Artificial intelligence-based fake news detection method and device, and computer device

Abstract

The embodiment of the application belongs to the field of artificial intelligence and the field of financial technology, and relates to a false news detection method based on artificial intelligence, which comprises the following steps: inputting pre-collected news training text data into a preset graph comparison learning model for training to obtain a trained graph comparison learning model; taking the trained graph comparison learning model as a false news detection model, and performing false prediction processing on to-be-detected news text data through the false news detection model. The application also provides a false news detection device based on artificial intelligence, a computer device and a storage medium. In addition, the application also relates to blockchain technology, and the false news detection model can be stored in the blockchain. The application can be applied to the scene of false news identification in the financial field, and through the use of the false news detection model, a prediction result corresponding to the news text data can be quickly and accurately generated, and the accuracy of the prediction result is improved.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence technology and fintech, and in particular to artificial intelligence-based methods, devices, computer equipment, and storage media for detecting fake news. Background Technology

[0002] Fake news detection aims to identify misleading and factually inaccurate news from a large volume of news texts, serving as a crucial technological means to maintain information security and social stability. Therefore, designing an effective fake news detection device to detect false and misleading news about financial companies widely distributed across various social media platforms is of great significance for fintech companies (such as insurance companies and banks) to protect their image. Through fake news detection, fintech companies can quickly understand the prevailing misinformation in the external environment, enabling them to rapidly formulate appropriate publicity strategies and clarification measures to effectively maintain a positive image and prevent damage to their interests.

[0003] Currently, the main methods used by fintech companies to detect fake news are non-graph structure modeling methods based on text semantic recognition, such as identifying keywords and writing style to determine whether news text is fake. However, these methods have the following problems: they primarily model the sequence information of news text and fail to uncover potential semantic structure information, making it difficult to characterize deeper semantic relationships between words and reducing the accuracy of the model in detecting fake news. Summary of the Invention

[0004] The purpose of this application is to propose an artificial intelligence-based method, apparatus, computer device, and storage medium for detecting fake news, in order to solve the technical problem that existing fake news detection methods used by fintech companies mainly model the sequence information of news text, which cannot uncover potential semantic structure information, making it difficult to depict deeper semantic relationships between words and reducing the accuracy of the model in detecting fake news.

[0005] To address the aforementioned technical problems, this application provides an artificial intelligence-based method for detecting fake news, employing the following technical solution:

[0006] The pre-collected news training text data is input into the input layer of a preset graph comparison learning model, and a heterogeneous graph corresponding to the news training text data is constructed in the input layer.

[0007] The heterogeneous graph is input into a preset construction layer within the graph alignment learning model, where an augmented graph corresponding to the heterogeneous graph is constructed.

[0008] The heterogeneous graph and the augmented graph are input into the encoder in the graph alignment learning model, and a first low-dimensional vector of news text corresponding to the heterogeneous graph and a second low-dimensional vector of news text corresponding to the augmented graph are generated in the encoder.

[0009] Entity recognition is performed on the news training text data to obtain the corresponding entity nouns, and attribute features corresponding to the entity nouns are constructed based on a preset corpus;

[0010] The first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features are input into a preset construction layer in the graph comparison learning model. In the construction layer, a first target news text low-dimensional vector corresponding to the heterogeneous graph is constructed, and a second target news text low-dimensional vector corresponding to the augmented graph is constructed.

[0011] Based on the low-dimensional vectors of the first and second target news texts, a final loss function is constructed, and the final loss function is continuously minimized iteratively until the preset iteration conditions are met, at which point the iteration stops, and a trained graph alignment learning model is obtained.

[0012] The trained graph alignment learning model is used as a fake news detection model, and the fake news detection model is used to perform false prediction processing on the news text data to be tested.

[0013] Furthermore, the step of inputting the heterogeneous graph into a preset construction layer within the graph alignment learning model, and constructing an augmented graph corresponding to the heterogeneous graph in the construction layer, specifically includes:

[0014] The heterogeneous graph is input into the construction layer within the graph alignment learning model, and the set of connected edges within the heterogeneous graph is obtained in the construction layer.

[0015] Get the preset value;

[0016] Based on the preset value, the edges of the edge set in the heterogeneous graph are randomly deleted to obtain the processed heterogeneous graph.

[0017] The processed heterogeneous graph is used as the augmented graph.

[0018] Furthermore, the step of constructing attribute features corresponding to the entity nouns based on a preset corpus specifically includes:

[0019] Get the preset quantity;

[0020] The corpus is searched based on the entity nouns, and target sentences corresponding to the preset number are extracted from the corpus.

[0021] The target sentence is input into a preset long short-term memory network, and the target vector corresponding to the target sentence is output based on the long short-term memory network.

[0022] The target vector is used as the attribute feature corresponding to the entity noun.

[0023] Furthermore, the step of inputting the first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features into a preset construction layer within the graph comparison learning model, and constructing a first low-dimensional vector of the target news text corresponding to the heterogeneous graph and a second low-dimensional vector of the target news text corresponding to the augmented graph within the construction layer, specifically includes:

[0024] The first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features are input into the construction layer of the graph comparison learning model. In the construction layer, a comparison vector corresponding to the attribute features is generated based on a preset comparison function.

[0025] Construct a comparison feature vector corresponding to the comparison vector;

[0026] Based on the contrast feature vector and the first news text low-dimensional vector, a first target news text low-dimensional vector corresponding to the heterogeneous graph is generated.

[0027] Based on the contrast feature vector and the low-dimensional vector of the second news text, a low-dimensional vector of the second target news text corresponding to the augmented image is generated.

[0028] Furthermore, the step of constructing the final loss function based on the low-dimensional vectors of the first and second target news texts specifically includes:

[0029] A first loss function is calculated based on the low-dimensional vectors of the first and second target news texts.

[0030] The low-dimensional vector of the first target news text is input into the fully connected layer in the graph alignment learning model, and a predicted label corresponding to the low-dimensional vector of the first target news text is generated based on a preset activation function.

