A method and device for detecting fake news based on reader behavior simulation

Abstract

The application provides a false news detection method and device based on reader behavior simulation. The method comprises: obtaining a news to be detected and extracting components; inputting the components into a trained false news detection model, the false news detection model comprising: two or more component-in-component feature extractors corresponding to the components one by one; a component-inter-component feature serialization module comprising one or more component-inter-component feature extractors and one sequence-based aggregator, the one or more component-inter-component feature extractors capturing component-inter-component features between any two components, and the sequence-based aggregator integrating all the component-inter-component features into a feature sequence according to a set order; and a detector obtaining true and false prediction probabilities. The false news detection model can discover the affinity between different components, comprehensively and effectively model the news, better mine the features within the components and between the components, learn better news representations by simulating the reading and verification behaviors of readers in the components, and improve the detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of news detection technology, and in particular to a method and device for detecting fake news based on reader behavior simulation. Background Technology

[0002] The veracity of information on social media is difficult to guarantee because every user can express their opinions and forward posts they believe, making social media an ideal environment for the spread of fake news. Since the widespread dissemination of fake news can lead to serious social consequences, it is necessary to remove fake news from social media.

[0003] To achieve automated detection of fake news on social media, recent research has made significant contributions by modeling fake news from both unimodal and multimodal perspectives. Unimodal methods use only one modality to detect fake news, such as text or images (text as one modality, images as another). However, determining the authenticity of news solely through text or images is challenging, as over 20% of fake news appears in image form. Furthermore, news on social media generally includes both text and images: text provides detailed descriptions of news events, while images enhance the narrative's vividness and credibility. Meanwhile, malicious fake news fabricators may exploit this psychology, meticulously crafting news text and images to create multimodal fake news to attract readers and gain more attention. In this context, multimodal fake news detection methods have been proposed to effectively integrate features from both intramodal and intermodal sources that are beneficial for detection. Some works explore the similarity between news text and images from a semantic perspective, neglecting the physical features of the images. Others, building upon text-image similarity, further introduce image tampering detection. In addition, some recent studies have enhanced textual and visual representations by extracting entities from images to improve fake news detection performance.

[0004] Generally, news on social media consists of different components, including headlines, images, comments, and the body text, such as... Figure 2The image illustrates four components of a news article: headline, image, comments, and body text, along with the preferred reading order among readers. While existing fake news detection methods have achieved commendable results, they fail to delve deeper into the behavior of users reading news on social media. Through extensive investigation and research, the inventors of this application discovered that readers follow a relatively fixed order when reading different components of a news article (i.e., sequentially searching for implicit clues). If necessary, readers typically verify adjacent components to enhance their understanding or assess the news's veracity. Since the reading order reflects the importance of each component, and the features between components (such as semantic differences) help strengthen semantics, the inventors found that both aspects play a crucial role in assessing the veracity of news. However, existing detection methods, whether unimodal or multimodal, neglect this behavior of readers typically utilizing differences between components and implicit clues when verifying news veracity. Summary of the Invention

[0005] The present invention aims to solve the technical problems existing in the prior art and provide a method and device for detecting fake news based on reader behavior simulation.

[0006] To achieve the above-mentioned objectives of the present invention, according to a first aspect of the present invention, the present invention provides a method for detecting fake news based on reader behavior simulation, comprising: acquiring news to be detected and extracting components from the news to be detected, the news to be detected including at least two of four components: title, comments, body text, and image; inputting the extracted components into a trained fake news detection model to obtain a true / false prediction probability, the fake news detection model comprising: two or more component-specific feature extractors corresponding one-to-one with the components, used to extract the component-specific features of each component; an inter-component feature serialization module, including one or more inter-component feature extractors and a sequence-based aggregator, wherein the one or more inter-component feature extractors respectively capture the inter-component features between any two components, and the sequence-based aggregator integrates all inter-component features into a feature sequence according to a set order; and a detector, which processes the feature sequence to obtain the true / false prediction probability of the news to be detected.

[0007] To achieve the above-mentioned objectives of the present invention, according to a second aspect of the present invention, an electronic device is provided, the electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the fake news detection method based on reader behavior simulation as described in the first aspect of the present invention.

