Fraud phone real-time identification method and device based on AI semantic understanding

Abstract

The application embodiment provides a fraud call real-time identification method and device based on AI semantic understanding, innovatively constructs a voice analysis mechanism, realizes accurate understanding of conversation content through grammar feature extraction and semantic role labeling, designs a scene discrimination model based on dialogue identification, combines semantic pattern matching and hierarchical clustering algorithm, establishes a fraud dialogue identification strategy for intelligent classification, introduces a residual fusion evaluation mechanism, realizes accurate evaluation and timely prevention and control of call risk through historical case feature fusion and risk scoring. The method effectively solves the shortcomings of traditional technologies in voice understanding, dialogue identification and risk assessment, and significantly improves the accuracy and reliability of fraud call identification.

Description

Technical Field

[0001] The present application relates to the field of data processing, and specifically to a method and device for real-time identification of fraudulent calls based on AI semantic understanding. Background Art

[0002] Existing fraud call identification methods have significant shortcomings. Traditional systems lack depth in semantic analysis and struggle to effectively integrate single-call identification with cross-user contextual analysis, impacting identification accuracy.

[0003] Furthermore, existing technologies face bottlenecks in monitoring and early warning. Most systems fail to continuously track reserved numbers and lack a feature-matching mechanism based on information components, resulting in inaccurate risk prevention and control.

[0004] The existing system has technical shortcomings in its handling mechanism. It lacks the ability to perform structured analysis of call content, making it difficult to implement a tiered approach based on identity information, cause description, and operational instructions, thus hindering its effectiveness. Addressing these issues is crucial for improving fraudulent call identification. Summary of the Invention

[0005] In response to the problems in the existing technology, this application provides a real-time identification method and device for fraudulent calls based on AI semantic understanding, which can effectively solve the shortcomings of traditional technologies in feature screening, speech recognition and protective measures, and significantly improve the intelligence level and protection effect of fraudulent call identification.

[0006] In order to solve at least one of the above problems, the present application provides the following technical solutions:

[0007] In the first aspect, the present application provides a method for real-time identification of fraudulent calls based on AI semantic understanding, comprising:

[0008] Perform speech-to-text processing on the call audio stream to generate conversation text. Use a word segmenter to segment and parse the syntactic structure of the conversation text, extract the dependencies between sentence components to obtain grammatical features, use a semantic annotation module to identify semantic roles and key entities in the sentence, and use a multi-head attention mechanism to perform contextual analysis on the grammatical features to generate a semantic representation vector.

[0009] Construct a speech recognition model, extract conversation intention features based on the semantic representation vector, extract typical semantic patterns from the fraud speech template library, calculate the similarity between the conversation intention features and the semantic patterns, classify the similarity results using a hierarchical clustering algorithm, input the classification results into a scene recognition network, and output a scene discrimination result based on a conditional random field structure;

[0010] The scenario discrimination result is input into a risk assessment model. The risk assessment model uses a residual connection structure to fuse the features of fraud cases in the historical annotated data to generate a risk feature vector. The risk feature vector is normalized to obtain a risk score. The calls are classified based on a preset risk threshold and prevention and control measures are triggered.

[0011] Furthermore, the method further includes: segmenting the call audio stream according to speech frames, inputting the segmented speech frame sequence into a speech enhancement network, wherein the speech enhancement network suppresses ambient noise using spectral subtraction, performs phoneme segmentation on the enhanced speech frame sequence, probabilistically decodes the phoneme sequence based on a hidden Markov model, maps the phoneme sequence into a corresponding text sequence using an acoustic model, and combines the text sequence into a conversation text according to punctuation mark positions;

[0012] The conversation text is input into a word segmenter, which segments the continuous text based on a maximum matching algorithm, performs part-of-speech tagging on the segmentation results to obtain a word sequence, and constructs a syntax tree using a dependency syntax analysis model. The dependency syntax analysis model calculates the dependency probability between sentence components through transfer arcs, selects the optimal dependency path according to the maximum spanning tree algorithm, and converts the path information into structured grammatical features.

[0013] Furthermore, the method further includes: inputting the grammatical features into a semantic annotation module, wherein the semantic annotation module uses a bidirectional long short-term memory network to perform sequence modeling on sentence components, annotates semantic roles based on a conditional random field, and constructs a named entity recognizer, wherein the named entity recognizer extracts text features through a character-level convolutional neural network, combines a dictionary matching method to identify key entities in the sentence, and combines the semantic role annotation results with the named entity recognition results to form a semantic label sequence;

[0014] A multi-head attention network is constructed for the semantic label sequence. The multi-head attention network maps the input sequence into a query vector, a key vector and a value vector, calculates the attention score by scaling the dot product, and obtains context features by weighted combination of the outputs of different attention heads. A feedforward neural network is used to perform nonlinear transformation on the context features, and a semantic representation vector is generated after layer normalization.

[0015] Furthermore, the method further includes: constructing an intent recognition network, inputting the semantic representation vector into a bidirectional gated recurrent unit, wherein the bidirectional gated recurrent unit controls information flow through an update gate and a reset gate to capture long-term dependencies in the sequence, employing an attention pooling layer to weightedly aggregate hidden layer states, and inputting the pooling result into a fully connected layer for dimensionality reduction mapping to obtain conversation intent features, wherein the conversation intent features include speech type and tone features;

[0016] The preset semantic pattern is loaded from the fraud speech template library, the semantic pattern is vectorized to represent the template feature matrix, the cosine similarity is used to calculate the distance score between the conversation intention feature and each vector in the template feature matrix, and a similarity calculation module is constructed. The similarity calculation module performs weighted fusion of the distance scores of features of different dimensions based on the attention mechanism and outputs the speech matching score.

[0017] Furthermore, the method further includes: constructing a hierarchical clustering model, calculating the Euclidean distance between samples based on the speech matching score, iteratively merging the clusters with the closest distance using the minimum distance method to generate a clustering tree structure, dynamically pruning the clustering tree structure to obtain speech clusters, calculating the inter-class distance and intra-class distance of the speech clusters, determining the optimal number of clusters based on the ratio of the inter-class distance to the intra-class distance, and converting the optimal clustering result into a speech type label;

[0018] The speech type label is input into the scene recognition network, and the scene recognition network uses a conditional random field structure to establish a state transfer matrix. The conditional probability distribution of the label sequence is calculated through a forward algorithm, and the optimal label path is searched based on the Viterbi algorithm. The optimal label path is decoded into a scene category, and the scene category is probability scored to obtain a scene discrimination result.

[0019] Furthermore, the method further includes: extracting fraud case features from historical annotated data, including voice features, text features, and behavioral features, constructing a feature processing network, wherein the feature processing network uses a multi-branch convolutional structure to map features of different dimensions, inputting the mapped features into a residual connection module, short-circuiting the features, retaining underlying feature information through skip connections, performing one-hot encoding on the scene discrimination results, and concatenating the encoded results with the residual features;

[0020] A feature fusion network is constructed. The feature fusion network uses a fully connected layer to reduce the dimension of the spliced ​​features, adjusts the feature distribution through a batch normalization layer, calculates the importance weights of different features using an attention mechanism, adaptively fuses the features based on the weight coefficients, and generates a risk feature vector through a nonlinear transformation of the fusion results.

[0021] Furthermore, the method further includes: constructing a risk scoring module, performing maximum and minimum value normalization processing on the risk feature vector, inputting the normalized features into a scoring network, the scoring network using a multi-layer perceptron structure to perform nonlinear mapping on the features, converting the mapping results into a probability distribution through a sigmoid activation function, synthesizing the multi-dimensional probability scores into a single risk score based on a weighted average method, and performing exponential smoothing processing on the risk score to eliminate short-term fluctuations;

[0022] Construct a risk grading module, compare the risk score with the preset multi-level risk threshold, divide the calls into different risk levels based on the threshold segmentation, set differentiated response strategies for different risk levels, input the response strategies into the task scheduler, and the task scheduler triggers the corresponding prevention and control instructions according to the priority configuration, and distributes the prevention and control instructions to the execution module based on the pipeline mechanism.

[0023] In a second aspect, the present application provides a real-time fraud call identification device based on AI semantic understanding, comprising:

[0024] A semantic processing module, which is used to perform speech-to-text processing on the call audio stream to obtain conversation text, perform word segmentation and syntactic structure analysis on the conversation text based on a word segmenter, extract the dependency relationships between sentence components to obtain grammatical features, identify semantic roles and key entities in the sentence through a semantic annotation module, and use a multi-head attention mechanism to perform contextual analysis on the grammatical features to generate a semantic representation vector;

[0025] A speech recognition module is used to build a speech recognition model, extract conversation intention features based on the semantic representation vector, extract typical semantic patterns from the fraud speech template library, calculate the similarity between the conversation intention features and the semantic patterns, classify the similarity results using a hierarchical clustering algorithm, input the classification results into a scene recognition network, and output a scene discrimination result based on a conditional random field structure;

[0026] The risk warning module is used to input the scenario discrimination results into the risk assessment model. The risk assessment model uses a residual connection structure to fuse the features of fraud cases in the historical annotated data to generate a risk feature vector. The risk feature vector is normalized to obtain a risk score. The call is classified based on the preset risk threshold and prevention and control measures are triggered.

[0027] In a third aspect, the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the method for real-time identification of fraudulent calls based on AI semantic understanding are implemented.

