Rumor detection method and device fusing in-sample curriculum learning and dual-encoding contrastive learning

Abstract

The application provides a rumor detection method and device fusing in-sample curriculum learning and double-encoding contrast learning, and belongs to the field of natural language processing; solves the problems of poor accuracy, low quality of extracted text features and mismatched feature distribution of existing rumor detection methods; the method comprises the following steps: constructing a rumor propagation graph; generating a new rumor propagation dataset by using an in-sample curriculum learning strategy; encoding the new rumor propagation dataset through a graph convolution network to obtain a new rumor propagation graph structure representation; extracting the theme of the historical posts of the source post user; performing text representation on the extracted historical posts by using double-encoding contrast learning; aligning the rumor propagation graph structure feature and the text feature of the historical posts of the user in two heterogeneous modalities by using a semantic alignment method; generating the specific category of the post; constructing a corresponding loss function to realize model training; and the application is applied to rumor detection.

Description

Technical Field

[0001] This application relates to the field of rumor detection technology in natural language processing, and in particular to a rumor detection method and apparatus that integrates in-sample curriculum learning and dual-encoding contrastive learning. Background Technology

[0002] With the advancement of artificial intelligence and big data technologies, social media has gradually reshaped people's daily lives. While the development of social media has provided new avenues for the public to access information, many rumors have also spread rapidly on these platforms. Research shows that rumors not only disrupt public opinion but can also incite significant social antagonism and even cause serious economic losses. To address this real challenge, the development of efficient rumor detection technologies is particularly urgent. Such methods will help mitigate the negative social impact of the spread of misinformation on social media platforms.

[0003] Rumor detection is a text classification and reasoning task aimed at verifying the authenticity of information. Its core objective is to identify unverified, false, or misleading information in online texts (such as social media posts, news, and forum comments), essentially a binary classification task. Early rumor detection methods primarily focused on designing manually generated features and combining them with traditional classification algorithms. However, these methods are generally time-consuming, rely heavily on human experience in specific scenarios, and have poor generalization capabilities. With the continuous development of society, higher demands have been placed on the efficiency and accuracy of information classification, rendering traditional manual feature extraction methods inadequate. Deep learning-based rumor detection, due to its ability to automatically extract features and efficiently identify rumors, has become one of the most efficient approaches to solving this problem. Early deep learning-based rumor detection methods mostly focused on the text modality. These methods rely on the semantics of the text and pay attention to the linguistic features of rumors. The neural networks used for this modality mainly include convolutional neural networks, recurrent neural networks, and Transformer models. However, in social networks, the generation and spread of rumors is a dynamic process. Research shows that rumors often spread faster and wider than genuine news, exhibiting different diffusion tree structures. To effectively capture this evolving propagation pattern over time, we need to construct a rumor propagation graph by organizing discrete comments and reposts into specific propagation paths. This graph-structured data can then be used to simulate the spread of rumors on social networks. In rumor detection research, graph convolutional networks are a typical example of analytical methods targeting rumor propagation graphs. These graph neural network models can effectively capture the topological information formed during the propagation of rumors, thereby accurately extracting structural features that contain propagation patterns.

[0004] Currently, most models use text and rumor propagation graphs, and a small number also incorporate information such as user preferences and historical behavior features. However, they have the following problems: (1) Since rumors have their own unique characteristics in different fields, and they show significant domain differences in terms of propagation graph structure, vocabulary composition and user preferences, rumor identification requires professional knowledge and multi-faceted verification to make accurate judgments. Therefore, the datasets collected from the Internet inevitably contain noise. This noise manifests as irrelevant text, missing content or incorrect labeling, which greatly affects the performance of the model. Especially in the rumor propagation graph, noise not only changes the diffusion trend, but also shows the characteristic of accumulating with depth. As the propagation structure extends, deep nodes are often filled with a lot of irrelevant information, which leads to a significant reduction in their semantic relevance to the source rumor. Although existing studies remove noise from nodes or edges through graph structure optimization, pruning or preprocessing, these methods often face problems such as weak generalization ability, strong data dependence and insufficient flexibility. (2) The quality of text feature extraction is not high, including the popular BERT model used to extract text features. Due to the large number of parameters in the BERT model, it suffers from high computational resource consumption. Furthermore, the top-level sentence representation of BERT is not good at capturing the semantics of sentences, which makes it easy to be affected by text representation learning through BERT. (3) Since the rumor propagation graph focuses on structural features, while the post text focuses on text semantics, the two belong to different data modalities, and direct fusion can easily lead to feature distribution mismatch. Therefore, it is necessary to find an efficient rumor detection method to solve these problems. Summary of the Invention

