Prompt learning and fuzzy label-based multi-field false news detection model detection method and system

Abstract

The invention belongs to the technical field of false news detection, and discloses a multi-field false news detection model detection method based on prompt learning and fuzzy tag.According to the model, a false news detection task is reconstructed into a blank filling problem, field information is embedded into a prompt template, and the false news detection efficiency is improved. And the potential knowledge of the pre-training language model is fully utilized to generate field enhanced representation for the original news text. Meanwhile, a fuzzy domain label is generated through a neural network fitting fuzzy reasoning process, guidance is provided for subsequent hybrid experts in a feature extraction process, semantic fine-grained features of news across domains are effectively captured, and the overall performance of the model in multi-domain false news detection is remarkably improved through the process. On a Weibo21 data set with domain annotation data and a Thu data set without annotation data, the overall performance of the model is better than that of a traditional baseline model and an existing advanced multi-domain false news detection method, and the accuracy and generalization ability of the false news detection model are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of fake news detection, and in particular relates to a multi-domain fake news detection model detection method and system based on prompt learning and fuzzy labels. Background Art

[0002] In today's digital age, the rapid development of the internet and social media has dramatically changed how people access information. However, this has also created a fertile ground for the proliferation of fake news. False news has spread widely across a wide range of sectors, including politics, economics, technology, health, and entertainment, negatively impacting public perception and social order. The World Health Organization (WHO) notes that false news related to healthcare has directly impacted public trust in vaccines, leading to a drop of over 30% in vaccination rates in some regions. In the political sphere, false news often contains inflammatory content. For example, approximately 19% of election-related tweets on Twitter have been labeled as misleading, and this false content has seriously disrupted the electoral process. In the economic sphere, false news can trigger market volatility and harm investor interests. Therefore, curbing the spread of false news has become a pressing issue. However, the diversity and timeliness of false news make it difficult to effectively detect.

[0003] In recent years, breakthroughs in artificial intelligence and natural language processing technologies have ushered in a new research paradigm in the field of fake news detection. Throughout this technological evolution, fake news detection methods have undergone two major stages of development. Early research primarily employed traditional machine learning methods, using artificial feature engineering to model a combination of news content (lexical, grammatical, and semantic aspects) and social context (communication structure, user behavior, etc.). However, when dealing with massive amounts of online false information disseminated in real time, these methods generally suffer from significant issues such as limited feature representation capabilities and insufficient model generalization. Advances in deep learning technology, particularly breakthroughs in pre-trained language models (PLMs), have provided new solutions for text representation learning. Through multi-stage pre-training strategies (such as masked language modeling) and large-scale corpus training, these models can automatically capture deep semantic features and contextual associations, significantly improving the richness and discriminability of text representations. It is worth noting that researchers are exploring multimodal fusion detection frameworks. For example, by integrating the local feature extraction capabilities of convolutional neural networks (CNNs), the sequence modeling advantages of long short-term memory networks (LSTMs), and the deep semantic understanding of the BERT model, researchers achieved higher F1 scores and accuracy on the Covid-19 fake news detection dataset. Although deep learning technology has significantly promoted the progress of fake news detection research, the subject still faces two core issues: cross-domain adaptability and poor performance in the absence of labeled data.

[0004] Through the above analysis, the problems and defects of the existing technology are as follows:

[0005] (1) Disinformation exhibits significant differences in language representation and propagation patterns across different vertical domains. This domain specificity leads to serious cross-domain generalization issues for detection models trained on a single domain. Models such as BERT-BiGRU and MMCNN fail to capture the subtle variations in semantics across different domains and typically treat domain labels as fixed classification labels rather than probability distribution models. Although they can achieve good results within a single domain, their performance degrades significantly when transferred to other domains.

[0006] (2) Current mainstream multi-domain datasets generally suffer from discretization of domain coverage and imbalanced sample distribution. The detection accuracy of existing deep learning models in few-sample domains is significantly lower than that in full-sample training, which exposes the high dependence of current methods on data size. In scenarios without domain data labels (such as the Thu dataset), existing methods have difficulty in performing domain-aware learning.

[0007] (3) How to build a domain-adaptive feature representation framework and realize fake news detection in small sample scenarios has become a key issue that needs to be broken through in this field.

[0008] (4) Compared with single-domain data, the accuracy of existing fake news detection models decreases when facing cross-domain data.

[0009] The significance of solving the above problems and defects is:

[0010] The present invention can limit the spread of false news on social network media, reduce the negative impact of false news on the public, maintain social order, protect public interests, and promote the healthy development of the Internet. Summary of the Invention

[0011] In response to the problems existing in the prior art, the present invention provides a multi-domain fake news detection model detection method based on prompt learning and fuzzy labels.

[0012] The present invention is implemented as follows: a multi-domain fake news detection model detection method based on prompt learning and fuzzy labels includes:

[0013] Step 1: Build original news and alert templates;

[0014] Step 2: Perform domain enhancement on the original news;

[0015] Step 3: Use membership functions to map news into a probability distribution and quantify its degree of membership in various fields;

[0016] Step 4: Use hybrid expert system to extract features;

[0017] Step 5: Adjust parameters, train the model, and learn and classify the extracted features.

