Multi-round problem generation method and device of risk control early warning system
By constructing early warning feature vectors and generating multi-round question lists, combined with a large language model and knowledge base, the problems of single question sets and low security in traditional risk control early warning systems are solved, thereby improving the flexibility and accuracy of the risk control system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG BANGSUN TECH CO LTD
- Filing Date
- 2024-11-15
- Publication Date
- 2026-05-15
AI Technical Summary
The fixed question set in traditional risk control and early warning systems is easily circumvented by attackers, resulting in low security and accuracy. Relying entirely on large language models to generate questions may lead to inaccurate or irrelevant information.
By acquiring transaction records to construct early warning feature vectors, using a preset early warning model to predict and, when the value exceeds a threshold, obtaining similar early warning features and script templates from the knowledge base, and combining them with a large language model to generate a multi-round question list and handling tree, the system is then connected to the outbound calling system.
It improves the flexibility and accuracy of the risk control system, ensures the diversity of issues, avoids data privacy leaks and inaccurate inferences, and effectively identifies potential risks.
Smart Images
Figure CN122045336A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of big data technology, and in particular to a method and apparatus for generating multiple rounds of questions for a risk control and early warning system. Background Technology
[0002] In traditional risk control and early warning systems, outbound calling systems typically use a fixed set of questions to collect customer feedback and information. While this method improves efficiency to some extent, the fixed questioning pattern also introduces significant risks. Attackers can design risk-avoiding answers to these standardized questions, thereby evading system monitoring and intervention. This predictability makes it easier for risk-prone individuals to identify and exploit system weaknesses, leading to potential financial losses or security vulnerabilities.
[0003] At the same time, relying solely on large language models to generate problems also faces challenges. Large language models may exhibit illusions when generating content, meaning that the generated information is inaccurate or irrelevant, which could affect the effectiveness of risk control decisions.
[0004] Therefore, how to improve the flexibility of outbound calling systems while ensuring the security and accuracy of information has become an urgent problem to be solved by current risk control and early warning systems. Summary of the Invention
[0005] In view of this, embodiments of the present invention provide a method and apparatus for generating multiple rounds of questions in a risk control and early warning system, in order to address the problem of single problem sets and low security and accuracy.
[0006] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:
[0007] The first aspect of this invention discloses a multi-round question generation method for a risk control and early warning system, the method comprising:
[0008] Obtain multiple transaction records and construct an early warning feature vector for each transaction record;
[0009] The prediction result is obtained by using a preset early warning model to predict the early warning feature vector;
[0010] If the prediction result is greater than the warning threshold, then k target warning features and k dialogue templates corresponding to each warning feature in the warning feature vector are obtained from the knowledge base. The target warning features are preset warning features similar to the warning features in the knowledge base. Each dialogue template includes multiple question examples.
[0011] For each of the aforementioned warning features, a question list is generated based on the large language model, the warning feature, k of the aforementioned target warning features, and multiple of the aforementioned question examples;
[0012] A handling tree is generated based on the list of questions, and the handling tree is connected to the outbound calling system so that users can answer multiple rounds of questions in the handling tree.
[0013] Preferably, the step of acquiring multiple transaction records and constructing a warning feature vector for each transaction record includes:
[0014] Retrieve multiple transaction records;
[0015] For each of the aforementioned transaction records, the values of multi-dimensional early warning features in the transaction record are obtained;
[0016] A warning feature vector is constructed based on the values of the multi-dimensional warning features.
[0017] Preferably, the step of obtaining k target warning features and k script templates corresponding to each warning feature in the warning feature vector from the knowledge base includes:
[0018] For each warning feature in the warning feature vector, calculate the cosine similarity between the warning feature and multiple preset warning features in the knowledge base;
[0019] Based on the order of the cosine similarity scores from high to low, obtain the k preset warning features corresponding to the first k cosine similarity scores and mark them as target warning features;
[0020] Obtain k script templates corresponding to the k target warning features from the knowledge base.