[0008] The beneficial technical effects of this invention are as follows: News is refined into components such as headlines, body text, images, and comments, providing a more granular and comprehensive approach compared to multimodal fake news detection. A corresponding number of in-component feature extractors can be configured based on the number of component types in the news to be detected. These in-component feature extractors simulate the reader's reading process from a local area to the entire component, capturing in-component features, specifically semantic or visual features within the component. The inter-component feature serialization module simulates the reader's understanding and verification process between components. Using an inter-component feature extractor with a refinement loss, inter-component features are obtained by simulating the reader's reading and verification behavior between two components. These inter-component features are then integrated into a feature sequence according to a set order to simulate the reader's reading order. The detector processes the feature sequence to obtain the predicted probability of the news being true or false. The fake news detection model provided by this invention can discover the affinity between different news components, thus comprehensively and effectively modeling news. It can better mine in-component and inter-component features, and learn better multi-component news representations by simulating the reader's reading and verification behavior on news components, thereby improving the accuracy of fake news detection. It has excellent practicality and versatility. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a method for detecting fake news based on reader behavior simulation in one embodiment of the present invention.

[0010] Figure 2 This is a diagram illustrating the reading stages and sequence among multiple components of a news article;

[0011] Figure 3 This is a schematic diagram of a fake news detection model in another embodiment of the present invention;

[0012] Figure 4 This is a schematic diagram of the generalization experiment verification results of the fake news detection model in another embodiment of the present invention;

[0013] Figure 5 This is a schematic diagram of the verification results of hyperparameter λ in another embodiment of the present invention;

[0014] Figure 6 This is a structural block diagram of an electronic device according to another embodiment of the present invention;

[0015] Figure 7 This is a performance comparison of different methods on four datasets in another embodiment of the present invention. Detailed Implementation

[0016] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0017] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0018] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0019] Inspired by readers' reading behavior on social media, the inventors of this application conducted extensive research. Based on previous research on modal similarity or consistency, they concluded that the components of authentic news possess high affinity and support, and used this to verify the authenticity of news, for example... Figure 2 All components in (a) mention actor Ben Affleck, while fake news does not. Figure 2Taking a news article containing four components (headline, image, comment, and body) as an example, the reading stages and sequences among multiple components are illustrated. Subgraph (a) shows two reading stages from the homepage to the details page, and subgraph (b) shows the reading order that readers generally prefer among the components. Therefore, the inventors aim to make their constructed fake news detection model (Emulate the Behavior of Readers, or Ember) conduct a comprehensive investigation of numerous news components, modeling news thoroughly and comprehensively, refining the fake news detection process from cross-modal to cross-component. This refinement faces an unavoidable challenge: how to simultaneously mine the inter-component features between different news components and effectively integrate them to facilitate fake news detection. Existing methods cannot directly solve these problems because they meticulously design different fusion modules to handle specific inter-component features while ignoring the hidden features between other components. To address this challenge, the inventors conducted a comprehensive investigation into how readers read and verify news on social media, defined the reading and verification order, and established a fake news detection model.

[0020] The inventors discovered that due to the layout of social media, the process by which readers obtain different news components can be represented in two stages: (1) first, obtaining the news headlines and images on the social media homepage; (2) after clicking on a news item, obtaining the corresponding text and reader comments on the details page. Under this layout, readers generally combine the headlines and images in stage (1) to initially assess the authenticity of the news. Then, they take into consideration the text and reader comments in stage (2) for further verification. In addition, the study found that readers usually read the quick information first and make decisions based on this information, such as deciding whether to forward a news post by reading only the headline. Since comments are much shorter than the text and are usually as concise as the headline, the inventors proposed that the reader's reading and verification behavior can be described as a sequence from quick information to long text: headline-image-comment-text.

[0021] Based on the inventors' above research and findings, this invention provides a method for detecting fake news based on reader behavior simulation. In one embodiment, such as... Figure 2 As shown, the method includes:

[0022] Step S1: Obtain the news to be detected and extract the components from the news to be detected. The news to be detected includes at least two of the four components: title, comment, body text and image.

[0023] In practical applications, the news to be detected can include any two of the following four components: title, comment, body text, and image; or any three of the four components; or all four components.

[0024] A news article with a title (H), a picture (I), a comment (C), and a body (B) can be written as New = (H, I, C, B).