[0018] Furthermore, the step 1 includes:

[0019] In the fake news detection task, for a given news text, when constructing a prompt template, first add special tags [CLS] and [SEP] to the text sequence, where [CLS] indicates the beginning of the sequence and [SEP] indicates the end of the sequence. Then, append the task prompt to classify the field to which the news belongs after the text; for example: This news belongs to the () field;

[0020] The powerful contextual semantic understanding ability of the pre-trained language model is used to fill in the blanks in the mask and predict the field to which the news belongs.

[0021] Furthermore, in step 2, the original news and prompt templates are first segmented to obtain a vocabulary sequence, and the vocabulary sequence is input into the RoBERTa pre-trained model to obtain a hidden state sequence containing the news context semantics; then, the model initializes a domain embedding vector, which is used to represent the annotated domain labels in the dataset.

[0022] Furthermore, in step three, the membership function assigns a membership level between zero and one to each news field, mapping it to a probability distribution, thereby quantifying its degree of membership in each field; the membership function includes three parts: 1) sequence feature extraction: using the gated recurrent unit (GRU) to perform sequence modeling on the text embedding of the input news to capture the temporal dependencies in the text; 2) nonlinear mapping: the output of the GRU is nonlinearly transformed through a multi-layer perceptron (MLP) to further extract high-level semantic features; 3) probability normalization: finally, the output is mapped to a probability distribution, i.e., a fuzzy field label, through the softmax layer.

[0023] Furthermore, in step 4, each expert network is a text convolutional neural network (Text-CNN), and each expert focuses on extracting features of news in a specific field. Nine experts from different fields capture the fine-grained features of news in different fields, and under the guidance of the weights of the membership function, the contributions of each expert are dynamically allocated according to the probability distribution of the news, thereby fusing a comprehensive feature representation to improve the performance of the multi-field fake news detection model.

[0024] Furthermore, in step five, the comprehensive feature representation obtained in the previous step is used as the input of the classifier, passes through a fully connected layer, and uses the sigmoid activation function to obtain a model prediction probability between 0 and 1. The larger the value, the more likely the model is to judge the news to be false; (category 1 is set to represent false news and 0 is set to represent true news); during the training process, the parameters of all modules are jointly optimized through the backpropagation algorithm, with the goal of minimizing the total loss, so that the model can more accurately capture the semantic information and domain characteristics of the news, and ultimately judge its authenticity.

[0025] Another object of the present invention is to provide a multi-domain fake news detection model detection system based on prompt learning and fuzzy labels, comprising:

[0026] Building blocks for constructing original news and alert templates;

[0027] Enhancement module, used to perform domain-enhanced representation of original news;

[0028] The quantification module is used to map news into a probability distribution using membership functions and quantify its degree of membership in various fields;

[0029] Feature extraction module, used for feature extraction using hybrid expert system;

[0030] The prediction module is used to adjust parameters, train models, and learn and classify the extracted features.

[0031] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the multi-domain fake news detection model detection method based on prompt learning and fuzzy labels.

[0032] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the multi-domain fake news detection model detection method based on prompt learning and fuzzy labels.

[0033] Another object of the present invention is to provide an information data processing terminal, which is used to implement the multi-domain fake news detection model detection system based on prompt learning and fuzzy labels.

[0034] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0035] First, the present invention proposes a multi-domain fake news detection model that integrates prompt learning and fuzzy sets. The model reconstructs the fake news detection task into a fill-in-the-blank problem and embeds domain information in the prompt template, making full use of the potential knowledge of the pre-trained language model to generate domain-enhanced representations for the original news text. At the same time, fuzzy domain labels are generated by fitting the fuzzy reasoning process through a neural network, providing guidance for subsequent hybrid experts in the feature extraction process, effectively capturing the fine-grained semantic features of news across domains. This process significantly improves the overall performance of the model in multi-domain fake news detection. Experimental results show that on the Weibo21 dataset with domain-labeled data and the Thu dataset without labeled data, the model proposed by the present invention outperforms the traditional baseline model and existing advanced multi-domain fake news detection methods in overall performance, significantly improving the accuracy and generalization ability of the fake news detection model. The specific advantages of the present invention are as follows:

[0036] (1) We deeply analyzed the problems existing in the current fake news detection task and designed a fake news domain information prompt learning paradigm, which was concatenated with the original news text and then input into the pre-trained RoBERTa encoder to more effectively extract the rich semantic features and domain characteristics contained in the news text.

[0037] (2) We evaluated and verified the limitations of the existing multi-domain fake news detection model, introduced the concept of fuzzy sets into the fake news detection model, labeled the multiple domains that each news piece may involve, and dynamically weighted its output, so as to more effectively capture the differences between news in different domains, realize a fake news detection method without explicit domain labels, and improve the detection performance of the model in the case of few samples and unlabeled data.

[0038] (3) For the first time, the prompt learning paradigm is combined with fuzzy set theory to construct a cross-domain soft labeling mechanism for fake news.

[0039] The model was simulated on the Weibo21 dataset annotated with domain labels and the Thu dataset without labels. The experimental results showed that the F1 score was improved by an average of 1.40 percentage points compared with the current advanced methods, verifying its feasibility in multi-domain fake news detection tasks.