[0031] Obtain the real tags corresponding to the news training text data;

[0032] Calculate a second loss function based on the predicted label and the true label;

[0033] The final loss function is constructed based on the first loss function and the second loss function.

[0034] Furthermore, the step of iteratively minimizing the final loss function until a preset iteration condition is met, thereby obtaining the trained graph alignment learning model, specifically includes:

[0035] The final loss function is minimized iteratively, and the current cumulative number of iterations is obtained.

[0036] Determine whether the cumulative number of iterations is greater than a preset iteration number threshold;

[0037] If the cumulative number of iterations is greater than the iteration threshold, the iteration condition is satisfied, the iteration stops, and the trained graph alignment learning model is obtained.

[0038] Furthermore, after the step of determining whether the cumulative number of iterations is greater than a preset iteration threshold, the method further includes:

[0039] If the cumulative number of iterations is less than the iteration number threshold, obtain the first loss value of the current iteration process;

[0040] Obtain the second loss value of the previous iteration process adjacent to the current iteration process;

[0041] Calculate the difference between the first loss value and the second loss value;

[0042] Determine whether the difference is less than a preset difference threshold;

[0043] If so, the iteration condition is met, the iteration stops, and the trained graph alignment learning model is obtained.

[0044] To address the aforementioned technical problems, this application also provides an artificial intelligence-based fake news detection device, employing the following technical solution:

[0045] The first construction module is used to input pre-collected news training text data into the input layer of a preset graph comparison learning model, and to construct a heterogeneous graph corresponding to the news training text data in the input layer.

[0046] The second construction module is used to input the heterogeneous graph into a preset construction layer in the graph alignment learning model, and to construct an augmented graph corresponding to the heterogeneous graph in the construction layer;

[0047] The generation module is used to input the heterogeneous graph and the augmented graph into the encoder in the graph alignment learning model, and generate a first low-dimensional vector of news text corresponding to the heterogeneous graph and a second low-dimensional vector of news text corresponding to the augmented graph in the encoder.

[0048] The third construction module is used to perform entity recognition on the news training text data to obtain the corresponding entity nouns, and to construct attribute features corresponding to the entity nouns based on a preset corpus.

[0049] The construction module is used to input the first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features into a preset construction layer in the graph comparison learning model, and to construct a first target news text low-dimensional vector corresponding to the heterogeneous graph and a second target news text low-dimensional vector corresponding to the augmented graph in the construction layer.

[0050] The first processing module is used to construct a final loss function based on the low-dimensional vector of the first target news text and the low-dimensional vector of the second target news text, and to continuously iterate and minimize the final loss function until the preset iteration condition is met and the iteration stops, so as to obtain a trained graph comparison learning model.

[0051] The second processing module is used to use the trained graph alignment learning model as a fake news detection model, and to perform false prediction processing on the news text data to be tested through the fake news detection model.

[0052] To address the aforementioned technical problems, this application also provides a computer device that employs the following technical solution:

[0053] The pre-collected news training text data is input into the input layer of a preset graph comparison learning model, and a heterogeneous graph corresponding to the news training text data is constructed in the input layer.

[0054] The heterogeneous graph is input into a preset construction layer within the graph alignment learning model, where an augmented graph corresponding to the heterogeneous graph is constructed.

[0055] The heterogeneous graph and the augmented graph are input into the encoder in the graph alignment learning model, and a first low-dimensional vector of news text corresponding to the heterogeneous graph and a second low-dimensional vector of news text corresponding to the augmented graph are generated in the encoder.

[0056] Entity recognition is performed on the news training text data to obtain the corresponding entity nouns, and attribute features corresponding to the entity nouns are constructed based on a preset corpus;

[0057] The first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features are input into a preset construction layer in the graph comparison learning model. In the construction layer, a first target news text low-dimensional vector corresponding to the heterogeneous graph is constructed, and a second target news text low-dimensional vector corresponding to the augmented graph is constructed.

[0058] Based on the low-dimensional vectors of the first and second target news texts, a final loss function is constructed, and the final loss function is continuously minimized iteratively until the preset iteration conditions are met, at which point the iteration stops, and a trained graph alignment learning model is obtained.

[0059] The trained graph alignment learning model is used as a fake news detection model, and the fake news detection model is used to perform false prediction processing on the news text data to be tested.

[0060] To address the aforementioned technical problems, this application also provides a computer-readable storage medium, employing the technical solution described below:

[0061] The pre-collected news training text data is input into the input layer of a preset graph comparison learning model, and a heterogeneous graph corresponding to the news training text data is constructed in the input layer.

[0062] The heterogeneous graph is input into a preset construction layer within the graph alignment learning model, where an augmented graph corresponding to the heterogeneous graph is constructed.

[0063] The heterogeneous graph and the augmented graph are input into the encoder in the graph alignment learning model, and a first low-dimensional vector of news text corresponding to the heterogeneous graph and a second low-dimensional vector of news text corresponding to the augmented graph are generated in the encoder.

[0064] Entity recognition is performed on the news training text data to obtain the corresponding entity nouns, and attribute features corresponding to the entity nouns are constructed based on a preset corpus;

[0065] The first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features are input into a preset construction layer in the graph comparison learning model. In the construction layer, a first target news text low-dimensional vector corresponding to the heterogeneous graph is constructed, and a second target news text low-dimensional vector corresponding to the augmented graph is constructed.

[0066] Based on the low-dimensional vectors of the first and second target news texts, a final loss function is constructed, and the final loss function is continuously minimized iteratively until the preset iteration conditions are met, at which point the iteration stops, and a trained graph alignment learning model is obtained.

[0067] The trained graph alignment learning model is used as a fake news detection model, and the fake news detection model is used to perform false prediction processing on the news text data to be tested.

