A social media early rumor detection method and system based on a PGQH-TE model

Abstract

This invention provides a method and system for early-stage rumor detection on social media based on the PGQH-TE model, belonging to the technical field of social media content security and information authenticity identification. It includes: encoding each early rumor event into a text modality and a grayscale image modality; extracting a global semantic prior vector from the text using a BERT model, and using this to guide an improved BiGRU network to extract text sequence features, an improved ResNet-18 network to extract visual structural features, and a specially designed parallel variational coding and bit-level adaptive fusion quantum hybrid network PVBF-QHN to extract quantum state features from the prior vector; projecting the text, visual, and quantum modal features into spatial point charges using a triangular electrostatic field Coulomb fusion method, and achieving lossless fusion of high-order cross-modal interactions based on electrostatic potential and potential gradient, thus realizing early-stage rumor detection on social media. This invention can achieve high-accuracy automatic identification of social media rumors in early-stage scenarios with highly sparse information.

Description

Technical Field

[0001] This invention relates to the field of social media content security and information authenticity verification technology, and in particular to a method and system for early rumor detection on social media based on the PGQH-TE model. Background Technology

[0002] Rumors are unverified or false information that spreads in society [1]. A complete social media rumor event usually includes the original post, related comments, retweets, the propagation network structure, and the user's social attributes. While social media improves the efficiency of information dissemination, it also enables false information to spread widely in a short period of time [2].

[0003] Depending on the timing of detection, social media rumor detection can be divided into late detection and early detection. Late detection[3] utilizes complete information of the propagation chain (including comments, reposts, propagation graphs, etc.), and has high detection accuracy, but is too delayed. Early detection[4] aims to intervene at the initial moment of rumor release, relying only on the limited information available at the time of release (such as the original content and information of the publisher), and strives to block the spread at the source. Existing early rumor detection methods can be roughly divided into three categories: (1) strict early detection, which uses only the original post and publisher information and makes judgments through multi-level text feature extraction or domain adaptive fusion[5], but lacks a high-confidence feature screening mechanism and fails to effectively capture high-order cross-modal associations; (2) early detection with limited propagation terms, which introduces early comments or shallow propagation graph structures[6], which improves accuracy but relies excessively on features that can only be observed after the propagation begins, which violates the original intention of blocking the source; (3) early detection based on data compensation, which introduces external knowledge graphs or generative content to supplement information[7], but the compensation evidence is mostly generated in a black box and lacks a transparent causal chain with the original post, resulting in poor interpretability.

[0004] In addition, quantum computing’s ability to represent high-dimensional semantic space through superposition and entanglement states provides a new paradigm for feature mining, and quantum hybrid networks have been initially introduced into rumor detection tasks [8]. However, due to the physical bottleneck of the current quantum hardware’s small number of qubits and high noise, existing quantum hybrid models generally adopt oversimplified circuit designs, which prevents the full release of the potential of quantum superposition and entanglement [9]; or blindly stacking quantum gates and entanglement layers, which leads to an exponential increase in circuit depth and noise accumulation, making inference delay unacceptable

[10] . Therefore, under the constraints of lightweight architecture, the potential of quantum feature mining has not been effectively utilized and it is difficult to support high-confidence classification decisions.

[0005] In summary, existing methods generally suffer from problems such as a lack of highly discriminative focus in feature extraction, difficulty in capturing and interpreting high-order cross-modal interactions through multimodal fusion, and insufficient expressive power of lightweight quantum networks in information-scarce scenarios. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a method and system for early social media rumor detection based on the PGQH-TE model, so as to solve the problems in the prior art, such as low feature confidence due to high information sparsity, difficulty in capturing multimodal high-order interactions and lack of interpretability, and limited feature mining capabilities of lightweight quantum hybrid networks.

[0007] The technical solution adopted by the embodiments of the present invention to solve its technical problem is as follows:

[0008] The first aspect of this invention provides a method for early rumor detection on social media based on the PGQH-TE model, comprising:

[0009] Step S1, construct an early rumor event dataset: select early social media rumor event data containing only the original published content and the publisher's public information, construct a bimodal early rumor detection dataset, and divide the dataset into training set, validation set, and test set;

[0010] Step S2: Construct the social media early rumor detection model PGQH-TE based on semantic prior guidance, triangular electrostatic field fusion, and quantum hybrid neural network, including: a preprocessing module, a global prior semantic extraction module BERT, a semantic prior guided sequence feature extraction network PG-BiGRU, a semantic prior guided visual structural feature extraction network PG-ResNet-18, a parallel variational coding and bit-level adaptive fusion quantum hybrid network PVBF-QHN, and a triangular electrostatic field Coulomb fusion module CoF-TE; train the PGQH-TE model using the aforementioned dual-modal early rumor detection dataset.

[0011] The preprocessing module encodes each early rumor event sample as a whole into a text modality, and generates a grayscale image modality of a specified size by filling text bytes and mapping rows and columns.

[0012] The global prior semantic extraction module performs feature extraction and classification on text modalities, and outputs a global semantic prior probability vector with high discriminative power.

[0013] The PG-BiGRU network concatenates the word vector sequence of the text modality with the global semantic prior probability vector, and extracts semantically guided sequence features through information transmission mechanism and feature-level self-attention.

[0014] The PG-ResNet-18 network receives grayscale image modalities, generates gating weights using the global semantic prior probability vector, and extracts semantically guided visual structural features through feature modulation.

[0015] The PVBF-QHN network receives the global semantic prior probability vector after dimensional transformation, and extracts high-dimensional quantum state features through a parallel dual quantum neural network block structure, a hybrid entanglement mechanism, and bit-level adaptive weighted fusion.

[0016] After the sequence features, visual structure features, and quantum state features are modulated to a unified dimension through a linear fully connected layer and an activation function, they are input into the CoF-TE module. By calculating the electrostatic potential and potential gradient under the triangular charge array, high-order cross-modal fusion features are obtained.

[0017] The fused features are input into the classification layer, and the early rumor detection results are output.

[0018] Step S3: Use the trained PGQH-TE model to perform real-time early rumor detection on newly published events on social media.

[0019] Preferably, the grayscale image modality generation in step S2 specifically includes: calculating the total number of bytes in the text modality file. Determine the side length of the smallest square image. Calculate the number of padding bytes Equidistant sampling from the original text Bytes are padded to the end; the padded data is then mapped byte by byte in line order. The initial grayscale image matrix; the initial grayscale image is uniformly scaled to a certain value using bilinear interpolation. Pixels, as grayscale image modalities.

[0020] Preferably, the specific processing flow of the global prior semantic extraction module in step S2 includes: inputting the text modality into the pre-trained BERT model to extract a 768-dimensional semantic feature vector; inputting the semantic feature vector into the logistic regression classification layer to output a 2-dimensional global semantic prior probability vector. .

[0021] Preferably, the PG-BiGRU network feature extraction in step S2 specifically includes: sampling and segmenting the text modality of early rumor event samples to obtain n word vector sequences. ; the global semantic prior probability vector The sequence is copied n times and alternately concatenated with n word vector sequences to form an augmented input sequence. The augmented input sequence is fed into a bidirectional GRU network, and the hidden states of the first and last time steps are extracted and concatenated to form sequence features. Based on sequence features Introduce learnable feature-level self-attention weight vectors Calculate semantically guided sequence features .

[0022] Preferably, the PG-ResNet-18 network feature extraction in step S2 specifically includes: inputting the grayscale image modality into the ResNet-18 network to extract an initial 512-dimensional visual feature vector. ; the global semantic prior probability vector Gated weight vectors are generated by sequentially passing through dimension-expanded fully connected layers and dimension-mapped fully connected layers. Introducing learnable weight vectors The semantically guided visual structural features are calculated using the feature modulation formula. .

