False news detection method and device

Abstract

The invention provides a false news detection method, which comprises the following steps of: constructing a large model prompt template, performing field specific entity concept analysis on an input news text and a field label through a large language model, and outputting a reconstructed news text; respectively constructing a domain specific prompt template and a domain general prompt template, respectively inputting the reconstructed news text into the domain specific prompt template and the domain general prompt template, and extracting a domain specific feature vector and a domain general feature vector of the news text through a mask language model; evaluating a first similarity between the domain-specific feature vector and the domain-specific true news prototype and the domain-specific false news prototype and a second similarity between the domain-general feature vector and the domain-general true news prototype and the domain-general false news prototype through cosine similarity, and evaluating a news type by using the first similarity and the second similarity; and calculating a loss value based on a loss calculation formula for model performance optimization. According to the method, the false news detection accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of information detection technology, and in particular, to a method for detecting fake news. Background Art

[0002] Fake news mainly uses text and pictures as information carriers and spreads rapidly on social platforms through interactive methods such as user forwarding and sharing. The most widely studied method currently is the text information-based method. On this basis, a multi-modal fake news detection model can be constructed by integrating news picture information. There are usually certain differences in the propagation characteristics of true and false news. Researchers mine the propagation patterns of fake news by modeling the propagation network. The complex entities contained in news are likely to affect the model's understanding of the semantics of news content, and the method of using external knowledge to expand news information has begun to attract more and more scholars' attention.

[0003] Most existing fake news detection methods do not distinguish the differences in data distributions in different fields and are difficult to adapt to the diverse and rapidly evolving real news environment.

[0004] To address the problems of differences in news word frequencies and propagation patterns in different fields, MDFEND (i.e., Multi-domain fake news detection) introduces a domain gating mechanism to allocate and aggregate the feature representations extracted by several expert networks. To address the problems of differences in news word frequencies, emotions, and writing styles in different fields, and incomplete domain labels, M3FEND (i.e., Memory-guided multi-view multi-domain fake news detection) models three views of semantics, emotion, and writing style, and constructs a domain memory bank to mine the potential domain information of news, and aggregates the feature representations under different views through a domain adapter.

[0005] However, due to the lack of conceptual parsing of domain-specific entities, existing fake news detection methods are difficult to effectively model the complex dependencies between news words in different fields; the above-mentioned multi-domain fake news detection methods perform poorly under few-shot conditions.

[0006] In view of this, how to provide a fake news detection method to achieve fake news detection with strong multi-domain adaptability under few-shot conditions has become an urgent technical problem to be solved. Summary of the Invention

[0007] Embodiments of this application provide a method for detecting fake news, a device for detecting fake news, a computer device, and a computer storage medium, which are used to solve the problem that the multi-domain adaptability of fake news detection methods is poor under current few-shot conditions.

[0008] In the first aspect of the embodiments of this application, a method for detecting fake news is provided, including:

[0009] Construct a large model prompt template, input the obtained news text and domain labels into the large model prompt template, perform domain-specific entity concept parsing through a large language model, and output the reconstructed news text. Among them, the large model prompt template includes task instructions, input specifications, and output format requirements. The task instructions cover entity screening, concept parsing processing, and enhanced news reconstruction;

[0010] Construct a domain-specific prompt template and a domain-general prompt template respectively. Input the reconstructed news text into the domain-specific prompt template and the domain-general prompt template respectively, and extract the domain-specific feature vector and domain-general feature vector of the news text through a masked language model;

[0011] Evaluate the first similarity between the domain-specific feature vector and the domain-specific true news prototype and the domain-specific false news prototype, and the second similarity between the domain-general feature vector and the domain-general true news prototype and the domain-general false news prototype through cosine similarity. Use the first similarity and the second similarity, based on the model prediction probability calculation formula, calculate the prediction probability of the model for the news text, and evaluate the news type based on the prediction probability;

[0012] Calculate the loss value based on the loss calculation formula for model performance optimization.