[0005] To address the problems of poor accuracy, low quality of extracted text features, and mismatched feature distribution in existing rumor detection methods, this application proposes a rumor detection method and apparatus that integrates in-sample curriculum learning and dual-encoding contrastive learning.

[0006] The technical solution adopted in this application is: a rumor detection method that integrates in-sample curriculum learning and dual-encoding contrastive learning, comprising the following steps:

[0007] Step 1: Construct a rumor propagation graph: Construct an undirected graph showing the relationship between each source post and its comments and retweets in the original dataset containing multiple rumor events. In the graph, nodes represent the content of the tweet, and edges represent the actions of retweeting / replying.

[0008] Step 2: Generate a new rumor propagation dataset using an in-sample course learning strategy;

[0009] Step 3: Encode the new rumor propagation dataset using a graph convolutional network to obtain a new graph structure representation of rumor propagation;

[0010] Step 4: Extract topics from the historical posts of the source post users using the Latent Dirichlet Distribution;

[0011] Step 5: Use dual-encoding contrastive learning to perform text representation on the extracted historical posts. This text representation includes word-level features and post-level features.

[0012] Step 6: Use a semantic alignment method to align the two heterogeneous modalities: the rumor propagation graph structural features from Step 3 and the textual features of user history posts extracted in Step 5.

[0013] Step 7: Generate the specific category of the post;

[0014] Step 8: Construct the corresponding loss function to train the model.

[0015] Furthermore, each rumor event is a triple, including the source post, the propagation graph structure of the source post, and the historical posts published by the user of the source post, with the source post serving as the root node in the rumor propagation graph.

[0016] Furthermore, the specific implementation steps of step 2 are as follows:

[0017] Step 2.1: Calculate the depth of all nodes in the rumor propagation graph using the breadth-first search method, and take the maximum depth among the nodes as the endpoint of the course learning;

[0018] Step 2.2: Determine the current training depth and the distance from each node to the root node;

[0019] Step 2.3: Dynamically prune the rumor propagation graph and generate a new rumor propagation dataset.

[0020] Furthermore, the specific implementation steps of step 3 are as follows:

[0021] Step 3.1: Use the word vectors of the tweet text as initial node features;

[0022] Step 3.2: Capture the information diffusion flow through a graph convolutional network, continuously strengthen the source tweets during the convolution process, and finally aggregate the node-level features into graph-level features.

[0023] Furthermore, the specific implementation steps of step 5 are as follows:

[0024] Step 5.1: Hierarchical feature extraction. The lightweight BERT encoder is used to process the text in a two-level structure of "word-post". First, word-level encoding is used to capture the contextual relationship between words to generate the feature sequence of each user's historical posts. Then, post-level encoding is used to process the feature sequence output by word level to capture the temporal or logical relationship between posts and generate the user's global historical features.

[0025] Step 5.2: Parallel processing with dual encoders. The same input is encoded by two BERT encoders to obtain two sets of word-level features and post-level features;

[0026] Step 5.3: Cross-level attention fusion. First, a linear layer is used to map post-level features to the same dimension as word-level features. Second, the word-level feature sequence is concatenated with the aligned post-level features in the sequence dimension. Then, multi-head attention is used to interact with the concatenated sequence, allowing the global features to be weighted and fused with each post feature. Finally, the fused sequence is averaged to generate the final fused representation for each BERT encoder.

[0027] Step 5.4: Use normalized temperature-scaled cross-entropy loss as the optimization objective and introduce a hierarchical contrastive learning mechanism: By calculating the similarity between sample embeddings in different network layers, the semantic discrimination ability of each layer's representation is systematically enhanced, thereby enabling the model to learn more accurate sentence semantic representations;

[0028] Step 5.5: Generate the final representation by concatenating the fused representations generated by the two BERT encoders together as the text representation of the user's historical posts.