[0023] Preferably, the PVBF-QHN network in step S2 adopts a 6-qubit structure, and the high-dimensional quantum state feature extraction specifically includes:

[0024] The 2D global semantic prior probability vector is first normalized and standardized, then scaled to 12D through a fully connected layer. After activation by the tanh activation function, it is fed into a fully connected layer, scaled to 6D, and then activated again by the tanh activation function to obtain the vector P = [p0, p1, p2, p3, p4, p5]. The vector P is then fed into two parallel quantum neural network blocks (QNB) with identical structures, outputting two 6D quantum features. and Each QNB sequentially includes an angle encoding layer, a learnable parameter layer, a locally adjacent entanglement layer, a globally annular entanglement layer, and a measurement layer. The angle encoding layer is initialized to... A Hadamard gate is applied to the qubit to generate a superposition state. The vector P is encoded into the superposition state using an RY rotation gate through angle encoding. Then, a trainable parameter layer consisting of RX, RY, and RZ rotation gates, a locally adjacent entangled layer consisting of fixed-structure CNOT gates, and a globally ring-shaped entangled layer consisting of learnable IsingZZ gates are sequentially applied. The measurement layer performs Pauli-Z measurements to obtain two 6-dimensional expectation vectors. and ;

[0025] Introduce learnable attention weight vectors and The 6-dimensional high-dimensional quantum state characteristics were calculated using a bit-level adaptive weighted fusion formula. .

[0026] Preferably, the CoF-TE processing procedure of the triangular electrostatic field coulomb fusion module in step S2 includes: , , Modulation to a uniform dimension The resulting sequence feature vector Visual feature vectors and quantum eigenvectors Further splicing into a charge sequence Define the three vertices of the triangular electrostatic field, calculate and insert the bilateral average charge (BMC) of the corresponding two adjacent edges, where the BMC is inserted at the vertex of the text edge-visual edge. , for The i-th element in the middle, for The j-th element is inserted at the vertex of the visual edge-quantum edge. , for Insertion at the i-th element of the quantum edge - text edge vertex Global mean normalization is performed on the charge sequence E; a Gaussian kernel function is used. Calculate the charge sequence E of the first charge. One charge electrostatic potential at ; Calculate the first Potential gradient at each charge The electrostatic potential sequence and the potential gradient sequence are concatenated to form a fused feature vector. =[ ,..., , ,..., ].

[0027] Preferably, the training strategy for the PGQH-TE model in step S2 specifically includes: first, independently pre-training the PVBF-QHN network and the PG-ResNet-18 network until the F1 score on the validation set converges and then freezing the parameters; then loading the BERT module, PG-BiGRU network, and CoF-TE module for joint training; using the AdamW optimizer and setting the joint training learning rate; using a cosine annealing learning rate scheduling strategy; using weighted cross-entropy loss as the loss function, applied to the output of the final classification layer; the classification layer outputs binary logits, which are converted into a probability distribution by the Softmax function, and the category corresponding to the maximum probability is taken as the early rumor detection result.

[0028] A second aspect of this invention provides a social media early rumor detection system based on the PGQH-TE model, comprising:

[0029] The dataset construction and preprocessing module is used to obtain the original content and metadata at the time of social media posting and generate bimodal data of text and grayscale images;

[0030] The model building module is used to construct the PGQH-TE model, which includes: a global prior semantic extraction submodule, a semantically guided sequence feature extraction submodule, a semantically guided visual feature extraction submodule, a parallel variational coding and bit-level adaptive fusion quantum hybrid network submodule, and a triangular electrostatic field Coulomb fusion submodule.

[0031] The model training module is used to perform phased pre-training and joint fine-tuning of the PGQH-TE model using the dataset;

[0032] The early rumor detection module is used to call the trained PGQH-TE model and output a classification result as to whether an early rumor event on the target social media is a rumor.

[0033] More preferably, it also includes:

[0034] The physical interpretability visualization module is used to extract the electrostatic potential distribution and potential gradient curve in the triangular electrostatic field Coulomb fusion submodule, generate the energy distribution difference map of rumor samples and non-rumor samples in the triangular charge field, and provide a physical interpretability basis for model detection decision.

[0035] As can be seen from the above technical solutions, the social media early rumor detection method and system based on the PGQH-TE model provided in this invention adopts a strict definition of early events, relying only on the original content at the time of publication; it extracts global semantic priors through BERT to guide PG-BiGRU and PG-ResNet-18 to achieve high-confidence feature extraction; it fully releases the potential for quantum state feature mining under a lightweight architecture through PVBF-QHN; and it captures high-order cross-modal interactions and endows them with physical interpretability through the CoF-TE triangular electrostatic field fusion method. This invention can quickly, accurately, and robustly achieve automatic identification of social media rumors in strictly early scenarios with highly sparse information, and is suitable for practical application scenarios such as network content security management, public opinion monitoring and early warning, and digital forensics. Attached Figure Description

[0036] Figure 1 The flowchart of the social media early rumor detection method based on the PGQH-TE model provided by the present invention.

[0037] Figure 2 This diagram illustrates the performance of the model for different grayscale image sizes.

[0038] Figure 3 A schematic diagram illustrating the visualization results of t-SNE for the features extracted by the model.

[0039] Figure 4 This is a schematic diagram showing the potential distribution and gradient curves of rumor and non-rumor samples in a triangular electrostatic field. Detailed Implementation

[0040] The technical solution and effects of the present invention will be further described in detail below with reference to the accompanying drawings.

[0041] This invention employs a strict definition of early events, utilizing only the original content and publicly available user information at the time of publication. It extracts global semantic priors using a BERT model, guiding an improved BiGRU network to extract text sequence features and an improved ResNet-18 network to extract visual structural features. A specially designed parallel variational coding and bit-level adaptive fusion quantum hybrid network (PVBF-QHN) extract high-quality quantum state features from the global semantic prior probability vector. An innovative triangular electrostatic field Coulomb fusion method (CoF-TE) projects the three-modal features into spatial point charges, achieving lossless fusion and physically interpretable representation of high-order cross-modal interactions based on electrostatic potential and potential gradient. Ultimately, it achieves high-precision end-to-end early rumor detection. This method achieves a peak accuracy of 98.39% on three public datasets: Twitter16, Weibo, and PHEME. In strictly early scenarios, it significantly outperforms existing state-of-the-art methods that rely on propagated information or external data, demonstrating significant advantages such as high classification accuracy, strong robustness, good physical interpretability, and strong cross-language generalization ability.

[0042] A method for early rumor detection on social media based on the PGQH-TE model, based on Figure 1 The implementation of the social media early rumor detection model PGQH-TE shown includes the following steps:

[0043] Step S1, construct an early rumor event dataset: select early social media rumor event data containing only the original published content and the publisher's public information, construct a bimodal early rumor detection dataset, and divide the dataset into training set, validation set, and test set;

[0044] Step S2, construct the PGQH-TE model based on semantic prior guidance, triangular electrostatic field fusion, and quantum hybrid neural network (where PG represents Prior Guided, QH represents Quantum Hybrid, and TE represents Triangular Electrostatic field), including: a preprocessing module ( Figure 1 Module a) Global Prior Semantic Extraction Module BERT Figure 1 Module b), Semantic Prior Guided Sequence Feature Extraction Network PG-BiGRU ( Figure 1 Module e), Semantic Prior Guided Visual Structural Feature Extraction Network PG-ResNet-18 ( Figure 1 Module d), Parallel variational coding and bit-level adaptive fusion quantum hybrid network PVBF-QHN ( Figure 1 Module c) and the triangular electrostatic field coulomb fusion module CoF-TE ( Figure 1 Module f); The PGC-TE model is trained using a bimodal early rumor detection dataset:

[0045] The preprocessing module encodes each early rumor event sample as a whole into a text modality, and generates a grayscale image modality of a specified size by padding text bytes and mapping rows and columns.

[0046] The global prior semantic extraction module extracts and classifies text modal features, and outputs a global semantic prior probability vector with high discriminative power.

[0047] The PG-BiGRU network concatenates the word vector sequence of the text modality with the global semantic prior probability vector, and extracts semantically guided sequence features through information transmission mechanism and feature-level self-attention;

[0048] The PG-ResNet-18 network receives grayscale image modalities, generates gating weights using global semantic prior probability vectors, and extracts semantically guided visual structural features through feature modulation.