[0068] Compared with the prior art, the embodiments of this application have the following main advantages:

[0069] This application's embodiments simulate the elaborate design process of fake news by constructing augmented graph data, and then introduce a contrastive learning process to characterize the construction features of such elaborately designed sentences into a low-dimensional vector representation. Furthermore, by designing an effective domain-external knowledge fusion technique, information from the internal corpus is integrated into the model learning process, characterizing the semantic inconsistencies between real and fake information in the low-dimensional representation vector. This improves the robustness and accuracy of the fake news detection model, enabling subsequent false prediction processing of the tested news text data using the fake news detection model to quickly and accurately generate prediction results corresponding to the news text data, thus improving the accuracy of the prediction results. Attached Figure Description

[0070] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 This is an exemplary system architecture diagram to which this application can be applied;

[0072] Figure 2 A flowchart of an embodiment of the artificial intelligence-based fake news detection method according to this application;

[0073] Figure 3 This is a schematic diagram of the structure of one embodiment of the artificial intelligence-based fake news detection device according to this application;

[0074] Figure 4 This is a schematic diagram of the structure of one embodiment of the computer device according to this application. Detailed Implementation

[0075] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0076] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0077] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0078] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0079] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0080] Terminal devices 101, 102, and 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 players (Moving Picture Experts Group Audio Layer IV), laptops, and desktop computers, etc.

[0081] Server 105 can be a server that provides various services, such as a backend server that supports the pages displayed on terminal devices 101, 102, and 103.

[0082] It should be noted that the AI-based fake news detection method provided in this application is generally executed by a server / terminal device, and correspondingly, the AI-based fake news detection device is generally installed in the server / terminal device.

[0083] It should be understood that Figure 1The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0084] Continue to refer to Figure 2 This document illustrates a flowchart of an embodiment of the AI-based fake news detection method according to this application. The order of steps in the flowchart can be changed, and some steps can be omitted, depending on different requirements. The AI-based fake news detection method provided in this application can be applied to any scenario requiring fake news detection, and thus can be applied to products in these scenarios, such as fake news identification in the financial insurance field. The AI-based fake news detection method includes the following steps:

[0085] Step S201: Input the pre-collected news training text data into the input layer of the preset graph comparison learning model, and construct a heterogeneous graph corresponding to the news training text data in the input layer.

[0086] In this embodiment, the AI-based fake news detection method operates on an electronic device (e.g., Figure 1 The server / terminal device shown can acquire news training text data via wired or wireless connection. It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G / 5G connections, WiFi connections, Bluetooth connections, WiMAX connections, Zigbee connections, UWB (ultra-wideband) connections, and other currently known or future-developed wireless connection methods. Among these, news text D = {d1, d2, ..., d...} collected by the target company from various social media platforms can be used. N The above news training text data serves as the basis for this heterogeneous graph, which can also be called a heterogeneous topology graph or topology graph, and can be used for each news item d. i Construct a heterogeneous topology graph based on text information. Specifically, for each news text d... i Construct a topology graph G i ={V i ,ε i ,X i For each graph G i It contains two types of nodes, namely sentence nodes S i ={s 1 ,s 2 ,…,s m} and entity noun node E i ={e 1 ,e 2 ,…,e n}, where sentence node S iThis refers to treating the entire sentence as a node, while an entity noun node refers to treating a specific word within the sentence as a node. Therefore, Graph G i The set of nodes V i =S i ∪E i ε represents the set of edges in the graph. The establishment of the edge set ε involves the following two steps: 1) Establishing edges between sentences: [The text d...] i Each sentence in the text is isolated independently, and each sentence is treated as a node. Then, edges are established between each pair of sentence nodes. 2) Establishing edges between entity nouns and sentences: First, from the news text d i n entity nouns were identified in the sentence S. Secondly, if sentence S... i Contains entity noun e j Then, an edge is created between the sentence node and the word node. i Let X be an attribute matrix. i The construction of this graph is based on the fact that each graph contains nodes of two types: sentences and entity nouns. Specifically, for the attributes of sentence nodes, the sentence S... i The input is fed into the LSTM, and the output feature vector is the sentence node S. i The initial features. For the attributes of entity nouns, e is obtained based on a real corpus of the target company. PA To initialize. Specifically, e PA The construction of this will be explained in later embodiments. Therefore, the initial feature vectors of all nodes can be constructed, thus completing the attribute matrix X in the graph. i The construction of the topology graph. If there are N news texts to be detected, then N such topology graphs can be constructed, i.e., {G} 1 G 2 ,…,G N The target company can refer to a fintech company, such as an insurance company or a bank. The aforementioned news text can refer to news information about fintech companies circulating on social media.

[0087] Step S202: Input the heterogeneous graph into a preset construction layer in the graph alignment learning model, and construct an augmented graph corresponding to the heterogeneous graph in the construction layer.

[0088] In this embodiment, the specific implementation process of inputting the heterogeneous graph into the preset construction layer within the graph alignment learning model and constructing the augmented graph corresponding to the heterogeneous graph in the construction layer will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0089] Step S203: Input the heterogeneous graph and the augmented graph into the encoder in the graph alignment learning model, and generate a first low-dimensional vector of news text corresponding to the heterogeneous graph and a second low-dimensional vector of news text corresponding to the augmented graph in the encoder.

[0090] In this embodiment, for a specific graph G i This yields nodes in the graph that have two types, namely, type. Here, τ1 and τ2 represent node type 1 and type 2, respectively. Based on this, given a news text d... i The constructed topology graph G i ={V i ,ε i ,X i}, enter into formula The encoder in the code can output a low-dimensional vector representation matrix H for all nodes. i Among them, H i It also contains news text d i Low-dimensional vector representation matrix of sentence nodes Vector representation matrix of entity noun nodes Among them, after obtaining the news text d i Low-dimensional vector representation matrix of sentence nodes Later (Note: The news text d is included i (representation of all sentences in the text), for this After performing max-pooling, the news text d can be obtained. i Low-dimensional vector representation σ() represents the ReLU activation function. Different types of nodes τ have different learnable parameter matrices. This is the attention weight matrix, where the number of rows represents the total number of nodes and the number of columns represents the number of nodes of type τ. Each element in the matrix This represents the attention weight value in the p-th row and q-th column. Specifically, this... It can be derived from the formula The calculation yields ν, where ν is the learnable attention vector. α τ It is a type-level learnable attention weight. and These are the low-dimensional vector representations of node p and its neighbor node q, respectively.