[0028] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for real-time identification of fraudulent calls based on AI semantic understanding.

[0029] In a fifth aspect, the present application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the method for real-time identification of fraudulent calls based on AI semantic understanding.

[0030] It can be seen from the above technical solution that this application provides a real-time identification method and device for fraudulent calls based on AI semantic understanding, which realizes accurate understanding of the content of the conversation by innovatively constructing a speech analysis mechanism, extracting grammatical features and labeling semantic roles. A scene discrimination model based on speech recognition is designed, and a fraud speech recognition strategy is established for intelligent classification by combining semantic pattern matching and hierarchical clustering algorithms. A residual fusion evaluation mechanism is introduced to achieve accurate assessment and timely prevention and control of call risks through historical case feature fusion and risk scoring. This method effectively solves the shortcomings of traditional technologies in speech understanding, speech recognition and risk assessment, and significantly improves the accuracy and reliability of fraudulent call identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0032] Figure 1 This is a flow chart of a method for real-time fraud call identification based on AI semantic understanding in an embodiment of the present application;

[0033] Figure 2 This is a structural diagram of a device for real-time fraudulent call identification based on AI semantic understanding in an embodiment of the present application;

[0034] Figure 3 Schematic diagram of the structure of the electronic device in the embodiment of the present application.

[0035] Reference numerals:

[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 program storage unit 9144, antenna 9111, speaker 9131, microphone 9132. DETAILED DESCRIPTION

[0037] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0038] The acquisition, storage, use, and processing of data in this application's technical solution comply with relevant national laws and regulations.

[0039] Taking into account the problems existing in the prior art, this application provides a real-time identification method and device for fraudulent calls based on AI semantic understanding. By innovatively constructing a speech analysis mechanism, grammatical feature extraction and semantic role labeling, accurate understanding of the content of the conversation is achieved. A scene discrimination model based on speech recognition is designed, and combined with semantic pattern matching and hierarchical clustering algorithms, a fraud speech recognition strategy is established for intelligent classification. A residual fusion evaluation mechanism is introduced to achieve accurate assessment and timely prevention and control of call risks through historical case feature fusion and risk scoring. This method effectively solves the shortcomings of traditional technologies in speech understanding, speech recognition and risk assessment, and significantly improves the accuracy and reliability of fraudulent call identification.

[0040] In order to effectively solve the shortcomings of traditional technologies in feature screening, speech recognition and protective measures, and significantly improve the intelligence level and protection effect of fraud call identification, this application provides an embodiment of a real-time fraud call identification method based on AI semantic understanding, see Figure 1 The real-time fraud call identification method based on AI semantic understanding specifically includes the following contents:

[0041] Step S101: Perform speech-to-text processing on the call audio stream to obtain a conversation text. Use a word segmenter to perform word segmentation and syntactic structure analysis on the conversation text, extract the dependency relationships between sentence components to obtain grammatical features, use a semantic annotation module to identify semantic roles and key entities in the sentence, and use a multi-head attention mechanism to perform contextual analysis on the grammatical features to generate a semantic representation vector.

[0042] Optionally, this embodiment addresses the issues of insufficient voice understanding accuracy and incomplete semantic feature extraction in fraud call identification by innovatively designing a speech semantic understanding solution based on deep learning. This embodiment first constructs a speech processing framework and implements a structured expression of call content through multi-level audio analysis. The system designs a semantic feature evaluation formula:

[0043] Semantic_Score=α Syntax_Weight+β Role_Weight+γ Entity_Weight+δ Context_Weight,

[0044] Syntax_Weight represents the weight of syntactic structure, Role_Weight represents the weight of semantic role, Entity_Weight represents the weight of entity importance, and Context_Weight represents the weight of contextual relevance. α, β, γ, and δ are dynamic adjustment coefficients. In fraud call identification scenarios, this multi-dimensional semantic evaluation method can effectively capture suspicious features in conversations.

[0045] This embodiment deeply optimizes the speech-to-text strategy. An adaptive audio enhancement mechanism is designed to address noise interference during phone calls. Through spectrum analysis and noise suppression, clear extraction of speech signals is achieved. Special attention is paid to the speaker's voice characteristics. When distortion of the speech signal is detected, the system initiates a signal reconstruction mechanism. For example, when processing call recordings in noisy environments, adaptive filtering can effectively improve the accuracy of speech recognition, which is crucial for subsequent text analysis.

[0046] This embodiment innovatively implements a word segmentation and syntactic analysis mechanism. To address the particularities of spoken language, the system constructs a word segmentation framework based on statistical learning. By combining dictionary matching and new word discovery, the accuracy of word segmentation is improved. Particular attention is paid to the processing of modal particles and catchphrases, and specific syntactic rules are designed to ensure the integrity of the analysis results. This rule-enhanced analysis method can effectively handle the special grammatical structures in spoken language.

[0047] This embodiment deeply optimizes the semantic role labeling strategy. The system builds a semantic role recognition framework based on conditional random fields and implements semantic analysis of sentence components through a sequence labeling model. It pays special attention to the logical relationships between roles and captures the deep semantic structure of sentences by constructing a semantic dependency graph. This comprehensive semantic analysis approach provides reliable feature support for intent recognition.

[0048] This embodiment achieves precise modeling of contextual features through a multi-head attention mechanism. The system builds a Transformer-based feature extraction framework, combining different attention heads to capture multi-scale contextual dependencies. It pays particular attention to long-range dependencies and, through the design of a positional encoding mechanism, ensures that long-range semantic associations are effectively captured. This systematic feature extraction approach provides deep semantic support for fraud detection.

[0049] The innovative design of this embodiment not only solves the speech comprehension issues of traditional methods but also establishes a continuously optimized semantic analysis framework. Through multi-level feature extraction and semantic modeling, the system is able to identify effective risk features from complex conversations. This deep learning-based analysis mechanism ensures that the system maintains effective understanding of diverse fraudulent tactics. In the identification of fraudulent calls, this intelligent semantic processing solution significantly improves recognition accuracy.

[0050] This embodiment achieves an intelligent upgrade for fraudulent call identification by establishing a complete semantic processing chain. The system dynamically adjusts processing strategies based on real-time voice features, avoiding the limitations of traditional fixed rule-based solutions. Through multi-dimensional semantic analysis and feature extraction, the accuracy and reliability of understanding are significantly improved, providing reliable feature support for subsequent risk assessment. This intelligent processing mechanism demonstrates strong adaptability and optimization effects in fraudulent call identification.

[0051] This embodiment not only improves the accuracy of semantic understanding but also establishes a continuously evolving recognition system through continuous strategy optimization and effectiveness analysis. This optimization mechanism, based on real-time feedback, ensures that the system can continuously improve as fraud methods evolve, providing users with increasingly precise protection. In practical applications, this self-optimization mechanism has significantly improved the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud efforts.

[0052] Step S102: Constructing a speech recognition model, extracting conversation intention features based on the semantic representation vector, extracting typical semantic patterns from a fraud speech template library, calculating the similarity between the conversation intention features and the semantic patterns, classifying the similarity results using a hierarchical clustering algorithm, inputting the classification results into a scene recognition network, and outputting a scene discrimination result based on a conditional random field structure;

[0053] Optionally, this embodiment innovatively designs a speech analysis solution based on deep learning to address the complex issues of speech recognition and scenario discrimination in fraudulent calls. This embodiment first constructs an intent recognition framework and achieves accurate modeling of fraudulent speech through multi-level feature extraction. The system designs a speech similarity evaluation formula:

[0054] Pattern_Score=α Intent_Sim+β Context_Rel+γ×Scene_Prob,

[0055] Intent_Sim represents intent similarity, Context_Rel represents context relevance, Scene_Prob represents scene probability, and α, β, and γ are dynamic weight coefficients. In fraud call identification scenarios, this multi-dimensional similarity assessment method can effectively capture the characteristics of the speech.

[0056] This embodiment deeply optimizes the intent feature extraction strategy. In response to the diverse nature of fraudulent rhetoric, a feature extraction mechanism based on semantic understanding was designed. Through in-depth analysis of semantic representation vectors, accurate identification of conversational intent is achieved. Special attention is paid to the coherence of the rhetoric. When suspicious semantic patterns are detected, the system increases the feature weights for the corresponding time period. For example, when analyzing fraudulent rhetoric such as "I am a public security, procuratorial, or judicial officer," intent feature extraction can accurately identify identity disguise features, which is of great significance for identifying fraudulent intent.

[0057] This embodiment innovatively implements a template matching mechanism. To address the evolving nature of fraudulent tactics, the system builds a dynamic template library update framework. By analyzing typical semantic patterns in historical cases, a comprehensive system of tactics templates is established. Special attention is paid to the timeliness of templates, and the template library is continuously updated through online learning to adapt to new fraudulent tactics. This template-based matching method effectively identifies variant tactics.

[0058] This embodiment deeply optimizes the clustering analysis strategy. The system builds a speech classification framework based on a hierarchical clustering algorithm. By calculating the semantic distance between samples, similar speech patterns are automatically categorized. Special attention is paid to clustering accuracy, and an adaptive distance metric is designed to improve the rationality of cluster division. This comprehensive clustering approach provides reliable feature support for scene recognition.