[0013] In the second aspect of the embodiments of the present application, a fake news detection device is provided, including:

[0014] An output module, configured to construct a large model prompt template, input the obtained news text and domain labels into the large model prompt template, perform domain-specific entity concept parsing through a large language model, and output the reconstructed news text. Among them, the large model prompt template includes task instructions, input specifications, and output format requirements. The task instructions cover entity screening, concept parsing processing, and enhanced news reconstruction;

[0015] A generation module, configured to construct a domain-specific prompt template and a domain-general prompt template respectively. Input the reconstructed news text into the domain-specific prompt template and the domain-general prompt template respectively, and extract the domain-specific feature vector and domain-general feature vector of the news text through a masked language model;

[0016] A prediction module, configured to evaluate a first similarity between the domain-specific feature vector and domain-specific true-news prototypes and domain-specific false-news prototypes, and a second similarity between the domain-general feature vector and domain-general true-news prototypes and domain-general false-news prototypes by cosine similarity, use the first similarity and the second similarity, calculate a prediction probability of the model for the news text based on a model prediction probability calculation formula, and evaluate the news type based on the prediction probability;

[0017] A training module, configured to calculate a loss value based on a loss calculation formula for optimizing the model performance.

[0018] The present application provides a fake news detection method, including: constructing a large model prompt template, inputting the obtained news text and domain label into the large model prompt template, performing domain-specific entity concept parsing through a large language model, and outputting a reconstructed news text, where the large model prompt template includes a task instruction, an input specification, and an output format requirement, and the task instruction covers entity screening, concept parsing processing, and enhanced news reconstruction; respectively constructing a domain-specific prompt template and a domain-general prompt template, inputting the reconstructed news text into the domain-specific prompt template and the domain-general prompt template respectively, and extracting a domain-specific feature vector and a domain-general feature vector of the news text through a masked language model; evaluating a first similarity between the domain-specific feature vector and domain-specific true-news prototypes and domain-specific false-news prototypes, and a second similarity between the domain-general feature vector and domain-general true-news prototypes and domain-general false-news prototypes by cosine similarity, use the first similarity and the second similarity, calculate a prediction probability of the model for the news text based on a model prediction probability calculation formula, and evaluate the news type based on the prediction probability; calculating a loss value based on a loss calculation formula for optimizing the model performance.

[0019] Applying the fake news detection method provided by the embodiments of the present application, by using a large language model to parse domain-specific entity concepts to reconstruct news, the model can better understand news in each domain, helps to effectively capture complex dependency relationships between news words, enhances the news representation effect, and improves the accuracy of fake news detection; constructing domain-specific category prototypes and domain-general category prototypes, by evaluating the similarity between the news feature vector and the category prototype and updating the category prototype, the model can effectively identify fake news in multiple domains under few-shot conditions, and improves the accuracy of fake news detection.

[0020] The above description is only an overview of the technical solution of the present application. In order to be able to understand the technical means of the present application more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features and advantages of the present application more obvious and understandable, the following specifically gives the specific implementation manners of the present application. Description of the Drawings

[0021] By reading the following detailed description of the preferred embodiments, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the present application. Moreover, throughout the drawings, the same reference numerals are used to represent the same components. In the drawings:

[0022] Figure 1 is a schematic flow chart of a fake news detection method provided by an embodiment of the present application;

[0023] Figure 2 is a top view of the model training stage of a fake news detection method provided by an embodiment of the present application;

[0024] Figure 3 is a top view of the model testing stage of a fake news detection method provided by an embodiment of the present application;

[0025] Figure 4 is a block diagram of a fake news detection device provided by an embodiment of the present application;

[0026] Figure 5 is a block diagram of a computing device structure provided by an embodiment of the present application. Detailed Embodiments

[0027] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be completely conveyed to those skilled in the art.

[0028] See Figure 1 , Figure 1 is a schematic flow chart of a fake news detection method provided by an embodiment of the present application. Specifically, it includes the following steps.

[0029] Step S102: Construct a large model prompt template, input the obtained news text and domain label into the large model prompt template, perform domain-specific entity concept parsing through a large language model, and output the reconstructed news text, where the large model prompt template includes a task instruction, an input specification, and an output format requirement, and the task instruction covers entity screening, concept parsing processing, and enhanced news reconstruction.

[0030] In the embodiment of the present application, the inputting the obtained news text and domain label into the large model prompt template, performing domain-specific entity concept parsing through a large language model, and outputting the reconstructed news text includes:

[0031] Identify the entities in the news text that are related to the domain tags, and filter out the core entities that require concept parsing;

[0032] Perform concept parsing processing on the core entities to generate concept parsing results, where the concept parsing results include basic definitions, domain relevance definitions, context adaptation explanations, and extended associations;

[0033] Based on the news text and the concept parsing results, obtain the reconstructed news text.