[0021] Preferably, for each of the aforementioned warning features, a question list is generated based on a large language model, the warning feature, k of the target warning features, and multiple question examples, including:
[0022] For each of the aforementioned warning features, the meaning corresponding to the warning feature is analyzed using a large language model;
[0023] A question list is generated by combining the meaning with k target warning features and multiple question examples using a large language model.
[0024] Preferably, the step of generating a handling tree based on the question list and connecting the handling tree to the outbound calling system so that users can answer multiple rounds of questions in the handling tree includes:
[0025] The large language model is used to generate a disposal tree based on the problem list, wherein the non-leaf nodes of the disposal tree are problems, and the leaf nodes of the disposal tree are the numbers of the disposal methods;
[0026] The processing tree is connected to the outbound calling system so that users can answer multiple rounds of questions in the processing tree.
[0027] A second aspect of this invention discloses a multi-round problem generation device for a risk control early warning system, the device comprising:
[0028] A construction unit is used to acquire multiple transaction records and construct an early warning feature vector for each of the transaction records;
[0029] The prediction unit is used to predict the warning feature vector using a preset warning model to obtain the prediction result;
[0030] The acquisition unit is configured to, if the prediction result is greater than the warning threshold, acquire k target warning features and k dialogue templates corresponding to each warning feature in the warning feature vector from the knowledge base, wherein the target warning features are preset warning features similar to the warning features in the knowledge base; each dialogue template includes multiple question examples;
[0031] The generation unit is used to generate a question list for each of the aforementioned warning features, based on a large language model, the warning feature, k of the aforementioned target warning features, and multiple question examples;
[0032] The access unit is used to generate a handling tree based on the question list and connect the handling tree to the outbound calling system so that users can answer multiple rounds of questions in the handling tree.
[0033] Preferably, the building unit includes:
[0034] The first acquisition module is used to acquire multiple transaction records;
[0035] The second acquisition module is used to acquire the values of multi-dimensional early warning features in each of the transaction records.
[0036] A construction module is used to construct an early warning feature vector based on the values of the multi-dimensional early warning features.
[0037] Preferably, the acquisition unit includes:
[0038] The calculation module is used to calculate the cosine similarity between each warning feature in the warning feature vector and multiple preset warning features in the knowledge base.
[0039] The third acquisition module is used to acquire k preset warning features corresponding to the first k cosine similarities in descending order of the multiple cosine similarities, and mark them as target warning features.
[0040] The fourth acquisition module is used to acquire k script templates corresponding to k target warning features from the knowledge base.
[0041] Preferably, the generation unit includes:
[0042] The parsing module is used to parse the meaning of each warning feature using a large language model.
[0043] The generation module is used to generate a question list by combining the meaning with k target warning features and multiple question examples through a large language model.
[0044] Preferably, the access unit includes:
[0045] A disposal tree generation module is used to generate a disposal tree based on the problem list using the large language model, wherein the non-leaf nodes of the disposal tree are problems, and the leaf nodes of the disposal tree are the numbers of disposal methods;
[0046] The access module is used to connect the processing tree to the outbound calling system so that users can answer multiple rounds of questions in the processing tree.
