Intelligent fraud-related short message identification method and device based on NLP

Abstract

The embodiment of the invention provides an NLP-based fraud-related short message intelligent identification method and device, and the method and device achieve the precise understanding of the short message content through the innovative construction of a text analysis mechanism, semantic structure extraction and scene classification. And designing a feature extraction model based on sequence labeling, and establishing a number website association analysis strategy for intelligent identification in combination with a bidirectional recurrent neural network and a conditional random field. A deep fusion evaluation mechanism is introduced, and accurate division and timely disposal of risk levels are realized through a residual neural network and nonlinear mapping. According to the method, the defects of the traditional technology in the aspects of text analysis, feature extraction, risk assessment and the like are effectively overcome, and the accuracy and reliability of fraud-related short message recognition are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to a method and apparatus for intelligent identification of fraudulent text messages based on NLP. Background Technology

[0002] Existing methods for identifying fraudulent text messages have significant shortcomings. Traditional systems lack a systematic approach to text analysis, making it difficult to effectively extract and analyze the semantic structure and contextual features of text messages, thus affecting the accuracy of identification.

[0003] Furthermore, existing technologies face bottlenecks in sequence feature extraction. Most systems fail to fully utilize temporal correlation information and lack labeling and modeling mechanisms based on conditional random fields, resulting in less than ideal analysis of the association between phone numbers and website addresses.

[0004] Existing systems have technical shortcomings in risk assessment. They lack the ability to deeply integrate features, making it difficult to accurately classify risks through residual networks and nonlinear mappings, thus affecting the effectiveness of responses. Solving these problems is crucial for improving the identification of fraudulent text messages. Summary of the Invention

[0005] To address the problems in existing technologies, this application provides a method and apparatus for intelligent identification of fraudulent text messages based on NLP, which can effectively solve the shortcomings of traditional technologies in feature fusion, sample processing and identification strategies, and significantly improve the intelligence level and prevention and control effect of fraudulent text message identification.

[0006] To solve at least one of the above problems, this application provides the following technical solution:

[0007] Firstly, this application provides a method for intelligent identification of fraudulent text messages based on NLP, including:

[0008] Collect SMS text data, perform word segmentation on the SMS text data to obtain word sequences, construct a text vectorization model, convert the word sequences into text feature vectors, perform syntactic dependency analysis on the text feature vectors to obtain semantic structure features, combine the semantic structure features with a scene tag library to classify the scene, and generate scene feature vectors.

[0009] A sequence labeling network is constructed, and the scene feature vector is input into a bidirectional recurrent neural network to extract temporal association features. Conditional random fields are used to label and model the temporal association features to identify the reserved number and website information in the SMS. Based on the association rule analysis model, the association degree score between the number and the website is calculated to generate an association feature matrix.

[0010] The associated feature matrix is ​​input into the risk assessment model, which uses a residual neural network structure to fuse the features, performs nonlinear mapping on the fused features through a multilayer perceptron, classifies the SMS messages based on a preset threshold, and outputs the classification results to the fraud handling module.

[0011] Furthermore, it also includes: receiving SMS text data streams from the SMS monitoring platform, performing encoding format conversion and cleaning on the SMS text data streams, removing special characters and redundant spaces, extracting text structure information based on regular expressions, segmenting the cleaned text into sentences according to punctuation marks, using the maximum matching algorithm to perform word segmentation on the sentence segmentation results, optimizing the word segmentation results by combining a custom dictionary, and combining the optimized word segmentation sequence into a word sequence;

[0012] A text vectorization model is constructed. The text vectorization model is based on the bag-of-words model to statistically analyze word frequency features. The word frequency-inverse document frequency algorithm is used to calculate word weights. The word sequence is represented by a sparse matrix. The sparse matrix is ​​input into a word embedding layer. The word embedding layer maps words into low-dimensional dense vectors through a lookup table. The mapping results are concatenated and combined to generate text feature vectors.

[0013] Furthermore, it also includes: constructing a syntactic analysis network, inputting the text feature vector into a dependency parser, wherein the dependency parser constructs a syntactic tree using a transfer arc mechanism, calculates the probability of dependency relationships between words based on a dynamic programming algorithm, selects the optimal dependency path to generate a syntactic structure graph, performs a depth-first traversal on the syntactic structure graph, extracts the hierarchical relationship between the head word and the modifier, and converts the hierarchical relationship information into semantic structure features;

[0014] The scene classification model is constructed by loading predefined scene type information from the scene tag library. The scene classification model uses an attention mechanism to weight the key components in the semantic structure features, performs nonlinear transformation on the weighted features based on the multilayer perceptron structure, calculates the scene category probability distribution through the softmax function, and concatenates the probability distribution results with the semantic structure features to obtain the scene feature vector.

[0015] Furthermore, it also includes: constructing a bidirectional recurrent neural network, organizing the scene feature vector into a feature sequence in chronological order, wherein the bidirectional recurrent neural network contains two hidden states, forward and backward, controlling the information flow through a gating mechanism, performing bidirectional scanning on the feature sequence, capturing long-term dependencies in the sequence based on long short-term memory units, and using an attention pooling layer to weight and aggregate the hidden states to generate temporally related features;

[0016] The temporal correlation features are input into a conditional random field model. The conditional random field model calculates the conditional probability of the label sequence based on the state transition matrix, constructs a potential function to model the constraint relationship between adjacent labels, and uses the Viterbi algorithm to search for the optimal path among all possible label paths. The optimal path is then decoded to obtain the labeling result.

[0017] Furthermore, it also includes: constructing an information extraction module, extracting phone number and URL fields from SMS text based on regular expression pattern matching, parsing the phone number segment location and standardizing the number format, using a URL parser to perform domain name segmentation and parameter extraction on the URL, and indexing the extracted number information and URL information according to their positional relationship in the text to generate a structured information object sequence;

[0018] The sequence of information objects is input into an association rule analysis model. The association rule analysis model uses a sliding time window to calculate the co-occurrence frequency of numbers and URLs, scores the co-occurrence patterns based on the association degree calculation formula, and constructs a feature combination network. The feature combination network models the association strength of different feature pairs through a multi-head attention mechanism and organizes the association strength scores into an association feature matrix.

[0019] Furthermore, it also includes: constructing a feature preprocessing module to normalize the associated feature matrix to eliminate dimensional differences, using principal component analysis to reduce the dimensionality of the features, inputting the dimensionality-reduced features into a residual neural network, wherein the residual neural network retains the original feature information through a skip connection structure, performs local feature extraction based on convolutional layers, adjusts the feature distribution using batch normalization layers, and generates a residual feature map;

[0020] A feature fusion network is constructed, which uses a multi-branch structure to process residual features at different scales in parallel. The importance weights of feature channels are calculated through a channel attention module, and the spatial dependencies of features are captured based on a spatial attention module. The multi-branch features are fused by weighted summation, and the fusion result is nonlinearly transformed to obtain the fused feature vector.

[0021] Furthermore, it also includes: constructing a risk mapping network, inputting the fused feature vector into a multilayer perceptron, wherein the multilayer perceptron transforms the features through hidden layer neurons, uses a batch normalization layer to standardize the intermediate features, introduces nonlinear characteristics based on the activation function, prevents gradient vanishing through the residual connection structure, and uses the sigmoid function to map the features to the risk probability space on the output of the last layer to generate a risk score result;

[0022] A tiered processing module is constructed to compare the risk score results with multi-level preset thresholds. Based on the threshold segmentation rules, the risk level is divided into three levels: high, medium, and low. Differentiated handling strategies are configured for different risk levels. The risk level label and handling strategy information are encapsulated into handling instructions. The handling instructions are distributed to the fraud handling module according to priority through a message queue.

[0023] Secondly, this application provides an NLP-based intelligent identification device for fraudulent text messages, comprising:

[0024] The text recognition module is used to collect SMS text data, perform word segmentation on the SMS text data to obtain word sequences, construct a text vectorization model, convert the word sequences into text feature vectors, perform syntactic dependency analysis on the text feature vectors to obtain semantic structure features, and combine the semantic structure features with a scene tag library to classify the scene and generate scene feature vectors.

[0025] The association analysis module is used to construct a sequence labeling network, input the scene feature vector into a bidirectional recurrent neural network to extract temporal association features, use a conditional random field to label and model the temporal association features, identify the reserved number and website information in the SMS, calculate the association degree score between the number and the website based on the association rule analysis model, and generate an association feature matrix.

[0026] The risk warning module is used to input the associated feature matrix into the risk assessment model. The risk assessment model uses a residual neural network structure to fuse the features, performs nonlinear mapping on the fused features through a multilayer perceptron, classifies the SMS messages based on a preset threshold, and outputs the classification results to the fraud handling module.

[0027] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the NLP-based intelligent identification method for fraudulent text messages.

[0028] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the NLP-based intelligent identification method for fraudulent text messages.

[0029] Fifthly, this application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the NLP-based intelligent identification method for fraudulent text messages.

