Anti-fraud prompt information generation method and device, computer equipment and storage medium

Abstract

The invention relates to an anti-fraud prompt information generation method and device, computer equipment and a storage medium, relates to the technical field of intelligent data processing, and is used for improving the accuracy of anti-fraud prompt information generation. The method mainly comprises the following steps: acquiring dialogue data between a target user and other people, wherein the target user is a user authorized to use an anti-fraud prompt function; inputting the dialogue data into a fraud recognition model to obtain anti-fraud prompt information, wherein the anti-fraud prompt information comprises a fraud type, a fraud probability, fraud means analysis and fraud prevention reminding; and sending the anti-fraud prompt information of which the fraud probability is greater than a first probability value to the target user so as to prompt the target user that the fraud risk exists.

Description

Technical Field

[0001] The present invention relates to the field of intelligent data processing technology, and in particular to a method, device, computer equipment and storage medium for generating anti-fraud prompt information. Background Art

[0002] Anti-fraud prompt information generation refers to generating corresponding prompt information according to different fraud types and scenarios to remind people to take precautions. Due to the rapid and diverse updates of fraud methods, fraudsters constantly change their rhetoric and methods to evade detection and crackdown, making it difficult for anti-fraud systems to keep up with the changes in fraud methods. In addition, since different types of fraud methods involve different rhetoric, the data that the anti-fraud system needs to process is highly diverse, which brings challenges to fraud identification and analysis. Summary of the invention

[0003] The present invention is intended to provide a method, device, computer equipment and storage medium for generating anti-fraud prompt information to solve the deficiencies in the prior art. The technical problem to be solved by the present invention is achieved through the following technical solutions.

[0004] An embodiment of the present invention provides a method for generating anti-fraud prompt information, the method comprising: Obtaining conversation data between the target user and other persons, where the target user is a user who authorizes to use the anti-fraud reminder function; Input the conversation data into a fraud identification model to obtain anti-fraud prompt information, wherein the anti-fraud prompt information includes: fraud type, fraud probability, fraud method analysis and fraud prevention reminder; The anti-fraud prompt information when the fraud probability is greater than the first probability value is sent to the target user to prompt the target user of the fraud risk.

[0005] In an optional embodiment, the training process of the fraud identification model is: Obtaining conversation sample data and conversation sample labels for training the fraud identification model, wherein the conversation sample labels include: fraud type, fraud probability, fraud method analysis, and fraud prevention reminder; Training the fraud identification model according to the conversation sample data and the conversation sample labels, and calculating the loss value of the fraud identification model through a loss function; When the loss value is less than a preset value, the training of the fraud identification model is completed.

[0006] In an optional embodiment, the obtaining of the conversation sample data and conversation sample labels for training the fraud identification model includes: Acquire fraud case record data, and input the fraud case record data into the dialogue generator model to obtain the dialogue sample data; The conversation sample data is labeled according to four dimensions: fraud type, fraud probability, fraud method analysis, and fraud prevention reminder to obtain the conversation sample label.

[0007] In an optional embodiment, the method further includes: performing data cleaning and data format unification on the speech sample data.

[0008] In an optional embodiment, the calculating the loss value of the fraud identification model by using a loss function includes: The loss function is: ; in, , , is the weight coefficient; is the cross entropy calculation factor, which is used to measure the difference between the anti-fraud prompt information generated by the model and the target anti-fraud prompt information; The overall evaluation calculation factor is used to measure the overall similarity between the generated anti-fraud prompt information and the target anti-fraud prompt information; A factor is calculated to penalize duplicate generation, to avoid generating lengthy, repetitive content.

[0009] In an optional embodiment, the formulas for the calculation factors are:

[0010] Where T is the length of the generated anti-fraud prompt information; is the true label of the target word at time step t; The probability of the model generating the target word at time step t; ;

[0011] is the ROUGE-L score between the generated anti-fraud prompt information and the target anti-fraud prompt information, where ROUGE-L represents the similarity of the longest common subsequence; ;

[0012] is the indicator function, when the tth generated word With any previous generated word The value is 1 if they are the same, otherwise the value is 0.