[0025] Step S2: Input the extracted components into the trained fake news detection model to obtain the true / false prediction probability. Figure 3 This diagram illustrates the fake news detection model structure when the news item to be detected contains four components: title, comments, body text, and image. In practical applications, when the number of component types in the news item to be detected is less than four, it can be referenced... Figure 3 A variant fake news detection model is obtained by removing a portion of the intra-component feature extractors and inter-component feature extractors. (See reference...) Figure 3 The fake news detection model includes: two or more component-specific feature extractors (FEs) corresponding one-to-one with each component, used to extract the intra-component features of each component; an inter-component feature serialization module, including one or more inter-component feature extractors and a sequence-based aggregator, wherein the inter-component feature extractors capture the inter-component features between any two components, and the sequence-based aggregator integrates all inter-component features into a feature sequence according to a set order; and a detector, which processes the feature sequence to obtain the probability of predicting whether the news to be detected is true or false.

[0026] In this embodiment, when the news to be detected contains 4 components, there are 4 in-component feature extractors (FEs), each corresponding to one of the 4 components, used to extract the in-component features of the corresponding components. Similarly, when the news to be detected contains 3 components, there are 3 in-component feature extractors (FEs), each corresponding to one of the 3 components, used to extract the in-component features of the corresponding components. When the components are titles, comments, or body text, the extracted in-component features are semantic features; when the component is an image, the extracted in-component features are visual features.

[0027] In this embodiment, the number of component feature extractors is also related to the number of components contained in the news to be detected. If the news to be detected contains 4 components, then it is necessary to set... Feature extractor between components.

[0028] In another embodiment, when the news to be detected includes the main text, the feature extractor within the component corresponding to the main text includes a first bidirectional GRU layer, a main text attention layer, and a second bidirectional GRU layer connected in sequence. The first bidirectional GRU layer encodes words in the main text, the main text attention layer redistributes word weights based on the importance of words in the main text, and the second bidirectional GRU layer encodes sentences in the main text.

[0029] Since all text information consists of words and sentences, the title, comment, and body text are formalized as follows: and Where I, J, and K represent the number of sentences. It is N k The main body sentence consists of 100 words. Similarly, H i and C j They respectively represent the contents of N i and N j A sentence of 1 word. Titles are typically limited to one sentence, so setting I=1 is more realistic. Furthermore, a GloVe model is introduced to vectorize each word in the sentence to d dimensions, achieving word vectorization.

[0030] Specifically, to avoid information loss during long sentence input, two bidirectional GRU layers (Bi-GRU, or bidirectional gated recurrent layer) are used to encode words and sentences respectively. The first Bi-GRU layer (i.e., the first text-based bidirectional GRU layer) is used for word encoding, which can be described as forward and backward reading:

[0031]

[0032] Where k represents the index of the k-th sentence in the text. It is the first Bi-GRU layer surrounding the b-th word of the k-th sentence. Output in both directions.

[0033] Because attention mechanisms have demonstrated powerful capabilities in natural language processing, a text-based attention layer was further introduced to redistribute word weights based on word importance. The weight of each word can be calculated as follows:

[0034]

[0035]

[0036] Among them, U b , These represent the learnable parameters of the first network, the second network, and the third network, respectively. It is the attention weight of the b-th word in the k-th sentence of the main text. It is obtained from a fully connected layer with a tanh activation function. The implicit representation of the sentence. Therefore, a weighted sentence representation vector. It can be calculated using the following formula:

[0037]

[0038] To further extract contextual information from sentences, a second Bi-GRU layer (i.e., a second bidirectional GRU layer) is introduced after the main text attention layer. This layer simulates the process of a user reading news sentence by sentence. Therefore, the features of the k-th sentence can be represented as:

[0039] h k =Bi-GRU2(s) k ), k∈{1,…,K}, 5

[0040] in, It is B i -GRU2 surrounds sentences k The output in both directions is the final sentence encoding.

[0041] This embodiment constructs a hierarchical attention network to capture the semantic information of text from words to sentences, just as readers read news text on social media from local words to whole text components, demonstrating significant performance in feature extraction.

[0042] In another embodiment, when the news to be detected includes a title, the in-component feature extractor corresponding to the title includes a first bidirectional LSTM layer for the title, a title attention layer, and a second bidirectional LSTM layer for the title, connected in sequence. The first bidirectional LSTM layer encodes words in the title, the title attention layer reassigns word weights based on the importance of words in the title, and the second bidirectional LSTM layer encodes sentences in the title. The in-component feature extractor for the title can be seen as a variant of the in-component feature extractor for the body text, using a bidirectional LSTM layer (Bi-LSTM) instead of a Bi-GRU layer to achieve better results, because the Bi-LSTM with more parameters can more accurately process the highly condensed information in the title.

