Fake news detection method and device based on large language model analysis guidance

Abstract

This paper proposes a fake news detection method guided by large language model analysis. By selectively injecting knowledge from large language model analysis into a small language model, this method enables collaborative detection of large and small language models. To further reduce detection costs, a fake news detection network based on large language model analysis knowledge transfer is derived from the trained network through knowledge transfer. This network does not require large language model analysis as input, but instead uses only the large model analysis knowledge condensed in the detection network for detection.

Description

Technical Field

[0001] The present invention relates to the fields of computer application and natural language understanding technology, and in particular to a method and device for detecting fake news based on large language model analysis and guidance. Background Art

[0002] Automatic fake news detection aims to train machine learning models to distinguish fake news from real news, and has become an effective solution in practice to quickly eliminate the negative impact of fake news.

[0003] Acquiring effective representations of fake news is key to achieving automatic fake news detection. However, due to the complexity of the news fabrication process, acquiring fake news representations remains challenging. This is because fake news creators may adopt various strategies and fabricate any part of the news based on different intentions and purposes. Therefore, to maintain effectiveness and versatility in fake news detection, automatic fake news detection methods need to have the following two capabilities: (1) accurate perception of multiple types of detection clues (such as style, facts, common sense, intention, etc.); (2) a deep understanding of the real-world context in which the news occurs.

[0004] Recently proposed automated fake news detection methods typically utilize small language models (typically with fewer than 5 billion parameters) pre-trained on large-scale corpora, such as BERT and RoBERTa, to understand news content and provide basic representations. While small language models can improve detection, their limited number of parameters and limited knowledge and capabilities still hinder their effectiveness.

[0005] Large language models (LLMs) typically refer to models with tens or hundreds of billions of parameters, trained on large corpora, and possessing powerful natural language understanding capabilities. Large language models, such as GPT-3.5, have demonstrated impressive performance across a wide range of tasks. Compared to small language models, large language models are typically pre-trained on larger corpora, have more parameters, and require more computational resources for training and inference. They perform well across a wide range of natural language processing tasks and hold promise as general-purpose problem solvers. However, due to their significant training costs, large language models struggle to adapt to downstream tasks through fine-tuning on downstream task datasets, as small language models do. Instead, they require in-context learning, which updates only a small portion of the model's parameters, to adapt to downstream tasks. Experimental results show that while large language models can provide reasonable and informative analysis of news from multiple perspectives, the performance of fake news detection using large language models alone is even weaker than that of methods based on small language models.

[0006] Therefore, how to combine the knowledge and ability level of a large language model with the trainability of a small language model is a technical challenge currently facing news authenticity detection. Summary of the Invention

[0007] To address the problem that existing large language models and small language models perform poorly in fake news detection tasks, this paper proposes an adaptive fake news detection method guided by large language model analysis, which fully leverages the advantages of large language models and small language models to improve fake news detection performance.

[0008] To address the shortcomings of existing technologies, the present invention proposes a fake news detection method based on large language model analysis and guidance, which includes:

[0009] Step 1: Construct a first fake news detection model including a classifier, an attention pooling module, and a news analysis interaction module; extract the first news features of the target news using a news encoder, and extract the first analysis features of the target news under each analysis perspective using a large language model and an analysis encoder;

[0010] Step 2: Perform double cross attention and average pooling on the combination of the first news feature and each first analysis feature to obtain the second news feature and the second analysis feature under each analysis angle;

[0011] Step 3: Construct a large language model judgment predictor for determining whether the target news is fake news under the analysis angle based on the first analysis feature, obtain a first fake detection result of the large language model under each analysis angle, and use the second fake detection result of the large language model for the target news as a label to construct a first loss function;

[0012] Step 4: Construct an analysis validity evaluator based on a multi-layer perceptron network. The analysis validity evaluator determines whether the target news is fake news based on each of the second analysis features, obtains the analysis accuracy of the large language model at each analysis angle, and constructs a second loss function based on the correctness label of the second fake news detection result. The second analysis feature is fed into the multi-layer perceptron to obtain the contribution weight of the second news feature at each angle to fake news detection. The second news feature is weighted by the contribution weight to obtain the weighted feature at each analysis angle.

