Dynamic interception system and interception method based on dynamic rules and deep learning

Abstract

The invention relates to a dynamic interception system and interception method based on a dynamic rule and deep learning, the system comprises an interception condition providing module, an interception module and an optimization module, the interception condition providing module comprises a dynamic rule sub-module and a call discrimination sub-module, the dynamic rule sub-module is used for combing business characteristics by risk control experts according to business experience and common characteristics of real cases to form a dynamic rule base, and providing a dynamically updated interception condition list for the interception module; the call judgment sub-module comprises a plurality of judgment models and is used for analyzing a call text to judge whether the call text needs to be intercepted or not; the interception module is used for screening and intercepting calls in real time according to the interception conditions of the interception condition providing module; and the optimization module is used for rotatably intercepting the call for quality inspection by adopting a sampling strategy, adjusting an interception rule parameter, correcting a discrimination label and carrying out parameter adjustment on the call discrimination sub-module. And intelligent discrimination and accurate interception of the call content are realized.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and in particular to a dynamic interception system and interception method based on dynamic rules and deep learning. Background Art

[0002] With the rapid development of communications technology, while telephone communications offer convenience, they also face challenges such as commercial or sales calls, which can impact users' communication experience and sense of security. Traditional interception systems often rely on pre-set rules, such as filtering based on number ranges, identifying specific keywords, or restricting calls to fixed frequencies. While simple and easy to implement, these methods have significant limitations: their rule bases are often static or semi-static, and once established, they struggle to flexibly and promptly respond to new risk patterns. When risk initiators employ new technical circumvention methods (such as changing call patterns or adjusting communication content), these systems are prone to failure, lacking overall adaptability and struggling to adapt to the dynamically changing communication environment.

[0003] In recent years, machine learning, particularly deep learning, has achieved significant breakthroughs in speech recognition and natural language understanding. Applying these technologies to communication content analysis can theoretically improve the accuracy and intelligence of identifying harmful information. However, current machine learning-based interception solutions still face key challenges:

[0004] Insufficient dynamic adaptability of rules: Existing systems rely heavily on pre-set rule bases (such as blacklists, keyword libraries, and call frequency thresholds). These rules lack efficient dynamic update and learning mechanisms. Consequently, they are slow to respond to evolving risk strategies and technical mitigation methods (such as adjusting call times and changing wording), resulting in a decrease in protection effectiveness over time.

[0005] Limited generalization capabilities of a single model: While deep learning offers advantages in content analysis, a single model struggles to fully capture the complex characteristics of diverse risk scenarios (such as varying communication content, tone, and word patterns). Risk types vary significantly, making it difficult for a single model to achieve high-precision, comprehensive recognition. Furthermore, the lack of an effective closed-loop model optimization system makes it difficult to leverage feedback data from actual operations for continuous iteration and performance improvement.

[0006] Manual review consumes significant resources: Existing systems generally require extensive manual intervention to review and analyze interception results during the verification phase. This approach is not only inefficient and limited in response time when processing massive amounts of call data, but also carries high labor costs. Furthermore, the consistency and real-time nature of manual judgments are difficult to guarantee, impacting the overall operational efficiency and scalability of the system.

[0007] Therefore, there is an urgent need for an interception system for commercial calls or sales calls that can solve the above problems. Summary of the Invention

[0008] The present invention aims to provide a dynamic interception system for privacy number risk control based on rules and deep learning. By dynamically adjusting the interception conditions and integrating multiple deep learning models, it can achieve intelligent judgment and precise interception of call content. At the same time, it uses manual quality inspection feedback to optimize the model, reduce the false alarm rate and missed alarm rate, and improve the overall interception efficiency and accuracy of the system.

[0009] The technical problem to be solved by the present invention is achieved through the following technical solutions: According to a first aspect of the present application, a dynamic interception system based on dynamic rules and deep learning is provided, comprising an interception condition providing module, an interception module, and an optimization module, wherein: The interception condition provision module includes a dynamic rule submodule and a call discrimination submodule. The dynamic rule submodule is composed of a manual or AI risk control system that sorts out business characteristics based on business experience and common characteristics of real cases to form a dynamic rule library, providing the interception module with a dynamically updated list of interception conditions; the call discrimination submodule includes several discrimination models for parsing call text to determine whether interception is required; The interception module is used to screen and intercept calls in real time according to the interception conditions of the interception condition providing module; The optimization module is used to adopt a sampling strategy to rotate and intercept calls for quality inspection, adjust interception rule parameters and correct discrimination labels, and adjust parameters of the call discrimination submodule.