[0043] In another embodiment, when the news to be detected includes comments, the intra-component feature extractor for the comments includes a comment attention layer and a comment bidirectional GRU layer connected in sequence. The intra-component feature extractor for comments can be seen as a variant of the intra-component feature extractor for the main text. To better extract the semantic features of the comments, the first Bi-GRU layer is discarded because the hierarchical structure is not suitable for short texts, and in most cases, comments are not as closely related to the news content as the headline.

[0044] In another embodiment, previous research has demonstrated that images can provide rich additional information for fake news detection. This embodiment simulates the reading process of a reader on an image component, aiming to capture image features from both semantic and physical perspectives to obtain the final visual features. When the news to be detected includes an image, the in-component feature extractor corresponding to the image includes:

[0045] The physical feature extraction module uses an error level analysis (ELA) algorithm to process images and obtain ELA images. This is to mine the physical features of the images and detect whether they have been tampered with. The error level analysis algorithm is existing technology; see the paper "Detecting fake news by exploring the consistency of mukimodal data" by Xue et al., Information Processing Management, vol.58, no.5, p.102610, 2021. The ELA algorithm can highlight malicious stitching parts in fake images to varying degrees by setting different error levels r. The error level r is set to 0.3. The ELA image is represented as follows:

[0046]

[0047] Among them, V l This represents the l-th original image of the news article to be tested.

[0048] The image vectorization module vectorizes the original image from the news to be detected to obtain the original image vector, and vectorizes the ELA image to obtain the ELA image vector. Specifically, it uses a pre-trained ResNet50 network to vectorize both the original image and the ELA image:

[0049]

[0050]

[0051] in, Let represent the original image vector of the l-th original image. Let L represent the ELA image vector of the l-th original image, where L represents the number of images in the news article to be detected.

[0052] The connection unit concatenates the original image vector and the ELA image vector to obtain the image component vector. and Connection as This represents the image component vector of the l-th original image, enabling the model to comprehensively process image features from both semantic and physical perspectives.

[0053] The image attention layer identifies the importance of different parts of the image component vectors and reweights each part to obtain a weighted image vector. Different parts refer to different regions, preferably, but not limited to, square or circular regions. Similar to the text feature extraction module, it utilizes a... An attention layer is used to identify the importance of different parts in the two images and to reweight them. Weighted image vectors. Represented as:

[0054]

[0055] Attention is an abbreviation for attention mechanism, and its detailed process can be found in formulas (2)-(4).

[0056] A bidirectional GRU layer is used to encode the weighted image vector. To align with text features, the weighted image vector... The image is input into a Bi-GRU to obtain the final image embedding. The image output after bidirectional GRU layer encoding is shown. Represented as:

[0057]

[0058] in, This indicates the l-th image (V) in the news article to be tested. l The final representation obtained after processing by the feature extractor within the corresponding component.

[0059] The inter-component feature serialization module simulates the reader's understanding and verification process. For example... Figure 2 As shown, different news components on social media should complement and support each other, which is one of the fundamental bases for people to better distinguish fake news. To replicate the reader's verification process among these components, this application combines them in pairs and mines the inter-component information within each combination, i.e., inter-component features. These inter-component features are concatenated in a predetermined order and provided to a sequence-based aggregator. This allows for a more comprehensive capture of features surrounding the news that are conducive to detection.

[0060] In another embodiment, the inter-component feature extractor in the inter-component feature serialization module captures inter-component features between two components based on a parallel co-attention method, specifically including:

[0061] Step A: Calculate the affinity matrix between the intra-group features of the first component and the intra-group features of the second component; both the first and second components are from the news items to be detected and are different. The intra-group features of the first component are represented as: N represents the number of sentences or images in the first component. The within-group features of the second component are represented as follows: Q represents the number of sentences or images in the second component. Calculate the affinity matrix. To measure P D and P E Affinity between them:

[0062]

[0063] in, Let P represent the learnable first parameter matrix. Specifically, when a news article has four accessible components, P... D P E ∈{title, image, comment, body} and P D ≠P E .

[0064] Step B: Use the affinity matrix to obtain the updated representation H of the within-group features of the first component. D And the update representation of the intra-group features of the second component H E :

[0065] H D =tanh(W D P D +(W E P E A), 12

[0066] H E =tanh(W E P E +(W D P D A T ), 13

[0067] Among them, W D W represents the sampling matrix of the first component. E W represents the sampling matrix of the second component. D , k′ represents the number of sentences or images sampled.