[0040] Second, as auxiliary evidence for the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

[0041] (1) The expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0042] 1. Content platforms (such as Weibo, Zhihu, Douyin, Toutiao, etc.): Improve the credibility of platform content and reduce public opinion and legal risks caused by the spread of false information.

[0043] 2. Government agencies and media organizations: This technology has application prospects in combating rumors and maintaining public safety, and is particularly useful in epidemic prevention and control, election supervision, and public event analysis.

[0044] 3. Financial and insurance companies: Automatically identify false financial news and manipulated market information, prevent investment fraud or financial misleading, and improve risk warning capabilities.

[0045] (2) The technical solution of the present invention fills the technical gap in the industry at home and abroad:

[0046] Currently, there are no published patents systematically introducing the integration of prompt learning and fuzzy sets for multi-domain fake news detection. This invention, through prompt learning and fuzzy labeling, effectively enhances the model's ability to understand and integrate features from different domains. Experimental results on multiple public datasets demonstrate superior accuracy and F1 scores compared to existing models. This invention maintains high performance, particularly in cross-domain detection and low-resource scenarios without domain labels, addressing the bottlenecks of traditional models, namely poor transferability and the need for large numbers of labels.

[0047] (3) Whether the technical solution of the present invention solves the technical problems that people have been eager to solve but have not been able to solve successfully:

[0048] This invention can be embedded in news platforms, social networking platforms, and other systems to monitor and identify disinformation in multiple fields in real time, assisting with intelligent upgrades to content review and risk control systems. Particularly during emergencies (such as epidemics, disasters, and political events), this invention's rapid adaptability can help enhance the platform's credibility and mitigate misleading public opinion.

[0049] (4) Whether the technical solution of the present invention overcomes technical prejudice:

[0050] This paper transforms fake news detection into a fill-in-the-blank task, embedding domain information with a prompt template to guide the pre-trained language model to actively perceive domain semantics. This allows the model to retain semantic context consistency when migrating between different domains, reducing overfitting to single-domain semantics and significantly improving cross-domain adaptability. This overcomes the domain bias of other models for fake news detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of a multi-domain fake news detection model detection method based on prompt learning and fuzzy labels provided by an embodiment of the present invention.

[0052] Figure 2 This is a structural block diagram of a multi-domain fake news detection model detection system based on prompt learning and fuzzy labels provided by an embodiment of the present invention.

[0053] Figure 3 This is a flowchart of multi-domain fake news detection based on prompt learning and fuzzy labels provided by an embodiment of the present invention.

[0054] Figure 4 This is a principle diagram of the membership function provided by an embodiment of the present invention.

[0055] Figure 5 Distribution map of real news and fake news topic types.

[0056] Figure 6 Performance comparison of multiple models in the fake news detection task.

[0057] Figure 7 Schematic diagram of the inference path for false news discrimination based on semantic matching. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0059] like Figure 1 As shown, the embodiment of the present invention provides a multi-domain fake news detection model detection method based on prompt learning and fuzzy labels, which includes the following steps:

[0060] S101: Build original news and alert templates;

[0061] S102: Perform domain enhancement on the original news;

[0062] S103: Use membership functions to map news into a probability distribution and quantify its degree of membership in various fields;

[0063] S104, feature extraction using a hybrid expert system;

[0064] S105, adjust parameters, train the model, and learn and classify the extracted features.

[0065] S101 provided in this embodiment of the present invention includes:

[0066] In the fake news detection task, for a given news text, when constructing a prompt template, first add special tags [CLS] and [SEP] to the text sequence, where [CLS] indicates the beginning of the sequence and [SEP] indicates the end of the sequence. Then, append the task prompt to classify the field to which the news belongs after the text; for example: This news belongs to the () field;

[0067] The powerful contextual semantic understanding ability of the pre-trained language model is used to fill in the blanks in the mask and predict the field to which the news belongs.

[0068] In S102 provided by an embodiment of the present invention, the original news and prompt templates are first segmented to obtain a vocabulary sequence, and the vocabulary sequence is input into the RoBERTa pre-trained model to obtain a hidden state sequence containing the news context semantics; then, the model initializes a domain embedding vector, which is used to represent the annotated domain labels in the dataset.

[0069] In S103 provided by the embodiment of the present invention, the membership function assigns a membership level between zero and one to each news field, maps it to a probability distribution, and thus quantifies its degree of membership in each field; the membership function includes three parts: 1) sequence feature extraction: using a gated recurrent unit (GRU) to perform sequence modeling on the text embedding of the input news to capture the temporal dependency in the text; 2) nonlinear mapping: the output of the GRU is nonlinearly transformed through a multi-layer perceptron (MLP) to further extract high-level semantic features; 3) probability normalization: finally, the output is mapped to a probability distribution, i.e., a fuzzy field label, through a softmax layer.

[0070] In S104 provided by the embodiment of the present invention, each expert network is a text convolutional neural network (Text-CNN), and each expert focuses on extracting features of news in a specific field. Nine experts in different fields respectively capture the fine-grained features of news in different fields, and under the guidance of the weights of the membership function, the contributions of each expert are dynamically allocated according to the probability distribution of the news, thereby fusing a comprehensive feature representation to improve the performance of the multi-field fake news detection model.

