A malicious semantic fusion and feature extraction method for a multi-lingual social network

Abstract

This invention relates to the field of network content security technology and discloses a method for malicious semantic fusion and feature extraction for multilingual social networks, comprising the following steps: S1, multilingual corpus construction and preprocessing stage; S2, cross-language semantic encoding stage; S3, cross-layer malicious semantic fusion stage; S4, cluster feature purification stage; S5, model discrimination and iterative optimization stage. This invention overcomes the limitations of monolingual malicious content detection technology, realizes unified representation and extraction of cross-language malicious semantics in multilingual social network scenarios, effectively eliminates the surface and deep heterogeneity between different language data, can adapt to the high-frequency multilingual mixed text and low-resource language content processing needs in social scenarios, and significantly improves the discriminability and representation ability of cross-language malicious semantic features by directionally strengthening malicious semantic-related features through an adaptive feature fusion mechanism.

Description

Technical Field

[0001] This invention relates to the field of network content security technology, and in particular to a method for malicious semantic fusion and feature extraction for multilingual social networks. Background Technology

[0002] With the rapid development of global social networks, multilingual social content has become one of the mainstream forms of online information dissemination, enabling users of different languages ​​to achieve cross-regional and cross-language information interaction through social platforms;

[0003] At the same time, malicious content such as hate speech, cyberbullying, disinformation, and terrorism-related content are also spreading across regions through multilingual social networks, posing a serious challenge to network content security and cyberspace governance.

[0004] Current detection technologies for malicious online content are mostly focused on monolingual scenarios, relying on monolingual corpus resources and semantic feature extraction models, which are difficult to adapt to the global governance needs of multilingual social networks.

[0005] Existing cross-language semantic processing technologies mostly rely on large-scale parallel corpora to complete model training. They are insufficient for processing short texts, multilingual mixed content, and low-resource languages ​​in social scenarios, and cannot achieve unified representation and accurate extraction of cross-language malicious semantics, making it difficult to meet the actual needs of malicious content control in multilingual social network scenarios. Summary of the Invention

[0006] The purpose of this invention is to provide a method for malicious semantic fusion and feature extraction for multilingual social networks, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for malicious semantic fusion and feature extraction for multilingual social networks includes the following steps:

[0009] S1. Multilingual Corpus Construction and Preprocessing Stage: Collect multilingual social network text data, complete language identification, text preprocessing, malicious semantic annotation and dataset partitioning, and construct a standardized multilingual social network malicious corpus;

[0010] S2, Cross-language semantic encoding stage: Based on the Transformer variant architecture, an encoding model is built to complete the cross-language word embedding mapping and multi-layer semantic encoding of the input text, and output the initial semantic features of the multilingual text.

[0011] S3, Cross-layer malicious semantic fusion stage: Receives semantic features from different network layers, completes layer feature dimension alignment, cross-layer attention weight calculation and multi-layer feature weighted fusion, combines malicious semantic prior knowledge to complete feature enhancement, and outputs cross-language malicious semantic fusion features;

[0012] S4. Cluster feature purification stage: Based on the fuzzy C-means clustering algorithm, the initialization of clustering parameters, iterative clustering calculation and feature selection and purification are completed, and language-independent malicious semantic discrimination features are output.

[0013] S5. In the model discrimination and iterative optimization stage, malicious semantic classification is completed based on the purified features, the classification loss is calculated and backpropagation parameter updates are performed to complete the model convergence verification and hyperparameter tuning.

[0014] As a further improvement to this technical solution: in the multilingual corpus construction and preprocessing stage, multilingual social network text data is collected through the corpus acquisition unit, covering multiple languages ​​and all types of social text content; the language identification unit completes the language determination of the input text and completes the language boundary annotation of multilingual mixed text; the text preprocessing unit completes text segmentation, sub-word segmentation, noise content cleaning and malicious semantic annotation; and the dataset partitioning unit completes the proportional partitioning of the training set, validation set and test set of the labeled data to construct a standardized multilingual social network malicious corpus.

[0015] As a further improvement to this technical solution: in the cross-lingual semantic encoding stage, the cross-lingual word embedding unit adopts the word embedding layer of the cross-lingual pre-trained model to complete the unified semantic space mapping of words in different languages; the local attention optimization Transformer encoding unit adjusts the self-attention mechanism and adds a local attention window to complete the multi-layer Transformer encoding of the input text and output the semantic features of different network layers; the initial feature output unit completes the standardized output of the encoded high-dimensional semantic features.