[0010] Preferably, in the dynamic rule submodule, the business characteristics are sorted out according to the business experience and the common characteristics of the actual cases to form a dynamic rule base, specifically: interception conditions are generated according to the number behavior, number identity attributes and customer identity attributes, and the rules are formulated in combination with legal policies to form the dynamic rule base; the dynamic rule base can be updated in real time as needed.

[0011] Preferably, in the call discrimination submodule, each call is output by several discrimination models and fused to form a call label, and interception clues are provided to the interception module according to the call label.

[0012] Preferably, the call discrimination submodule outputs discrimination and confidence for each call using several discrimination models. The specific method is as follows: The system implements recording and uses deep learning NLP technology to convert the recording into text data; The converted text data is preprocessed, including at least removing stop words, word segmentation and part-of-speech tagging; Using the risk keyword identification model and the BERT model in parallel to identify and output bad information from the text data; The outputs of the risk keyword identification model and the BERT model are integrated to obtain a comprehensive identification result as the call label.

[0013] Preferably, the specific method for the risk keyword discrimination model to identify and output bad information from the text data is: Calculate the TF-IDF values ​​of all words in the document collection in the preprocessed text data and construct a TF-IDF matrix; Calculate the TF-IDF values ​​of bad texts and normal texts, select words in the bad texts with TF-IDF values ​​higher than a threshold as bad keywords, classify the bad keywords into high-risk keywords, medium-risk keywords, and low-risk keywords, store them in a risk keyword library, and assign weights; The risk score for each call is calculated using the following formula: ; Where: k is the adjustment coefficient, ranging from 0 to 1, which is used to control the impact of the number of hit keywords on the call score; N is the number of keywords in the call hit risk keyword library; It is determined whether the risk score exceeds a threshold, and if so, the call is determined to be a bad call.

[0014] Preferably, the method further includes extracting keywords from the pre-processed text data using a deepseek or qwen model, and selecting keywords that intersect with the results of the TF-IDF method screening as the final bad keywords.

[0015] Preferably, the specific method of the BERT model for identifying and outputting bad information from the text data is: Convert the text data into a subword sequence T: The subword sequence T is input into the bert model to obtain the context representation. The result after the fully connected layer is z, and the probability is output through the Softmax function: P(y∣X)=Softmax(z)=[p0,p1]; Where y∈{0(normal),1(bad)}, P(y|X) indicates the probability p0 that text X is a normal text, and the probability p1 that the text is a bad text.

[0016] Preferably, the outputs of the risk keyword discrimination model and the BERT model are integrated to obtain a comprehensive discrimination result as the call label. The specific method is as follows: for the same text data, when the risk keyword discrimination model and the BERT model both determine that the call is a bad call, the call is marked as a call label of a bad call.

[0017] Preferably, the specific working steps of the optimization module are: Review and analyze intercepted call records, and mark and correct any erroneous or missed interceptions; Based on the annotations and corrections, the interception rules are adjusted and the model is retrained, including BERT model optimization and dynamic adjustment of risk keyword weights.

[0018] According to a second aspect of the present application, there is provided an interception method using the above-mentioned dynamic interception system based on dynamic rules and deep learning, comprising the following steps: Step 1: The manual or AI risk control system sorts out business features based on business experience and common characteristics of real cases to form a dynamic rule base, which stores a list of interception conditions; Step 2: Several discrimination models in the call discrimination submodule parse the call text to determine whether it needs to be intercepted; Step 3: Screen and intercept calls in real time based on interception conditions; Step 4: Use a sampling strategy to rotate the intercepted calls for quality inspection, adjust the interception rule parameters and correct the discrimination labels, and adjust the parameters of the call discrimination submodule.

