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

Abstract

The application relates to a dynamic interception system and method based on dynamic rules and deep learning, which comprises an interception condition providing module, an interception module and an optimization module, wherein the interception condition providing module comprises a dynamic rule submodule and a call discrimination submodule; the dynamic rule submodule is formed by combing business characteristics into a dynamic rule library according to business experience and common characteristics of real cases by risk control experts, and provides an interception condition list dynamically updated for the interception module; the call discrimination submodule comprises a plurality of discrimination models, which are used for analyzing call texts to determine whether interception is needed; the interception module is used for screening and intercepting calls in real time according to the interception conditions of the interception condition providing module; the optimization module is used for rotating intercepted calls for quality inspection by adopting a sampling strategy, adjusting interception rule parameters and correcting discrimination labels, and adjusting parameters of the call discrimination submodule. Intelligent discrimination and accurate interception of call contents are realized.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to a dynamic interception system and method based on dynamic rules and deep learning. Background Technology

[0002] With the rapid development of communication technology, while telephone communication brings convenience, it also faces challenges such as business calls and sales calls, which may affect users' communication experience and sense of security. Traditional interception systems typically rely on preset rules, such as filtering based on number ranges, identifying specific keywords, or restricting fixed-frequency calls. While these methods are simple and easy to implement, they have significant limitations: their rule bases are often static or semi-static, and once set, they are difficult to flexibly and promptly respond to changes in new risk patterns. When risk initiators adopt new technical circumvention methods (such as changing call patterns or adjusting communication content), such systems are prone to failure, lacking overall adaptability and struggling to match dynamically changing communication environments.

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

[0004] Insufficient dynamic adaptability of rules: Existing systems rely heavily on preset rule bases (such as blacklists, keyword databases, and call frequency thresholds). These rules lack efficient dynamic update and learning mechanisms, resulting in slow response and poor flexibility when faced with constantly evolving risk strategies and technical evasion methods (such as adjusting call times and changing wording), leading to a decline in protection effectiveness over time.

[0005] Limited generalization ability of a single model: While deep learning has advantages in content analysis, a single model struggles to fully cover the complex features of diverse risk scenarios (such as varying communication content, tone, and word patterns). Different risk types differ significantly, making it difficult for a single model to achieve high-precision, comprehensive identification. Furthermore, the lack of an effective model optimization loop hinders the system from utilizing feedback data from actual operation for continuous iteration and performance improvement.

[0006] Manual verification is resource-intensive: In the result verification stage, existing systems generally require a large amount of manual intervention to review and analyze the intercepted results. This approach is not only inefficient and has limited response speed when processing massive amounts of call data, but also has high labor costs. At the same time, the consistency and real-time performance of manual judgments are difficult to guarantee, affecting the overall operating efficiency and scalability of the system.

[0007] Therefore, there is an urgent need for a system to block business or telemarketing calls that can solve the above problems. Summary of the Invention

[0008] This 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 identification and accurate interception of call content. At the same time, it uses manual quality inspection feedback to optimize the model, reduce the false alarm rate and the false negative rate, and improve the overall interception efficiency and accuracy of the system.

[0009] The technical problem to be solved by this invention is achieved through the following technical solution:

[0010] According to a first aspect of this 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:

[0011] The interception condition providing module includes a dynamic rule submodule and a call discrimination submodule. The dynamic rule submodule is formed by a manual or AI risk control system sorting out business characteristics based on business experience and common characteristics of real cases to form a dynamic rule base, providing the interception module with a dynamically updated list of interception conditions. The call discrimination submodule includes several discrimination models used to parse call text to determine whether interception is necessary.

[0012] The interception module is used to provide the module's interception conditions based on the interception conditions, and to filter and intercept calls in real time;

[0013] The optimization module is used to perform quality inspection on the intercepted calls by using a sampling strategy, and to adjust the interception rule parameters and correction discrimination labels, thereby adjusting the parameters of the call discrimination submodule.

