A multi-domain fake news detection method based on multi-perspective collaboration

Abstract

The present invention provides a multi-domain fake news detection method based on multi-perspective collaboration, aiming to solve the problems of existing fake news detection methods in domain transfer and improve the performance of the model in multi-domain scenarios. This method is based on multi-perspective collaboration and expert network feature extraction. By introducing a domain gate network to learn the relationship between domains and perspectives, it realizes the effective utilization of domain features. The multi-domain fake news detection method provided by the present invention can be applied to social media, news media, online question and answer and other fields, providing users with reliable and accurate information, and has broad application prospects.

Description

Technical Field

[0001] The present invention relates to the field of fake news detection, and in particular to a multi-field fake news detection method based on multi-perspective collaboration. Background Art

[0002] With the widespread use of the internet and social media, fake news has become a serious problem. It not only misleads people but also has a serious impact on society. Therefore, fake news detection has become a hot research topic. The goal of fake news detection is to classify news content into true and false categories. Existing methods can be mainly divided into content-based methods and social context-based methods.

[0003] In content-based approaches, researchers primarily analyze the textual content of news articles and extract features to detect fake news. These features include lexical, semantic, and statistical features. Some researchers also leverage external evidence, such as knowledge graphs or information from fact-checking websites, to detect fake news. The advantage of content-based approaches is that they can independently analyze news text, but the disadvantage is that they may overlook social context.

[0004] Methods based on social context primarily model the news dissemination process and exploit structural signals of news dissemination. These methods can capture social context by analyzing interactions between social media entities. On the other hand, some researchers have leveraged collective intelligence, such as sentiment and stance, to detect fake news.

[0005] Multi-domain fake news detection is an important branch of fake news detection. News from different domains has distinct characteristics, necessitating the use of different models to detect fake news. Multi-domain fake news detection methods aim to learn common features from data from different domains to improve the accuracy and generalization performance of fake news detection.

[0006] Emotion-based fake news detection is another approach. Research has shown that emotional features are crucial for fake news detection. Some researchers have used multi-task learning to improve fake news detection performance by leveraging multiple tasks, such as emotional features, novelty, and sentiment.

[0007] In summary, fake news detection is an important research direction that can be addressed through a variety of approaches, including content-based, social context-based, multi-domain, and sentiment-based approaches. These approaches can be used individually or in combination to improve the accuracy and generalization performance of fake news detection. Research on fake news detection will continue to evolve, and new technologies and methods will need to be continuously explored to address new challenges posed by fake news. Summary of the Invention

[0008] The purpose of the present invention is to address the shortcomings of the existing technology and provide a multi-domain fake news detection method based on multi-perspective collaboration to solve the problems raised by the above background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a multi-domain fake news detection method based on multi-perspective collaboration, which determines whether the news is fake news according to the following steps:

[0010] S1: Receive news content input, process the input news content through the BERT model, and obtain word embedding vectors;

[0011] S2: Process the word embedding vector through a bidirectional LSTM to extract the sequential features of the news;

[0012] S3: Use semantic networks and domain networks to process news content respectively to obtain semantic features and domain-specific features of news;

[0013] S4: Process news content through a hybrid expert system to obtain emotional and style features;

[0014] S5: Input semantic features, domain-specific features, sentiment features, and style features into the cross-view fusion module to achieve adaptive cross-view representation;

[0015] S6: According to the weights obtained by the domain network, the fused features are weighted summed to obtain the overall feature representation;

[0016] S7: Input the total feature representation into the classifier module to judge the authenticity of the news content;

[0017] S8: Output the news authenticity judgment result.

