A method for detecting fake news based on self-supervised learning and consistency with propagation.
By optimizing the fake news detection method through self-supervised learning and propagation consistency features, this method solves the problem of fake news detection in data-scarce scenarios and achieves efficient and accurate fake news detection, especially demonstrating superior performance in the identification of new and variant fake news.
Patent Information
- Application Number
- CN202410738920.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-07
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-06-07
AI Technical Summary
Existing fake news detection methods are difficult to generalize effectively in data-scarce scenarios, consume large amounts of computational resources, struggle to capture complex social communication environment information, and do not fully utilize the theory of communication consistency, resulting in insufficient detection accuracy and robustness.
We employ self-supervised learning to optimize sentence embedding representations, combine adversarial generative networks and news dissemination consistency features, construct positive and negative contrast sample pairs, adversarial generative networks and news dissemination subgraphs, integrate neighbor node features, and utilize a pre-trained language model to detect fake news.
It significantly improves the accuracy and robustness of fake news detection in data-scarce scenarios, enhances the model's ability to identify fake news, and demonstrates superior performance, especially in the detection of new and variant fake news.
Smart Images

Figure CN118673404B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of fake news detection technology in data-scarce scenarios, specifically involving a fake news detection method based on self-supervised learning and consistency of propagation. Background Technology
[0002] With the rapid popularization of social media, fake news has proliferated online, seriously jeopardizing the healthy development of mainstream social consciousness. Fake news detection, as a solution that utilizes machine learning to automatically identify fake news, is attracting increasing attention from academia and researchers. Currently, mainstream fake news detection methods rely on deep learning, primarily involving a fully supervised fine-tuning paradigm based on pre-trained language models (PLMs). These methods depend on large amounts of labeled data. However, in many real-world scenarios, obtaining high-quality labeled datasets is time-consuming, requires expert knowledge, and is costly, thus limiting the competitiveness of automated fake news detection systems. Therefore, developing and enhancing fake news detection methods for data-scarce scenarios has become increasingly important.
[0003] Despite the numerous advantages of current mainstream fake news detection methods, they still face some difficulties and challenges in practical applications:
[0004] 1. High difficulty in obtaining labeled data: Current mainstream fake news detection methods rely on large-scale labeled datasets. However, obtaining these high-quality labeled corpora is time-consuming, labor-intensive, and requires professional knowledge, making it difficult to obtain sufficient training data in many real-world scenarios.
[0005] 2. Model generalization ability: Traditional fully supervised learning methods rely entirely on labeled data during training, making it difficult for the model to generalize well when faced with scarce labels. Especially in new events or domains, the model's performance may drop significantly.
[0006] 3. High computational resource consumption: Deep learning-based fake news detection methods typically require a large amount of computational resources for training and inference, especially when dealing with large-scale social media data, where the consumption of computational resources becomes a bottleneck.
[0007] 4. Complexity of the social environment: The environment in which fake news spreads is complex and ever-changing, involving multiple media and platforms. Traditional methods struggle to effectively capture and utilize this complex social context information, thus affecting the accuracy of detection.
[0008] 5. Insufficient utilization of the consistency of communication theory: In the process of spreading fake news, there is often a phenomenon of "consistency of news truthfulness," that is, news circulated within the same social group usually has a consistent degree of truthfulness. Traditional methods often ignore this social communication law and fail to make full use of this information to improve detection performance. Summary of the Invention
[0009] The purpose of this invention is to provide a fake news detection method based on self-supervised learning and propagation consistency, aiming to improve the accuracy and robustness of fake news detection, especially in scenarios with scarce data. This method optimizes sentence embedding representations through self-supervised contrastive learning, combines adversarial embedding generation techniques from generative adversarial networks, and utilizes news propagation consistency features for authenticity verification, significantly improving the performance of fake news detection.
[0010] To achieve the above objectives, the present invention adopts the following technical solution:
[0011] The method for detecting fake news based on self-supervised learning and propagation consistency described in this invention mainly includes the following steps:
[0012] (1) Sentence Representation Optimization Based on Self-Supervised Contrastive Learning: First, a pre-trained language model (PLM) is used as the detection model for this method. Then, a self-supervised contrastive learning method is used to optimize the sentence embedding representation generated by the pre-trained language model. By constructing positive and negative contrastive sample pairs, the model's ability to understand and distinguish sentence semantics is enhanced. The specific operations are as follows:
[0013] 1. Construct positive contrast sample pairs, selecting the representation of the same news text in different instances;
[0014] 2. Construct negative contrast sample pairs and select representations of different news texts;
[0015] 3. The sentence embedding representation is optimized by contrastive learning loss function to maximize the similarity of positive contrastive sample pairs and minimize the similarity of negative contrastive sample pairs.