[0091] Step S204: Entity recognition is performed on the news training text data to obtain the corresponding entity nouns, and attribute features corresponding to the entity nouns are constructed based on a preset corpus.

[0092] In this embodiment, the specific implementation process of constructing attribute features corresponding to the entity nouns based on the preset corpus will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0093] Step S205: Input the first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features into a preset construction layer in the graph comparison learning model. In the construction layer, construct the first low-dimensional vector of the target news text corresponding to the heterogeneous graph, and construct the second low-dimensional vector of the target news text corresponding to the augmented graph.

[0094] In this embodiment, the specific implementation process of constructing the first target news text low-dimensional vector corresponding to the heterogeneous graph and the second target news text low-dimensional vector corresponding to the augmented graph within the construction layer will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0095] Step S206: Construct a final loss function based on the low-dimensional vectors of the first and second target news texts, and continuously iterate to minimize the final loss function until the preset iteration conditions are met, at which point the iteration stops, and a trained graph alignment learning model is obtained.

[0096] In this embodiment, the specific implementation process of constructing a final loss function based on the low-dimensional vectors of the first and second target news texts and iteratively minimizing the final loss function until the preset iteration condition is met, thereby obtaining the trained graph alignment learning model, will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0097] Step S207: Use the trained graph alignment learning model as a fake news detection model, and use the fake news detection model to perform false prediction processing on the news text data to be tested.

[0098] In this embodiment, the news text data to be determined as fake news can be input into the trained fake news detection model, which then outputs predicted labels for all news texts. in w represents the number of news texts in the news text data, where 0 represents that the news is real news and 1 represents that the news is fake news.

[0099] This application simulates the elaborate design process of fake news by constructing augmented graph data, and then introduces a contrastive learning process to characterize the construction features of such elaborately designed sentences into a low-dimensional vector representation. Furthermore, by designing an effective domain-external knowledge fusion technique, information from the internal corpus is integrated into the model learning process, characterizing the semantic inconsistencies between real and fake information in the low-dimensional representation vector. This improves the robustness and accuracy of the fake news detection model, enabling subsequent false prediction processing of the tested news text data using the fake news detection model to quickly and accurately generate prediction results corresponding to the news text data, thus improving the accuracy of the prediction results.

[0100] In some alternative implementations, step S202 includes the following steps:

[0101] The heterogeneous graph is input into the construction layer within the graph alignment learning model, and the set of edges within the heterogeneous graph is obtained in the construction layer.

[0102] In this embodiment, for a given heterogeneous graph G i ={V i ,ε i ,X i}, where the set of edges is ε. i .

[0103] Get the preset value.

[0104] In this embodiment, the value of the above-mentioned preset value is not specifically limited and can be set according to actual usage needs, for example, it can be set to 10%.

[0105] Based on the preset value, the edges of the edge set in the heterogeneous graph are randomly deleted to obtain the processed heterogeneous graph.

[0106] In this embodiment, if the preset value is 10%, then it can be obtained from the heterogeneous graph G. i ={V i ,ε i ,X i}ε i Randomly delete 10% of the edges in the set to obtain a new set of edges ε. 2i This allows us to obtain a new augmented graph G. 2i ={V i ,ε 2i ,X i}

[0107] The processed heterogeneous graph is used as the augmented graph.

[0108] This application inputs the heterogeneous graph into the construction layer of the graph alignment learning model, where the set of edges within the heterogeneous graph is obtained. Then, a preset value is acquired. Subsequently, based on the preset value, edges in the set of edges within the heterogeneous graph are randomly deleted to obtain a processed heterogeneous graph, which is then used as the augmented graph. By constructing the augmented graph, this application can simulate situations in fake news where individuals intentionally remove keywords, rearrange characters, or change sentence order to evade model detection. This provides richer data input for subsequent training and effectively enhances the robustness of the model.

[0109] In some optional implementations of this embodiment, the step S204 of constructing attribute features corresponding to the entity noun based on a preset corpus includes the following steps:

[0110] Get the preset quantity.

[0111] In this embodiment, the value of the preset quantity is not specifically limited and can be set according to actual usage needs, for example, it can be set to 3.

[0112] The corpus is searched based on the entity nouns, and target sentences corresponding to the preset number are extracted from the corpus.

[0113] In this embodiment, the aforementioned corpus can be a database constructed based on the target company's internal news, announcements, and other materials. For example, if the preset quantity is 3, then the aforementioned target sentences are three sentences randomly extracted from the corpus that contain the aforementioned entity nouns.

[0114] The target sentence is input into a preset long short-term memory network, and the target vector corresponding to the target sentence is output based on the long short-term memory network.

[0115] In this embodiment, the aforementioned Long Short-Term Memory (LSTM) network is also referred to as LSTM. By inputting the target sentence into the preset LSTM network, the LSTM network outputs the ePA vector, which is the attribute feature of the corresponding entity noun.

[0116] The target vector is used as the attribute feature corresponding to the entity noun.

[0117] This application obtains a preset number of entity names; then, it searches the corpus based on these entity names to extract target sentences corresponding to the preset number of entity names; subsequently, it inputs the target sentences into a preset Long Short-Term Memory (LSTM) network, outputs a target vector corresponding to the target sentences based on the LSM network, and uses the target vector as the attribute feature corresponding to the entity names. This application, based on the use of the corpus, can quickly and intelligently generate the attribute features corresponding to the entity names, which is beneficial for accurately constructing a first low-dimensional vector of target news text corresponding to the heterogeneous graph and a second low-dimensional vector of target news text corresponding to the augmented graph based on the obtained attribute features.