[0047] This invention provides a method and apparatus for generating multiple rounds of questions for a risk control and early warning system, based on the above embodiments. The method involves acquiring multiple transaction records and constructing an early warning feature vector; predicting the early warning feature vector using a preset early warning model to obtain a prediction result; if the prediction result is greater than an early warning threshold, retrieving k target early warning features and k dialogue templates corresponding to each early warning feature in the early warning feature vector from a knowledge base; generating a question list for each early warning feature based on a large language model, the early warning feature, the k target early warning features, and multiple question examples; generating a handling tree based on the question list and connecting the handling tree to an outbound calling system so that users can answer multiple rounds of questions in the handling tree. This invention first determines whether a transaction carries risk by analyzing the characteristics of the transaction records. The questions and handling tree generated by the offline large language model can raise different follow-up questions according to different situations, ensuring the diversity of questions and effectively guaranteeing the accuracy of the risk control system's inferences. Attached Figure Description
[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0049] Figure 1 A flowchart illustrating a multi-round problem generation method for a risk control and early warning system provided in an embodiment of the present invention;
[0050] Figure 2 This is a flowchart illustrating the acquisition of target warning features and script templates provided in an embodiment of the present invention;
[0051] Figure 3 A flowchart for generating a problem list provided in an embodiment of the present invention;
[0052] Figure 4 A schematic diagram illustrating a multi-round problem generation method for a risk control and early warning system provided in an embodiment of the present invention;
[0053] Figure 5 This is a structural block diagram of a multi-round problem generation device for a risk control and early warning system provided in an embodiment of the present invention. Detailed Implementation
[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0055] In this application, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0056] As the background technology indicates, a fixed set of questions allows attackers to design risk-avoiding answers, thereby evading system monitoring and intervention, resulting in low security. Furthermore, relying entirely on large language models to generate questions may lead to inaccurate or irrelevant information, potentially impacting the effectiveness of risk control decisions.
[0057] Therefore, this invention provides a method and apparatus for generating multiple rounds of questions for a risk control and early warning system. The method involves acquiring multiple transaction records and constructing an early warning feature vector; predicting the early warning feature vector using a preset early warning model to obtain a prediction result; if the prediction result is greater than an early warning threshold, obtaining k target early warning features and k dialogue templates corresponding to each early warning feature in the early warning feature vector from a knowledge base; generating a question list for each early warning feature based on a large language model, the early warning feature, the k target early warning features, and multiple question examples; generating a handling tree based on the question list and connecting the handling tree to an outbound calling system so that users can answer multiple rounds of questions in the handling tree. This invention first determines whether a transaction has risk by analyzing the characteristics of the transaction records. The questions and handling tree generated by the offline large language model can raise different follow-up questions according to different situations, ensuring the diversity of questions and effectively guaranteeing the accuracy of the risk control system's inferences.
[0058] See Figure 1 The flowchart illustrates a multi-round question generation method for a risk control and early warning system provided by an embodiment of the present invention. The method includes:
[0059] Step S101: Obtain multiple transaction records and construct a warning feature vector for each transaction record.
[0060] In the specific implementation step S101, multiple transaction records of the user are obtained, including information such as transaction time, transaction amount, account balance, loan and credit tags, and transaction channels; for each transaction record, the values of multi-dimensional warning features in the transaction record are obtained; and a warning feature vector is constructed based on the values of the multi-dimensional warning features.
[0061] It should be noted that the warning feature vector of the transaction record is, for example: X = (x1, x2, ..., x...). n ), where n represents the number of warning features; x1, x2, ..., x n This represents the value of the warning features used to assess risk in the transaction records. The meanings of the warning features in each dimension are f1, f2, ..., f n This indicates information such as the transaction time, transaction amount, account balance, loan / credit flag, and transaction channel for that transaction.
[0062] Understandably, constructing a warning feature vector for transaction records helps determine whether a transaction carries risk. For specific judgment methods, please refer to the steps below.
[0063] Step S102: Use the preset early warning model to predict the early warning feature vector and obtain the prediction result.
[0064] It should be noted that the pre-set early warning model is trained based on historical transaction records. Specifically, the pre-set early warning model uses Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and other ensemble learning frameworks supported by decision trees.
[0065] In the specific implementation step S102, for each transaction record, the warning feature vector of the transaction record is input into the preset warning model, and the preset warning model M is used to predict the warning feature vector to obtain the prediction result pred.
[0066] In other words, the probability that this transaction record carries risk is: pred = M(X).