[0013] Step 5: After attention pooling, the first news feature is aggregated with the weighted features from all analysis angles to obtain an aggregated feature. The classifier obtains a fourth falsehood detection result, indicating whether the target news is false news, based on the aggregated feature. The third loss function is constructed in combination with the correct category label of the target news.

[0014] Step 6: Based on the total loss formed by all the first loss function, the second loss function, and the third loss function, the first fake news detection model is trained, and the fake news detection task is performed using the trained first fake news detection model to obtain a detection result of whether the task news is fake news.

[0015] The fake news detection method based on large language model analysis guidance includes:

[0016] Step 7: Using the settings of the news encoder and the classifier to initialize model parameters, a second fake news detection model including the news encoder, the classifier, and an analysis and perception feature simulator is constructed;

[0017] Step 8: extracting features of the target news through the news encoder and sending them to the analysis perception feature simulator to obtain simulated features. The simulated features are subjected to attention pooling processing and then sent to the classifier to obtain a fifth false detection result.

[0018] Step 9: The fifth false detection result is combined with the correct category label of the target news to construct a fourth loss function; a fifth loss function is constructed based on the multi-head attention feature and the aggregated feature after attention pooling processing;

[0019] Step 10: Train the second fake news detection model based on the total loss formed by the fourth loss function and the fifth loss function, and use the trained second fake news detection model to perform a fake news detection task to obtain a detection result of whether the task news is fake news.

[0020] The fake news detection method based on large language model analysis guidance, wherein the large language model is a closed source language model or an open source language model.

[0021] The fake news detection method guided by large language model analysis, wherein the analysis angles include common sense, writing style, and writing intention.

[0022] The present invention also proposes a fake news detection device based on large language model analysis and guidance, which includes:

[0023] Module 1: Construct a first fake news detection model including a classifier, an attention pooling module, and a news analysis interaction module; extract the first news features of the target news using a news encoder, and extract the first analysis features of the target news under each analysis perspective using a large language model and an analysis encoder;

[0024] Module 2: Performing double cross attention and average pooling on the combination of the first news feature and each first analysis feature to obtain the second news feature and second analysis feature from each analysis perspective;

[0025] Module 3: Constructing a large language model judgment predictor for determining whether the target news is fake news under the analysis angle based on the first analysis feature, obtaining a first fake news detection result of the large language model under each analysis angle, and combining a second fake news detection result of the large language model for the target news as a label to construct a first loss function;

[0026] Module 4: Constructing an analysis validity evaluator based on a multi-layer perceptron network. The analysis validity evaluator determines whether the target news is fake news based on each of the second analysis features, obtains the analysis accuracy of the large language model at each analysis angle, and constructs a second loss function based on the correctness label of the second fake news detection result. The second analysis features are fed into the multi-layer perceptron to obtain the contribution weight of the second news features at each angle to fake news detection. The second news features are weighted by the contribution weights to obtain weighted features at each analysis angle.

[0027] Module 5: Attention pooling is performed on the first news feature and then aggregated with the weighted features from all analysis angles to obtain an aggregated feature. The classifier obtains a fourth falsehood detection result, indicating whether the target news is false news, based on the aggregated feature. The third loss function is constructed in combination with the correct category label of the target news.

[0028] Module 6: Train the first fake news detection model based on the total loss formed by all the first loss function, the second loss function, and the third loss function, and use the trained first fake news detection model to perform a fake news detection task to obtain a detection result of whether the task news is fake news.

[0029] The fake news detection device based on large language model analysis guidance includes:

[0030] Module 7, using the setting parameters of the news encoder and the classifier to initialize model parameters, and constructing a second fake news detection model including the news encoder, the classifier, and an analysis and perception feature simulator;

[0031] Module 8, extracting features of the target news through the news encoder, and sending the features to the analysis perception feature simulator to obtain simulated features, performing attention pooling processing on the simulated features, and then sending them to the classifier to obtain a fifth false detection result;

[0032] Module 9: constructing a fourth loss function based on the fifth false detection result and the correct category label of the target news; constructing a fifth loss function based on the multi-head attention feature and the aggregated feature after the attention pooling process;

[0033] Module 10 trains the second fake news detection model based on the total loss formed by the fourth loss function and the fifth loss function, and uses the trained second fake news detection model to perform a fake news detection task to obtain a detection result of whether the task news is fake news.