[0030] As described above, this application provides a method and apparatus for intelligent identification of fraudulent text messages based on NLP. By innovatively constructing a text analysis mechanism, it achieves accurate understanding of text message content through semantic structure extraction and scene classification. A sequence labeling-based feature extraction model is designed, combined with bidirectional recurrent neural networks and conditional random fields, to establish a number / website association analysis strategy for intelligent identification. A deep fusion evaluation mechanism is introduced, using residual neural networks and nonlinear mapping to achieve accurate risk level classification and timely handling. This method effectively addresses the shortcomings of traditional technologies in text analysis, feature extraction, and risk assessment, significantly improving the accuracy and reliability of fraudulent text message identification. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart illustrating the NLP-based intelligent identification method for fraudulent text messages in this application embodiment;

[0033] Figure 2 This is a structural diagram of the NLP-based intelligent identification device for fraudulent text messages in this application embodiment;

[0034] Figure 3 This is a schematic diagram of the structure of the electronic device in the embodiments of this application.

[0035] Figure label:

[0036] Electronic device 9600, central processing unit 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver storage unit 9144, antenna 9111, speaker 9131, microphone 9132. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0038] The acquisition, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0039] To address the shortcomings of existing technologies, this application provides a method and apparatus for intelligent identification of fraudulent text messages based on NLP. By innovatively constructing a text analysis mechanism, it achieves accurate understanding of text message content through semantic structure extraction and scene classification. A sequence labeling-based feature extraction model is designed, combining bidirectional recurrent neural networks and conditional random fields to establish a number / website association analysis strategy for intelligent identification. A deep fusion evaluation mechanism is introduced, using residual neural networks and nonlinear mapping to achieve accurate risk level classification and timely handling. This method effectively solves the deficiencies of traditional technologies in text analysis, feature extraction, and risk assessment, significantly improving the accuracy and reliability of fraudulent text message identification.

[0040] To effectively address the shortcomings of traditional technologies in feature fusion, sample processing, and identification strategies, and to significantly improve the intelligence level and prevention effectiveness of fraudulent SMS identification, this application provides an embodiment of an NLP-based intelligent identification method for fraudulent SMS. See [link to embodiment]. Figure 1 The NLP-based intelligent identification method for fraudulent text messages specifically includes the following:

[0041] Step S101: Collect SMS text data, perform word segmentation on the SMS text data to obtain word sequences, construct a text vectorization model, convert the word sequences into text feature vectors, perform syntactic dependency analysis on the text feature vectors to obtain semantic structure features, combine the semantic structure features with a scene tag library to classify the scene, and generate scene feature vectors.

[0042] Optionally, this embodiment addresses the problems of low text processing efficiency and inaccurate semantic understanding in the identification of fraudulent SMS messages by innovatively designing a text analysis scheme based on natural language processing. This embodiment first constructs a text preprocessing framework, achieving a structured expression of the SMS content through multi-level text analysis. The system designs a text feature extraction formula: Text_Score=αWord_Freq+βSemantic_Rel+γ×Scene_Type, where Word_Freq represents word frequency features, Semantic_Rel represents semantic relevance, Scene_Type represents scene type features, and α, β, and γ are dynamic weight coefficients. In the scenario of fraudulent SMS message identification, this multi-dimensional feature extraction method can effectively capture the semantic information of the SMS text.

[0043] An adaptive word segmentation rule base was designed to address the unique expressions used in fraudulent text messages. By combining a domain dictionary and a new word discovery mechanism, the accuracy of word segmentation is improved. Special attention is paid to mixed expressions of numbers and letters; when a suspicious character combination is detected, the system activates a specific segmentation rule. For example, when processing mixed encodings like "For details, consult 4OO8208820", variant forms of the phone number are identified using feature templates to ensure that key information is not incorrectly segmented. This intelligent word segmentation scheme provides a reliable foundation for subsequent feature extraction.

[0044] To address the characteristics of SMS text, the system constructs a feature extraction framework based on word embeddings. A pre-trained word vector model maps the word segmentation results to a high-dimensional semantic space. Particular attention is paid to the representation of domain-specific vocabulary, and word vectors are fine-tuned through transfer learning to better adapt to the semantic characteristics of fraud-related scenarios. This deep learning-based text representation method effectively captures the semantic relationships between words.

[0045] The system constructs a semantic structure extraction framework based on dependency relation analysis, building a complete syntactic tree by identifying the modification and dominance relationships between words. It pays particular attention to key semantic structures in text messages, improving the ability to identify fraudulent sentence patterns by designing specific dependency relation templates. This comprehensive syntactic analysis method provides structured feature support for scene classification.

[0046] The system constructs a scene recognition framework based on a tag library, combining an attention mechanism to weight semantic features. It pays particular attention to the dynamic changes in scenes, continuously updating the scene feature library through online learning to adapt to the evolution of new fraud methods. This systematic scene analysis approach provides a semantic basis for risk identification.

[0047] This embodiment constructs a scenario tag library for high-risk SMS messages, focusing on three major categories of high-risk scenarios: fraud, pornography, and gambling. Among them, the fraud scenario is particularly complex, including specific scenario types such as impersonating e-commerce customer service (e.g., falsely claiming that an order is abnormal and needs to be processed), impersonating express delivery customer service (e.g., falsely claiming that a package is abnormal and needs a refund), fake investment and wealth management (e.g., fabricating high-yield investment projects), fake educational institution refunds (e.g., falsely claiming to handle tuition refunds), and fake part-time job scams (e.g., committing fraud under the guise of boosting online reputation).

[0048] During the processing, the key fraud features in the text are first identified, including words related to financial operations (such as "transfer", "remittance", "refund", etc.), words related to identity spoofing, misleading expressions (such as "urgent", "process immediately", "missed and void"), and link address features.

[0050] The feature vector comprises textual semantic features (reflecting the linguistic characteristics of fraudulent rhetoric), behavioral features (describing the specific methods of inducement), and relational features (representing similarity to known fraud patterns). The system captures the complex relationships between these features using a deep learning model and dynamically adjusts feature weights based on the characteristics of different scenarios. For newly emerging fraud variants, the system can quickly construct corresponding feature representations based on existing knowledge, maintaining strong scenario adaptability.

[0051] The innovation of this embodiment lies primarily in its refined classification of fraud scenarios and dynamic feature extraction. By establishing a complete risk scenario classification system, from scenario feature identification to risk level assessment, accurate identification of high-risk SMS messages is achieved. In practical applications, this scenario classification scheme based on deep semantic understanding significantly enhances the protection capabilities of the anti-fraud system. For example, when identifying new types of fraudulent SMS messages, the system can quickly summarize fraud characteristics and update protection strategies in a timely manner, providing users with more reliable security.

[0052] This innovative design not only solves the text processing problems of traditional methods but also establishes a continuously optimizable feature extraction framework. Through multi-level semantic analysis, the system can extract effective risk features from complex SMS text. This natural language processing-based analysis mechanism ensures that the system maintains effective identification capabilities when faced with diverse fraudulent texts. In fraudulent SMS identification, this intelligent text analysis scheme significantly improves the accuracy of feature extraction.

[0053] This embodiment achieves an intelligent upgrade in identifying fraudulent SMS messages by establishing a complete text analysis chain. The system can dynamically adjust its analysis strategy based on real-time text features, avoiding the limitations of traditional fixed-rule solutions. Through multi-dimensional feature extraction and scenario analysis, the accuracy and reliability of SMS classification are significantly improved, providing reliable feature support for subsequent risk assessment. This intelligent analysis mechanism demonstrates strong adaptability and optimization effects in identifying fraudulent SMS messages.

[0054] This embodiment not only improves the accuracy of text analysis but also establishes a continuously evolving identification system through ongoing strategy optimization and effect analysis. This real-time feedback-based optimization mechanism ensures that the system can continuously improve as fraud methods change, providing users with increasingly accurate protection. In practical applications, this self-optimization mechanism significantly improves the system's long-term service quality and protective effectiveness, providing reliable technical support for anti-fraud efforts.

[0055] Step S102: Construct a sequence labeling network, input the scene feature vector into a bidirectional recurrent neural network to extract temporal association features, use a conditional random field to label and model the temporal association features, identify the reserved number and website information in the SMS, calculate the association degree score between the number and the website based on the association rule analysis model, and generate an association feature matrix;

[0056] Optionally, this embodiment addresses the challenge of identifying and analyzing key information in fraudulent text messages by innovatively designing a sequence labeling scheme based on deep learning. This embodiment first constructs a bidirectional recurrent neural network framework, capturing contextual semantic dependencies through forward and backward states. The system designs a sequence feature extraction formula: H_t=αLSTM_f(X_t)+βLSTM_b(X_t)+γ×Context, where LSTM_f and LSTM_b represent the outputs of the forward and backward LSTM units, respectively, X_t is the input feature at time t, Context is the contextual information, and α, β, and γ are adaptive weight coefficients. In the scenario of fraudulent text message identification, this bidirectional modeling method can effectively capture the contextual semantic association between phone numbers and website addresses.

