Intelligent anti-fraud information construction system based on telephone traffic information

Abstract

The invention relates to the technical field of telephone traffic information identification, in particular to an intelligent anti-fraud information construction system based on telephone traffic information, and the system comprises the steps: obtaining fraud call information and historical fraud call information, and constructing an anthropomorphic telephone traffic virtual body; the anthropomorphic telephone traffic virtual body performs a session according to the fraud telephone information and generates fraud session content; performing session voice expectation analysis according to the session fraud content to generate response session content and instructing a preset session response model to perform a session according to the session content; performing semantic recognition analysis on the response session content and generating a dialogue result deviation value; according to the dialogue result deviation value, correcting a preset dialogue response model, and generating a deviation element; performing keyword marking on each telephone traffic content, and identifying and generating an abnormal keyword sequence; and substituting the abnormal keyword sequence into the anti-fraud feature recognition model to correct the corresponding weight. The method and the device have the effect of improving fraud early warning of the dialogue traffic information.

Description

Technical Field

[0001] This application relates to the technical field of call information recognition, and in particular to an intelligent anti-fraud information construction system based on call information. Background Art

[0002] With the popularization of communication and Internet technologies, fraudsters use online platforms, mobile terminals, etc. to carry out fraud activities. Traditional anti-fraud means mainly rely on manual publicity and case warnings. With the diversification and concealment of fraud methods, traditional means are difficult to meet the prevention needs, so an intelligent anti-fraud system based on call data analysis has emerged.

[0003] In related technologies, when analyzing fraud information, the call intelligent anti-fraud system often focuses on simple keyword matching, behavior pattern comparison, etc., so as to analyze whether there is a possibility of fraud in the call, and then issue a fraud warning prompt for the call. However, it lacks the analysis of complex and changeable language behavior.

[0004] In view of the above related technologies, when lacking the analysis of complex and changeable language behavior, the call intelligent fraud system can only prevent some fraud means that have been used in the past, but cannot effectively identify the continuously developing fraud call information, which is not conducive to improving the anti-fraud effect. Summary of the Invention

[0005] In order to improve the fraud warning effect for call information, this application provides an intelligent anti-fraud information construction system based on call information.

[0006] In a first aspect, this application provides an intelligent anti-fraud information construction system based on call information, adopting the following technical solution:

[0007] An intelligent anti-fraud information construction system based on call information, comprising:

[0008] Step 1, obtain fraud call information and historical fraud call information, and extract fraud features according to the historical fraud call information to construct an anthropomorphic call virtual entity;

[0009] Step 2, the anthropomorphic call virtual entity conducts a conversation according to the fraud call information and generates fraud conversation content;

[0010] Step 3, conduct a conversation voice expectation analysis according to the conversation fraud content to generate a response conversation content, and instruct a preset conversation response model to conduct a conversation according to the conversation content;

[0011] Step 4, conduct a semantic recognition analysis on the response conversation content and generate a conversation result deviation value;

[0012] Step 5: Modify the preset conversation response model according to the conversation result deviation value, and generate deviation factors for feedback adjustment;

[0013] Step 6: Mark keywords for each traffic content, and identify and generate an abnormal keyword sequence;

[0014] Step 7: Input the abnormal keyword sequence into the anti-fraud feature recognition model to modify the corresponding weights.

[0015] Optionally, when performing conversation voice expectation analysis in Step 3, it also includes:

[0016] Perform voice recognition analysis on the conversation to obtain conversation voice features, and analyze the speech rate, tone, and response pause ratio coefficient of the conversation voice features;

[0017] Sort the weight values based on the speech rate, tone, and response pause ratio coefficient and generate a conversation response weight value;

[0018] Adjust and optimize the content response ratio of the conversation content in the conversation response model based on the conversation response weight value.

[0019] Optionally, when analyzing the speech rate, tone, and response pause ratio coefficient of the conversation voice features, it also includes:

[0020] Analyze the conversation framework of the conversation voice content and compare it with the preset historical fraud conversation framework to determine the conversation framework similarity;

[0021] When the conversation framework similarity is less than the preset similarity of the same type of framework, match the extension duration corresponding to the conversation framework similarity in the preset duration adjustment database;

[0022] Based on the extension duration adjustment, instruct the preset conversation response model to increase the expected conversation content;

[0023] When the conversation framework similarity is greater than or equal to the similarity of the same type of framework, instruct the preset conversation response model to reduce the conversation content and end the conversation according to the preset conversation ending strategy.

[0024] Optionally, when the conversation response model in Step 3 conducts a conversation according to the conversation content, it also includes:

[0025] Extract gender, age, and voice color characteristics according to the conversation object and generate conversation personnel information;

[0026] Compare the conversation personnel information with the historical conversation personnel information, and screen out different preset conversation role parameters of the traffic virtual entity;

[0027] Adjust the anthropomorphic traffic virtual entity based on the screened conversation role parameters.