[0049] The PVBF-QHN network receives the global semantic prior probability vector after dimensional transformation, and extracts high-dimensional quantum state features through a parallel dual quantum neural network block structure, a hybrid entanglement mechanism, and bit-level adaptive weighted fusion.

[0050] After the sequence features, visual structure features, and quantum state features are modulated to a unified dimension through a linear fully connected layer and activation function, they are input into the CoF-TE module. By calculating the electrostatic potential and potential gradient under the triangular charge array, high-order cross-modal fusion features are obtained.

[0051] The fused features are input into the classification layer, and the output is the early rumor detection result;

[0052] Step S3: Use the trained PGQH-TE model to perform real-time early rumor detection on newly published events on social media.

[0053] The specific implementation of step S1 in constructing the dataset is as follows: A strict definition of early rumor events is adopted, meaning only the original content (text, images, videos, audio) available at the time of event publication and the publisher's publicly available social media attributes are used, excluding any post-publication comments, reposts, propagation graphs, and external knowledge. Early rumor event data from social media is selected to construct a strictly defined bimodal early rumor detection dataset. Rumor event samples are used as the smallest unit, with the first 70% of the total dataset used as the training set, the middle 15% as the validation set, and the last 15% as the test set. Each early rumor event is treated as a single input sample, with both rumor and non-rumor labels.

[0054] Step S2, construct the PGQH-TE model and train it using the dataset:

[0055] The PGQH-TE model contains six core processing modules:

[0056] Module a: Unified event coding and cross-modal generation, including text modality generation and image modality generation:

[0057] In terms of text modality generation, only the original text, images, videos, audio, and the publisher's public social attributes at the time of posting are extracted. URLs, HTML tags, control characters, and redundant whitespace are removed from the original data, while the original tweet text, hashtags (#), and mentions (@) are retained. All images (including emojis), videos, and audio are converted to Base64 encoding. The publisher's social attributes (username, age, bio, authentication status, location, number of followers, etc.) are precisely extracted from the metadata. The cleaned text content, corresponding social attributes, and corresponding multimedia Base64 encodings are then sequentially organized and merged into a single text modality file.

[0058] In image modality generation, the total number of bytes in the text modality file is calculated. Determine the side length of the smallest square image. Calculate the number of padding bytes Equidistant sampling from the original text Bytes are padded to the end; the padded data is then mapped byte by byte in line order. The initial grayscale image matrix; the initial grayscale image is uniformly scaled to [size value] using bilinear interpolation. Pixels, as grayscale image modalities.

[0059] Module b: Global Prior Semantic Extraction Module (BERT). The text modality is input into the pre-trained BERT model. For English text, bert-base-uncased

[28] is used, and for Chinese text, bert-base-chinese

[29] is used to extract a 768-dimensional [CLS] semantic feature vector. This feature vector is input into the logistic regression classification layer, and a 2-dimensional global semantic prior probability vector is output. , , This represents the binary classification probability value. This vector will serve as a global prior, guiding the subsequent feature extraction process.

[0060] Module e: Semantic prior-guided sequence feature extraction module (PG-BiGRU). It samples and segments the text modality of early rumor event samples (using Jieba segmenter for Chinese text and standard space segmentation for English text) to obtain n word vector sequences. ; the global semantic prior probability vector The sequence is copied n times and alternately concatenated with n word vector sequences to form an augmented input sequence. The augmented input sequence is fed into a bidirectional GRU network, and the hidden states of the first and last time steps are extracted and concatenated to form sequence features. Based on sequence features Introduce learnable feature-level self-attention weight vectors Calculate semantically guided sequence features .

[0061] Module d: Semantic prior-guided visual structural feature extraction module (PG-ResNet-18). The grayscale image modality is input into the ResNet-18 network to extract a 512-dimensional initial visual feature vector. ; the global semantic prior probability vector The gated weight vector is generated by sequentially passing a dimension-expanded fully connected layer (2→128, ReLU activation) and a dimension-mapped fully connected layer (128→512, Sigmoid activation). Introducing learnable weight vectors (Randomly generated trainable weights) are used to calculate semantically guided visual structural features through a feature modulation formula. .

[0062] Module c: Parallel Variational Coding and Bit-Level Adaptive Fusion Quantum Hybrid Network (PVBF-QHN): PVBF-QHN employs a 6-qubit design. The 2-dimensional global semantic prior probability vector is first normalized and standardized, then scaled to 12 dimensions through a fully connected layer. After activation by the tanh activation function, it is fed into a fully connected layer and scaled to 6 dimensions, then activated again by the tanh activation function, resulting in a vector P = [p0, p1, p2, p3, p4, p5]. Vector P is then fed into two identical parallel quantum neural network blocks (QNB blocks), outputting two 6-dimensional quantum features. and Each QNB sequentially includes an angle encoding layer, a learnable parameter layer, a locally adjacent entanglement layer, a globally annular entanglement layer, and a measurement layer. The angle encoding layer is initialized to... A Hadamard gate is applied to the qubit to generate a superposition state. The vector P is encoded into the superposition state using an RY rotation gate through angle encoding. Then, a trainable parameter layer (18 trainable parameters in total) consisting of RX, RY, and RZ rotation gates, a locally adjacent entangled layer consisting of fixed-structure CNOT gates, and a globally ring-shaped entangled layer consisting of learnable IsingZZ gates are sequentially applied. The measurement layer performs Pauli-Z measurements to obtain two 6-dimensional expectation vectors. and Introduce learnable attention weight vectors and The 6-dimensional high-dimensional quantum state characteristics were calculated using a bit-level adaptive weighted fusion formula. softmax() is the normalized exponential function.

[0063] Module f: Triangular electrostatic field Coulomb fusion module (CoF-TE) and classification. First, the quantum features extracted from modules c, d, and e are... Visual features Sequence features Modulation to a uniform dimension is achieved through a linear fully connected layer and the tanh activation function. (In this embodiment) ), to obtain the sequence feature vector ={ Visual feature vectors Quantum eigenvectors The three are then sequentially assembled into a charge sequence. Define the three vertices of the triangular electrostatic field, calculate and insert the bilateral average point charge (BMC) of the corresponding two adjacent edges, where the BMC is inserted at the vertex of the text edge-visual edge. , for The i-th element in the middle, for The j-th element is inserted at the vertex of the visual edge-quantum edge. , for Insertion at the i-th element of the quantum edge - text edge vertex Global mean normalization is performed on the charge sequence E; a Gaussian kernel function is used. Calculate the charge sequence E of the first charge. One charge electrostatic potential at , For input parameters, For learnable scale parameters, Represents the i-th and i-th elements on the plane. The geometric distance between point charges It is a point charge The charge quantity (eigenvalue); the central difference is used to calculate the first... Potential gradient at each charge . This represents the directed distance (with direction, sign) between the i-th charge and its adjacent charges; the electrostatic potential sequence and the potential gradient sequence are concatenated to form a fused feature vector. =[ ,..., , ,..., ].

[0064] Finally, the fused feature vector is input into the fully connected classification layer, which outputs binary logits. The logits are then converted into a probability distribution using the Softmax function, and the category corresponding to the maximum probability is taken as the early rumor detection result.

[0065] Step S2, which employs a phased training strategy to train the PGQH-TE model, includes the following: First, the PVBF-QHN and PG-ResNet-18 networks are pre-trained independently until the F1 score on the validation set converges, at which point the parameters are frozen. Then, the BERT, PG-BiGRU, and CoF-TE modules are loaded for joint training. The AdamW optimizer is used, and the joint training learning rate is set. A cosine annealing learning rate scheduling strategy is employed. The weighted cross-entropy loss function is used as the loss function and applied to the output of the final classification layer. The classification layer outputs binary logits, which are converted into a probability distribution using the Softmax function. The category corresponding to the maximum probability is taken as the early rumor detection result.