[0059] This embodiment achieves precise modeling of scene discrimination through conditional random fields. The system constructs a scene recognition framework based on sequence annotation and combines state transition probabilities to predict scene categories. It pays special attention to the continuity of scenes and captures the regularity of scene changes by establishing a state transition matrix. This systematic scene recognition solution provides a scenario-level basis for risk assessment.

[0060] The innovative design of this embodiment not only solves the problem of speech recognition in traditional methods but also establishes a continuously optimized scenario analysis framework. Through multi-level feature extraction and scenario modeling, the system is able to identify effective risk characteristics from complex conversation content. This deep learning-based analysis mechanism ensures that the system maintains effective recognition capabilities even when faced with a variety of fraudulent speech. In the prevention and control of fraudulent calls, this intelligent speech analysis solution significantly improves recognition accuracy.

[0061] This embodiment achieves an intelligent upgrade for fraudulent call identification by establishing a complete speech analysis chain. The system dynamically adjusts analysis strategies based on real-time conversation features, avoiding the limitations of traditional fixed-rule solutions. Through multi-dimensional speech analysis and scenario recognition, the accuracy and reliability of prevention and control are significantly improved, providing more comprehensive feature support for subsequent risk assessment. This intelligent analysis mechanism demonstrates strong adaptability and optimization effects in fraudulent call identification.

[0062] This embodiment not only improves the accuracy of speech recognition but also establishes a continuously evolving recognition system through continuous strategy optimization and effectiveness analysis. This optimization mechanism, based on real-time feedback, ensures that the system can continuously improve as fraud methods evolve, providing users with increasingly precise protection. In practical applications, this self-optimization mechanism has significantly improved the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud efforts.

[0063] Step S103: Input the scenario discrimination result into the risk assessment model. The risk assessment model uses a residual connection structure to fuse the fraud case features in the historical annotated data to generate a risk feature vector. The risk feature vector is normalized to obtain a risk score. The call is classified based on the preset risk threshold and prevention and control measures are triggered.

[0064] Optionally, this embodiment innovatively designs a risk assessment solution based on deep learning to address issues such as insufficient feature fusion and inaccurate risk warnings in fraudulent call risk assessment. This embodiment first constructs a risk feature fusion framework to achieve a comprehensive analysis of risk factors through multi-level feature integration. The system designs a risk assessment formula:

[0065] Risk_Score=α Scene_Weight+β History_Pattern+γ×Feature_Fusion,

[0066] Here, Scene_Weight represents the scene weight, History_Pattern indicates the historical pattern matching degree, Feature_Fusion represents the feature fusion score, and α, β, and γ are dynamic adjustment coefficients. In the fraud call identification scenario, this multi-dimensional risk assessment method can effectively predict fraud risk.

[0067] This embodiment deeply optimizes the feature fusion strategy. In response to the diversity of fraud features, a feature processing mechanism based on a residual network is designed. The jump connection structure ensures that the original feature information is effectively retained in the deep network. Special attention is paid to the temporal nature of features. When a sudden change in risk features is detected, the system increases the feature weights of the corresponding time period. For example, when processing risk features in consecutive calls, the original feature information is retained through short-circuit connections in the residual blocks, while high-order feature patterns are extracted through the deep network. This dual feature extraction mechanism can more accurately portray the risk evolution process.

[0068] This embodiment innovatively implements a risk scoring mechanism. To address the complex distribution of risk characteristics, the system constructs a normalized scoring framework. Through feature normalization, a unified measurement of risk factors across different dimensions is achieved. Special attention is paid to the interpretability of the scores, and a tiered scoring strategy is designed to ensure that risk assessment results have clear explanations. This normalized scoring approach effectively quantifies fraud risk.

[0069] This embodiment deeply optimizes the risk classification strategy. The system builds a dynamic grading framework based on preset thresholds, using an adaptive threshold adjustment mechanism to balance coverage across different risk levels. Special attention is paid to the real-time nature of classification, and a warning queue management mechanism is established to ensure that high-risk warnings are prioritized. This comprehensive classification approach provides a reliable basis for the dissemination of warning information.

[0070] This embodiment achieves precise control over risk management through prevention and control measures. The system establishes a risk-based response strategy framework, configuring policies based on the characteristics of different scenarios. Special attention is paid to the timeliness of response, and a multi-level caching mechanism is designed to improve the response speed of prevention and control instructions. This systematic response solution provides comprehensive technical support for the prevention and control of fraudulent calls.

[0071] The innovative design of this embodiment not only solves the risk assessment issues in traditional methods but also establishes a sustainable and optimized prevention and control framework. Through multi-level feature fusion and risk grading, the system is able to generate accurate risk assessment results from complex call characteristics. This deep learning-based assessment mechanism ensures that the system maintains effective identification capabilities when faced with diverse fraud methods. In the prevention and control of fraudulent calls, this intelligent assessment solution significantly improves the targeted protection.

[0072] This embodiment achieves an upgraded intelligent prevention and control system for fraudulent calls by establishing a complete risk management chain. The system dynamically adjusts its response strategy based on real-time risk assessment, avoiding the limitations of traditional fixed-rule solutions. Through multi-dimensional risk grading and response distribution, the accuracy and timeliness of prevention and control are significantly improved, providing users with more proactive fraud protection. This intelligent prevention and control mechanism demonstrates strong adaptability and optimization effects in fraudulent call protection.

[0073] This implementation not only improves the accuracy of risk assessment but also establishes a continuously evolving prevention and control system through continuous strategy optimization and effectiveness analysis. This optimization mechanism, based on real-time feedback, ensures that the system can continuously improve as fraud methods evolve, providing users with increasingly precise protection. In practical applications, this self-optimization mechanism has significantly improved the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud efforts.

[0074] In one embodiment of the method for real-time identification of fraudulent calls based on AI semantic understanding of the present application, the following contents may also be specifically included:

[0075] Step S201: Segment the call audio stream into speech frames, input the segmented speech frame sequence into a speech enhancement network, and use spectral subtraction to suppress ambient noise. The speech enhancement network then segments the enhanced speech frame sequence into phonemes, performs probabilistic decoding on the phoneme sequence based on a hidden Markov model, maps the phoneme sequence into a corresponding text sequence using an acoustic model, and combines the text sequence into a conversation text based on punctuation mark positions.

[0076] Step S202: The conversation text is input into a word segmenter. The word segmenter segments the continuous text based on a maximum matching algorithm, performs part-of-speech tagging on the segmentation results to obtain word sequences, and constructs a syntax tree using a dependency syntax analysis model. The dependency syntax analysis model calculates dependency probabilities between sentence components through transition arcs, selects an optimal dependency path based on a maximum spanning tree algorithm, and converts the path information into structured grammatical features.

[0077] Optionally, this embodiment innovatively designs a speech processing and text analysis solution based on deep learning to address the problems of noise interference and inaccurate semantic understanding in the voice understanding of fraudulent calls. This embodiment first constructs a speech processing framework to achieve clear extraction of call content through multi-level signal analysis. The system designs a speech quality assessment formula:

[0078] Voice_Score=α SNR+β Clarity+γ×Continuity,

[0079] Here, SNR represents signal-to-noise ratio, Clarity represents clarity, Continuity represents continuity, and α, β, and γ are dynamic weighting coefficients. In fraud call identification scenarios, this multi-dimensional voice evaluation method can effectively improve voice processing quality.

[0080] This embodiment deeply optimizes the speech segmentation strategy. An adaptive frame length adjustment mechanism is designed to address the continuity of speech in telephone calls. Dynamic frame length selection enables precise segmentation of speech signals. Special attention is paid to the recognition of speech boundaries. When a transition in the speech signal is detected, the system activates the boundary optimization mechanism. For example, when processing multi-person conversation scenarios, accurate speech segmentation can effectively distinguish between speech segments of different speakers, which is crucial for subsequent speech recognition. This intelligent segmentation solution provides a reliable processing unit for speech enhancement.

[0081] This embodiment innovatively implements a noise suppression mechanism. The system builds a noise cancellation framework based on spectral subtraction to address various types of noise present in telephone calls. By estimating the noise spectrum in real time, it accurately suppresses non-speech components. With particular attention to the fidelity of the speech signal, an adaptive suppression threshold is designed to ensure maximum preservation of speech details while eliminating noise. This spectrum-based noise reduction method effectively improves speech clarity.

[0082] This embodiment deeply optimizes the phoneme recognition strategy. The system constructs a phoneme decoding framework based on a hidden Markov model, modeling the temporal dependencies of phoneme sequences through state transition probabilities. It pays special attention to the continuity of phonemes and uses the Viterbi algorithm to search for the optimal state sequence, achieving accurate decoding of speech content. This comprehensive decoding method provides a reliable phoneme foundation for text conversion.

[0083] This example achieves precise speech-to-text mapping through acoustic modeling. The system builds a deep learning-based acoustic model framework that incorporates contextual information to convert phonemes into text. It pays special attention to text coherence and, by establishing language model constraints, improves the rationality of text sequences. This systematic conversion solution provides high-quality input for subsequent text analysis.

[0084] The innovative design of this embodiment not only solves the speech processing issues of traditional methods but also establishes a continuously optimized semantic analysis framework. Through multi-layered speech processing and text analysis, the system is able to extract clear semantic information from noisy call environments. This deep learning-based processing mechanism ensures that the system maintains effective comprehension capabilities even in complex call environments. In the identification of fraudulent calls, this intelligent speech processing solution significantly improves the accuracy of semantic understanding.