[0034] In practical applications, construct a large model prompt template, the content is as follows,

[0035] First, construct a task instruction. According to the provided news text and domain tags, perform the following professional processing procedures:

[0036] (1) Entity screening. Identify the entities in the input text that are related to <domain tags>; filter out the core entities that require concept parsing (technical terms / domain-specific nouns / industry-specific expressions)

[0037] (2) Concept parsing processing. For each entity screened out, perform the following steps: Determine the entity name: basic definition (i.e., concise industry consensus definition); domain relevance (i.e., explain the special meaning of this entity in <domain tags>); context adaptation explanation (i.e., the specific role or impact in the current news scenario); extended association (i.e., list 2-3 related concepts or technical terms)

[0038] (3) Enhanced news reconstruction. Based on the original text and concept parsing results: retain the core facts and narrative structure of the original text; insert concept parsing content at the position where the professional term first appears; ensure that the entity parsing content is naturally integrated with the news context; maintain the objectivity and readability of the news.

[0039] Second, determine the input specifications. Original news: [news text]; Specified domain: [domain tags, such as: economy / politics / entertainment / technology].

[0040] Finally, determine the output format requirements, that is, the reconstructed news text.

[0041] Inject the news text and domain tags into the large model prompt template, perform domain-specific entity concept parsing through the large language model, and output the reconstructed news text.

[0042] For example, input: News text: OpenAI recently released a new multi-modal model GPT-4o, which has made a breakthrough in cross-modal alignment and has particularly improved the cross-modal attention mechanism; Domain tag: technology.

[0043] Output: News text containing the parsing of domain-specific entity concepts: OpenAI recently released a new multi-modal model, GPT-4o (the fourth generation of generative pre-trained model developed by OpenAI, where "o" represents omni-modal, supporting the joint processing of text, images, and audio). This model has made a breakthrough in cross-modal alignment, particularly improving the cross-modal attention mechanism (a core technology in the field of artificial intelligence that realizes the alignment and fusion of features in different modalities through an attention weight matrix).

[0044] Step S104: Construct a domain-specific prompt template and a domain-general prompt template respectively. Input the reconstructed news text into the domain-specific prompt template and the domain-general prompt template respectively. Through the masked language model, extract the domain-specific feature vector and the domain-general feature vector of the news text.

[0045] In practical applications, the domain-specific prompt template: {Domain label} News: {News text}. To sum up, this news is [MASK] news.

[0046] Domain-general prompt template: News: {News text}. To sum up, this news is [MASK] news.

[0047] After filling the news text and the domain label into the domain-specific prompt template and the domain-general prompt template respectively, use the masked language model BERT for in-depth semantic understanding, and output the prediction vector at the [MASK] position of the domain-specific template The prediction vector at the [MASK] position of the domain-general template Use a multi-layer perceptron MLP for The conversion of the vector dimension, the formula is as follows:

[0048]

[0049] where, v spe is the domain-specific feature vector of the news, and v gen is the domain-general feature vector of the news.

[0050] Step S106: Through cosine similarity, evaluate the first similarity between the domain-specific feature vector and the domain-specific true news prototype and the domain-specific false news prototype, and the second similarity between the domain-general feature vector and the domain-general true news prototype and the domain-general false news prototype. Use the first similarity and the second similarity, based on the model prediction probability calculation formula, calculate the prediction probability of the model for the news text, and based on the prediction probability, evaluate the news type.

[0051] In practical applications, use learnable parameters The category prototype representing news in each field, using learnable parameters g t , g f represents the category prototype common to the field,

[0052] wherein is the true news prototype of field D, is the fake news prototype of field D, g t is the true news prototype common to the field, g f is the fake news prototype common to the field.

[0053] By evaluating the similarity between the news feature vector and the true news prototype and the fake news prototype through the cosine similarity M(,), the news type can be better evaluated under few-shot conditions.

[0054] Based on the model prediction probability calculation formula, calculate the prediction probability of the model for the news text, wherein the model prediction probability calculation formula includes:

[0055]

[0056] where α is a hyperparameter, is the category prototype of the field d where the sample i is located i of, g k , g k' are the category prototypes common to the field, g k The corresponding news category is k, g k' The corresponding news category k' is opposite to k.

[0057] Input the prediction probability of the news for category k into the round function to return the prediction result (fake news / true news).

[0058] Step S108: Calculate the loss value based on the loss calculation formula for model performance optimization.