[0118] In some alternative implementations, step S205 includes the following steps:

[0119] The first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features are input into the construction layer of the graph comparison learning model. In the construction layer, a comparison vector corresponding to the attribute features is generated based on a preset comparison function.

[0120] In this embodiment, to effectively utilize accurate data from within the target company to assist in the accurate detection of fake news, this solution further introduces an effective external domain knowledge fusion technology strategy to characterize the semantic inconsistency between potential fake news and the target company's internal real information. Specifically, e c A low-dimensional vector representation of a certain entity noun node (e c (Also a row vector in a matrix He). e PA These are entity noun vectors obtained from a pre-defined corpus of the target company. For comparison with e... c and e PA The difference between these two is determined by calculating the comparison vector u. i To depict it, the specific formula is as follows: u i =fcom(e c W b ·e PA ), where fcom() is the function to calculate the comparison vector value, W b This is a learnable parameter matrix. Here, fcom() is... The formula is constructed using the given information. Where W... k For learnable matrices, It is the hadamard product.

[0121] Construct a comparison feature vector corresponding to the comparison vector.

[0122] In this embodiment, the final comparison feature vector C is defined in U = [u1, u2, ..., u...]. n The result is obtained by performing max-pooling on [u1, u2, ..., u]. n [E] represents the entity noun node E = {e1, e2, ..., e} in all news texts. n The comparison vector of}.

[0123] Based on the contrast feature vector and the first news text low-dimensional vector, a first target news text low-dimensional vector corresponding to the heterogeneous graph is generated.

[0124] In this embodiment, based on this, the final news text d can be obtained. i Low-dimensional vector representation [] represents vector concatenation operations.

[0125] Based on the contrast feature vector and the low-dimensional vector of the second news text, a low-dimensional vector of the second target news text corresponding to the augmented image is generated.

[0126] In this embodiment, the low-dimensional vector of the second target news text The generation process can be referenced from the generation process of the low-dimensional vector of the first target news text.

[0127] This application inputs the first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features into the construction layer of the graph comparison learning model. Within the construction layer, a comparison vector corresponding to the attribute features is generated based on a preset comparison function; then, a comparison feature vector corresponding to the comparison vector is constructed; subsequently, a first target news text low-dimensional vector corresponding to the heterogeneous graph is generated based on the comparison feature vector and the first low-dimensional vector of the news text; and a second target news text low-dimensional vector corresponding to the augmented graph is generated based on the comparison feature vector and the second low-dimensional vector of the news text. Based on the use of the construction layer, this application can quickly and accurately construct the first target news text low-dimensional vector corresponding to the heterogeneous graph, and construct the second target news text low-dimensional vector corresponding to the augmented graph, which is beneficial for the rapid construction of the final loss function based on the obtained first and second target news text low-dimensional vectors.

[0128] In some optional implementations, the step S206 of constructing the final loss function based on the first target news text low-dimensional vector and the second target news text low-dimensional vector includes the following steps:

[0129] The first loss function is calculated based on the low-dimensional vectors of the first and second target news texts.

[0130] In this embodiment, the first loss function described above can also be called the construction of a contrastive learning loss function. For a news text d i (Anchor point), its label is y i ∈{0,1}, where y i =0 indicates that the tag belongs to the real news category, y i =1 indicates that the tag belongs to the fake news category. First, the positive samples of the anchor point are determined to be the low-dimensional vector representations of their corresponding augmented graphs. And the set of news texts P(k) = {d1,…,d1} that belong to the same label column as news text di. k For ease of description, we will use d. k The news text representing the same tag as news text di is represented by its low-dimensional vector as follows: Negative samples are those related to the news text d. i The set of news texts that do not belong to the same tag category is Q(a)={d1,…,d a For ease of description, we will use d. a The low-dimensional vector representation of news text that does not belong to the same tag as news text di is: Based on this, the contrastive learning loss function can be constructed as shown in the following formula. (or the first loss function): Here, sim(x,y) represents the calculation of the cosine similarity between vectors x and y. δ is a manually adjustable parameter. During training, since the input is training data, the true label y is... i It is known that, on real social media platforms, those who publish fake news often deliberately modify it to further mimic the structure of real news, thereby evading detection models. To address this issue, this solution introduces a graph contrastive learning strategy. By simulating this "deliberate behavior" and characterizing it in the learned low-dimensional vector representation, this situation can be effectively incorporated into the training process, enabling the model to effectively identify such disguised fake news and improving its robustness.

[0131] The low-dimensional vector of the first target news text is input into the fully connected layer in the graph alignment learning model, and a predicted label corresponding to the low-dimensional vector of the first target news text is generated based on a preset activation function.

[0132] In this embodiment, the activation function is specifically the softmax activation function. This is achieved by using the low-dimensional vector of the first target news text. The input is activated by the softmax function to obtain the predicted label. As shown in the formula below: Where Wo is the learnable weight matrix and bo is the learnable bias vector.

[0133] Obtain the real labels corresponding to the news training text data.

[0134] A second loss function is calculated based on the predicted label and the true label.

[0135] In this embodiment, the second loss function is specifically the cross-entropy loss. Specifically, the second loss function includes: Where l() represents calculating the cross-entropy loss between the two vectors within the parentheses, y i This represents the news text d in the training data. i The true label. V N This represents the set of nodes used for training.

[0136] The final loss function is constructed based on the first loss function and the second loss function.

[0137] In this embodiment, the final loss function specifically includes: Where α is a manually adjustable parameter. For the first loss function, This is the second loss function.

[0138] This application calculates a first loss function based on the low-dimensional vectors of the first and second target news texts; then, it inputs the low-dimensional vector of the first target news text into a fully connected layer within the graph alignment learning model, generating predicted labels corresponding to the low-dimensional vector of the first target news text based on a preset activation function; subsequently, it obtains the ground truth labels corresponding to the news training text data; next, it calculates a second loss function based on the predicted labels and the ground truth labels; finally, it constructs a final loss function quickly and accurately based on the first and second loss functions, which is beneficial for accurately completing the training process of the graph alignment learning model based on the obtained final loss function.