[0016] As a further improvement to this technical solution: in the cross-layer malicious semantic fusion stage, the hierarchical feature adaptation unit receives semantic features output from different network layers of the Transformer encoding unit to complete the dimensional alignment of features at different levels; the cross-layer attention weight calculation unit adaptively calculates the fusion weights of features at different levels, and the corresponding calculation formula is as follows: , in the formula, The fusion weights are the encoded features of the l-th Transformer layer. This represents the total number of layers in the Transformer encoder. The semantic feature vector output by the l-th layer Transformer encoder. This is a trainable attention weight mapping matrix. This is the semantic baseline vector corresponding to prior knowledge of cross-language malicious keywords. The function is a natural exponential function. Based on the calculated fusion weights, the cross-layer attention weight calculation unit performs a weighted summation of semantic features at each level to obtain the initial cross-language malicious semantic fusion features. The malicious semantic feature enhancement unit introduces prior knowledge of cross-language malicious keywords to enhance the malicious dimension of the fusion features. The fusion feature output unit completes the standardized output of the enhanced cross-language malicious semantic fusion features.

[0017] As a further improvement to this technical solution: In the clustering feature purification stage, the clustering parameter initialization unit initializes the cluster centers and membership matrices based on preset malicious semantic categories; the fuzzy C-means clustering iterative calculation unit performs iterative optimization calculations with the objective function as the goal, and the calculation formula corresponding to the objective function is: , in the formula, Let C be the objective function for fuzzy C-means clustering. For fuzzy weighted index, This represents the total number of input feature vectors. The preset total number of cluster centers. For the first The i-th feature vector corresponds to the i-th eigenvector. Membership degree of each cluster center For the first A cross-language malicious semantic fusion feature vector of each input. For the first Feature vectors of cluster centers The Euclidean norm is used; the membership matrix and cluster centers are updated synchronously in each iteration by the fuzzy C-means clustering iterative calculation unit. The iteration process continues until the change in the objective function is lower than the preset convergence threshold, thus completing the soft clustering of the input fusion features; the feature selection and purification unit removes noisy features and outliers based on the membership threshold, and filters redundant feature dimensions; the final feature output unit completes the standardized output of the purified language-independent malicious semantic discrimination features.

[0018] As a further improvement to this technical solution: in the model discrimination and iterative optimization stage, the semantic classification unit receives the purified malicious semantic discrimination features and outputs the malicious semantic category and confidence level of the input text; the loss calculation unit calculates the cross-entropy classification loss based on the labeled real labels and the model prediction results; the backpropagation optimization unit updates the backpropagation of the model's full-link trainable parameters based on the calculated loss value; and the model convergence verification unit performs hyperparameter tuning and determines the model convergence state based on the index data of the verification set.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. This invention breaks through the limitations of single-language malicious content detection technology, realizes the unified representation and extraction of cross-language malicious semantics in multilingual social network scenarios, effectively eliminates the surface and deep heterogeneity between different language data, can adapt to the processing needs of multilingual mixed text and low-resource language content that frequently occur in social scenarios, and significantly improves the distinguishability and representation ability of cross-language malicious semantic features by directionally strengthening malicious semantic-related features through an adaptive feature fusion mechanism.

[0021] 2. This invention has optimized the model architecture to suit the characteristics of short text scenarios on social networks, avoiding the problem of insufficient semantic capture of short texts by native models. At the same time, it has achieved effective feature purification through soft clustering mechanism, reducing the interference of noisy content on feature extraction results. It has strong scenario adaptability and generalization ability, and can be directly deployed in the content security management scenario of global social networks, providing stable technical support for malicious content identification and online public opinion management in multilingual environments.

[0022] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0024] Figure 1 This is a schematic diagram of the method structure for a malicious semantic fusion and feature extraction method for multilingual social networks. Detailed Implementation

[0025] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are for illustrative purposes only and are not intended to limit the scope of the invention. The invention is described more specifically in the following paragraphs by way of example with reference to the accompanying drawings. It should be noted that the drawings are in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.