[0029] Furthermore, the hierarchical contrastive learning mechanism in step 5.4 is trained using the following loss function.

[0030] ;

[0031] in, This indicates the selected layer blocks used for contrastive learning. Indicates the first Hyperparameters of each layer block and They represent word level respectively. and Or post level and , and Representing attention fusion features and ;

[0032] Let represent the normalized temperature-scale cross-entropy loss function, which is defined as follows:

[0033] ;

[0034] in, It is a temperature coefficient that controls the model's ability to distinguish negative samples. Represent two vectors and The cosine similarity values ​​are generated by Transformer's standard dropout. and They represent word-level features respectively. and or post-level features and or attention fusion features and , This represents the index of all samples in the batch. This represents the index of the specific sample for which the loss is currently being calculated.

[0035] Furthermore, the specific implementation steps of step 6 are as follows:

[0036] Step 6.1: Represent the two modal features, namely the rumor propagation graph structure feature from Step 3 and the text feature of user history posts extracted in Step 5, using vector representation;

[0037] Step 6.2: Map the two modal features to a unified shared semantic space. By maximizing the cosine similarity between the two vectors, the model learns the semantic consistency between the two modalities, thereby achieving complementary enhancement of information from different perspectives.

[0038] Furthermore, the specific implementation steps of step 7 are as follows:

[0039] Step 7.1: Combine the rumor propagation graph features extracted in Step 3.2 with the user history text features extracted in Step 5.5;

[0040] Step 7.2: Apply the ReLU activation function to the concatenated feature vectors to introduce nonlinearity;

[0041] Step 7.3: Input the activated features into the fully connected layer to map the high-dimensional features to the classification representation space.

[0042] Furthermore, the total loss function in step 8 is expressed as follows:

[0043] ;

[0044] in, Represents classification loss. Represents semantic alignment loss. Indicates the layered contrast loss. , , This represents the hyperparameters that balance the three loss functions.

[0045] A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0046] The advantages of this application over the prior art are as follows:

[0047] 1. This application proposes a rumor detection method that integrates in-sample curriculum learning and dual-encoding contrastive learning. Based on social media posts, it uses in-sample curriculum learning and dual-encoding contrastive learning to obtain post representations of different dimensions. Then, it uses semantic alignment operation and feature fusion to obtain a comprehensive post representation, thereby realizing rumor detection.

[0048] 2. This application designs an in-sample course learning mechanism. This mechanism distinguishes the easy and difficult parts of the propagation graph based on the node difficulty coefficient, retaining only the shallow structure in the early stage of training, and then gradually introducing the deep structure as the training progresses. This method of dynamically aligning graph topology growth with model capabilities simulates the rumor propagation process and effectively improves the model's ability to handle noisy data.

[0049] 3. This application proposes a dual-encoder contrastive learning module. This module utilizes two lightweight BERT encoders to extract word-level contextual features and post-level temporal logical features from the same user's historical posts, respectively, and dynamically fuses them through a multi-head attention mechanism to capture user historical behavior. Simultaneously, this method constructs a multi-level contrastive mechanism, aligning features between the two encoders at the word level, post level, and the fused global level. Compared to existing techniques that only focus on document-level representations, this multi-granularity contrastive strategy enables the model to consider both local keywords and global semantics, generating more discriminative text representations.

[0050] 4. The semantic alignment method designed in this application can unify different modalities into a shared feature space, ensuring that the rumor representation learned by the model contains both rich content semantics and propagation topology information, thereby improving the accuracy of classification. Attached Figure Description

[0051] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0052] Figure 1 A dual-modal fusion framework diagram for implementing the overall method provided in the embodiments of this application;

[0053] Figure 2 A flowchart of in-sample course learning provided for embodiments of this application;

[0054] Figure 3 A flowchart of dual-encoding comparison learning provided in the embodiments of this application;