[0016] (2) Sample-level adversarial learning based on adversarial generative networks: Construct a generative adversarial network (GAN), use random noise and negative fake news samples as input, generate rich adversarial embedding representations through a multilayer perceptron (MLP) and an independent pre-trained language model encoder, and combine the adversarial embeddings to further optimize the output embedding representation of the detection model in step one.
[0017] The specific steps are as follows:
[0018] 1. Negative sequence and language encoder-driven engine: Input random word sequences, negative fake news samples, and semantically rewritten news text to generate adversarial embedding representations;
[0019] 2. Noise and Multilayer Perceptron Driven Engine: Input random noise and generate text perturbation embedding representation;
[0020] 3. Binary Classification and Contrast Learning Discriminator: The embedding of the classification input represents whether it is real or generated, so as to enhance the model's ability to identify fake samples.
[0021] (3) Authenticity Verification Based on News Dissemination Consistency Features: A news dissemination consistency network is constructed. Based on the adjacency subgraph feature aggregation algorithm, the dissemination features of news nodes are integrated, and the authenticity features of neighboring news nodes are injected into the current news authenticity prediction process to further improve the accuracy of fake news detection. The specific operation is as follows:
[0022] 1. Construct a news dissemination subgraph: Based on news dissemination records, calculate and identify neighboring news nodes that have a direct or indirect dissemination relationship with the current news;
[0023] 2. Calculate the credibility of neighboring news nodes: Calculate the credibility of each neighboring news node based on the frequency of news dissemination and the propagation weight between nodes;
[0024] 3. Optimize the prediction results of the current news by combining the prediction results of the authenticity of neighboring news nodes.
[0025] (4) Optimization of the pre-trained language model based on pseudo-prompt learning: Based on the detection model optimized in steps one and two, a masked language model training task (MLM) based on pseudo-prompt learning is adopted in the optimization process of the above model. By constructing task-oriented prompt templates, the fake news detection task is transformed into a text fill-in-the-blank question to make full use of the linguistic knowledge of the pre-trained model. Combined with the post-processing optimization process in step three, the performance of the model in scenarios with few samples is improved. The specific operation is as follows:
[0026] 1. Develop a prompt template to transform the fake news detection task into a text fill-in-the-blank question;
[0027] 2. Use a pre-trained language model to process the prompt template and generate the word probability distribution at the [MASK] position;
[0028] 3. Based on a predefined set of tags, select the most suitable tags from the word probability distribution to complete the classification and prediction of fake news.
[0029] Through the above process, the fake news detection method based on self-supervised learning and propagation consistency proposed in this invention can achieve efficient and accurate fake news detection in environments with limited sample data.
[0030] Preferably, step one of the method execution flow, the process of optimizing sentence representation based on self-supervised contrastive learning, employs the SimCSE framework. This self-supervised contrastive learning optimizes sentence embedding representation by constructing positive and negative contrastive sample pairs. Positive contrastive sample pairs are obtained by encoding the same news text multiple times, while negative contrastive sample pairs are selected from different news texts. The optimization objective is to maximize the similarity of positive contrastive sample pairs while minimizing the similarity of negative contrastive sample pairs. The formula used in this self-supervised contrastive learning is as follows:
[0031]
[0032] Among them, L contrast The optimization objective of this contrastive learning is represented by the loss function; Sim(zi,zj) represents the cosine similarity between two embedding vectors, τ is the temperature coefficient; k represents the index of the sample to be compared; N represents N pairs of comparison samples; the function log(.) represents the logarithmic function; exp(.) represents the exponential function.
[0033] Preferably, step two in the method execution flow: the process of sample-level adversarial learning based on adversarial generative networks, specifically includes:
[0034] First, a Generative Adversarial Network (GAN) was constructed, consisting of two generators and one discriminator. The generators are responsible for generating adversarial examples, while the discriminator distinguishes between real and generated examples. This approach enhances the model's ability to identify fake examples.