[0026] Please see Figure 1 In this embodiment of the invention, a method for malicious semantic fusion and feature extraction for multilingual social networks includes the following steps:

[0027] S1. Multilingual Corpus Construction and Preprocessing Stage: Collect multilingual social network text data, complete language identification, text preprocessing, malicious semantic annotation and dataset partitioning, and construct a standardized multilingual social network malicious corpus;

[0028] S2, Cross-language semantic encoding stage: Based on the Transformer variant architecture, an encoding model is built to complete the cross-language word embedding mapping and multi-layer semantic encoding of the input text, and output the initial semantic features of the multilingual text.

[0029] S3, Cross-layer malicious semantic fusion stage: Receives semantic features from different network layers, completes layer feature dimension alignment, cross-layer attention weight calculation and multi-layer feature weighted fusion, combines malicious semantic prior knowledge to complete feature enhancement, and outputs cross-language malicious semantic fusion features;

[0030] S4. Cluster feature purification stage: Based on the fuzzy C-means clustering algorithm, the initialization of clustering parameters, iterative clustering calculation and feature selection and purification are completed, and language-independent malicious semantic discrimination features are output.

[0031] S5. In the model discrimination and iterative optimization stage, malicious semantic classification is completed based on the purified features, the classification loss is calculated and backpropagation parameter update is performed to complete the model convergence verification and hyperparameter tuning.

[0032] Specifically, stage S1 provides a standardized data foundation for the method, completing the collection, cleaning, annotation, and dataset partitioning of multilingual social texts to provide compliant input data for subsequent model training; stage S2 is the core encoding stage, using a variant Transformer architecture to complete semantic space mapping and deep feature extraction of texts in different languages, achieving initial semantic alignment of texts in different languages; stage S3 is the core fusion stage, completing adaptive fusion of semantic features at different levels of Transformer, strengthening the malicious semantic-related dimensions in the features, and outputting cross-language malicious semantic fusion features; stage S4 is the core purification stage, using a fuzzy C-means clustering algorithm to complete soft clustering of cross-language features, removing language-specific redundant features, and outputting language-independent malicious semantic discrimination features; stage S5 is the model optimization stage, completing the classification and discrimination of malicious semantics and iterative optimization of the model's full-link parameters to ensure the method's recognition accuracy and generalization ability.

[0033] In the multilingual corpus construction and preprocessing stage, multilingual social network text data is collected through the corpus acquisition unit, covering multiple languages ​​and all types of social text content; the language identification unit determines the language of the input text and completes language boundary annotation for multilingual mixed text; the text preprocessing unit completes text segmentation, sub-word segmentation, noise content cleaning, and malicious semantic annotation; and the dataset partitioning unit completes the proportional partitioning of the training set, validation set, and test set of the labeled data to construct a standardized multilingual social network malicious corpus.

[0034] Specifically, the corpus collection unit is used to collect text data from multilingual social networks. The collection scope covers mainstream common languages ​​and high-frequency minor languages ​​in social scenarios, as well as all types of social text content such as short text comments, topic posts, and user private messages, ensuring the scenario coverage and language coverage of the corpus.

[0035] The language recognition unit is used to determine the language attribute of the input text. At the same time, for multilingual mixed texts that frequently appear in social scenarios, it completes the language boundary labeling of each paragraph and word, providing the language attribute basis for subsequent word segmentation and encoding.

[0036] The text preprocessing unit is used to perform standardized preprocessing operations on the input text, perform adaptive word segmentation and sub-word segmentation according to the language characteristics of different languages, solve the problem of out-of-vocabulary words in low-resource languages, clean up meaningless noise content such as @users, invalid topic tags, and special symbols in social texts, and at the same time perform malicious semantic annotation of the text, clarify the malicious semantic category of the text, and provide annotation ground truth for model training.

[0037] The dataset partitioning unit is used to divide the labeled effective data into training, validation and test sets according to a preset ratio. These sets are used for model training, hyperparameter tuning and model performance verification, respectively, and finally a standardized multilingual social network malicious corpus is constructed.

[0038] In the cross-linguistic semantic encoding stage, the word embedding layer of the cross-linguistic pre-trained model is used by the cross-linguistic word embedding unit to complete the unified semantic space mapping of words from different languages; the self-attention mechanism is adjusted by the Transformer encoding unit through local attention optimization and the addition of a local attention window to complete the multi-layer Transformer encoding of the input text and output the semantic features of different network layers; the standardized output of the high-dimensional semantic features after encoding is completed by the initial feature output unit.