[0067] In practical applications, for each transaction record, the prediction result `pred` in the preset early warning model is compared with the early warning threshold `thres`. When the prediction result `pred` is greater than the early warning threshold `thres`, an early warning is triggered, and subsequent steps S103 to S105 are executed to further assess the user's risk anomaly. If the prediction result `pred` is not greater than the early warning threshold `thres`, the process ends.
[0068] Understandably, the optimal choice of the warning threshold thres is determined by the evaluation index of the preset warning model M.
[0069] Step S103: If the prediction result is greater than the warning threshold, then obtain the k target warning features and k dialogue templates corresponding to each warning feature in the warning feature vector from the knowledge base.
[0070] It should be noted that the target warning features are preset warning features similar to the warning features in the knowledge base. Each script template includes multiple question examples.
[0071] It is understandable that the knowledge base (KB) stores multiple preset warning features x. i (x i ∈X). Indexed by preset warning features, each preset warning feature corresponds to a multi-round question template for historical manual verification in a given scenario. i (Temp i ∈T). Each script template Temp i It includes q1, q2, ..., q ti Total t i Example of a problem: Temp i ={q n n = 1, 2, ..., ti}. The knowledge base KB stores m pairs of preset warning features and script templates. The set of preset warning features in the knowledge base KB is X = {x i i = 1, 2, ..., m}, the set of script templates T = {Temp} i i = 1, 2, ..., m}, knowledge base set KB = {X} i Temp i i = 1, 2, ..., m}.
[0072] In the specific implementation step S103, when the prediction result pred is greater than the warning threshold thres, k target warning features and k speech templates that are most similar to the warning features are obtained from the knowledge base according to each warning feature in the warning feature vector.
[0073] For details of the acquisition process, please refer to the embodiments of this invention. Figure 2 The content in Figure 2 Includes:
[0074] Step S201: For each warning feature in the warning feature vector, calculate the cosine similarity between the warning feature and multiple preset warning features in the knowledge base.
[0075] It is understandable that cosine similarity is used to measure the vector distance D(X,X) between each warning feature in the warning feature vector and multiple preset warning features in the knowledge base. i ).
[0076] In the specific implementation step S201, the cosine similarity between each early warning feature x in the early warning feature vector and multiple preset early warning features xi in the knowledge base KB is calculated.
[0077] Step S202: Based on the order of multiple cosine similarities from high to low, obtain the k preset warning features corresponding to the first k cosine similarities and mark them as target warning features.
[0078] In the specific implementation step S202, multiple cosine similarities are sorted in descending order, and the k preset warning features corresponding to the top k cosine similarities are marked as target warning features Xr1, Xr2, ..., Xr K .
[0079] Step S203: Obtain k script templates corresponding to k target warning features from the knowledge base.
[0080] In the specific implementation step S203, using the target warning features as indexes, k dialogue templates Tempr1, Tempr2, ..., Tempr corresponding to k target warning features are retrieved from the knowledge base KB. k.
[0081] The specific acquisition process is shown in formula (1).
[0082]
[0083] Step S104: For each warning feature, generate a list of questions based on the large language model, the warning feature, k target warning features, and multiple question examples.
[0084] In the specific implementation step S104, for each early warning feature X in the early warning feature vector, the early warning feature X and its corresponding k target early warning features Xr1, Xr2, ..., Xr K and k script templates Tempr1, Tempr2, ..., Tempr k Multiple problem examples are used to generate a QList corresponding to each warning feature.
[0085] For details of the generation process, please refer to the embodiments of this invention. Figure 3 The content in Figure 3 Includes:
[0086] Step S301: For each warning feature, analyze the meaning of the warning feature using a large language model.
[0087] In the specific implementation step S301, for each warning feature in the warning feature vector, a large language model is used to parse the meaning corresponding to the warning feature.
[0088] For example: the current warning feature is {{X}}, and the meaning of each dimension in the warning feature vector is {{f1,f2,...,f}, respectively. n}}.
[0089] Step S302: Generate a list of questions by combining the meaning with k target warning features and multiple question examples through a large language model.
