A false news detection method and system based on multi-dimensional perspective rationale fusion

Abstract

This invention discloses a method and system for detecting fake news based on multi-dimensional perspective fusion of reasoning, belonging to the field of natural language processing technology. This invention utilizes a large language model to generate four types of reasoning based on perspectives: common sense reasoning, logical structure, sentiment tendency, and fact-checking, and constructs a dataset accordingly. It achieves semantic interaction between news and reasoning through a dual cross-attention mechanism, introduces a reasoning prediction task to distill knowledge from the large language model, and dynamically filters high-quality reasoning based on a reasoning evaluation mechanism. Finally, it adaptively fuses news features and reasoning features through a dynamic gating mechanism for classification. This invention solves the problems of insufficient interpretability, high dependence on manual annotation, and noise in reasoning from large language models in existing technologies, effectively improving the accuracy, robustness, and interpretability of fake news detection.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, and specifically relates to a method and system for detecting fake news based on multi-dimensional perspective fusion. Background Technology

[0002] The booming development of social media platforms has significantly improved the efficiency of users' information acquisition and dissemination, but it has also provided a convenient channel for the breeding and spread of fake news. In the massive flood of online information, the proliferation of false content often leaves users struggling to distinguish truth from falsehood, posing a serious challenge to the healthy development of the current media ecosystem. Even more serious is the fact that large language models (LMs), with their powerful content generation and simulation capabilities, have significantly lowered the barrier to fabricating fake news. Therefore, building an efficient and accurate fake news detection mechanism is of great practical significance for maintaining order in the information environment and safeguarding public cognitive security.

[0003] The evolution of fake news detection technology has gone through several stages. Early research mainly relied on traditional feature engineering and machine learning methods, such as decision tree models and shallow semantic analysis using support vector machines, which initially constructed the basic framework for fake news identification. With breakthroughs in deep learning technology, the research paradigm gradually shifted to representation learning, such as mining local text features based on convolutional neural networks and modeling information propagation relationships based on graph neural networks, driving a leap in detection capabilities towards deep semantic understanding. Today, this field has entered a paradigm shift centered on large language models. This transformation not only breaks through the limitations of traditional rule systems but also achieves a leap from surface pattern matching to end-to-end semantic parsing through tasks such as generative validation, multimodal contextual reasoning, and fact-checking. For example, by using fusion retrieval-augmented generation (RAG) technology, the accuracy and reliability of the output results of large language models have been significantly improved, and the fine-tuning strategy of large language models has effectively enhanced the model's sensitivity to deceptive content. Large language models have shown unique advantages in complex text analysis and stylistic feature identification.

[0004] However, existing technologies still have the following key problems:

[0005] (1) Insufficient interpretability: Traditional deep learning models are mostly "black boxes", which make it difficult to provide clear rationale for judging the authenticity of news, thus limiting their credibility and practicality in real-world scenarios.

[0006] (2) Dependence on manual annotation: High-quality news analysis requires a large amount of manual annotation, which is costly and time-consuming, making it difficult to keep up with the rapid iteration of social media news.

[0007] (3) The rationale generated by large language models may contain "illusions" and uncertainties. Direct use will introduce noise and affect detection performance. When utilizing the capabilities of large models, existing methods often only use single or dual-perspective rationale (such as only combining text description and common sense reasoning), which is difficult to fully cover the key dimensions of news analysis. Moreover, they generally lack dynamic evaluation and screening mechanisms for the quality of generated rationale, which makes it impossible for the model to effectively distinguish between high-value rationale and noisy information.

[0008] To address the aforementioned issues, there is an urgent need for an efficient and reliable method and system for detecting fake news. This system should be able to generate multi-dimensional perspectives using large language models, achieve comprehensive semantic enhancement, and adaptively filter and utilize high-value arguments based on input. It should also dynamically fuse news features and argument features to improve the accuracy, robustness, and interpretability of fake news detection, thereby providing accurate and interpretable decision support for identifying fake news on social media. Summary of the Invention

[0009] In view of this, the present invention provides a fake news detection method and system based on multi-dimensional perspective evidence fusion, in order to solve the technical problems of insufficient interpretability of fake news detection, high dependence on manual annotation, difficulty in making full use of multi-dimensional semantic evidence, and noise in the evidence generated by large language models in the prior art.