[0035] The first generator is specifically designed as follows: It is a negative sequence and language encoder-driven engine that takes random word sequences, negative fake news samples, and semantically rewritten news text as input to generate adversarial embeddings. The second generator is specifically designed as a noise and multilayer perceptron-driven engine that takes random noise as input to generate text perturbation embeddings. The discriminator, acting as a binary classification contrastive learning discriminator, is responsible for classifying the input embeddings and determining whether they are real or generated samples.
[0036] Subsequently, the objective function of adversarial learning can be specifically equated to: the optimization objective of adversarial learning is equivalent to defining the optimization objectives of the generator and the discriminator. The generator's goal is to generate adversarial examples that are as realistic as possible, so that the discriminator cannot distinguish between real and fake examples; the discriminator's goal is to accurately distinguish between real and generated examples.
[0037] Finally, the specific adversarial learning objective function is defined as follows:
[0038]
[0039] in, is the objective function of GAN, where D represents the discriminator, G represents the generator, pdata represents the distribution of the real data, and pz(z) represents the distribution of the noise. It is the discriminator's expected value for the real data x; It is the discriminator's expectation of the generated data; the sample G(z) generated by the generator G is input into the discriminator D for evaluation.
[0040] In general, the discriminator D aims to maximize the sum of the two expected values in the formula to make the most accurate classification between real and generated data. The generator G aims to minimize this objective function so that the generated data can fool the discriminator, making the output of D(G(z)) close to 1.
[0041] Based on the adversarial generative network architecture and optimization objectives described above, the sample-level adversarial learning process of this method model based on adversarial generative networks is realized.
[0042] Preferably, step three in the method execution flow—authentication based on news dissemination consistency features—aims to construct a news dissemination consistency network. Through an adjacency subgraph feature aggregation algorithm, it integrates the dissemination features of news nodes, injecting the authenticity features of neighboring news nodes into the current news authenticity prediction process to further improve the accuracy of fake news detection. The specific operation of this authenticity verification based on news dissemination consistency features is as follows:
[0043] First, a news dissemination subgraph is constructed. Based on news dissemination records, neighboring news nodes with direct or indirect dissemination relationships to the current news are calculated and identified. By analyzing the dissemination path of news on social media platforms, a dissemination subgraph is generated, which includes the current news, news nodes related to its dissemination, and their connection relationships.
[0044] Furthermore, the credibility of neighboring news nodes is calculated based on the frequency of news dissemination and the propagation weight between nodes. The frequency of dissemination reflects the activity of news within the network, while the propagation weight reflects the trust and influence between news nodes. By combining these two factors, the credibility of each news node can be quantified.
[0045] The preferred formula for calculating credibility is as follows:
[0046]
[0047] Where Credibility(Ni) represents the credibility of node Ni; Frequency(Ni) represents the propagation frequency of node Ni; Weight(Ni,Nj) represents the propagation weight between node Ni and its neighboring node Nj; Neighbors(Ni) represents the set of neighboring nodes that have a direct propagation relationship with node Ni; α and β are parameters used to balance frequency and propagation weight.
[0048] Finally, the prediction results of the current news item are optimized by aggregating the prediction results of neighboring news nodes. The specific steps are as follows:
[0049] The propagation characteristics of each neighboring news node are summarized using an adjacency subgraph feature aggregation algorithm.
[0050] 2. Using a propagation consistency network, the features of neighboring news nodes are fused to generate a comprehensive propagation consistency feature vector.
[0051] 3. Integrate the consistency feature vector of the dissemination with the feature vector of the current news, and use a machine learning model to make the final prediction of its authenticity.
[0052] Preferably, the integration formula between feature vectors is as follows:
[0053] Final_Prediction=γ·Current_News_Feature+δ·Propagation_Consistency_Feature
[0054] Where Current_News_Feature represents the feature vector of the current news, Propagation_Consistency_Feature represents the propagation consistency feature vector, and γ and δ are parameters used to balance the current news feature and the propagation consistency feature.
[0055] This authenticity verification step, based on the consistency characteristics of news dissemination, makes full use of the dissemination characteristics of news in social networks. Through the credibility analysis of dissemination paths and neighboring nodes, it enhances the accuracy and robustness of fake news detection.