[0059] Specifically, calculating the model loss value based on the model loss calculation formula includes:

[0060] Calculate the model loss value based on the model loss calculation formula, wherein the model loss calculation formula includes:

[0061]

[0062]

[0063] where σ, γ are hyperparameters, y i is the actual category of news i, N is the number of training samples, is the corresponding field t of news i iThe category prototype, is the domain - general news prototype, The corresponding news category is the actual category y of news i i , The corresponding news category is opposite to the actual category of news i.

[0064] See Figure 2 , Figure 2 which is the top - level view of the model training stage of a fake news detection method provided by an embodiment of this application.

[0065] See Figure 3 , Figure 3 which is the top - level view of the model testing stage of a fake news detection method provided by an embodiment of this application.

[0066] Existing methods do not fully explore the semantic meaning of domain - specific entities, making it difficult to effectively model the dependencies between words in different domains, which limits the multi - domain adaptability of the model. To address this problem, a large - model prompt template is constructed, and the large model is used for domain - specific entity concept parsing to enhance the model's semantic understanding of news and improve the multi - domain adaptability of the method.

[0067] There are phenomena of theme and writing style deviation in news from different domains, making it difficult for the model to better distinguish true and false news in each domain under the condition of limited data resources. To address this problem, two types of prompt templates (domain - specific prompt template, domain - general prompt template) are constructed. By converting the classification task into a task in the pre - training stage of the masked language model, the difference between the training stage and the pre - training stage tasks is reduced, and the parameter update amplitude of the masked language model is decreased, enabling the model to better adapt to the fake news detection task under resource - constrained conditions; two types of category prototypes (domain - specific news category prototype, domain - general news category prototype) are constructed. By optimizing and updating the category prototype with a small number of parameters, the model's ability to distinguish true and false news in the same domain can be effectively improved under few - sample conditions.

[0068] By using the large - language model to parse domain - specific entity concepts to reconstruct news, the model can better understand news in each domain, which helps to effectively capture the complex dependencies between news words, enhance the news representation effect, and improve the accuracy of fake news detection; by constructing domain - specific category prototypes and domain - general category prototypes, and evaluating the similarity between the news feature vector and the category prototype and updating the category prototype, the model can effectively identify fake news under few - sample conditions, improving the accuracy of fake news detection.

[0069] Corresponding to the above - mentioned method embodiment, this specification also provides an embodiment of a fake news detection device,

[0070] Figure 4 which is the block diagram of a fake news detection device provided by an embodiment of this application. AsFigure 4 As shown in the figure, it specifically includes the following modules.

[0071] The output module 402 is configured to construct a large model prompt template, input the obtained news text and domain labels into the large model prompt template, perform domain-specific entity concept parsing through a large language model, and output the reconstructed news text. Among them, the large model prompt template includes task instructions, input specifications, and output format requirements. The task instructions cover entity screening, concept parsing processing, and enhanced news reconstruction;

[0072] The generation module 404 is configured to construct a domain-specific prompt template and a domain-general prompt template respectively, input the reconstructed news text into the domain-specific prompt template and the domain-general prompt template respectively, and extract the domain-specific feature vector and domain-general feature vector of the news text through a masked language model;

[0073] The prediction module 406 is configured to evaluate the first similarity between the domain-specific feature vector and the domain-specific true news prototype and domain-specific false news prototype through cosine similarity, and the second similarity between the domain-general feature vector and the domain-general true news prototype and domain-general false news prototype. Utilize the first similarity and the second similarity, based on the model prediction probability calculation formula, calculate the prediction probability of the model for the news text, and evaluate the news type based on the prediction probability;

[0074] The training module 408 is configured to calculate a loss value based on the loss calculation formula for optimizing the model performance.

[0075] In an optional embodiment, the output module 402 is further configured to:

[0076] Identify the entities related to the domain labels in the news text, and screen out the core entities that require concept parsing;

[0077] Perform concept parsing processing on the core entities to generate concept parsing results, where the concept parsing results include basic definitions, domain relevance definitions, context adaptation explanations, and extended associations;

[0078] Based on the news text and the concept parsing results, obtain the reconstructed news text.

[0079] In an optional embodiment, the prediction module 406 is further configured to:

[0080] Based on the model prediction probability calculation formula, calculate the probability that the news text is false news. The model prediction probability calculation formula includes:

[0081]

[0082] Among them, α is a hyperparameter, is the class prototype of the field d where the sample i is located i of is the general class prototype of the field The corresponding news category is k, The corresponding news category k' is opposite to k.