[0071] In S105 provided by the embodiment of the present invention, the comprehensive feature representation obtained in the previous step is used as the input of the classifier, passes through a fully connected layer, and uses the sigmoid activation function to obtain a model prediction probability between 0 and 1. The larger the value, the more likely the model is to judge the news as false; (category 1 is set to represent false news and 0 is set to represent true news); during the training process, the parameters of all modules are jointly optimized through the back propagation algorithm, with the goal of minimizing the total loss, so that the model can more accurately capture the semantic information and domain characteristics of the news, and ultimately judge its authenticity.

[0072] like Figure 2 As shown, an embodiment of the present invention provides a multi-domain fake news detection model detection system based on prompt learning and fuzzy labels, including:

[0073] Building blocks for constructing original news and alert templates;

[0074] Enhancement module, used to perform domain-enhanced representation of original news;

[0075] The quantification module is used to map news into a probability distribution using membership functions and quantify its degree of membership in various fields;

[0076] Feature extraction module, used for feature extraction using hybrid expert system;

[0077] The prediction module is used to adjust parameters, train models, and learn and classify the extracted features.

[0078] Existing cross-domain fake news detection methods often rely on data distribution in a single domain or a fixed labeling system, resulting in a lack of sufficient generalization capabilities when faced with diverse news content. First, traditional supervised learning models are severely affected by domain bias when migrating between different domains, making it difficult to account for the heterogeneity of features in each domain. Second, the hard label (one-hot label) format cannot characterize the ambiguous attribution of news between different domains and is prone to introducing label noise. Third, the lack of effective integration and utilization of domain information makes the detection model insufficiently robust against the complex and ever-changing news contexts in the real world. These deficiencies directly affect the detection performance of the model in new or cross-domain scenarios, and make it difficult to adapt to the diversity and dynamic evolution of news corpora in actual applications.

[0079] In order to solve the above problems, the present invention introduces the idea of ​​prompt learning, which embeds domain prior information into the model input in a templated form. The construction module fuses the original news text with domain keywords by designing a universal and flexible prompt template, and then uses the context interaction capability of the pre-trained language model to realize the joint encoding of news semantics and domain clues. In this process, the prompt template not only contains question-based guidance on the statement of news facts, but also incorporates label descriptions related to each target field, such as "finance", "entertainment", "technology", etc., so that the pre-trained model can automatically explore the common and different features between fields in the fine-tuning stage, thereby reducing the representation gap during domain migration.

[0080] On this basis, the enhancement module further performs domain-enhanced representation on the original news to alleviate potential domain bias in the corpus. Specifically, the module uses domain-adaptive embedding or adversarial training mechanisms to inject information fragments from the same or similar domains into the news text to generate a number of semantically expanded samples; at the same time, for each news item, it constructs multiple "domain hybrid views", such as encoding the news title under the "technology" and "finance" prompt templates respectively, and then integrating the outputs of each view through attention fusion or residual fusion strategies to obtain a more robust multi-domain semantic representation. In this way, the model can not only capture the news content in the original semantic space, but also adaptively learn domain-related features in the elastic space, making it more generalizable in the subsequent quantification and classification processes.

[0081] The quantification module introduces a membership function to generate fuzzy labels, addressing the fact that news items do not necessarily belong to the same domain. Using predefined membership mapping rules, the domain strength of each news item is mapped to a probability distribution vector in the interval [0, 1], representing the news item's degree of belonging to each domain. For example, if a news item has both financial and political attributes, its membership in the "finance" domain might be 0.7 and its membership in the "politics" domain might be 0.3. This allows the model to balance the subjective domain characteristics of news items during training while retaining a certain degree of ambiguity, avoiding label bias caused by artificial classification. Membership functions can be designed using Gaussian, triangular, or Bell membership functions. Membership values ​​are derived based on the similarity between news items and domain keywords, dynamically providing soft label guidance for subsequent feature extraction.

[0082] The feature extraction module uses a mixture of experts (MoE) to achieve deep fusion of multi-domain features. This module consists of several domain expert sub-networks and a routing gating network. Each expert sub-network focuses on capturing the deep semantics and feature patterns of its corresponding domain. The gating network dynamically assigns weights to the outputs of each expert based on the enhanced representation and fuzzy label distribution of the input news, achieving selective integration of feature context. In layman's terms, when a piece of news has both scientific and entertainment attributes, the gating network automatically adjusts the activation ratio of "science and technology experts" and "entertainment experts" based on the quantized membership probability, so that the final extracted features contain both deep semantics in the scientific and technological fields and surface expressions in the entertainment field. This hybrid expert architecture not only preserves the expertise of experts in each field, but also enables flexible interaction of cross-domain information through gating strategies, improving the model's discriminative efficiency and accuracy.

[0083] After feature extraction, the prediction module classifies and optimizes the parameters of the extracted multi-domain hybrid features. The module first dimensionalizes and normalizes the hybrid expert output, and then connects to a series of gradient descent-based optimization processes to jointly train the parameters of the feedforward network and the gated network. The training objective not only includes the traditional cross-entropy loss, but also introduces fuzzy label consistency loss to ensure that the probability distribution of the model output is as aligned as possible with the soft labels generated by the membership function. In addition, in order to enhance the stability under adversarial samples, an adversarial loss term can be introduced during the training process, so that the model can still maintain good classification performance for inputs containing minor perturbations. After training is completed, the prediction module can use the learned weights and classification thresholds to perform real-time in-domain and out-of-domain fake news discrimination on the news input in the test phase, and output the authenticity score and final binary classification decision for each domain.