[0085] This embodiment achieves intelligent analysis upgrades for fraudulent calls by establishing a complete voice processing chain. The system can dynamically adjust processing strategies based on real-time voice features, avoiding the limitations of traditional fixed rule-based solutions. Through multi-dimensional voice enhancement and text conversion, the accuracy and reliability of semantic understanding are significantly improved, providing a reliable semantic foundation for subsequent intent recognition. This intelligent processing mechanism demonstrates strong adaptability and optimization effects in fraudulent call analysis.

[0086] This embodiment not only improves the accuracy of speech understanding but also establishes a continuously evolving processing system through continuous strategy optimization and effect analysis. This optimization mechanism based on real-time feedback ensures that the system can continuously improve as the call environment changes, providing increasingly accurate semantic input for subsequent analysis. In actual application, this self-optimization mechanism significantly improves the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud work.

[0087] In one embodiment of the method for real-time identification of fraudulent calls based on AI semantic understanding of the present application, the following contents may also be specifically included:

[0088] Step S301: Inputting the grammatical features into a semantic annotation module, the semantic annotation module uses a bidirectional long short-term memory network to perform sequence modeling on sentence components, annotates semantic roles based on conditional random fields, and constructs a named entity recognizer. The named entity recognizer extracts text features through a character-level convolutional neural network, combines a dictionary matching method to identify key entities in the sentence, and combines the semantic role annotation results with the named entity recognition results to form a semantic label sequence;

[0089] Step S302: Construct a multi-head attention network for the semantic tag sequence, which maps the input sequence into a query vector, a key vector, and a value vector. The attention score is calculated by scaling the dot product, and the outputs of different attention heads are weightedly combined to obtain context features. A feedforward neural network is used to perform nonlinear transformation on the context features, and a semantic representation vector is generated after layer normalization.

[0090] Optionally, this embodiment innovatively designs a set of semantic annotation and representation learning solutions based on deep learning to address the problems of insufficient semantic understanding and ambiguous context association in fraudulent calls. This embodiment first constructs a semantic analysis framework and achieves a deep understanding of the conversation content through multi-level feature extraction. The system designs a semantic understanding evaluation formula:

[0091] Semantic_Score=α Role_Weight+β Entity_Weight+γ Context_Weight+δ Attention_Weight,

[0092] Here, Role_Weight represents the semantic role weight, Entity_Weight represents the entity importance weight, Context_Weight represents the contextual relevance weight, Attention_Weight represents the attention score, and α, β, γ, and δ are dynamic adjustment coefficients. In fraud call identification scenarios, this multi-dimensional semantic evaluation method can effectively capture suspicious features in conversations.

[0093] This embodiment deeply optimizes the sequence modeling strategy. In view of the semantic coherence characteristics in fraudulent speech, a sequence processing mechanism based on a bidirectional long short-term memory network is designed. Through the bidirectional information flow of forward and backward states, the complete capture of contextual semantics is achieved. Special attention is paid to long-distance dependencies. When cross-sentence associations of key information are detected, the system will retain the information through memory units. For example, when analyzing fraudulent speech such as "You are involved in a money laundering case" that appears at a distance after "I am a public security, procuratorial and judicial staff", the semantic association between the two can be accurately captured through bidirectional modeling, which is crucial for identifying complete fraudulent intentions.

[0094] This embodiment innovatively implements a semantic role labeling mechanism. To address the diverse roles found in fraudulent conversations, the system constructs a labeling framework based on conditional random fields. Using a sequence labeling model, the semantic roles of sentence components are identified. Particular attention is paid to the coherence of role transitions, and a state transition matrix is ​​designed to ensure the overall consistency of the role labeling results. This probabilistic labeling method effectively identifies various semantic roles in fraudulent conversations.

[0095] This embodiment deeply optimizes the entity recognition strategy. The system builds an entity extraction framework based on a character-level convolutional neural network, capturing local features of text through multi-layer convolution operations. It pays special attention to variant entity expressions and improves the robustness of entity recognition by combining dictionary matching and deep learning methods. This comprehensive entity recognition approach provides entity-level support for semantic understanding.

[0096] This embodiment achieves precise modeling of contextual features through a multi-head attention mechanism. The system constructs an attention calculation framework based on scaled dot products, combining multiple attention heads to capture semantic dependencies across different dimensions. It pays special attention to distinguishing the importance of features, and through the design of an attention score calculation mechanism, it automatically focuses on key information. This systematic attention scheme provides deep feature support for semantic representation.

[0097] The innovative design of this embodiment not only solves the semantic understanding issues of traditional methods but also establishes a continuously optimized feature learning framework. Through multi-level semantic annotation and attention modeling, the system is able to extract effective semantic features from complex conversation content. This deep learning-based analysis mechanism ensures that the system maintains effective understanding capabilities even when faced with diverse fraudulent tactics. In fraud call identification, this intelligent semantic processing solution significantly improves the accuracy of understanding.

[0098] This embodiment achieves an intelligent upgrade for fraudulent call analysis by establishing a complete semantic processing chain. The system can dynamically adjust processing strategies based on real-time conversation features, avoiding the limitations of traditional fixed rule-based solutions. Through multi-dimensional semantic analysis and feature extraction, the accuracy and reliability of understanding are significantly improved, providing reliable semantic support for subsequent intent recognition. This intelligent processing mechanism demonstrates strong adaptability and optimization effects in fraudulent call analysis.

[0099] This embodiment not only improves the accuracy of semantic understanding but also establishes a continuously evolving processing system through continuous strategy optimization and effectiveness analysis. This optimization mechanism, based on real-time feedback, ensures that the system can continuously improve as fraudulent tactics evolve, providing increasingly accurate semantic representations for subsequent analysis. In practical applications, this self-optimization mechanism has significantly improved the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud efforts.

[0100] In one embodiment of the method for real-time identification of fraudulent calls based on AI semantic understanding of the present application, the following contents may also be specifically included:

[0101] Step S401: Construct an intent recognition network and input the semantic representation vector into a bidirectional gated recurrent unit. The bidirectional gated recurrent unit controls the flow of information through update gates and reset gates to capture long-term dependencies in the sequence. An attention pooling layer is used to perform weighted aggregation on the hidden layer states. The pooling result is input into a fully connected layer for dimensionality reduction mapping to obtain dialogue intent features. The dialogue intent features include speech type and tone features.

[0102] Step S402: Load a preset semantic pattern from the fraud speech template library, vectorize the semantic pattern to obtain a template feature matrix, use cosine similarity to calculate the distance score between the conversation intention feature and each vector in the template feature matrix, and construct a similarity calculation module. The similarity calculation module performs weighted fusion on the distance scores of features of different dimensions based on the attention mechanism, and outputs a speech matching score.

[0103] Optionally, this embodiment innovatively designs a deep learning-based intent analysis solution to address issues such as insufficient accuracy in intent recognition and unsatisfactory speech matching in fraudulent calls. This embodiment first constructs an intent recognition framework and accurately captures fraudulent intent through multi-level feature extraction. The system designs an intent matching evaluation formula:

[0104] Intent_Score=α Semantic_Sim+β Pattern_Match+γ×Tone_Weight,

[0105] Semantic_Sim represents semantic similarity, Pattern_Match represents pattern matching, Tone_Weight represents tone weight, and α, β, and γ are dynamic weight coefficients. In fraud call detection scenarios, this multi-dimensional intent assessment method can effectively identify suspicious speech.

[0106] This embodiment deeply optimizes the sequence processing strategy. In view of the temporal dependency characteristics in fraudulent language, an information processing mechanism based on a bidirectional gated recurrent unit is designed. Through the synergistic effect of the update gate and the reset gate, the selective retention of long-term dependent information is achieved. Special attention is paid to the control of information flow. When key information is detected, the system will enhance the feature weight of the corresponding position through the gating mechanism. For example, when analyzing fraudulent language such as "Your bank account has an abnormality" and "You need to cooperate with the investigation" that appears later, the gating mechanism can accurately capture the logical connection between the language, which is crucial for identifying the complete fraudulent intention.

[0107] This embodiment innovatively implements an attention pooling mechanism. The system constructs an attention-based feature aggregation framework based on the importance differences of hidden states. By calculating attention weights at different time steps, key information in the sequence is automatically identified. Special attention is paid to the aggregation method of features, and an adaptive pooling strategy is designed to ensure that important information is fully preserved. This attention-based feature aggregation method can effectively extract the core intent features in the sequence.

[0108] This embodiment deeply optimizes the template matching strategy. The system builds a dynamic matching framework based on a preset template library, achieving a structured representation of semantic patterns through vectorized representation. Special attention is paid to the expressive power of templates, and by designing diverse template features, the matching coverage is improved. This comprehensive template matching approach provides a reliable reference benchmark for speech recognition.

[0109] This embodiment achieves accurate evaluation of speech matching through similarity calculation. The system constructs a distance calculation framework based on cosine similarity and incorporates an attention mechanism to weight the importance of feature dimensions. It pays special attention to differences between dimensions and establishes a multi-dimensional similarity assessment system to capture different aspects of speech characteristics. This systematic similarity calculation scheme provides a quantitative evaluation basis for speech recognition.