[0018] As a preferred technical solution of the present invention, the modeling of the fake news detection problem includes the following steps:

[0019] K1: Encode the text content of news P into a token sequence of length T using the BERT pre-trained model;

[0020] K2: Extract sentiment features E and style features S from news P, where both sentiment features E and style features S are numerical features;

[0021] K3: Take the domain label g of news P as input, combine it with the sentiment feature E and style feature S, and use multi-task learning to train a multi-domain fake news detection model;

[0022] K4: For news P, input its text tag sequence, sentiment feature E and style feature S, combined with its domain label g, use the trained multi-domain fake news detection model, and output its true or false label y;

[0023] K5: Repeat steps K3 and K4 for multiple domain labels to obtain the true or false label y under each domain. Finally, combine the true or false labels y under multiple domains to obtain the final true or false label of news P;

[0024] K6: For each domain label of news P, use a set of indicators such as confusion matrix, precision, recall, F1 score, etc. to evaluate the fake news detection performance under this domain.

[0025] As a preferred technical solution of the present invention, the specific extraction process of the multi-view collaboration includes the following steps:

[0026] (a): Set the hyperparameter T, which represents the number of experts in the expert network;

[0027] (b): Construct a hybrid expert network, including semantic network, sentiment network, style network and domain network;

[0028] (c): Convert the input news text into word vector W;

[0029] (d): For each expert network Do the following:

[0030] (d1): Determine the structure of each expert network model and the learnable parameters θ i ;

[0031] (d2): Using word vector W and learnable parameters θ i , Computational Expert Network The output representation r i ;

[0032] (e): According to the output representation of each expert network r i , obtain multi-view feature representation of input news text;

[0033] Each expert network They all have their own areas of expertise and are good at extracting the characteristics of a certain field.

[0034] As a preferred technical solution of the present invention, the specific process steps of cross-view fusion in S5 are:

[0035] S51: receiving input data of a plurality of views, wherein each view represents a specific data feature, including but not limited to semantics, sentiment, and style;

[0036] S52: Calculate the corresponding weight coefficient for each view, where w sem , w emo , w stlRepresent the weight coefficients of semantic, sentiment and style views respectively;

[0037] S53: Calculate the cross-view interaction representation z by multiplying the weight coefficients of different views with the corresponding view representations and summing them up, where the calculation formula is:

[0038]

[0039] where k sem , k emo , k stl Represents the number of experts in the semantic network, sentiment network and style network respectively, where lnr sem , lnr emo , lnr stl The view representations for semantic, sentiment and style views, w domain and lnr domain Representation of domain weights and domain local view representation;

[0040] S54: Set up multi-head cross-view fusion, each head adaptively learns a cross-view representation, and generates a set of cross-view representations Where H represents the number of cross-view representations;

[0041] S55: Based on the generated cross-view representation set Perform classification or regression tasks on the input data and output the results.

[0042] As a preferred technical solution of the present invention: the specific features of the classifier module in S7 are:

[0043] S71: Using different expert networks to obtain cross-view representations of news articles;

[0044] S71: Input the domain label into the domain gate to model the domain difference and obtain the weight score. The weight function is expressed as softmax(MLP(g));

[0045] S71: Aggregate cross-view representations based on the calculated weight scores. The formula is: w = softmax(MLP(g));

[0046] S71: The aggregated cross-view representation is fed into a multi-layer perceptron classifier with a softmax output layer for the binary classification problem of fake news.

[0047] S71: The network is trained using a binary cross entropy loss function. The loss function is expressed as y i represents the true label, Represents the predicted label.

[0048] As a preferred technical solution of the present invention: the specific steps of constructing the hybrid expert network in (b) are as follows:

[0049] (b1): Perform BERT tagging and LSTM sequence information extraction on the news text to obtain the semantic feature representation r of the news text sem ;

[0050] (b2): The five aspects of sentiment category, sentiment dictionary, sentiment intensity, sentiment score and auxiliary features are input into the sentiment network as sentiment features, and the hybrid expert network is used to extract the sentiment features to obtain the sentiment feature representation r of the news text. emo ;

[0051] (b3): Readability, logic, credibility, standardization, interactivity, fun, emotion, and integrity are input into the style network as style features, and the hybrid expert network is used to extract style features to obtain the style feature representation of the news text. stl ;

[0052] (b4): Use a specific domain feature extraction network for each domain, use TextCNN to extract domain features, and obtain domain-specific feature representations of news texts. domain ;

[0053] (b5): The three sets of features from different sources are represented as r sem 、r emo 、r stl and domain-specific feature representation r domain By combining them, we can obtain a multi-dimensional feature representation, which can fully describe the content and characteristics of the news text.