[0014] Preferably, in the dynamic rules submodule, the dynamic rule base is formed by sorting out business characteristics based on business experience and common characteristics of actual cases. Specifically, this involves generating interception conditions based on number behavior, number identity attributes, and customer identity attributes, and then formulating rules in conjunction with laws and policies to form the dynamic rule base. The dynamic rule base can be updated in real time as needed.

[0015] Preferably, in the call discrimination submodule, each call is output by several discrimination models and fused to form a call tag, and the call tag provides interception clues to the interception module.

[0016] Preferably, the call discrimination submodule outputs a discrimination and confidence score for each call using several discrimination models. The specific method is as follows:

[0017] The system records audio and uses deep learning NLP technology to convert the recordings into text data.

[0018] The converted text data undergoes preprocessing, including at least the removal of stop words, word segmentation, and part-of-speech tagging.

[0019] The text data is used to identify and output harmful information by employing a risk keyword discrimination model and a BERT model in parallel.

[0020] By combining the outputs of the risk keyword discrimination model and the BERT model, a comprehensive discrimination result is obtained as the call tag.

[0021] Preferably, the specific method for the risk keyword discrimination model to identify and output harmful information from the text data is as follows:

[0022] Calculate the TF-IDF values ​​of all words in the preprocessed text data within the document set, and construct a TF-IDF matrix.

[0023] Calculate the TF-IDF values ​​of malicious text and normal text, select words in malicious text with TF-IDF values ​​higher than the threshold as malicious keywords, classify the malicious keywords into high-risk keywords, medium-risk keywords and low-risk keywords, store them in the risk keyword library and assign weights;

[0024] The risk score for each call is calculated using the following formula:

[0025] ;

[0026] Where: k is the adjustment coefficient, which takes a value between 0 and 1, and is used to control the degree of influence of the number of hit keywords on the call score; N is the number of keywords in the call hit risk keyword library.

[0027] Determine whether the risk score exceeds the threshold; if so, the call is determined to be a bad call.

[0028] Preferably, the method further includes extracting keywords from the preprocessed text data using Deepseek or Qwen large-scale models, and selecting keywords that overlap with the TF-IDF method screening results as the final bad keywords.

[0029] Preferably, the specific method for the BERT model to identify and output inappropriate information from the text data is as follows:

[0030] Convert the text data into a word sequence T:

[0031] The word sequence T is input into the BERT model to obtain a contextual representation. The result after the fully connected layer is z, and the probability is output through the Softmax function.

[0032] P(y∣X)=Softmax(z)=[p0,p1];

[0033] Where y∈{0(normal),1(bad)}, P(y|X) represents the probability p0 that text X belongs to normal text and the probability p1 that text X belongs to bad text.

[0034] Preferably, the specific method for combining the outputs of the risk keyword discrimination model and the BERT model to obtain the comprehensive discrimination result as the call tag is as follows: when both the risk keyword discrimination model and the BERT model determine the same text data as a bad call, then the call is labeled as a bad call.

[0035] Preferably, the specific working steps of the optimization module are as follows:

[0036] Review and analyze the blocked call records, and mark and correct any cases of false or missed blocking;

[0037] 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.

[0038] According to a second aspect of this application, an interception method employing the aforementioned dynamic interception system based on dynamic rules and deep learning is provided, comprising the following steps:

[0039] Step 1: The manual or AI risk control system sorts out business characteristics based on business experience and common characteristics of real cases to form a dynamic rule base, which stores a list of interception conditions.

[0040] Step 2: Several discrimination models in the call discrimination submodule parse the call text to determine whether interception is necessary;

[0041] Step 3: Based on the blocking criteria, filter and block calls in real time;

[0042] Step 4: Use a sampling strategy to rotate and intercept calls for quality inspection, and adjust the interception rule parameters and correction labels to adjust the parameters of the call discrimination submodule.