[0110] The innovative design of this embodiment not only solves the intent recognition problem in traditional methods but also establishes a continuously optimized speech analysis framework. Through multi-level feature extraction and pattern matching, the system is able to identify effective fraud characteristics from complex conversation content. This deep learning-based analysis mechanism ensures that the system maintains effective recognition capabilities even when faced with a variety of fraudulent speech. In the identification of fraudulent calls, this intelligent intent analysis solution significantly improves recognition accuracy.

[0111] This embodiment achieves an intelligent upgrade for fraudulent call identification by establishing a complete intent analysis chain. The system dynamically adjusts analysis strategies based on real-time conversation features, avoiding the limitations of traditional fixed rule-based solutions. Through multi-dimensional intent extraction and speech matching, the accuracy and reliability of identification are significantly improved, providing reliable feature support for subsequent risk assessment. This intelligent analysis mechanism demonstrates strong adaptability and optimization effects in fraudulent call identification.

[0112] This embodiment not only improves the accuracy of intent recognition but also establishes a continuously evolving recognition system through continuous strategy optimization and effectiveness analysis. This optimization mechanism, based on real-time feedback, ensures that the system can continuously improve as fraudulent tactics evolve, providing users with increasingly precise protection. In practical applications, this self-optimization mechanism has significantly improved the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud efforts.

[0113] In one embodiment of the method for real-time identification of fraudulent calls based on AI semantic understanding of the present application, the following contents may also be specifically included:

[0114] Step S501: Construct a hierarchical clustering model, calculate the Euclidean distance between samples based on the speech matching score, iteratively merge the clusters with the closest distance using the minimum distance method to generate a cluster tree structure, dynamically prune the cluster tree structure to obtain speech clusters, calculate the inter-class distance and intra-class distance of the speech clusters, determine the optimal number of clusters based on the ratio of the inter-class distance to the intra-class distance, and convert the optimal clustering result into a speech type label;

[0115] Step S502: Input the speech type label into the scene recognition network. The scene recognition network uses a conditional random field structure to establish a state transfer matrix, calculates the conditional probability distribution of the label sequence through a forward algorithm, searches for the optimal label path based on the Viterbi algorithm, decodes the optimal label path into a scene category, and performs probability scoring on the scene category to obtain a scene discrimination result.

[0116] Optionally, this embodiment addresses the problems of inaccurate clustering of speech types and unsatisfactory scene recognition in fraudulent calls by innovatively designing a set of clustering and scene recognition solutions based on deep learning. This embodiment first constructs a cluster analysis framework to achieve accurate classification of speech types through multi-level feature aggregation. The system designs a cluster evaluation formula:

[0117] Cluster_Score=α×(Inter_Dist / Intra_Dist)+β Purity+γ Coherence,

[0118] Here, Inter_Dist represents the inter-class distance, Intra_Dist represents the intra-class distance, Purity represents cluster purity, Coherence represents cluster consistency, and α, β, and γ are dynamic weight coefficients. In the case of fraudulent call identification, this multi-dimensional clustering evaluation method can effectively identify speech patterns.

[0119] This embodiment deeply optimizes the distance calculation strategy. To address the diversity of speech features, a sample similarity calculation mechanism based on Euclidean distance was designed. By considering geometric distances in feature space, accurate measurement of speech samples is achieved. Special attention is paid to feature standardization. When uneven feature distribution is detected, the system activates an adaptive normalization mechanism. For example, when analyzing scam speech related to "online loans" and "financial management," distance calculation can accurately identify similar patterns, which is crucial for speech clustering.

[0120] This embodiment innovatively implements a hierarchical clustering mechanism. Targeting the hierarchical nature of speech types, the system constructs a clustering framework based on the minimum distance method. Through an iterative merging strategy, similar speech types are automatically aggregated. Special attention is paid to the hierarchical nature of clustering, and a dynamic pruning strategy is designed to ensure the rationality of the clustering results. This hierarchical clustering method effectively identifies the type structure of speech types.

[0121] This embodiment deeply optimizes the scene recognition strategy. The system constructs a scene classification framework based on conditional random fields, modeling the temporal dependencies of scene sequences through state transition matrices. It pays special attention to the continuity of scenes, calculating the conditional probability distribution through a forward algorithm to accurately model scene changes. This comprehensive scene recognition approach provides a reliable scenario-based basis for risk assessment.

[0122] This embodiment achieves accurate prediction of scene labels through Viterbi decoding. The system constructs a path search framework based on dynamic programming, incorporating state transition probabilities to select the optimal path. Special attention is paid to the coherence of the label sequence, and a probabilistic scoring mechanism is established to improve the reliability of scene identification. This systematic scene recognition solution provides scenario-level support for subsequent risk assessment.

[0123] The innovative design of this embodiment not only solves the problem of speech clustering in traditional methods, but also establishes a continuously optimized scene recognition framework. Through multi-level feature clustering and scene modeling, the system is able to identify effective scene features from complex conversation content. This deep learning-based analysis mechanism ensures that the system maintains effective recognition capabilities when faced with diverse fraud scenarios. In the prevention and control of fraudulent calls, this intelligent scene analysis solution significantly improves recognition accuracy.

[0124] This embodiment achieves an intelligent upgrade for fraudulent call identification by establishing a complete scenario analysis chain. The system dynamically adjusts analysis strategies based on real-time conversation features, avoiding the limitations of traditional fixed rule-based solutions. Through multi-dimensional speech clustering and scenario recognition, the accuracy and reliability of prevention and control are significantly improved, providing more comprehensive feature support for subsequent risk assessment. This intelligent analysis mechanism demonstrates strong adaptability and optimization effects in fraudulent call identification.

[0125] This embodiment not only improves the accuracy of scene recognition but also establishes a continuously evolving recognition system through continuous strategy optimization and effectiveness analysis. This optimization mechanism, based on real-time feedback, ensures that the system can continuously improve as fraud methods evolve, providing users with increasingly precise protection. In practical applications, this self-optimization mechanism has significantly improved the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud efforts.

[0126] In one embodiment of the method for real-time identification of fraudulent calls based on AI semantic understanding of the present application, the following contents may also be specifically included:

[0127] Step S601: Extract fraud case features from historical annotated data, including voice features, text features, and behavioral features, and construct a feature processing network. The feature processing network uses a multi-branch convolutional structure to map features of different dimensions. The mapped features are input into a residual connection module, which short-circuits the features and retains the underlying feature information through skip connections. The scene discrimination results are one-hot encoded and the encoded results are concatenated with the residual features.

[0128] Step S602: Construct a feature fusion network, which uses a fully connected layer to reduce the dimension of the spliced ​​features, adjusts the feature distribution through a batch normalization layer, uses an attention mechanism to calculate the importance weights of different features, adaptively fuses the features based on the weight coefficients, and generates a risk feature vector through a nonlinear transformation of the fusion results.

[0129] Optionally, this embodiment innovatively designs a feature processing solution based on deep learning to address the problems of insufficient feature utilization and unsatisfactory fusion effects in fraud call risk assessment. This embodiment first constructs a feature extraction framework to fully capture fraud features through multi-level feature analysis. The system designs a feature fusion evaluation formula:

[0130] Feature_Score=α Voice_Weight+β Text_Weight+γ Behavior_Weight+δ Scene_Weight,

[0131] Here, Voice_Weight represents the weight of voice features, Text_Weight represents the weight of text features, Behavior_Weight represents the weight of behavioral features, and Scene_Weight represents the weight of scene features. α, β, γ, and δ are dynamic adjustment coefficients. In fraud call identification scenarios, this multi-dimensional feature assessment method can effectively extract risk characteristics.

[0132] This embodiment deeply optimizes the historical data processing strategy. A comprehensive feature extraction mechanism is designed for the multimodal features in historical cases. By analyzing the interrelationships between voice, text, and behavioral features, a complete portrayal of the fraud pattern is achieved. Special attention is paid to the temporal nature of features. When evolution of feature patterns is detected, the system updates the feature extraction strategy through an adaptive learning mechanism. For example, when analyzing fraud cases involving "impersonating public security, procuratorial, and judicial officials," multidimensional feature analysis can accurately capture key features such as voice disguise, textual induction, and behavioral coercion, which is crucial for building a complete risk assessment model.

[0133] This embodiment innovatively implements a multi-branch feature processing mechanism. To address the heterogeneity of features across different dimensions, the system constructs a convolution-based feature mapping framework. By designing specialized convolution branches, each type of feature can be independently processed. Special attention is paid to the expressive power of features, and the effectiveness of feature mapping is ensured by optimizing the convolution structure. This multi-branch feature processing method fully preserves the unique information of features across each dimension.

[0134] This embodiment deeply optimizes the residual connection strategy. The system builds a feature transfer framework based on short-circuit connections, preserving original feature information through a skip connection mechanism. Special attention is paid to the integrity of the information flow, and by designing a multi-layer residual structure, it ensures that deep networks can effectively learn detailed feature representations. This comprehensive residual learning approach provides a reliable foundation for feature fusion.

[0135] This example achieves precise modeling of risk characteristics through feature fusion. The system builds an attention-based feature fusion framework, combining batch normalization and nonlinear transformations for feature integration. It pays special attention to distinguishing feature importance and establishes an adaptive weight calculation mechanism to highlight key features. This systematic fusion solution provides comprehensive feature support for risk assessment.