[0028] Optionally, when adjusting the anthropomorphic call virtual entity, it further includes:

[0029] Statistically analyze the number of session fraud feature acquisitions for different session role parameters to determine the value of fraud session feature extraction;

[0030] Sort based on the session role parameters corresponding to the fraud session feature extraction rate to determine the session role parameter with the highest fraud session feature extraction value and mark it as the priority session role parameter;

[0031] Adjust the session role parameters of the anthropomorphic call virtual entity according to the priority session role parameters.

[0032] Optionally, when obtaining fraud call information in step 1, it further includes:

[0033] Instruct the preset information screening module to classify according to the preset group characteristics to obtain classified fraud call information;

[0034] Perform a classification feature consistency check on the classified fraud call information and analyze the number of similar fraud call information;

[0035] Based on the comparison and analysis of the number of similar fraud call information and the preset quantity condition, issue an anthropomorphic call virtual entity construction instruction when the preset quantity condition is met.

[0036] Optionally, when generating fraud session content in step 2, it further includes:

[0037] Extract dialogue structure features based on the dialogue content, including dialogue turns, topic transitions, and session duration features;

[0038] Analyze based on the dialogue structure features and the preset fraud session intensive time analysis strategy to determine the fraud session intensive time period;

[0039] Increase the number of sessions of the anthropomorphic call virtual entity based on the fraud session intensive time period.

[0040] Optionally, the fraud session intensive time analysis strategy includes:

[0041] Analyze the comparison between the dialogue turn, topic transition, and session duration features and the preset benchmark dialogue features to obtain the dialogue turn suspicious score, topic transition suspicious score, and time suspicious score;

[0042] Based on the dialogue turn suspicious score, topic transition suspicious score, and time suspicious score, perform weighted calculation to obtain the comprehensive suspicious score;

[0043] When the comprehensive score exceeds the preset benchmark suspicious score, perform time analysis on the session and form a fraud intensive session time period mark.

[0044] Optionally, step 7 further includes:

[0045] Constructing a fraud knowledge graph based on the abnormal keyword sequence and a preset recognition module;

[0046] Analyzing the conversation content based on the fraud knowledge graph to determine new fraud patterns;

[0047] Adding the new fraud patterns to a preset deep learning semantic understanding model, and analyzing the implicit intention of the response content to determine implicit fraud features;

[0048] Based on the implicit fraud features, correcting the conversation content of a preset conversation response model to generate implicit deviation elements;

[0049] Based on the implicit deviation elements, adjusting the feedback of the conversation content of the preset conversation response model.

[0050] In summary, the present application includes at least one of the following beneficial technical effects:

[0051] 1. Establishing an anthropomorphic call virtual entity and conducting different fraud conversations, and analyzing the fraud conversation content to form a deviation value of the dialogue result for correcting the conversation response model, so as to further improve the accuracy of the anthropomorphic call virtual entity in identifying fraud conversation features, which helps to increase the reliability of fraud early warning of call information. Compared with traditional anti-fraud publicity, it can timely identify whether there is a fraud activity during the call, so that people are not easily caused economic losses due to call fraud;

[0052] 2. The continuous correction of the model improves the recognition accuracy of the system for fraud conversation techniques. The keyword recognition assisted by the knowledge graph can discover complex fraud patterns, enhancing the adaptability and accuracy of the system;

[0053] 3. The optimized virtual entity can more realistically simulate fraud conversation techniques, improving the quality of conversation generation, providing a more reliable data basis for subsequent conversation analysis, and enhancing the system's understanding and response capabilities for fraud conversation techniques. Description of the Drawings

[0054] Figure 1 is a flowchart of the method of steps S100 to S700 in the present application.

[0055] Figure 2 is a flowchart of the method of steps S301 to S303 in the present application.

[0056] Figure 3 is a flowchart of the method of steps S304 to S307 in the present application.

[0057] Figure 4 is a flowchart of the method of steps S308 to S310 in the present application.

[0058] Figure 5 is the flowchart of the method from step S101 to S103 in this application.

[0059] Figure 6 is the flowchart of the method from step S104 to S206 in this application.

[0060] Figure 7 is the flowchart of the method from step S201 to S203 in this application.

[0061] Figure 8 is the flowchart of the method from step S2021 to S2023 in this application.

[0062] Figure 9 is the flowchart of the method from step S701 to S705 in this application. Detailed implementation manners

[0063] In order to make the purpose, technical solutions and advantages of this application clearer, the following further describes this application in detail with reference to the attached Figures 1-9 drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not used to limit this application.

[0064] The following further describes the embodiments of the present invention in detail with reference to the drawings of the specification.

[0065] The embodiments of this application disclose an intelligent anti-fraud information construction system based on traffic information. By constructing an anthropomorphic traffic virtual body for fraud call information and enabling the anthropomorphic traffic virtual body to conduct continuous fraud conversations according to historical fraud call information, and correcting the anthropomorphic traffic virtual body to optimize and enhance the recognition of fraud content, thereby improving the recognition accuracy of fraud conversations.