[0055] Figure 4This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0056] like Figures 1 to 4 As shown, this application provides a rumor detection method that integrates in-sample curriculum learning and dual-encoding contrastive learning, mainly addressing the problem that current text feature extraction methods produce low-quality text features, and that noise in the dataset significantly impacts the representation of the rumor propagation graph. Therefore, this method first employs in-sample curriculum learning and the in-sample curriculum learning strategy within the GCN module, along with a graph convolutional network, to obtain the graph structure representation of the post. Secondly, it utilizes dual-encoding contrastive learning in the dual-encoding contrastive learning module to extract user historical post information from the source post, generating user historical tendency features. Then, a semantic alignment method maps the two heterogeneous features to a unified shared semantic space and concatenates them into the final vector representation. This method is a dual-modal fusion framework, including embedding representation, in-sample curriculum learning, graph convolutional networks, dual-encoding contrastive learning, semantic alignment, and fully connected layers. Figure 1 The corresponding framework diagram is shown.

[0057] The method proposed in this application is based on an in-sample course learning strategy and a dual-encoding contrastive learning method. The specific steps are as follows:

[0058] Step 1: Construct a rumor propagation graph. Based on the relationship between the source post and its comments and reposts, build an undirected graph where nodes represent the content of the tweet and edges represent repost / reply actions. The specific implementation steps are as follows.

[0059] First, obtain the original dataset. ,in Representing the A rumor incident The total number of events. For each event... Characterize it as a triple = , , ,in This refers to the original post. This is the propagation graph structure of the post. These are historical posts published by the user who created the original post. (Based on the original post...) Construct an undirected graph based on the relationship between its comments and reposts. As a rumor-spreading image, in the rumor-spreading image, the post It is the root node, the node set Other nodes in the set consist of interactive posts such as comments and reposts. It is formed by interactive behaviors such as replying and quoting between posts.

[0060] Step 2: Generate a new rumor propagation graph dataset using an in-sample course learning strategy, specifically including:

[0061] Step 2.1: Calculate the depth of all nodes in the rumor propagation graph using the breadth-first search (BFS) method, and take the maximum depth among the nodes as the endpoint of the course learning;

[0062] Step 2.2: Determine the current training depth and the distance from each node to the root node;

[0063] Step 2.3: Dynamically prune the propagation graph and generate a new dataset.

[0064] Step 2 is implemented in the following specific way in this embodiment:

[0065] For each rumor propagation graph sample in the original dataset The depth of each node is calculated using a breadth-first search. According to hyperparameters (The diagram shows increased depth) and the propagation depth of the previous training round. Calculate the current maximum allowed propagation depth Based on the current maximum allowed propagation depth Preserve the root node and Less than or equal to The nodes and edges associated with these nodes are used to generate a new rumor propagation graph. ,in This represents the set of nodes in the new rumor propagation graph. This represents the edge set in the new rumor propagation graph.

[0066] Step 3: Encode the new dataset using a graph convolutional network to obtain a graph structure representation of the new dataset, specifically including:

[0067] Step 3.1: Use the word vectors of the tweet text as initial node features;

[0068] Step 3.2: Capture the information diffusion flow through a graph convolutional network, continuously strengthen the source tweets during the convolution process, and finally aggregate the node-level features into graph-level features.

[0069] Step 3 is implemented in the following specific way in this embodiment:

[0070] Regarding the new rumor spread diagram , where nodes Node feature vectors , This refers to the feature dimension. Let be the set of real numbers. Considering the _th Layered graph convolutional network, nodes In the eigenvectors of the layer It can be represented as:

[0071] ;

[0072] in, Representing a parameter matrix, the feature transformation involves two steps: element-wise mean pooling. and activation function , For nodes In the Feature vectors of layer -1 For nodes The neighboring nodes, To contain only nodes The collection of itself, Represents the set of neighboring nodes, where each node in the set is related to another node. Adjacent. Initialization Then, aggregate the first The node features of the layer are analyzed to obtain a global representation. :

[0073] .

[0074] Step 4: Extract topics from the user's historical posts of the source post using the Latent Dirichlet Distribution, specifically including:

[0075] Step 4.1: Filter and organize user's historical posts;

[0076] Step 4.2: Extract posts related to the topic of the source post from the user's historical posts using the Latent Dirichlet Distribution Model.

