature representation acceptable to the model, thereby facilitating subsequent model processing.

[0046] 7) Fully Connected Layer: A basic layer type in a neural network, where each neuron is connected to every neuron in the previous layer. Fully connected layers map input features to the output space and are a key component of the model's final classification or regression task.

[0047] 8) Concatenate: concatenate two or more vectors into one vector.

[0048] 9) Binary Classification Model: A machine learning model used to solve binary classification tasks. Its goal is to divide samples into two categories based on input features, such as predicting whether a piece of news is fake news (true or false).

[0049] Example 1

[0050] This embodiment discloses a fake news detection method based on time interval and knowledge fusion.

[0051] like Figure 1 As shown in FIG, the fake news detection method based on time interval and knowledge fusion includes:

[0052] Step S1: Acquire target news data, wherein the target news data includes a target news image and a target news text;

[0053] Step S2: performing image feature extraction on the target news image based on the VGG-19 convolutional neural network, and converting the obtained image features into image feature vectors; performing semantic representation on the target news text based on the BERT text encoder, and converting the obtained semantic representation into a text feature vector; extracting entities from the target news text and performing entity linking to obtain external knowledge supplements, and converting the obtained external knowledge supplements into external knowledge feature vectors; calculating the time interval by extracting the occurrence time of the target news event and the news release time, and converting the obtained time interval into a time interval feature vector;

[0054] Step S3: performing feature fusion on the image feature vector, text feature vector, external knowledge feature vector and time interval feature vector, and inputting the fused features into a multi-layer perceptron classifier to judge the authenticity of the target news data.

[0055] Based on the above process, the present invention is able to take into account the complex semantic associations and temporal characteristics of news data, thereby improving the accuracy and timeliness of false news identification. To facilitate understanding of the technical solution of the present invention, the specific implementation steps of the technical solution of the present invention are further explained and illustrated below.

[0056] Step S1: Acquire target news data; wherein the target news data includes target news images and target news texts.

[0057] In this embodiment, a public multimodal dataset is used as the basic dataset; at the same time, new news data is obtained from several large websites with rich data volumes by using web crawler technology to amplify the data; finally, target news data is obtained from the amplified multimodal dataset.

[0058] Furthermore, after acquiring the target news data and before feature extraction, it is necessary to determine whether there is noise data in the target news text. If so, the target news text is subjected to denoising to reduce the interference of irrelevant information. Noise data generally refers to special symbols and expressions in the news text.

[0059] Step S2: Extract image features from the target news image using the VGG-19 convolutional neural network and convert the resulting image features into an image feature vector; perform semantic representation on the target news text using the BERT text encoder and convert the resulting semantic representation into a text feature vector; extract entities from the target news text and perform entity linking to obtain external knowledge supplements, and convert the resulting external knowledge supplements into an external knowledge feature vector; calculate the time interval by extracting the occurrence time of the target news event and the news release time, and convert the resulting time interval into a time interval feature vector. This can be achieved specifically through the following steps:

[0060] Step S2-1: Extract image features of the target news image based on the VGG-19 convolutional neural network, and convert the obtained image features into image feature vectors.

[0061] VGG-19 is a classic deep convolutional neural network consisting of 19 weighted layers, including 16 convolutional layers and 3 fully connected layers. Specifically, the convolutional layers are used for feature extraction, achieving efficient feature extraction by stacking small convolution kernels. The pooling layers are used for downsampling to reduce the size of feature maps while retaining important features. The fully connected layers are used to map the extracted features to the classification space. The softmax output layer can output a probability distribution for classification.

[0062] Furthermore, if Figure 2 As shown in the figure, a pre-trained VGG-19 network is used to extract the image features of the target news. That is, the target news image is input into the VGG-19-based convolutional neural network for feature extraction. The deep features of the image are extracted through the convolutional layer in the network, and the extracted image features are vectorized to generate an image feature vector. To ensure that the multimodal features of subsequent news can be aligned and spliced, a fully connected layer is added to the last layer of the VGG-19-based convolutional neural network to adjust the dimension of the final output image features. Specifically, the calculation formula of the image feature vector is as follows:

[0063] R v =σ(W v ·F vgg );

[0064] Among them, R v represents the image feature vector, W v represents the weight matrix of the fully connected layer attached to the output layer of the VGG-19 network; F vgg Represents the output features of the last layer of the VGG-19 network. The output features F vgg After the linear transformation, the tanh activation function (denoted by the symbol σ) is applied to perform nonlinear processing on the result of the linear transformation, and finally the image feature vector R with unified dimension is obtained. v , for subsequent splicing operations.