[0066] A second aspect of this invention provides a social media early rumor detection system based on the PGQH-TE model, comprising:

[0067] The dataset construction and preprocessing module is used to obtain the original content and metadata at the time of social media posting and generate bimodal data of text and grayscale images;

[0068] The model building module is used to construct the PGQH-TE model, which includes: a global prior semantic extraction submodule, a semantically guided sequence feature extraction submodule, a semantically guided visual feature extraction submodule, a parallel variational coding and bit-level adaptive fusion quantum hybrid network submodule, and a triangular electrostatic field Coulomb fusion submodule.

[0069] The model training module is used to perform phased pre-training and joint fine-tuning of the PGQH-TE model using the dataset.

[0070] The early rumor detection module is used to call the trained PGQH-TE model and output a classification result to determine whether an early rumor event on the target social media is a rumor.

[0071] The physical interpretability visualization module is used to extract the electrostatic potential distribution and potential gradient curve from the triangular electrostatic field Coulomb fusion submodule, generate the energy distribution difference map of rumor samples and non-rumor samples in the triangular charge field, and provide a physical interpretability basis for model detection decisions.

[0072] The essence of rumor semantics stems from the complex interaction between weak signals such as vague references, emotional incitement, and unreliable user information. These weak signals cannot independently carry the complete semantics of rumors. Only through purposeful combination and global interaction can discriminative rumor patterns be revealed.

[0073] Existing multimodal fusion methods, such as linear concatenation, weighted summation, gated fusion, and semantic alignment, can only achieve shallow feature combinations and are unable to effectively model the high-order collaborative interactions between a large number of weak signals in the semantics of rumors. Furthermore, these methods struggle to support the interpretability requirements of rumor detection criteria. Under the core constraints of early rumor detection—characterized by "highly sparse information, dominant weak signals, and strong reliance on collaboration"—traditional fusion methods cannot effectively model global feature interactions or provide clear decision-making logic, severely limiting detection accuracy and result reliability. To address this, this invention proposes the Coulomb Fusion of Triangular Electrostatic Field (CoF-TE) method as the core multimodal feature fusion module of the PGQH-TE model, specifically targeting the aforementioned core scientific problems of "difficulty in capturing high-order cross-modal interactions and lack of interpretability in classification criteria."

[0074] CoF-TE is a trimodal feature fusion framework built upon the laws of physical fields. Its core idea is to map semantically prior-guided sequence features, visual structural features, and quantum state features as point charges in a two-dimensional triangular charge array. It globally superimposes and models arbitrary-order cross-modal interactions of all features using Gaussian kernel-weighted electrostatic potentials, capturing potential gradients to quantify the global collaborative interaction relationships between features. Ultimately, it generates fused features with both high discriminative power and physical interpretability, providing a decision-making basis for early rumor classification. The essential difference from traditional fusion methods lies in the fact that CoF-TE does not directly perform algebraic operations on feature vectors. Instead, it naturally fuses high-order cross-modal interaction information through the global energy distribution of the electrostatic field, while fully preserving modal heterogeneity and the traceability of the interaction process. Its core design is based on three key dimensions. First, the triangle is the naturally optimal geometric carrier for trimodal fusion. Its asymmetric vertex distances naturally adapt to the semantic space heterogeneity of the three modalities: sequence, vision, and quantum state. It can model the cyclic bidirectional interaction between modalities, overcoming the shortcomings of linear layouts which limit unidirectional information flow and annular layouts where symmetrical distances cannot reflect modal differences. Second, although the methodological idea originates from Coulomb's electrostatic field theory, it uses a learnable Gaussian kernel. Alternating divergent Coulomb potential This ensures both the numerical stability of the backpropagation process and the ability to utilize the learnable bandwidth parameter. By dynamically adjusting the receptive field of interaction, adaptive focusing on semantically related feature interactions can be achieved. Finally, by inserting the bilateral average point charge (BMC) of the corresponding neighboring side at the vertex of the triangle, the local cooperation between adjacent modes can be strengthened, and the information discontinuity problem of pure discrete charge sequences at the mode boundary can be made up.

[0075] CoF-TE uses sequence features extracted guided by semantic priors. Visual structural features and quantum state characteristics As input, in the specific implementation, it first goes through a linear fully connected layer and The activation function modulates the three types of features to a unified dimension K (in this embodiment of the invention, K is taken as K). The normalized sequence characteristic charge vector is obtained. Visual feature charge vector and quantum state characteristic charge vector The three components are then assembled in sequence to form an initial charge sequence of length 3K. Then, at the three vertices of the triangle, the bilateral average point charge of the corresponding two adjacent sides is calculated and inserted. The calculation formula is as follows: , , After insertion, the total charge sequence length is updated to Next, a global mean zeroing process is performed on the complete charge sequence to eliminate the interference of global potential shift on the classification results. The formula is as follows: ,in Let be the charge of the i-th element in the charge sequence. Then, the charge sequence is uniformly distributed across the three sides of a triangle, with the fundamental charge located inside the sides and BMC at the vertex. Calculate the electrostatic potential at each point charge location. The electrostatic potential is obtained by weighted superposition of all other charges through a Gaussian kernel, as shown in the formula: ,in Let J be the charge of the j-th point charge. Let be the geometric distance between the i-th and j-th point charges on the triangular plane. The learnable Gaussian kernel scale parameters are automatically optimized during training, and the electrostatic potential is... Integrating global energy information from all features at that location effectively enhances weak but crucial cross-modal cooperative signals. Based on this, the central difference method is used to calculate the potential gradient at each charge location. The formula for the intensity and direction of local interactions between quantitative features is as follows: ,in Let be the directed distance between the i-th charge and its immediate and adjacent charges. Simultaneously, the first and last elements are processed using boundary closure rules, i.e., when... hour ,when hour The physical meaning of the gradient is that a positive gradient indicates that energy increases with position, corresponding to a synergistic enhancement effect between features, while a negative gradient indicates that energy decreases with position, corresponding to a conflict cancellation effect between features. Finally, the electrostatic potential sequence and the potential gradient sequence at all positions are concatenated to form the final high-order cross-modal fusion feature. The fused features are then fed into a fully connected classification layer, which outputs a binary classification result for early rumor detection.

[0076] CoF-TE possesses three core characteristics and significant technical advantages. First, it offers lossless capture of high-order cross-modal interactions. By globally superimposing electrostatic potentials, it naturally models arbitrary-order interactions between all features, eliminating the need for explicitly designing complex interaction layers. Experimental results show that compared to traditional fusion methods such as direct concatenation, weighted summation, and gated fusion, CoF-TE improves the average accuracy by 1.14% and the F1 score by 1.42% on three public benchmark datasets. Second, it offers physical interpretability. The fusion process can be intuitively explained through visualizing the electrostatic potential distribution and gradient curves. Rumor samples exhibit a horizontal single-core potential distribution with complementary and smooth gradients, indicating high cross-modal semantic synergy. Non-rumor samples exhibit a vertical multi-core potential distribution with significant gradient conflicts, indicating relatively independent semantics between modalities. This provides a clear and traceable physical basis for model decisions. Finally, it naturally enhances weak signals. The exponential decay of the Gaussian kernel ensures that weakly correlated but crucial feature interactions are not drowned out by strong noise, making it particularly suitable for information-sparse scenarios in strict early rumor detection where "weak signals dominate and strong signals are scarce."

[0077] Example 1: Dataset Construction and Preprocessing

[0078] This embodiment uses three publicly available datasets: the Twitter16 dataset (English, containing 207 rumor events and 611 non-rumor events), the Weibo dataset (Chinese, containing 2,351 rumor events and 2,313 non-rumor events), and the PHEME dataset (primarily English, containing 1,972 rumor events and 3,830 non-rumor events), with both rumor and non-rumor sample labels. Each early rumor event is treated as a whole input sample. The original tweets, comments, images, videos, audio, and user social attributes in the event are uniformly encoded to form a single text file. The text file encoding process includes: removing URLs, HTML tags, control characters, and redundant whitespace; converting images (emojis), videos, and audio to Base64 encoding; extracting user social attributes (username, timestamp, hashtags, mentions, emojis, retweets, likes, verification status, and number of followers); and organizing all content in chronological order of tweet posting.