[0010] Currently, large language models offer diverse solutions and achieve significant performance for natural language processing tasks. This invention designs a multi-dimensional perspective reasoning fusion module, aiming to integrate explanatory reasoning from different dimensions of news to achieve more comprehensive semantic enhancement. Simultaneously, this invention proposes a novel dynamic gating module, enabling the model to adaptively filter and utilize high-value reasoning based on input, thereby improving the accuracy and interpretability of news authenticity verification.

[0011] The technical solution adopted in this invention is as follows: On the one hand, this invention provides a method for detecting fake news based on multi-dimensional perspective reasoning fusion, including the following steps: The news texts to be detected are preprocessed, and a large language model is used to generate multi-dimensional perspective analysis of the rationale for each news item. Predictions based on each perspective are generated to construct a news rationale dataset. Semantically encode the news text and the arguments from various perspectives to generate corresponding semantic representations; By using a dual cross-attention mechanism, each semantic representation of the rationale is interacted with the semantic representation of the news, resulting in supplementary features of the rationale to the news content and guiding features of the news to the rationale analysis. The evaluation results of news credibility by a large language model based on the semantic representation of rationale from various perspectives are independently predicted. Based on the guiding characteristics of the news in the rationale analysis, the validity weights of the rationale in each dimension are calculated. The validity weights are then used to weight the supplementary features of the rationale to the news content. Finally, the weighted features are extracted to obtain enhanced news-rational features. The news feature representation is fused with various enhanced news-justification features. The weights of each justification feature are adaptively adjusted through a dynamic gating mechanism to obtain the final feature representation. The news authenticity prediction result is obtained by classification based on the final feature representation.

[0012] Furthermore, the data preprocessing includes: deleting news entries containing missing items, tokenizing the data, and constructing a thesaurus.

[0013] Furthermore, the process of generating multi-dimensional perspective analysis justifications for each news item using a large language model, and generating prediction results based on each perspective to construct a news justification dataset includes: Using a large language model, four types of perspective justifications are generated for each news item through prompt word engineering. These four types of perspective justifications include common sense reasoning, logical structure, sentiment tendency, and fact verification. The large language model is used to predict the corresponding results based on the four types of perspectives generated, forming a news reasoning dataset with the prediction labels of the large language model.

[0014] Furthermore, the step of semantically encoding the news text and the rationale from each perspective to generate corresponding semantic representations includes: Representing news items as The four-dimensional perspective rationale is expressed as follows: , , , Two BERT models were used to independently encode semantic representations. , , , , Two BERT models were used for news encoding and reasoning encoding, respectively.

[0015] Furthermore, by employing a dual cross-attention mechanism, each semantic representation of the rationale is interacted with the semantic representation of the news, resulting in supplementary features of the rationale to the news content and guiding features of the news for rationale analysis, including: Given news content and reasoning The supplementary features of the rationale to the news content are obtained through cross-attention calculation. And the guiding nature of news in reasoned analysis The calculation process is as follows:

[0016]

[0017]

[0018] in, , , , , As a dimension, This indicates that the lexical representations of the cross-attention output are averaged.

[0019] Furthermore, the evaluation results of the large language model based on the independent prediction of semantic representations of various perspectives on news credibility include: Based on given rationale, the rationale predictor predicts the rationale. The independent prediction large language model provides the original evaluation result of news credibility. The rationale predictor is parameterized using a multilayer perceptron, and the calculation process is as follows:

[0020] in, For large language models to provide rationale The prediction results For the rationale predictor The prediction results, loss function Using cross-entropy loss .