[0066] Referring to Figure 1 , the method flow of the intelligent anti-fraud information construction system based on traffic information includes the following steps:

[0067] Step S100: Obtain fraud call information and historical fraud call information, and extract fraud features according to historical fraud call information to construct an anthropomorphic traffic virtual body;

[0068] The fraud call information is obtained through the fraud call database interface established with the operator, and the corresponding marked fraud phone numbers are obtained. The historical fraud call information is the fraud conversation voice data stored in the historical traffic database. By extracting the fraud voice features in the fraud conversation voice data, a virtual conversation intelligent body for identifying relevant fraud conversations can be established. The virtual conversation intelligent body is defined as an anthropomorphic traffic virtual body and implanted into the virtual conversation module for making calls and having conversations with fraud phones, so as to collect the corresponding conversation content as samples for optimizing the model.

[0069] When constructing an anthropomorphic call virtual entity, a hybrid neural network architecture can be used as a framework to implant BiLSTM+CNN (BiLCNN): the bidirectional LSTM extracts temporal features (such as the inductive dialogue process in fraud calls), and the CNN captures local keywords (such as high-frequency fraud terms like "secure account" and "transfer"). The Transformer-CNN hybrid model: uses the Transformer encoder to extract long-range dependence features (such as fraud logic across multiple sentences), combines the local feature extraction ability of the CNN, and classifies after fusing through vector concatenation. Attention mechanism enhancement: Add an attention layer (BiLSTM-Attention) after the BiLSTM, or introduce a self-attention mechanism after the CNN to strengthen the weight of key fraud features (such as descriptions of time urgency). By feeding a large amount of data and training the model with the hybrid neural network architecture for dialogue and feature recognition, the dialogue effect of the anthropomorphic call virtual entity is made close to the dialogue logic of the deceived.

[0070] In addition, when conducting model training, data collection can cooperate with telecommunications operators to establish a real-time data interface, collect a large number of call records daily, and at the same time obtain detailed information on a certain number of fraud calls from the reporting platform. At the same time, text information related to fraud, such as fraud dialogue discussion posts and victims' self-narratives, can be collected through the social media data interface, and fraud text message data can be obtained from the SMS gateway.

[0071] Step S200: The anthropomorphic call virtual entity conducts a conversation based on the fraud call information and generates fraud conversation content;

[0072] The anthropomorphic call virtual entity makes a call according to the fraud call and generates voice data and text data for the conversation content, and generates fraud conversation content from the voice data and text data for subsequent call and analysis.

[0073] Step S300: Conduct a conversation voice expectation analysis based on the conversation fraud content to generate a response conversation content and instruct a preset conversation response model to conduct a conversation according to the conversation content;

[0074] By conducting a semantic expectation analysis on the fraud conversation content, it can be known what the response content expected by the conversation partner is. For example, when content related to winning a prize and information registration appears in the conversation, the semantic expectation in the conversation is to obtain the identity and property-related information of the person, so that the conversation response model generates the response conversation content to enable the conversation to continue.

[0075] Step S400: Conduct a semantic recognition analysis on the response conversation content and generate a conversation result deviation value;

[0076] By performing semantic recognition on the response session content, analyzing the semantic overlap between the generated conversation content and the conversation content for voice expectation, it can be determined whether the response conversation content meets the expectation. The degree of compliance is used as the deviation value of the conversation result. When the deviation value of the conversation result is larger, it indicates that the generated conversation content does not meet the expectation, and the authenticity of the anthropomorphic call virtual entity is relatively poor. On the contrary, it indicates that the conversation authenticity effect of the anthropomorphic call virtual entity is good.

[0077] Step S500: Modify the preset session response model according to the deviation value of the conversation result, and generate deviation elements for feedback adjustment;

[0078] The session response model is a session automatic generation module that generates conversations based on voice expectation analysis. It is pre-embedded in the anthropomorphic call virtual entity by the staff, and forms deviation content elements by screening the deviation content between the conversation content and the voice expectation content. This element is defined as the deviation element, and the session response model is fed back according to the deviation element, so that the session response model can supplement the same type of conversation content in the deviation element.

[0079] Step S600: Mark keywords for each call content and identify and generate an abnormal keyword sequence;

[0080] By detecting and marking fraud keywords in the text of the call content, screening abnormal keyword combinations, and generating an abnormal keyword sequence, it helps to improve the recognition accuracy of different fraud characteristics.

[0081] Step S700: Substitute the abnormal keyword sequence into the anti-fraud feature recognition model to correct the corresponding weights.

[0082] By substituting the abnormal keyword sequence into the anti-fraud feature recognition model, the recognition content of the anti-fraud feature recognition model for fraud conversation features is expanded, so that the recognition weight of the anti-fraud feature recognition model for abnormal keywords increases, thereby improving the recognition accuracy.