[0068] Step C, using the updated representation H of the intra-group features of the first component. D Calculate the attention score vector a of the within-group features of the first component. D H is represented by the update of the intra-group features of the second component. E Calculate the attention score vector a of the within-group features of the second component. E :

[0069]

[0070] Among them, W DE W represents the learnable weight matrix of the first component relative to the second component. ED W represents the learnable weight matrix of the second component relative to the first component. DE ,

[0071] Step D, utilize the attention score vector a of the intra-group features of the first component.D Calculate the weighted representation of the first component relative to the second component. Attention score vector a using the intra-group features of the second component E Calculate the weighted representation of the second component relative to the first component. The weighted representation of the first component relative to the second component Weighted representation of the second component relative to the first component Obtain the inter-component feature O between the first component and the second component. DE The specific process is as follows:

[0072]

[0073]

[0074] in, The intra-group feature P of the first component D The output contains the within-group features P from the second component. E Weighted information, Similarly, Furthermore, the inter-component characteristics between the first component and the second component It is the final output of the co-attention operation, integrating two news-related components.

[0075] In this embodiment, lazy thinking about news content may make it difficult for readers to judge the authenticity of news, while analytical thinkers can better distinguish between fake news and real news. Based on this, this application further introduces a news overall refinement loss module to refine the overall representation of news. Specifically, this refinement process is performed on the set last reading component (i.e., the last component of the set reading sequence), because only at this time can readers obtain all news components (i.e., read the entire news article). In order to simulate the behavior of readers rereading the entire news to enhance their understanding, the three enhanced weighted representations of the set last reading component calculated by formula (15) are connected to obtain the overall refined representation of news. For example, when the set last reading component is the main text component, the overall refined representation of news is: in and These are the main text representations that integrate news headlines, images, and comments. Therefore, Fea R It is a news report that needs to be presented in more detail.

[0076] In order to ascertain the veracity of news by repeatedly reading it, just like a reader, Fea R The second predicted probability G is obtained by inputting a second multilayer perceptron (preferably, but not limited to, a two-layer MLP). R :

[0077] G R =MLP R (Fea R ), 17

[0078] Preferably, to learn high-quality news representations during the training of the fake news detection model, a news overall refinement loss module is introduced to refine the overall news representation. A news overall refinement loss function is established, and an inter-component feature extractor with refinement loss is used to simulate the reader's reading and verification behavior for every two components. The refinement loss function further refines the verification process after the reader has obtained all news components. Specifically, during training, a training set is first established, where each training sample contains one news item, and each training sample is associated with a real label representing true or false. Specifically, the real label is replaced by 0 and 1, where 0 represents true and 1 represents false. The news overall refinement loss module performs the following:

[0079] Step 1: For each training sample, merge the weighted representation of the last read component relative to other components in the training sample to obtain the overall refined representation of the news in the training sample; other components are the components in the training sample other than the last read component.

[0080] Step 2: The news as a whole of the training sample is refined and input into the second multilayer perceptron to obtain the second prediction probability of the training sample; the second prediction probability can be the probability that the training sample is false, and the value range of the second prediction probability is 0 to 1.

[0081] Step 3: During the training process on the training set, calculate the overall news refinement loss according to the following formula:

[0082]

[0083] Where, θ R Indicates passage The parameters of the optimized fake news detection model are: N′ represents the number of samples in the training set, i′ represents the index of the training sample, i′∈[1,N′], Y i′ This represents the true label of the i′-th training sample. This represents the second predicted probability of the i′-th training sample.

[0084] In this embodiment, and more preferably, to improve the training effect, the total loss function of the fake news detection model is set as follows during sample training:

[0085]

[0086] Wherein, λ represents the balancing parameter, specifically a hyperparameter that balances the level of detail in the overall representation of the news; θ represents the loss obtained based on the detector output. gru Indicates the process Optimized parameters for the fake news detection model.

[0087]

[0088] Let represent the first predicted probability obtained by the detector for the i′-th training sample. The first predicted probability can represent the probability that the training sample is false, and the value range of the first predicted probability is 0 to 1.