[0077] To accurately extract event source text through dual-encoding contrastive learning, a novel source post model is constructed by using latent Dirichlet distribution to filter historical posts by topic. The application uses a user's historical information dataset as its training basis. Due to the finite length of the sequences, it is impossible to process all posts within the same sequence. Therefore, this application uses each user's historical post set... Perform independent encoding and represent it as a matrix. As model input, where Indicates the user's first One post, Represents the number of posts. This represents the number of words in each post.

[0078] Step 5: Use dual-encoding contrastive learning to perform text representation on the extracted user history posts:

[0079] This application employs two lightweight BERT encoders to process the user's historical posting matrix. Each BERT encoder contains two algorithms based on word-level encoding and post-level encoding. The encoder architecture (word-level and post-level) consists of N stacked encoder layers. Each encoder layer contains two sub-layers: a multi-head self-attention mechanism (…). ) and location-based feedforward neural networks ( ).

[0080] Step 5.1: Hierarchical feature extraction. The lightweight BERT encoder is used to process the text in a two-level structure of "word-post". First, word-level encoding is used to capture the contextual relationship between words to generate the feature sequence of each user's historical posts. Then, post-level encoding is used to process the feature sequence output by word level to capture the temporal or logical relationship between posts and generate the user's global historical features.

[0081] The specific implementation principle of step 5.1 is as follows:

[0082] The first BERT encoder passed by word level Generate word-level features This involves capturing the contextual relationships between words within a single post. First, it involves analyzing the user's historical posting matrix. Initial input for generating word-level Transformers :

[0083] ;

[0084] ;

[0085] in, For the dimension of word embedding, It is The word vector matrix is ​​converted into a continuous vector and then scaled. It is a sine / cosine position encoding matrix, matrix Even-numbered indices use the sine function, and odd-numbered indices use the cosine function.

[0086] Assuming the current level is word level No. Layer encoder ( ), its input is the first -1 layer encoder output ,pass generate .

[0087] ;

[0088] in, Representing the Intermediate variables within the layer encoder.

[0089] Then, the output representation of the current layer is generated using FFN. :

[0090] ;

[0091] Vectors are generated after N layers of stacking. Then, layer normalization is performed to obtain the context representation. :

[0092] ;

[0093] in, This represents the number of layers in the encoder;

[0094] Key information is extracted and word-level features are generated using max pooling. :

[0095] ;

[0096] in, , L represents the number of posts, and L represents the length of words in each post.

[0097] Step 5.2: Parallel processing with dual encoders. The same input is encoded from two different perspectives to provide positive sample pairs for subsequent comparative learning. The data is processed by two BERT encoders to obtain two sets of word-level features and post-level features.

[0098] The specific implementation principle of step 5.2 is as follows:

[0099] Word-level features Perform post-level encoding to capture the temporal order or logical relationship between posts, making = As the initial input for a post-level Transformer:

[0100] Assuming the current step is the post-level Transformer... Layer encoder ( ), its input is the first -1 layer output ,pass generate .

[0101] ;

[0102] in, Representing the Intermediate variables within the layer encoder.

[0103] Then, the output representation of the current layer is generated using FFN. :

[0104] ;

[0105] Vectors are generated after N layers of stacking. Then, a final normalization is performed to obtain the context representation. :

[0106] ;

[0107] Key information is extracted and post-level features are generated using max pooling. :

[0108] ;

[0109] in, For word-level features, As a post-level feature, For post dimensions.

[0110] Another encoder generates word-level features and post level features The steps are the same as steps 5.1 and 5.2.

[0111] Step 5.3: Cross-level attention fusion. In order to combine local details and global information, firstly, a linear layer is used to map post-level features to the same dimension as word-level features. Secondly, the word-level feature sequence and the aligned post-level feature sequence are concatenated on the same dimension. Then, multi-head attention is used to interact with the concatenated sequence, allowing global features to be weighted and fused with each post feature. Finally, the fused sequence is averaged to generate the final fused representation of each BERT encoder.

[0112] Specifically, first use a linear layer to apply the global... , Mapping to , Using the same dimensions, the word-level features are then concatenated with the aligned post-level features:

[0113] = ;

[0114] = ;

[0115] in, , These are the features fused from two BERT encoders. Used to concatenate the input features.