[0083] In addition, referring to Figure 2 , when performing session voice expectation analysis in step 3, it also includes:

[0084] Step S301: Perform voice recognition analysis on the session to obtain session voice features, and analyze the speech rate, intonation, and response pause ratio coefficient of the session voice features;

[0085] Pre-build a voice recognition analysis network model and implant it into the anthropomorphic call virtual entity, so that the anthropomorphic call virtual entity can analyze the speech rate, intonation, and response pause of the voice in the session, and form a speech rate ratio coefficient, an intonation ratio coefficient, and a response pause ratio coefficient, which helps to identify whether the session object has an induced tendency to obtain financial and personal privacy information.

[0086] Step S302: Sort the weight values based on the speech rate, intonation, and response pause ratio coefficient, and generate the session response weight value;

[0087] By analyzing the speech rate ratio coefficient, intonation ratio coefficient, and response pause ratio coefficient in the session voice, the proportion weight of the three can be obtained. Sorting the proportion weight can obtain the emphasis value of the session object on the session content, and this emphasis value is used as the session response weight value.

[0088] Step S303: Adjust and optimize the content response ratio of the session content of the session response model based on the session response weight value.

[0089] Input the session response weight value into the session response model to adjust the proportion adjustment of the generated session content, so that the responded session content is more in line with the conversation logic and expectations, can increase the conversation duration, and helps to dig out more fraud session content.

[0090] Refer to Figure 3 , when analyzing the speech rate, intonation, and response pause ratio coefficient of the session voice characteristics, it also includes:

[0091] Step S304: Analyze the dialogue framework of the session voice content and compare it with the preset historical fraud session framework to determine the dialogue framework similarity;

[0092] Using the pre-built recurrent neural network as the analysis model to analyze the framework structure of the dialogue in the session voice content, the dialogue framework can be obtained, and the dialogue framework is compared with the historical fraud session framework to further verify the fraud in the session content of the fraud call.

[0093] Step S305: When the dialogue framework similarity is less than the preset similarity of the same type of framework, match the extended duration corresponding to the dialogue framework similarity in the preset duration adjustment database;

[0094] The similarity of the same type of framework indicates that the coincidence degree of the two dialogue frameworks is within a consistent range. When the dialogue framework similarity is less than the similarity of the same type of framework, it indicates that the two dialogue frameworks belong to different types, which means that the fraud session belongs to a new dialogue framework and is a valuable reply. By extending the conversation duration, more fraud session characteristics of the new fraud session framework can be further extracted. Among them, the duration adjustment database is a pre-built extended duration database, which stores different dialogue framework similarities and the corresponding extended durations, and automatically matches and outputs the corresponding extended duration when the dialogue framework similarity is input.

[0095] Step S306: Based on the extended duration adjustment, instruct the preset session response model to increase the expected session content;

[0096] Increase the expected conversation content according to the conversation response model with the indicated delay duration, so as to increase the duration of the conversation, obtain sufficient conversation content as an analysis sample, and further improve the recognition effect of the fraud conversation characteristics of the anthropomorphic call virtual entity.

[0097] Step S307: When the conversation frame similarity is greater than or equal to the similarity of the same type of frame, instruct the preset conversation response model to end the conversation by reducing the conversation content according to the preset conversation ending strategy.

[0098] When the conversation frame similarity is equal to or greater than the similarity of the same type of frame, it indicates that the compared conversation does not belong to the content of a new type of fraud conversation. As a conversation frame that has been recognized, it does not have a high conversation value. Therefore, by instructing the conversation response model to reduce and end the conversation, new conversations can be launched, which helps to improve the efficiency of optimizing the conversation of the anthropomorphic call virtual entity.

[0099] Refer to Figure 4 , when the conversation response model conducts a conversation according to the conversation content in step 3, it also includes:

[0100] Step S308: Extract gender, age, and voice color characteristics based on the conversation object and generate conversation personnel information;

[0101] During the conversation process, the conversation response model extracts the gender, age, and voice color characteristics of the conversation object and generates conversation personnel information for subsequent further analysis.

[0102] Step S309: Compare based on the conversation personnel information and the historical conversation personnel information, and screen out different preset conversation role parameters of the call virtual entity;

[0103] The anthropomorphic call virtual entity has different conversation role parameters, enabling the anthropomorphic call virtual entity to conduct conversations with different virtual roles, thereby simulating the conversations and psychology of different people as needed, and thus being able to more accurately prompt the responses during conversations with different groups of people, warning the anthropomorphic call virtual entity against information leakage behaviors that are likely to occur among different groups of people, and reducing the possibility of the corresponding personnel group being scammed by calls.

[0104] Therefore, by establishing a conversation role parameter database, when the anthropomorphic call virtual entity conducts a conversation, it analyzes the conversation object of the fraud call. When comparing the gender, age, and voice color of the conversation object with the historical conversation personnel, different role parameters are screened out for the conversation to increase the comprehensiveness of collecting fraud conversation test samples for different groups.

[0105] Step S310: Adjust the anthropomorphic call virtual entity based on the screened conversation role parameters.

[0106] When the corresponding session role parameters are screened out, the anthropomorphic call virtual entity is adjusted with the corresponding role parameters.