[0084] The entire system can be integrated into a general-purpose computing device and connected to upper-level applications through a unified API interface. The computer device includes a standard processor and storage unit, which can run a deep learning framework and is responsible for the management of prompt templates, corpus preprocessing, model training and inference deployment; computer-readable storage media (such as SSD or cloud storage) are used to save pre-trained models, prompt template sets, membership function parameters and trained multi-domain hybrid expert networks; the information data processing terminal can be used as a front-end access layer to achieve real-time collection, preprocessing and visual display of news data. In this way, the present invention not only solves the noise problems caused by poor cross-domain generalization and label binarization in the existing technology, but also ensures the scalability and maintainability of the system through modular design and professional terminology description.

[0085] The present invention is specifically implemented:

[0086] 1. The prompt learning paradigm constructs a template containing the original news text, then sets the task prompt and masks the keywords in the task, allowing the pre-trained language model to predict the masked parts of the text. In the fake news detection task, for a given news text X, when constructing the prompt template, first add special tags [CLS] and [SEP] to the text sequence, where [CLS] indicates the beginning of the sequence and [SEP] indicates the end of the sequence. Then, the task prompt for classifying the field to which the news belongs is appended to the text. Its expression is as follows:

[0087] X=[[CLS],X1,X2,...,X n .This news may belong to the [DOMAIN] domain. [SEP]] (1)

[0088] Among them, [DOMAIN] is a domain placeholder, and its representation will be updated later through the domain enhanced representation module.

[0089] 2. First, perform word segmentation on the prompt template to obtain the vocabulary sequence Tokens = [[CLS], token1, token2, ..., token n , Domain, [SEP]]. Then, the vocabulary sequence is input into the RoBERTa pre-training model to obtain the hidden state sequence H x , whose expression is as follows:

[0090]

[0091] Next, the model initializes a domain embedding vector, which is used to represent the annotated domain labels in the dataset. In order to effectively integrate global text information with domain labels, this model introduces an attention mechanism to calculate the correlation between domain information and global representation:

[0092]

[0093] Among them, d k is the scaling factor of the vector dimension; α is the attention weight. Then, the fusion vector h that integrates the global semantic information of the text and the domain prompt information is obtained through weighted summation. α , whose expression is as follows:

[0094] h α =αh d (4)

[0095] For the [DOMAIN] placeholder in the original prompt template, its initial embedding is represented as Combine the fusion vector h obtained in the above steps α , get the updated field representation:

[0096]

[0097] Updated h [DOMAIN] The representation carries global information and domain-related knowledge from the original text, so that the prompt template can better reflect the domain characteristics when it is subsequently encoded through the pre-trained model. After replacing the placeholder in the original template with the updated [DOMAIN] representation, the text embedding representation is obtained as follows:

[0098] H′ x =[h [CLS] ,h1,...h n ,h [DOMAIN] ,h [SEP] ] (6)

[0099] At this point, the sequence input into the feature extraction module not only contains the semantic information of the news text, but also embeds dynamically updated domain information, thereby providing a more domain-aware contextual representation for subsequent fake news detection.

[0100] 3. This paper introduces the concept of fuzzy sets to better describe the fields to which news belongs. Specifically, the membership function assigns a membership level between zero and one to each object, mapping the news into a probability distribution, thereby quantifying its degree of membership in various fields. The specific framework is as follows: Figure 3 As shown:

[0101] Figure 4 The membership function consists of three parts: 1) Sequence feature extraction: Using a gated recurrent unit (GRU) to perform sequence modeling on the text embedding of the input news, capturing the temporal dependencies in the text. 2) Nonlinear mapping: The GRU output is nonlinearly transformed through a multi-layer perceptron (MLP) to further extract high-level semantic features. 3) Probability normalization: Finally, the softmax layer maps the output to a probability distribution, i.e., a fuzzy domain label. Its working principle is as follows:

[0102] The word embedding matrix W = [W[CLS], W1, W2, ..., Wn, W[SEP]] obtained by the original article through the RoBERTa pre-training model

[0103] And input it into GRU to obtain the hidden sequence state, which is expressed as follows:

[0104] h=GRU(W;θ1)(7)

[0105] Here, θ1 represents the parameters of the GRU, and h is a vector representation that captures the global semantic information of the text. Next, h is input into the MLP, and a 9-dimensional vector is obtained through nonlinear mapping:

[0106] z=MLP(h;θ2)(8)

[0107] Among them, θ2 represents the parameters of MLP, z∈R K is the unnormalized domain score.

[0108] To obtain the fuzzy domain label, z is normalized by softmax, and its expression is as follows:

[0109]

[0110] Among them, g represents a probability distribution, reflecting the membership of the news in each candidate field.