[0065] The advantage of using VGG-19 as the image feature extractor in this embodiment is that by stacking multiple convolutional layers, high-dimensional deep features in the image can be extracted. This network structure can capture the details and high-level semantic information of the image, while also reducing the dependence on large-scale annotated data and improving the accuracy of feature extraction.

[0066] Step S2-2: semantically represent the target news text based on the BERT text encoder, and convert the obtained semantic representation into a text feature vector.

[0067] First, the target news text is vectorized and BERT is used as a text encoder. When input, the text is processed into a word sequence T = [T1, T2, ..., T n ]; where n represents the length of the text. Then, the word sequence T (i.e., the input sequence) is passed into the pre-trained BERT model and converted into an input feature matrix H. The input feature matrix H contains the contextual features of each word, namely:

[0068] H = BERT(T);

[0069] Where H represents the input feature matrix, T represents the input sequence, BERT(·) represents encoding the input sequence T, and uses BERT’s aggregate sequence representation as the text feature; the text feature BERT is input into the fully connected layer, which can be further reduced to a fixed-dimensional text feature vector R T .

[0070] In this embodiment, the advantage of using BERT as a text feature extractor is that it can capture the deep contextual semantics of words from the input text and model the text semantics more accurately. At the same time, BERT can also well capture the logical relationships across sentences and long-distance word associations.

[0071] Step S2-3: extract entities from the target news text and perform entity linking to obtain external knowledge supplements, and convert the obtained external knowledge supplements into external knowledge feature vectors.

[0072] For the entities extracted from the target news text content, the entity linking tool is used to obtain external knowledge information supplement and perform entity association to eliminate the ambiguity between the text and the knowledge. The acquired external knowledge is processed into a feature vector by the same BERT text encoder as in step S2-2, and further processed into a fixed-dimensional external knowledge feature R K .

[0073] Furthermore, entity linking tools are used to obtain additional external knowledge information, including:

[0074] Entities in the target news text are extracted using named entity recognition technology. Here, the BERT-Chinese-NER model pre-trained on Hugging Face is used. After extracting the entities, the API of CN-DBpedia (Fudan University's large-scale structured encyclopedia) is used to call the conceptual knowledge corresponding to the extracted entities and write it into a document. BERT is then used for processing, and the process is consistent with processing news text.

[0075] Furthermore, entity association is used to eliminate ambiguity between text and knowledge, including:

[0076] Extract the sentence containing the entity and its preceding and following sentences as context, obtain the knowledge description of the candidate entity through the API, use BERT to encode the context and candidate knowledge description into a vector, calculate the cosine similarity between the context vector and each candidate description vector, and select the best candidate entity and its conceptual knowledge. Specifically, the cosine similarity formula is expressed as:

[0077]

[0078] Among them, cos_similarity represents cosine similarity, v 上下文 represents the context vector, v 知识描述 Represents the candidate knowledge description vector.

[0079] Step S2-4: Calculate the time interval by extracting the occurrence time of the target news event and the news release time, and convert the obtained time interval into a time interval feature vector.

[0080] For each article (target news event), record its release time and parse the time set of all events mentioned in the target news; calculate the time interval by selecting the earliest time in the time set mentioned in the article. Specifically, the time interval calculation formula is:

[0081] tinterval =t post -min{t1,t2,…,t n};

[0082] Among them, t interval represents the time interval, t post represents the news release time, {t1,t2,…,t n} represents the time set of all events mentioned in the target news; min{t1,t2,…,t n} represents the minimum value of all event time sets mentioned in the target news, that is, the earliest time.

[0083] Furthermore, the sigmoid function and tanh function are used to normalize and encode the time interval information of the target news, so as to convert the obtained time interval into a time interval feature vector. Specifically:

[0084] t rep =tanh(Sigmoid(t interval ));

[0085] Among them, t rep represents the time interval feature vector. tanh() represents the activation function, which maps the input value to the range -1 to 1. Sigmoid() also represents the activation function, which maps the input value to the range 0 to 1, suitable for expressing probabilities or normalized results. In the specific implementation, the time interval value is first normalized to the range (0, 1) using the Sigmoid() activation function. Then, the tanh activation function is used to map it to the range (-1, 1) to further enhance the feature's expressiveness.

[0086] In this embodiment, the advantage of introducing the time interval feature is that it can reflect the time difference between the occurrence and reporting of a news event, obtain additional contextual information from the time dimension, thereby enhancing the overall expression ability and improving the accuracy and timeliness of fake news detection.