[0107] Refer to Figure 5 , in step 1, when adjusting the anthropomorphic call virtual entity, it further includes:

[0108] Step S101: Statistically analyze the number of session fraud feature acquisitions for different session role parameters to determine the value of fraud session feature extraction;

[0109] Extract fraud features from the session content during the conversation with different session role parameters, count the number of acquired features for proportion calculation, and use the calculation result as the value of fraud session feature extraction for further analysis and invocation.

[0110] Step S102: Sort the session role parameters based on the fraud session feature extraction rate to determine the session role parameter with the highest fraud session feature extraction value and mark it as the priority session role parameter;

[0111] Sort different session role parameters according to the fraud session feature extraction rate, and mark the session role parameter with the highest fraud session feature extraction value. When the fraud session feature extraction value is higher, it means that the corresponding population group is more likely to be the target of fraud, so the session role parameter with the highest value is marked as the priority session role parameter.

[0112] Step S103: Adjust the session role parameters of the anthropomorphic call virtual entity according to the priority session role parameters.

[0113] By using the priority session role parameters as the session role parameters of the anthropomorphic call virtual entity for conversation, the recognition effect of current frequent fraud sessions can be improved, which helps to collect the session content of current main fraud calls for the optimization training of the anthropomorphic call virtual entity.

[0114] Refer to Figure 6 , when obtaining fraud call information in step 1, it further includes:

[0115] When obtaining fraud call information in step 1, it further includes:

[0116] Step S104: Instruct the preset information screening module to classify according to the preset group characteristics to obtain classified fraud call information;

[0117] The preset group characteristics are groups of personnel with different work types, which are classified according to different occupational staff and ages, and the corresponding group characteristics are constructed, so that the fraud call information can be classified according to different personnel groups, so that the anthropomorphic call virtual entity can quickly conduct conversations according to different categories of fraud calls and collect fraud conversation characteristics, improving the recognition efficiency of fraud calls received by the same group.

[0118] Step S105: Perform a classification feature consistency check on the classified fraud call information and analyze the number of fraud call information of the same type;

[0119] By analyzing and verifying the classification consistency of the classified fraud call information, it helps to improve the classification accuracy. In addition, by analyzing the number of fraud call information of the same type, it can be known whether there is sufficient fraud call information for the anthropomorphic call virtual entity to conduct conversations.

[0120] Step S106: Based on the comparison and analysis of the number of fraud call information of the same type and the preset quantity condition, when the preset quantity condition is met, issue an anthropomorphic call virtual entity construction instruction.

[0121] The preset quantity condition represents the benchmark quantity for the anthropomorphic call virtual entity to conduct conversations. When conducting conversations according to this benchmark quantity, sufficient conversation samples can be obtained. Therefore, when the number of fraud call information of the same type meets or is greater than the preset quantity condition, an anthropomorphic call virtual entity construction instruction is issued to construct an anthropomorphic call virtual entity for conversations and make calls according to the obtained fraud call information of the same type.

[0122] The information acquisition module is a pre-constructed fraud call information screening model. By connecting to an external media data interface, it can receive fraud call information data stored in the network media terminal. By screening and classifying the received fraud call information, it can screen out the deceived calls of different regions and identity groups, enabling the fraud call information to be distinguished according to different categories, and enabling centralized conversations with the fraud calls deceived by specific groups, so as to obtain more fraud conversation details and improve the comprehensiveness of fraud recognition of the anthropomorphic conversation virtual entity.

[0123] Refer to Figure 7 , when generating fraud conversation content in step 2, it also includes:

[0124] Step S201: Extract dialogue structure features based on the dialogue content, including dialogue turns, topic transitions, and conversation duration features;

[0125] Extract the conversation structure features through the recurrent neural network embedded in the anthropomorphic conversation virtual entity, so as to extract the dialogue turn features, topic transition features, and conversation duration features in the conversation, and the corresponding language structure features constitute the fraud conversation structure features.

[0126] Step S202: Analyze based on the dialogue structure features and a preset fraud conversation intensive time analysis strategy to determine the fraud conversation intensive time period;

[0127] The fraud conversation intensive time period analysis strategy is a preset time period analysis method for analyzing fraud conversations. The specific analysis method will be further elaborated in the subsequent steps. By analyzing the fraud conversation intensive time period, it is possible to know the time when a conversation can be effectively carried out with a fraud call during a conversation call, improving the efficiency of obtaining fraud conversation features and data.

[0128] Step S203: Increase the number of conversations of the anthropomorphic call virtual entity based on the fraud conversation intensive time period.

[0129] By increasing the number of conversations of the anthropomorphic conversation virtual entity during the fraud conversation intensive time period, the data acquisition volume for fraud conversations is increased.

[0130] Refer to Figure 8 , the fraud conversation intensive time analysis strategy includes:

[0131] Step S2021: Analyze the conversation turn, topic conversion, and conversation duration features and compare them with the preset benchmark conversation features to obtain the conversation turn suspicious score, topic conversion suspicious score, and time suspicious score;

[0132] The basic conversation features are the conversation feature ratios stored in a pre-established fraud conversation feature database and corresponding suspicious scores are stored. By analyzing the conversation turn, topic conversion, and conversation duration feature ratios between both parties in a conversation and comparing them with the basic conversation feature ratios, the corresponding conversation turn suspicious score, topic conversion suspicious score, and time suspicious score can be found.