[0111] Combining the above steps, the membership function can be expressed as:

[0112] g=M(W; θ)=softmax(MLP(GRU(W; θ1); θ2)) (10)

[0113] 4. Each expert network in the hybrid expert system is a text convolutional neural network (Text-CNN), which focuses on extracting features from a specific domain or type. Nine experts from different fields capture fine-grained features of news text in different areas. Guided by the weights of the membership functions, a domain gating mechanism dynamically allocates the contributions of each expert based on the probability distribution of news, thereby fusing them into a comprehensive feature representation, thereby improving the effectiveness of cross-domain fake news detection. The specific process is as follows:

[0114] For the hidden state sequence obtained by the RoBERTa pre-training model, there are T experts who map the input into a feature vector, which is expressed as follows:

[0115] f i =E i (H′ x ;φ i )=TextCNN i (H′ x ;φ i ) (11)

[0116] in, represents the parameters of the i-th expert network. The domain gating module uses the membership level of the news to assign a weight to each expert. The membership function is used as input, passed through a multi-layer perceptron and normalized to obtain the weight vector of each expert. Its expression is as follows:

[0117]

[0118] Where φ is the parameter of the domain gating module and satisfies And each α i ≥ 0. Finally, the model performs weighted fusion based on the feature outputs of each expert and the corresponding weights to obtain the final comprehensive feature representation, which is expressed as follows:

[0119]

[0120] 5. The comprehensive feature representation F obtained by the feature extraction module is used as the input of the classifier. After passing through a fully connected layer, the sigmoid activation function is used to obtain the model's predicted probability y, which is expressed as follows:

[0121]

[0122] The output y is between 0 and 1. The larger the value, the more likely the model is to judge the news as false (category 1 represents false news and 0 represents true news).

[0123] This model uses a binary cross entropy loss function to measure the gap between the predicted result and the true label. For each sample in the training set, the loss function is defined as:

[0124]

[0125] Where N is the total number of training samples.

[0126] During the training process, the parameters of all modules are jointly optimized through the back-propagation algorithm, with the goal of minimizing the total loss L, so that the model can more accurately capture the semantic information and domain characteristics of the news and ultimately judge its authenticity.

[0127] This experiment mainly uses two data sets for experimental verification:

[0128] Weibo21 is a Chinese multi-domain fake news dataset, containing news from nine fields: politics, society, education, health, finance, technology, military, disasters, and entertainment. The dataset contains 4,640 real news items and 4,488 fake news items, each with a manually annotated domain label.

[0129] The Thu dataset is a Chinese fake news dataset released by the Tsinghua Natural Language Processing Laboratory. It is a partial merger of the CEO_Dataset and the rumors dataset. It contains 1,849 real news items and 1,560 fake news items without domain label annotations. This paper uses this dataset to test the model's detection performance and domain knowledge transfer capabilities in the absence of labeled data.

[0130] This experiment used the pretrained RoBERTa model as a text feature extractor, with word embedding dimensions set to 1024 and a maximum sentence length of 170. The Adam optimizer was used for training, with a learning rate of 5e-4. The model was trained for 40 epochs. The experiment was conducted five times and the average F1 score was taken. To test the model's performance, this experiment used four evaluation metrics: accuracy (Acc), precision (P), recall (R), and F1 value. The specific calculation formulas are as follows:

[0131] (1) Accuracy: This represents the proportion of correct predictions made by the model among all predictions. Although simple, accuracy may mask the model’s prediction performance for minority classes when positive and negative samples are unbalanced.

[0132]

[0133] (2) Precision: This is mainly used to measure the accuracy of the model in predicting “false” information. A high precision means fewer false positives in false information identification.

[0134]

[0135] (3) Recall: This measures the model’s ability to identify actual “fake” information samples. A high recall indicates that the model can capture more false information, but may increase false positives.

[0136]

[0137] (4) F1 value: The harmonic mean of precision and recall, which can comprehensively evaluate the performance of the model in balancing precision and recall, and is particularly suitable for data imbalance scenarios.

[0138]

[0139] 1. Specific application fields or related products of the present invention.

[0140] 1. Content review on social media platforms

[0141] Applicable to information content risk control systems on platforms such as Weibo, Zhihu, Douyin, Kuaishou, and WeChat official accounts. It enables real-time false news identification, automatic blocking, and labeling or user notifications.

[0142] 2. Content screening system for news media organizations

[0143] This technology is used by mainstream news publishers and information aggregation platforms (e.g., Toutiao, Tencent News, and The Paper). It pre-screens suspicious content and verifies its domain labels and authenticity before it is pushed. This improves the credibility of information released on the platform and mitigates the reputational and legal risks associated with the spread of misinformation.

[0144] 3. Public opinion monitoring and crisis warning by government / regulatory agencies

[0145] This system is suitable for government agencies such as the Cyberspace Administration of China, market supervision departments, and public security / propaganda departments to monitor internet public opinion and provide early warnings about rumors. It can automatically aggregate, categorize, and identify false online statements; it is used for analyzing government public opinion and generating early warning reports.

[0146] 2. Relevant evidence of the technical effects obtained by the embodiments of the present invention.

[0147] 1 Experimental Environment

[0148] In terms of hardware configuration, this paper uses a Xeon(R) Platinum 8362 processor and an NVIDIA GeForce RTX3090 graphics card. In terms of software environment, the operating system is Ubuntu 18.04, the deep learning framework uses PyTorch 1.8.0, and all code is developed based on Python 3.8.