[0054] The multi-domain fake news detection method based on multi-perspective collaboration described in the present invention has the following technical effects compared with the existing technology by using the above technical solution:

[0055] The beneficial effects of the present invention are: the multi-domain fake news detection method based on multi-perspective collaboration proposed by the present invention can better utilize domain information to realize fake news correction in multi-domain scenarios. With the expert network entering the feature extraction, the domain portal network is introduced to learn the relationship between the domain and the perspective, and the multi-perspective fusion and BiLSTM module are used to effectively capture the multi-perspective feature representation of the news. The comparison of multiple virtual fake news detection methods and the experiment prove the effectiveness and excellence of this method. The beneficial effect of the present invention is to improve the accuracy and reliability of fake news detection, and it can be carried out in multiple fields such as news media and social networks in the hope of reducing the impact of false information on the public and society, and protecting information security and social stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a model architecture diagram of the multi-perspective collaborative multi-domain fake news detection method described in the present invention. DETAILED DESCRIPTION

[0057] The preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more precise definition of the protection scope of the present invention.

[0058] Example: See Figure 1 The present invention provides a technical solution: a multi-domain fake news detection method based on multi-perspective collaboration, comprising the following steps S1 to S8 to determine whether the news is fake news: S1: receiving news content input, processing the input news content through the BERT model to obtain a word embedding vector; S2: processing the word embedding vector through a bidirectional LSTM to extract the sequential features of the news; S3: using a semantic network and a domain network to process the news content respectively to obtain semantic features and domain-specific features of the news; S4: processing the news content through a hybrid expert system to obtain sentiment features and style features; S5: inputting the semantic features, domain-specific features, sentiment features, and style features into a cross-perspective fusion module for feature fusion; S6: performing a weighted summation of the fused features based on the weights obtained from the domain network to obtain an overall feature representation; S7: inputting the overall feature representation into a classifier module to perform a truth-or-fake judgment on the news content; S8: outputting the news truth-or-fake judgment result. The present invention conducts experiments on public datasets, including Chinese and English datasets.

[0059] The modeling of the fake news detection problem of this method includes the following steps: encoding the text content of news P into a token sequence of length T using the BERT pre-trained model; extracting sentiment features E and style features S from news P, where both sentiment features E and style features S are numerical features; taking the domain label g of news P as input, combining sentiment features E and style features S, and using a multi-perspective collaborative method to train a multi-domain fake news detection model; for news P, input its text token sequence, sentiment features E and style features S, combined with its domain label g, use the trained multi-domain fake news detection model to output its true or false label y; for multiple domain labels, repeat the steps to obtain the true or false label y under each domain, and finally combine the true or false labels y under multiple domains to obtain the final true or false label of news P; for each domain label of news P, use a set of confusion matrix, precision, recall rate, F1 score and other indicators to evaluate the fake news detection performance in that domain.

[0060] The specific extraction of multi-perspective collaboration in this method includes the following steps: setting the hyperparameter T to represent the number of experts in the expert network; constructing a hybrid expert network, including semantic network, sentiment network, style network and domain network; converting the input news text into a word vector W; for each expert network Perform the following operations: Determine the learnable parameters θ in each expert network i ; Using word vector W and learnable parameters θ i , Computational Expert Network The output representation r i ; According to the output of each expert network, r i , obtain the multi-view feature representation of the input news text; where each expert network They all have their own areas of expertise and are good at extracting the characteristics of a certain field.