[0057] To address the sequence dependency characteristics in SMS text messages, a feature extraction mechanism based on Long Short-Term Memory (LSTM) was designed. Through the synergistic action of input gates, forget gates, and output gates, selective retention of long-term dependency information is achieved. Particular attention is paid to the contextual environment of key information; when a suspicious semantic pattern is detected, the system enhances the feature weights at the corresponding location through a gating mechanism. For example, when analyzing contact information with suggestive descriptions, selective updates of the memory unit can accurately capture the semantic association between the number or website and the suggestive text, which is crucial for identifying the integrity of fraudulent messages.

[0058] To address the overall constraints of sequence labeling, the system constructs a probabilistic model based on state transitions. By calculating the transition probabilities between adjacent labels, global consistency of the labeling results is established. Particular attention is paid to the rationality of the label sequences; the system designs feature functions and transition functions to ensure that the labeling results conform to predefined grammatical rules. This probabilistic sequence labeling method can effectively identify key information fragments in SMS messages.

[0059] The system constructs an adaptive correlation calculation framework based on co-occurrence statistics, analyzing local co-occurrence patterns of phone numbers and website addresses through a sliding window mechanism. Particular attention is paid to the timeliness of correlation patterns, dynamically weighting historical co-occurrence information by designing a time decay factor. This comprehensive correlation analysis method provides reliable feature support for risk assessment.

[0060] The system constructs a feature combination framework based on multi-head attention, fusing different types of correlated features. Particular attention is paid to feature complementarity; by establishing a feature interaction layer, the correlations between features of different dimensions are captured. This systematic feature representation scheme provides comprehensive data support for subsequent risk assessment.

[0061] This embodiment's innovative design not only solves the information extraction problem in traditional methods but also establishes a continuously optimizing feature analysis framework. Through multi-level sequence labeling and association analysis, the system can extract effective risk features from complex SMS text. This deep learning-based analysis mechanism ensures that the system maintains effective identification capabilities when faced with diverse fraudulent texts. In anti-fraud applications, this intelligent feature extraction scheme significantly improves the accuracy of risk identification.

[0062] This embodiment achieves an intelligent upgrade in identifying fraudulent text messages by establishing a complete feature analysis chain. The system can dynamically adjust its analysis strategy based on real-time text features, avoiding the limitations of traditional fixed-rule solutions. Through multi-dimensional information extraction and correlation analysis, the accuracy and reliability of identification are significantly improved, providing more comprehensive feature support for risk assessment. This intelligent analysis mechanism demonstrates strong adaptability and optimization effects in fraudulent text message identification scenarios.

[0063] This embodiment not only improves the accuracy of information extraction but also establishes a continuously evolving identification system through ongoing strategy optimization and effect analysis. This real-time feedback-based optimization mechanism ensures that the system can continuously improve as fraud methods change, providing users with increasingly accurate protection. In practical applications, this self-optimization mechanism significantly improves the system's long-term service quality and protective effectiveness, providing reliable technical support for anti-fraud efforts.

[0064] Step S103: Input the associated feature matrix into the risk assessment model. The risk assessment model uses a residual neural network structure to fuse the features, performs nonlinear mapping on the fused features through a multilayer perceptron, classifies the SMS messages based on a preset threshold, and outputs the classification results to the fraud handling module.

[0065] Optionally, this embodiment addresses the problems of insufficient feature fusion and low classification accuracy in fraudulent SMS risk assessment by innovatively designing a deep learning-based risk assessment scheme. This embodiment first constructs a feature fusion framework, achieving effective integration of multi-dimensional features through a residual network. The system designs a risk assessment formula: Risk_Score=σ(W·F(x)+x), where F(x) is the residual mapping function, x is the input feature, W is the weight matrix, and σ is the activation function. In the scenario of fraudulent SMS identification, this feature fusion method based on residual learning can effectively improve the accuracy of risk assessment.

[0066] A multi-layered feature processing flow was designed for different types of associated features. A residual connection structure ensures the effective transmission of important feature information within the deep network. Special attention is paid to the temporal dependencies of features; when a sudden change in a risk feature is detected, the system enhances the feature weights for the corresponding time period. For example, when processing associated features in consecutive SMS messages, the original feature information is preserved through skip connections in the residual blocks, while high-order feature patterns are extracted through the deep network. This dual feature extraction mechanism can more accurately characterize the risk evolution process.

[0067] To address the complex distribution characteristics of risk features, a multilayer perceptron network is constructed. Through multilayer nonlinear transformations, the high-dimensional feature space is mapped to the risk probability space. With particular attention to the model's generalization ability, the system employs regularization techniques such as batch normalization and dropout to improve the model's adaptability to novel fraud methods. This deep learning-based mapping method effectively captures the complex correlations between risk features.

[0068] The system constructs a dynamic hierarchical framework based on preset thresholds, balancing the coverage of different risk levels through an adaptive threshold adjustment mechanism. Particular attention is paid to the real-time nature of the hierarchical classification, ensuring that high-risk warnings are prioritized through a warning queue management mechanism. This comprehensive hierarchical approach provides a reliable basis for the distribution of warning information.

[0069] The system constructs a result coding framework based on risk levels, combining scenario characteristics to create personalized risk descriptions. Particular attention is paid to the operability of the proposed actions, and risk handling templates are designed to help the handling module respond quickly. This systematic risk assessment approach provides reliable guidance for subsequent handling work.

[0070] This embodiment's innovative design not only solves the risk assessment problem in traditional methods but also establishes a continuously optimizing assessment framework. Through multi-level feature fusion and risk mapping, the system can generate accurate risk assessment results from a complex feature matrix. This deep learning-based assessment mechanism ensures that the system maintains effective identification capabilities when facing diverse fraud tactics. In fraudulent SMS identification, this intelligent assessment scheme significantly improves the accuracy of risk identification.

[0071] This embodiment achieves an intelligent upgrade in protection against fraudulent SMS messages by establishing a complete risk assessment chain. The system can dynamically adjust the assessment strategy based on real-time characteristics, avoiding the limitations of traditional fixed threshold schemes. Through multi-dimensional feature fusion and risk classification, the accuracy and reliability of early warnings are significantly improved, providing more precise decision support for the handling module. This intelligent assessment mechanism demonstrates strong adaptability and optimization effects in the protection against fraudulent SMS messages.

[0072] This embodiment not only improves the accuracy of risk assessment but also establishes a continuously evolving assessment system through ongoing strategy optimization and effect analysis. This real-time feedback-based optimization mechanism ensures that the system continuously improves as fraud methods evolve, providing users with increasingly precise protection. In practical applications, this self-optimization mechanism significantly enhances the system's long-term service quality and protective effectiveness, providing reliable technical support for anti-fraud efforts.

[0073] As described above, the NLP-based intelligent identification method for fraudulent text messages provided in this application can achieve accurate understanding of text message content through the innovative construction of a text analysis mechanism, semantic structure extraction, and scene classification. A sequence labeling-based feature extraction model is designed, combined with bidirectional recurrent neural networks and conditional random fields, to establish a number / website association analysis strategy for intelligent identification. A deep fusion evaluation mechanism is introduced, using residual neural networks and nonlinear mapping to achieve accurate risk level classification and timely handling. This method effectively solves the shortcomings of traditional technologies in text analysis, feature extraction, and risk assessment, significantly improving the accuracy and reliability of fraudulent text message identification.

[0074] In one embodiment of the NLP-based intelligent identification method for fraudulent text messages in this application, it may further include the following:

[0075] Step S201: Receive SMS text data stream from SMS monitoring platform, perform encoding format conversion and cleaning on the SMS text data stream, remove special characters and redundant spaces, extract text structure information based on regular expressions, divide the cleaned text into sentences according to punctuation marks, use the maximum matching algorithm to process the sentence segmentation results into words, optimize the word segmentation results by combining a custom dictionary, and combine the optimized word segmentation sequence into a word sequence.

[0076] Step S202: Construct a text vectorization model. The text vectorization model is based on the bag-of-words model to statistically analyze word frequency features. It uses the word frequency-inverse document frequency algorithm to calculate word weights, performs sparse matrix representation on the word sequence, inputs the sparse matrix into the word embedding layer, and maps words into low-dimensional dense vectors through a lookup table. The mapping results are then concatenated and combined to generate text feature vectors.

[0077] Optionally, this embodiment innovatively designs a text processing scheme to address issues such as unstable data quality and insufficient feature representation in the preprocessing and vectorized representation of fraudulent SMS messages. This embodiment first constructs a data cleaning framework, achieving standardized representation of SMS content through multi-level text processing. The system designs a text quality evaluation formula: Quality_Score=αFormat_Norm+βInfo_Complete+γ×Token_Accuracy, where Format_Norm represents the format standardization, Info_Complete represents the information completeness, Token_Accuracy represents the word segmentation accuracy, and α, β, and γ are dynamic weight coefficients. In the scenario of processing fraudulent SMS messages, this multi-dimensional quality evaluation method can effectively ensure the reliability of data processing.