[0013] In an optional embodiment, the method further includes: Determine the conversation data with a fraud probability between the first probability value and the second probability value as target conversation data, wherein the second probability value is less than the first probability value; Clustering the target conversation data to obtain a cluster label, where the cluster label is used to indicate the fraud type of the target conversation data; Annotate the target conversation data with fraud probability, fraud method analysis, and fraud prevention reminders; The fraud identification model continues to be trained based on the target conversation data and its corresponding fraud type, fraud probability, fraud method analysis and fraud prevention reminder.

[0014] An embodiment of the present invention provides an anti-fraud prompt information generating device, the device comprising: An acquisition module, used to acquire conversation data between a target user and other persons, where the target user is a user who is authorized to use the anti-fraud prompt function; A prediction module, used to input the conversation data into a fraud identification model to obtain anti-fraud prompt information, wherein the anti-fraud prompt information includes: fraud type, fraud probability, fraud method analysis and fraud prevention reminder; The sending module is used to send the anti-fraud prompt information when the fraud probability is greater than the first probability value to the target user to prompt the target user of the fraud risk.

[0015] The embodiments of the present invention include the following advantages: The embodiment of the present invention provides a method, device, computer equipment and storage medium for generating anti-fraud prompt information. First, the conversation data between the target user and other people is obtained, and then the conversation data is input into the fraud recognition model to obtain anti-fraud prompt information. Finally, the anti-fraud prompt information with a fraud probability greater than a first probability value is sent to the target user to remind the target user of the fraud risk. Since the fraud recognition model in this embodiment is fine-tuned and trained based on a large semantic model, it solves the problems of the traditional model generating anti-fraud prompt information, such as the rigid technology with weak generalization, inflexible constraints, low adaptability, and simple extraction without analysis to give high-level semantic understanding; and the anti-fraud prompt information obtained by the fraud recognition model in this embodiment includes fraud type, fraud probability, fraud means analysis and fraud prevention reminder. The fraud means analysis and fraud prevention reminder are obtained by the fraud recognition model based on the in-depth understanding and analysis of the conversation data, thereby improving the accuracy of the anti-fraud prompt information generation through the present invention, and providing users with more effective anti-fraud suggestions. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flowchart of the steps of an embodiment of a method for generating anti-fraud prompt information of the present invention; Figure 2 is a flowchart of a method for training a fraud identification model according to an embodiment of the present invention; Figure 3 It is a structural schematic diagram of an embodiment of an anti-fraud prompt information generating device of the present invention. DETAILED DESCRIPTION

[0017] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0018] Embodiment 1: See also Figure 1 , is a method for generating anti-fraud prompt information provided by an embodiment of the present invention, the method specifically includes S101-S103: S101, obtaining conversation data between the target user and other people.

[0019] The conversation data may be call data, social software chat data, SMS data, etc., which is not specifically limited in this embodiment. It should be noted that the target user in this embodiment is a user who authorizes the use of the anti-fraud reminder function, that is, by turning on the anti-fraud reminder function, the application corresponding to this embodiment can obtain the target user's authorization to obtain the conversation data with other people in the target user's terminal device, thereby determining whether the target user is at risk of being defrauded through the target user's conversation data.

[0020] S102, inputting the conversation data into a fraud identification model to obtain anti-fraud prompt information, wherein the anti-fraud prompt information includes: fraud type, fraud probability, fraud method analysis and fraud prevention reminder.

[0021] Among them, the fraud types may be online shopping fraud, fake customer service fraud, false lottery fraud, impersonating relatives and friends fraud, online part-time job fraud, etc. This embodiment does not limit the specific manifestations of the fraud types.

[0022] In this embodiment, the fraud identification model is trained based on sample conversation data and corresponding sample labels. By using fine-tuning methods such as Lora, the parameters of the pre-trained model are adjusted to guide the model to generate anti-fraud prompt information in fraud-related scenarios to improve model performance. It performs better in the three tasks of generating anti-fraud prompt information in terms of fraud-related scene identification, fraud-related means analysis, and prevention advice guidance. Specifically, the fraud identification model in this embodiment uses the language understanding, context understanding, and summary extraction capabilities of the large model base and various expansion tools to achieve distributed and efficient fine-tuning training based on high-quality fine-tuning corpus in the context of fraud-related call scenarios.

[0023] For example, the conversation data input to the fraud detection model is: [Scammer]: Hello, I am your leader. I have an important project that requires your help. Can you help me? [Answer]: Hello, which leader are you? [Scammer]: I am Leader XX, don’t worry, this project is very important, if you help me, I will give you a generous reward.