[0087] Step S3: perform feature fusion on the obtained image feature vector, text feature vector, external knowledge feature vector and time interval feature vector, and input the fused features into a multi-layer perceptron classifier to judge the authenticity of the target news data.

[0088] The image feature vector, text feature vector, external knowledge feature vector, and time interval feature vector are concatenated (feature fusion) and input into a pre-trained classifier to determine whether the target news is fake news. Specifically:

[0089] The text feature vector R is obtained through step S2 T , image feature vector R v, external knowledge feature vector R K and the time interval characteristic t rep By combining the four multimodal features, we can obtain the joint feature representation R of the target news. This joint feature representation R is the fused feature. Specifically, the calculation method of the joint feature representation R is as follows:

[0090]

[0091] in, represents a concatenation operation. Finally, the joint feature representation R is used to determine the authenticity of the target news data through a fully connected multilayer perceptron classifier. Furthermore, when the fused features are input into the multilayer perceptron classifier to determine the authenticity of the target news data, a softmax logistic regression function is used to obtain a binary classification result of whether the target news is fake news. Specifically, the calculation formula is as follows:

[0092] p=softmax(W R ·R);

[0093] Among them, p represents the binary classification result, W R represents the parameters of the fully connected layer in a multilayer perceptron, softmax() represents the calculation of the logistic regression function, and R represents the joint feature representation. This invention, through the combined modeling of multimodal features and the efficient classification strategy of a binary classification detector, not only improves detection accuracy but also significantly increases detection efficiency, enabling rapid identification of fake news.

[0094] Differentiating from existing fake news detection models, this paper proposes a multimodal fake news detection method that comprehensively considers time interval features, external knowledge features, text features, and image features. This method utilizes an innovative time interval feature extraction strategy, combined with text, image, and external knowledge, to provide a more comprehensive and robust fake news detection mechanism from the perspective of temporal dynamics and semantic enhancement. This approach effectively avoids the information loss and modality imbalance that can plague existing models, significantly improving the accuracy and timeliness of fake news detection through more efficient feature extraction and fusion.

[0095] Example 2

[0096] This embodiment discloses a fake news detection system based on time interval and knowledge fusion.

[0097] The fake news detection system based on time interval and knowledge fusion includes:

[0098] The data acquisition module is configured to: acquire target news data, wherein the target news data includes a target news image and a target news text;

[0099] The feature vector acquisition module is configured to: extract image features of the target news image based on the VGG-19 convolutional neural network and convert the obtained image features into an image feature vector; perform semantic representation of the target news text based on the BERT text encoder and convert the obtained semantic representation into a text feature vector; extract entities from the target news text and perform entity linking to obtain external knowledge supplements, and convert the obtained external knowledge supplements into an external knowledge feature vector; calculate the time interval by extracting the occurrence time of the target news event and the news release time, and convert the obtained time interval into a time interval feature vector;

[0100] The fake news detection module is configured to: perform feature fusion on the image feature vector, text feature vector, external knowledge feature vector and time interval feature vector, and input the fused features into a multi-layer perceptron classifier to judge the authenticity of the target news data.

[0101] Example 3

[0102] The purpose of this embodiment is to provide a computer-readable storage medium.

[0103] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the fake news detection method based on time interval and knowledge fusion as described in the first embodiment of the present disclosure.

[0104] Example 4

[0105] The purpose of this embodiment is to provide an electronic device.

[0106] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps of the fake news detection method based on time interval and knowledge fusion as described in the first embodiment of the present disclosure are implemented.

[0107] The steps involved in the apparatuses of Examples 2, 3, and 4 above correspond to those of Method Example 1. For detailed implementations, please refer to the relevant description of Example 1. The term "computer-readable storage medium" should be understood to mean a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and causing the processor to perform any method of the present invention.