[0078] This embodiment deeply optimizes the text cleaning strategy. An adaptive cleaning rule base is designed to address noise issues in SMS text. Regular expression templates are used to accurately filter special characters and redundant information. Special attention is paid to protecting information integrity; when critical information is detected as potentially being incorrectly cleaned, the system activates a protection mechanism. For example, when processing text containing distorted phone numbers, feature pattern matching ensures that mixed encoding of numbers and letters is not incorrectly processed. This intelligent cleaning scheme provides a high-quality data foundation for subsequent text analysis.

[0079] This embodiment innovatively implements sentence and word segmentation mechanisms. Targeting the characteristics of SMS text, the system constructs a word segmentation framework based on maximum matching. By combining a custom dictionary and a new word discovery mechanism, the accuracy of word segmentation is improved. Particular attention is paid to the identification of domain-specific vocabulary, and the dictionary is dynamically updated to adapt to changes in new fraudulent language. This dictionary-enhanced word segmentation method can effectively handle the special expressions in fraudulent SMS messages.

[0080] This embodiment deeply optimizes the vectorization processing strategy. The system constructs a text feature extraction framework based on the bag-of-words model and calculates word weights using the TF-IDF algorithm. Special attention is paid to the sparsity of features; a sparse matrix compression strategy is designed to improve storage and computational efficiency. This complete vectorization method provides structured feature support for subsequent deep learning analysis.

[0081] This embodiment achieves semantic representation of text through word embedding. The system constructs a word vector mapping framework based on a pre-trained model and combines domain transfer learning for model fine-tuning. Particular attention is paid to the expressive power of the semantic space; by optimizing word vector dimensions and training strategies, the discriminative power of features is improved. This systematic semantic representation scheme provides deep feature-based support for risk identification.

[0082] This embodiment's innovative design not only solves the text processing problems of traditional methods but also establishes a continuously optimizing feature extraction framework. Through multi-level text analysis and vectorized representation, the system can extract effective semantic features from the original SMS text. This deep learning-based processing mechanism ensures that the system maintains effective expressive power when faced with diverse fraudulent texts. In fraudulent SMS identification, this intelligent text processing scheme significantly improves the accuracy of feature extraction.

[0083] This embodiment establishes a complete text processing chain, achieving an intelligent upgrade for analyzing fraudulent SMS messages. The system can dynamically adjust its processing strategy based on real-time text features, avoiding the limitations of traditional fixed-rule solutions. Through multi-dimensional text processing and feature extraction, the quality and reliability of vectorized representation are significantly improved, providing reliable feature support for subsequent risk identification. This intelligent processing mechanism demonstrates strong adaptability and optimization effects in the analysis of fraudulent SMS messages.

[0084] This embodiment not only improves the accuracy of text processing but also establishes a continuously evolving processing system through ongoing strategy optimization and effect analysis. This real-time feedback-based optimization mechanism ensures that the system can continuously improve as fraud methods change, providing increasingly accurate feature representations for subsequent analysis. In practical applications, this self-optimization mechanism significantly improves the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud efforts.

[0085] In one embodiment of the NLP-based intelligent identification method for fraudulent text messages in this application, it may further include the following:

[0086] Step S301: Construct a syntactic analysis network, input the text feature vector into a dependency parser, the dependency parser uses a transfer arc mechanism to construct a syntactic tree, calculates the probability of dependency relationships between words based on a dynamic programming algorithm, selects the optimal dependency path to generate a syntactic structure graph, performs a depth-first traversal on the syntactic structure graph, extracts the hierarchical relationship between the headword and modifiers, and converts the hierarchical relationship information into semantic structure features;

[0087] Step S302: Load predefined scene type information from the scene tag library, construct a scene classification model. The scene classification model uses an attention mechanism to weight the key components in the semantic structure features, performs nonlinear transformation on the weighted features based on a multilayer perceptron structure, calculates the scene category probability distribution through the softmax function, and concatenates the probability distribution results with the semantic structure features to obtain the scene feature vector.

[0088] Optionally, this embodiment innovatively designs a deep learning-based semantic analysis scheme to address the semantic understanding and scene recognition problems in fraudulent text messages. This embodiment first constructs a syntactic analysis framework, achieving a structured representation of the text through dependency relation modeling. The system designs a dependency relation scoring formula: Dep_Score=αArc_Prob+βPath_Valid+γ×Tree_Balance, where Arc_Prob represents the dependency arc probability, Path_Valid represents the path validity, Tree_Balance represents the tree structure balance, and α, β, and γ are dynamic weight coefficients. In the scenario of fraudulent text message analysis, this multi-dimensional syntactic analysis method can effectively capture the structured features of the text.

[0089] This embodiment deeply optimizes the dependency parsing strategy. An adaptive transition arc mechanism is designed to address the unique grammatical structure of SMS text. A dynamic programming algorithm is used to accurately model the dependency relationships between words. Particular attention is paid to the identification of key semantic components; when a suspicious semantic pattern is detected, the system strengthens the weight of the corresponding dependency path. For example, when analyzing text with suggestive expressions, the structural features of the dependency tree can accurately capture the semantic connections between subject, verb, and object, which is crucial for understanding fraudulent intent. This intelligent syntactic analysis scheme provides structured feature support for subsequent scene recognition.

[0090] This embodiment innovatively implements a semantic structure extraction mechanism. Taking advantage of the hierarchical characteristics of syntactic trees, the system constructs a feature extraction framework based on depth-first traversal. By analyzing the hierarchical relationship between the head word and modifiers, a complete semantic representation is constructed. Particular attention is paid to the integrity of semantic chunks; specific traversal rules are designed to ensure that key semantic information is not fragmented. This hierarchical semantic extraction method effectively maintains the semantic integrity of the text.

[0091] This embodiment deeply optimizes the scene classification strategy. The system constructs a dynamic classification framework based on a predefined scene library, achieving intelligent selection of key features through an attention mechanism. Particular attention is paid to the timeliness of scene features, continuously updating the scene feature library through online learning to adapt to the evolution of new fraud methods. This comprehensive scene classification method provides a semantic basis for risk identification.

[0092] This embodiment achieves deep feature representation through nonlinear transformation. The system constructs a feature transformation framework based on a multilayer perceptron and combines it with a softmax function for probability distribution calculation. Particular attention is paid to the discriminative power of features; by optimizing the network structure and training strategy, the recognition accuracy for different scene categories is improved. This systematic feature processing scheme provides reliable support for scene classification.

[0093] This embodiment's innovative design not only solves the semantic understanding problem in traditional methods but also establishes a continuously optimizing scene recognition framework. Through multi-level semantic analysis and scene classification, the system can extract effective scene features from complex SMS text. This deep learning-based analysis mechanism ensures that the system maintains effective recognition capabilities when facing diverse fraud scenarios. In fraudulent SMS analysis, this intelligent semantic processing scheme significantly improves the accuracy of scene recognition.

[0094] This embodiment achieves an intelligent upgrade in identifying fraudulent text messages by establishing a complete semantic analysis chain. The system can dynamically adjust its analysis strategy based on real-time text features, avoiding the limitations of traditional fixed-rule solutions. Through multi-dimensional semantic analysis and scenario classification, the accuracy and reliability of identification are significantly improved, providing more comprehensive feature support for subsequent risk assessment. This intelligent analysis mechanism demonstrates strong adaptability and optimization effects in fraudulent text message identification scenarios.

[0095] This embodiment not only improves the accuracy of semantic understanding but also establishes an ever-evolving identification system through continuous strategy optimization and effect analysis. This optimization mechanism based on real-time feedback ensures that the system can continuously improve as fraud methods change, providing users with increasingly accurate protection. In practical applications, this self-optimization mechanism significantly improves the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud efforts.

[0096] In one embodiment of the NLP-based intelligent identification method for fraudulent text messages in this application, it may further include the following:

[0097] Step S401: Construct a bidirectional recurrent neural network, organize the scene feature vector into a feature sequence according to the time order. The bidirectional recurrent neural network contains two hidden states, forward and backward. The information flow is controlled by a gating mechanism. The feature sequence is scanned bidirectionally. Long short-term memory units are used to capture long-term dependencies in the sequence. Attention pooling layers are used to weight and aggregate the hidden states to generate temporally related features.

[0098] Step S402: Input the temporal correlation features into the conditional random field model. The conditional random field model calculates the conditional probability of the label sequence based on the state transition matrix, constructs a potential function to model the constraint relationship between adjacent labels, uses the Viterbi algorithm to search for the optimal path among all possible label paths, and decodes the optimal path to obtain the labeling result.

[0099] Optionally, this embodiment addresses the challenge of sequence labeling and tag prediction in fraudulent text messages by innovatively designing a sequence modeling scheme based on bidirectional recurrent neural networks and conditional random fields. This embodiment first constructs a bidirectional temporal modeling framework, capturing contextual semantic dependencies through forward and backward states. The system designs a sequence feature extraction formula: H_t = LSTM_f(X_t) + LSTM_b(X_t) + α × Attention(H_t), where LSTM_f and LSTM_b represent the outputs of the forward and backward LSTM units, respectively, X_t is the input feature at time t, α is the attention weight, and Attention is the attention calculation function. In the scenario of fraudulent text message analysis, this bidirectional modeling method can effectively capture long-term dependencies in the text.