[0056] Preferably, step four in the method execution process is: optimization of the pre-trained language model based on pseudo-cue learning. By using a masked language model (MLM), a task-oriented cue template is constructed, converting the fake news detection task into a text fill-in-the-blank question. The cue template is processed using the pre-trained language model to generate the word probability distribution at the [MASK] position. Based on a predefined set of tag words, the most suitable tag word is selected from the word probability distribution to complete the classification and prediction of fake news.
[0057] The specific steps are as follows:
[0058] Construct task-oriented prompt templates: Transform the fake news detection task into a text fill-in-the-blank question. By designing appropriate prompt templates, embed the news text to be detected into the template and insert [MASK] tags at key positions. For example, the following template can be designed for the news text to be detected, "News X":
[0059] The veracity of this news item is: [MASK]. News content: News X.
[0060] Using a pre-trained language model to process cue templates: The constructed cue template is input into a pre-trained language model (such as BERT or GPT-3) to generate the word probability distribution at the [MASK] position. The pre-trained language model predicts the word at the [MASK] position based on contextual information and outputs the probability of each word.
[0061] The specific formulas for the above steps are shown below:
[0062] P(word|context) = MLM([MASK]|prompt template)
[0063] Selecting the most suitable tag: Based on a predefined set of tag words, select the most suitable tag words from the lexical probability distribution to complete the classification prediction of fake news. Assuming the predefined set of tag words is {"true", "false"}, select the word with the highest probability as the final classification result according to the probability distribution. The specific steps for this probability tag word selection are as follows:
[0064] First, the prompt template is input into the pre-trained language model to obtain the word probability distribution P(word|context) at the [MASK] position;
[0065] Secondly, extract the probabilities P(true|context) and P(false|context) of each word in the predefined set of tag words from the probability distribution;
[0066] Finally, the word with the highest probability is selected as the final classification result.
[0067] The corresponding specific formula is shown below:
[0068]
[0069] Comprehensive evaluation and model tuning: Through multiple rounds of training and evaluation, the prompt templates and pre-trained models are tuned to improve the model's classification accuracy in low-sample scenarios. This specifically includes:
[0070] First, the structure and content of the prompt template were adjusted to better guide the model in understanding the fake news detection task;
[0071] Secondly, experiments were conducted using different pre-trained language models to compare their performance and select the optimal model.
[0072] Finally, by combining cross-validation and model evaluation metrics (such as accuracy, recall, F1 score, etc.), the model parameters are optimized to ensure that efficient and accurate detection capabilities can still be maintained even with a small number of samples.
[0073] Through the detailed steps described above, this invention realizes a pre-trained language model optimization method based on pseudo-prompt learning, which effectively improves the accuracy and robustness of fake news detection, especially in scenarios with scarce data, it can still maintain efficient and accurate detection performance.
[0074] The present invention has at least the following beneficial technical effects:
[0075] 1. Self-supervised contrastive learning optimizes embedding representations: By introducing a self-supervised contrastive learning method, the sentence embedding representations generated by the pre-trained language model are optimized, solving the problem of insufficient model generalization ability in data-scarce scenarios. Self-supervised contrastive learning utilizes unlabeled data for training, improving the model's adaptability to different domains and new events, and significantly reducing the dependence on large amounts of labeled data;
[0076] 2. Enhanced Robustness of the Model through Generative Adversarial Networks: A Generative Adversarial Network (GAN) consisting of two generators and one discriminator was constructed. By generating adversarial examples and distinguishing between real and generated examples, the model's ability to identify fake news samples was enhanced. The GAN improves the model's detection accuracy in complex and variable environments, particularly its ability to identify new and variant forms of fake news.
[0077] 3. Improved Detection Accuracy Through Propagation Consistency Features: Based on the theory of news propagation consistency, a news propagation consistency network is constructed, and the propagation features of news nodes are integrated using an adjacency subgraph feature aggregation algorithm. By introducing the credibility analysis of news propagation paths and neighboring nodes, the model's accuracy in detecting fake news is enhanced. The introduction of propagation consistency features enables the model to effectively utilize propagation information in social networks, improving overall detection performance.
[0078] 4. Pseudo-prompt learning improves performance in low-sample scenarios: A masked language model (MLM) based on pseudo-prompt learning is employed. By constructing task-oriented prompt templates, the fake news detection task is transformed into a text completion problem. Processing the prompt templates using a pre-trained language model significantly improves detection performance in low-sample scenarios. The pseudo-prompt learning method fully utilizes the linguistic knowledge of the pre-trained model, enhancing the model's classification ability under data-scarce conditions.