[0024] [Pickup]: OK, what do you need me to do? [Scammer]: You need to pay a contract deposit for me. This deposit is required. After you pay it, I will transfer the money to you immediately.

[0025] [Agent]: OK, how much can I pay? [Scammer]: You need to pay 98,000 yuan. This is a must. After you pay, I will transfer the money to you immediately.

[0026] [Answer]: OK, I understand.

[0027] [Scammer]: OK, you transfer the money to this account and I will transfer the money to you immediately.

[0028] [Answer]: OK, I have transferred the money.

[0029] [Scammer]: OK, I received it, I will transfer the money to you immediately.

[0030]

Pickup

[0031] [Scammer]: You’re welcome. If you have any questions, feel free to contact me.

[0032]

Pickup

[0033] The anti-fraud prompt information obtained through the fraud identification model is: {'Fraud type': 'Fraud by impersonating a leader or acquaintance', 'Fraud probability':'0.95', 'Analysis of fraud methods':'1. Impersonating a leader or acquaintance as bait to try to gain the victim's trust. 2. Pulling the victim into a QQ group or WeChat group and asking the victim to pay a deposit or other fees. 3. Asking the victim to transfer money multiple times under the pretext of transferring money, in fact, is to defraud the victim's property by inducing the victim to transfer money to the fraudster's account. 4. Once the victim's property is defrauded, they will refuse to refund, and even ridicule and threaten the victim to evade legal responsibility. ',

[0034] 'Fraud prevention reminder':'1. When encountering such fraud, you should be vigilant and do not easily believe the information provided by strangers. 2. Do not disclose user information, especially sensitive information such as bank accounts and passwords. 3. Do not trust so-called leaders or acquaintances, and do not easily pay deposits or other fees. 4. In case of suspicious situations, consult local law enforcement agencies or bank customer service centers in time. If you encounter fraud, report it to the police or relevant departments in time to avoid property losses.}"}

[0035] S103, sending anti-fraud prompt information indicating that the fraud probability is greater than the first probability value to the target user to prompt the target user of the fraud risk.

[0036] The first probability value is set according to actual needs, such as 80%, 85%, 90%, etc. The anti-fraud prompt information can be sent to the target customer in the form of a text message or in the form of a voice broadcast, which is not specifically limited in this embodiment.

[0037] An anti-fraud prompt information generation method provided by an embodiment of the present invention first obtains the conversation data between the target user and other people, then inputs the conversation data into a fraud identification model to obtain anti-fraud prompt information, and finally sends the anti-fraud prompt information with a fraud probability greater than a first probability value to the target user to remind the target user of the fraud risk. Since the fraud identification model in this embodiment is fine-tuned and trained based on a large semantic model, it solves the problems of the traditional model generating anti-fraud prompt information, such as the rigid technology with weak generalization, inflexible constraints, low adaptability, and simple extraction without analysis to give high-level semantic understanding; and the anti-fraud prompt information obtained by the fraud identification model in this embodiment includes fraud type, fraud probability, fraud means analysis and fraud prevention reminder, which is obtained by the fraud identification model through in-depth understanding and analysis based on the conversation data, thereby improving the accuracy of anti-fraud prompt information generation through the present invention, and providing users with more effective anti-fraud suggestions.

[0038] Embodiment 2: See also Figure 2 , is a flow chart of a method for training a fraud identification model provided by an embodiment of the present invention, the method specifically includes S201-S203: S201, obtaining conversation sample data and conversation sample labels for training a fraud identification model, wherein the conversation sample labels include: fraud type, fraud probability, fraud method analysis, and fraud prevention reminder.

[0039] The conversation sample data is extracted from the Internet or provided by the user. In an optional embodiment provided in the present application, the conversation sample data and conversation sample labels for training the fraud recognition model are obtained, including:

[0040] S2011, obtaining fraud case record data, and inputting the fraud case record data into the dialogue generator model to obtain dialogue sample data.