[0034] The fake news detection device based on large language model analysis guidance, wherein the large language model is a closed source language model or an open source language model.

[0035] The fake news detection device guided by large language model analysis, wherein the analysis angles include common sense, writing style, and writing intention.

[0036] The present invention also provides a server, which includes any one of the false news detection devices described above.

[0037] The present invention also provides a storage medium for storing a computer program for executing the fake news detection method.

[0038] It can be seen from the above scheme that the advantages of the present invention are: by using the analysis of the large language model to guide the detection process of the small language model, a reasonable combination of the respective advantages of the large language model and the small language model is achieved, thereby improving the performance of false news detection. Specifically, the present invention proposes an adaptive false news detection method guided by the analysis of the large language model, which selectively injects the knowledge contained in the analysis of the large language model into the small language model (classifier) to achieve collaborative detection of the large language model and the small language model. In order to further reduce the detection cost, through the method of knowledge transfer, a false news detection network based on the knowledge transfer of the large language model analysis is obtained based on the above-mentioned trained network. The characteristic of this network is that it does not require the large language model analysis as input, and only uses the large model analysis knowledge condensed in the detection network for detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A block diagram of the adaptive fake news detection network guided by large language model analysis of the present invention;

[0040] Figure 2 Network graph for fake news detection based on knowledge transfer of large language model analysis. DETAILED DESCRIPTION

[0041] In order to achieve the above technical effects, the present invention includes the following key technical points:

[0042] Key Point 1: Obtaining Large Language Model Analysis and Constructing News and Analysis Representations. First, the news text is input into the large language model. Using appropriate prompt learning strategies, the large language model analyzes it from different perspectives. Then, two independent text representation encoders are used to obtain representations of the news and its corresponding analysis text, respectively.

[0043] Key Point 2: News-Analysis Collaboration Design. Based on the representation in Key Point 1, news-analysis collaboration is performed to obtain analytically perceived news features. This method designs three modules: a news-analysis interaction module, a large language model judgment and prediction module, and an analysis effectiveness assessment module. These provide the model with effective interaction features between news and analysis, allowing the model to adaptively select effective analysis references for final detection during training.

[0044] Key point 3: Feature aggregation and prediction. Aggregate the news features perceived by analysis and the original news features for the final news classification task. This method is based on the features obtained in key point 2 and aggregates the news vector x and the analysis-perceived news vector f′. xi Aggregation, used for final judgment.

[0045] Key Point 4: Knowledge Transfer from Large Language Model Analysis. This method uses knowledge distillation to build a fake news detection network based on knowledge transfer from large language model analysis, based on an adaptive fake news detection network guided by the analysis of a trained large language model. This network does not require the analysis text of the large language model as input and can be used in low-cost scenarios where large language models cannot be deployed or accessed.

[0046] To illustrate the above-mentioned features and effects of the present invention more clearly and easily, the following embodiments are specifically described below with reference to the accompanying drawings. This specification discloses one or more embodiments incorporating the features of the present invention. The disclosed embodiments are for illustrative purposes only. The scope of protection of the present invention is not limited to the disclosed embodiments; the present invention is defined by the appended claims.

[0047] Figure 1 and Figure 2 The adaptive fake news detection method guided by large language model analysis of the present invention is demonstrated. Figure 1 The adaptive fake news detection network, guided by analysis of large language models (language models with parameter sizes typically in the tens or hundreds of billions, such as closed-source large language models like GPT3.5 and GPT4, and open-source large language models like Llama), consists of three parts: 1. Representation acquisition, which acquires news and its corresponding large language model analysis, and obtains their representations through an independent text encoder; 2. News-analysis collaboration, which interacts features between news and analysis, and uses prediction tasks on the judgment results of the large language model and evaluation tasks on the effectiveness of the large language model analysis as auxiliary training tasks to improve the quality of feature interaction; 3. Authenticity prediction, which aggregates the features obtained from the interaction with the news features processed by the attention mechanism to predict the authenticity of the news.