[0149] 2 Datasets

[0150] This paper mainly uses two data sets for experimental verification:

[0151] (1) Weibo21 dataset

[0152] Weibo21 is a Chinese multi-domain fake news dataset that contains news from nine fields: politics, society, education, health, finance, technology, military, disaster, and entertainment. The dataset contains 4,640 real news items and 4,488 fake news items, each with a manually annotated domain label. This dataset is used to verify the model's detection performance and cross-domain feature extraction capabilities in scenarios with labeled data. The specific data is as follows: Figure 5 As shown:

[0153] (2) Thu dataset

[0154] The Thu dataset is a Chinese fake news dataset released by the Tsinghua Natural Language Processing Laboratory. It is a partial merger of the CEO_Dataset and the rumors dataset. It contains 1,849 real news items and 1,560 fake news items without domain label annotations. This paper uses this dataset to test the model's detection performance and domain knowledge transfer capabilities in the absence of labeled data.

[0155] 3 Experimental setup

[0156] This paper uses a pretrained RoBERTa model as a text feature extractor, with a word embedding dimension of 1024 and a maximum sentence length of 170. The Adam optimizer is used for training, with a learning rate of 5e-4. The model is trained for 40 epochs. Experiment 1 is conducted five times and the average F1 score is taken. The specific parameters for PLFuFND model training are shown in Table 1.

[0157] Table 1. PLMFEND training parameters

[0158]

[0159]

[0160] 4.5 Comparative Experiment

[0161] 4.5.1 Experimental Results on Weibo21 Dataset

[0162] Table 2 shows the experimental results of the PLMFEND model in this paper and various baseline models trained and tested using all labeled data in the Weibo21 dataset, comprehensively evaluating the performance of the model in the fake news detection task with labeled data.

[0163] Table 2. Experimental results on Weibo21 dataset (F1 score)

[0164]

[0165]

[0166] 4.5.2 Experimental results on the Thu dataset

[0167] Figure 6 The experimental results of the PLFuFND model proposed in this paper and various baseline models are shown on the Thu dataset without any labeled data, directly using pre-trained models and prompt templates to detect false information.

[0168] From Table 2 and Figure 6 As can be seen, the proposed PLFuFND model outperforms baseline models on both the Weibo21 and Thu datasets. On the Weibo21 dataset, PLFuFND achieves higher F1 scores than all baseline models in six domains: science, military, education exams, disasters and accidents, politics, and finance and commerce. Its overall F1 score improves by 1.40 percentage points over existing state-of-the-art models. However, in the three domains of medicine and health, entertainment, and social life, the high similarity of news content and the repetitive use of domain-specific vocabulary in these domains cause the model to generate domain embeddings that do not match the true labels during the prompt learning process, resulting in poor overall performance and requiring further improvement in future research. On the Thu dataset, PLFuFND achieves higher precision, recall, and accuracy than other multi-domain fake news detection models, demonstrating the effectiveness of the prompt learning template in improving the performance of fake news detection tasks by providing domain embeddings and representation enhancement for news text in the absence of labeled data.

[0169] Comprehensive analysis shows that PLFuFND outperforms existing baseline models in overall performance on both domain-labeled data and unlabeled data, verifying the effectiveness of hint learning in multi-domain fake news detection tasks; and the combination of hint learning and fuzzy sets better captures the semantic relationship between news in different domains, improves the generalization ability of the model, and achieves more effective fake news detection.

[0170] 4.6 Ablation Learning

[0171] To verify the effectiveness of each module in this experiment, this paper conducted an ablation experiment on the Weibo21 dataset, which mainly includes the following configurations:

[0172] (1) PLFuFND w / o-Fu removes the membership function and the weight assignment guidance for each expert, that is, removes the concept of fuzzy set and only uses the single domain label generated by the prompt template for testing.

[0173] (2) PLFuFND w / o-expert removes the hybrid expert system and uses the pre-trained RoBERTa model to directly extract the features of fake news.

[0174] (3) PLFuFND w / o-PL removes the hint learning module, removes the domain information embedding and representation enhancement process, and directly extracts features from the original news text and conducts testing.

[0175] Table 3. Impact of each module on PLFuFND (F1 score)

[0176]

[0177]

[0178] As shown in Table 3, on the Weibo21 dataset, removing the membership function (Fu) or prompt learning (PL) resulted in a decrease in the model's F1 score across all domains. Removing the membership function and fuzzy labels significantly degraded model performance, with the total F1 score dropping by 4.04 percentage points, demonstrating the importance of fuzzy labels and weight distribution in the multi-domain fake news detection task. Removing the hybrid expert system resulted in a 0.341 percentage point decrease in the total F1 score, demonstrating the importance of the hybrid expert system in the multi-domain fake news detection task. Removing the prompt learning and domain embedding modules also significantly impacted model performance, with the total F1 score dropping by 2.51 percentage points, demonstrating the necessity of prompt learning and domain-enhanced representations. Ablation experiments further demonstrate the positive contribution of each module to the overall model performance and emphasize the importance of fuzzy set multi-domain labels and prompt learning in cross-domain fake news detection.