[0039] Specifically, the cross-linguistic word embedding unit uses the word embedding layer of cross-linguistic pre-trained models such as XLM-RoBERTa and LaBSE to complete the unified semantic space mapping of words in different languages, transforming words in different languages ​​into vector representations in the same semantic space, achieving basic semantic alignment at the word level of different languages, and eliminating the heterogeneity of different languages ​​at the word surface.

[0040] The locally attention-optimized Transformer encoding unit, built on a variant Transformer architecture, is used to perform deep semantic encoding of the input text. Addressing the semantic sparseness of short social texts, this unit adjusts the self-attention mechanism of the native Transformer by adding a local attention window to limit the computational scope of self-attention, enhancing the ability to capture the contextual semantics of short texts and reducing redundant attention computation. Through the stacking of multiple Transformer encoders, the unit completes deep feature extraction from the lexical, syntactic, and semantic levels, outputting semantic features corresponding to different network layers. The lower-level encoders output basic lexical and syntactic features, while the higher-level encoders output deep features at the semantic and malicious attribute levels, providing multi-dimensional input features for subsequent feature fusion.

[0041] The initial feature output unit is used to standardize the output of the encoded high-dimensional semantic features, and transmits the semantic features of different levels to the subsequent cross-layer malicious semantic fusion stage in accordance with a unified format specification.

[0042] In the cross-layer malicious semantic fusion stage, the hierarchical feature adaptation unit receives semantic features output from different network layers of the Transformer encoding unit to complete the dimensional alignment of features at different levels; the cross-layer attention weight calculation unit adaptively calculates the fusion weights of features at different levels, and the corresponding calculation formula is as follows: , in the formula, The fusion weights are the encoded features of the l-th Transformer layer. This represents the total number of layers in the Transformer encoder. The semantic feature vector output by the l-th layer Transformer encoder. This is a trainable attention weight mapping matrix. This is the semantic baseline vector corresponding to prior knowledge of cross-language malicious keywords. The function is a natural exponential function. Based on the calculated fusion weights, the cross-layer attention weight calculation unit performs a weighted summation of semantic features at each level to obtain the initial cross-lingual malicious semantic fusion features. The malicious semantic feature enhancement unit introduces prior knowledge of cross-lingual malicious keywords to enhance the malicious dimension of the fusion features. The fusion feature output unit completes the standardized output of the enhanced cross-lingual malicious semantic fusion features.

[0043] Specifically, the hierarchical feature adaptation unit is used to receive semantic features output from different network layers of the Transformer coding unit, and to complete the unified alignment of feature dimensions for the dimensional differences of different levels of features, so as to ensure the feasibility of subsequent fusion calculations.

[0044] The cross-layer attention weight calculation unit is used to adaptively calculate the fusion weights of semantic features at different levels, as well as the weighted fusion of multi-layer features. This unit calculates the corresponding fusion weights for the semantic features output by each Transformer layer using the above calculation formula and combined with prior knowledge of cross-language malicious keywords. The weights are positively correlated with the correlation between the features of that layer and the malicious semantic baseline vector, so as to focus on strengthening the high-level features with high correlation to malicious semantics, while retaining the cross-language common syntactic information in the lower-level features. After the weight calculation is completed, the unit performs a weighted summation of the semantic features at all levels based on the calculated fusion weights of each level to obtain the initial cross-language malicious semantic fusion features.

[0045] The core function of the formula is to achieve adaptive allocation of fusion weights, replacing the fixed-weight fusion method. This allows the model to automatically adjust the fusion ratio of different levels of features based on the characteristics of the input text, thereby improving the ability of the fused features to represent malicious semantics. The annotations of each symbol in the formula are clearly defined in the standard text, including the trainable attention weight mapping matrix. It will be updated synchronously during the model iteration and optimization phase to ensure the accuracy of weight calculation;

[0046] The malicious semantic feature enhancement unit is used to introduce prior knowledge of cross-language malicious keywords, perform targeted enhancement of the malicious dimension of the initial fused features, amplify the feature dimensions that are strongly related to malicious semantics in the fused features through feature mapping operations, suppress language-specific irrelevant feature dimensions, further remove language-specific noise, and improve the cross-language generalization ability of the fused features.