[0048] Figure 2 The fake news detection network based on large language model analysis knowledge transfer is based on Figure 1 The knowledge is transferred. Specifically, it includes four parts: news encoder, analysis and perception feature simulator, attention mechanism and classifier. The news encoder and classifier are used Figure 1 The parameters of the corresponding modules of the network are initialized. In addition to the news authenticity prediction task, a feature imitation task is added, requiring the classifier to input features and Figure 1 The corresponding features are similar, which enables knowledge transfer of large model analysis and serves low-cost scenarios where large language models cannot be deployed or accessed.

[0049] Obtain large language model analysis and build news and analysis representations:

[0050] In this step, the present invention first inputs the news into the large language model and obtains its analysis from various perspectives. Specifically, this solution selects various perspectives required for fake news detection and analysis, such as common sense, writing style, and writing intention. It constructs a prompt from a specific perspective, combines it with the target news, and then feeds it into the large model. This prompt template then generates the large language model's analysis from that perspective and the corresponding judgment result. The prompt template includes the task description and the text of the news to be tested. An example is shown below:

[0051] Q: Given the following message, determine whether it is true or false. If it is more likely to be false, return "1", otherwise return "0". Be careful not to make vague judgments such as "cannot be determined":

News Content

[0052] A: Let’s start with

Analysis angle

[0053] in,

News Content

Analysis angle

[0054] Afterwards, the present invention uses two independent text representation encoders to obtain the representation (feature vector) of the news and its corresponding analysis text respectively. We obtain the news representation X and its corresponding representations of n different analysis angles For example, R1 stands for "Analyzing from the perspective of the first style, this sentence uses overly emotional expressions and does not conform to the common news style."

[0055] News-Analysis Collaboration:

[0056] This solution aims to provide the model with rich interactions between news and analysis, and enable the model to adaptively select effective analysis as a reference. To achieve this goal, this solution has designed three modules: a news-analysis interaction module, a large language model judgment and prediction module, and an analysis effectiveness assessment module.

[0057] 1. News-Analysis Interactive Module

[0058] In order to achieve comprehensive information interaction between news and analysis, this scheme uses a dual cross-attention mechanism as a news-analysis interactor to explore the interaction between features. Cross-attention can be described as:

[0059]

[0060] Where Q′=W Q Q, K′=W K K, V′=WV V, W Q 、W K and W V are all learnable parameter matrices, and d is the dimension of each row vector in Q′ and K′. Given the representation X of news x and the large language model analysis R of the i-th angle i , through double cross attention and average pooling, the news representation vector f perceived by analysis i can be obtained respectively xi Analysis of news perception i represents vector f ri :

[0061] f xi =AvgPool(CA(R i ,X,X))

[0062] f ri =AvgPool(CA(X,R i ,R i ))

[0063] Among them, AvgPool(·) is the average pooling function.

[0064] 2. Large language model judgment and prediction module

[0065] Understanding the judgments that the analysis points to is a prerequisite for fully utilizing the analytical information. To this end, this solution constructs a large language model judgment prediction task, which requires predicting the large language model's judgment on the authenticity of the news based on the given analysis. We hope that this will deepen the model's understanding of the analyzed text. For the analysis of angle i, this solution represents it as R i Input large language model judgment predictor, which uses multi-layer perceptron (multi-layer perceptron network MLP i1 ) is parameterized and used to predict the judgment of the large language model, thereby assisting the overall prediction:

[0066]

[0067]

[0068] Among them, m i ∈{0,1} and They are the actual judgment of the large language model (extracted from the response of the large language model, 1 represents the large model judges the news x to be false, otherwise it is 0) and the predicted result. pi is the cross entropy loss

[0069] 3. Analysis effectiveness evaluation module

[0070] The effectiveness of analyses from different angles in different news projects is also different, and improper integration may lead to a decline in model performance. In order to enable the model to adaptively select effective large-model analyses, this solution designs an analysis effectiveness evaluation module. In this module, this solution evaluates the contributions of different analyses and adjusts their weights for subsequent authenticity predictions. In other words, it determines whether the news detection results evaluated from different angles of the large language model may be incorrect, and it is necessary to evaluate each angle. This module consists of two stages, namely evaluation and reweighting. In the evaluation stage, this solution uses the news perception analysis vector f ri (i∈{1,2,…,n}) input analysis validity evaluator (Multilayer Perceptron Network MLP i2 ) to predict its effectiveness, and the prediction result is Based on the assumption that “analysis leading to correct judgment is more effective”, the present invention analyzes the corresponding large language model based on the category label of the input news to determine whether it is correct as the validity label u i ∈{0,1}, where 1 represents the correct judgment of the large language model and 0 represents incorrect judgment:

[0071]

[0072]

[0073] In the reweighting stage, f ri Input a standalone MLP i3 Get the weight value w i , and then use this weight to classify the reason perception news vector f xi Re-empower. The specific steps are as follows:

[0074] f′ xi =w i .f xi

[0075] The present invention also uses attention pooling to reduce the dimension of the feature matrix X into a vector x, and the process is as follows:

[0076] x=softmax(W X X)·X

[0077] Where W X Is a learnable transformation matrix. Attention pooling is divided into two steps. One is to calculate the attention weight for each element in the matrix, and the other is to weight each element in the matrix based on the weight. The above softmax (W X X) corresponds to the attention weight calculation step.

[0078] predict:

[0079] Based on the output of the previous step, this scheme converts the news vector x and the analysis perception news vector f′ xi Aggregation is used for final judgment. For news x with label y∈{0,1}, this scheme aggregates these vectors using different weights:

[0080]

[0081] in, is a learnable weight parameter, The weight corresponding to the news vector x, is the weight of analysis angle i in the current news (the weight value is between 0 and 1, and the sum of all weights is 1), f cls is the fusion vector.

[0082] Then f cls Input classifier (Multi-layer Perceptron Network, MLP y ), used for news authenticity prediction:

[0083]

[0084] in if Then predict x as fake news, otherwise it is real news, thr is the pre-set judgment threshold. The cross entropy loss function is used in network training:

[0085]

[0086] Where y is the true or false label of the news during training.

[0087] The total loss function of the above process is the weighted sum of the above loss terms:

[0088]

[0089] Where β1 and β2 are pre-set hyperparameters. The loss function L includes three optimization objectives: the large language model judgment and prediction module in the news analysis and interaction module, the analysis effectiveness evaluation module, and the classifier in the prediction step.

[0090] Large language model analysis knowledge transfer:

[0091] The adaptive fake news detection network guided by the large language model analysis mentioned above needs to send a request to the large language model for each prediction, which is difficult to use in scenarios with limited cost budgets. Therefore, the present invention further attempts to build a fake news detection network based on large language model analysis knowledge transfer based on the trained adaptive fake news detection network guided by the large language model analysis through knowledge transfer. The basic idea is to simulate and internalize the large model analysis knowledge through parameterized modules. Figure 2 As shown in the figure, the scissor symbol indicates that the initialization parameters of this part come from the previously completed training, that is, Figure 1 "Adaptive Fake News Detection Network Guided by Large Language Model Analysis", this scheme uses the corresponding modules in the Adaptive Fake News Detection Network Guided by Large Language Model Analysis to initialize the news encoder and classifier, and trains an analysis-perception feature simulator based on a multi-head attention mechanism and an attention module to internalize knowledge. In addition to using the cross entropy loss L ce The supervised model acquires the ability to classify true and false news. This scheme also uses auxiliary loss L kd Guide Features Imitation network f cls :

[0092]

[0093] In actual use, you need to complete Figure 1 After training the model, Figure 2 Training of the model. Figure 1 and Figure 2 The methods are all proposed by the present invention and are applicable to different scenarios. Figure 1 This method requires the participation of a large language model, has better performance, and is suitable for cost-insensitive scenarios; Figure 2 The method does not require the participation of a large language model, and its performance is weaker than Figure 1 , suitable for cost-sensitive scenarios.

[0094] The following is a system embodiment corresponding to the above method embodiment. This embodiment can be implemented in conjunction with the above embodiment. The relevant technical details mentioned in the above embodiment are still valid in this embodiment and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiment.