[0041] In this embodiment, a dialogue generator model is first constructed, and sample data of fraud-related dialogues is generated in batches through the dialogue generator model, thereby solving the problem of insufficient telecommunications fraud corpus. Then, by using the ChatGPI API interface, the dialogue generator model is trained by using a few-shot method to build a prompt fine-tuning large model. This achieves the restoration of corresponding call dialogue corpus based on public telecommunications fraud reporting case samples and typical fraud-related scenarios. Among them, before training the dialogue generator model, it is necessary to manually construct 1-2 samples to fine-tune the guiding dialogue generator model. These 1-2 sample data are constructed by manually extracting and transforming real fraud-related call samples.

[0042] For example, the open-source telecom fraud case transcript data is listed below:

[0043] A small amount of dialogue sample data for fewshot fine-tuning is manually constructed based on the above-mentioned telecom fraud case transcript data. The example is as follows: {Fraud type label: 'impersonating a leader, acquaintance', Call content restoration: [Scammer]: Hello, I am your leader. I have an important project that requires your help. Can you help me? [Answer]: Hello, which leader are you? [Scammer]: I am Leader XX, don’t worry, this project is very important, if you help me, I will give you a generous reward.

[0044] [Pickup]: OK, what do you need me to do? [Scammer]: You need to pay a contract deposit for me. This deposit is required. After you pay it, I will transfer the money to you immediately.

[0045] [Agent]: OK, how much can I pay? [Scammer]: You need to pay 98,000 yuan. This is a must. After you pay, I will transfer the money to you immediately.

[0046] [Answer]: OK, I understand.

[0047] [Scammer]: OK, you transfer the money to this account and I will transfer the money to you immediately.

[0048] [Answer]: OK, I have transferred the money.

[0049] [Scammer]: OK, I received it, I will transfer the money to you immediately.

[0050]

Pickup

[0051] [Scammer]: You’re welcome. If you have any questions, feel free to contact me.

[0052]

Pick up

[0053] Furthermore, after obtaining the conversation sample data and conversation sample labels for training the fraud identification model, this embodiment can clean the conversation sample data and unify the data format. Among them, data cleaning is to remove the noise and inconsistent data such as inconsistent conversations and too few information rounds in the data set to ensure the quality of the data; data formatting is to convert the data into a format that the model can process, such as the original data is usually xxxx, which is constructed into the training format of the model; data segmentation and data enhancement are to segment the data set into a training set, a validation set, and a test set to facilitate the evaluation of the performance of the model during the training process. At the same time, the samples of low-frequency scenes are enhanced by using synonyms and other methods.

[0054] S2012, the conversation sample data is labeled according to four dimensions: fraud type, fraud probability, fraud method analysis and fraud prevention reminder to obtain conversation sample labels.

[0055] For example, the dialogue sample labels obtained by annotating the dialogue sample data obtained in S2011 are as follows: {'Fraud type': 'Fraud by impersonating a leader or acquaintance' 'Fraud probability':'0.95' 'Analysis of fraud methods': '1. Impersonating a leader or acquaintance as bait to try to gain the victim's trust. 2. Pulling the victim into a QQ group or WeChat group and asking the victim to pay a deposit or other fees. 3. Asking the victim to transfer money multiple times under the pretext of transferring money, in fact, is to defraud the victim's property by inducing the victim to transfer money to the fraudster's account. 4. Once the victim's property is defrauded, they will refuse to refund, and even ridicule and threaten the victim to evade legal responsibility.'

[0056] 'Fraud prevention reminder':'1. When encountering such fraud, you should be vigilant and do not easily believe the information provided by strangers. 2. Do not disclose personal information, especially sensitive information such as bank accounts and passwords. 3. Do not trust so-called leaders or acquaintances, and do not easily pay deposits or other fees. 4. In case of suspicious situations, consult local law enforcement agencies or bank customer service centers in time. If you encounter fraud, report it to the police or relevant departments in time to avoid property losses. '}

[0057] S202, training the fraud identification model according to the conversation sample data and the conversation sample labels, and calculating the loss value of the fraud identification model through the loss function.

[0058] It should be noted that the cross-entropy loss is usually used in the large model fine-tuning training loss function to measure the difference between the summary generated by the model and the real summary. The traditional cross-entropy loss focuses on the accuracy of word-by-word prediction, but for the summary generation task, it is hoped that the content of the summary is as consistent as possible with the real summary.