[0079] 5. Superior Overall Detection Performance: This invention comprehensively utilizes self-supervised learning, generative adversarial networks, and propagation consistency features in the task of fake news detection, providing an efficient, accurate, and robust solution. Regardless of whether data is abundant or scarce, this method demonstrates superior detection performance, providing strong technical support for fake news detection.
[0080] Of course, product instances implementing this invention in any scenario do not necessarily need to achieve all the advantages and benefits described above at the same time. Attached Figure Description
[0081] Figure 1 This is a flowchart of a method for detecting fake news based on self-supervised learning and consistency of propagation, according to the present invention. Detailed Implementation
[0082] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the embodiments described below are merely some embodiments of the present invention, and not embodiments compatible with all application scenarios. Referring to the embodiments demonstrated in this invention, all other embodiments obtained by those skilled in the art without any inventive step are also within the protection scope of this invention.
[0083] This invention provides a method for detecting fake news based on self-supervised learning and consistency of propagation, comprising the following steps:
[0084] First, refer to the appendix Figure 1 Step one in the process, 'Sentence Representation Optimization Based on Self-Supervised Comparative Learning,' is shown in the following implementation flow:
[0085] First, a pre-trained language model (PLM) is used as the detection model for this method. Then, a self-supervised contrastive learning method is employed to optimize the sentence embedding representation generated by the pre-trained language model, as follows:
[0086] 1. Constructing Contrast Sample Pairs: By introducing the SimCSE framework, positive and negative contrast sample pairs are constructed. Positive contrast sample pairs select representations of the same news text in different instances, while negative contrast sample pairs are selected from different news texts.
[0087] 3. Optimization Objective: To optimize the sentence embedding representation through a contrastive learning loss function, maximizing the similarity of positive contrastive sample pairs and minimizing the similarity of negative contrastive sample pairs. The specific formula is as follows:
[0088]
[0089] Among them, L contrast The optimization objective of this contrastive learning is represented by the loss function; Sim(zi,zj) represents the cosine similarity between two embedding vectors, τ is the temperature coefficient; k represents the index of the sample to be compared; N represents N pairs of comparison samples; the function log(.) represents the logarithmic function; exp(.) represents the exponential function.
[0090] Further, refer to the appendix. Figure 1 Step two in the process, sample-level adversarial learning based on adversarial generative networks, is shown in the following specific implementation flow:
[0091] A Generative Adversarial Network (GAN) is constructed to generate adversarial examples and distinguish between real and generated examples, thereby further enhancing the detection model's ability to identify fake news. The specific steps are as follows:
[0092] First, we define Generator 1: a negative sequence and language encoder-driven engine. It takes a random word sequence, negative fake news samples, and semantically rewritten news text as input to generate adversarial embedding representations.
[0093] Secondly, we define Generator 2: a noise and multilayer perceptron-driven engine. It takes random noise as input and generates a text perturbation embedding representation.
[0094] Next, we define the discriminator: a binary classification contrastive learning discriminator. The embedding of the classification input represents whether it is a real or generated sample.
[0095] Finally, the objective function for adversarial learning is defined as follows:
[0096]
[0097] in, is the objective function of GAN, where D represents the discriminator, G represents the generator, pdata represents the distribution of the real data, and pz(z) represents the distribution of the noise. It is the discriminator's expected value for the real data x; It is the discriminator's expectation of the generated data; the sample G(z) generated by the generator G is input into the discriminator D for evaluation.
[0098] Further, refer to the appendix. Figure 1 Step three, authenticity verification based on news dissemination consistency features, aims to construct a news dissemination consistency network. Through an adjacency subgraph feature aggregation algorithm, it integrates the dissemination features of news nodes, injecting the authenticity features of neighboring news nodes into the current news authenticity prediction process. The specific implementation process is as follows:
[0099] First, a news dissemination subgraph is constructed: based on news dissemination records, neighboring news nodes with direct or indirect dissemination relationships to the current news are calculated and identified. By analyzing the dissemination path of news on social media platforms, a dissemination subgraph is generated, which includes the current news, news nodes related to its dissemination, and their connection relationships.