[0136] The innovative design of this embodiment not only solves the feature processing issues of traditional methods but also establishes a continuously optimized feature learning framework. Through multi-level feature extraction and fusion, the system is able to extract effective risk features from complex historical data. This deep learning-based analysis mechanism ensures that the system maintains effective feature expression capabilities when facing diverse fraud methods. In the identification of fraudulent calls, this intelligent feature processing solution significantly improves the accuracy of risk assessment.

[0137] This embodiment achieves an intelligent upgrade for fraudulent call analysis by establishing a complete feature processing chain. The system dynamically adjusts processing strategies based on real-time features, avoiding the limitations of traditional fixed-feature solutions. Through multi-dimensional feature extraction and fusion, the quality and reliability of feature expression are significantly improved, providing reliable feature support for subsequent risk assessment. This intelligent processing mechanism demonstrates strong adaptability and optimization effects in fraudulent call analysis.

[0138] This embodiment not only improves the accuracy of feature processing but also establishes a continuously evolving processing system through continuous strategy optimization and effectiveness analysis. This optimization mechanism based on real-time feedback ensures that the system can continuously improve as fraud methods evolve, providing increasingly accurate feature representations for subsequent analysis. In practical applications, this self-optimization mechanism has significantly improved the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud efforts.

[0139] In one embodiment of the method for real-time identification of fraudulent calls based on AI semantic understanding of the present application, the following contents may also be specifically included:

[0140] Step S701: Construct a risk scoring module, perform maximum and minimum normalization processing on the risk feature vector, input the normalized features into the scoring network, the scoring network uses a multi-layer perceptron structure to perform nonlinear mapping on the features, converts the mapping results into a probability distribution through a sigmoid activation function, and synthesizes the multi-dimensional probability scores into a single risk score based on a weighted average method. The risk score is then subjected to exponential smoothing to eliminate short-term fluctuations.

[0141] Step S702: Construct a risk grading module, compare the risk score with the preset multi-level risk threshold, divide the calls into different risk levels based on the threshold segmentation, set differentiated response strategies for different risk levels, input the response strategies into the task scheduler, and the task scheduler triggers the corresponding prevention and control instructions according to the priority configuration, and distributes the prevention and control instructions to the execution module based on the pipeline mechanism.

[0142] Optionally, this embodiment innovatively designs a risk scoring and grading solution based on deep learning to address issues such as inaccurate risk assessment and untimely prevention and control measures in fraudulent calls. This embodiment first constructs a risk assessment framework and achieves accurate determination of risk levels through multi-level feature processing. The system designs a risk assessment formula:

[0143] Risk_Score=α Base_Score+β Trend_Weight+γ×Exp_Smooth,

[0144] Base_Score represents the base risk score, Trend_Weight represents the trend weight, Exp_Smooth represents the exponential smoothing value, and α, β, and γ are dynamic weight coefficients. This multi-dimensional risk assessment method can effectively predict fraud risk in fraud call identification scenarios.

[0145] This embodiment deeply optimizes the feature normalization strategy. A normalization mechanism based on maximum and minimum values ​​is designed to address the scale differences of risk features. Through feature normalization, a unified measurement of features across different dimensions is achieved. Special attention is paid to the handling of outliers. When a significant shift in the feature distribution is detected, the system initiates a robust normalization strategy. For example, when dealing with different types of fraud features, normalization ensures feature comparability, which is crucial for accurate risk assessment.

[0146] This embodiment innovatively implements a risk scoring mechanism. To address the complex distributional characteristics of risk features, the system constructs a scoring framework based on a multi-layer perceptron. Through nonlinear mapping, it achieves in-depth modeling of risk patterns. Special attention is paid to the interpretability of the scores, and a hierarchical scoring strategy is designed to ensure that risk assessment results have clear explanations. This deep learning-based scoring method can effectively quantify fraud risk.

[0147] This example deeply optimizes the smoothing strategy. The system builds a dynamic smoothing framework based on exponential smoothing, effectively suppressing short-term fluctuations through time-series weighting. It pays special attention to trend preservation, balancing short-term responsiveness with long-term stability through the design of adaptive smoothing parameters. This comprehensive smoothing approach provides a stable numerical foundation for risk assessment.

[0148] This embodiment achieves precise risk management through tiered processing. The system builds a tiered framework based on multi-level thresholds, configuring policies based on the characteristics of different scenarios. Special attention is paid to timely handling, and a priority mechanism is designed to improve response speed for high-risk cases. This systematic tiered approach provides comprehensive technical support for fraud prevention and control.

[0149] The innovative design of this embodiment not only solves the risk assessment issues inherent in traditional methods but also establishes a continuously optimized prevention and control framework. Through multi-level risk scoring and tiered processing, the system is able to generate accurate risk assessment results from complex call characteristics. This deep learning-based assessment mechanism ensures that the system maintains effective identification capabilities despite diverse fraudulent methods. This intelligent assessment solution significantly enhances the targeted nature of fraudulent call prevention and control.

[0150] This embodiment achieves an upgraded intelligent prevention and control system for fraudulent calls by establishing a complete risk management chain. The system dynamically adjusts its response strategy based on real-time risk assessment, avoiding the limitations of traditional fixed-rule solutions. Through multi-dimensional risk grading and response distribution, the accuracy and timeliness of prevention and control are significantly improved, providing users with more proactive fraud protection. This intelligent prevention and control mechanism demonstrates strong adaptability and optimization effects in fraudulent call protection.

[0151] This implementation not only improves the accuracy of risk assessment but also establishes a continuously evolving prevention and control system through continuous strategy optimization and effectiveness analysis. This optimization mechanism, based on real-time feedback, ensures that the system can continuously improve as fraud methods evolve, providing users with increasingly precise protection. In practical applications, this self-optimization mechanism has significantly improved the system's long-term service quality and protection effectiveness, providing reliable technical support for anti-fraud efforts.

[0152] In order to effectively solve the shortcomings of traditional technologies in feature screening, speech recognition and protective measures, and significantly improve the intelligence level and protective effect of fraud call identification, this application provides an embodiment of a fraud call real-time identification device based on AI semantic understanding for realizing all or part of the content of the fraud call real-time identification method based on AI semantic understanding, see Figure 2 The real-time fraud call identification device based on AI semantic understanding specifically includes the following contents:

[0153] The semantic processing module 10 is configured to perform speech-to-text processing on the call audio stream to obtain a conversation text, perform word segmentation and syntactic structure analysis on the conversation text based on a word segmenter, extract dependencies between sentence components to obtain grammatical features, identify semantic roles and key entities in the sentence through a semantic annotation module, and perform contextual analysis on the grammatical features using a multi-head attention mechanism to generate a semantic representation vector;

[0154] The speech recognition module 20 is used to build a speech recognition model, extract conversation intention features based on the semantic representation vector, extract typical semantic patterns from the fraud speech template library, calculate the similarity between the conversation intention features and the semantic patterns, classify the similarity results using a hierarchical clustering algorithm, input the classification results into the scene recognition network, and output the scene discrimination results based on the conditional random field structure;

[0155] The risk warning module 30 is used to input the scenario discrimination results into the risk assessment model. The risk assessment model uses a residual connection structure to fuse the features of fraud cases in the historical annotation data to generate a risk feature vector, normalize the risk feature vector to obtain a risk score, classify the calls based on the preset risk threshold, and trigger prevention and control measures.

[0156] From the above description, it can be seen that the real-time fraud call identification device based on AI semantic understanding provided by the embodiment of the present application can achieve accurate understanding of the content of the conversation through innovative construction of a speech analysis mechanism, grammatical feature extraction and semantic role labeling. A scene discrimination model based on speech recognition is designed, and combined with semantic pattern matching and hierarchical clustering algorithms, a fraud speech recognition strategy is established for intelligent classification. A residual fusion evaluation mechanism is introduced to achieve accurate assessment and timely prevention and control of call risks through historical case feature fusion and risk scoring. This method effectively solves the shortcomings of traditional technologies in speech understanding, speech recognition and risk assessment, and significantly improves the accuracy and reliability of fraud call identification.

[0157] From a hardware perspective, in order to effectively address the deficiencies of traditional technologies in feature screening, speech recognition, and protective measures, and to significantly improve the intelligence level and protective effect of fraudulent call identification, this application provides an embodiment of an electronic device for implementing all or part of the content of the real-time fraudulent call identification method based on AI semantic understanding. The electronic device specifically includes the following content:

[0158] Processor (processor), memory (memory), communication interface (CommunicationsInterface) and bus; wherein, the processor, memory, and communication interface complete communication with each other through the bus; the communication interface is used to realize information transmission between the real-time identification device for fraudulent calls based on AI semantic understanding and related equipment such as core business systems, user terminals and related databases; the logic controller can be a desktop computer, tablet computer and mobile terminal, etc., but this embodiment is not limited to this. In this embodiment, the logic controller can be implemented with reference to the embodiment of the real-time identification method for fraudulent calls based on AI semantic understanding and the embodiment of the real-time identification device for fraudulent calls based on AI semantic understanding in the embodiment, and the contents thereof are incorporated herein, and the repeated parts are not repeated.

[0159] It is understandable that the user terminal may include a smart phone, a tablet electronic device, a network set-top box, a portable computer, a desktop computer, a personal digital assistant (PDA), a vehicle-mounted device, a smart wearable device, etc. Among them, the smart wearable device may include smart glasses, a smart watch, a smart bracelet, etc.