[0079] Convert a text file to a grayscale image: Calculate the total number of bytes in the text. Determine the side length of the smallest square image. Calculate the number of padding bytes Equidistant sampling from text After padding the original text with bytes, it is mapped byte by byte in line-major order. Grayscale images were uniformly scaled using bilinear interpolation. Pixel.

[0080] The dataset is divided into training, validation, and test sets in a ratio of 70%-15%-15% to ensure a balanced distribution of samples across all categories.

[0081] Example 2: Model Training and Evaluation

[0082] PG-BiGRU employs 100-dimensional word embeddings, a two-layer bidirectional GRU, 128 hidden units, and 4 feature-level self-attention heads. BERT uses pre-trained weights bert-base-uncased (English) or bert-base-chinese (Chinese), followed by a logistic regression classifier with a maximum of 1000 iterations. ResNet-18 uses ImageNet pre-trained weights, takes a 128×128 grayscale image as input, and outputs 512-dimensional visual features. The PVBF-QHN quantum network uses 6 qubits, a parallel dual-quantum-block structure, each block containing 18 trainable rotation parameters and learnable entanglement parameters, outputting a 6-dimensional quantum feature vector. The CoF-TE fusion module uses a Gaussian kernel function with a kernel mapping dimension of 256.

[0083] The training process consists of two phases: pre-training and joint training. PVBF-QHN pre-training uses a batch size of 16, a learning rate of 0.005, a StepLR scheduler (decreasing by 0.7 every 15 epochs), weight decay of 1e-4, and gradient clipping with a maximum norm of 1.0, for 50 epochs. ResNet-18 pre-training uses a batch size of 32, a learning rate of 0.001, for 50 epochs, and an early stopping patience value of 7. In the joint training phase, the parameters of PVBF-QHN and ResNet-18 are frozen, using a batch size of 32, the AdamW optimizer, a learning rate of 2e-5, and cosine annealing learning rate scheduling, for 20 epochs. Experiments were conducted on a server equipped with an AMD EPYC 7542 32-core CPU, an NVIDIA GeForce RTX 4090 24GB GPU, and 503GB of RAM, based on PyTorch 2.7.1 and CUDA 12.6. Quantum simulations were performed using the PennyLane framework.

[0084] Model evaluation uses accuracy ( ), accuracy ( ), recall rate ), F1 score ( ) are used as indicators. TP, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives, respectively. Accuracy is calculated using micro-average, while all other indicators are calculated using macro-average to balance the evaluation weights across different categories.

[0085] ;

[0086] ;

[0087] ;

[0088] ;

[0089] Example 3: Model intrinsic validity verification experiment:

[0090] Ablation experiments: To evaluate the contribution of each module, the following experimental configurations were designed: BERT, BiGRU, ResNet-18, PVBF-QHN, PG-BiGRU, PG-ResNet-18, PG-BG-RN (direct concatenation of two modalities), PG-BG-RN-Q (direct concatenation of three modalities), and PG-BG-RN-QE (the complete model, i.e., PGQH-TE). Experimental results show that the complete model achieves the best performance on all three datasets, with a maximum accuracy of 98.39%. The semantic prior guidance mechanism (PG) improves the average accuracy of BiGRU and ResNet-18 by 7.55% and 15.45%, respectively. PVBF-QHN's single-modal performance significantly outperforms BERT and BiGRU. Compared to direct concatenation, CoF-TE fusion improves the accuracy by 1.62%, 0.57%, and 2.39% on the three datasets, respectively.

[0091] Table 1. Ablation study results of the model on three datasets.

[0092] Optimal parameter experiment: Figure 2The impact of different grayscale image sizes (16×16 to 1024×1024) on model performance is demonstrated. Results show that a 128×128 image size achieves the best balance between information preservation and computational efficiency. Furthermore, as shown in Table 2, in experiments evaluating the impact of different qubit numbers (4, 8, 16, 20) on the LIL-QHN branch performance, the 6-qubit configuration achieved 93.55% accuracy, a negligible difference (less than 0.5%) compared to higher qubit numbers (93.65%), and an inference latency (Lat.) of only 4 ms / sample, far lower than the 28 ms / sample of the 20-qubit configuration. This negligible feature extraction latency signifies that this invention fully unleashes the feature mining potential of lightweight quantum neural networks. Therefore, this invention employs a 128×128 image size and a 6-qubit configuration.

[0093] Table 2. Effect of the number of qubits on the performance of PVBF-QHN on the Twitter16 dataset.

[0094] Feature visualization experiment: Figure 3 The results of dimensionality reduction visualization (bottom right) of text features (top left), visual features (top right), quantum features (bottom left), and CoF-TE fused features from the Weibo dataset using t-SNE are presented. Experiments show that quantum features exhibit good clustering performance, while the features fused by CoF-TE form a clear crescent-shaped separation boundary between rumor and non-rumor samples, verifying the significant advantages of the fusion method of this invention in capturing high-order cross-modal interactions and improving feature discriminativeness.

[0095] Physical interpretability verification: Figure 4 This visualization shows the gradient curves of the average potential distribution of rumors (top left) and non-rumor samples (bottom right) on the Weibo dataset. Rumor samples in the triangular electrostatic field exhibit a horizontal, single-core potential distribution with complementary and smooth gradients, indicating high cross-modal cooperation. Non-rumor samples, on the other hand, show a vertical potential distribution with significant gradient conflicts, indicating weak cross-modal cooperation. This visualization provides a physically interpretable basis for model decision-making.

[0096] Example 4: Experiment to evaluate the practicality of the model

[0097] Performance Comparison Experiment: The PGQH-TE model of this invention was compared with 16 existing state-of-the-art models in the field of early-stage social media rumor detection. Referring to Table 3, the experimental results show that, in the strictly early-stage scenario (S, using only the original post), PGQH-TE's average F1 score is 9.01% higher than the best method of the same period, MMFSL (2026), 6.40% higher than MTAS (2026) with propagation period (T), and 4.16% higher than D² (2025) generated using data compensation (D). PGQH-TE achieves superior detection performance with less information input, verifying the advancement of the method of this invention in the strictly early-stage rumor detection task.

[0098] Table 3. Performance comparison of the PGQH-TE model with state-of-the-art methods on three datasets.

[0099] Impact of different multimodal fusion methods on model performance: To verify the superiority of the CoF-TE fusion method, it was replaced with direct concatenation (DC), weighted summation (WS), gated fusion (GF), and semantic guidance (SG). Referring to Table 4, experimental results show that CoF-TE achieves the best performance on all three datasets, with an average accuracy improvement of 1.14% and an average F1 score improvement of 1.42% compared to the second-best methods, verifying the effectiveness of electrostatic field fusion in capturing high-order cross-modal interactions.

[0100] Table 4. Impact of different feature fusion strategies on the PGQH-TE model

[0101] Example 5:

[0102] Cross-lingual generalization experiment: To verify the generalization ability of the model in cross-lingual scenarios, the three datasets were mixed in a 70%-15%-15% ratio to construct a multilingual training set, validation set and test set, and the multilingual BERT model bert-base-multilingual-cased

[27] was adopted. The experimental results in Table 5 show that the model achieved an accuracy of 90.83% on the mixed test set, proving that it can learn language-independent rumor discrimination patterns and has good cross-lingual generalization ability.

[0103] Table 5. Model performance on cross-linguistic datasets

[0104] Example 6: Boundaries and Limitations of Quantum Gain

[0105] Quantum Component vs. Classical Component: To evaluate the gains of the quantum component, we replaced two parallel QNB modules in PVBF-QHN with two multilayer perceptrons (MLPs) while keeping the rest of the architecture unchanged. As shown in Table 6, the quantum variant achieved a significant improvement in accuracy (up to 5.20%) and reduced the number of parameters, although the inference latency increased slightly (4 ms per sample), but remained within acceptable limits. This demonstrates that our quantum hybrid network can capture complex patterns that are difficult to capture with classical modules with fewer model parameters, while maintaining high inference efficiency on classical hardware.