[0021] Furthermore, based on the guiding characteristics of the news in the rationale analysis, the validity weights of the rationale analysis in each dimension are calculated. These validity weights are then used to weight the supplementary features of the rationale to the news content. Finally, feature extraction is performed on the weighted features to obtain enhanced news-rational features, including: The guiding nature of news in reasoned analysis Input the reasoning evaluator module to predict the validity of the reasoning. Useful label based on the judgment of correctness , will vector The input is fed into a multilayer perceptron to obtain the weight coefficients. Using this weighting coefficient to supplement the news content with the rationale Reweighting yields a weighted news-justification vector. The calculation process is as follows:

[0022]

[0023] in, For large language models, make reasoned judgments. Indicate the validity of the reasoning. Using a binary cross-entropy loss function ,and ; Weighted news-justification vectors were processed using a feature extractor. The process is performed to obtain enhanced news-rational data after feature extraction. The calculation process is as follows:

[0024] in, This represents the weight matrix of the linear layer. This is a bias term.

[0025] Furthermore, the news feature representation is fused with various enhanced news-justification features. A dynamic gating mechanism adaptively adjusts the weights of each justification feature to obtain the final feature representation. Classification is then performed based on this final feature representation to obtain the news authenticity prediction result, including: News key feature vectors And enhanced news-justification feature vector The input is fed into the fusion unit to obtain the fused features. :

[0026] Fusion features The final feature vector is obtained by inputting the dynamic gating unit. :

[0027] in, and For linear layer parameters, and It is a bias term, and ; The final feature vector The input is fed into a multilayer perceptron classifier, where a sigmoid activation function is used to obtain the prediction result, and the classification loss is calculated using the binary cross-entropy loss function.

[0028] The total loss function is a weighted sum of classification loss, rationale prediction loss, and rationale evaluation loss:

[0029] in, and It's a hyperparameter.

[0030] Furthermore, this invention also provides a fake news detection system based on multi-dimensional perspective reasoning fusion, comprising: Data preprocessing module: Used to preprocess the news text to be detected, generate multi-dimensional analysis of the rationale for each news item using a large language model, and generate prediction results based on each perspective to build a news rationale dataset; Encoder module: Used to semantically encode the news text and the arguments from various perspectives, generating corresponding semantic representations; News-Reasoning Interaction Module: This module uses a dual cross-attention mechanism to interact each reasoning semantic representation with the news semantic representation, thereby obtaining supplementary features of reasoning to news content and guiding features of news to reasoning analysis. Reasoning Prediction Module: Used to independently predict the evaluation results of news credibility by the large language model based on the semantic representation of reasoning from various perspectives; The rationale evaluation module is used to calculate the validity weight of each dimension of the rationale analysis based on the guiding features of the news on the rationale analysis, use the validity weight to weight the supplementary features of the rationale to the news content, and extract features from the weighted features to obtain enhanced news-rational features. Dynamic gating module: It is used to fuse news feature representation with various enhanced news-justification features, and adaptively adjust the weight of each justification feature through dynamic gating mechanism to obtain the final feature representation. Based on the final feature representation, classification is performed to obtain the news authenticity prediction result.

[0031] Compared with the prior art, the beneficial effects of the present invention are: (1) By generating multi-dimensional perspectives (common sense reasoning, logical structure, sentiment tendency, and fact verification) through a large language model, it covers multiple key dimensions of news analysis, provides a more comprehensive explanation basis for fake news detection, and significantly improves the interpretability of the detection.

[0032] (2) A news-reasoning dynamic interaction module was designed, and a dual cross-attention mechanism was introduced to realize the deep semantic interaction between news and reasoning. At the same time, the reasoning knowledge of the large language model was distilled through the reasoning prediction task, and a reasoning evaluation mechanism was used to dynamically screen high-quality reasoning, which effectively suppressed the noise and "illusion" problem in the generation of reasoning and enhanced the robustness of the model.