[0090] In the specific implementation step S302, a problem list QList is generated by combining the meaning of the warning features with the warning features of k targets and multiple problem examples through a large language model.
[0091] Understandably, the question samples are historical manual verification questions, and the questions in the QList are all questions with only a yes or no answer.
[0092] It should be noted that during the process of generating the question list, the temperature and top_p parameters of the large language model are adjusted to increase the randomness and diversity of the generated results.
[0093] Specifically, an example of generating a list of questions using a large language model is as follows:
[0094] {% for case in KB%}## Warning characteristics:
[0095] {{case.X}};
[0096] The feature meanings corresponding to each dimension of the vector are {{f1,f2,...,f...} n}};
[0097] ## List of Issues:
[0098] {%for question in case.Temp%}{{question}};
[0099] {%endfor%};
[0100] {%endfor%}.
[0101] Step S105: Generate a handling tree based on the question list and connect the handling tree to the outbound calling system so that users can answer multiple rounds of questions in the handling tree.
[0102] In the specific implementation step S105, based on the CoT (Cooperation of Thought) technology, a large language model is used to generate a DTree based on the question list QList, and the DTree is connected to the outbound calling system so that users can answer multiple rounds of questions in the DTree.
[0103] It should be noted that, in conjunction with the embodiments of the present invention Figure 4 As shown, the disposition tree (DTree) is a binary tree. Each node in the binary tree represents a round of questions, and the yes / no answers connect different branches of the binary tree. In other words, the non-leaf nodes of the disposition tree (DTree) are questions, and the leaf nodes are the numbers of the disposition methods.
[0104] It is understandable that users answer multiple rounds of questions, and the methods for dealing with users can be inferred from the answers to these multiple rounds of questions. For example... Figure 4 As shown, the final answer points to the leaf node with the final disposal method number.
[0105] In practical applications, based on the already answered question-and-answer stream (QAStream), the disposal tree is generated iteratively using the following large language model prompt word templates:
[0106] Candidate question list: {% for question in QList%}{{question}};
[0107] {%endfor%};
[0108] Answered questions: {% for question, answer in QAStream%};
[0109] Question: {{question}};
[0110] Answer: {{answer}};
[0111] {%endfor%}.
[0112] Possible actions:
[0113] Start: If there are no answered questions, select the first question in the candidate question list to start asking your question.
[0114] Continue asking questions: Based on the previous answer in the answered question stream, select the next question from the candidate question list to continue asking questions.
[0115] End: If the answered question flow can lead to a solution, or if all questions in the candidate question list have been answered, then the questioning ends.
[0116] In some specific embodiments, the user is initially asked questions based on the root node of the processing tree. When a user's response exceeds a set range (i.e., it is neither a "yes" nor a "no" answer), the user is prompted that the response is incorrect and the question is repeated until the user provides a valid answer. Depending on the user's answer, different branches are selected to continue asking questions about the child nodes on that branch. If the child node is a leaf node, the process stops, and the user's processing method is returned.
[0117] In this embodiment of the invention, by combining preset warning features and dialogue templates, an offline large language model is used to generate a disposal tree, and diverse questions are raised based on different responses, thereby effectively identifying potential risks. In this process, the characteristics of transaction records are first analyzed to determine whether a transaction carries risk. The questions and disposal tree generated by the offline large language model can raise different follow-up questions based on different situations, ensuring the diversity of the questions. Furthermore, based on historical dialogue templates, the application of the offline large language model solves data privacy and illusion problems, ensuring that sensitive information is not leaked and inaccurate inferences are avoided when generating similar questions and disposal trees.
[0118] Corresponding to the multi-round problem generation method for a risk control and early warning system provided in the above embodiments of the present invention, see also... Figure 5 The diagram shows a structural block diagram of a multi-round problem generation device for a risk control and early warning system provided by an embodiment of the present invention. The device includes: a construction unit 501, a prediction unit 502, an acquisition unit 503, a generation unit 504, and an access unit 505.