[0061] This method constructs a hybrid expert network, which specifically includes the following steps: BERT tagging and LSTM sequence information extraction of news text to obtain the semantic feature representation of news text. sem The five aspects of sentiment category, sentiment dictionary, sentiment intensity, sentiment score and auxiliary features are input into the sentiment network as sentiment features, and the sentiment features are extracted using the hybrid expert network to obtain the sentiment feature representation of the news text. emo Eight aspects of readability, logic, credibility, standardization, interactivity, fun, emotion and integrity are input into the style network as style features, and the style features are extracted using the hybrid expert network to obtain the style feature representation of the news text. stl ; Use a specific domain feature extraction network for each domain, use TextCNN to extract domain features, and obtain domain-specific feature representations of news texts. domain ; The three sets of features from different sources are represented by r sem 、r emo 、r stl and domain-specific feature representation r domain By combining them, we can obtain a multi-dimensional feature representation, which can fully describe the content and characteristics of the news text.

[0062] The specific process steps of cross-view fusion of this method are: receiving input data from multiple views, where each view represents a specific data feature, including but not limited to semantics, emotion, and style; calculating the corresponding weight coefficient for each view, where w sem , w emo , w stl Represent the weight coefficients of semantic, sentiment and style views respectively; calculate the cross-view interaction representation z, which is obtained by multiplying the weight coefficients of different views with the corresponding view representation and summing them up, where the calculation formula is: Set up multi-head cross-view fusion, each head adaptively learns a cross-view representation, and generates a set of cross-view representations Where H represents the number of cross-view representations; according to the generated cross-view representation set Perform classification or regression tasks on the input data and output the results.

[0063] The specific features of the classifier module of this method are as follows: different expert networks are used to obtain cross-view representations of news articles; domain labels are input into domain gates to model domain differences and obtain weight scores, with the weight function represented as softmax(MLP(g)); cross-view representations are aggregated based on the calculated weight scores, as follows: The aggregated cross-view representation is input into a multi-layer perceptron classifier with a softmax output layer for the binary classification problem of fake news. The network is trained using the binary cross entropy loss function, which is expressed as y i represents the true label, Represents the predicted label.

[0064] Through the above process, a multi-domain fake news detection method based on multi-perspective collaboration can effectively detect fake news in multiple fields, improving accuracy and robustness. This method fully utilizes the various perspective characteristics of news text, such as semantics, sentiment, and style, to help improve the accuracy of news authenticity judgment. Furthermore, multi-perspective feature fusion can further improve detection performance.

[0065] Example: This invention uses the commonly used evaluation metrics AUC, F1, and accuracy to assess the performance of fake news detection models. AUC is the area under the visual operating characteristic curve (AUC), which is plotted in two dimensions with the false positive rate as the x-coordinate and the true positive rate as the y-coordinate. AUC is widely used to evaluate the performance of different models because it is unaffected by class imbalance and independent of the prediction threshold. F1 combines precision and recall, with accuracy indicating the proportion of correct classifications.

[0066] The baselines can be divided into three categories. The first category is single-domain detection methods, which build models and train them for each domain. TextCNN_s: This is a commonly used text classification model. We implement TextCNN with 5 kernels. The 5 kernels with the same 64 channels have different step sizes of 1, 2, 3, 5, and 10. BiGRU_s: The news text is modeled using BiGRU to obtain the sequential information of the text, and the prediction results are obtained using MLP. The hidden layer size of GRU is set to 300. BERT_s: We use BERT to encode the tokens of the news text and embed the extracted average value into the MLP to obtain the final prediction results.

[0067] The second category is mixed-domain baselines, which mix data from all domains for training. The implementations of BiGRU_a, TextCNN_a, and BERT_a in this group are the same as those in the first group. The other two baselines in this category are: StyleLSTM: This first uses a BiLSTM to extract news representations from the content. Then, the news representations and style features are combined and fed into an MLP to produce the final prediction results. DualEmo: This uses a BiGRU to extract feature representations for news and extracts sentiment features from news text and comments, fusing the two to achieve fake news detection.

[0068] The third category is multi-domain methods. MMoE is a multi-task model that treats each domain as a separate task by sharing a Mixture of Experts (MoE) across domains, with each domain having its own specific head. MoSE adds an LSTM before the MMoE's experts to capture sequential information from the text. Mdfend is a multi-domain fake news detection model that utilizes domains to select useful experts for the MoE.