[0019] Based on one embodiment of the present application, the advantages of using the dynamic interception system based on dynamic rules and deep learning are: Dynamic interception condition management: By combining rules and deep learning technology, interception conditions can be dynamically adjusted and managed. The system can automatically update the interception condition list based on real-time call data and user feedback, ensuring the accuracy and timeliness of interception strategies. Multi-model fusion discrimination: Multiple models are used in parallel to process call discrimination tasks. By comparing the results of different models, the optimal discrimination is selected as the final result. This multi-model fusion method can significantly improve discrimination accuracy and reduce false positive and false negative rates. Closed-loop optimization mechanism: A closed-loop feedback system combining manual quality inspection and model iteration is built. Through manual review and analysis of interception results, correctly labeled data is fed back to the model for retraining, enabling iterative model optimization. This closed-loop optimization mechanism can continuously improve the model's recognition accuracy and generalization capabilities.

[0020] Scalability and Adaptability: The system design supports the dynamic addition and subtraction of interception rules and discrimination models, making it easy to expand to more application scenarios. Furthermore, the system can adapt to changing communication environments and marketing methods, maintaining high recognition accuracy and interception efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1It is a structural diagram of a dynamic interception system based on dynamic rules and deep learning of the present invention; Figure 2 This is a step flow chart of a dynamic interception method based on dynamic rules and deep learning of the present invention. DETAILED DESCRIPTION

[0022] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0023] It should be noted that the above detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the art to which this application belongs.

[0024] like Figure 1 As shown, the dynamic interception system based on dynamic rules and deep learning in one embodiment of the present application includes an interception condition provision module, an interception module, and an optimization module. The interception module is centrally scheduled and interacts with the interception condition provision module to obtain interception basis, based on which calls are screened and intercepted. After interception, the data flows into the optimization module for quality inspection and optimization. Each module complements each other. Among them:

[0025] The interception condition provision module is responsible for providing the interception module with a dynamically updated interception condition list. This module includes two submodules: the dynamic rule submodule and the call discrimination submodule; Dynamic rule submodule: A manual or AI risk control system sorts out the characteristics of negative business information based on business experience and the common characteristics of real sales calls. For example, interception conditions are generated based on characteristics such as number behavior (high-frequency calls in a short period of time, calls to multiple locations in a short period of time), number identity attributes (blacklist association), and customer identity attributes (with a certain probability of sales calls). (The interception condition is set to whether the probability|frequency of a certain feature is greater than a certain threshold, or whether a certain feature is in the blacklist and whitelist library of the feature, etc.) Rules are formulated in combination with laws and policies, and the rule library is stored and supports online modification and expansion to adapt to the evolution of risks. These interception conditions can be added or deleted as the business is updated and new features are discovered. These rules can be based on call metadata, such as number, time, duration, etc., and application personnel can freely choose which rule combinations to use as interception conditions. For example, interception rules are: Blacklist numbers: Add known sales calls, harassment numbers, or numbers belonging to illegal and irregular customers to the blacklist, and set the interception condition to "intercept calls from blacklisted numbers."

[0026] Frequent calls: Set a rule to block calls if the same number is called more than a certain number of times within a certain period of time. This rule can be used to address dangerous scenarios involving information security.

[0027] Call Duration: Set a rule to block calls with a duration less than or greater than a certain threshold. This rule can be used to filter out unusually short calls, such as those lasting less than a minute. These calls are often associated with sales calls or harassment.

[0028] The dynamic rules submodule supports user-defined rules. Users can add, modify, or delete rules based on their needs and experience. At the same time, the system automatically adjusts rules based on real-time call data and user feedback to improve the accuracy and timeliness of interception.