[0083] In an optional embodiment, the training module 408 is further configured to:

[0084] Based on the model loss calculation formula, calculate the model loss value, where the model loss calculation formula includes:

[0085]

[0086] Among them, σ and γ are hyperparameters, y i is the actual category of news i, N is the number of training samples, is the class prototype of the field t corresponding to news i i of is the general news prototype of the field The corresponding news category is the actual category y of news i i , The corresponding news category is opposite to the actual category of news i.

[0087] By using the large language model to parse domain-specific entity concepts to reconstruct news, the model can better understand news in each field, which helps to effectively capture the complex dependency relationships between news words, enhance the news representation effect, and improve the accuracy of fake news detection; constructing domain-specific class prototypes and general domain class prototypes, and by evaluating the similarity between news feature vectors and class prototypes and updating the class prototypes, the model can effectively identify fake news in multiple fields under few-shot conditions, improving the accuracy of fake news detection.

[0088] Each embodiment in this specification is described in a progressive manner. The same or similar parts among the embodiments can be referred to each other, and the key points of each embodiment are the differences from other embodiments. In particular, for the fake news detection device, since it is basically similar to the fake news detection method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the fake news detection method embodiment.

[0089] Figure 5A structural block diagram of a computing device provided by an embodiment of the present application. Components of the computing device 500 include, but are not limited to, a memory 510 and a processor 520. The processor 520 is connected to the memory 510 via a bus 530, and a database 550 is used to store data.

[0090] The computing device 500 further includes an access device 540, which enables the computing device 500 to communicate via one or more networks 560. Examples of these networks include the Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 540 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0091] In one embodiment of the present specification, the above components of the computing device 500 and Figure 5 other components not shown may also be connected to each other, for example, via a bus. It should be understood that Figure 5 the shown structural block diagram of the computing device is for illustrative purposes only and is not a limitation on the scope of the present specification. Those skilled in the art can add or replace other components as needed.

[0092] The computing device 500 can be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smartphone), a wearable computing device (e.g., a smartwatch, smart glasses, etc.) or other types of mobile devices, or a stationary computing device such as a desktop computer or a personal computer (PC). The computing device 500 can also be a mobile or stationary server.

[0093] Among them, the processor 520 is used to execute the following computer-executable instructions, and when the computer-executable instructions are executed by the processor, the steps of the above-mentioned fake news detection method are implemented.

[0094] Each embodiment in this specification is described in a progressive manner. For the same or similar parts between each embodiment, reference can be made to each other, and the key point of each embodiment is to illustrate the differences from other embodiments. In particular, for the embodiment of the computing device, since it is basically similar to the embodiment of the fake news detection method, the description is relatively simple, and for the relevant parts, reference can be made to the partial description of the embodiment of the fake news detection method.

[0095] An embodiment of this specification also provides a computer-readable storage medium, which stores computer-executable instructions, and when the computer-executable instructions are executed by the processor, the steps of the above-mentioned fake news detection method are implemented.

[0096] Each embodiment in this specification is described in a progressive manner. For the same or similar parts between each embodiment, reference can be made to each other, and the key point of each embodiment is to illustrate the differences from other embodiments. In particular, for the embodiment of the computer-readable storage medium, since it is basically similar to the embodiment of the fake news detection method, the description is relatively simple, and for the relevant parts, reference can be made to the partial description of the embodiment of the fake news detection method.

[0097] An embodiment of this specification also provides a computer program, wherein when the computer program is executed on a computer, the computer is made to execute the steps of the above-mentioned fake news detection method.

[0098] Each embodiment in this specification is described in a progressive manner. For the same or similar parts between each embodiment, reference can be made to each other, and the key point of each embodiment is to illustrate the differences from other embodiments. In particular, for the embodiment of the computer program, since it is basically similar to the embodiment of the fake news detection method, the description is relatively simple, and for the relevant parts, reference can be made to the partial description of the embodiment of the fake news detection method.

[0099] The above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve the desired result. Additionally, the processes depicted in the figures do not necessarily require the particular order or sequential order shown to achieve the desired result. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0100] The computer instructions include computer program code, which may be in the form of source code, object code, executable files, or some intermediate forms, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the content included in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice within the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0101] It should be noted that the above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired result. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments of this specification.

[0102] In the above embodiments, the descriptions of each embodiment have their own focuses. For the parts not detailed in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

[0103] The preferred embodiments of this specification disclosed above are only used to help explain this specification. The alternative embodiments do not elaborate on all details and do not limit the invention to the specific embodiments described. Obviously, many modifications and variations can be made according to the content of the embodiments of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the embodiments of this specification, so that those skilled in the art can well understand and utilize this specification. This specification is only limited by the claims and their full scope and equivalents.