[0106] Table 6. Effects of QNB and MLP on the performance of PVBF-QHN

[0107] Practical application scenarios:

[0108] The PGQH-TE model of this invention, after being trained on early rumor events on a specific social media platform, can perform real-time early rumor detection on newly published social media events on the platform and determine whether they are rumors. In practical deployment, it can be applied to the following scenarios:

[0109] Network content security management: Deploy the PGQH-TE model in the content moderation system of social media platforms to detect rumors in real time on user-posted content, identify and block suspected rumors at the source of information dissemination, assist moderators in quick handling, and minimize the social harm of rumor spread.

[0110] Public opinion monitoring and early warning: Integrate the PGQH-TE model into the public opinion monitoring system to perform batch analysis of the initial content released on social media related to hot events, providing early public opinion warning support for government agencies and enterprises.

[0111] Digital Evidence Collection and Judicial Applications: In cybercrime investigations and judicial evidence collection, the PGQH-TE model is used to identify rumors in initial information related to a case on social media platforms. The physical interpretability evidence (potential distribution and gradient) it provides can offer traceable technical support for the construction of evidence chains and case analysis.

[0112] Technical effectiveness verification:

[0113] The technical effects of this invention were verified through systematic experiments:

[0114] Accuracy verification: The highest accuracy rate reached 98.39% on three public datasets, which significantly outperforms existing state-of-the-art models that rely on propagated information or external data in strict early-stage scenarios, verifying the superior detection performance of the method of this invention.

[0115] Robustness verification: Maintaining an accuracy of over 90% on a cross-lingual mixed test set demonstrates that the model has strong cross-lingual generalization ability.

[0116] Interpretability verification: By visualizing the electric potential field and gradient, the differential energy distribution of rumor and non-rumor samples in the electrostatic field is intuitively displayed, providing a physically traceable explanation for model decision-making.

[0117] Efficiency verification: PVBF-QHN has only 0.96K parameters and an inference latency of about 4.17ms / sample, which meets the lightweight deployment requirements for real-time detection.

[0118] Compared with the prior art, the present invention has the following beneficial effects:

[0119] Breakthrough detection performance: The highest accuracy of 98.39% was achieved on three public datasets: Twitter16

[24] , Weibo

[25] and PHEME

[26] , significantly outperforming existing state-of-the-art models that rely on more information in strict early-stage scenarios.

[0120] Innovative architecture design: A semantic prior guidance mechanism (PG) is proposed to enable the feature extractor to focus on high discriminative information. In the existing technology, quantum hybrid networks are limited by the physical bottleneck of the small number of qubits and high noise of current quantum hardware. Existing quantum hybrid models generally adopt oversimplified circuit design [8], which leads to the inability to fully release the potential of quantum superposition and entanglement [9]. This invention proposes a PVBF-QHN quantum network, which fully releases the potential of quantum feature mining through parallel dual-block, hybrid entanglement and qubit-level adaptive fusion under a lightweight architecture. A CoF-TE triangular electrostatic field fusion method is proposed to capture high-order cross-modal interactions through electrostatic potential and gradient, and to give physical interpretability.

[0121] Strong generalization and robustness: By preserving semantic integrity through strict event definition and combining multimodal complementary information with physical fusion mechanism, high accuracy is maintained in cross-language scenarios.

[0122] Physical interpretability: Through visualization of electrostatic potential distribution and gradient curves, a physical model-based decision explanation is provided for multimodal rumor detection for the first time, enhancing the credibility of the model in practical applications.

[0123] Practical Engineering Value: The model design balances performance and efficiency. PVBF-QHN can be efficiently simulated on classic hardware to meet real-time detection requirements. It can be deployed in practical application scenarios such as network content security management, public opinion monitoring and early warning, and digital forensics.

[0124] The PGQH-TE model proposed in this invention effectively solves key technical challenges in early-stage rumor detection on social media, such as information sparsity, difficulty in capturing high-order interactions, and lack of interpretability. This is achieved through rigorous early event definition, semantic prior-guided feature enhancement, lightweight quantum feature mining, and a physically inspired interpretable fusion mechanism. Extensive experiments demonstrate that this method significantly outperforms existing technologies in terms of accuracy, generalization, interpretability, and practicality, providing a reliable solution for the engineering application of early-stage rumor detection technology on social media.

[0125] According to embodiments disclosed herein, the present invention also provides an electronic device, a readable storage medium, and a computer program product. The electronic device is intended to represent various forms of digital computers, including a computing unit that can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) or a computer program loaded from a storage unit into a random access memory (RAM). Various programs and data required for device operation may also be stored in the RAM. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus. Multiple components in the device are connected to the I / O interfaces, including: input units, such as a keyboard, mouse, etc.; output units, such as various types of displays, speakers, etc.; storage units, such as disks, optical disks, etc.; and communication units, such as network interface cards, modems, wireless transceivers, etc. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks. The computing unit can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Examples of computing units include, but are not limited to, central processing units (CPUs), graphics processing units (GPUs), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processors, controllers, microcontrollers, etc. The computing unit performs the various methods and processes described above, such as the social media early rumor detection method based on the PGQH-TE model. For example, in some embodiments, the social media early rumor detection method based on the PGQH-TE model can be implemented as a computer software program tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the computing unit, one or more steps of the social media early rumor detection method based on the PGQH-TE model described above can be performed. Alternatively, in other embodiments, the computing unit can be configured to perform the social media early rumor detection method based on the PGQH-TE model by any other suitable means (e.g., by means of firmware).

[0126] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0127] Program code for implementing the methods disclosed in this invention may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a standalone software package, or entirely on a remote machine or server.

[0128] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0129] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0130] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0131] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0132] The above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the invention. Those skilled in the art will understand that implementing all or part of the above-described embodiments and making equivalent changes in accordance with the claims of the present invention are still within the scope of the invention.

[0133] [1] Gao, Y., Liang, G., Jiang, F., Xu, C., Yang, J., Chen, J., &Wang, H. (2020). Social network rumor detection: A survey. Acta ElectronicaSinica, 48(7), 1421–1435. https: / / doi.org / 10.3969 / j.issn.0372-2112.2020.07.023;

[0134] [2] Zannettou, S., Sirivianos, M., Blackburn, J., & Kourtellis, N.(2019). The web of false information: Rumors, fake news, hoaxes, clickbait,and various other shenanigans. *Journal of Data and Information Quality*, *11*(3), 1–37. https: / / doi.org / 10.1145 / 3309699;

[0135] [3] Lu, K., Zhang, H., Sun, T., Yang, Y., Meng, C., Yin, G., & Fang,B. (2025). GMCL: Graph-enhanced multimodal contrastive learning for rumordetection. *2025 IEEE International Conference on Acoustics, Speech andSignal Processing (ICASSP)*. IEEE. https: / / doi.org / 10.1109 / ICASSP49660.2025.10890376;

[0136] [4] Zhou, H., Ma, T., Rong, H., Qian, Y., Tian, Y., & Al-Nabhan, N.(2022). MDMN: Multi-task and domain adaptation based multi-modal network forearly rumor detection. Expert Systems with Applications, 195, 116517.https: / / doi.org / 10.1016 / j.eswa.2022.116517;

[0137] [5] Huang, X., Ma, T., Rong, H., Wang, G., Liao, G., & Liu, D.(2025). Multimodal rumor detection with external knowledge and evidence basedscene graph attention network. *Chinese Journal of Computers*, *48*(9), 2159–2180. https: / / doi.org / 10.11897 / SP.J.1016.2025.02159;

[0138] [6] Le-Pham, H.-T., Nguyen, T.-P., Tran, A.-D., & Le, K.-H. (2026).MTAS: A temporal-aware multiview graph attention framework for early rumordetection on social media. *Information Fusion*, *126*, 103598. https: / / doi.org / 10.1016 / j.inffus.2025.103598;