[0119] Construction unit 501 is used to acquire multiple transaction records and construct an early warning feature vector for each transaction record.
[0120] The prediction unit 502 is used to predict the early warning feature vector using a preset early warning model to obtain the prediction result.
[0121] The acquisition unit 503 is used to acquire k target warning features and k dialogue templates corresponding to each warning feature in the warning feature vector from the knowledge base if the prediction result is greater than the warning threshold. The target warning features are preset warning features similar to the warning features in the knowledge base; each dialogue template includes multiple question examples.
[0122] The generation unit 504 is used to generate a list of questions for each warning feature, based on a large language model, the warning feature, k target warning features, and multiple question examples.
[0123] Access unit 505 is used to generate a handling tree based on the problem list and connect the handling tree to the outbound calling system so that users can answer multiple rounds of questions in the handling tree.
[0124] In this embodiment of the invention, by combining preset warning features and dialogue templates, an offline large language model is used to generate a disposal tree, and diverse questions are raised based on different responses, thereby effectively identifying potential risks. In this process, the characteristics of transaction records are first analyzed to determine whether a transaction carries risk. The questions and disposal tree generated by the offline large language model can raise different follow-up questions based on different situations, ensuring the diversity of the questions. Furthermore, based on historical dialogue templates, the application of the offline large language model solves data privacy and illusion problems, ensuring that sensitive information is not leaked and inaccurate inferences are avoided when generating similar questions and disposal trees.
[0125] Combination Figure 5 The content shown, the construction unit 501, includes: a first acquisition module, a second acquisition module and a construction module.
[0126] The first acquisition module is used to acquire multiple transaction records.
[0127] The second acquisition module is used to acquire the values of multi-dimensional warning features in each transaction record.
[0128] The module is used to construct early warning feature vectors based on the values of multi-dimensional early warning features.
[0129] Combination Figure 5 The content shown, the acquisition unit 503, includes: a calculation module, a third acquisition module and a fourth acquisition module.
[0130] The calculation module is used to calculate the cosine similarity between each warning feature in the warning feature vector and multiple preset warning features in the knowledge base.
[0131] The third acquisition module is used to acquire the k preset warning features corresponding to the first k cosine similarities in descending order of multiple cosine similarities, and mark them as target warning features.
[0132] The fourth acquisition module is used to retrieve k script templates corresponding to k target warning features from the knowledge base.
[0133] Combination Figure 5 The content shown, generation unit 504, includes: a parsing module and a generation module.
[0134] The parsing module is used to parse the meaning of each warning feature using a large language model.
[0135] The generation module is used to generate a list of questions by combining the meaning with k target warning features and multiple question examples through a large language model.
[0136] Combination Figure 5 The content shown, access unit 505, includes: a processing tree generation module and an access module.
[0137] The disposition tree generation module is used to generate a disposition tree based on a list of questions using a large language model. The non-leaf nodes of the disposition tree are questions, and the leaf nodes are the numbers of the disposition methods.
[0138] The access module is used to connect the handling tree to the outbound calling system so that users can answer multiple rounds of questions in the handling tree.
[0139] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0140] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.
[0141] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A multi-round problem generation method for a risk control and early warning system, characterized in that, The method includes: Obtain multiple transaction records and construct an early warning feature vector for each transaction record; The prediction result is obtained by using a preset early warning model to predict the early warning feature vector; If the prediction result is greater than the warning threshold, then k target warning features and k dialogue templates corresponding to each warning feature in the warning feature vector are obtained from the knowledge base. The target warning features are preset warning features similar to the warning features in the knowledge base. Each dialogue template includes multiple question examples. For each of the aforementioned warning features, a question list is generated based on the large language model, the warning feature, k of the aforementioned target warning features, and multiple of the aforementioned question examples; A handling tree is generated based on the list of questions, and the handling tree is connected to the outbound calling system so that users can answer multiple rounds of questions in the handling tree.