[0029] The call identification submodule integrates multiple identification models and analyzes call text in parallel. Each model outputs the identification and confidence level of each call. The model with the highest confidence level is selected as the call label to determine whether the call is suspected of dangerous behavior. The label provides interception clues to the interception module. The model can be switched and updated as needed to optimize the identification accuracy. This submodule utilizes deep learning NLP technology to convert call recordings into text and identify scenarios (such as loan scenarios and game top-up scenarios). By integrating multiple models and selecting the optimal results, it provides intelligent interception conditions for the interception module. The call discrimination submodule utilizes deep learning NLP technology to label calls with different labels, such as negative information and normal information. These discrimination labels provide intelligent interception conditions for the interception module. The specific processing methods of this submodule are as follows:

[0030] 1. The system converts real-time recordings into text, using speech recognition technology to convert audio signals into processable text data; 2. Then preprocess the converted text data, including removing stop words, word segmentation, part-of-speech tagging, etc. The preprocessing step helps improve the accuracy of subsequent discrimination; 3. Use the risk keyword identification model and the BERT model to identify whether the call text contains bad information in parallel. The identification methods of different models are as follows: The workflow of the risk keyword identification model is as follows: Pre-build a dynamic keyword library: Based on historical call texts, after pre-processing the data of bad information call texts and normal call texts, use the TF-IDF method and large models such as qwen and deepseek to obtain bad information keywords. In order to adapt to the ever-changing communication environment, the risk keyword library will be expanded at regular time intervals using the historical call texts within that time interval.

[0031] Calculate the TF-IDF values ​​of all words in the document collection after preprocessing and construct the TF-IDF matrix. The formula is as follows: ; The term frequency (TF) of word t indicates how often a word appears in a single document. The inverse document frequency (IDF) of word t measures the general importance of the word. The calculation formula is: ; The TF-IDF value calculation formula for word t is: TF-IDF(t,d,D)=TF(t,d)×IDF(t,D) After calculating the TF-IDF values ​​of the bad text and the normal text, we select words with significantly higher TF-IDF values ​​in the bad text as bad keywords, and classify the bad keywords into high-risk keywords, medium-risk keywords, and low-risk keywords based on their TF-IDF thresholds. In order to make the detected risk keywords more understandable, the same text is used to extract keywords related to bad text using the DeepSeek or Qwen large model. The intersection of the risk keywords extracted by the TF-IDF method and the keywords extracted by the large model is taken as the final risk keyword result. The results are stored in the risk keyword library according to the high, medium and low risk levels. Keywords of different levels have different scores and weights, with scores ranging from 0 to 100 and weights ranging from 0 to 1.

[0032] Calculate the risk score for each call using the following formula: ; Where k is the adjustment coefficient, which ranges from 0 to 1 and is used to control the impact of the number of hit keywords on the call score. N is the number of keywords in the call hit risk keyword library.

[0033] If the risk score of the call is greater than a certain threshold A, the call is considered to be a sales call.

[0034] The workflow of the Bert model is as follows: First, train the BERT model: preprocess the input text. Suppose the input text is X={x1,x2,...,xn}, and BERT's Tokenization converts it into a subword sequence: T=Tokenizer(X)=[t1,t2,...,tm] Where m is the sequence length (padded or truncated to the maximum length), and n is the number of words in the text.

[0035] Input T into the Bert model to obtain the context representation. The result after the fully connected layer is z, and the probability is output through the Softmax function: P(y|X)=Softmax(z)=[p0,p1], where y∈{0(normal),1(bad text)}. P(y|X) indicates the probability p0 that text X is normal and the probability p1 that text X is bad text.

[0036] Then use the trained bert model to make predictions: Predictive text Use the trained BERT model To make a judgment, ; Predictive text The probability of being good and bad text.

[0037] 4. The BERT model outputs probabilities for normal and bad text based on the input text. The risk keyword discrimination model outputs the probability of the text being bad. If the bad text probability from the BERT model is above a certain threshold B, and the bad text probability from the risk keyword discrimination model is above a certain threshold C, the call is comprehensively judged to be a sales call. If the bad text probability from the BERT model is above a certain threshold D, and the bad text probability from the risk keyword discrimination model is below a certain threshold C, the call is comprehensively judged to be normal. The results of these two models provide a comprehensive judgment result.

[0038] This multi-model fusion method can significantly improve the discrimination accuracy and reduce the false alarm rate and missed alarm rate.

[0039] The interception module of this system is responsible for providing modules based on dynamically updated interception conditions, screening and intercepting qualified calls in real time. The interception conditions can be flexibly increased or decreased to ensure that the system responds quickly and is highly adaptable. The specific functions of this module are as follows: Real-time interception: The interception module receives call data in real time and filters and intercepts calls based on a list of interception criteria. The system supports multiple interception methods, such as hanging up the call directly, hanging up after playing a prompt tone, and transferring the call to customer service.