[0139] [7] Xu, H., Gao, C., Li, X., & Wang, Z. (2025). *D²: Customizing two-stage graph neural networks for early rumor detection through cascadediffusion prediction*. Proceedings of the 18th ACM International Conferenceon Web Search and Data Mining (WSDM), 568–576. ACM. https: / / doi.org / 10.1145 / 3701551.3703589;

[0140] [8] Jia, K., Meng, F., Chen, Z., Du, M., & Liang, J. (2025). LQMF-RD:A lightweight quantum-driven multi-modal fusion framework for rumordetection. Knowledge-Based Systems, 330, 116517. https: / / doi.org / 10.1016 / j.knosys.2025.114633;

[0141] [9] Bosco, D. L., Portelli, B., & Serra, G. (2026). Integratedencoding and quantization to enhance quanvolutional neural networks. IEEETransactions on Quantum Engineering, 7, 3100719. https: / / doi.org / 10.1109 / TQE.2025.3646040;

[0142]

[10] Chen, Y., Maulana, F., & others. (2026). Impact of circuit depthversus qubit count on variational quantum classifiers for Higgs boson signaldetection. arXiv preprint arXiv:2601.11937. https: / / doi.org / 10.48550 / arXiv.2601.11937;

[0143]

[11] Yin, P., Pan, W., Peng, C., & Zhang, H. (2020). Research onearly detection of Weibo rumors based on user characteristic analysis. *Journal of Intelligence*, *39*(7), 81–86. https: / / doi.org / 10.3969 / j.issn.1002-1965.2020.07.014;

[0144]

[12] Zhou, H., Ma, T., Rong, H., Qian, Y., Tian, Y., & Al-Nabhan, N.(2022). MDMN: Multi-task and domain adaptation based multi-modal network forearly rumor detection. *Expert Systems with Applications*, *195*, 116517.https: / / doi.org / 10.1016 / j.eswa.2022.116517;

[0145] Hu, Y., Ju, X., Ye, Z., Khan, S., Yuan, C., Lai, Q., & Liu, J.(2022). Early rumor detection based on data augmentation and pre-trainingtransformer. 2022 IEEE 12th Annual Computing and Communication Workshop andConference (CCWC), 152–158. IEEE. https: / / doi.org / 10.1109 / CCWC54503.2022.9720776;

[0146]

[13] Dong, F., Zhu, Y., Wan, S., & Xu, Y. (2022). Improving the earlyrumor detection performance of the deep learning models by CGAN. Proceedingsof the International Conference on Software Engineering and KnowledgeEngineering (SEKE), 580–585. Knowledge Systems Institute Graduate School.https: / / doi.org / 10.18293 / SEKE2022-030;

[0147]

[14] Zeng, F., & Gao, W. (2022). Early rumor detection using neuralhawkes process with a new benchmark dataset. Proceedings of the 2022Conference of the North American Chapter of the Association for ComputationalLinguistics: Human Language Technologies, 4105–4117. https: / / doi.org / 10.18653 / v1 / 2022.naacl-main.302;

[0148]

[15] Thota, N. R., Sun, X., & Dai, J. (2023). Early rumor detectionin social media based on graph convolutional networks. 2023 InternationalConference on Computing, Networking and Communications (ICNC), 516–522. IEEE.https: / / doi.org / 10.1109 / ICNC57223.2023.10074242;

[0149]

[16] Ouyang, Q., Chen, H., Liu, S., Wang, K., & Li, X. (2024). Earlyrumor detection based on Bert-GNNs heterogeneous graph attention network.Acta Electronica Sinica, 52(1), 311–323. https: / / doi.org / 10.12263 / DZXB.20220882;

[0150]

[17] Zhang, S., Wei, J., Zhao, Z., Li, B., & Wong, K.-F. (2024).TGAN: Temporal-aware graph attention network for early rumor detection insocial media. In Natural Language Processing and Chinese Computing (NLPCC2024) (Vol. 15362, pp. 454–468). Springer. https: / / doi.org / 10.1007 / 978-981-97-9440-9_35;

[0151]

[18] Zhu, Y., Wang, G., Jin, W., Huang, X., & Li, S. (2024). Networkrumor detection based on enhanced textual semantics and weighted commentstance. Journal of Frontiers of Computer Science and Technology, 18(12),3311–3323. https: / / doi.org / 10.3778 / j.issn.1673-9418.2402056;

[0152]

[19] Tao, X., Wang, L., Liu, Q., Wu, S., & Wang, L. (2024). Semanticevolvement enhanced graph autoencoder for rumor detection. In Proceedings ofthe ACM Web Conference 2024 (WWW) (pp. 4150–4159). Association for ComputingMachinery. https: / / doi.org / 10.1145 / 3589334.3645478;

[0153]

[20] Yang, Z., Pang, Y., Li, Q., Wei, S., Wang, R., & Xiao, Y.(2024). A model for early rumor detection based on topic-derived domaincompensation and multi-user association. Expert Systems with Applications,250, 123951. https: / / doi.org / 10.1016 / j.eswa.2024.123951;

[0154]

[21] Yang, C., Yu, X., Wu, J., Xue, D., & Duan, Y. (2024). RumorSAGE:Semantic augmentation graph for early rumor detection. *Proceedings of the2024 3rd International Conference on Cyber Security, Artificial Intelligenceand Digital Economy*, 406–412. ACM. https: / / doi.org / 10.1145 / 3672919.3672994;

[0155]

[22] Wu, Y., Sun, J., Yuan, X., Huang, Z., & Dai, J. (2024). Dual-channel early rumor detection based on factual evidence. Expert Systems withApplications, 238, 121928. https: / / doi.org / 10.1016 / j.eswa.2023.121928;

[0156]

[23] Ye, Y., Chen, H., Cai, Q., & Su, H. (2026). Multimodal meta-learning for early rumor detection based on few-shot learning. IEEETransactions on Computational Social Systems, 13(1), 475–484. https: / / doi.org / 10.1109 / TCSS.2025.3592660;

[0157]

[24] Ma, J., Gao, W., & Wong, K.-F. (2017). Detect rumors inmicroblog posts using propagation structure via kernel learning. InProceedings of the 55th Annual Meeting of the Association for ComputationalLinguistics (Long Papers) (pp. 708–717). Association for ComputationalLinguistics.

[0158]

[25] Ma, J., Gao, W., Mitra, P., Kwon, S., Jansen, B. J., Wong, K.-F., & Cha, M. (2016). Detecting rumors from microblogs with recurrent neuralnetworks. In Proceedings of the Twenty-Fifth International Joint Conferenceon Artificial Intelligence (IJCAI) (pp. 3818–3824). International JointConferences on Artificial Intelligence.

[0159]

[26] Zubiaga, A., Liakata, M., & Procter, R. (2017). Exploitingcontext for rumour detection in social media. In Proceedings of the 9thInternational Conference on Social Informatics (SocInfo) (Vol. 10539, pp.109–123). Springer.

[0160]

[27] Google AI. (2018). bert-base-multilingual-cased: Pre-trainedmultilingual BERT model (cased) [Computer software]. ModelScope Model Hub.Retrieved March 20, 2026,

[0161]

[28] Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT:Pre-training of deep bidirectional transformers for language understanding.In Proceedings of the Conference of the North American Chapter of theAssociation for Computational Linguistics: Human Language Technologies(NAACL-HLT) (pp. 4171–4186). Association for Computational Linguistics.

[0162]

[29] Google AI. (2018). bert-base-chinese: Pre-trained Chinese BERTmodel (simplified and traditional). ModelScope.