[0089] In another embodiment, when the news to be detected includes four components: title, comments, body text, and image, the sequence-based aggregator integrates all inter-component features into a feature sequence in the following order: inter-component features between the title and image. HI Inter-component features between titles and comments (O) HC Inter-component features O between image comments IC Inter-component features between title and body text (O) HB Inter-component features O between images and text IB Inter-component features O between the comment text and the main body of the text CB The feature integration result among all components can be represented as Fea = [O HI O HC O IC O HB O IB O CB To reflect the reader's reading and verification order, a sequence model (preferred but not limited to GRU) is used as an aggregator to aggregate features from different combinations. Since GRU suffers from long-range dependency issues, this application utilizes a reverse GRU for feature integration, allowing more critical parts to become later inputs, thus mitigating the long-range dependency problem from a data perspective, resulting in the feature sequence Fea. gru for:

[0090]

[0091] Feature sequence Fea gru This represents the final news representation that integrates features from different news components.

[0092] In another embodiment, to detect fake news with multiple components, a GRU-based approach is used. seq Fea learned gru The detector includes a first multilayer perceptron, preferably a two-layer MLP (Multilayer Perceptron), which obtains the true / false prediction probability of the news to be detected.

[0093] Ggru =MLP gru (Fea gru ), 20

[0094] Among them, G gru MLP represents the predicted probability of a news item being true or false. gru Indicates the use of Fea for prediction gru The MLP for tags. The true / false prediction probability represents the probability that the news to be detected is false, and its value ranges from 0 to 1.

[0095] Experiments were conducted to verify the effectiveness of the fake news detection model (hereinafter referred to as Ember) provided in this application:

[0096] 1. Selection of validation dataset

[0097] We selected two fake news datasets from FakeNewsNet, collected from the fact-checking platforms GossipCop and PolitiFact, and the Fakeddit fake news dataset from the social media platform Reddit. GossipCop is a dataset in the gossip domain, while PolitiFact is a dataset in the political domain. Both include news headlines, images, comments, and body text. The PolitiFact dataset only has 286 images; therefore, we built two subsets based on it, named PolitiFact2 and PolitiFact7. PolitiFact2 contains 286 news items with images, and PolitiFact7 contains 748 news items, of which only 268 have images. Fakeddit contains headlines, images, and comments, and includes five types of fake news: satirical (ironic tone), misleading (deliberately fabricated content), artificially manipulated (human-manipulated images), broken links (text and image mismatch), and bot-generated content. We constructed six sub-datasets based on the aforementioned fake news categories, including headlines, images, and comments, named: Satire, MisCon, ManCon, FalCon, ImpCon, and Compre. Compre is a comprehensive dataset containing all five types of fake news. GossipCop and PolitiFact are domain-specific datasets, not category-specific datasets, while Satire, MisCon, FalCon, ImpCon, and Compre are category-specific datasets, not domain-specific datasets. Specifically, Compre is neither domain-specific nor category-specific.

[0098] 2. Baseline Selection

[0099] To evaluate Ember's advantages, we compared it with six state-of-the-art baseline methods, where 'S(n)' and 'M(n)' represent unimodal and multimodal methods using n news-related components, respectively. Furthermore, while the first five methods use body text as the information, SAFE also considers news headlines. To further evaluate the effectiveness of headlines in fake news detection, we replaced the body text used in the first five baselines with headlines and renamed these methods to "baseline_H", for example, "HAN_H".

[0100] •S(1)HAN: HAN uses the body text to detect fake news. This method mines word-level and sentence-level features by constructing a hierarchical attention network.

[0101] •S(1)Text_CNN: Text_CNN models news using the news text itself. They combine traditional kernels to capture features from different dimensions.

[0102] S(2)dEFEND: dEFEND proposes an interpretable fake news detection model that uses the text and user comments to find k interpretable comments, thereby improving the performance of fake news detection.

[0103] ·M(2)HMCAN: HMCAN uses BERT and ResNet50 to extract textual and visual features, constructs a multimodal contextual attention network, and integrates cross-modal features.

[0104] M(2)CAFE: CAFE constructs a cross-modal alignment module that transforms features from different modalities into a shared semantic space and then evaluates the ambiguity between different modalities.

[0105] M(3)SAFE: SAFE uses the similarity between text and images to detect fake news. Specifically, they use a pre-trained image-to-text model to convert images into text and then calculate their similarity.

[0106] 3. Performance Comparison

[0107] Four metrics were used to measure model performance: accuracy, precision, recall, and F1 score. Performance comparisons of different methods on four datasets are shown below. Figure 7In most cases, Ember outperforms the other six baseline methods on four datasets. Specifically, Ember achieves 4.16% higher accuracy and 5.10% higher F1 score than the second-best results on Compre, and at least 5.58% higher across all metrics on the PolitiFact2 and PolitiFact7 sub-datasets, demonstrating Ember's effectiveness in detecting fake news. Furthermore, Ember can detect fake news from different domains (politics and gossip) and handle complex scenarios involving various domains and fake news types (as demonstrated in the Compre dataset), indicating that Ember is more practical and versatile than the baselines.