[0133] Step S2022: Calculate the weighted comprehensive suspicious score based on the conversation turn suspicious score, topic conversion suspicious score, and time suspicious score;

[0134] By summing up the conversation turn suspicious score, topic conversion suspicious score, and time suspicious score and taking the average value, this average value is defined as the comprehensive suspicious score, which is used to represent the comprehensive performance of multiple weighted item deliberate scores.

[0135] Step S2023: When the comprehensive score exceeds the preset benchmark suspicious score, perform time analysis on the conversation and form a fraud intensive conversation time period mark.

[0136] The benchmark suspicious score represents the suspicious scores of the weights corresponding to the session structure features in multiple aspects of the session, and marks the corresponding time period. Different sessions will generate different fraud session times. By recording and marking a large number of different fraud session time periods, a more accurate fraud session intensive occurrence time can be obtained, enabling the anthropomorphic traffic virtual entity to make call sessions during the intensive occurrence of fraud sessions, thereby increasing the probability of successful call sessions.

[0137] Reference Figure 9 , Step 7 also includes the following steps:

[0138] Step S701: Construct a fraud knowledge graph based on the abnormal keyword sequence and the preset recognition module;

[0139] By constructing a fraud knowledge graph, a fraud recognition module is embedded in the fraud knowledge graph to identify different fraud contents.

[0140] Step S702: Analyze the fraud patterns of the conversation content based on the fraud knowledge graph to determine new fraud patterns;

[0141] Through the corresponding conversation content recognition and analysis by the fraud knowledge graph, the fraud types in the conversation content are identified. When the fraud type is inconsistent with the historical fraud type, it indicates the existence of a new fraud type. Then, the fraud session content is marked, and a conversation structure record of the fraud pattern is generated as a reference for the new fraud pattern.

[0142] Step S703: Add the new fraud pattern to the preset deep learning semantic understanding model, and perform implicit intention analysis on the response content to determine implicit fraud features;

[0143] By integrating and adding the new fraud pattern to the preset deep learning semantic understanding model, the anthropomorphic traffic virtual entity can understand the new fraud session features. At the same time, perform implicit intention analysis on the response content collected by the anthropomorphic traffic virtual entity to collect implicit fraud features for subsequent calls.

[0144] Step S704: Based on the implicit fraud features, correct the conversation content of the preset conversation response model to generate implicit deviation elements;

[0145] By inputting the implicit fraud features into the conversation response model, the conversation response model adjusts the conversation content according to the implicit fraud features. At the same time, by comparing the adjusted content with the initial conversation content, the missing response content can be known, and the missing response content is marked and used as text data for implicit deviation elements for calls.

[0146] Step S705: Adjust the session content feedback of the preset session response model based on the implicit deviation factors.

[0147] By inputting the implicit deviation factors into the session response model, the session response model further adjusts the generation ratio of the content corresponding to the implicit deviation factors, so that the response content can better meet the actual needs.

[0148] And a risk warning module, which, according to the recognition result of the anti-fraud feature recognition model, when detecting high-risk fraud features, sends warning information to users or relevant institutions. The warning information includes fraud types, risk levels, and possible loss estimates.

[0149] Meanwhile, the information acquisition module is also used to access text information related to fraud and fraud SMS data on social media, develop corresponding data interfaces, and establish a data cleaning and integration mechanism to ensure the quality and consistency of the data.

[0150] The anthropomorphic call virtual entity construction module adopts the introduction of sentiment analysis technology to analyze the sentiment tendency in historical fraud texts and integrates it into the dialogue generation model of the virtual entity. At the same time, it uses reinforcement learning technology to enable the virtual entity to continuously optimize its conversation skills during multiple conversations with the preset "opponent".

[0151] The session analysis module integrates a real-time speech recognition module to analyze the speech of fraud calls in real time, extract speech features, and combine these features with text information for joint session speech expectation analysis.

[0152] The model correction module adopts a more complex semantic analysis model, such as a semantic understanding model based on deep learning, to perform multi-dimensional analysis on the response content, and combines pragmatics theory to construct a pragmatic analysis module to identify implicit intentions in the dialogue.

[0153] The keyword marking and recognition module uses knowledge graph technology to construct a knowledge graph related to fraud, integrates keywords and their semantic relationships and logical relationships into the graph, and mines complex fraud patterns through graph analysis.

[0154] The anti-fraud feature recognition model correction module develops a feature extraction module to extract dialogue structure features from dialogue data, such as the number of dialogue turns, topic transitions, and session duration features, such as the time pattern of the appearance of fraud calls, and adopts an ensemble learning framework, such as random forest, gradient boosting tree, etc., to fuse the prediction results of multiple models.

[0155] It also includes a risk warning module. Based on the recognition results of the anti-fraud feature recognition model, when detecting high-risk fraud features, the risk warning module sends warning messages to users or relevant institutions. The warning messages include fraud types, risk levels, and possible loss estimates.