[0095] The present invention also proposes a fake news detection device based on large language model analysis and guidance, which includes:

[0096] Module 1: Construct a first fake news detection model including a classifier, an attention pooling module, and a news analysis interaction module; extract the first news features of the target news using a news encoder, and extract the first analysis features of the target news under each analysis perspective using a large language model and an analysis encoder;

[0097] Module 2: Performing double cross attention and average pooling on the combination of the first news feature and each first analysis feature to obtain the second news feature and second analysis feature from each analysis perspective;

[0098] Module 3: Constructing a large language model judgment predictor for determining whether the target news is fake news under the analysis angle based on the first analysis feature, obtaining a first fake news detection result of the large language model under each analysis angle, and combining a second fake news detection result of the large language model for the target news as a label to construct a first loss function;

[0099] Module 4: Constructing an analysis validity evaluator based on a multi-layer perceptron network. The analysis validity evaluator determines whether the target news is fake news based on each of the second analysis features, obtains the analysis accuracy of the large language model at each analysis angle, and constructs a second loss function based on the correctness label of the second fake news detection result. The second analysis features are fed into the multi-layer perceptron to obtain the contribution weight of the second news features at each angle to fake news detection. The second news features are weighted by the contribution weights to obtain weighted features at each analysis angle.

[0100] Module 5: Attention pooling is performed on the first news feature and then aggregated with the weighted features from all analysis angles to obtain an aggregated feature. The classifier obtains a fourth falsehood detection result, indicating whether the target news is false news, based on the aggregated feature. The third loss function is constructed in combination with the correct category label of the target news.

[0101] Module 6: Train the first fake news detection model based on the total loss formed by all the first loss function, the second loss function, and the third loss function, and use the trained first fake news detection model to perform a fake news detection task to obtain a detection result of whether the task news is fake news.

[0102] The fake news detection device based on large language model analysis guidance includes:

[0103] Module 7, using the setting parameters of the news encoder and the classifier to initialize model parameters, and constructing a second fake news detection model including the news encoder, the classifier, and an analysis and perception feature simulator;

[0104] Module 8, extracting features of the target news through the news encoder, and sending the features to the analysis perception feature simulator to obtain simulated features, performing attention pooling processing on the simulated features, and then sending them to the classifier to obtain a fifth false detection result;

[0105] Module 9: constructing a fourth loss function based on the fifth false detection result and the correct category label of the target news; constructing a fifth loss function based on the multi-head attention feature and the aggregated feature after the attention pooling process;

[0106] Module 10 trains the second fake news detection model based on the total loss formed by the fourth loss function and the fifth loss function, and uses the trained second fake news detection model to perform a fake news detection task to obtain a detection result of whether the task news is fake news.

[0107] The fake news detection device based on large language model analysis guidance, wherein the large language model is a closed source language model or an open source language model.

[0108] The fake news detection device guided by large language model analysis, wherein the analysis angles include common sense, writing style, and writing intention.

[0109] The present invention also provides a server, which includes any one of the false news detection devices described above.

[0110] The present invention also provides a storage medium for storing a computer program for executing the fake news detection method.

[0111] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A fake news detection method based on large language model analysis, characterized in that: include: Step 1: Build a first fake news detection model including a classifier, an attention pooling module, and a news analysis interaction module; Extract the first news feature of the target news through the news encoder, and extract the first analysis feature of the target news under each analysis angle through the large language model and the analysis encoder; Step 2: Perform double cross attention and average pooling on the combination of the first news feature and each first analysis feature to obtain the second news feature and the second analysis feature under each analysis angle; Step 3: Construct a large language model judgment predictor for determining whether the target news is fake news under the analysis angle based on the first analysis feature, obtain a first fake detection result of the large language model under each analysis angle, and use the second fake detection result of the large language model for the target news as a label to construct a first loss function; Step 4: Construct an analysis validity evaluator based on a multi-layer perceptron network. The analysis validity evaluator determines whether the target news is fake news based on each of the second analysis features, obtains the analysis accuracy of the large language model at each analysis angle, and constructs a second loss function based on the correctness label of the second fake news detection result. The second analysis feature is fed into the multi-layer perceptron to obtain the contribution weight of the second news feature at each angle to fake news detection. The second news feature is weighted by the contribution weight to obtain the weighted feature at each analysis angle. Step 5: After attention pooling, the first news feature is aggregated with the weighted features from all analysis angles to obtain an aggregated feature. The classifier obtains a fourth falsehood detection result, indicating whether the target news is false news, based on the aggregated feature. The third loss function is constructed in combination with the correct category label of the target news. Step 6: Based on the total loss formed by all the first loss function, the second loss function, and the third loss function, the first fake news detection model is trained, and the fake news detection task is performed using the trained first fake news detection model to obtain a detection result of whether the task news is fake news.