[0160] In practical applications, part of the real-time fraud call identification method based on AI semantic understanding can be executed on the electronic device side as described above, or all operations can be completed in the client device. The specific selection can be based on the processing power of the client device and the limitations of the user's usage scenario. This application does not limit this. If all operations are completed in the client device, the client device may also include a processor.

[0161] The aforementioned client device may include a communication module (i.e., a communication unit) capable of establishing a communication connection with a remote server to facilitate data transmission with the server. The server may include a server at the task scheduling center or, in other implementation scenarios, a server on an intermediate platform, such as a server on a third-party server platform that is communicatively linked to the task scheduling center server. The server may comprise a single computer device, a server cluster consisting of multiple servers, or a distributed server configuration.

[0162] Figure 3Schematic block diagram of the system structure of the electronic device 9600 according to an embodiment of the present 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 the Figure 3 is exemplary; other types of structures may also be used to supplement or replace this structure to implement telecommunication functions or other functions.

[0163] In one embodiment, the real-time fraud call identification method based on AI semantic understanding can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following control:

[0164] Step S101: Perform speech-to-text processing on the call audio stream to obtain a conversation text. Use a word segmenter to perform word segmentation and syntactic structure analysis on the conversation text, extract the dependency relationships between sentence components to obtain grammatical features, use a semantic annotation module to identify semantic roles and key entities in the sentence, and use a multi-head attention mechanism to perform contextual analysis on the grammatical features to generate a semantic representation vector.

[0165] Step S102: Constructing a speech recognition model, extracting conversation intention features based on the semantic representation vector, extracting typical semantic patterns from a fraud speech template library, calculating the similarity between the conversation intention features and the semantic patterns, classifying the similarity results using a hierarchical clustering algorithm, inputting the classification results into a scene recognition network, and outputting a scene discrimination result based on a conditional random field structure;

[0166] Step S103: Input the scenario discrimination result into the risk assessment model. The risk assessment model uses a residual connection structure to fuse the fraud case features in the historical annotated data to generate a risk feature vector. The risk feature vector is normalized to obtain a risk score. The call is classified based on the preset risk threshold and prevention and control measures are triggered.

[0167] As can be seen from the above description, the electronic device provided in the embodiment of the present application realizes accurate understanding of the content of the conversation by innovatively constructing a speech analysis mechanism, extracting grammatical features and labeling semantic roles. A scenario discrimination model based on speech recognition is designed, and a fraud speech recognition strategy is established for intelligent classification by combining semantic pattern matching and hierarchical clustering algorithms. A residual fusion evaluation mechanism is introduced to achieve accurate assessment and timely prevention and control of call risks through historical case feature fusion and risk scoring. This method effectively solves the shortcomings of traditional technologies in speech understanding, speech recognition and risk assessment, and significantly improves the accuracy and reliability of fraud call identification.

[0168] In another embodiment, the real-time fraud call identification device based on AI semantic understanding can be configured separately from the central processing unit 9100. For example, the real-time fraud call identification device based on AI semantic understanding can be configured as a chip connected to the central processing unit 9100, and the function of the real-time fraud call identification method based on AI semantic understanding can be realized through the control of the central processing unit.

[0169] 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 have to include Figure 3 In addition, the electronic device 9600 may also include all components shown in Figure 3 For components not shown, reference may be made to the prior art.

[0170] like Figure 3 As shown, the central processing unit 9100 is sometimes also referred to as a controller or operation control, and may include a microprocessor or other processor device and / or logic device. The central processing unit 9100 receives input and controls the operation of various components of the electronic device 9600.

[0171] Memory 9140 can 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 can store the aforementioned failure-related information and also store programs that execute the relevant information. The CPU 9100 can execute the programs stored in memory 9140 to implement information storage or processing.

[0172] The input unit 9120 provides input to the central processing unit 9100. The input unit 9120 may be, for example, a keypad or touch input device. The power supply 9170 is used to provide power to the electronic device 9600. The display 9160 is used to display objects such as images and text. The display may be, for example, an LCD display, but is not limited thereto.

[0173] The memory 9140 may be a solid-state memory, such as a read-only memory (ROM), random access memory (RAM), or SIM card. Alternatively, it may be a memory that retains information even when power is off, can be selectively erased, and is capable of storing additional data. Examples of such memory are sometimes referred to as EPROMs. The memory 9140 may 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 used by the central processing unit 9100 to execute operations of the electronic device 9600.

[0174] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, images, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various driver programs for communication functions of the electronic device and / or for executing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0175] The communication module 9110 is a transmitter / receiver that transmits and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processor 9100 to provide input signals and receive output signals, which may be the same as the case of a conventional mobile communication terminal.

[0176] Based on different communication technologies, multiple communication modules 9110 may be provided in the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless local area network modules. The communication module 9110 (transmitter / receiver) is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130, providing audio output via the speaker 9131 and receiving audio input from the microphone 9132, thereby implementing common telecommunication functions. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. Furthermore, the audio processor 9130 is coupled to the central processing unit 9100, enabling local recording via the microphone 9132 and playback of stored audio via the speaker 9131.

[0177] The embodiments of the present application also provide a computer-readable storage medium capable of implementing all steps of the method for real-time identification of fraudulent calls based on AI semantic understanding, where the execution subject is a server or a client in the above embodiments. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements all steps of the method for real-time identification of fraudulent calls based on AI semantic understanding, where the execution subject is a server or a client in the above embodiments. For example, when the processor executes the computer program, the following steps are implemented:

[0178] Step S101: Perform speech-to-text processing on the call audio stream to obtain a conversation text. Use a word segmenter to perform word segmentation and syntactic structure analysis on the conversation text, extract the dependency relationships between sentence components to obtain grammatical features, use a semantic annotation module to identify semantic roles and key entities in the sentence, and use a multi-head attention mechanism to perform contextual analysis on the grammatical features to generate a semantic representation vector.

[0179] Step S102: Constructing a speech recognition model, extracting conversation intention features based on the semantic representation vector, extracting typical semantic patterns from a fraud speech template library, calculating the similarity between the conversation intention features and the semantic patterns, classifying the similarity results using a hierarchical clustering algorithm, inputting the classification results into a scene recognition network, and outputting a scene discrimination result based on a conditional random field structure;

[0180] Step S103: Input the scenario discrimination result into the risk assessment model. The risk assessment model uses a residual connection structure to fuse the fraud case features in the historical annotated data to generate a risk feature vector. The risk feature vector is normalized to obtain a risk score. The call is classified based on the preset risk threshold and prevention and control measures are triggered.

[0181] As can be seen from the above description, the computer-readable storage medium provided in the embodiment of the present application realizes accurate understanding of the content of the conversation by innovatively constructing a speech analysis mechanism, extracting grammatical features and labeling semantic roles. A scenario discrimination model based on speech recognition is designed, and a fraud speech recognition strategy is established for intelligent classification by combining semantic pattern matching and hierarchical clustering algorithms. A residual fusion evaluation mechanism is introduced to achieve accurate assessment and timely prevention and control of call risks through historical case feature fusion and risk scoring. This method effectively solves the shortcomings of traditional technologies in speech understanding, speech recognition and risk assessment, and significantly improves the accuracy and reliability of fraud call identification.

[0182] The embodiments of the present application also provide a computer program product capable of implementing all steps of the method for real-time identification of fraudulent calls based on AI semantic understanding, where the execution subject is a server or a client in the above-mentioned embodiment. When the computer program / instructions are executed by a processor, the steps of the method for real-time identification of fraudulent calls based on AI semantic understanding are implemented. For example, the computer program / instructions implement the following steps:

[0183] Step S101: Perform speech-to-text processing on the call audio stream to obtain a conversation text. Use a word segmenter to perform word segmentation and syntactic structure analysis on the conversation text, extract the dependency relationships between sentence components to obtain grammatical features, use a semantic annotation module to identify semantic roles and key entities in the sentence, and use a multi-head attention mechanism to perform contextual analysis on the grammatical features to generate a semantic representation vector.

[0184] Step S102: Constructing a speech recognition model, extracting conversation intention features based on the semantic representation vector, extracting typical semantic patterns from a fraud speech template library, calculating the similarity between the conversation intention features and the semantic patterns, classifying the similarity results using a hierarchical clustering algorithm, inputting the classification results into a scene recognition network, and outputting a scene discrimination result based on a conditional random field structure;

[0185] Step S103: Input the scenario discrimination result into the risk assessment model. The risk assessment model uses a residual connection structure to fuse the fraud case features in the historical annotated data to generate a risk feature vector. The risk feature vector is normalized to obtain a risk score. The call is classified based on the preset risk threshold and prevention and control measures are triggered.

[0186] As can be seen from the above description, the computer program product provided in the embodiment of the present application realizes accurate understanding of the content of the conversation by innovatively constructing a speech analysis mechanism, extracting grammatical features and labeling semantic roles. A scenario discrimination model based on speech recognition is designed, and a fraud speech recognition strategy is established for intelligent classification by combining semantic pattern matching and hierarchical clustering algorithms. A residual fusion evaluation mechanism is introduced to achieve accurate assessment and timely prevention and control of call risks through historical case feature fusion and risk scoring. This method effectively solves the shortcomings of traditional technologies in speech understanding, speech recognition and risk assessment, and significantly improves the accuracy and reliability of fraud call identification.