[0043] Based on one embodiment of this application, the advantages of using this dynamic interception system based on dynamic rules and deep learning are as follows:

[0044] Dynamic blocking condition management: By combining rules and deep learning technology, the blocking conditions can be dynamically adjusted and managed. The system can automatically update the blocking condition list based on real-time call data and user feedback to ensure the accuracy and timeliness of the blocking strategy.

[0045] Multi-model fusion discrimination: Multiple models are used to process the call discrimination task in parallel. By comparing the results of different models, the optimal discrimination is selected as the final result. This multi-model fusion method can significantly improve the discrimination accuracy and reduce the false alarm rate and false negative rate.

[0046] Closed-loop optimization mechanism: A closed-loop feedback system is constructed for manual quality inspection and model iteration. By manually reviewing and analyzing the interception results, correctly labeled data is fed back to the model for retraining, achieving iterative optimization. This closed-loop optimization mechanism can continuously improve the model's recognition accuracy and generalization ability.

[0047] Scalability and Adaptability: The system design supports the dynamic addition and removal of interception rules and discrimination models, making it easy to expand to more application scenarios. At the same time, the system can adapt to constantly changing communication environments and sales tactics, maintaining high identification accuracy and interception efficiency. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the structure of a dynamic interception system based on dynamic rules and deep learning according to the present invention;

[0049] Figure 2 This is a flowchart of the steps of a dynamic interception method based on dynamic rules and deep learning according to the present invention. Detailed Implementation

[0050] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0051] It should be noted that the above detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0052] like Figure 1 As shown, a dynamic interception system based on dynamic rules and deep learning in one embodiment of this application includes an interception condition providing module, an interception module, and an optimization module. The interception module is centrally scheduled and interacts with the interception condition providing module to obtain interception criteria, thereby filtering and intercepting calls. After interception, the data flows into the optimization module for quality inspection and optimization. Each module complements the others. Wherein:

[0053] The interception condition provision module is responsible for providing the interception module with a dynamically updated list of interception conditions. This module includes two sub-modules: a dynamic rules sub-module and a call discrimination sub-module.

[0054] The dynamic rules submodule: Based on business experience and common characteristics of real telemarketing calls, the manual or AI-driven risk control system identifies problematic business information features. These features include number behavior (high-frequency calls in a short period, calls from multiple locations in a short time), number identity attributes (blacklist association), and customer identity attributes (a certain probability of receiving telemarketing calls). Interception conditions are generated based on these features (e.g., whether the probability / frequency of a certain feature exceeds a certain threshold, or whether a feature is in the blacklist / whitelist for that feature). Rules are formulated in conjunction with regulations and policies, stored in a rule base, and supported for online modification and expansion to adapt to evolving risks. These interception conditions can be added or deleted as business updates and new features are discovered. These rules can be based on call metadata such as number, time, and duration, allowing application users to freely choose which rule combinations to use as interception conditions. Example interception rules:

[0055] Blacklisted Numbers: Add known telemarketing numbers, harassing numbers, or numbers registered under the names of customers with illegal or irregular activities to the blacklist, and set the blocking condition to "block calls from blacklisted numbers".

[0056] Frequent calls: Set the rule to "block calls from the same number more than a certain number of times within a certain period of time". This rule can deal with some dangerous scenarios involving information security.

[0057] Call duration: Set the rule to "block calls with a duration less than or greater than a specific threshold". This rule can be used to filter abnormally short calls, such as calls where the caller and the called party talk for less than one minute. These calls are often associated with telemarketing or harassment.

[0058] The dynamic rules submodule supports user-defined rules, allowing users to add, modify, or delete rules based on their needs and experience. Simultaneously, the system automatically adjusts rules based on real-time call data and user feedback to improve the accuracy and timeliness of blocking.

[0059] The call discrimination submodule integrates multiple discrimination models, parses call text in parallel, and each model outputs the discrimination and confidence score of each call. The one with the highest confidence score is selected as the label of this call to determine whether the call involves dangerous behavior. The label provides interception clues to the interception module. The model can be switched and updated as needed to optimize the discrimination accuracy.