[0047] The fusion feature output unit is used to complete the standardized output of enhanced cross-language malicious semantic fusion features, and transmits the fusion features to the subsequent clustering feature purification stage according to a unified format specification.

[0048] In the clustering feature purification stage, the clustering parameter initialization unit initializes the cluster centers and membership matrices based on preset malicious semantic categories; the fuzzy C-means clustering iterative calculation unit performs iterative optimization calculations with the objective function as the goal, and the corresponding calculation formula for the objective function is: , in the formula, Let C be the objective function for fuzzy C-means clustering. For fuzzy weighted index, This represents the total number of input feature vectors. The preset total number of cluster centers. For the first The i-th feature vector corresponds to the i-th eigenvector. Membership degree of each cluster center For the first A cross-language malicious semantic fusion feature vector of each input. For the first Feature vectors of cluster centers The Euclidean norm is used; the membership matrix and cluster centers are updated synchronously in each iteration through the fuzzy C-means clustering iterative calculation unit. The iteration process continues until the change in the objective function is lower than the preset convergence threshold, thus completing the soft clustering of the input fusion features; the feature selection and purification unit removes noisy features and outliers based on the membership threshold, and filters redundant feature dimensions; the final feature output unit completes the standardized output of the purified language-independent malicious semantic discrimination features.

[0049] Specifically, the clustering parameter initialization unit is used to complete the parameter initialization settings of the fuzzy C-means clustering algorithm. Based on the preset number of malicious semantic categories, it sets the corresponding total number of cluster centers, and at the same time completes the assignment of initial cluster centers and the initialization of the membership matrix, providing initial parameters for subsequent iterative clustering calculations.

[0050] The fuzzy C-means clustering iterative computation unit is used to perform soft clustering calculations on the input fused features. Its core function is to iteratively optimize and minimize the aforementioned objective function, achieving cluster alignment of malicious semantic features across different languages. In each iteration, this unit synchronously updates the membership matrix and cluster centers. The iteration continues until the change in the objective function falls below a preset convergence threshold, at which point the iteration terminates. Through this soft clustering mechanism, the unit assigns feature vectors from different languages ​​and with different surface expressions, but sharing the same core malicious semantics, to the same cluster center, achieving deep alignment of cross-lingual malicious semantic features and further mitigating the heterogeneity of feature distribution in multilingual data.

[0051] The core function of the formula is to serve as the objective function for iterative optimization of clustering. By minimizing the sum of weighted squared distances from all feature vectors to their corresponding cluster centers, it achieves the optimal clustering effect, where the fuzzy weighted exponent... Used to control the fuzziness of clustering, ensuring the flexibility of soft clustering and adapting to the distribution characteristics of cross-linguistic features;

[0052] The feature selection and purification unit is used to select and purify the features after clustering. Based on the preset membership threshold, it removes noisy features and outliers with membership below the threshold, and filters out redundant feature dimensions that are irrelevant to malicious semantics and are language-specific, so as to obtain malicious semantic discrimination features with strong discriminativeness and language independence.

[0053] The final feature output unit is used to standardize the output of language-independent malicious semantic discrimination features after purification, and transmits the features to the subsequent model discrimination and iterative optimization stages according to a unified format specification.

[0054] In the model discrimination and iterative optimization phase, the semantic classification unit receives the purified malicious semantic discrimination features and outputs the malicious semantic category and confidence level of the input text; the loss calculation unit calculates the cross-entropy classification loss based on the labeled real labels and the model prediction results; the backpropagation optimization unit updates the trainable parameters of the entire model chain based on the calculated loss value; and the model convergence verification unit performs hyperparameter tuning and determines the model convergence state based on the index data of the validation set.

[0055] Specifically, the semantic classification unit consists of a fully connected layer and a Softmax classification layer. It is used to receive the purified malicious semantic discrimination features, complete the dimensional mapping and classification calculation of the features, and finally output the malicious semantic category corresponding to the input text, as well as the confidence level of the corresponding category, thus completing the final discrimination of malicious semantics.