[0139] In some optional implementations of this embodiment, step S206, which iteratively minimizes the final loss function until a preset iteration condition is met, stops the iteration to obtain a trained graph alignment learning model, includes the following steps:

[0140] The final loss function is minimized iteratively, and the current cumulative number of iterations is obtained.

[0141] In this embodiment, the cumulative number of iterations refers to the sum of all iterations.

[0142] Determine whether the cumulative number of iterations is greater than a preset iteration threshold.

[0143] In this embodiment, the value of the above-mentioned iteration number threshold θ is not specifically limited and can be set according to actual usage requirements.

[0144] If the cumulative number of iterations is greater than the iteration threshold, the iteration condition is satisfied, the iteration stops, and the trained graph alignment learning model is obtained.

[0145] In this embodiment, the model has been trained when the iteration stops.

[0146] This application iteratively minimizes the final loss function and obtains the current cumulative iteration count; then it determines whether the cumulative iteration count is greater than a preset iteration count threshold; if the cumulative iteration count is greater than the iteration count threshold, it is determined that the iteration condition is met, the iteration stops, and a trained graph alignment learning model is obtained. Based on the comparison of the final loss function, the cumulative iteration count, and the iteration count threshold, this application can accurately train and generate the required fake news detection model for identifying fake news.

[0147] In some optional implementations of this embodiment, after the step of determining whether the cumulative number of iterations is greater than a preset iteration number threshold, step S206 further includes:

[0148] If the cumulative number of iterations is less than the iteration number threshold, obtain the first loss value of the current iteration process.

[0149] In this embodiment, the first loss value mentioned above refers to the loss value calculated based on the final loss function during the current iteration.

[0150] Obtain the second loss value of the previous iteration process adjacent to the current iteration process.

[0151] In this embodiment, the second loss value refers to the loss value calculated based on the final loss function in the previous iteration process adjacent to the current iteration process.

[0152] Calculate the difference between the first loss value and the second loss value.

[0153] In this embodiment, the aforementioned difference may specifically refer to the absolute value of the difference between the first loss value and the second loss value.

[0154] Determine whether the difference is less than a preset difference threshold.

[0155] In this embodiment, the value of the above-mentioned difference threshold is not specifically limited and can be set according to actual usage requirements.

[0156] If so, the iteration condition is met, the iteration stops, and the trained graph alignment learning model is obtained.

[0157] In this embodiment, the model has been trained when the iteration stops.

[0158] This application, when detecting that the cumulative number of iterations is less than the iteration threshold, obtains a first loss value for the current iteration process; then obtains a second loss value for the previous iteration process adjacent to the current iteration process; subsequently calculates the difference between the first loss value and the second loss value; and then determines whether the difference is less than a preset difference threshold; if so, it determines that the iteration condition is met, the iteration stops, and a trained graph comparison learning model is obtained. This application, based on the comparison of the final loss function, the cumulative number of iterations, the iteration threshold, and the difference between the loss values ​​of two adjacent iteration processes, can accurately train and generate the required fake news detection model for identifying fake news.

[0159] It should be emphasized that, to further ensure the privacy and security of the aforementioned fake news detection model, the model can also be stored in a blockchain node.

[0160] The blockchain referred to in this application is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0161] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware with computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When executed, the program can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0162] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0163] Further reference Figure 3 As a response to the above Figure 2 The implementation of the method shown in this application provides an embodiment of an artificial intelligence-based fake news detection device, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0164] like Figure 3 As shown, the artificial intelligence-based fake news detection device 300 described in this embodiment includes: a first construction module 301, a second construction module 302, a generation module 303, a third construction module 304, a construction module 305, a first processing module 306, and a second processing module 307. Wherein:

[0165] The first construction module 301 is used to input pre-collected news training text data into the input layer of a preset graph comparison learning model, and to construct a heterogeneous graph corresponding to the news training text data in the input layer.

[0166] The second construction module 302 is used to input the heterogeneous graph into a preset construction layer in the graph alignment learning model, and to construct an augmented graph corresponding to the heterogeneous graph in the construction layer;

[0167] The generation module 303 is used to input the heterogeneous graph and the augmented graph into the encoder in the graph alignment learning model, generate a first low-dimensional vector of news text corresponding to the heterogeneous graph, and generate a second low-dimensional vector of news text corresponding to the augmented graph in the encoder.

[0168] The third construction module 304 is used to perform entity recognition on the news training text data to obtain the corresponding entity nouns, and to construct attribute features corresponding to the entity nouns based on a preset corpus.

[0169] The construction module 305 is used to input the first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features into a preset construction layer in the graph comparison learning model, and construct a first target news text low-dimensional vector corresponding to the heterogeneous graph and a second target news text low-dimensional vector corresponding to the augmented graph in the construction layer.

[0170] The first processing module 306 is used to construct a final loss function based on the low-dimensional vector of the first target news text and the low-dimensional vector of the second target news text, and to continuously iterate and minimize the final loss function until the preset iteration condition is met and the iteration stops, so as to obtain a trained graph comparison learning model.

[0171] The second processing module 307 is used to use the trained graph alignment learning model as a fake news detection model, and to perform false prediction processing on the news text data to be tested through the fake news detection model.

[0172] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based fake news detection method in the aforementioned implementation method, and will not be repeated here.

[0173] In some optional implementations of this embodiment, the second construction module 302 includes:

[0174] The first acquisition submodule is used to input the heterogeneous graph into the construction layer of the graph alignment learning model, and to acquire the set of edges in the heterogeneous graph in the construction layer.

[0175] The second acquisition submodule is used to acquire preset values;

[0176] The deletion submodule is used to randomly delete edges in the edge set within the heterogeneous graph based on the preset value, so as to obtain the processed heterogeneous graph.