[0104] The preferred embodiments of the present specification disclosed above are only used to help explain the present specification. The alternative embodiments do not describe all the details in detail, nor do they limit the invention to the specific embodiments described. Obviously, according to the content of the embodiments of the present specification, many modifications and variations can be made. The present specification selects and specifically describes these embodiments to better explain the principles and practical applications of the embodiments of the present specification, so that those skilled in the art can well understand and utilize the present specification. The present specification is only limited by the claims and their full scope and equivalents.

Claims

1. A fake news detection method, characterized in that, Including: Construct a large model prompt template, input the obtained news text and domain labels into the large model prompt template, perform domain-specific entity concept parsing through a large language model, and output the reconstructed news text. The large model prompt template includes task instructions, input specifications, and output format requirements. The task instructions cover entity screening, concept parsing processing, and enhanced news reconstruction; Construct a domain-specific prompt template and a domain-general prompt template respectively. Input the reconstructed news text into the domain-specific prompt template and the domain-general prompt template respectively, and extract the domain-specific feature vector and domain-general feature vector of the news text through a masked language model; Evaluate the first similarity between the domain-specific feature vector and the domain-specific true news prototype and domain-specific false news prototype, and the second similarity between the domain-general feature vector and the domain-general true news prototype and domain-general false news prototype through cosine similarity. Use the first similarity and the second similarity, based on the model prediction probability calculation formula, calculate the prediction probability of the model for the news text, and evaluate the news type based on the prediction probability; Calculate the loss value based on the loss calculation formula for model performance optimization.

2. The method according to claim 1, characterized in that, The step of inputting the obtained news text and domain labels into the large model prompt template and performing domain-specific entity concept parsing through a large language model to output the reconstructed news text includes: Identify the entities related to the domain labels in the news text and screen out the core entities that need concept parsing; Perform concept parsing processing on the core entities to generate concept parsing results. The concept parsing results include basic definitions, domain relevance definitions, context adaptation explanations, and extended associations; Based on the news text and the concept parsing results, obtain the reconstructed news text.

3. The method according to claim 1, wherein The step of calculating the prediction probability of the model for the news text based on the model prediction probability calculation formula includes: Calculate the prediction probability of the model for the news text based on the model prediction probability calculation formula. The model prediction probability calculation formula includes: where α is a hyperparameter, is the class prototype of the field d where sample i is located i g k , g k′ are domain - general class prototypes, g k the corresponding news category is k, g k′ the corresponding news category k′ is opposite to k.

4. The method according to claim 1, wherein The step of calculating the loss value based on the loss calculation formula includes: Calculate the model loss value based on the model loss calculation formula. The model loss calculation formula includes: where σ and γ are hyperparameters, and y i is the actual category of news i, N is the number of training samples, is the category prototype of the corresponding field t i of news i, is the domain-general news prototype, and the corresponding news category is the actual category y i of news i, while the corresponding news category is opposite to the actual category of news i.

5. A fake news detection device, characterized in that, Including: An output module configured to construct a large model prompt template, input the obtained news text and domain labels into the large model prompt template, perform domain-specific entity concept parsing through a large language model, and output the reconstructed news text. The large model prompt template includes task instructions, input specifications, and output format requirements. The task instructions cover entity screening, concept parsing processing, and enhanced news reconstruction; A generation module configured to construct a domain-specific prompt template and a domain-general prompt template respectively. Input the reconstructed news text into the domain-specific prompt template and the domain-general prompt template respectively, and extract the domain-specific feature vector and domain-general feature vector of the news text through a masked language model; A prediction module, configured to evaluate a first similarity between the domain-specific feature vector and domain-specific true-news prototypes and domain-specific false-news prototypes through cosine similarity, and a second similarity between the domain-general feature vector and domain-general true-news prototypes and domain-general false-news prototypes, use the first similarity and the second similarity, based on a model prediction probability calculation formula, calculate a prediction probability of the model for the news text, and evaluate the news type based on the prediction probability; A training module, configured to calculate a loss value based on a loss calculation formula for optimizing the model performance.

6. A computer device, characterized in that, The computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the method according to any one of claims 1-4 are implemented.

7. A computer-readable storage medium, characterized in that, An implementation program for information transmission is stored on the computer-readable storage medium. When the program is executed by the processor, the steps of the method according to any one of claims 1-4 are implemented.

Images