[0056] The loss calculation unit is used to calculate the model classification loss. Based on the real labels of the text annotation and the prediction results output by the semantic classification unit, it calculates the cross-entropy classification loss. The loss value is used to measure the degree of deviation between the model prediction results and the real labels, and provides a basis for subsequent parameter updates.

[0057] The backpropagation optimization unit is used to update the trainable parameters of the entire model. Based on the loss value output by the loss calculation unit, it calculates the gradient of the trainable parameters of the entire model through the backpropagation algorithm. According to the preset optimizer and learning rate, it completes the iterative update of parameters, including all trainable parameters such as the network parameters of the Transformer encoding unit and the mapping matrix of the cross-layer attention weight calculation unit.

[0058] The model convergence verification unit is used to complete the convergence verification and hyperparameter tuning of the model. It calculates the model's classification accuracy, F1 score and other indicators through the validation set data, completes the hyperparameter tuning of the model, and monitors the changes in the model indicators. When the model's indicators on the validation set do not show significant improvement and tend to stabilize, the model is determined to have converged and the model training is terminated.

[0059] The method of use and working principle of this invention are as follows:

[0060] Usage: First, collect, identify, preprocess, and annotate malicious semantics of multilingual text data for the target multilingual social network scenario. Construct a standardized multilingual malicious corpus and divide the dataset. Then, input the divided training data into a cross-language semantic coding model based on a Transformer variant to complete the semantic space mapping and multi-layer semantic feature extraction of texts in different languages. Subsequently, perform dimension alignment and adaptive cross-layer attention weighted fusion on the extracted semantic features at different levels. Combine with prior knowledge of malicious semantics to complete feature enhancement and obtain cross-language malicious semantic fusion features. Then, use the fuzzy C-means clustering algorithm to perform iterative clustering calculation and feature purification on the fusion features to obtain language-independent malicious semantic discrimination features. Finally, complete the malicious semantic classification and discrimination based on the purified features. The backpropagation update and convergence verification of the model's full-link parameters are completed through classification loss. The trained model can be directly deployed on multilingual social networks to complete the extraction of malicious semantic features and identification of malicious content from input text.

[0061] Working principle: Based on the cross-language pre-trained Transformer variant architecture, the model achieves basic alignment of texts in different languages ​​in a unified semantic space, eliminating the superficial heterogeneity of different languages ​​at the lexical and syntactic levels. Through a cross-layer attention adaptive fusion mechanism, it achieves weighted fusion of semantic features at different levels of the Transformer, specifically strengthening feature dimensions strongly correlated with malicious semantics, removing language-specific irrelevant noise, and constructing malicious semantic fusion features with cross-language generalization capabilities. Then, through a soft clustering mechanism of fuzzy C-means clustering, features from different languages ​​but with consistent core malicious semantics are mapped to the same feature subspace, further eliminating the deep feature distribution heterogeneity of multilingual data, and refining language-independent, highly discriminative malicious semantic features. Finally, the model achieves accurate discrimination of malicious semantics through a classification module, while continuously optimizing the model's feature extraction and fusion capabilities through end-to-end iterative optimization.

[0062] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Those skilled in the art can readily implement the present invention based on the description and drawings above. However, any modifications, alterations, and variations made by those skilled in the art without departing from the scope of the present invention using the disclosed technical content are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, and variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.

Claims

1. A method for malicious semantic fusion and feature extraction for multilingual social networks, characterized in that, Includes the following steps: S1. Multilingual Corpus Construction and Preprocessing Stage: Collect multilingual social network text data, complete language identification, text preprocessing, malicious semantic annotation and dataset partitioning, and construct a standardized multilingual social network malicious corpus; S2, Cross-language semantic encoding stage: Based on the Transformer variant architecture, an encoding model is built to complete the cross-language word embedding mapping and multi-layer semantic encoding of the input text, and output the initial semantic features of the multilingual text. S3, Cross-layer malicious semantic fusion stage: Receives semantic features from different network layers, completes layer feature dimension alignment, cross-layer attention weight calculation and multi-layer feature weighted fusion, combines malicious semantic prior knowledge to complete feature enhancement, and outputs cross-language malicious semantic fusion features; S4. Cluster feature purification stage: Based on the fuzzy C-means clustering algorithm, the initialization of clustering parameters, iterative clustering calculation and feature selection and purification are completed, and language-independent malicious semantic discrimination features are output. S5. In the model discrimination and iterative optimization stage, malicious semantic classification is completed based on the purified features, the classification loss is calculated and backpropagation parameter updates are performed to complete the model convergence verification and hyperparameter tuning.