[0060] This submodule utilizes deep learning NLP technology to convert call recordings into text and perform discrimination (e.g., loan scenarios, game recharge scenarios, etc.). By integrating multiple models and selecting the optimal result, it provides intelligent interception conditions for the interception module. The call discrimination submodule uses deep learning NLP technology to label calls with different tags, such as malicious information and normal information. These discrimination tags can provide intelligent interception conditions for the interception module. The specific processing method of this submodule is as follows:

[0061] 1. The system will convert real-time recordings into text, using speech recognition technology to convert audio signals into processable text data;

[0062] 2. Then, the converted text data is preprocessed, including stop word removal, word segmentation, and part-of-speech tagging. The preprocessing step helps improve the accuracy of subsequent discrimination.

[0063] 3. Then, the risk keyword discrimination model and the BERT model are used in parallel to identify whether the call text contains inappropriate information. The identification methods of the different models are as follows:

[0064] The workflow of the risk keyword identification model is as follows:

[0065] Pre-build a dynamic keyword library: Based on historical call texts, after preprocessing the texts of calls containing harmful information and normal call texts, use the TF-IDF method and large models such as qwen and deepseek to obtain keywords containing harmful information. In order to adapt to the constantly changing communication environment, the risk keyword library is expanded at regular intervals using historical call texts within that time interval.

[0066] Calculate the TF-IDF values ​​of all words in the document set after preprocessing, and construct the TF-IDF matrix as follows: ;

[0067] The term frequency (TF) of word t represents the frequency with which the word appears in a single document, while the inverse document frequency (IDF) of word t measures the general importance of the word. The formula for calculating it is:

[0068] ;

[0069] The formula for calculating the TF-IDF value of word t is:

[0070] TF-IDF(t,d,D)=TF(t,d)×IDF(t,D)

[0071] After calculating the TF-IDF values ​​of malicious and normal text, words with significantly higher TF-IDF values ​​in the malicious text are selected as malicious keywords. Based on their TF-IDF thresholds, the malicious keywords are divided into high-risk keywords, medium-risk keywords, and low-risk keywords.

[0072] To make the detected risk keywords more understandable, the same text is used to extract keywords related to the malicious text using deepseek or qwen large models. 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 risk keywords are stored in the risk keyword library according to the risk level of high, medium and low. Keywords of different levels have different scores and weights, with scores ranging from 0 to 100 and weights ranging from 0 to 1.

[0073] The risk score for each call is calculated using the following formula:

[0074] ;

[0075] Where k is an adjustment coefficient, ranging from 0 to 1, used to control the impact of the number of hit keywords on the call score, and N is the number of keywords in the call hit risk keyword library.

[0076] If the risk score of a call is greater than a certain threshold A, the call is considered a sales call.

[0077] The workflow of the BERT model is as follows:

[0078] First, the BERT model is trained: Input text preprocessing, assuming the input text is X={x1,x2,...,xn}, BERT's tokenization converts it into a sequence of subwords:

[0079] T = Tokenizer(X) = [t1, t2, ..., tm]

[0080] Where m is the sequence length (padding or truncation to the maximum length), and n is the number of words in the text.

[0081] Inputting T into the BERT model yields a contextual representation. The result after passing through a fully connected layer is z, which is then output as a probability using the Softmax function.

[0082] P(y|X)=Softmax(z)=[p0,p1], where y∈{0(normal),1(inappropriate text)}. P(y|X) represents the probability p0 that text X is normal and the probability p1 that text X is inappropriate.

[0083] Then, the trained BERT model is used for prediction:

[0084] Predicted text Using the trained Bert model To make a judgment,

[0085] ;

[0086] That is, the predicted text The probability of being normal or bad text.

[0087] 4. The BERT model outputs the probability of the input text being normal or inappropriate, while the risk keyword discrimination model outputs the probability of the text being inappropriate. If the probability of inappropriate text in the BERT model is higher than a certain threshold B, and the probability of inappropriate text in the risk keyword discrimination model is higher than a certain threshold C, then the call is judged as a sales call. If the probability of inappropriate text in the BERT model is higher than a certain threshold D, and the probability of inappropriate text in the risk keyword discrimination model is lower than a certain threshold C, then the call is judged as a normal call. A comprehensive discrimination result is given based on the results of the two models.