[0033] (3) The dynamic gating mechanism is adopted to adaptively fuse news features and various rationale features. The weight of each rationale feature is dynamically adjusted according to the input news, which realizes more accurate feature fusion, improves the accuracy of fake news detection, and solves the problem that traditional methods rely on single text features and have insufficient ability to identify hidden false information.

[0034] (4) The high cost of manual annotation is transformed into a controllable computational cost of large language model. The large language model is used to automatically generate multi-dimensional perspective analysis of the rationale and its prediction results through prompt word engineering, providing an economical and efficient solution for rapid iteration of fake news detection. It overcomes the problems of high dependence on manual annotation, high cost and long cycle of existing methods.

[0035] This invention addresses the problems of insufficient interpretability, high dependence on manual annotation, and difficulty in fully utilizing multi-dimensional semantic evidence in traditional models for current fake news detection tasks. At the same time, it effectively suppresses noise in the reasoning generated by large language models and enhances the robustness of the model. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a flowchart illustrating a method for detecting fake news based on multi-dimensional perspective fusion, as described in an embodiment of the present invention.

[0038] Figure 2 This is a model architecture diagram of a fake news detection system based on multi-dimensional perspective fusion in an embodiment of the present invention. Detailed Implementation

[0039] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0040] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0041] This invention constructs a collaborative detection framework based on multi-dimensional reasoning generated by a large language model. Its core lies in introducing external explanatory evidence beyond news text. Through a two-way interaction mechanism between news and reasoning, a reasoning credibility prediction and evaluation mechanism, and a dynamic gating fusion mechanism, it addresses the technical problems of insufficient interpretability of detection results, the reasoning illusion of the large language model, and high manual annotation costs. Specifically, at the input information source level, existing technologies typically rely solely on news text content to mine feature representations from different perspectives. This invention, in addition to news text, introduces explanatory reasoning automatically generated by a large language model, covering multiple dimensions such as common sense, logic, sentiment, and fact-checking, as an independent information source. At the level of multi-source information connotation, the multi-perspective in existing technologies is based on semantic representation perspectives extracted through feature engineering, while the multi-dimensional aspect of this invention refers to explanatory reasoning dimensions geared towards truth / falsehood determination. At the feature fusion logic level, existing technologies typically employ a direct fusion strategy of multi-perspective features, while this invention adopts a mechanism of filtering credible reasoning before fusion, i.e., first predicting the task and validity through reasoning. The evaluation performs credibility modeling and noise suppression on the rationale, and then assigns differentiated weights based on the evaluation results. A dynamic gating mechanism adaptively adjusts the fusion ratio of news features and rationale features to avoid noise propagation caused by low-quality rationale directly participating in classification. At the model structure level, this invention achieves coordinated convergence of rationale generation and utilization, quality control, and final classification tasks by jointly optimizing classification loss, rationale prediction loss, and rationale evaluation loss. It can enhance the interpretability of detection results while controlling the quality of rationale without relying on a domain adaptation module, enabling the model to distinguish between helpful and misleading rationale, thereby significantly improving the robustness and generalization performance of the system in complex scenarios.

[0042] The technical method of this invention is as follows: First, a large language model is used to generate multi-dimensional rationale for news through prompting engineering; then, the news content and each rationale are semantically encoded; next, a news-rational dynamic interaction module is used to realize bidirectional interaction, rationale prediction and evaluation, and obtain weighted rationale features; finally, a dynamic gating mechanism is used to fuse news features and rationale features, and input them into a classifier to obtain detection results.

[0043] like Figure 1 As shown in the figure, a method for detecting fake news based on multi-dimensional perspective fusion in an embodiment of the present invention specifically includes the following steps: S1: Perform data preprocessing on the news text to be detected, use a large language model to generate multi-dimensional perspective analysis of the rationale for each news item, and generate prediction results based on each perspective to construct a news rationale dataset.

[0044] Specifically, the fake news data is cleaned and tagged, news entries with missing information are deleted, and a thesaurus of news and news authenticity is constructed.