[0040] Dynamic Condition Management: The blocking condition list is dynamically updated, and the system can automatically adjust the blocking conditions based on real-time call data and user feedback. For example, if the system detects a number frequently making harassing calls, it can automatically add the number to the blacklist and update the blocking condition list.

[0041] Dynamic selection of interception: Application personnel can freely select the rules of the interception condition provision module below to generate a list of combined interception conditions. Once a match is made, the call will be intercepted immediately to ensure a quick response to suspected risk calls.

[0042] Dynamic condition management: A rule update engine is set up to flexibly increase or decrease interception conditions based on system performance indicators (false interception rate, missed interception rate) and business risk control strategies (such as relaxed interception during promotional periods). For example, when encountering a batch of abnormal calls during peak hours, number frequency interception rules can be quickly added to strengthen prevention and control.

[0043] Interception log records: The interception module records detailed information about each interception operation, including call time, call duration, blocked number, interception reason, etc. This information can be used for subsequent analysis and optimization.

[0044] Multiple Condition Combinations: The interception module supports combining multiple interception conditions. For example, you can set "blacklisted number + call duration greater than 30 minutes" as a combined interception condition to handle scenarios involving extremely long sales calls from historically blacklisted numbers.

[0045] The specific function of the optimization module is: the professional quality inspection team selects intercepted calls for quality inspection according to the preset sampling strategy, carefully examines the semantics and context to determine whether the interception is right or wrong, adjusts the interception rule parameters based on the conclusion (such as relaxing the blacklist association level), corrects the discrimination labels, and returns the data after quality inspection to the call discrimination sub-module, uses manually labeled data to optimize the model, precipitate new regulations, and enrich the system's risk control capabilities.

[0046] The specific workflow of this module is as follows: Manual quality control: Manual quality control is the foundation of the optimization module. The system reviews and analyzes intercepted call records to determine whether the interception was correct. Any inaccurate or missed interceptions are marked and corrected by quality control personnel.

[0047] Rule Adjustment: Based on manual quality inspection results, the system will adjust interception rules. For example, if a rule is found to be causing a large number of false interceptions, the threshold or conditions of that rule can be adjusted. If a new sales tactic is discovered, a new rule can be added to address it.

[0048] Model iteration: The optimization module feeds correctly labeled data back to the discriminant model for retraining.

[0049] (1) BERT model optimization: After data preprocessing, the manually annotated bad text and normal text are retrained with the BERT model to improve the recognition accuracy of the BERT model with high-quality data.

[0050] (2) Dynamic adjustment of risk keyword weights: In order to adapt to the ever-changing communication environment and means, the scores and weights of the keywords in the risk keyword library will be updated at regular intervals, and the hit ratio of each keyword in the keyword library during that period will be calculated. The formula is as follows: ; If the keyword ratio is greater than threshold a, the keyword weight is increased (for example, by 0.1). If it is less than 1, the score is increased (for example, by 10). If the keyword ratio is less than threshold b, the keyword weight is decreased (for example, by 0.1) and the score is decreased (for example, by 10). If the weight change exceeds the weight range of each level, the risk level (keyword score) is updated.

[0051] The above iterative training process can continuously improve the recognition accuracy and generalization ability of the model. At the same time, the system can also introduce new deep learning models or algorithms according to actual needs to further improve the accuracy of discrimination.

[0052] like Figure 2 As shown, in one embodiment of the present application, an interception method using the above-mentioned dynamic interception system based on dynamic rules and deep learning is disclosed, comprising the following steps: Step 1: The manual or AI risk control system sorts out business characteristics based on business experience and common features of real cases to form a dynamic rule base, which stores a list of interception conditions; In this step, the risk control team customizes the initial interception rules based on business characteristics and historical risk data, enters the rule submodule, selects multiple suitable model sets to deploy the call discrimination submodule, sets parameters such as the optimization module quality inspection cycle and sampling ratio, and starts the system operation. Step 2: When a call is connected, the interception conditions are compared. If they match, the call is intercepted. If they do not match, the call is forwarded to the several discrimination models in the call discrimination submodule to parse the call text and then decide whether to intercept or not. In the steps, the moment the call is connected, the interception module retrieves the condition library for comparison. If a match is found, the call is intercepted and the details are recorded for quality inspection. If a match is found, the recording is fast-forwarded to text and sent to the discrimination sub-module. Multiple models are operated in parallel, and labels are determined based on confidence levels. Risky calls are intercepted and sent for inspection, and a rapid response is given throughout the process to ensure timeliness.