[0177] The first determining submodule is used to use the processed heterogeneous graph as the augmented graph.

[0178] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based fake news detection method in the aforementioned implementation method, and will not be repeated here.

[0179] In some optional implementations of this embodiment, the third construction module 304 includes:

[0180] The third submodule is used to obtain a preset quantity;

[0181] The retrieval submodule is used to retrieve the corpus based on the entity nouns and extract the target sentences corresponding to the preset number from the corpus;

[0182] The output submodule is used to input the target sentence into a preset long short-term memory network and output a target vector corresponding to the target sentence based on the long short-term memory network.

[0183] The second determining submodule is used to use the target vector as the attribute feature corresponding to the entity noun.

[0184] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based fake news detection method in the aforementioned implementation method, and will not be repeated here.

[0185] In some optional implementations of this embodiment, the construction module 305 includes:

[0186] The first generation submodule is used to input the first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features into the construction layer of the graph comparison learning model. In the construction layer, a comparison vector corresponding to the attribute features is generated based on a preset comparison function.

[0187] The first construction submodule is used to construct a comparison feature vector corresponding to the comparison vector;

[0188] The second generation submodule is used to generate a low-dimensional vector of the first target news text corresponding to the heterogeneous graph based on the comparison feature vector and the first news text low-dimensional vector.

[0189] The third generation submodule is used to generate a low-dimensional vector of the second target news text corresponding to the augmented image based on the comparison feature vector and the low-dimensional vector of the second news text.

[0190] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based fake news detection method in the aforementioned implementation method, and will not be repeated here.

[0191] In some optional implementations of this embodiment, the first processing module 306 includes:

[0192] The first calculation submodule is used to calculate a first loss function based on the low-dimensional vector of the first target news text and the low-dimensional vector of the second target news text.

[0193] The fourth generation submodule is used to input the low-dimensional vector of the first target news text into the fully connected layer in the graph comparison learning model, and generate a predicted label corresponding to the low-dimensional vector of the first target news text based on a preset activation function.

[0194] The fourth acquisition submodule is used to acquire the real tags corresponding to the news training text data;

[0195] The second calculation submodule is used to calculate a second loss function based on the predicted label and the true label;

[0196] The second construction submodule is used to construct the final loss function based on the first loss function and the second loss function.

[0197] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based fake news detection method in the aforementioned implementation method, and will not be repeated here.

[0198] In some optional implementations of this embodiment, the first processing module 306 includes:

[0199] The fifth submodule is used to continuously iterate and minimize the final loss function, and obtain the current cumulative number of iterations;

[0200] The first judgment submodule is used to determine whether the cumulative number of iterations is greater than a preset iteration number threshold;

[0201] The first determination submodule is used to determine that the iteration condition is met if the cumulative number of iterations is greater than the iteration number threshold, and to stop the iteration to obtain the trained graph alignment learning model.

[0202] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based fake news detection method in the aforementioned implementation method, and will not be repeated here.

[0203] In some optional implementations of this embodiment, the first processing module 306 further includes:

[0204] The sixth acquisition submodule is used to acquire the first loss value of the current iteration process if the cumulative number of iterations is less than the iteration number threshold.

[0205] The seventh acquisition submodule is used to acquire the second loss value of the previous iteration process adjacent to the current iteration process;

[0206] The third calculation submodule is used to calculate the difference between the first loss value and the second loss value;

[0207] The second judgment submodule is used to determine whether the difference is less than a preset difference threshold.

[0208] The second decision submodule is used to determine if the iteration condition is met, stop the iteration, and obtain the trained graph alignment learning model.

[0209] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based fake news detection method in the aforementioned implementation method, and will not be repeated here.

[0210] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0211] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected via a system bus. It should be noted that only the computer device 4 with components 41-43 is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0212] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0213] The memory 41 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 41 may be an internal storage unit of the computer device 4, such as the hard disk or memory of the computer device 4. In other embodiments, the memory 41 may also be an external storage device of the computer device 4, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 4. Of course, the memory 41 may also include both the internal storage unit and its external storage device of the computer device 4. In this embodiment, the memory 41 is typically used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions for artificial intelligence-based fake news detection methods. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or will be output.

[0214] In some embodiments, the processor 42 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. The processor 42 is typically used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to execute computer-readable instructions stored in the memory 41 or to process data, for example, to execute computer-readable instructions for the artificial intelligence-based fake news detection method.

[0215] The network interface 43 may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the computer device 4 and other electronic devices.

[0216] This application also provides another embodiment, namely, providing a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor to cause the at least one processor to perform the steps of the artificial intelligence-based fake news detection method described above.

[0217] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A method for detecting fake news based on artificial intelligence, characterized in that, Includes the following steps: The pre-collected news training text data is input into the input layer of a preset graph comparison learning model, and a heterogeneous graph corresponding to the news training text data is constructed in the input layer. The heterogeneous graph is input into a preset construction layer within the graph alignment learning model, where an augmented graph corresponding to the heterogeneous graph is constructed. The heterogeneous graph and the augmented graph are input into the encoder in the graph alignment learning model, and a first low-dimensional vector of news text corresponding to the heterogeneous graph and a second low-dimensional vector of news text corresponding to the augmented graph are generated in the encoder. Entity recognition is performed on the news training text data to obtain the corresponding entity nouns, and attribute features corresponding to the entity nouns are constructed based on a preset corpus; The first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features are input into a preset construction layer in the graph comparison learning model. In the construction layer, a first target news text low-dimensional vector corresponding to the heterogeneous graph is constructed, and a second target news text low-dimensional vector corresponding to the augmented graph is constructed. Based on the low-dimensional vectors of the first and second target news texts, a final loss function is constructed, and the final loss function is continuously minimized iteratively until the preset iteration conditions are met, at which point the iteration stops, and a trained graph alignment learning model is obtained. The trained graph alignment learning model is used as a fake news detection model, and the fake news detection model is used to perform false prediction processing on the news text data to be tested. The steps of inputting the first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features into a preset construction layer within the graph comparison learning model, constructing a first low-dimensional vector of the target news text corresponding to the heterogeneous graph, and constructing a second low-dimensional vector of the target news text corresponding to the augmented graph within the construction layer, specifically include: The first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features are input into the construction layer of the graph comparison learning model. In the construction layer, a comparison vector corresponding to the attribute features is generated based on a preset comparison function. Construct a comparison feature vector corresponding to the comparison vector; Based on the contrast feature vector and the first news text low-dimensional vector, a first target news text low-dimensional vector corresponding to the heterogeneous graph is generated. Based on the contrast feature vector and the low-dimensional vector of the second news text, a low-dimensional vector of the second target news text corresponding to the augmented image is generated.