[0100] This embodiment deeply optimizes the temporal feature extraction strategy. Targeting the sequential dependency characteristics of SMS text, a feature extraction mechanism based on Long Short-Term Memory (LSTM) units is designed. Through the synergistic effect of input gates, forget gates, and output gates, selective retention of long-term dependency information is achieved. Particular attention is paid to the contextual environment of key information; when a suspicious semantic pattern is detected, the system enhances the feature weights at the corresponding location through a gating mechanism. For example, when analyzing contact information with misleading descriptions, selective updates of the memory units can accurately capture the semantic connections between the preceding and following text, which is crucial for identifying the integrity of fraudulent information.

[0101] This embodiment innovatively implements an attention pooling mechanism. Addressing the differences in the importance of sequence features, the system constructs an attention-based feature enhancement framework. By calculating attention weights at different time steps, key information in the sequence is automatically identified. Particular attention is paid to modeling long-distance dependencies; a multi-head attention mechanism is designed to ensure that long-distance semantic associations can be effectively captured. This attention-based feature aggregation method can effectively extract key information patterns from sequences.

[0102] This embodiment deeply optimizes the conditional random field modeling strategy. The system constructs a sequence labeling framework based on state transitions and establishes global consistency of the labeling results by calculating the label transition probability matrix. Particular attention is paid to the rationality of the label sequences; feature functions and transition functions are designed to ensure that the labeling results conform to predefined grammatical rules. This complete sequence labeling method provides a reliable theoretical foundation for information extraction.

[0103] This embodiment achieves the finding of the optimal annotation path through Viterbi decoding. The system constructs a path search framework based on dynamic programming and combines state transition probabilities for path scoring. Special attention is paid to decoding efficiency, and a pruning strategy is used to reduce the search space for invalid paths. This systematic decoding scheme provides a reliable guarantee for the generation of annotation results.

[0104] This embodiment's innovative design not only solves the sequence modeling problem in traditional methods but also establishes a continuously optimizable annotation framework. Through multi-level feature extraction and sequence annotation, the system can identify key information from complex text sequences. This deep learning-based annotation mechanism ensures that the system maintains effective identification capabilities when faced with diverse fraudulent texts. In fraudulent SMS analysis, this intelligent sequence processing scheme significantly improves the accuracy of information extraction.

[0105] This embodiment achieves an intelligent upgrade in the analysis of fraudulent text messages by establishing a complete sequence processing chain. The system can dynamically adjust its processing strategy based on real-time text features, avoiding the limitations of traditional fixed-rule solutions. Through multi-dimensional feature extraction and sequence labeling, the accuracy and reliability of information identification are significantly improved, providing reliable feature support for subsequent risk assessment. This intelligent processing mechanism demonstrates strong adaptability and optimization effects in the analysis of fraudulent text messages.

[0106] This embodiment not only improves the accuracy of sequence labeling but also establishes a continuously evolving processing system through ongoing strategy optimization and effect analysis. This real-time feedback-based optimization mechanism ensures that the system can continuously improve as fraud methods evolve, providing increasingly accurate labeling results for subsequent analysis. In practical applications, this self-optimization mechanism significantly enhances the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud efforts.

[0107] In one embodiment of the NLP-based intelligent identification method for fraudulent text messages in this application, it may further include the following:

[0108] Step S501: Construct an information extraction module, extract phone number and URL fields from SMS text based on regular expression pattern matching, perform number segment location parsing and number format standardization on the phone number, use a URL parser to perform domain name segmentation and parameter extraction on the URL, and index the extracted number information and URL information according to their positional relationship in the text to generate a structured information object sequence;

[0109] Step S502: Input the information object sequence into the association rule analysis model. The association rule analysis model uses a sliding time window to calculate the co-occurrence frequency of numbers and URLs, scores the co-occurrence patterns based on the association degree calculation formula, and constructs a feature combination network. The feature combination network models the association strength of different feature pairs through a multi-head attention mechanism and organizes the association strength scores into an association feature matrix.

[0110] Optionally, this embodiment addresses the challenge of extracting and analyzing key information in fraudulent text messages by innovatively designing an information extraction and association analysis scheme. This embodiment first constructs an information extraction framework, achieving accurate identification of key information through multi-level pattern matching. The system designs a formula for calculating association strength: Association_Score=αCo_occurrence+βPosition_Relation+γ×Pattern_Similarity, where Co_occurrence represents co-occurrence frequency, Position_Relation represents positional relationship strength, Pattern_Similarity represents pattern similarity, and α, β, and γ are dynamic weighting coefficients. In fraudulent text message analysis scenarios, this multi-dimensional association analysis method can effectively identify suspicious information combinations.

[0111] This embodiment deeply optimizes the information extraction strategy. An adaptive regular expression template library is designed to handle modified expressions in SMS text. Through fuzzy matching and variant recognition, accurate extraction of spoofed phone numbers and modified URLs is achieved. Special attention is paid to protecting information integrity; when deliberate segmentation of key information is detected, the system initiates an information reconstruction mechanism. For example, when processing disguised expressions like "contact number 4**8**82", feature pattern matching can accurately reconstruct the complete number, which is crucial for subsequent association analysis.

[0112] This embodiment innovatively implements an information standardization mechanism. A standardized processing framework is constructed to address the diverse expressions of phone numbers and website addresses. Location parsing is performed using a phone number segment database, while structured parsing is conducted on website addresses. Particular attention is paid to information validity verification; multiple verification rules are designed to ensure the accuracy of extracted information. This rule-based information processing method effectively improves the reliability of subsequent analysis.

[0113] This embodiment deeply optimizes the association analysis strategy. The system constructs a dynamic co-occurrence analysis framework based on a sliding window, identifying potential associations by calculating information co-occurrence patterns within a time window. Particular attention is paid to temporal correlations; a time decay factor is designed to dynamically weight historical co-occurrence information. This comprehensive association analysis method provides reliable feature support for risk assessment.

[0114] This embodiment achieves accurate modeling of feature associations through multi-head attention. The system constructs a feature combination framework based on an attention mechanism, fusing different types of associated features. Particular attention is paid to the interaction relationships between features, capturing complex feature dependency patterns by establishing a multi-layer attention network. This systematic feature combination scheme provides a deep feature basis for association analysis.

[0115] This innovative design not only solves the information extraction problem in traditional methods but also establishes a continuously optimizing association analysis framework. Through multi-level information processing and association modeling, the system can identify effective risk features from complex SMS text. This deep learning-based analysis mechanism ensures that the system maintains effective identification capabilities when facing diverse fraud methods. In fraudulent SMS analysis, this intelligent information processing scheme significantly improves the accuracy of feature extraction.

[0116] This embodiment achieves an intelligent upgrade in the analysis of fraudulent text messages by establishing a complete information processing chain. The system can dynamically adjust its processing strategy based on real-time text features, avoiding the limitations of traditional fixed-rule solutions. Through multi-dimensional information extraction and correlation analysis, the accuracy and reliability of feature recognition are significantly improved, providing reliable feature support for subsequent risk assessment. This intelligent processing mechanism demonstrates strong adaptability and optimization effects in the analysis of fraudulent text messages.

[0117] This embodiment not only improves the accuracy of information extraction but also establishes a continuously evolving processing system through ongoing strategy optimization and effect analysis. This real-time feedback-based optimization mechanism ensures that the system can continuously improve as fraud methods change, providing increasingly accurate feature representations for subsequent analysis. In practical applications, this self-optimization mechanism significantly improves the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud efforts.

[0118] In one embodiment of the NLP-based intelligent identification method for fraudulent text messages in this application, it may further include the following:

[0119] Step S601: Construct a feature preprocessing module to normalize the associated feature matrix to eliminate dimensional differences, use principal component analysis to reduce the dimensionality of the features, input the dimensionality-reduced features into a residual neural network, the residual neural network retains the original feature information through a skip connection structure, performs local feature extraction based on convolutional layers, adjusts the feature distribution using batch normalization layers, and generates a residual feature map.

[0120] Step S602: Construct a feature fusion network. The feature fusion network adopts a multi-branch structure to process residual features at different scales in parallel. The importance weights of feature channels are calculated through the channel attention module. The spatial dependencies of features are captured based on the spatial attention module. The multi-branch features are fused by weighted summation. The fusion result is nonlinearly transformed to obtain the fused feature vector.

[0121] Optionally, this embodiment innovatively designs a deep learning-based feature processing scheme to address the problems of feature redundancy and information loss in feature fusion and representation learning of fraudulent SMS messages. This embodiment first constructs a feature preprocessing framework, achieving optimized representation of associated features through multi-level feature transformations. The system designs a feature importance evaluation formula: Feature_Score=αPCA_Weight+βResidual_Contrib+γ×Channel_Attention, where PCA_Weight represents the principal component weight, Residual_Contrib represents the residual contribution, Channel_Attention represents the channel attention score, and α, β, and γ are dynamic weight coefficients. In the scenario of fraudulent SMS message analysis, this multi-dimensional feature evaluation method can effectively extract key information.