[0045] Using a large language model, four types of perspectives (common sense reasoning, logical structure, sentiment, and fact-checking) are generated for each news item through prompt word engineering. The large language model is used to predict the corresponding results based on the four types of perspectives generated, forming a news reasoning dataset with the prediction labels of the large language model. By constructing a multi-task learning framework based on a news reasoning dataset for prediction, the semantic analysis capabilities of a large language model are leveraged to enhance the model's representation and prediction performance.

[0046] S2 performs semantic encoding on the news text and the rationale from each perspective to generate corresponding semantic representations.

[0047] Specifically, news items are represented as The four-dimensional perspective rationale is expressed as follows: , , , It uses two BERT models (one for news encoding and one for justification encoding) to independently encode semantic representations. , , , , .

[0048] S3, through a dual cross-attention mechanism, interacts each semantic representation of the rationale with the semantic representation of the news to obtain the supplementary features of the rationale to the news content and the guiding features of the news to the rationale analysis.

[0049] Specifically, the model uses an attention mechanism to extract news content. Key features.

[0050] To achieve deep semantic interaction between news and multidimensional reasoning, the model employs a news-reasoning interactor with a dual cross-attention mechanism. This mechanism dynamically captures news through bidirectional information transmission. with reason The implicit connections between them.

[0051] Given news content and reasoning The news-rational vector is obtained after the interaction through the interactor. (Representing the supplementary characteristics of the rationale to the news content) and the rationale-news vector (This represents the guiding nature of news in the analysis of reasoning), and the calculation process is as follows:

[0052]

[0053]

[0054] in, , , , , As a dimension, This indicates that the lexical representations output by cross-attention are averaged to obtain a single-vector text representation. .

[0055] S4, based on the semantic representation of rationale from various perspectives, independently predicts the evaluation results of news credibility by a large language model.

[0056] Specifically, to deeply explore the potential value of the rationale generated by large language models, it is necessary to explicitly model their implicit evaluation signals. To this end, a rationale prediction task is introduced, requiring the model to predict rationale based solely on given rationale. Independently predict the original evaluation results of news credibility by a large language model. The core objectives of this task include signal decoupling (forcing the model to focus on the discriminative features of the argument itself, rather than relying on the news content) and knowledge distillation (transferring deep semantic understanding capabilities to a lightweight detection model by fitting the evaluation behavior of the large language model).

[0057] The rationale prediction task is implemented through a rationale predictor, whose structure is parameterized using a multilayer perceptron (MLP). The calculation process is as follows:

[0058] in, It is a large language model for rationale The prediction results It is a rationale predictor for The prediction results, loss function Using cross-entropy loss .

[0059] The introduction of the rationale prediction task aims to distill the reasoning logic of the large language model, enabling the small model to understand and evaluate the credibility of each rationale, thereby ensuring the reliability of the rationales used.

[0060] S5. Calculate the validity weights of the reasoning for each dimension based on the guiding features of the news for reasoning analysis. Use the validity weights to weight the supplementary features of the reasoning for the news content, and extract features from the weighted features to obtain enhanced news-reasoning features.

[0061] Specifically, given the complexity and diversity of news content, the utility value of different perspectives and arguments varies for various types of news. Therefore, an argument evaluation mechanism is introduced to first assess the contribution of each argument and then dynamically adjust its weight.

[0062] The guiding nature of news in reasoned analysis Input each argument into the argument evaluator module to predict the validity of each argument. (Using the correctness of the judgment as a usefulness label), the vector The input is fed into a multilayer perceptron to obtain the weight coefficients. Using this weighting coefficient on the news-justification vector Reweighting yields a weighted news-justification vector. The calculation process is as follows:

[0063]

[0064] in, For large language models, make reasoned judgments. The loss function represents the validity of the rationale. Using binary cross-entropy loss ,and .

[0065] The rationale evaluation mechanism dynamically assesses the contribution weight of different rationales to the current news, which not only improves performance but also enables the model to highlight the most critical explanatory dimensions for the current decision, thereby enhancing the transparency of the decision.