[0059] This embodiment improves the loss function during the training of the fraud identification model, and constructs a joint loss function by introducing evaluation indicators and penalizing repeated generation calculation factors. At the same time, there are high requirements for the conciseness and information density of the summary. By constructing a joint loss function, on the one hand, the cross entropy loss is combined with the ROUGE evaluation indicator, which can take into account the accuracy of word-by-word prediction while taking into account the overall summary quality. On the other hand, by combining it with the penalty for repeated generation calculation factor, the model can be prevented from generating lengthy and repetitive content, thereby generating a more concise and information-dense summary. Therefore, this embodiment can calculate the loss value of the fraud identification model through the following loss function:

[0060] ; in, , , is the weight coefficient; is the cross entropy calculation factor, which is used to measure the difference between the anti-fraud prompt information generated by the model and the target anti-fraud prompt information; The overall evaluation calculation factor is used to measure the overall similarity between the generated anti-fraud prompt information and the target anti-fraud prompt information; A factor is calculated to penalize duplicate generation, to avoid generating lengthy, repetitive content.

[0061] Among them, the calculation formula of the cross entropy calculation factor is:

[0062] Where T is the length of the generated anti-fraud prompt information; is the true label of the target word at time step t (one-hot encoding); The probability of the model generating the target word at time step t; The calculation formula for the overall evaluation factor is: ;

[0063] is the ROUGE-L score between the generated anti-fraud prompt information and the target anti-fraud prompt information, where ROUGE-L represents the similarity of the longest common subsequence; The calculation formula for the penalty duplicate generation calculation factor is:

[0064] is the indicator function, when the tth generated word With any previous generated word The value is 1 if they are the same, otherwise the value is 0.

[0065] S203, when the loss value is less than the preset value, the training of the fraud recognition model is completed.

[0066] The preset value is a value set according to actual needs.

[0067] In this embodiment, in order to further improve the accuracy of the fraud identification model, the fraud identification model can also be updated at a preset time interval to enhance the identification ability of the fraud identification model. Specifically, in this embodiment, the conversation data whose fraud probability is between the first probability value and the second probability value is determined as the target conversation data, and the second probability value is less than the first probability value; then the target conversation data is clustered to obtain cluster labels, and the cluster labels are used to indicate the fraud type of the target conversation data; the fraud probability, fraud means analysis and fraud prevention reminder of the target conversation data are marked, and then the final target conversation data and its corresponding fraud type, fraud probability, fraud means analysis and fraud prevention reminder are added to the fraud segment knowledge base, and finally the fraud identification model is continuously trained according to the preset time interval for the target conversation data added to the fraud segment knowledge base and its corresponding fraud type, fraud probability, fraud means analysis and fraud prevention reminder.

[0068] This embodiment adopts a self-learning mechanism, and obtains cluster labels representing fraud types based on semantic understanding clustering through accumulated target conversation data. Such labels include "fraud by impersonating a leader", "fraud by false lottery winning", "fraud by impersonating a customer service representative", "fraud by false loan fraud", etc. Then, the fraud probability, fraud method analysis and fraud prevention reminder are labeled through manual review to form new fraud conversation data, which is then added to the fraud segment knowledge base. In the process of generating anti-fraud prompt information, the fraud recognition model forms new semantic prompt guidance features by searching the knowledge base, thereby realizing the self-update capability of the model and enhancing the model recognition capability.

[0069] By adopting a self-learning mechanism, traditional models often find very novel fraud methods online but the model does not recognize them. Due to the small number of bad cases, the data volume required for model optimization cannot be reached. The conventional data collection / generation-model optimization path is time-consuming and unpredictable (usually 1-2 months), and it is impossible to quickly solve and timely feedback new forms of fraud. By clustering a small amount of text by analogy judgment, a semantic prompt is generated and stored in the feedback update knowledge base. By adding semantic prompts to the model, the model is updated to timely identify fraud scenarios, solving the problem of fast sample updates and timely updates of data diversification. Improve the system's ability to handle diverse data, so that it can more accurately identify and prevent various types of fraud.