[0108] 4. Model generalization verification

[0109] To adapt to six subsets built on the Fakeddit dataset, Ember was reconstructed, consisting of three intra-component feature extractors (HFE, CFE, and IFE, and three inter-component feature extractors) and three inter-component feature extractors. The generalization effect of Ember in detecting various fake news items was evaluated, and the impact of fake news categories on model detection performance was explored. Since these subsets do not contain text information, HMCAN_H, dEFEND_H, and CAFE_H were used as comparison baselines, and the comment component was also evaluated. This refinement is performed because it is the last component in the reading sequence. Only methods using at least two news components are considered here, and SAFE is not feasible on these datasets because it requires both body text and headlines.

[0110] The results are as follows Figure 4 As shown, Ember outperforms other methods in generalization, meaning it can detect various types of fake news better than baseline methods. Furthermore, it's easy to see that the performance of the same algorithm varies significantly when detecting different categories of fake news, and the results on Compre demonstrate that Ember can comprehensively model all five types of fake news and outperforms other baselines.

[0111] 5. Hyperparameter λ sensitivity experiment

[0112] Experiments were conducted on four datasets, and the results are as follows: Figure 5 As shown. Figure 5 The study presents the λ values ​​and their corresponding F1 scores on four datasets. With different degrees of refinement, Ember demonstrates varying performance improvements in fake news detection (in most cases), particularly on PolitiFact2 and GossipCop, suggesting that refining the overall news representation can help learn high-quality news representations.

[0113] The present invention also discloses an electronic device, in one embodiment of which the electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the fake news detection method based on reader behavior simulation provided in the above embodiments.

[0114] like Figure 6 The diagram shown is a schematic representation of an electronic device implementing a fake news detection method based on reader behavior simulation, according to an embodiment of the present invention. The electronic device may include a processor 10, a memory 11, a communication bus 12, and a communication interface 13. It may also include a computer program stored in the memory 11 and executable on the processor 10, such as a fake news detection method program based on reader behavior simulation.

[0115] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the memory 11 (e.g., a fake news detection method based on reader behavior simulation) and calls data stored in the memory 11 to perform various functions of the electronic device and process data.

[0116] The memory 11 includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of an electronic device, such as a portable hard drive. In other embodiments, the memory 11 can be an external storage device of the electronic device, such as a plug-in portable hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc. Furthermore, the memory 11 can include both internal and external storage units of the electronic device. The memory 11 can be used not only to store application software and various types of data installed on the electronic device, such as the code of a fake news detection method program based on reader behavior simulation, but also to temporarily store data that has been output or will be output.

[0117] The communication bus 12 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into an address bus, a data bus, a control bus, etc. The bus is configured to enable communication between the memory 11 and at least one processor 10, etc.

[0118] Communication interface 13 is used for communication between the aforementioned electronic device and other devices, including a network interface and a user interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, Bluetooth interface, etc.), typically used to establish communication connections between the electronic device and other electronic devices. The user interface may be a display, an input unit (such as a keyboard), or, optionally, a standard wired or wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the electronic device and to display a visual user interface.

[0119] Figure 6 Only electronic devices with components are shown; it will be understood by those skilled in the art that... Figure 6The structure shown does not constitute a limitation on the electronic device and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0120] For example, although not shown, the electronic device may also include a power supply (such as a battery) to power various components. Preferably, the power supply can be logically connected to at least one processor 10 via a power management device, thereby enabling functions such as charging management, discharging management, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0121] It should be understood that the embodiments are for illustrative purposes only and are not limited to this structure in the scope of the patent application.

[0122] Furthermore, if the modules / units integrated in electronic device 1 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile. For example, a computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).

[0123] This invention also discloses a mobile robot equipped with the electronic device provided by this invention, which plans a three-dimensional path for the mobile robot. This electronic device is preferably, but not limited to, the central processor of the mobile robot, or a newly added processor.

[0124] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0125] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0126] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0127] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0128] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0129] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0130] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a system claim may also be implemented by a single unit or device through software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any specific order.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included within the present invention. No appended diagram markings in the claims should be considered as limiting the scope of the claims.