[0066] In order to extract the weighted news-justification vector The main features are processed by the model through a feature extractor to obtain the news-reasoning after feature extraction. The calculation process is as follows:

[0067] in, This represents the weight matrix of the linear layer. It is a bias term, and .

[0068] S6 integrates the news feature representation with each enhanced news-justification feature, and adaptively adjusts the weights of each justification feature through a dynamic gating mechanism to obtain the final feature representation. Based on the final feature representation, classification is performed to obtain the news authenticity prediction result.

[0069] Specifically, aggregate key feature vectors of news and feature extraction vector Furthermore, a dynamic gating mechanism was designed to optimize feature selection within the aggregation vector.

[0070] News key feature vectors and feature extraction vector The input is fed into the fusion module to obtain the fusion features. :

[0071] Aggregation features The input is a dynamic gating unit, which dynamically adjusts the weights of each rationale vector to achieve adaptive fusion, resulting in the final feature vector. :

[0072] in, and For linear layer parameters, and It is a bias term, and .

[0073] The final feature vector The input is fed into a multilayer perceptron classifier, where a sigmoid activation function is used to obtain the prediction result, and the classification loss is calculated using the binary cross-entropy loss function.

[0074] The total loss function is a weighted sum of classification loss, rationale prediction loss, and rationale evaluation loss:

[0075] in and It's a hyperparameter.

[0076] The fake news detection method in the above embodiments generates multi-dimensional rationale through a large language model, designs a news-rational dynamic interaction module to introduce rationale prediction tasks and rationale evaluation mechanisms, and adopts a dynamic gating mechanism to adaptively fuse news features and various rationale features, thereby achieving accurate identification of false information. It effectively solves the problems of traditional fake news detection methods relying on single text features, insufficient ability to identify hidden false information, and difficulty in making full use of multi-dimensional semantic evidence.

[0077] More specifically, the above embodiments have the following beneficial effects: (1) The above embodiments automatically generate multi-dimensional reasoning covering common sense, logic, sentiment and fact-checking by introducing a large language model, which can provide an explicit chain of evidence for the determination of fake news to enhance the interpretability of the results, while reducing the reliance on manually constructed explanatory labels and expert rules, and helping the model to capture false signals that are difficult to identify by surface semantics alone in complex or implicitly misleading texts.

[0078] (2) The above embodiments realize bidirectional information interaction between news and reasoning through a dual cross-attention mechanism, which enables reasoning to supplement the semantic information of news and news to constrain the generation of reasoning, thereby improving the semantic alignment between the two and avoiding the reasoning being coarsely spliced ​​as static external features, thereby enhancing the discriminative ability of fused features.

[0079] (3) The above embodiments introduce a reasoning prediction module to model the credibility of the content generated by the large language model, so that the model can distinguish between helpful reasoning and misleading reasoning, thereby effectively mitigating the interference of large model illusion, reasoning bias and noise propagation on the final classification results and improving the robustness and stability of the system.

[0080] (4) The above embodiments, by evaluating the effectiveness of different justifications and assigning differentiated weights, can automatically highlight effective justifications and suppress low-quality justifications when the quality of justifications is inconsistent, thereby reducing the negative impact of irrelevant or conflicting justifications on the classification results and achieving fine-grained controllability of justification utilization.

[0081] (5) The above embodiments adaptively fuse news features and reasoning features through dynamic gating mechanism. It can dynamically adjust the information contribution ratio according to the characteristics of different samples, reduce its negative impact when the reasoning quality is weak, and give full play to its auxiliary role when the reasoning quality is strong, thereby avoiding information imbalance caused by fixed fusion strategy, thus improving classification accuracy and generalization ability.

[0082] (6) The above embodiments jointly optimize the classification loss, the rationale prediction loss and the rationale evaluation loss, so that the rationale generation and utilization, the rationale quality control and the final classification task form a synergistic constraint, promote the synergistic convergence of the rationale module and the classification target, thereby improving the consistency and stability of the overall system.