[0070] This embodiment provides a method for training a fraud identification model to improve the quality and practicality of anti-fraud prompt information generation. By using an NLP model to restore the conversation sample data and corresponding conversation sample labels of the telecommunications fraud call process based on reported cases or typical fraud scenarios, the problem of insufficient telecommunications fraud corpus is solved; the conversation sample labels include fraud type, fraud probability, fraud method analysis and fraud prevention reminders, and then the model is trained through the conversation sample data and the corresponding conversation sample labels to obtain a fraud identification model. The large model is used to generate high accuracy and flexibility to solve the problem that the traditional model generates anti-fraud prompt information with rigid technology, weak generalization, inflexible constraints, low adaptability, and simple extraction without analysis to give high-level semantic understanding. The fraud identification model in this embodiment not only extracts key information, but also has a deeper understanding and analysis of the content, thereby providing users with more effective anti-fraud suggestions.

[0071] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0072] Embodiment 3: In one embodiment, an anti-fraud prompt information generating device is provided, and the anti-fraud prompt information generating device corresponds to the anti-fraud prompt information generating method in the above embodiment. Figure 3 As shown, the functional modules of the anti-fraud prompt information generation and acquisition device are described in detail as follows:

[0073] An acquisition module 31 is used to acquire conversation data between a target user and other persons, where the target user is a user who is authorized to use the anti-fraud prompt function; Prediction module 32, used for inputting the conversation data into a fraud identification model to obtain anti-fraud prompt information, wherein the anti-fraud prompt information includes: fraud type, fraud probability, fraud method analysis and fraud prevention reminder; The sending module 33 is used to send the anti-fraud prompt information that the fraud probability is greater than the first probability value to the target user to prompt the target user of the fraud risk.

[0074] In an optional embodiment, the anti-fraud prompt information generating device further includes a training module 34: The acquisition module 31 is further used to acquire the conversation sample data and conversation sample labels for training the fraud identification model, wherein the conversation sample labels include: fraud type, fraud probability, fraud method analysis and fraud prevention reminder; The training module 34 is used to train the fraud identification model according to the conversation sample data and the conversation sample label, and calculate the loss value of the fraud identification model through a loss function; when the loss value is less than a preset value, the training of the fraud identification model is completed.

[0075] In an optional embodiment, the acquisition module 31 is specifically configured to: Acquire fraud case record data, and input the fraud case record data into the dialogue generator model to obtain the dialogue sample data; The conversation sample data is labeled according to four dimensions: fraud type, fraud probability, fraud method analysis, and fraud prevention reminder to obtain the conversation sample label.

[0076] In an optional embodiment, the anti-fraud prompt information generating device further includes a data processing module 35 for performing data cleaning and data format unification on the speech sample data.

[0077] In an optional embodiment, the loss function is: ; in, , , is the weight coefficient; is the cross entropy calculation factor, which is used to measure the difference between the anti-fraud prompt information generated by the model and the target anti-fraud prompt information; The overall evaluation calculation factor is used to measure the overall similarity between the generated anti-fraud prompt information and the target anti-fraud prompt information; A factor is calculated to penalize duplicate generation, to avoid generating lengthy, repetitive content.

[0078] In an optional embodiment, the formulas for the calculation factors are: ;

[0079] Where T is the length of the generated anti-fraud prompt information; is the true label of the target word at time step t; The probability of the model generating the target word at time step t; ;

[0080] is the ROUGE-L score between the generated anti-fraud prompt information and the target anti-fraud prompt information, where ROUGE-L represents the similarity of the longest common subsequence; ;

[0081] is the indicator function, when the tth generated word With any previous generated word The value is 1 if they are the same, otherwise the value is 0.

[0082] In an optional embodiment, the training module 34 is further configured to: Determine the conversation data with a fraud probability between the first probability value and the second probability value as target conversation data, wherein the second probability value is less than the first probability value; Clustering the target conversation data to obtain a cluster label, where the cluster label is used to indicate the fraud type of the target conversation data; Annotate the target conversation data with fraud probability, fraud method analysis, and fraud prevention reminders; The fraud identification model continues to be trained based on the target conversation data and its corresponding fraud type, fraud probability, fraud method analysis and fraud prevention reminder.

[0083] It should be noted that the above detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the art to which the present application belongs.

[0084] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should also be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0085] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein.

[0086] In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusions. 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 explicitly listed, but may include other steps or units that are not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0087] For ease of description, spatially relative terms, such as "above", "above", "on the upper surface of", "above", etc., may be used herein to describe the spatial positional relationship between a device or feature and other devices or features as shown in the figure. It should be understood that spatially relative terms are intended to include different orientations of the device in use or operation in addition to the orientation described in the figure. For example, if the device in the accompanying drawings is inverted, the device described as "above other devices or structures" or "above other devices or structures" will be positioned as "below other devices or structures" or "below other devices or structures". Thus, the exemplary term "above" may 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 spatially relative descriptions used herein are interpreted accordingly.