[0179] In summary, the overall performance of the complete PLFuFND model is better than all baseline models in various fields, which fully demonstrates the advantages of cross-domain feature extraction and the effectiveness of the overall model design in the multi-domain fake news detection task.

[0180] In order to more intuitively demonstrate the effect of prompt templates on improving the judgment of news authenticity, this paper selected three representative news texts for case analysis. Figure 7The relationship between news content, evidence text, and the output of the prompt learning model is shown. In the process of model authenticity judgment, the closer the output result is to 1, the more likely it is fake news, and the closer it is to 0, the more likely it is real news.

[0181] Through analysis Figure 7 As can be seen from the examples, the model first identifies the possible domains of the news content based on the prompt template, where the output words are limited to the prompt template by the nine domains divided by the Weibo21 dataset. During the evidence screening process, the model identified the keywords "nanoparticles" and "targeted therapy" and determined their domain to be "healthcare." However, due to the extensive use of medical terminology, these could easily mislead the public without professional medical knowledge, and thus were judged as false. For "limited-time stocks" and "limited-time redemption," the model determined their domain to be "finance and commerce." Due to the urgency of the content, which could induce public decision-making, these were judged as false. For "country 1 and country 2," the model determined their domain to be "politics," and the use of the possibility expression "close," thus judging their authenticity as true. In these three cases, the model's veracity was fully consistent with the authenticity of the news, validating the feasibility and accuracy of the proposed multi-domain fake news detection model.

[0182] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0183] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A multi-domain fake news detection method based on prompt learning and fuzzy labeling, characterized by: The method comprises the following steps: Step 1: Build original news and tips templates; Step 2: Perform domain enhancement on the original news; Step 3: Use membership functions to map news into probability distributions and quantify their membership in various fields. Step 4: Use hybrid expert system to extract features; Step 5: Adjust the parameters and train the model to learn and classify the extracted features.

2. A multi-domain fake news detection system based on prompt learning and fuzzy labels, characterized by: The system includes the following modules: Building blocks for constructing original news and alert templates; Enhancement module, used to perform domain-enhanced representation of original news; The quantification module is used to map news into probability distribution using membership functions and quantify its degree of membership in various fields; Feature extraction module, used for feature extraction using hybrid expert system; The prediction module is used to adjust parameters and train the model to learn and classify the extracted features.

3. A computer-executable instruction product, characterized in that: The instructions are stored in a non-volatile storage medium and, when executed, cause a computing device to execute the multi-domain fake news detection method based on prompt learning and fuzzy labels as claimed in claim 1.

4. A storage medium, characterized in that The device stores a program code capable of enabling a computing device to execute a multi-domain fake news detection method based on prompt learning and fuzzy labels, wherein the method includes the steps described in claim 1.

5. The detection method according to claim 1, wherein In the step 1, a special marker CLS is added to the text sequence as a sequence start marker and SEP is added as a sequence end marker, and a field classification prompt is appended after the text. The mask is filled in with a pre-trained language model to predict the field to which the news belongs.

6. The detection method according to claim 1, wherein In the second step, the original news and the prompt template are segmented to obtain a vocabulary sequence, and the vocabulary sequence is input into the RoBERTa pre-trained model to obtain a hidden state sequence containing the news context semantic information; The model then initializes the domain embedding vector to represent the annotated domain labels in the dataset.

7. The detection method according to claim 1, wherein In step 3, the membership function includes the following parts: In the sequence feature extraction stage, the gated recurrent unit is used to perform sequence modeling on news text embedding; In the nonlinear mapping stage, the output of the gated recurrent unit is nonlinearly mapped through a multilayer perceptron to extract high-level semantic features; In the probability normalization stage, the mapping results are mapped into probability distributions through normalization operations to form fuzzy domain labels.

8. The detection method according to claim 1, wherein In step 4, the hybrid expert system includes multiple expert networks, each of which is a text convolutional neural network, used to extract fine-grained features of news in a specific field; the system dynamically assigns output weights to each expert network based on the field affiliation probabilities obtained in step 3, and fuses them into a comprehensive feature representation.

9. The detection method according to claim 1, wherein In step five, the comprehensive feature representation is input into a classifier, which outputs a predicted probability through a fully connected layer and a sigmoid activation function, with the predicted probability being between zero and one. During the training process, the parameters of each module are jointly optimized through backpropagation, with the goal of minimizing the total loss to improve the accuracy of the model in multi-domain fake news detection tasks.

10. The detection system according to claim 2, wherein: The construction module is specifically used to add special tags CLS and SEP to the text sequence and append domain classification prompts to generate a prompt template and predict the domain label through the pre-training model; the enhancement module is specifically used to segment the original news and the prompt template and input the RoBERTa pre-training model to obtain the hidden state sequence and initialize the domain embedding vector; the quantization module is specifically used to use the gated recurrent unit and the multi-layer perceptron and the normalization operation to obtain the fuzzy domain label of the news; the feature extraction module is specifically used to construct multiple text convolutional neural network experts and weightedly fuse them according to the fuzzy domain labels to obtain a comprehensive feature representation; the prediction module is specifically used to input the comprehensive feature representation into the fully connected layer, output the prediction probability through the Sigmoid activation function, and perform parameter joint optimization training.

Images