[0088] This multi-model fusion approach can significantly improve discrimination accuracy and reduce false positive and false negative rates.

[0089] The interception module of this system is responsible for filtering and blocking calls that meet the dynamically updated interception conditions in real time. The interception conditions can be flexibly added or removed to ensure that the system responds quickly and is highly adaptable.

[0090] The specific functions of this module are as follows:

[0091] Real-time blocking: The blocking module can receive call data in real time and filter and block calls based on a list of blocking criteria. The system supports multiple blocking methods, such as hanging up the call directly, hanging up after playing a prompt tone, or transferring the call to customer service.

[0092] 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, when the system detects that a number is frequently making harassing calls, it can automatically add the number to the blacklist and update the blocking condition list.

[0093] Dynamic blocking selection: Application users can freely select the blocking conditions provided by the module below to generate a list of combined blocking conditions. Once a match is found, the block will be blocked immediately, ensuring a rapid response to suspected risky calls.

[0094] Dynamic condition management: A rule update engine is set up to flexibly add or remove blocking conditions based on system performance indicators (false blocking rate, missed blocking rate) and business risk control strategies (such as relaxing blocking during promotional periods). For example, when encountering a batch of abnormal calls during peak periods, number frequency blocking rules can be quickly added to strengthen prevention and control.

[0095] Interception Log Recording: The interception module records detailed information for each interception operation, including call time, call duration, blocked number, and reason for interception. This information can be used for subsequent analysis and optimization.

[0096] Multiple Condition Combinations: The blocking module supports the combined use of multiple blocking conditions. For example, you can set "blacklisted number + call duration greater than 30 minutes" as a combined blocking condition to deal with excessively long telemarketing calls from historical blacklisted numbers.

[0097] The specific function of the optimization module is as follows: The professional quality inspection team selects intercepted calls for quality inspection according to the preset sampling strategy, examines the semantics and context to judge whether the interception is right or wrong, adjusts the interception rule parameters (such as relaxing the blacklist association level) and corrects the judgment labels according to the conclusions, and feeds back the quality inspection data to the call judgment sub-module. The model is optimized and new rules are accumulated using manually labeled data to enrich the system's risk control capabilities.

[0098] The specific workflow of this module is as follows:

[0099] Manual quality inspection: Manual quality inspection is the foundation of the optimization module. The system will review and analyze the blocked call records to determine whether the blocking was correct. For cases of incorrect or missed blocking, quality inspectors will mark and correct them.

[0100] Rule Adjustment: Based on the results of manual quality inspection, the system will adjust the blocking rules. For example, when a certain rule is found to be causing a large number of false blocks, the threshold or conditions of that rule can be adjusted; when a new type of sales tactic is discovered, a new rule can be added to deal with it.

[0101] Model iteration: The optimization module feeds back the correctly labeled data to the discriminant model for retraining.

[0102] (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.

[0103] (2) Dynamic adjustment of risk keyword weights: In order to adapt to the ever-changing communication environment and methods, the scores and weights of the keywords in the risk keyword database will be updated at regular intervals. The hit rate of each keyword in the keyword database during that period will be calculated using the following formula:

[0104] ;

[0105] If the keyword percentage is greater than threshold a, the keyword weight is increased (e.g., increased by 0.1); if it is less than 1, the score is increased (e.g., increased by 10). If the keyword percentage is less than threshold b, the keyword weight is decreased (e.g., decreased by 0.1), and the score is decreased (e.g., decreased by 10). If the weight change exceeds the weight range of each level, the risk level (keyword score) is updated.

[0106] The iterative training process described above can continuously improve the model's recognition accuracy and generalization ability. Furthermore, the system can introduce new deep learning models or algorithms based on actual needs to further enhance the accuracy of the judgment.