[0100] Secondly, the credibility of neighboring news nodes is calculated using the following formula:
[0101]
[0102] Where Credibility(Ni) represents the credibility of node Ni; Frequency(Ni) represents the propagation frequency of node Ni; Weight(Ni,Nj) represents the propagation weight between node Ni and its neighboring node Nj; Neighbors(Ni) represents the set of neighboring nodes that have a direct propagation relationship with node Ni; α and β are parameters used to balance frequency and propagation weight.
[0103] Finally, the true / false prediction results of neighboring news nodes are combined: The propagation features of each neighboring news node are aggregated using an adjacency subgraph feature aggregation algorithm. A propagation consistency network is then used to fuse the features of neighboring news nodes, generating a comprehensive propagation consistency feature vector. This propagation consistency feature vector is integrated with the feature vector of the current news article, and a machine learning model is used for the final true / false prediction. The specific integration formula is as follows:
[0104] Final_Prediction=γ·Current_News_Feature+δ·Propagation_Consistency_Feature
[0105] Where Current_News_Feature represents the feature vector of the current news, Propagation_Consistency_Feature represents the propagation consistency feature vector, and γ and δ are parameters used to balance the current news feature and the propagation consistency feature.
[0106] Further, refer to the appendix. Figure 1 Step four, "Pre-trained Language Model Tuning Based on Pseudo-Prompt Learning," aims to optimize the detection model based on the models optimized in steps one and two. During this tuning process, a Masked Language Model (MLM) based on pseudo-prompt learning is employed. By constructing task-oriented prompt templates, the fake news detection task is transformed into a text completion problem. Combined with the post-processing optimization process from step three, this improves the model's performance in scenarios with few samples of fake news. The specific implementation process is as follows:
[0107] First, construct a task-oriented prompt template: embed the news text to be detected into the prompt template and insert [MASK] tags at key positions. For example, "The veracity of this news is: [MASK]. News content: News X."
[0108] Secondly, a pre-trained language model is used to process the cue template: the constructed cue template is input into a pre-trained language model (such as BERT or GPT-3) to generate the word probability distribution at the [MASK] position. The pre-trained language model predicts the word at the [MASK] position based on contextual information and outputs the probability of each word.
[0109] The specific formulas for the above steps are shown below:
[0110] P(word|context) = MLM([MASK]|prompt template)
[0111] Next, the most suitable tag words are selected: based on the predefined tag word set, the most suitable tag words are selected from the lexical probability distribution to complete the classification prediction of fake news. Assuming the predefined tag word set is {"true", "false"}, the word with the highest probability is selected as the final classification result according to the probability distribution. The specific steps for this probability tag word selection are as follows:
[0112] 1. Input the prompt template into the pre-trained language model to obtain the word probability distribution P(word|context) at the [MASK] position;
[0113] 2. Extract the probabilities P(true|context) and P(false|context) of each word in the predefined set of tag words from the probability distribution;
[0114] 3. Select the word with the highest probability as the final classification result to complete the classification prediction of fake news. The specific formula for selecting tag words is as follows:
[0115]
[0116] 4. Comprehensive Evaluation and Model Tuning: Through multiple rounds of training and evaluation, the prompt templates and pre-trained models were tuned to improve the model's classification accuracy in low-sample scenarios. Specifically, this included: first, adjusting the structure and content of the prompt templates to better guide the model in understanding the fake news detection task; second, conducting experiments using different pre-trained language models, comparing their performance, and selecting the optimal model; and finally, optimizing model parameters by combining cross-validation and model evaluation metrics (such as precision, recall, and F1 score) to ensure efficient and accurate detection capabilities even with low-sample data.
[0117] A preferred example of the tuning steps for this pre-trained language model based on pseudo-cue learning is as follows:
[0118] 1. Create a prompt template: "The authenticity of this news is: [MASK]. News content: News X."
[0119] 2. Input the pre-trained language model to generate the word probability distribution at the [MASK] position:
[0120] P(\text{true}|\text{context})=0.7;
[0121] P(\text{false}|\text{context})=0.3;
[0122] 3. Select the word with the highest probability as the final classification result:
[0123] \text{Prediction}=argmax{0.7,0.3}=\text{True};
[0124] Through the above specific implementation steps, this invention provides a fake news detection method based on self-supervised learning and propagation consistency, which achieves efficient and accurate fake news detection in environments with few sample data.
[0125] To reiterate, the above embodiments are merely some examples of the present invention, and not embodiments compatible with all application scenarios. All other embodiments obtained by those skilled in the art without any inventive step, referring to the embodiments demonstrated in this invention, are also within the scope of protection of this invention.