[0088] In the above detailed description, reference is made to the accompanying drawings, which form a part of this document. In the accompanying drawings, similar symbols typically identify similar components unless the context indicates otherwise. The illustrated embodiments described in the detailed description, drawings, and claims are not meant to be limiting. Other embodiments may be used, and other changes may be made, without departing from the spirit or scope of the subject matter presented herein.

[0089] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for generating anti-fraud prompt information, characterized in that: The method comprises: Obtaining conversation data between the target user and other persons, where the target user is a user who authorizes to use the anti-fraud reminder function; Input the conversation data into a fraud identification model to obtain anti-fraud prompt information, wherein the anti-fraud prompt information includes: fraud type, fraud probability, fraud method analysis and fraud prevention reminder; The anti-fraud prompt information when the fraud probability is greater than the first probability value is sent to the target user to prompt the target user of the fraud risk.

2. The method according to claim 1, characterized in that The training process of the fraud identification model is as follows: Obtaining conversation sample data and conversation sample labels for training the fraud identification model, wherein the conversation sample labels include: fraud type, fraud probability, fraud method analysis, and fraud prevention reminder; Training the fraud identification model according to the conversation sample data and the conversation sample labels, and calculating the loss value of the fraud identification model through a loss function; When the loss value is less than a preset value, the training of the fraud identification model is completed.

3. The method according to claim 2, characterized in that The obtaining of the conversation sample data and conversation sample labels for training the fraud identification model includes: Acquire fraud case record data, and input the fraud case record data into the dialogue generator model to obtain the dialogue sample data; The conversation sample data is labeled according to four dimensions: fraud type, fraud probability, fraud method analysis, and fraud prevention reminder to obtain the conversation sample label.

4. The method according to claim 3, characterized in that The method further comprises: The sample data is cleaned and the data format is unified.

5. The method according to claim 2, characterized in that: The calculating the loss value of the fraud identification model by using a loss function includes: The loss function is: ; in, , , is the weight coefficient; is the cross entropy calculation factor, which is used to measure the difference between the anti-fraud prompt information generated by the model and the target anti-fraud prompt information; The overall evaluation calculation factor is used to measure the overall similarity between the generated anti-fraud prompt information and the target anti-fraud prompt information; A factor is calculated to penalize duplicate generation, to avoid generating lengthy, repetitive content.

6. The method according to claim 5, characterized in that The formulas for each calculation factor are: ； Where T is the length of the generated anti-fraud prompt information; is the true label of the target word at time step t; The probability of the model generating the target word at time step t; ； is the ROUGE-L score between the generated anti-fraud prompt information and the target anti-fraud prompt information, where ROUGE-L represents the similarity of the longest common subsequence; ； is the indicator function, when the tth generated word With any previous generated word The value is 1 if they are the same, otherwise the value is 0.

7. The method according to any one of claims 1 to 6, characterized in that: The method further comprises: Determine the conversation data with a fraud probability between the first probability value and the second probability value as target conversation data, wherein the second probability value is less than the first probability value; Clustering the target conversation data to obtain a cluster label, where the cluster label is used to indicate the fraud type of the target conversation data; Annotate the target conversation data with fraud probability, fraud method analysis, and fraud prevention reminders; The fraud identification model continues to be trained based on the target conversation data and its corresponding fraud type, fraud probability, fraud method analysis and fraud prevention reminder.

8. An anti-fraud prompt information generating device, characterized in that: The device comprises: An acquisition module, used to acquire conversation data between a target user and other persons, where the target user is a user who is authorized to use the anti-fraud prompt function; A prediction module, used to input the conversation data into a fraud identification model to obtain anti-fraud prompt information, wherein the anti-fraud prompt information includes: fraud type, fraud probability, fraud method analysis and fraud prevention reminder; The sending module is used to send the anti-fraud prompt information when the fraud probability is greater than the first probability value to the target user to prompt the target user of the fraud risk.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the anti-fraud prompt information generating method as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the anti-fraud prompt information generating method as described in any one of claims 1 to 7 is implemented.

Images