[0107] like Figure 2 As shown, one embodiment of this application discloses an interception method using the above-mentioned dynamic interception system based on dynamic rules and deep learning, comprising the following steps:

[0108] Step 1: The manual or AI risk control system sorts out business characteristics based on business experience and common characteristics of real cases to form a dynamic rule base, which stores a list of interception conditions.

[0109] In this step, based on business characteristics and historical risk data, the risk control team customizes the initial interception rule input 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.

[0110] Step 2: When a call is connected, compare it with the blocking conditions. If the conditions are met, the call is blocked. If the conditions are not met, the call is transferred to several discrimination models in the call discrimination submodule to parse the call text and then decide whether to block or not.

[0111] In this process, the moment a call is connected, the interception module retrieves a condition database for comparison. If a match is found, the call is intercepted and the details are recorded and forwarded to quality inspection. If no match is found, the recording is quickly converted to text and sent to the judgment submodule. Multiple models are used for parallel processing, and labels are assigned based on confidence level. Risk-related calls are intercepted and sent for inspection, ensuring a rapid response throughout the process and guaranteeing timeliness.

[0112] Step 3: Use a sampling strategy to rotate and intercept calls for quality inspection, and adjust the interception rule parameters and correction labels to adjust the parameters of the call discrimination submodule.

[0113] In this module, the quality inspection team analyzes intercepted samples periodically, corrects rules, calibrates judgments, updates the model weight structure, adds new rules to the database, and launches the model. This cycle repeats, allowing the system to continuously enhance its risk control efficiency through practical refinement.

[0114] The advantages of using this method are as follows:

[0115] Highly efficient and accurate interception: Dynamic interception captures emerging risks, multi-model fusion for in-depth semantic analysis, reduces false positives, improves interception accuracy, and safeguards the purity of business communications.

[0116] Continuous system evolution: Closed-loop optimization leverages quality inspection insights to refine rules, enhances model generalization capabilities, copes with complex and ever-changing business needs, and extends the effective lifespan of the system.

[0117] Scalability and Adaptability: The system design supports the dynamic addition and removal of interception rules and discrimination models, making it easy to expand to more application scenarios. Simultaneously, the system can adapt to constantly changing communication environments and propagation methods, maintaining high identification accuracy and interception efficiency.

[0118] The above technical solution effectively solves many problems existing in the current privacy number risk control and interception system, and provides strong support for building a safer and more efficient communication environment.

[0119] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0120] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that the embodiments of this application described herein can be implemented in sequences other than those illustrated or described herein.

[0121] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0122] For ease of description, spatial relative terms such as "above," "on top of," "on the upper surface of," "above," etc., are used herein to describe the spatial positional relationship of a device or feature as shown in the figures to other devices or features. It should be understood that spatial relative terms are intended to encompass different orientations in use or operation beyond the orientation of the device as described in the figures. For example, if the device in the figures were inverted, a device described as "above" or "on top of" other devices or structures would subsequently be positioned as "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both "above" and "below." The device may also be positioned in other different ways, such as rotated 90 degrees or in other orientations, and the spatial relative descriptions used herein will be interpreted accordingly.

[0123] In the detailed description above, reference has been made to the accompanying drawings, which form part of this document. In the drawings, similar symbols typically identify similar parts unless the context otherwise indicates otherwise. The illustrated embodiments described in the detailed specification, drawings, and claims are not intended 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.