Claims
1. A method for detecting fake news based on self-supervised learning and consistency of propagation, characterized in that, Includes the following steps: Step 1: Optimization of sentence representation based on self-supervised contrastive learning: A pre-trained language model is used as the initial detection model; Subsequently, the sentence embedding representation generated by the initial detection model was optimized using a self-supervised contrastive learning method. By constructing positive and negative contrastive sample pairs, the initial detection model's ability to understand and distinguish sentence semantics was enhanced. Step 2: Sample-level adversarial learning based on adversarial generative networks: Construct a generative adversarial network, using random noise and negative fake news samples as input, generate rich adversarial embedding representations through a multilayer perceptron and an independent PLM encoder, and combine the adversarial embeddings to optimize the output embedding representation of the initial detection model; Step 3: Authenticity verification based on news dissemination consistency characteristics: Based on a large number of user news forwarding records collected from social media, a news dissemination consistency network is constructed. Based on the adjacency subgraph feature aggregation algorithm, the dissemination characteristics of news nodes are integrated, and the authenticity characteristics of neighboring news nodes are injected into the current news authenticity prediction process to further improve the accuracy of fake news detection. Step 4: Optimization of the pre-trained language model based on pseudo-prompt learning: Based on the detection model optimized in Step 1 and Step 2, a mask language model based on pseudo-prompt learning is used to train the task during the detection model optimization process. By constructing task-oriented prompt templates, the fake news detection task is transformed into a text fill-in-the-blank question to make full use of the linguistic knowledge of this pre-trained detection model. Combined with the post-processing optimization process in Step 3, the performance of the detection model in scenarios with few samples is improved, thereby achieving the detection of fake news.
2. The method for detecting fake news based on self-supervised learning and consistency of propagation as described in claim 1, characterized in that, Step one: Sentence representation optimization based on self-supervised contrastive learning includes: Construct positive contrast sample pairs and select the representations of the same news text in different instances; Construct negative contrast sample pairs and select representations of different news texts; By optimizing the sentence embedding representation through a contrastive learning loss function, the similarity of positive contrastive sample pairs is maximized, while the similarity of negative contrastive sample pairs is minimized.
3. The method for detecting fake news based on self-supervised learning and consistency of propagation as described in claim 2, characterized in that, The process of optimizing sentence representation based on self-supervised contrastive learning employs the SimCSE framework. It optimizes sentence embedding representation by constructing positive and negative contrastive sample pairs. Positive contrastive sample pairs are obtained by encoding the same news text multiple times, while negative contrastive sample pairs are selected from different news texts. The optimization objective is to maximize the similarity of positive contrastive sample pairs while minimizing the similarity of negative contrastive sample pairs. The formula used in this self-supervised contrastive learning is as follows: Among them, L contrast The optimization objective of this contrastive learning is represented by the loss function; Sim(zi,zj) represents the cosine similarity between two embedding vectors, τ is the temperature coefficient; k represents the index of the sample to be compared; N represents N pairs of comparison samples; the function log(.) represents the logarithmic function; exp(.) represents the exponential function.
4. The method for detecting fake news based on self-supervised learning and consistency of propagation as described in claim 1, characterized in that, Step two: Sample-level adversarial learning based on adversarial generative networks includes: Negative sequence and language encoder-driven engine: Input random word sequences, negative fake news samples, and semantically rewritten news text, and generate adversarial embedding representations; Noise and Multilayer Perceptron Driven Engine: Input random noise and generate text perturbation embedding representation; Binary classification contrastive learning discriminator: The embedding of the classification input represents whether it is real or generated, so as to enhance the detection model's ability to identify fake samples.