[0069] Experimental analysis: In order to verify the effectiveness of the model proposed in this paper, we conducted comparative experiments on the model and the baseline method on Chinese and English datasets, with the comparison indicators being F1, Acc and AUC. The experimental results are shown in Tables 1 and 2.

[0070] Table 1: Performance comparison of model methods on Chinese dataset

[0071]

[0072]

[0073] Table 2: Performance comparison of model methods on English dataset

[0074]

[0075] The various indicators of the model proposed in this paper and its variants on the Chinese dataset are shown in Table 3. In order to verify the importance of multi-view fusion and BiLSTM modules to the proposed model, these two modules are analyzed separately.

[0076] 1) The impact of multiple perspectives. First, we experimentally validated the contribution of different views and compared the proposed model with its variants MMFND-sem, MMFND-emo, and MMFND-stl, which respectively remove the semantic view, sentiment view, and style view from MMFND. We found that all views are beneficial for fake news detection, especially the semantic view, which is the core of most existing methods. Since sentiment and style features are manually extracted from text content, these features are usually used as auxiliary information for semantic view modeling. In addition, we observed that the sentiment view is more effective than the style view. The reason may be that sentiment features contain publisher and social features, while style features only represent publisher preferences.

[0077] 2) The Impact of BiLSTM. To model the sequential information of news text, a BiLSTM module was added before the semantic network. To evaluate the effectiveness of the proposed sequential modeling module, a variant of the model, MMFND-BiLSTM, was designed. This variant does not include the BiLSTM module and simply inputs text features directly into the semantic network, ignoring the sequential connections in the text. Experimental verification shows that MMFND performs better than its variant, MMFND-BiLSTM. This indicates that BiLSTM can effectively model the sequential information in news text, helping to improve detection performance.

[0078] Table 3: Ablation experiments

[0079]

[0080] Experimental Conclusions: It can be seen that our model, MMFND, outperforms the comparison models on multiple metrics. It achieves very good performance on both Chinese and English datasets. On most tasks, the mixed-domain model outperforms the single-domain model, indicating that combining multi-domain training data not only helps improve overall performance across multiple domains, but also improves performance within a single domain. StyleLSTM and DualEmo also outperform BiGRU in detection. Since these two models combine text features extracted by BiGRU with style and sentiment features, this suggests that introducing more views is beneficial for multi-domain fake news detection. We found that MMFND significantly outperforms the comparison models on most tasks, demonstrating that MMDFND not only improves overall detection performance but also performance in specific domains. This is primarily due to the fact that MMFND enriches domain information and explicitly models various domain differences by aggregating useful cross-view interactions across different domains.

[0081] The above embodiments merely illustrate several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. A multi-domain fake news detection method based on multi-perspective collaboration, characterized by: It uses a multi-perspective collaborative approach to combine domain information to solve the problem of multi-domain fake news detection. Specifically, it includes the following steps to determine whether the news is fake news: S1: Receive news content P, process the input news content through the BERT model, and obtain word embedding vectors; S2: Process the word embedding vector through a bidirectional LSTM to extract the sequential features of the news; S3: Use semantic networks and domain networks to process news content respectively to obtain semantic features and domain-specific features of news; S4: Process the news content P through a hybrid expert system to obtain the emotional features E and style features S; S5: Input semantic features, domain-specific features, sentiment features, and style features into the cross-view fusion module to achieve adaptive cross-view representation; The specific operation steps of the cross-view fusion module in step S5 include: S51: receiving input data of a plurality of views, wherein each view represents a specific data feature, including but not limited to semantics, sentiment, and style; S52: Calculate the corresponding weight coefficient for each view, where w sem , w emo , w stl Represent the weight coefficients of semantic, sentiment and style views respectively; S53: Calculate the cross-view interaction representation z by multiplying the weight coefficients of different views with the corresponding view representations and summing them up, where the calculation formula is: where k sem , k emo , k stl Represents the number of experts in the semantic network, sentiment network and style network respectively, where lnr sem , lnr emo , lnr stl The view representations for semantic, sentiment and style views, w domain and lnr domain Representation of domain weights and domain local view representation; S54: Set up multi-head cross-view fusion, each head adaptively learns a cross-view representation, and generates a set of cross-view representations Where H represents the number of cross-view representations; S55: Based on the generated cross-view representation set Perform classification or regression tasks on the input data and output the results; S6: According to the weights obtained by the domain network, the fused features are weighted summed to obtain the overall feature representation; S7: Input the total feature representation into the classifier module to judge the authenticity of the news content; S8: Output the result of judging the authenticity of the news.