[0122] This embodiment deeply optimizes the feature preprocessing strategy. An adaptive normalization mechanism is designed to address the issue of dimensional differences in the associated feature matrix. By calculating the statistical properties of the feature distribution, standardized representations of different types of features are achieved. Particular attention is paid to handling outliers; when a significant shift in the feature distribution is detected, the system initiates a robust normalization strategy. For example, when processing associated features at different time scales, adaptive normalization ensures feature comparability, which is crucial for subsequent feature fusion.

[0123] This embodiment innovatively implements a dimensionality reduction optimization mechanism. Addressing the high-dimensionality of the feature space, the system constructs a dimensionality reduction framework based on principal component analysis. By analyzing the feature covariance matrix, the main directions of variation are identified. Particular attention is paid to information retention rate; an adaptive dimensionality selection strategy is designed to retain key information to the greatest extent possible during dimensionality reduction. This statistically based dimensionality reduction method effectively reduces feature redundancy.

[0124] This embodiment deeply optimizes the residual learning strategy. The system constructs a feature extraction framework based on skip connections and preserves the original features through the residual structure. Particular attention is paid to the multi-scale representation of features, and a multi-layer convolutional network is designed to capture feature patterns of different granularities. This complete feature learning method provides a reliable foundation for subsequent fusion processing.

[0125] This embodiment achieves intelligent feature selection through an attention mechanism. The system constructs a feature enhancement framework based on channel and spatial attention, combining feature relationships from different dimensions for modeling. Particular attention is paid to feature dependencies; a multi-head attention network is established to capture complex feature interaction patterns. This systematic feature processing scheme provides deep feature support for risk identification.

[0126] This embodiment's innovative design not only solves the feature processing problem in traditional methods but also establishes a continuously optimizing feature fusion framework. Through multi-level feature transformation and attention enhancement, the system can extract effective risk features from a complex feature space. This deep learning-based processing mechanism ensures that the system maintains effective expressive power when faced with diverse fraud features. In the analysis of fraudulent text messages, this intelligent feature processing scheme significantly improves the accuracy of risk identification.

[0127] This embodiment achieves an intelligent upgrade in the analysis of fraudulent SMS messages by establishing a complete feature processing chain. The system can dynamically adjust its processing strategy based on real-time features, avoiding the limitations of traditional fixed feature schemes. Through multi-dimensional feature processing and fusion, the quality and reliability of feature representation are significantly improved, providing reliable feature support for subsequent risk assessment. This intelligent processing mechanism demonstrates strong adaptability and optimization effects in the analysis of fraudulent SMS messages.

[0128] This embodiment not only improves the accuracy of feature processing but also establishes a continuously evolving processing system through ongoing strategy optimization and effect analysis. This real-time feedback-based optimization mechanism ensures that the system can continuously improve as fraud methods change, providing increasingly accurate feature representations for subsequent analysis. In practical applications, this self-optimization mechanism significantly improves the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud efforts.

[0129] In one embodiment of the NLP-based intelligent identification method for fraudulent text messages in this application, it may further include the following:

[0130] Step S701: Construct a risk mapping network, input the fused feature vector into a multilayer perceptron, the multilayer perceptron transforms the features through hidden layer neurons, uses a batch normalization layer to standardize the intermediate features, introduces nonlinear characteristics based on the activation function, prevents gradient vanishing through the residual connection structure, and uses the sigmoid function to map the features to the risk probability space on the output of the last layer to generate a risk score result.

[0131] Step S702: Construct a graded processing module, compare the risk score results with multi-level preset thresholds, divide the risk level into three levels (high, medium, and low) based on threshold segmentation rules, configure differentiated handling strategies for different risk levels, encapsulate the risk level label and handling strategy information into handling instructions, and distribute the handling instructions to the fraud handling module according to priority through a message queue.

[0132] Optionally, this embodiment innovatively designs a risk mapping and hierarchical processing scheme based on deep learning to address issues such as inaccurate scoring and untimely handling in the risk assessment of fraudulent text messages. This embodiment first constructs a risk assessment framework, achieving a nonlinear mapping from features to the risk space through a multilayer perceptron. The system designs a risk scoring formula: Risk_Score=σ(W_n·f(BN(W_(n-1)·f(...))))+X, where W is the weight matrix, f is the activation function, BN is the batch normalization operation, σ is the sigmoid function, and X is the residual connection term. In the scenario of fraudulent text message analysis, this deep learning-based mapping method can effectively capture the complex correlations between features.

[0133] This embodiment deeply optimizes the feature transformation strategy. A multi-level feature transformation process is designed to address the high-dimensionality of fused features. Layer-by-layer abstraction of the feature space is achieved through nonlinear transformations of hidden layer neurons. Special attention is paid to feature standardization; when a shift in the intermediate feature distribution is detected, the system adaptively adjusts through a batch normalization layer. For example, when processing feature combinations from different sources, standardization ensures the consistency of feature distribution, which is crucial for maintaining the stability of risk assessment.

[0134] This embodiment innovatively implements a nonlinear feature extraction mechanism. Addressing the complex distribution characteristics of risk features, the system constructs a nonlinear transformation framework based on activation functions. By selecting appropriate activation functions, the model's ability to express nonlinear patterns is enhanced. Particular attention is paid to gradient propagation; the system ensures the training stability of deep networks through residual connection structures. This deep learning-based feature extraction method can effectively identify complex risk patterns.

[0135] This embodiment deeply optimizes the risk classification strategy. The system constructs a dynamic classification framework based on preset thresholds, achieving fine-grained risk level division through multi-level thresholds. Particular attention is paid to the real-time nature of the classification; a warning queue management mechanism is built to ensure that high-risk warnings are prioritized. This comprehensive classification method provides a reliable basis for the distribution of warning information.

[0136] This embodiment achieves precise risk management through differentiated handling. The system constructs a handling strategy framework based on risk levels and configures strategies according to different scenario characteristics. Particular attention is paid to the timeliness of handling, and a multi-level caching mechanism is designed to improve the response speed of handling commands. This systematic handling solution provides comprehensive technical support for the prevention and control of fraudulent text messages.

[0137] This embodiment's innovative design not only solves the risk assessment problem in traditional methods but also establishes a continuously optimized prevention and control framework. Through multi-level feature mapping and risk classification, the system can generate accurate risk assessment results from a complex feature space. This deep learning-based assessment mechanism ensures that the system maintains effective identification capabilities when facing diverse fraud methods. In the prevention and control of fraudulent text messages, this intelligent assessment scheme significantly improves the targeted nature of protection.

[0138] This embodiment achieves an intelligent upgrade in the prevention and control of fraudulent text messages by establishing a complete risk handling chain. The system can dynamically adjust its handling strategy based on real-time risk assessment, avoiding the limitations of traditional fixed-rule solutions. Through multi-dimensional risk classification and handling distribution, the accuracy and timeliness of prevention and control are significantly improved, providing users with more proactive fraud protection support. This intelligent prevention and control mechanism demonstrates strong adaptability and optimization effects in the protection against fraudulent text messages.

[0139] This embodiment not only improves the accuracy of risk assessment but also establishes a continuously evolving prevention and control system through ongoing strategy optimization and effect analysis. This optimization mechanism based on real-time feedback ensures that the system can continuously improve as fraud methods change, providing users with increasingly precise protection. In practical applications, this self-optimization mechanism significantly improves the system's long-term service quality and protective effectiveness, providing reliable technical support for anti-fraud efforts.

[0140] To effectively address the shortcomings of traditional technologies in feature fusion, sample processing, and identification strategies, and to significantly improve the intelligence level and prevention effectiveness of fraudulent SMS identification, this application provides an embodiment of an NLP-based intelligent SMS identification device for implementing all or part of the aforementioned NLP-based intelligent SMS identification method. See [link to embodiment]. Figure 2 The NLP-based intelligent identification device for fraudulent text messages specifically includes the following components:

[0141] The text recognition module 10 is used to collect SMS text data, perform word segmentation on the SMS text data to obtain word sequences, construct a text vectorization model, convert the word sequences into text feature vectors, perform syntactic dependency analysis on the text feature vectors to obtain semantic structure features, and combine the semantic structure features with a scene tag library to classify the scene and generate scene feature vectors.

[0142] The association analysis module 20 is used to construct a sequence labeling network, input the scene feature vector into a bidirectional recurrent neural network to extract temporal association features, use a conditional random field to label and model the temporal association features, identify the reserved number and website information in the SMS, calculate the association degree score between the number and the website based on the association rule analysis model, and generate an association feature matrix.

[0143] The risk warning module 30 is used to input the associated feature matrix into the risk assessment model. The risk assessment model uses a residual neural network structure to fuse the features, performs nonlinear mapping on the fused features through a multilayer perceptron, classifies the SMS messages based on a preset threshold, and outputs the classification results to the fraud handling module.