2. The method for malicious semantic fusion and feature extraction for multilingual social networks according to claim 1, characterized in that, In the multilingual corpus construction and preprocessing stage, the corpus acquisition unit collects multilingual social network text data, covering multiple languages ​​and all types of social text content; the language identification unit determines the language of the input text and marks the language boundaries of multilingual mixed text; the text preprocessing unit performs text segmentation, sub-word segmentation, noise content cleaning, and malicious semantic annotation; and the dataset partitioning unit divides the labeled data into training, validation, and test sets to construct a standardized multilingual social network malicious corpus.

3. The method for malicious semantic fusion and feature extraction for multilingual social networks according to claim 1, characterized in that, In the cross-linguistic semantic encoding stage, the cross-linguistic word embedding unit uses the word embedding layer of the cross-linguistic pre-trained model to complete the unified semantic space mapping of words from different languages; the local attention optimization Transformer encoding unit adjusts the self-attention mechanism and adds a local attention window to complete the multi-layer Transformer encoding of the input text and outputs semantic features of different network layers; the initial feature output unit completes the standardized output of the encoded high-dimensional semantic features.

4. The method for malicious semantic fusion and feature extraction for multilingual social networks according to claim 1, characterized in that, In the cross-layer malicious semantic fusion stage, the hierarchical feature adaptation unit receives semantic features output from different network layers of the Transformer encoding unit to complete the dimensional alignment of features at different levels; the cross-layer attention weight calculation unit adaptively calculates the fusion weights of features at different levels, and the corresponding calculation formula is as follows: , in the formula, The fusion weights are the encoded features of the l-th Transformer layer. This represents the total number of layers in the Transformer encoder. The semantic feature vector output by the l-th layer Transformer encoder. This is a trainable attention weight mapping matrix. This is the semantic baseline vector corresponding to prior knowledge of cross-language malicious keywords. The function is a natural exponential function. Based on the calculated fusion weights, the cross-layer attention weight calculation unit performs a weighted summation of semantic features at each level to obtain the initial cross-language malicious semantic fusion features. The malicious semantic feature enhancement unit introduces prior knowledge of cross-language malicious keywords to enhance the malicious dimension of the fusion features. The fusion feature output unit completes the standardized output of the enhanced cross-language malicious semantic fusion features.

5. The method for malicious semantic fusion and feature extraction for multilingual social networks according to claim 1, characterized in that, In the clustering feature purification stage, the clustering parameter initialization unit initializes the cluster centers and membership matrices based on preset malicious semantic categories; the fuzzy C-means clustering iterative calculation unit performs iterative optimization calculations with the objective function as the goal, and the corresponding calculation formula for the objective function is: , in the formula, Let C be the objective function for fuzzy C-means clustering. For fuzzy weighted index, This represents the total number of input feature vectors. The preset total number of cluster centers. For the first The i-th feature vector corresponds to the i-th eigenvector. Membership degree of each cluster center For the first A cross-language malicious semantic fusion feature vector of each input. For the first Feature vectors of cluster centers The Euclidean norm is used; the membership matrix and cluster centers are updated synchronously in each iteration by the fuzzy C-means clustering iterative calculation unit. The iteration process continues until the change in the objective function is lower than the preset convergence threshold, thus completing the soft clustering of the input fusion features. The feature selection and purification unit removes noisy and outlier features and filters redundant feature dimensions based on the membership threshold. The final feature output unit outputs the purified language-independent malicious semantic discrimination features in a standardized manner.

6. The method for malicious semantic fusion and feature extraction for multilingual social networks according to claim 1, characterized in that, In the model discrimination and iterative optimization stage, the semantic classification unit receives the purified malicious semantic discrimination features and outputs the malicious semantic category and confidence level of the input text; the loss calculation unit calculates the cross-entropy classification loss based on the labeled real labels and the model prediction results; the backpropagation optimization unit updates the trainable parameters of the entire model chain based on the calculated loss value; and the model convergence verification unit performs hyperparameter tuning and determines the model convergence state based on the index data of the verification set.

Images