2. The multi-domain fake news detection method based on multi-perspective collaboration according to claim 1 is characterized by: The modeling of the multi-domain fake news detection problem includes the following steps: K1: Encode the text content of news content P into a text token sequence of length T using the BERT pre-trained model; K2: Extract emotional features E and style features S from news content P, where both emotional features E and style features S are numerical features; K3: Take the domain label g of news P as input, combine the sentiment feature E and style feature S, and use multi-task learning to train a multi-domain fake news detection model, where the domain label g∈{Domian1,…,Domian N },Domian represents the specific value of the domain; K4: For news P, input its text tag sequence, sentiment feature E and style feature S, combined with its domain label g, use the trained multi-domain fake news detection model, and output its true or false label y; K5: Repeat steps K3 and K4 for multiple domain labels to obtain the true or false label y under each domain. Finally, combine the true or false labels y under multiple domains to obtain the final true or false label of news P; K6: For each domain label g of news P, a set of evaluation metrics, including confusion matrix, precision, recall, and F1 score, are used to evaluate the fake news detection performance in that domain.

3. The multi-domain fake news detection method based on multi-perspective collaboration according to claim 1 is characterized by: The multi-perspective collaborative method: Its specific extraction process includes the following steps: (a): Set the hyperparameter T, which represents the number of experts in the expert network; (b): Construct a hybrid expert network, including semantic network, sentiment network, style network and domain network; (c): Convert the text content of the input news content P into a word vector W; (d): For each expert network Do the following: (d1): Determine the model structure of each expert network and the learnable parameters θ i ; (d2): Using word vector W and learnable parameters θ i , Computational Expert Network The output representation r i ; (e): According to the output representation of each expert network r i , obtain multi-view feature representation of input news text; Each expert network Used to extract features from different fields.

4. The multi-domain fake news detection method based on multi-perspective collaboration according to claim 1 is characterized by: The specific workflow of the classifier module in S7 is as follows: S71: Using different expert networks to obtain cross-view representations of news articles; S72: Input the domain label into the domain gate to model the domain difference and obtain the weight score. The weight function is expressed as softmax(MLP(g)); S73: Aggregate cross-view representations based on the calculated weight scores. The formula is: S74: Input the aggregated cross-view representation into a multi-layer perceptron classifier with a softmax output layer to perform binary classification of fake news; S75: The network is trained using a binary cross entropy loss function. The loss function is expressed as y i represents the true label, Represents the predicted label.

5. The multi-domain fake news detection method based on multi-perspective collaboration according to claim 3 is characterized by: The specific steps of constructing the hybrid expert network in (b) are as follows: (b1): Perform BERT tagging and LSTM sequence information extraction on the news text to obtain the semantic feature representation r of the news text sem ; (b2): Input the emotional features into the emotional network. The emotional features include emotional category, emotional dictionary, emotional intensity, emotional score and auxiliary features. Use the hybrid expert network to extract the emotional features and obtain the emotional feature representation of the news text. emo ; (b3): Input the style features into the style network. The style features include readability, logic, credibility, standardization, interactivity, fun, emotion and integrity. Use the hybrid expert network to extract the style features and obtain the style feature representation r of the news text. stl ; (b4): Use a specific domain feature extraction network for each domain, use TextCNN to extract domain features, and obtain domain-specific feature representations of news texts. domain ; (b5): The three sets of features from different sources are represented as r sem 、r emo 、r stl and domain-specific feature representation r domain By combining them, a multidimensional feature representation is obtained, which is used to comprehensively describe the content and characteristics of the news text.

Images