[0156] Among them, it should be further explained that for data cleaning and integration: using data cleaning algorithms to remove duplicate, incorrect, and incomplete data, unifying and standardizing the formats of data from different sources, establishing a data warehouse, and integrating and storing various types of data for convenient subsequent analysis.

[0157] Through the acquisition and integration of multi-source data, the data dimension is greatly enriched, the recognition ability of fraud behaviors is improved, more potential fraud patterns can be discovered, and the risk of missed judgments is reduced.

[0158] When constructing and optimizing the anthropomorphic call virtual entity, for feature extraction and model construction: using the TF-IDF algorithm to extract keywords from tens of thousands of historical fraud call information, obtaining hundreds of keywords that meet the expected quantity, such as "transfer", "handling fee", "unfreeze", etc. Based on these keywords, using a neural network to construct an anthropomorphic call virtual entity model for preliminary training.

[0159] For sentiment analysis and reinforcement learning: introducing sentiment analysis tools to conduct sentiment tendency analysis on historical fraud texts and integrating sentiment features into the virtual entity model.

[0160] When generating a conversation, when obtaining a new fraud call message, such as "You have won a huge prize and need to pay a handling fee first to receive the bonus", the anthropomorphic call virtual entity generates a conversation according to the learned conversation patterns: "Customer service: The bonus amount is huge. Please pay the handling fee as soon as possible and we will handle the prize collection for you immediately. Victim: Could it be a scam? Customer service: We are a regular institution with government certification. How could we deceive you?"

[0161] When analyzing the conversation: The real-time speech recognition module converts the conversation speech into text, combines the text information, and analyzes the speech features. For example, the customer service has a fast speaking speed, an urgent tone, and the statement of "government certification" is misleading. Through semantic and pragmatic analysis, it is judged that there is a fraud risk in this conversation. The combined analysis of real-time speech and text improves the accuracy of conversation analysis, can timely discover the key features of fraud conversation patterns, and provides strong support for subsequent model correction and risk warning.

[0162] Model correction and keyword recognition: When performing model correction, compare the generated response session content with the preset correct dialogue result to calculate the deviation value. For example, in the preset dialogue, the customer service should not mention "government certification" to induce the user. Use the gradient descent algorithm to correct the session response model and adjust the model parameters. Keyword recognition: Use the TextRank algorithm to extract dialogue keywords, such as "grand prize", "handling fee", "government certification", and analyze the semantic and logical relationships of the keywords in combination with the knowledge graph to identify that this abnormal keyword sequence conforms to the false winning scam pattern. The continuous correction of the model improves the system's recognition accuracy of scam words, and the keyword recognition assisted by the knowledge graph can detect complex scam patterns, enhancing the system's adaptability and accuracy.

[0163] Anti-fraud feature recognition model correction and risk warning include: Feature extraction and model fusion: Extract features such as the number of dialogue rounds being 3, the topic always revolving around the prize and the handling fee, and such scam calls mostly occurring between 10 am and 12 pm from the dialogue data. Use the random forest algorithm to fuse multiple anti-fraud feature recognition models to improve the recognition accuracy. Risk warning: When the anti-fraud feature recognition model detects high-risk fraud features, send a warning text message to the user's mobile phone, the content of which includes "Detected false winning scam risk, high risk level, may result in significant losses, please be cautious about winning information mentioned in strange calls", and at the same time send a warning message to the public security anti-fraud department.

[0164] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the above-mentioned division of each functional module is used as an example. In actual applications, the above functions can be allocated to different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. The specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0165] The embodiment of the present invention provides a computer-readable storage medium storing a computer program that can be loaded and executed by a processor for an intelligent anti-fraud information construction system based on traffic information.

[0166] Computer storage media include, for example: USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs and other media that can store program codes.

[0167] Based on the same inventive concept, the embodiment of the present invention provides an intelligent terminal, including a memory and a processor, and a computer program that can be loaded and executed by the processor for an intelligent anti-fraud information construction system based on traffic information is stored on the memory.

[0168] Those skilled in the art can clearly understand that, for the convenience and conciseness of description, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. For the specific working processes of the system, device, and unit described above, reference can be made to the corresponding processes in the foregoing method embodiments, which will not be elaborated here.

[0169] The above are all preferred embodiments of the present application. The protection scope of the present application is not limited thereby. Any feature disclosed in this specification (including the abstract and drawings), unless specifically described, can be replaced by other equivalent or similar-purpose alternative features. That is, unless specifically described, each feature is only an example of a series of equivalent or similar features.