[0053] Step 3: Use a sampling strategy to rotate the intercepted calls for quality inspection, adjust the interception rule parameters and correct the discrimination labels, and adjust the parameters of the call discrimination submodule.

[0054] In this module, the quality inspection team periodically analyzes intercepted samples, corrects rules, calibrates judgments, updates the model weight structure, enters new rules into the database, and puts the model online, repeating the cycle so that the system can continuously enhance its risk control effectiveness through practical polishing.

[0055] This method has the following advantages: Efficient and accurate interception: Dynamic interception captures emerging risks, and multi-model fusion and discrimination deeply explore semantics, reducing misjudgments, improving interception accuracy, and protecting the purity of business communications.

[0056] Continuous system evolution: Closed-loop optimization leverages quality inspection insights to tune rules, enhance model generalization, cope with complex business changes, and extend the system's effective lifecycle.

[0057] Scalability and Adaptability: The system design supports the dynamic addition and subtraction of interception rules and discrimination models, making it easy to expand to more application scenarios. Furthermore, the system can adapt to changing communication environments and transmission methods, maintaining high recognition accuracy and interception efficiency.

[0058] The adoption of the above technical solution effectively solves many problems existing in the existing privacy number risk control and interception system, and provides strong support for building a more secure and efficient communication environment.

[0059] It should be noted that the terms used herein are intended only to describe specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0060] It should be noted that the terms "first," "second," and the like in the specification and claims of this application and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, such that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0061] In addition, the terms "comprises" and "comprising" and any variations thereof are intended to cover a non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or elements is not necessarily limited to those steps or elements expressly listed but may include other steps or elements not expressly listed or inherent to such process, method, product, or apparatus.

[0062] For ease of description, spatially relative terms such as "above", "above", "on the upper surface of", "above", etc. may be used herein to describe the spatial positional relationship of a device or feature to other devices or features as shown in the figures. It should be understood that spatially relative terms are intended to include different orientations of the device in use or operation in addition to the orientation described in the figures. For example, if the device in the drawings is inverted, the device described as "above other devices or structures" or "above other devices or structures" will be subsequently positioned as "below other devices or structures" or "below other devices or structures". Thus, the exemplary term "above" can include both "above" and "below". The device can also be positioned in other different ways, such as rotated 90 degrees or in other orientations, and the spatially relative descriptions used herein are interpreted accordingly.

[0063] In the above detailed description, reference is made to the accompanying drawings, which form a part hereof. In the drawings, similar symbols typically identify similar components, unless the context dictates otherwise. The illustrated embodiments described in the detailed description, drawings, and claims are not meant to be limiting. Other embodiments may be used, and other changes may be made, without departing from the spirit or scope of the subject matter presented herein.

[0064] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. Dynamic interception system based on dynamic rules and deep learning, characterized by: It includes interception condition provision module, interception module and optimization module, among which: The interception condition provision module includes a dynamic rule submodule and a call discrimination submodule. The dynamic rule submodule is composed of a manual or AI risk control system that sorts out business characteristics based on business experience and common characteristics of real cases to form a dynamic rule library, providing the interception module with a dynamically updated list of interception conditions; the call discrimination submodule includes several discrimination models for parsing call text to determine whether interception is required; The interception module is used to screen and intercept calls in real time according to the interception conditions of the interception condition providing module; The optimization module is used to adopt a sampling strategy to rotate and intercept calls for quality inspection, adjust interception rule parameters and correct discrimination labels, and adjust parameters of the call discrimination submodule.