[0124] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic interception system based on dynamic rules and deep learning, characterized in that, Comprise an interception condition providing module, an interception module and an optimization module, wherein: The interception condition providing module comprises a dynamic rule submodule and a call discrimination submodule, the dynamic rule submodule is formed by human or AI risk control system according to business experience and common characteristics of real cases to form a dynamic rule library, and provides a dynamically updated interception condition list for the interception module; the call discrimination submodule comprises a plurality of discrimination models for analyzing call text to determine whether interception is needed; in the call discrimination submodule, each call is outputted by a plurality of discrimination models to form a call label by fusing discrimination and confidence, and the call label is provided to the interception module as an interception clue; each call is outputted by a plurality of discrimination models to form a call label by fusing discrimination and confidence, and the specific method is as follows: The system implements recording, and converts the recording into text data by using deep learning NLP technology; The converted text data is preprocessed, at least including removing stop words, word segmentation and part-of-speech tagging; The risk keyword discrimination model and the bert model are used in parallel to identify the text data and output; the specific method is as follows: The TF-IDF values of all words in the preprocessed text data in the document set are calculated, and a TF-IDF matrix is constructed; The TF-IDF values of the bad text and the normal text are calculated, the words with TF-IDF values higher than the threshold value in the bad text are selected as bad keywords, the bad keywords are divided into high-risk keywords, medium-risk keywords and low-risk keywords, and the bad keywords are stored in the risk keyword library and assigned weights; The risk score of each call is calculated by using the following formula: wherein: k is an adjustment coefficient, taking a value between 0 and 1, for controlling the influence degree of the number of hit keywords on the call score, and N is the number of keywords in the call hit risk keyword library. It is judged whether the risk score exceeds the threshold value, if yes, the call is determined as a bad call; The output of the risk keyword discrimination model and the bert model is fused to obtain a comprehensive discrimination result as the call label; the interception module is used to real-time screen and intercept calls according to the interception conditions of the interception condition providing module; The optimization module is used to rotate the intercepted calls for quality inspection by using a sampling strategy, adjust the interception rule parameters and correct the discrimination label, and adjust the parameters of the call discrimination submodule.

2. The dynamic rule and deep learning based dynamic interception system according to claim 1, wherein, In the dynamic rule submodule, the dynamic rule library is formed according to business experience and common characteristics of real cases, specifically: the interception conditions are generated according to number behavior, number identity attribute and customer identity attribute, and the rules are formed by combining legal policies to form the dynamic rule library; the dynamic rule library can be updated in real time as needed.

3. The dynamic rule and deep learning based dynamic interception system of claim 1, wherein, The preprocessed text data is extracted by using a deepseek or qwen large model to obtain keywords, and the keywords intersecting with the screening results of the TF-IDF method are selected as the final bad keywords.

4. The dynamic rules and deep learning based dynamic interception system of claim 1, wherein, The specific method for the bert model to identify the text data and output is as follows: The text data is converted into a subword sequence T: The subword sequence T is input into the bert model to obtain context representation, and the result of the full connection layer is z, and the probability is output by the Softmax function: P(y|X) = Softmax(z) = [p0, pi] where y e {0 (normal), 1 (abnormal)}, P(y|X) indicates the probability p0 that the text X belongs to normal text and the probability pi that the text belongs to abnormal text.

5. The dynamic rule and deep learning based dynamic interception system of claim 1, wherein, The specific method for 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, if the risk keyword discrimination model and the bert model are both determined as abnormal calls, then the call is marked as an abnormal call call label.

6. The dynamic rules and deep learning based dynamic interception system of claim 1, wherein, The specific working steps of the optimization module are: Review and analyze the intercepted call records, and label and correct the misinterception or missed interception; According to the labeling and correction, adjust the interception rule and retrain the model, including optimizing the bert model and dynamically adjusting the risk keyword weight.

7. An intercepting method employing the dynamic rule and deep learning based dynamic intercepting system according to any one of claims 1 to 6, characterized in that, Comprise the following steps: Step 1: The artificial or AI risk control system forms a dynamic rule library according to business experience and common characteristics of real cases, and the dynamic rule library stores an interception condition list; Step 2: When the call is accessed, compare the interception conditions. If it hits, intercept the call. If it does not hit, go to the call discrimination submodule, analyze the call text, and then intercept or not intercept; Step 3: Rotate the intercepted calls for quality inspection using a sampling strategy, adjust the interception rule parameters, and correct the discrimination label to adjust the parameters of the call discrimination submodule.

Images