Claims

1. An intelligent anti-fraud information construction system based on traffic information, characterized in that: include: Step 1, obtaining fraudulent call information and historical fraudulent call information, and extracting fraud features based on the historical fraudulent call information to construct an anthropomorphic call virtual body; Step 2, the anthropomorphic call virtual body conducts a conversation based on the fraudulent call information and generates fraudulent conversation content; Step 3, performing conversation voice expectation analysis based on the conversation fraud content to generate response conversation content and instructing a preset conversation response model to conduct a conversation based on the conversation content; Step 4, performing semantic recognition analysis on the response conversation content and generating a conversation result deviation value; Step 5, modifying the preset conversation response model according to the dialogue result deviation value, and generating deviation elements for feedback adjustment; Step 6, tagging each call content with keywords and identifying and generating abnormal keyword sequences; Step 7: Bring the abnormal keyword sequence into the anti-fraud feature recognition model to correct the corresponding weights.

2. The intelligent anti-fraud information construction system based on traffic information according to claim 1 is characterized in that: When performing conversation voice expectation analysis in step 3, it also includes: Perform speech recognition analysis on the conversation to obtain the conversation speech features, and analyze the speech rate, intonation and response pause ratio coefficient of the conversation speech features; The weight values ​​are sorted based on the speech rate, intonation and response pause ratio coefficient and a conversation response weight value is generated; The content response ratio of the conversation content of the conversation response model is adjusted and optimized based on the conversation response weight value.

3. The intelligent anti-fraud information construction system based on traffic information according to claim 2 is characterized in that: When analyzing the speech rate, intonation, and response pause ratio of conversational speech features, the following factors are also included: Analyze the conversation frame of the conversation voice content and compare it with the preset historical fraud conversation frame to determine the similarity of the conversation frame; When the dialogue frame similarity is less than the preset similarity of the same frame, the extended time corresponding to the dialogue frame similarity in the database is adjusted by matching the preset time length; Increasing the expected conversation content based on the preset conversation response model of the extended duration adjustment indication; When the dialogue frame similarity is greater than or equal to the similarity of the same frame, the preset conversation response model is instructed to reduce the conversation content and end the conversation according to the preset conversation ending strategy.

4. The intelligent anti-fraud information construction system based on traffic information according to claim 1 is characterized in that: When the conversation response model in step 3 conducts a conversation according to the conversation content, it also includes: Extract gender, age, and voice features based on the conversation partner and generate conversation partner information; Based on the comparison between the conversation person information and the historical conversation person information, different preset conversation role parameters of the traffic virtual body are screened out; The anthropomorphic conversation virtual body is adjusted based on the filtered conversation role parameters.

5. The intelligent anti-fraud information construction system based on traffic information according to claim 4 is characterized in that: When adjusting the anthropomorphic traffic virtual body, it also includes: Conduct statistical analysis on the number of session fraud features obtained for different session role parameters to determine the value of extracting fraud session features; Sorting the session role parameters corresponding to the fraud session feature extraction rate to determine the session role parameter with the highest fraud session feature extraction value and marking it as the priority session role parameter; The conversation role parameters of the anthropomorphic traffic virtual body are adjusted according to the priority conversation role parameters.

6. The intelligent anti-fraud information construction system based on traffic information according to claim 1 is characterized in that: When obtaining fraudulent phone information in step 1, it also includes: Instructing a preset information screening module to classify according to preset group characteristics to obtain classified fraudulent phone call information; Perform consistency check on classified fraudulent phone call information and analyze the number of similar fraudulent phone call information; A comparative analysis is conducted based on the number of similar fraudulent phone calls and the preset number conditions, and when the preset number conditions are met, an instruction to construct an anthropomorphic call virtual body is issued.

7. The intelligent anti-fraud information construction system based on traffic information according to claim 1 is characterized in that: When generating fraudulent session content in step 2, it also includes: Extract dialogue structure features based on the dialogue content, including dialogue turns, topic transitions, and conversation duration features; Analyze based on the dialogue structure characteristics and the preset fraud conversation intensive time analysis strategy to determine the fraud conversation intensive time period; Increase the number of conversations with the anthropomorphic traffic virtual body based on the time period with dense fraud conversations.

8. The intelligent anti-fraud information construction system based on traffic information according to claim 7 is characterized in that: The fraud session intensive time analysis strategy includes: Analyze the conversation turn, topic conversion and conversation duration features and compare them with the preset benchmark conversation features to obtain the conversation turn suspicious score, topic conversion suspicious score and time suspicious score; A comprehensive suspicious score is obtained by weighting the suspicious score of the conversation turn, the suspicious score of the topic change and the suspicious score of the time; When the comprehensive score exceeds the preset benchmark suspicious score, the session is analyzed over time and a time period marker for intensive fraud sessions is generated.

9. The intelligent anti-fraud information construction system based on traffic information according to claim 1 is characterized in that: Step 7 also includes: Construct a fraud knowledge graph based on abnormal keyword sequences and preset identification modules; Analyze the fraud patterns of the conversation content based on the fraud knowledge graph to identify new fraud patterns; Add new fraud patterns to the preset deep learning semantic understanding model and perform implicit intent analysis on the response content to determine implicit fraud characteristics; Based on the implicit fraud features, the preset conversation response model is modified for conversation content to generate implicit bias elements; The preset conversation response model is adjusted for conversation content feedback based on implicit bias factors.

Images