2. The dynamic interception system based on dynamic rules and deep learning according to claim 1 is characterized in that: In the dynamic rule submodule, business characteristics are sorted out according to business experience and common characteristics of real cases to form a dynamic rule base. Specifically, interception conditions are generated according to number behavior, number identity attributes and customer identity attributes, and the rules are formulated in combination with legal policies to form the dynamic rule base. The dynamic rule base can be updated in real time as needed.

3. The dynamic interception system based on dynamic rules and deep learning according to claim 1 is characterized in that: In the call discrimination submodule, each call is output by several discrimination models and fused to form a call label, and the call label provides an interception clue for the interception module.

4. The dynamic interception system based on dynamic rules and deep learning according to claim 3 is characterized in that: The call discrimination submodule outputs discrimination and confidence for each call using several discrimination models. The specific method is as follows: The system implements recording and uses deep learning NLP technology to convert the recording into text data; The converted text data is preprocessed, including at least removing stop words, word segmentation and part-of-speech tagging; Using the risk keyword identification model and the BERT model in parallel to identify and output bad information from the text data; The outputs of the risk keyword identification model and the BERT model are integrated to obtain a comprehensive identification result as the call label.

5. The dynamic interception system based on dynamic rules and deep learning according to claim 4 is characterized in that: The specific method of the risk keyword identification model to identify and output bad information from the text data is as follows: Calculate the TF-IDF values ​​of all words in the document collection in the preprocessed text data and construct a TF-IDF matrix; Calculate the TF-IDF values ​​of bad texts and normal texts, select words in the bad texts with TF-IDF values ​​higher than a threshold as bad keywords, classify the bad keywords into high-risk keywords, medium-risk keywords, and low-risk keywords, store them in a risk keyword library, and assign weights; The risk score for each call is calculated using the following formula: ； Where: k is the adjustment coefficient, ranging from 0 to 1, which is used to control the impact of the number of hit keywords on the call score; N is the number of keywords in the call hit risk keyword library; It is determined whether the risk score exceeds a threshold, and if so, the call is determined to be a bad call.

6. The dynamic interception system based on dynamic rules and deep learning according to claim 5 is characterized in that: It also includes extracting keywords from the preprocessed text data using the deepseek or qwen model, and selecting keywords that intersect with the TF-IDF method screening results as the final bad keywords.

7. The dynamic interception system based on dynamic rules and deep learning according to claim 4 is characterized in that: The specific method of the BERT model to identify and output bad information from the text data is as follows: Convert the text data into a subword sequence T: The subword sequence T is input into the bert model to obtain the context representation. The result after the fully connected layer is z, and the probability is output through the Softmax function: P(y∣X)=Softmax(z)=[p0,p1]; Where y∈{0(normal),1(bad)}, P(y|X) indicates the probability p0 that text X is a normal text, and the probability p1 that the text is a bad text.

8. The dynamic interception system based on dynamic rules and deep learning according to claim 4 is characterized in that: The specific method of fusing the outputs of the risk keyword discrimination model and the BERT model to obtain a comprehensive discrimination result as the call label is: for the same text data, when the risk keyword discrimination model and the BERT model both determine that the call is a bad call, the call is marked as a call label of a bad call.

9. The dynamic interception system based on dynamic rules and deep learning according to claim 1, characterized in that: The specific working steps of the optimization module are: Review and analyze intercepted call records, and mark and correct any erroneous or missed interceptions; Based on the annotations and corrections, the interception rules are adjusted and the model is retrained, including BERT model optimization and dynamic adjustment of risk keyword weights.

10. An interception method using the dynamic interception system based on dynamic rules and deep learning according to any one of claims 1 to 9, characterized in that: The steps include: Step 1: The manual or AI risk control system sorts out business features based on business experience and common characteristics of real cases to form a dynamic rule base, which stores a list of interception conditions; Step 2: When a call is connected, the interception conditions are compared. If they match, the call is intercepted. If they do not match, the call is forwarded to the several discrimination models in the call discrimination submodule to parse the call text and then decide whether to intercept or not. Step 3: Use a sampling strategy to rotate the intercepted calls for quality inspection, adjust the interception rule parameters and correct the discrimination labels, and adjust the parameters of the call discrimination submodule.

Images