2. The artificial intelligence-based fake news detection method according to claim 1, characterized in that, The step of inputting the heterogeneous graph into a preset construction layer within the graph alignment learning model, and constructing an augmented graph corresponding to the heterogeneous graph in the construction layer, specifically includes: The heterogeneous graph is input into the construction layer within the graph alignment learning model, and the set of connected edges within the heterogeneous graph is obtained in the construction layer. Get the preset value; Based on the preset value, the edges of the edge set in the heterogeneous graph are randomly deleted to obtain the processed heterogeneous graph. The processed heterogeneous graph is used as the augmented graph.

3. The artificial intelligence-based fake news detection method according to claim 1, characterized in that, The step of constructing attribute features corresponding to the entity nouns based on a preset corpus specifically includes: Get the preset quantity; The corpus is searched based on the entity nouns, and target sentences corresponding to the preset number are extracted from the corpus. The target sentence is input into a preset long short-term memory network, and the target vector corresponding to the target sentence is output based on the long short-term memory network. The target vector is used as the attribute feature corresponding to the entity noun.

4. The artificial intelligence-based fake news detection method according to claim 1, characterized in that, The step of constructing the final loss function based on the low-dimensional vectors of the first and second target news texts specifically includes: A first loss function is calculated based on the low-dimensional vectors of the first and second target news texts. The low-dimensional vector of the first target news text is input into the fully connected layer in the graph alignment learning model, and a predicted label corresponding to the low-dimensional vector of the first target news text is generated based on a preset activation function. Obtain the real tags corresponding to the news training text data; Calculate a second loss function based on the predicted label and the true label; The final loss function is constructed based on the first loss function and the second loss function.

5. The artificial intelligence-based fake news detection method according to claim 1, characterized in that, The step of iteratively minimizing the final loss function until a preset iteration condition is met, and then stopping the iteration to obtain the trained graph alignment learning model, specifically includes: The final loss function is minimized iteratively, and the current cumulative number of iterations is obtained. Determine whether the cumulative number of iterations is greater than a preset iteration number threshold; If the cumulative number of iterations is greater than the iteration threshold, the iteration condition is satisfied, the iteration stops, and the trained graph alignment learning model is obtained.

6. The artificial intelligence-based fake news detection method according to claim 5, characterized in that, After the step of determining whether the cumulative number of iterations is greater than a preset iteration number threshold, the method further includes: If the cumulative number of iterations is less than the iteration number threshold, obtain the first loss value of the current iteration process; Obtain the second loss value of the previous iteration process adjacent to the current iteration process; Calculate the difference between the first loss value and the second loss value; Determine whether the difference is less than a preset difference threshold; If so, the iteration condition is met, the iteration stops, and the trained graph alignment learning model is obtained.

7. A fake news detection device based on artificial intelligence, characterized in that, include: The first construction module is used to input pre-collected news training text data into the input layer of a preset graph comparison learning model, and to construct a heterogeneous graph corresponding to the news training text data in the input layer. The second construction module is used to input the heterogeneous graph into a preset construction layer in the graph alignment learning model, and to construct an augmented graph corresponding to the heterogeneous graph in the construction layer; The generation module is used to input the heterogeneous graph and the augmented graph into the encoder in the graph alignment learning model, and generate a first low-dimensional vector of news text corresponding to the heterogeneous graph and a second low-dimensional vector of news text corresponding to the augmented graph in the encoder. The third construction module is used to perform entity recognition on the news training text data to obtain the corresponding entity nouns, and to construct attribute features corresponding to the entity nouns based on a preset corpus. The construction module is used to input the first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features into a preset construction layer in the graph comparison learning model, and to construct a first target news text low-dimensional vector corresponding to the heterogeneous graph and a second target news text low-dimensional vector corresponding to the augmented graph in the construction layer. The first processing module is used to construct a final loss function based on the low-dimensional vector of the first target news text and the low-dimensional vector of the second target news text, and to continuously iterate and minimize the final loss function until the preset iteration condition is met and the iteration stops, so as to obtain a trained graph comparison learning model. The second processing module is used to use the trained graph alignment learning model as a fake news detection model, and to perform false prediction processing on the news text data to be tested through the fake news detection model. The construction module includes: The first generation submodule is used to input the first low-dimensional vector of the news text, the second low-dimensional vector of the news text, and the attribute features into the construction layer of the graph comparison learning model. In the construction layer, a comparison vector corresponding to the attribute features is generated based on a preset comparison function. The first construction submodule is used to construct a comparison feature vector corresponding to the comparison vector; The second generation submodule is used to generate a low-dimensional vector of the first target news text corresponding to the heterogeneous graph based on the comparison feature vector and the first news text low-dimensional vector. The third generation submodule is used to generate a low-dimensional vector of the second target news text corresponding to the augmented image based on the comparison feature vector and the low-dimensional vector of the second news text.

8. A computer device comprising a memory and a processor, the memory storing computer-readable instructions, wherein the processor, when executing the computer-readable instructions, implements the steps of the artificial intelligence-based fake news detection method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the artificial intelligence-based fake news detection method as described in any one of claims 1 to 6.

Images