Claims

1. A method for detecting fake news based on reader behavior simulation, characterized in that, include: The process involves acquiring news articles to be detected and extracting their components, including four components: title, comments, body text, and images. The extracted components are input into the trained fake news detection model to obtain the probability of true or false news prediction. The fake news detection model includes: Two or more intra-component feature extractors, each corresponding to a component, are used to extract intra-component features for each component; inter-component feature extractors capture inter-component features between two components based on a parallel co-attention method, specifically including: Calculate the affinity matrix between the intra-group features of the first component and the intra-group features of the second component; both the first and second components are from the news to be detected and are different. Call the affinity matrix to obtain the updated representations of the within-group features of the first component and the updated representations of the within-group features of the second component; The attention score vector of the intra-group features of the first component is calculated using the updated representation of the intra-group features of the first component, and the attention score vector of the intra-group features of the second component is calculated using the updated representation of the intra-group features of the second component. The weighted representation of the first component relative to the second component is calculated using the attention score vector of the intra-group features of the first component. The weighted representation of the second component relative to the first component is calculated using the attention score vector of the intra-group features of the second component. The inter-component features between the first component and the second component are obtained by combining the weighted representation of the first component relative to the second component and the weighted representation of the second component relative to the first component. The inter-component feature serialization module includes one or more inter-component feature extractors and a sequence-based aggregator. The inter-component feature extractors capture inter-component features between any two components, and the sequence-based aggregator integrates all inter-component features into a feature sequence according to a set order. The sequence-based aggregator integrates all inter-component features into a feature sequence in the following order: inter-component features between title images, inter-component features between title comments, inter-component features between image comments, inter-component features between title text, inter-component features between image text, and inter-component features between comment text. The detector processes the feature sequence to obtain the probability of predicting whether the news to be detected is true or false.

2. The method for detecting fake news based on reader behavior simulation as described in claim 1, characterized in that, When the news to be detected includes the body text, the feature extractor within the component corresponding to the body text includes a first bidirectional GRU layer for the body text, a body text attention layer, and a second bidirectional GRU layer for the body text, connected in sequence.

3. The method for detecting fake news based on reader behavior simulation as described in claim 1 or 2, characterized in that, When the news to be detected includes a title, the feature extractor within the component corresponding to the title includes a first bidirectional LSTM layer for the title, a title attention layer, and a second bidirectional LSTM layer for the title, which are connected in sequence.

4. The method for detecting fake news based on reader behavior simulation as described in claim 3, characterized in that, When the news to be detected includes comments, the feature extractor within the component corresponding to the comments includes a comment attention layer and a comment bidirectional GRU layer connected in sequence.

5. The method for detecting fake news based on reader behavior simulation as described in claim 1, 2, or 4, characterized in that, When the news to be detected includes images, the feature extractor within the corresponding component for the images includes: The physical feature extraction module uses an error level analysis algorithm to process images to obtain ELA images; The image vectorization module vectorizes the original image in the news to be detected to obtain the original image vector, and vectorizes the ELA image to obtain the ELA image vector; The connection unit connects the original image vector and the ELA image vector to obtain the image component vector; The image attention layer identifies the importance of different parts of the image component vector and reassigns weights to each part to obtain a weighted image vector. A bidirectional GRU layer for images is used to encode weighted image vectors.

6. The method for detecting fake news based on reader behavior simulation as described in claim 5, characterized in that, The detector includes a first multilayer perceptron.

7. The method for detecting fake news based on reader behavior simulation as described in claim 5 or 6, characterized in that, During the training process of the fake news detection model, a news overall refinement loss module is also included. The news overall refinement loss module is executed as follows: For each training sample, the weighted representation of the last reading component relative to other components in the training sample is fused to obtain the overall refined representation of the news in the training sample; The news content of the training samples is refined and input into the second multilayer perceptron to obtain the second predicted probability of the training samples; The overall news detailing loss is calculated using the following formula: in, Indicates passage Parameters of the optimized fake news detection model. This indicates the number of samples in the training set. Indicates the training sample index. , Indicates the first The true labels of each training sample Indicates the first The second predicted probability of each training sample; And / or, the total loss function of the fake news detection model is: in, Represents the balance parameters. This indicates that the loss is obtained based on the detector output. Indicates the process Parameters of the optimized fake news detection model. , Indicates the first The first predicted probability is obtained by the detector for each training sample.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the fake news detection method based on reader behavior simulation as described in any one of claims 1 to 7.

Images