5. The method for detecting fake news based on self-supervised learning and consistency of propagation as described in claim 4, characterized in that, The process of sample-level adversarial learning based on adversarial generative networks specifically includes: First, an adversarial generative network was constructed, consisting of two generators and one discriminator. The generators are used to generate adversarial examples, and the discriminator is used to distinguish whether these examples are real or generated. The first generator is specifically designed as follows: it is a negative sequence and language encoder driving engine that takes random word sequences, negative fake news samples, and semantically rewritten news text as input to generate adversarial embedding representations. The second generator is specifically designed as follows: it is a noise and multilayer perceptron driving engine that takes random noise as input to generate text perturbation embedding representations. The discriminator is a binary classification contrastive learning discriminator responsible for classifying the input embedding representations and determining whether they are real samples or generated samples. Subsequently, the objective function of adversarial learning can be specifically equated to: the optimization objective of adversarial learning is equivalent to defining the optimization objectives of the generator and the discriminator; the generator's objective is to generate adversarial samples that are as realistic as possible, so that the discriminator cannot distinguish between real and fake samples; the discriminator's objective is to accurately distinguish between real samples and generated samples. Finally, the specific adversarial learning objective function is defined as follows: in, is the objective function of GAN, where D represents the discriminator, G represents the generator, pdata represents the distribution of the real data, and pz(z) represents the distribution of the noise. It is the discriminator's expected value for the real data x; It is the discriminator's expectation of the generated data; the sample G(z) generated by the generator G is input into the discriminator D for evaluation.
6. The method for detecting fake news based on self-supervised learning and consistency of propagation as described in claim 1, characterized in that, Step three: Verification of authenticity based on the consistency characteristics of news dissemination includes: Construct a news dissemination subgraph: Based on news dissemination records, calculate and identify neighboring news nodes that have a direct or indirect dissemination relationship with the current news; Calculate the credibility of neighboring news nodes: Based on the frequency of news dissemination and the propagation weight between nodes, calculate the credibility of each neighboring news node; By combining the truth prediction results of neighboring news nodes, the truth prediction results of the current news are optimized.
7. The method for detecting fake news based on self-supervised learning and consistency of propagation as described in claim 6, characterized in that, The specific steps for verifying the authenticity of news based on the consistency characteristics of news dissemination are as follows: First, a news dissemination subgraph is constructed. Based on news dissemination records, neighboring news nodes that have direct or indirect dissemination relationships with the current news are calculated and identified. By analyzing the dissemination path of news on social media platforms, a dissemination subgraph is generated, which includes the current news and news nodes related to its dissemination and their connection relationships. Furthermore, the credibility of neighboring news nodes is calculated based on the frequency of news dissemination and the propagation weight between nodes. The formula for calculating credibility is as follows: Where Credibility(Ni) represents the credibility of node Ni; Frequency(Ni) represents the propagation frequency of node Ni; Weight(Ni,Nj) represents the propagation weight between node Ni and its neighboring nodes Nj; Neighbors(Ni) represents the set of neighboring nodes that have a direct propagation relationship with node Ni; α and β are parameters used to balance frequency and propagation weight. Finally, the prediction results of the news are combined with those of neighboring news nodes: by aggregating the prediction results of the news nodes, the prediction results of the current news are optimized.
8. The method for detecting fake news based on self-supervised learning and consistency of propagation as described in claim 1, characterized in that, Step four: Optimizing the pre-trained language model based on pseudo-cue learning includes: Develop a prompt template to transform the fake news detection task into a text fill-in-the-blank question; The cue template is processed using a pre-trained language model to generate the word probability distribution at the [MASK] position; Based on a predefined set of tags, the most suitable tags are selected from the lexical probability distribution to complete the classification and prediction of fake news.
9. The method for detecting fake news based on self-supervised learning and consistency of propagation as described in claim 8, characterized in that, The specific implementation steps for step four are as follows: 1) Construct task-oriented prompt templates: Transform the fake news detection task into a text fill-in-the-blank question; embed the news text to be detected into the template through the designed prompt template, and insert [MASK] tags in key positions; 2) Processing cue templates using a pre-trained language model: Input the constructed cue templates into the pre-trained language model to generate the word probability distribution at the [MASK] position; the pre-trained language model predicts the words at the [MASK] position based on contextual information and outputs the probability of each word; 3) Selecting tag words: Based on the predefined set of tag words, select the most suitable tag words from the word probability distribution to complete the classification prediction of fake news; assuming the predefined set of tag words is {"true", "false"}, select the word with the highest probability as the final classification result according to the probability distribution; 4) Comprehensive evaluation and model tuning: Through multiple rounds of training and evaluation, the prompt templates and pre-trained models are tuned to improve the classification accuracy of the model in scenarios with few samples.
Citation Information
Patent Citations
Method for detecting false information based on multi-modal fusion mechanism of common attention
CN114662586A
Graph neural network false news detection method based on comparative learning
CN117195080A