2. The method according to claim 1, characterized in that, The step of acquiring multiple transaction records and constructing a warning feature vector for each transaction record includes: Retrieve multiple transaction records; For each of the aforementioned transaction records, the values of multi-dimensional early warning features in the transaction record are obtained; A warning feature vector is constructed based on the values of the multi-dimensional warning features.
3. The method according to claim 1, characterized in that, The step of retrieving k target warning features and k dialogue templates corresponding to each warning feature in the warning feature vector from the knowledge base includes: For each warning feature in the warning feature vector, calculate the cosine similarity between the warning feature and multiple preset warning features in the knowledge base; Based on the order of the cosine similarity scores from high to low, obtain the k preset warning features corresponding to the first k cosine similarity scores and mark them as target warning features; Obtain k script templates corresponding to the k target warning features from the knowledge base.
4. The method according to claim 1, characterized in that, For each of the aforementioned warning features, a question list is generated based on the large language model, the warning feature, k of the target warning features, and multiple question examples, including: For each of the aforementioned warning features, the meaning corresponding to the warning feature is analyzed using a large language model; A question list is generated by combining the meaning with k target warning features and multiple question examples using a large language model.
5. The method according to claim 1, characterized in that, The step of generating a handling tree based on the question list and connecting the handling tree to the outbound calling system, so that users can answer multiple rounds of questions in the handling tree, includes: The large language model is used to generate a disposal tree based on the problem list, wherein the non-leaf nodes of the disposal tree are problems, and the leaf nodes of the disposal tree are the numbers of the disposal methods; The processing tree is connected to the outbound calling system so that users can answer multiple rounds of questions in the processing tree.
6. A multi-round problem generation device for a risk control and early warning system, characterized in that, The device includes: A construction unit is used to acquire multiple transaction records and construct an early warning feature vector for each of the transaction records; The prediction unit is used to predict the warning feature vector using a preset warning model to obtain the prediction result; The acquisition unit is configured to, if the prediction result is greater than the warning threshold, acquire k target warning features and k dialogue templates corresponding to each warning feature in the warning feature vector from the knowledge base, wherein the target warning features are preset warning features similar to the warning features in the knowledge base; each dialogue template includes multiple question examples; The generation unit is used to generate a question list for each of the aforementioned warning features, based on a large language model, the warning feature, k of the aforementioned target warning features, and multiple question examples; The access unit is used to generate a handling tree based on the question list and connect the handling tree to the outbound calling system so that users can answer multiple rounds of questions in the handling tree.
7. The apparatus according to claim 6, characterized in that, The building unit includes: The first acquisition module is used to acquire multiple transaction records; The second acquisition module is used to acquire the values of multi-dimensional early warning features in each of the transaction records. A construction module is used to construct an early warning feature vector based on the values of the multi-dimensional early warning features.
8. The apparatus according to claim 6, characterized in that, The acquisition unit includes: The calculation module is used to calculate the cosine similarity between each warning feature in the warning feature vector and multiple preset warning features in the knowledge base. The third acquisition module is used to acquire k preset warning features corresponding to the first k cosine similarities in descending order of the multiple cosine similarities, and mark them as target warning features. The fourth acquisition module is used to acquire k script templates corresponding to k target warning features from the knowledge base.
9. The apparatus according to claim 6, characterized in that, The generation unit includes: The parsing module is used to parse the meaning of each warning feature using a large language model. The generation module is used to generate a question list by combining the meaning with k target warning features and multiple question examples through a large language model.
10. The apparatus according to claim 6, characterized in that, The access unit includes: A disposal tree generation module is used to generate a disposal tree based on the problem list using the large language model, wherein the non-leaf nodes of the disposal tree are problems, and the leaf nodes of the disposal tree are the numbers of disposal methods; The access module is used to connect the processing tree to the outbound calling system so that users can answer multiple rounds of questions in the processing tree.