[0116] Features fused from two BERT encoders , use The concatenated sequence is interactively processed to perform weighted fusion of global features and features from each local post, and finally the average is taken to obtain two fused representations. and :

[0117] ;

[0118] ;

[0119] Generate a text representation of the user's historical posts :

[0120] .

[0121] Step 5.4: In order to systematically evaluate and improve the semantic discriminative ability of each layer of representation, this application proposes a hierarchical contrastive learning method, which calculates the distance between the embedding vectors of sample pairs, enabling the model to learn more accurate semantic representations of sentences.

[0122] This hierarchical contrastive learning method uses normalized temperature-scaled cross-entropy loss as the optimization objective and introduces a hierarchical contrastive learning mechanism: by calculating the similarity between sample embeddings in different network layers, the semantic discriminative ability of each layer's representation is systematically enhanced, thereby enabling the model to learn more accurate sentence semantic representations.

[0123] The loss function for this method is shown below:

[0124] ;

[0125] in, Indicates the layered contrast loss. This indicates the selected layer blocks used for contrastive learning. Indicates the first Hyperparameters of each layer block and They represent word level respectively. and Or post level and Similarly, and Two augmented representations of the attention block output and . Let represent the normalized temperature-scale cross-entropy loss function, which is defined as follows:

[0126] ;

[0127] in, It is a temperature coefficient that controls the model's ability to distinguish negative samples. Represent two vectors and The cosine similarity values ​​are generated by the standard dropout of the Transformer. and These respectively represent the word-level features mentioned above. and or post-level features and or attention fusion features and . This represents the index of all samples in the batch. This represents the index of the specific sample for which the loss is currently being calculated.

[0128] Step 5.5: Generate the final representation by concatenating the fused representations generated by the two BERT encoders together as the user's historical text representation.

[0129] Step 6: Use semantic alignment to align the two heterogeneous modalities: the rumor propagation graph structural features and the user's historical text features.

[0130] Specifically, it includes the following steps:

[0131] Step 6.1: The two modal features about the rumor event obtained from the in-sample course learning and GCN module and the dual-encoding contrastive learning module are represented by vectors.

[0132] Step 6.2: Map the two modal features to a unified shared semantic space. By maximizing the cosine similarity between the two vectors (i.e. minimizing the alignment loss), the model learns the semantic consistency between the two modalities, thereby achieving complementary enhancement of information from different perspectives.

[0133] Step 7: Generate the specific category of the post.

[0134] Specifically, it includes the following steps:

[0135] Step 7.1: Combine the rumor propagation graph features extracted by GCN in Step 3.2 with the user history text features extracted by dual-encoding contrastive learning in Step 5.5;

[0136] Step 7.2: Apply the concatenated feature vectors Activation function, introducing nonlinear factors;

[0137] Step 7.3: Input the activated features into the fully connected layer to map the high-dimensional features to the classification representation space.

[0138] Specifically, the representations learned from the propagation graph structure Compared with the above text global representation By concatenating the components, we obtain the final hidden embedding representation of the event:

[0139] .

[0140] Hidden embedding of learned events Through a included Activation function and The network structure of the classification layer is mapped to the final classification representation space:

[0141] ;

[0142] in, This represents the probability distribution of labels across all categories. This represents the number of categories. and Let represent the learnable matrix and bias vector of the fully connected layer, respectively.

[0143] Step 8: Construct the corresponding loss function to train the model.

[0144] Step 8.1: Apply cross-entropy to obtain the classification loss:

[0145] ;

[0146] in, This represents the true probability under category c. Indicates category The predicted probability is as follows. This refers to the batch size.

[0147] Step 8.2: To improve model performance, it is necessary to effectively fuse heterogeneous features extracted from the rumor propagation graph and historical propagation trends. To achieve this goal, this application introduces semantic alignment loss. This loss constructs a shared semantic space by constraining the representations of the two types of features. Its form is as follows:

[0148] ;

[0149] in, Represented as a user's historical preference vector, This is represented as a vector diagram of rumor propagation.

[0150] Step 8.3: Total Loss The form is:

[0151] ;

[0152] in , , This represents the hyperparameters that balance the three loss functions.