[0083] Corresponding to the fake news detection method in the above embodiments, this invention also provides a fake news detection system based on multi-dimensional perspective reasoning fusion, including: Data preprocessing module: Used to preprocess the news text to be detected, generate multi-dimensional analysis of the rationale for each news item using a large language model, and generate prediction results based on each perspective to build a news rationale dataset; Encoder module: Used to semantically encode the news text and the arguments from various perspectives, generating corresponding semantic representations; News-Reasoning Interaction Module: This module uses a dual cross-attention mechanism to interact each reasoning semantic representation with the news semantic representation, thereby obtaining supplementary features of reasoning to news content and guiding features of news to reasoning analysis. Reasoning Prediction Module: Used to independently predict the evaluation results of news credibility by the large language model based on the semantic representation of reasoning from various perspectives; The rationale evaluation module is used to calculate the validity weight of each dimension of the rationale analysis based on the guiding features of the news on the rationale analysis, use the validity weight to weight the supplementary features of the rationale to the news content, and extract features from the weighted features to obtain enhanced news-rational features. Dynamic gating module: It is used to fuse news feature representation with various enhanced news-justification features, and adaptively adjust the weight of each justification feature through dynamic gating mechanism to obtain the final feature representation. Based on the final feature representation, classification is performed to obtain the news authenticity prediction result.

[0084] The fake news detection system based on multi-dimensional perspective fusion in this embodiment of the invention is described simply because it corresponds to the fake news detection method based on multi-dimensional perspective fusion in the above embodiment. For related similarities, please refer to the description of the fake news detection method based on multi-dimensional perspective fusion in the above embodiment, which will not be described in detail here.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting fake news based on multi-dimensional perspective reasoning fusion, characterized in that, Includes the following steps: The news texts to be detected are preprocessed, and a large language model is used to generate multi-dimensional perspective analysis of the rationale for each news item. Predictions based on each perspective are generated to construct a news rationale dataset. Semantically encode the news text and the arguments from various perspectives to generate corresponding semantic representations; By using a dual cross-attention mechanism, each semantic representation of the rationale is interacted with the semantic representation of the news, resulting in supplementary features of the rationale to the news content and guiding features of the news to the rationale analysis. The evaluation results of news credibility by a large language model based on the semantic representation of rationale from various perspectives are independently predicted. Based on the guiding characteristics of the news in the rationale analysis, the validity weights of the rationale in each dimension are calculated. The validity weights are then used to weight the supplementary features of the rationale to the news content. Finally, the weighted features are extracted to obtain enhanced news-rational features. The news feature representation is fused with various enhanced news-justification features. The weights of each justification feature are adaptively adjusted through a dynamic gating mechanism to obtain the final feature representation. The news authenticity prediction result is obtained by classification based on the final feature representation.

2. The method according to claim 1, characterized in that, The data preprocessing includes: deleting news entries containing missing items, tokenizing the data, and constructing a thesaurus.

3. The method according to claim 1 or 2, characterized in that, The process involves using a large language model to generate multi-dimensional perspective analysis of each news item, generating prediction results based on each perspective, and constructing a news reasoning dataset, including: Using a large language model, four types of perspective justifications are generated for each news item through prompt word engineering. These four types of perspective justifications include common sense reasoning, logical structure, sentiment tendency, and fact verification. The large language model is used to predict the corresponding results based on the four types of perspectives generated, forming a news reasoning dataset with the prediction labels of the large language model.

4. The method according to claim 1, characterized in that, The step of semantically encoding the news text and the rationale from each perspective to generate corresponding semantic representations includes: Representing news items as The four-dimensional perspective rationale is expressed as follows: , , , Two BERT models were used to independently encode semantic representations. , , , , Two BERT models were used for news encoding and reasoning encoding, respectively.