[0108] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0109] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it 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 on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A fake news detection method based on time interval and knowledge fusion, characterized by: include: Acquire target news data, wherein the target news data includes a target news image and a target news text; Image features of the target news image are extracted based on the VGG-19 convolutional neural network, and the obtained image features are converted into image feature vectors; semantic representation of the target news text is performed based on the BERT text encoder, and the obtained semantic representation is converted into a text feature vector; entities are extracted from the target news text and entity linking is performed to obtain external knowledge supplements, and the obtained external knowledge supplements are converted into external knowledge feature vectors; the time interval is calculated by extracting the occurrence time of the target news event and the news release time, and the obtained time interval is converted into a time interval feature vector. Specifically: for each target news event, its release time is recorded and all event time sets mentioned in the target news are parsed; the time interval is calculated by selecting the earliest time in the time set mentioned in the event; the time interval calculation formula is: ； in, Indicates the time interval, Indicates the time when the news was released. represents the time set of all events mentioned in the target news; Represents the minimum value of all event time sets mentioned in the target news, that is, the earliest time; The time interval information of the target news is normalized and encoded using the Sigmoid function and the tanh function, and the obtained time interval is converted into a time interval feature vector, namely: ； in, represents the time interval feature vector; Represents the activation function, which is used to map the input value to between -1 and 1; It also represents the activation function, which is used to map the input value to between 0 and 1. Specifically, the activation function is first used Normalize the time interval value to (0,1); then, map it to the (-1,1) interval through the activation function tanh to enhance the feature’s expressiveness; The image feature vector, the text feature vector, the external knowledge feature vector and the time interval feature vector are subjected to feature fusion, and the fused features are input into a multi-layer perceptron classifier to judge the authenticity of the target news data.

2. The fake news detection method based on time interval and knowledge fusion as claimed in claim 1, characterized in that: After acquiring the target news data and before extracting the features, it is necessary to determine whether there is noise data in the target news text. If so, the target news text is subjected to denoising processing; wherein, the noise data includes special symbols and expressions in the news text.

3. The fake news detection method based on time interval and knowledge fusion as claimed in claim 1, characterized in that: The last layer of the VGG-19 convolutional neural network is a fully connected layer, which is used to adjust the dimension of the output image features.

4. The fake news detection method based on time interval and knowledge fusion as claimed in claim 1, characterized in that: The target news text is semantically represented based on the BERT text encoder, including: vectorizing the target news text and processing it into a word sequence, and inputting the word sequence into the BERT text encoder; encoding the word sequence based on the BERT text encoder, and using the aggregated sequence of the BERT text encoder as the semantic representation of the target news text.

5. The fake news detection method based on time interval and knowledge fusion as claimed in claim 1, characterized in that: Use entity linking tools to obtain external knowledge supplements and eliminate ambiguity between text and knowledge based on entity associations.

6. The fake news detection method based on time interval and knowledge fusion as claimed in claim 1, characterized in that: When the fused features are input into the multi-layer perceptron classifier to judge the authenticity of the target news data, Logistic regression function is used to obtain the binary classification result of whether the target news is fake news.

7. A fake news detection system based on time interval and knowledge fusion, characterized by: include: The data acquisition module is configured to: acquire target news data, wherein the target news data includes a target news image and a target news text; The feature vector acquisition module is configured to: extract image features of the target news image based on the VGG-19 convolutional neural network, and convert the obtained image features into image feature vectors; perform semantic representation of the target news text based on the BERT text encoder, and convert the obtained semantic representation into a text feature vector; extract entities from the target news text and perform entity linking to obtain external knowledge supplements, and convert the obtained external knowledge supplements into external knowledge feature vectors; calculate the time interval by extracting the occurrence time of the target news event and the news release time, and convert the obtained time interval into a time interval feature vector. Specifically: for each target news event, record its release time and parse all event time sets mentioned in the target news; calculate the time interval by selecting the earliest time in the time set mentioned in the event; the time interval calculation formula is: ； in, Indicates the time interval, Indicates the time when the news was released. represents the time set of all events mentioned in the target news; Represents the minimum value of all event time sets mentioned in the target news, that is, the earliest time; The time interval information of the target news is normalized and encoded using the Sigmoid function and the tanh function, and the obtained time interval is converted into a time interval feature vector, namely: ； in, represents the time interval feature vector; Represents the activation function, which is used to map the input value to between -1 and 1; It also represents the activation function, which is used to map the input value to between 0 and 1. Specifically, the activation function is first used Normalize the time interval value to (0,1); then, map it to the (-1,1) interval through the activation function tanh to enhance the feature’s expressiveness; The fake news detection module is configured to: perform feature fusion on the image feature vector, text feature vector, external knowledge feature vector and time interval feature vector, and input the fused features into a multi-layer perceptron classifier to judge the authenticity of the target news data.

8. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps of the fake news detection method based on time interval and knowledge fusion as described in any one of claims 1 to 6 are implemented.

9. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the fake news detection method based on time interval and knowledge fusion as described in any one of claims 1 to 6 are implemented.

Images