[0187] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, apparatuses, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0188] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (apparatus), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts 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, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0189] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0190] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0191] Specific embodiments are used in the present invention to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. A real-time fraud call identification method based on AI semantic understanding, characterized in that: The method comprises: Perform speech-to-text processing on the call audio stream to generate conversation text. Use a word segmenter to segment and parse the syntactic structure of the conversation text, extract the dependencies between sentence components to obtain grammatical features, use a semantic annotation module to identify semantic roles and key entities in the sentence, and use a multi-head attention mechanism to perform contextual analysis on the grammatical features to generate a semantic representation vector. Construct a speech recognition model, extract conversation intention features based on the semantic representation vector, extract typical semantic patterns from the fraud speech template library, calculate the similarity between the conversation intention features and the semantic patterns, classify the similarity results using a hierarchical clustering algorithm, input the classification results into a scene recognition network, and output a scene discrimination result based on a conditional random field structure; Fraud case features are extracted from historical annotated data, including voice features, text features and behavioral features, and a feature processing network is constructed. The feature processing network uses a multi-branch convolution structure to map features of different dimensions, and the mapped features are input into the residual connection module. The residual connection module short-circuits the features, retains the underlying feature information through jump connections, performs unique hot encoding on the scene discrimination results, and splices the encoding results with the residual features; a feature fusion network is constructed, which uses a fully connected layer to reduce the dimension of the spliced ​​features, adjusts the feature distribution through a batch normalization layer, uses an attention mechanism to calculate the importance weights of different features, adaptively fuses the features based on the weight coefficients, and generates a risk feature vector through a nonlinear transformation of the fusion results; constructs a risk A scoring module performs maximum and minimum value normalization on the risk feature vector, and inputs the normalized features into a scoring network. The scoring network uses a multi-layer perceptron structure to perform nonlinear mapping on the features, and converts the mapping results into a probability distribution through a sigmoid activation function. The multi-dimensional probability scores are synthesized into a single risk score based on a weighted average method, and the risk scores are subjected to exponential smoothing to eliminate short-term fluctuations. A risk grading module is constructed to compare the risk scores with preset multi-level risk thresholds, and calls are divided into different risk levels based on threshold segmentation. Differentiated response strategies are set for different risk levels, and the response strategies are input into a task scheduler. The task scheduler triggers corresponding prevention and control instructions according to the priority configuration, and distributes the prevention and control instructions to the execution module based on a pipeline mechanism.

2. The method for real-time identification of fraudulent calls based on AI semantic understanding according to claim 1 is characterized in that: The speech-to-text processing of the call audio stream is performed to obtain a conversation text, and the conversation text is segmented and parsed into syntactic structures based on a word segmenter to extract dependency relationships between sentence components to obtain grammatical features, including: The call audio stream is segmented into speech frames, and the segmented speech frame sequence is input into a speech enhancement network. The speech enhancement network uses spectral subtraction to suppress ambient noise, performs phoneme segmentation on the enhanced speech frame sequence, performs probabilistic decoding on the phoneme sequence based on a hidden Markov model, and uses an acoustic model to map the phoneme sequence into a corresponding text sequence. The text sequence is then combined into a conversation text according to the position of punctuation marks. The conversation text is input into a word segmenter, which segments the continuous text based on a maximum matching algorithm, performs part-of-speech tagging on the segmentation results to obtain a word sequence, and constructs a syntax tree using a dependency syntax analysis model. The dependency syntax analysis model calculates the dependency probability between sentence components through transfer arcs, selects the optimal dependency path according to the maximum spanning tree algorithm, and converts the path information into structured grammatical features.

3. The method for real-time identification of fraudulent calls based on AI semantic understanding according to claim 1 is characterized in that: The semantic tagging module identifies semantic roles and key entities in a sentence, uses a multi-head attention mechanism to perform contextual analysis on the grammatical features, and generates a semantic representation vector, including: Inputting the grammatical features into a semantic annotation module, the semantic annotation module uses a bidirectional long short-term memory network to perform sequence modeling on sentence components, annotates semantic roles based on conditional random fields, and constructs a named entity recognizer. The named entity recognizer extracts text features through a character-level convolutional neural network, combines a dictionary matching method to identify key entities in the sentence, and combines the semantic role annotation results with the named entity recognition results to form a semantic label sequence; A multi-head attention network is constructed for the semantic label sequence. The multi-head attention network maps the input sequence into a query vector, a key vector and a value vector, calculates the attention score by scaling the dot product, and obtains context features by weighted combination of the outputs of different attention heads. A feedforward neural network is used to perform nonlinear transformation on the context features, and a semantic representation vector is generated after layer normalization.

4. The method for real-time identification of fraudulent calls based on AI semantic understanding according to claim 1 is characterized in that: The method of constructing a speech recognition model, extracting conversation intention features based on the semantic representation vector, extracting typical semantic patterns from a fraud speech template library, and calculating the similarity between the conversation intention features and the semantic patterns includes: Construct an intent recognition network and input the semantic representation vector into a bidirectional gated recurrent unit. This bidirectional gated recurrent unit controls information flow through update and reset gates to capture long-term dependencies in the sequence. An attention pooling layer is used to perform weighted aggregation on the hidden layer states. The pooling result is input into a fully connected layer for dimensionality reduction mapping to obtain conversation intent features. These features include speech type and tone characteristics. The preset semantic pattern is loaded from the fraud speech template library, the semantic pattern is vectorized to represent the template feature matrix, the cosine similarity is used to calculate the distance score between the conversation intention feature and each vector in the template feature matrix, and a similarity calculation module is constructed. The similarity calculation module performs weighted fusion of the distance scores of features of different dimensions based on the attention mechanism and outputs the speech matching score.

5. The method for real-time identification of fraudulent calls based on AI semantic understanding according to claim 4 is characterized in that: The hierarchical clustering algorithm is used to classify the similarity results, the classification results are input into the scene recognition network, and the scene discrimination results are output based on the conditional random field structure, including: Construct a hierarchical clustering model, calculate the Euclidean distance between samples based on the speech matching score, iteratively merge the clusters with the closest distance using the minimum distance method to generate a clustering tree structure, dynamically prune the clustering tree structure to obtain speech clusters, calculate the inter-class distance and intra-class distance of the speech clusters, determine the optimal number of clusters based on the ratio of the inter-class distance to the intra-class distance, and convert the optimal clustering result into a speech type label; The speech type label is input into the scene recognition network, and the scene recognition network uses a conditional random field structure to establish a state transfer matrix. The conditional probability distribution of the label sequence is calculated through a forward algorithm, and the optimal label path is searched based on the Viterbi algorithm. The optimal label path is decoded into a scene category, and the scene category is probability scored to obtain a scene discrimination result.

6. A real-time fraud call identification device based on AI semantic understanding, characterized in that: The device comprises: A semantic processing module, which is used to perform speech-to-text processing on the call audio stream to obtain conversation text, perform word segmentation and syntactic structure analysis on the conversation text based on a word segmenter, extract the dependency relationships between sentence components to obtain grammatical features, identify semantic roles and key entities in the sentence through a semantic annotation module, and use a multi-head attention mechanism to perform contextual analysis on the grammatical features to generate a semantic representation vector; A speech recognition module is used to build a speech recognition model, extract conversation intention features based on the semantic representation vector, extract typical semantic patterns from the fraud speech template library, calculate the similarity between the conversation intention features and the semantic patterns, classify the similarity results using a hierarchical clustering algorithm, input the classification results into a scene recognition network, and output a scene discrimination result based on a conditional random field structure; The risk warning module is used to extract fraud case features from historical annotated data, including voice features, text features and behavioral features, and construct a feature processing network. The feature processing network uses a multi-branch convolution structure to map features of different dimensions, and the mapped features are input into the residual connection module. The residual connection module short-circuits the features, retains the underlying feature information through jump connections, performs unique hot encoding on the scene discrimination results, and splices the encoded results with the residual features; constructs a feature fusion network, which uses a fully connected layer to reduce the dimension of the spliced ​​features, adjusts the feature distribution through a batch normalization layer, uses the attention mechanism to calculate the importance weights of different features, adaptively fuses the features based on the weight coefficients, and generates a risk feature vector through a nonlinear transformation of the fusion results. ; Construct a risk scoring module, perform maximum and minimum value normalization on the risk feature vector, input the normalized features into the scoring network, the scoring network uses a multi-layer perceptron structure to perform nonlinear mapping on the features, and converts the mapping results into probability distributions through the sigmoid activation function. The multi-dimensional probability scores are synthesized into a single risk score based on the weighted average method, and the risk scores are exponentially smoothed to eliminate short-term fluctuations; Construct a risk grading module, compare the risk score with the preset multi-level risk threshold, divide the calls into different risk levels based on threshold segmentation, set differentiated response strategies for different risk levels, and input the response strategies into the task scheduler. The task scheduler triggers the corresponding prevention and control instructions according to the priority configuration, and distributes the prevention and control instructions to the execution module based on the pipeline mechanism.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the real-time fraud call identification method based on AI semantic understanding as described in any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for real-time identification of fraudulent calls based on AI semantic understanding as described in any one of claims 1 to 5 are implemented.

Images