[0144] As described above, the NLP-based intelligent identification device for fraudulent text messages provided in this application can achieve accurate understanding of text message content through an innovative text analysis mechanism, semantic structure extraction, and scene classification. A sequence labeling-based feature extraction model is designed, combined with bidirectional recurrent neural networks and conditional random fields, to establish a number / website association analysis strategy for intelligent identification. A deep fusion evaluation mechanism is introduced, using residual neural networks and nonlinear mapping to achieve accurate risk level classification and timely handling. This method effectively solves the shortcomings of traditional technologies in text analysis, feature extraction, and risk assessment, significantly improving the accuracy and reliability of fraudulent text message identification.

[0145] From a hardware perspective, in order to effectively address the shortcomings of traditional technologies in feature fusion, sample processing, and recognition strategies, and significantly improve the intelligence level and prevention effect of fraudulent SMS identification, this application provides an embodiment of an electronic device for implementing all or part of the aforementioned NLP-based intelligent identification method for fraudulent SMS. The electronic device specifically includes the following components:

[0146] The system comprises a processor, memory, a communications interface, and a bus; wherein the processor, memory, and communications interface communicate with each other via the bus; the communications interface is used to realize information transmission between the NLP-based intelligent identification device for fraudulent SMS messages and core business systems, user terminals, and related databases and other related devices; the logic controller can be a desktop computer, tablet computer, or mobile terminal, etc., and this embodiment is not limited to these. In this embodiment, the logic controller can be implemented with reference to the embodiments of the NLP-based intelligent identification method for fraudulent SMS messages and the embodiments of the NLP-based intelligent identification device for fraudulent SMS messages in the embodiments, the content of which is incorporated herein, and repeated details will not be described again.

[0147] It is understood that the user terminal may include smartphones, tablet computers, network set-top boxes, portable computers, desktop computers, personal digital assistants (PDAs), in-vehicle devices, smart wearable devices, etc. Among these, the smart wearable devices may include smart glasses, smartwatches, smart bracelets, etc.

[0148] In practical applications, the NLP-based intelligent identification method for fraudulent SMS messages can be partially executed on the electronic device side as described above, or all operations can be completed on the client device. The choice can be made based on the processing power of the client device and the limitations of the user's usage scenario. This application does not impose any limitations on this. If all operations are completed on the client device, the client device may further include a processor.

[0149] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0150] Figure 3 This is a schematic block diagram illustrating the system configuration of the electronic device 9600 according to an embodiment of this application. Figure 3 As shown, the electronic device 9600 may include a central processing unit 9100 and a memory 9140; the memory 9140 is coupled to the central processing unit 9100. It is worth noting that... Figure 3 This is an example; other types of structures can also be used to supplement or replace this structure to achieve telecommunications functions or other functions.

[0151] In one embodiment, the NLP-based intelligent identification method for fraudulent text messages can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following control:

[0152] Step S101: Collect SMS text data, perform word segmentation on the SMS text data to obtain word sequences, construct a text vectorization model, convert the word sequences into text feature vectors, perform syntactic dependency analysis on the text feature vectors to obtain semantic structure features, combine the semantic structure features with a scene tag library to classify the scene, and generate scene feature vectors.

[0153] Step S102: Construct a sequence labeling network, input the scene feature vector into a bidirectional recurrent neural network to extract temporal association features, use a conditional random field to label and model the temporal association features, identify the reserved number and website information in the SMS, calculate the association degree score between the number and the website based on the association rule analysis model, and generate an association feature matrix;

[0154] Step S103: Input the associated feature matrix into the risk assessment model. The risk assessment model uses a residual neural network structure to fuse the features, performs nonlinear mapping on the fused features through a multilayer perceptron, classifies the SMS messages based on a preset threshold, and outputs the classification results to the fraud handling module.

[0155] As described above, the electronic device provided in this application innovatively constructs a text analysis mechanism to achieve accurate understanding of SMS content through semantic structure extraction and scene classification. It designs a feature extraction model based on sequence labeling, combining bidirectional recurrent neural networks and conditional random fields to establish a number / website association analysis strategy for intelligent identification. A deep fusion evaluation mechanism is introduced, using residual neural networks and nonlinear mapping to achieve accurate risk level classification and timely handling. This method effectively solves the shortcomings of traditional technologies in text analysis, feature extraction, and risk assessment, significantly improving the accuracy and reliability of identifying fraudulent SMS messages.

[0156] In another embodiment, the NLP-based intelligent identification device for fraudulent text messages can be configured separately from the central processing unit 9100. For example, the NLP-based intelligent identification device for fraudulent text messages can be configured as a chip connected to the central processing unit 9100, and the NLP-based intelligent identification method for fraudulent text messages can be implemented through the control of the central processing unit.

[0157] like Figure 3 As shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily need to include these components. Figure 3 All components shown; in addition, the electronic device 9600 may also include Figure 3 For components not shown, please refer to existing technologies.

[0158] like Figure 3 As shown, the central processing unit 9100, sometimes also referred to as a controller or operating control, may include a microprocessor or other processor device and / or logic device, which receives inputs and controls the operation of various components of the electronic device 9600.

[0159] The memory 9140 may be, for example, one or more of a cache, flash memory, hard drive, removable media, volatile memory, non-volatile memory, or other suitable devices. It may store the aforementioned failure-related information, and also store a program for executing that information. The central processing unit 9100 may execute the program stored in the memory 9140 to perform information storage or processing, etc.

[0160] Input unit 9120 provides input to central processing unit 9100. Input unit 9120 may be, for example, a keypad or touch input device. Power supply 9170 provides power to electronic device 9600. Display 9160 displays images and text. Display may be, for example, an LCD display, but is not limited thereto.

[0161] The memory 9140 can be a solid-state memory, such as a read-only memory (ROM), random access memory (RAM), a SIM card, etc. It can also be a memory that retains information even when power is off, can be selectively erased, and contains more data; examples of this type of memory are sometimes referred to as EPROMs. The memory 9140 can also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142 for storing application programs and function programs or processes for executing the operation of the electronic device 9600 via the central processing unit 9100.

[0162] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various drivers for the electronic device's communication functions and / or for performing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0163] The communication module 9110 is a transmitter / receiver that sends and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processing unit 9100 to provide input signals and receive output signals, which is the same as in a conventional mobile communication terminal.

[0164] Based on different communication technologies, multiple communication modules 9110 can be configured in the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless LAN modules. The communication module 9110 (transmitter / receiver) is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide audio output via the speaker 9131 and receive audio input from the microphone 9132, thereby realizing typical telecommunications functions. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. Additionally, the audio processor 9130 is coupled to a central processing unit 9100, enabling on-device recording via the microphone 9132 and on-device playback of stored audio via the speaker 9131.

[0165] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the NLP-based intelligent identification method for fraudulent SMS messages with a server or client as the execution subject in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the NLP-based intelligent identification method for fraudulent SMS messages with a server or client as the execution subject in the above embodiments. For example, when the processor executes the computer program, it implements the following steps:

[0166] Step S101: Collect SMS text data, perform word segmentation on the SMS text data to obtain word sequences, construct a text vectorization model, convert the word sequences into text feature vectors, perform syntactic dependency analysis on the text feature vectors to obtain semantic structure features, combine the semantic structure features with a scene tag library to classify the scene, and generate scene feature vectors.

[0167] Step S102: Construct a sequence labeling network, input the scene feature vector into a bidirectional recurrent neural network to extract temporal association features, use a conditional random field to label and model the temporal association features, identify the reserved number and website information in the SMS, calculate the association degree score between the number and the website based on the association rule analysis model, and generate an association feature matrix;

[0168] Step S103: Input the associated feature matrix into the risk assessment model. The risk assessment model uses a residual neural network structure to fuse the features, performs nonlinear mapping on the fused features through a multilayer perceptron, classifies the SMS messages based on a preset threshold, and outputs the classification results to the fraud handling module.

[0169] As described above, the computer-readable storage medium provided in this application, through an innovative text analysis mechanism, achieves accurate understanding of SMS content via semantic structure extraction and scene classification. A sequence-label-based feature extraction model is designed, combining bidirectional recurrent neural networks and conditional random fields to establish a number / website association analysis strategy for intelligent identification. A deep fusion evaluation mechanism is introduced, using residual neural networks and nonlinear mapping to achieve accurate risk level classification and timely handling. This method effectively addresses the shortcomings of traditional technologies in text analysis, feature extraction, and risk assessment, significantly improving the accuracy and reliability of fraudulent SMS identification.

[0170] Embodiments of this application also provide a computer program product capable of implementing all steps in the NLP-based intelligent identification method for fraudulent SMS messages, where the execution subject is a server or client, as described in the above embodiments. When executed by a processor, this computer program / instruction implements the steps of the NLP-based intelligent identification method for fraudulent SMS messages. For example, the computer program / instruction implements the following steps:

[0171] Step S101: Collect SMS text data, perform word segmentation on the SMS text data to obtain word sequences, construct a text vectorization model, convert the word sequences into text feature vectors, perform syntactic dependency analysis on the text feature vectors to obtain semantic structure features, combine the semantic structure features with a scene tag library to classify the scene, and generate scene feature vectors.