5. The method according to claim 1, characterized in that, By employing a dual cross-attention mechanism, each semantic representation of the rationale interacts with the semantic representation of the news, resulting in supplementary features of the rationale to the news content and guiding features of the news for rationale analysis, including: Given news content and reasoning The supplementary features of the rationale to the news content are obtained through cross-attention calculation. And the guiding nature of news in reasoned analysis The calculation process is as follows: in, , , , , As a dimension, This indicates that the lexical representations of the cross-attention output are averaged.

6. The method according to claim 1, characterized in that, The evaluation results of the news credibility based on the independent prediction of semantic representations of various perspectives by the large language model include: Based on given rationale, the rationale predictor predicts the rationale. The independent prediction large language model provides the original evaluation result of news credibility. The rationale predictor is parameterized using a multilayer perceptron, and the calculation process is as follows: in, For large language models to provide rationale The prediction results For the rationale predictor The prediction results, loss function Using cross-entropy loss .

7. The method according to claim 1, characterized in that, Based on the guiding characteristics of the news in the rationale analysis, the validity weights of the rationale analysis in each dimension are calculated. These validity weights are then used to weight the supplementary features of the rationale to the news content. Finally, feature extraction is performed on the weighted features to obtain enhanced news-rational features, including: The guiding nature of news in reasoned analysis Input the reasoning evaluator module to predict the validity of the reasoning. Useful label based on the judgment of correctness , will vector The input is fed into a multilayer perceptron to obtain the weight coefficients. Using this weighting coefficient to supplement the news content with the rationale Reweighting yields a weighted news-justification vector. The calculation process is as follows: in, For large language models, make reasoned judgments. Indicate the validity of the reasoning. Using a binary cross-entropy loss function ,and ; Weighted news-justification vectors were processed using a feature extractor. The process is performed to obtain enhanced news-rational data after feature extraction. The calculation process is as follows: in, This represents the weight matrix of the linear layer. This is a bias term.

8. The method according to claim 1, characterized in that, The news feature representation is fused with various enhanced news-justification features. A dynamic gating mechanism adaptively adjusts the weights of each justification feature to obtain the final feature representation. Classification is then performed based on this final feature representation to obtain news authenticity prediction results, including: News key feature vectors And enhanced news-justification feature vector The input is fed into the fusion unit to obtain the fused features. : Fusion features The final feature vector is obtained by inputting the dynamic gating unit. : in, and For linear layer parameters, and It is a bias term, and ; The final feature vector The input is fed into a multilayer perceptron classifier, where a sigmoid activation function is used to obtain the prediction result, and the classification loss is calculated using the binary cross-entropy loss function. The total loss function is a weighted sum of classification loss, rationale prediction loss, and rationale evaluation loss: in, and It's a hyperparameter.

9. A fake news detection system based on multi-dimensional perspective reasoning fusion, characterized in that, include: Data preprocessing module: Used to preprocess the news text to be detected, generate multi-dimensional analysis of the rationale for each news item using a large language model, and generate prediction results based on each perspective to build a news rationale dataset; Encoder module: Used to semantically encode the news text and the arguments from various perspectives, generating corresponding semantic representations; News-Reasoning Interaction Module: This module uses a dual cross-attention mechanism to interact each reasoning semantic representation with the news semantic representation, thereby obtaining supplementary features of reasoning to news content and guiding features of news to reasoning analysis. Reasoning Prediction Module: Used to independently predict the evaluation results of news credibility by the large language model based on the semantic representation of reasoning from various perspectives; The rationale evaluation module is used to calculate the validity weight of each dimension of the rationale analysis based on the guiding features of the news on the rationale analysis, use the validity weight to weight the supplementary features of the rationale to the news content, and extract features from the weighted features to obtain enhanced news-rational features. Dynamic gating module: It is used to fuse news feature representation with various enhanced news-justification features, and adaptively adjust the weight of each justification feature through dynamic gating mechanism to obtain the final feature representation. Based on the final feature representation, classification is performed to obtain the news authenticity prediction result.

Images