2. The method for detecting fake news based on large language model analysis and guidance according to claim 1, characterized in that: include: Step 7: Using the settings of the news encoder and the classifier to initialize model parameters, a second fake news detection model including the news encoder, the classifier, and an analysis and perception feature simulator is constructed; Step 8: extracting features of the target news through the news encoder and sending them to the analysis perception feature simulator to obtain simulated features. The simulated features are subjected to attention pooling processing and then sent to the classifier to obtain a fifth false detection result. Step 9: The fifth false detection result is combined with the correct category label of the target news to construct a fourth loss function; and a fifth loss function is constructed based on the simulated features and the aggregated features after the attention pooling process; Step 10: Train the second fake news detection model based on the total loss formed by the fourth loss function and the fifth loss function, and use the trained second fake news detection model to perform a fake news detection task to obtain a detection result of whether the task news is fake news.

3. The fake news detection method based on large language model analysis and guidance according to claim 1, characterized in that: The large language model is a closed-source language model or an open-source language model.

4. The method for detecting fake news based on large language model analysis and guidance according to claim 1, characterized in that: The analysis angles include common sense, writing style, and writing intention.

5. A fake news detection device based on large language model analysis guidance, characterized in that: include: Module 1: Construct a first fake news detection model including a classifier, an attention pooling module, and a news analysis interaction module; extract the first news features of the target news using a news encoder, and extract the first analysis features of the target news under each analysis perspective using a large language model and an analysis encoder; Module 2: Performing double cross attention and average pooling on the combination of the first news feature and each first analysis feature to obtain the second news feature and second analysis feature from each analysis perspective; Module 3: Constructing a large language model judgment predictor for determining whether the target news is fake news under the analysis angle based on the first analysis feature, obtaining a first fake news detection result of the large language model under each analysis angle, and combining a second fake news detection result of the large language model for the target news as a label to construct a first loss function; Module 4: Constructing an analysis validity evaluator based on a multi-layer perceptron network. The analysis validity evaluator determines whether the target news is fake news based on each of the second analysis features, obtains the analysis accuracy of the large language model at each analysis angle, and constructs a second loss function based on the correctness label of the second fake news detection result. The second analysis features are fed into the multi-layer perceptron to obtain the contribution weight of the second news features at each angle to fake news detection. The second news features are weighted by the contribution weights to obtain weighted features at each analysis angle. Module 5: Attention pooling is performed on the first news feature and then aggregated with the weighted features from all analysis angles to obtain an aggregated feature. The classifier obtains a fourth falsehood detection result, indicating whether the target news is false news, based on the aggregated feature. The third loss function is constructed in combination with the correct category label of the target news. Module 6: Train the first fake news detection model based on the total loss formed by all the first loss function, the second loss function, and the third loss function, and use the trained first fake news detection model to perform a fake news detection task to obtain a detection result of whether the task news is fake news.

6. The fake news detection device based on large language model analysis and guidance according to claim 5, characterized in that: include: Module 7, using the setting parameters of the news encoder and the classifier to initialize model parameters, and constructing a second fake news detection model including the news encoder, the classifier, and an analysis and perception feature simulator; Module 8, extracting features of the target news through the news encoder, and sending the features to the analysis perception feature simulator to obtain simulated features, performing attention pooling processing on the simulated features, and then sending them to the classifier to obtain a fifth false detection result; Module 9: constructing a fourth loss function based on the fifth false detection result and the correct category label of the target news; constructing a fifth loss function based on the simulated features and the aggregated features after the attention pooling process; Module 10 trains the second fake news detection model based on the total loss formed by the fourth loss function and the fifth loss function, and uses the trained second fake news detection model to perform a fake news detection task to obtain a detection result of whether the task news is fake news.

7. The fake news detection device based on large language model analysis and guidance according to claim 5, characterized in that: The large language model is a closed-source language model or an open-source language model.

8. The fake news detection device based on large language model analysis and guidance according to claim 5, characterized in that: The analysis angles include common sense, writing style, and writing intention.

9. A server, characterized in that: The invention comprises the fake news detection device according to any one of claims 5 to 8.

10. A storage medium for storing any computer program for executing the fake news detection method according to any one of claims 1 to 4.

Images