[0172] Step S102: Construct a sequence labeling network, input the scene feature vector into a bidirectional recurrent neural network to extract temporal association features, use a conditional random field to label and model the temporal association features, identify the reserved number and website information in the SMS, calculate the association degree score between the number and the website based on the association rule analysis model, and generate an association feature matrix;

[0173] Step S103: Input the associated feature matrix into the risk assessment model. The risk assessment model uses a residual neural network structure to fuse the features, performs nonlinear mapping on the fused features through a multilayer perceptron, classifies the SMS messages based on a preset threshold, and outputs the classification results to the fraud handling module.

[0174] As described above, the computer program product provided in this application innovatively constructs a text analysis mechanism to achieve accurate understanding of SMS content through semantic structure extraction and scene classification. It designs a feature extraction model based on sequence labeling, combining bidirectional recurrent neural networks and conditional random fields to establish a number / website association analysis strategy for intelligent identification. A deep fusion evaluation mechanism is introduced, using residual neural networks and nonlinear mapping to achieve accurate risk level classification and timely handling. This method effectively addresses the shortcomings of traditional technologies in text analysis, feature extraction, and risk assessment, significantly improving the accuracy and reliability of identifying fraudulent SMS messages.

[0175] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0176] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0177] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0178] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0179] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. An NLP-based intelligent identification method for fraud-related short messages, characterized in that, The method comprises: Collecting short message text data, performing word segmentation processing on the short message text data to obtain a word sequence, constructing a text vectorization model, converting the word sequence into a text feature vector, performing syntax dependency analysis on the text feature vector to obtain semantic structure features, classifying the semantic structure features according to a scene label library to generate a scene feature vector; Constructing a sequence labeling network, inputting the scene feature vector into a bidirectional recurrent neural network to extract timing correlation features, using a conditional random field to model and label the timing correlation features, identifying reserved numbers and website information in the short message, calculating the correlation degree score of the numbers and the website based on an association rule analysis model to generate an association feature matrix; Inputting the association feature matrix into a risk assessment model, the risk assessment model using a residual neural network structure to fuse the features, using a multilayer perceptron to perform nonlinear mapping on the fused features, classifying the short message based on a preset threshold, and outputting the classification result to a fraud handling module.

2. The NLP-based fraud message intelligent identification method according to claim 1, characterized in that, The collecting short message text data, performing word segmentation processing on the short message text data to obtain a word sequence, constructing a text vectorization model, converting the word sequence into a text feature vector, comprises: Receiving a short message text data stream from a short message monitoring platform, performing encoding format conversion and cleaning on the short message text data stream, removing special characters and redundant spaces, extracting text structured information based on a regular expression, dividing the cleaned text into sentences according to punctuation marks, performing word segmentation processing on the sentence division result using a maximum matching algorithm, optimizing the word segmentation result combined with a self-defined dictionary, and combining the optimized word segmentation sequence into a word sequence; Constructing a text vectorization model, the text vectorization model based on a bag-of-words model to count word frequency features, using a term frequency-inverse document frequency algorithm to calculate word weights, representing the word sequence as a sparse matrix, inputting the sparse matrix into a word embedding layer, the word embedding layer mapping words to low-dimensional dense vectors through a lookup table, and concatenating the mapping results to generate a text feature vector. 3.The NLP-based fraud message intelligent identification method according to claim 1, characterized in that, The text feature vector is analyzed by syntax dependency analysis to obtain semantic structure features, and the semantic structure features are classified according to a scene label library to generate a scene feature vector, which comprises: Constructing a syntax analysis network, inputting the text feature vector into a dependency syntax analyzer, the dependency syntax analyzer using a transition arc mechanism to construct a syntax tree, calculating the dependency relationship probability between words based on a dynamic programming algorithm, selecting the optimal dependency path to generate a syntax structure graph, performing depth-first traversal on the syntax structure graph, extracting the hierarchical relationship between center words and modifier words, and converting the hierarchical relationship information into semantic structure features; Loading predefined scene type information from a scene label library, constructing a scene classification model, the scene classification model using an attention mechanism to weight key components in the semantic structure features, performing nonlinear transformation on the weighted features based on a multilayer perceptron structure, calculating the scene class probability distribution through a softmax function, and concatenating the probability distribution result with the semantic structure features to obtain a scene feature vector.

4. The NLP-based fraud message intelligent identification method according to claim 1, characterized in that, The sequence labeling network is constructed, the scene feature vector is input into a bidirectional recurrent neural network to extract time correlation features, and a conditional random field is used to label and model the time correlation features, including: A bidirectional recurrent neural network is constructed, the scene feature vector is organized into a feature sequence in chronological order, the bidirectional recurrent neural network includes two hidden layer states in the forward and backward directions, information flow is controlled through a gating mechanism, the feature sequence is scanned in both directions, long-term dependencies in the sequence are captured based on a long short-term memory unit, hidden layer states are weighted and aggregated using an attention pooling layer, and time correlation features are generated; The time correlation features are input into a conditional random field model, the conditional random field model calculates the conditional probability of a label sequence based on a state transition matrix, a potential function is constructed to model the constraint relationship between adjacent labels, a Viterbi algorithm is used to search for an optimal path among all possible label paths, and the optimal path is decoded to obtain a labeling result. 5.The NLP-based fraud message intelligent identification method according to claim 1, characterized in that, The reserved number and website information in the short message are identified, the correlation degree score of the number and the website is calculated based on an association rule analysis model, and an association feature matrix is generated, including: An information extraction module is constructed, phone numbers and website fields are extracted from short message texts based on regular expression pattern matching, the phone numbers are parsed for number segment attribution and number format normalization, a URL parser is used to split domain names and extract parameters from the website, the extracted number information and website information are indexed according to their positional relationship in the text, and a structured information object sequence is generated; The information object sequence is input into an association rule analysis model, the association rule analysis model calculates the co-occurrence frequency of the number and the website using a sliding time window, scores the co-occurrence mode based on a correlation degree calculation formula, and constructs a feature combination network, the feature combination network models the association strength of different feature pairs through a multi-head attention mechanism, and organizes the association strength scores into an association feature matrix.

6. The NLP-based fraud message intelligent identification method according to claim 1, characterized in that, The association feature matrix is input into a risk assessment model, the risk assessment model uses a residual neural network structure to fuse features, including: A feature preprocessing module is constructed, the association feature matrix is normalized to eliminate dimensional differences, principal component analysis is used to reduce the dimensionality of the features, and the reduced features are input into a residual neural network, the residual neural network retains the original feature information through a skip connection structure, performs local feature extraction on the features based on a convolutional layer, adjusts the feature distribution using a batch normalization layer, and generates a residual feature mapping; A feature fusion network is constructed, the feature fusion network uses a multi-branch structure to process residual features of different scales in parallel, calculates the importance weights of feature channels through a channel attention module, captures the spatial dependency of features based on a spatial attention module, fuses the multi-branch features through weighted summation, and performs nonlinear transformation on the fusion result to obtain a fusion feature vector.

7. The NLP-based fraud message intelligent identification method according to claim 1, wherein, The fusion features are nonlinearly mapped by a multilayer perceptron, the short messages are risk graded based on a preset threshold, and the grading result is output to a fraud handling module, including: A risk mapping network is constructed, the fusion feature vector is input into a multilayer perceptron, the multilayer perceptron transforms features through hidden layer neurons, adopts a batch normalization layer to standardize intermediate features, introduces a nonlinear property based on an activation function, prevents gradient disappearance through a residual connection structure, uses a sigmoid function on the output of the last layer to map features to a risk probability space, and generates a risk score result; A hierarchical processing module is constructed, the risk score result is compared with a plurality of preset threshold values, risk levels are divided into three levels of high, medium and low based on threshold segmentation rules, different risk levels are configured with differentiated disposal strategies, a risk level label and disposal strategy information are encapsulated as disposal instructions, and the disposal instructions are distributed to a fraud disposal module according to priority through a message queue.

8. An NLP-based fraud message intelligent identification device, characterized in that, The device comprises: A text recognition module is configured to collect short message text data, perform word segmentation processing on the short message text data to obtain a word sequence, construct a text vectorization model, convert the word sequence into a text feature vector, perform syntax dependency analysis on the text feature vector to obtain semantic structure features, perform scene classification on the semantic structure features in combination with a scene label library, and generate a scene feature vector; An association analysis module is configured to construct a sequence labeling network, input the scene feature vector into a bidirectional recurrent neural network to extract time sequence association features, use a conditional random field to model and label the time sequence association features, identify reserved numbers and website information in the short message, calculate an association degree score of the numbers and the website information based on an association rule analysis model, and generate an association feature matrix; A risk early warning module is configured to input the association feature matrix into a risk assessment model, the risk assessment model uses a residual neural network structure to fuse features, performs nonlinear mapping on the fused features through a multilayer perceptron, classifies the short message based on a preset threshold value, and outputs a classification result to a fraud disposal module.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the NLP-based intelligent fraud short message identification method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the NLP-based intelligent fraud short message identification method of any one of claims 1 to 7.

Images