[0153] To verify the effectiveness of the proposed method, this embodiment conducted comparative experiments with current mainstream rumor detection methods on two real social media datasets (such as Twitter15 and Twitter16). The baseline models cover a variety of technical approaches, including: GCN-based models BiGCN and DDGCN, contrastive learning-based models GACL, UICL, and RAGCL, causal inference-based model CausalRD, semantic mining-integrated model GSMRD, real-world scenario-oriented model AD-GSCL, and the latest dynamic architecture model DMPRN. Table 1 shows the performance of each model on these two datasets, with the best results highlighted in bold. The results demonstrate that the proposed method... 2 -RD achieves best performance across multiple metrics.

[0154] Table 1. Comparative experimental results.

[0155] .

[0156] This application proposes a rumor detection method that integrates in-sample curriculum learning and dual-encoding contrastive learning. Based on social media post information, it first constructs a rumor propagation graph using comments and reposts related to the original post. A new training sample set is generated through an in-sample curriculum learning strategy, which starts learning from shallow graph structures and gradually increases the depth of the graph structure as training progresses. Next, a graph convolutional network is used to extract features from the new propagation graph structure. Then, a hidden Dirichlet distribution model is used to extract topics from the historical posts of the source post's users, and dual-encoding contrastive learning is used to extract textual features from the user's historical posts. Following this, semantic alignment operations are used to share the propagation graph features and textual features into a unified feature space and concatenate them to generate a representation. Finally, a network structure containing a ReLU activation function and a Softmax classification layer is used to map these features to a classification representation space, improving the model's performance.

[0157] Figure 4 A structural block diagram of a computer device according to a specific embodiment of this application is shown. Figure 4As shown, the computer device includes a memory and a processor, the memory storing instructions executable on the processor. When the processor executes the instructions, it implements the methods described in the above embodiments. The number of memories and processors can be one or more. This computer device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The computer device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present application described and / or claimed herein.

[0158] The computer device may also include a communication interface for communicating with external devices and exchanging data. The devices are interconnected using different buses and can be mounted on a common motherboard or otherwise as needed. The processor can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as a display device coupled to the interface). In other embodiments, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple electronic devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). The bus can be divided into address buses, data buses, control buses, etc. For ease of illustration, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0159] Optionally, in a specific implementation, if the memory, processor, and communication interface are integrated on a single chip, then the memory, processor, and communication interface can communicate with each other through an internal interface.

[0160] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. General-purpose processors can be microprocessors or any conventional processor. It is worth noting that the processor can be a processor supporting advanced RISC machines (ARM) architecture.

[0161] This application provides a computer-readable storage medium (such as the memory described above) storing computer instructions that, when executed by a processor, implement the method provided in this application.

[0162] Optionally, the memory may include a stored program area and a stored data area, wherein the stored program area may store the operating system and application programs required for at least one function; the stored data area may store data created based on the use of the computer device for mapping. Furthermore, the memory may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the computer device for mapping via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0163] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A rumor detection method fusing in-sample curriculum learning and dual-encoding contrastive learning, characterized in that: Includes the following steps: Step 1: Construct a rumor propagation graph: Construct an undirected graph showing the relationship between each source post and its comments and retweets in the original dataset containing multiple rumor events. In the graph, nodes represent the content of the tweet, and edges represent the actions of retweeting / replying. Step 2: Generate a new rumor propagation dataset using an in-sample course learning strategy. The specific implementation steps are as follows: Step 2.1: Calculate the depth of all nodes in the rumor propagation graph using the breadth-first search method, and take the maximum depth among the nodes as the endpoint of the course learning; Step 2.2: Determine the current training depth and the distance from each node to the root node; Step 2.3: Dynamically prune the rumor propagation graph and generate a new rumor propagation dataset; Step 3: Encode the new rumor propagation dataset using a graph convolutional network to obtain a new graph structure representation of rumor propagation; Step 4: Extract the themes from the source post user's historical posts; Step 5: Use dual-encoding contrastive learning to perform text representation on the extracted historical posts. This text representation includes word-level features and post-level features. The specific implementation steps are as follows: Step 5.1: Hierarchical feature extraction. The lightweight BERT encoder is used to process the text in a two-level structure of "word-post". First, word-level encoding is used to capture the contextual relationship between words to generate the feature sequence of each user's historical posts. Then, post-level encoding is used to process the feature sequence output by word level to capture the time or logical relationship between posts and generate the user's global historical features. Step 5.2: Parallel processing with dual encoders. The same input is encoded by two BERT encoders to obtain two sets of word-level features and post-level features; Step 5.3: Cross-level attention fusion. First, a linear layer is used to map post-level features to the same dimension as word-level features. Second, the word-level feature sequence is concatenated with the aligned post-level features in the sequence dimension. Then, multi-head attention is used to interact with the concatenated sequence, allowing the global features to be weighted and fused with each post feature. Finally, the fused sequence is averaged to generate the final fused representation for each BERT encoder. Step 5.4: Use normalized temperature-scaled cross-entropy loss as the optimization objective and introduce a hierarchical contrastive learning mechanism: By calculating the similarity between sample embeddings in different network layers, the semantic discrimination ability of each layer's representation is systematically enhanced, thereby enabling the model to learn more accurate sentence semantic representations; Step 5.5: Generate the final representation by concatenating the fused representations generated by the two BERT encoders together as the text representation of the user's historical posts; Step 6: Use a semantic alignment method to align the two heterogeneous modalities: the rumor propagation graph structural features from Step 3 and the textual features of user history posts extracted in Step 5. Step 7: Generate the specific category of the post; Step 8: Construct the corresponding loss function to train the model.