Claims

1. A method for early rumor detection on social media based on the PGQH-TE model, characterized in that, include: Step S1, construct an early rumor event dataset: select early social media rumor event data containing only the original published content and the publisher's public information, construct a bimodal early rumor detection dataset, and divide the dataset into training set, validation set, and test set; Step S2: Construct the social media early rumor detection model PGQH-TE based on semantic prior guidance, triangular electrostatic field fusion, and quantum hybrid neural network, including: a preprocessing module, a global prior semantic extraction module BERT, a semantic prior guided sequence feature extraction network PG-BiGRU, a semantic prior guided visual structural feature extraction network PG-ResNet-18, a parallel variational coding and bit-level adaptive fusion quantum hybrid network PVBF-QHN, and a triangular electrostatic field Coulomb fusion module CoF-TE; train the PGQH-TE model using the aforementioned dual-modal early rumor detection dataset. The preprocessing module encodes each early rumor event sample as a whole into a text modality, and generates a grayscale image modality of a specified size by filling text bytes and mapping rows and columns. The global prior semantic extraction module performs feature extraction and classification on text modalities, and outputs a global semantic prior probability vector with high discriminative power. The PG-BiGRU network concatenates the word vector sequence of the text modality with the global semantic prior probability vector, and extracts semantically guided sequence features through information transmission mechanism and feature-level self-attention. The PG-ResNet-18 network receives grayscale image modalities, generates gating weights using the global semantic prior probability vector, and extracts semantically guided visual structural features through feature modulation. The PVBF-QHN network receives the global semantic prior probability vector after dimensional transformation, and extracts high-dimensional quantum state features through a parallel dual quantum neural network block structure, a hybrid entanglement mechanism, and bit-level adaptive weighted fusion. After the sequence features, visual structure features, and quantum state features are modulated to a unified dimension through a linear fully connected layer and an activation function, they are input into the CoF-TE module. By calculating the electrostatic potential and potential gradient under the triangular charge array, high-order cross-modal fusion features are obtained. The fused features are input into the classification layer, and the early rumor detection results are output. Step S3: Use the trained PGQH-TE model to perform real-time early rumor detection on newly published events on social media.

2. The method for early social media rumor detection based on the PGQH-TE model according to claim 1, characterized in that, The grayscale image modality generation in step S2 specifically includes: calculating the total number of bytes in the text modality file. Determine the side length of the smallest square image. Calculate the number of padding bytes Equidistant sampling from the original text Bytes are padded to the end; the padded data is then mapped byte by byte in line order. The initial grayscale image matrix; the initial grayscale image is uniformly scaled to a certain value using bilinear interpolation. Pixels, as grayscale image modalities.

3. The method for early social media rumor detection based on the PGQH-TE model according to claim 2, characterized in that, The specific processing flow of the global prior semantic extraction module in step S2 includes: inputting the text modality into the pre-trained BERT model to extract a 768-dimensional semantic feature vector; inputting the semantic feature vector into the logistic regression classification layer to output a 2-dimensional global semantic prior probability vector. .

4. The method for early social media rumor detection based on the PGQH-TE model according to claim 3, characterized in that, The PG-BiGRU network feature extraction in step S2 specifically includes: sampling and segmenting the text modality of early rumor event samples to obtain n word vector sequences. ; the global semantic prior probability vector copy This is repeated several times, and the sequence is alternately concatenated with n word vector sequences to form an augmented input sequence. The augmented input sequence is fed into a bidirectional GRU network, and the hidden states of the first and last time steps are extracted and concatenated to form sequence features. Based on sequence features Introduce learnable feature-level self-attention weight vectors Calculate semantically guided sequence features .

5. The method for early social media rumor detection based on the PGQH-TE model according to claim 4, characterized in that, The PG-ResNet-18 network feature extraction in step S2 specifically includes: inputting the grayscale image modality into the ResNet-18 network to extract an initial 512-dimensional visual feature vector. ; the global semantic prior probability vector Gated weight vectors are generated by sequentially passing through dimension-expanded fully connected layers and dimension-mapped fully connected layers. Introducing learnable weight vectors The semantically guided visual structural features are calculated using the feature modulation formula. .

6. The method for early social media rumor detection based on the PGQH-TE model according to claim 5, characterized in that, The PVBF-QHN network in step S2 adopts a 6-qubit structure, and the high-dimensional quantum state feature extraction specifically includes: The 2D global semantic prior probability vector is first normalized and standardized, then scaled to 12D through a fully connected layer. After activation by the tanh activation function, it is fed into a fully connected layer, scaled to 6D, and then activated again by the tanh activation function to obtain the vector P = [p0, p1, p2, p3, p4, p5]. The vector P is then fed into two parallel quantum neural network blocks (QNB) with identical structures, outputting two 6D quantum features. and Each QNB sequentially includes an angle encoding layer, a learnable parameter layer, a locally adjacent entanglement layer, a globally annular entanglement layer, and a measurement layer. The angle encoding layer is initialized to... A Hadamard gate is applied to the qubit to generate a superposition state. The vector P is encoded into the superposition state using an RY rotation gate through angle encoding. Then, a trainable parameter layer consisting of RX, RY, and RZ rotation gates, a locally adjacent entangled layer consisting of fixed-structure CNOT gates, and a globally ring-shaped entangled layer consisting of learnable IsingZZ gates are sequentially applied. The measurement layer performs Pauli-Z measurements to obtain two 6-dimensional expectation vectors. and ; Introduce learnable attention weight vectors and The 6-dimensional high-dimensional quantum state characteristics were calculated using a bit-level adaptive weighted fusion formula. .

7. The method for early social media rumor detection based on the PGQH-TE model according to claim 6, characterized in that, The CoF-TE processing procedure of the triangular electrostatic field coulomb fusion module in step S2 includes: ... , , Modulation to a uniform dimension The resulting sequence feature vector Visual feature vectors and quantum eigenvectors Further splicing into a charge sequence Define the three vertices of the triangular electrostatic field, calculate and insert the bilateral average charge (BMC) of the corresponding two adjacent edges, where the BMC is inserted at the vertex of the text edge-visual edge. , for The i-th element in the middle, for The j-th element is inserted at the vertex of the visual edge-quantum edge. , for Insertion at the i-th element of the quantum edge - text edge vertex Global mean normalization is performed on the charge sequence E; a Gaussian kernel function is used. Calculate the charge sequence E of the first charge. One charge electrostatic potential at ; Calculate the first Potential gradient at each charge The electrostatic potential sequence and the potential gradient sequence are concatenated to form a fused feature vector. =[ ,..., , ,..., ].

8. The method for early social media rumor detection based on the PGQH-TE model according to claim 7, characterized in that, The training strategy for the PGQH-TE model in step S2 specifically includes: first, independently pre-training the PVBF-QHN network and the PG-ResNet-18 network until the F1 score on the validation set converges, then freezing the parameters; subsequently, loading the BERT module, PG-BiGRU network, and CoF-TE module for joint training; using the AdamW optimizer and setting the joint training learning rate; employing a cosine annealing learning rate scheduling strategy; using weighted cross-entropy loss as the loss function, applied to the output of the final classification layer; the classification layer outputs binary logits, which are converted into a probability distribution by the Softmax function, and the category corresponding to the maximum probability is taken as the early rumor detection result.

9. A social media early rumor detection system based on the PGQH-TE model, characterized in that, For implementing the method according to any one of claims 1-8, comprising: The dataset construction and preprocessing module is used to obtain the original content and metadata at the time of social media posting and generate bimodal data of text and grayscale images; The model building module is used to construct the PGQH-TE model, which includes: a global prior semantic extraction submodule, a semantically guided sequence feature extraction submodule, a semantically guided visual feature extraction submodule, a parallel variational coding and bit-level adaptive fusion quantum hybrid network submodule, and a triangular electrostatic field Coulomb fusion submodule. The model training module is used to perform phased pre-training and joint fine-tuning of the PGQH-TE model using the dataset; The early rumor detection module is used to call the trained PGQH-TE model and output a classification result as to whether an early rumor event on the target social media is a rumor.

10. The social media early rumor detection system based on the PGQH-TE model according to claim 9, characterized in that, Also includes: The physical interpretability visualization module is used to extract the electrostatic potential distribution and potential gradient curve in the triangular electrostatic field Coulomb fusion submodule, generate the energy distribution difference map of rumor samples and non-rumor samples in the triangular charge field, and provide a physical interpretability basis for model detection decision.

Images