2. The rumor detection method of claim 1, wherein: Each rumor event is a triple, which includes the source post, the propagation graph structure of the source post, and the historical posts published by the user who posted the source post. The source post is the root node in the rumor propagation graph.

3. The rumor detection method of claim 2, wherein: The specific implementation steps for step 3 are as follows: Step 3.1: Use the word vectors of the tweet text as initial node features; Step 3.2: Capture the information diffusion flow through a graph convolutional network, continuously strengthen the source tweets during the convolution process, and finally aggregate the node-level features into graph-level features.

4. The rumor detection method according to claim 3, which integrates in-sample curriculum learning and dual-encoding contrastive learning, is characterized in that: The hierarchical contrastive learning mechanism in step 5.4 is trained using the following loss function: ； in, This indicates the selected layer blocks used for contrastive learning. Indicates the first Hyperparameters of each layer block and They represent word level respectively. and Or post level and , and Representing attention fusion features and ; Let represent the normalized temperature-scale cross-entropy loss function, which is defined as follows: ； in, It is a temperature coefficient that controls the model's ability to distinguish negative samples. Represent two vectors and The cosine similarity values ​​are generated by Transformer's standard dropout. and They represent word-level features respectively. and or post-level features and or attention fusion features and , This represents the index of all samples in the batch. This represents the index of the specific sample for which the loss is currently being calculated.

5. The rumor detection method according to claim 4, which integrates in-sample curriculum learning and dual-encoding contrastive learning, is characterized in that: The specific implementation steps for step 6 are as follows: Step 6.1: Represent the two modal features, namely the rumor propagation graph structure feature from Step 3 and the text feature of user history posts extracted in Step 5, using vector representation; Step 6.2: Map the two modal features to a unified shared semantic space. By maximizing the cosine similarity between the two vectors, the model learns the semantic consistency between the two modalities, thereby achieving complementary enhancement of information from different perspectives.

6. The rumor detection method according to claim 5, which integrates in-sample curriculum learning and dual-encoding contrastive learning, is characterized in that: The specific implementation steps for step 7 are as follows: Step 7.1: Combine the rumor propagation graph features extracted in Step 3.2 with the user history text features extracted in Step 5.5; Step 7.2: Apply the ReLU activation function to the concatenated feature vectors to introduce nonlinearity; Step 7.3: Input the activated features into the fully connected layer to map the high-dimensional features to the classification representation space.

7. A rumor detection method integrating in-sample curriculum learning and dual-encoding contrastive learning according to any one of claims 1-6, characterized in that: The total loss function in step 8 is expressed as follows: ； in, Represents classification loss. Represents semantic alignment loss. Indicates the layered contrast loss. , , This represents the hyperparameters that balance the three loss functions.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method of claim 7.

Images