Bank customer group screening instruction intent recognition method based on contrast learning

By using a comparative learning method to generate path fragments and calculate similarity, the problem of subjective judgment in the correspondence between bank customer managers' screening operations and customer group tags is solved, realizing automated filtering and accurate generation of customer lists, and reducing the repeated flow and delay of screening tasks.

CN123118232APending Publication Date: 2026-11-17FUJIAN ZHUOFONG INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611600720.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-16
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, the correspondence between bank account managers' screening operations and customer group tags relies on subjective judgment, resulting in a lack of record association between customer list extraction and screening operations. Furthermore, when customer lists are submitted for approval, they are not filtered according to account management relationships and scope of authority, leading to repeated processing of screening tasks and delays in customer list generation.

Method used

A contrastive learning-based method for recognizing bank customer screening instructions is adopted. By collecting the correspondence between screening operations and customer tags, path segments are generated and positive and negative sample pairs are constructed. The similarity is calculated using the path feature table to filter the customer list and reduce repeated transfers.

Benefits of technology

By automatically comparing path feature records with customer group prototype records, the repeated transfer of screening tasks between the customer manager terminal and the approval terminal is reduced, the customer list generation time is shortened, and the accuracy and efficiency of the operation are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118232A_ABST
    Figure CN123118232A_ABST
Patent Text Reader

Abstract

The application discloses a bank customer group screening instruction intention recognition method based on contrast learning, belongs to the technical field of customer group screening, and specifically comprises the following steps: collecting screening operations and customer group labels, comparing a customer identifier set to establish a corresponding record set; generating path segments according to operation elements and sequences and composing positive and negative sample pairs, inputting path feature tables to obtain path feature records, and obtaining customer group prototype records by weighted average according to task execution times; calculating path contrast loss and updating the path feature table and the customer group prototype record; receiving a current screening operation to obtain a current path feature record, calculating the similarity with each customer group prototype record to determine a recognition result; filtering to generate a customer list according to a relationship with a bank account holder, a permission range and an approval node, writing the customer list into a screening task work order and sending the customer list to a customer manager terminal after approval. The application is used for reducing the repeated circulation of screening tasks between a customer manager terminal and an approval terminal and shortening the generation time of a customer list.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of customer screening technology, and more specifically to a method for recognizing the intent of bank customer screening instructions based on contrastive learning. Background Technology

[0002] As the scale of banking business continues to expand, bank account managers need to filter target customer groups from the customer management system according to different business objectives, such as identifying loan loss customers, finding customers who have recently settled loans, and locating customers holding specific products. The customer management system has multiple filter controls and filter criteria. Bank account managers select filter controls, fill in filter criteria, and arrange the operation sequence according to business objectives to form a filtering operation. When different bank account managers perform filtering operations on the same customer group tag, there are differences in the selection of filter controls, the filling in of filter criteria, and the order of operation. Even when the same bank account manager performs filtering operations on different customer group tags at different times, there is overlap in filter controls and filter criteria. The customer management system stores customer group tag records, which contain a set of customer group tags and customer identifiers. The filtering task, from the bank account manager initiating the filtering operation to generating the customer list and submitting it for approval at the approval terminal, involves multiple stages of flow between the account manager's terminal, the customer management system, and the approval terminal.

[0003] In existing technologies, after a bank account manager completes a screening operation, they typically determine the corresponding customer group tags and then extract the customer list from the customer management system based on these tags. However, the correspondence between the screening operation and the customer group tags relies on the bank account manager's judgment. The process of extracting the customer list lacks verifiable records linking the screening operation and the customer group tags. When a discrepancy occurs, there is no basis for comparison between the customer list and the screening operation, necessitating a re-execution of the screening operation. Furthermore, screening operation records, customer group tag records, account management relationship records, authority scope records, and approval node records are stored in different data tables. When a screening task order is submitted to the approval terminal, the customer list is directly extracted based on the bank account manager's screening results, without any filtering step between the customer management relationship records and the authority scope records. Consequently, the screening task order received by the approval terminal may contain customer identifiers outside the bank account manager's authority scope, requiring the approval terminal to return the screening task order for the bank account manager to re-execute the screening operation. The above process causes the screening task to be repeatedly transferred between the account manager's terminal and the approval terminal, which prolongs the time for generating the customer list. Account managers are unable to obtain the customer list in a timely manner after the screening operation is completed and carry out subsequent business follow-up. Summary of the Invention

[0004] The purpose of this invention is to provide a method for recognizing the intent of bank customer screening instructions based on contrastive learning, and to solve the following technical problems: In existing technologies, the correspondence between screening operations and customer tags relies on subjective judgment. Customer list extraction and screening operations lack record association, and customer list submission for approval does not filter according to management relationship, scope of authority, and approval node, resulting in repeated processing of screening tasks and delays in customer list generation.

[0005] The objective of this invention can be achieved through the following technical solutions: A method for recognizing the intent of bank customer screening instructions based on contrastive learning includes the following steps: S1. Collect the filtering operations and customer group tags in the customer management system, compare them with the customer identifier set, and establish the correspondence between the filtering operations and customer group tags when the set ratio is reached, so as to obtain the corresponding record set. S2. Based on the corresponding record set, for the filtering operation of the same customer group label, generate path segments according to the operation elements and order. Path segments of the same customer group label form positive sample pairs, and path segments of different customer group labels form negative sample pairs. S3. Input positive and negative sample pairs into the path feature table to obtain path feature records. Calculate the customer group prototype records by weighted average according to the number of task executions and store them in the customer group label library. S4. Calculate the path comparison loss based on the distance between path feature records and the distance with the customer group prototype records, and update the path feature table and the customer group prototype records in the customer group tag library. S5. Receive the current filtering operation, obtain the current path feature record after updating the path feature table, calculate the similarity with the prototype record of each customer group, and determine the customer group label with the highest similarity as the recognition result. S6. Based on the identification results, filter the customer list according to the relationship between the account manager and the scope of authority and the approval node, write it into the screening task work order, and send it to the account manager terminal after approval. Add the current screening operation and identification results to the corresponding record set.

[0006] As a further aspect of the present invention, the specific process of S1 is as follows: The system reads the customer identifier set from the filtering operation and the customer identifier set from the customer group tag record from the customer management system. It then compares the intersection of the filtering result customer identifier set and the customer identifier set, and calculates the ratio of the number of customer identifiers in the intersection comparison result to the number of customer identifiers in the union comparison result. When the ratio reaches a set ratio, it establishes a correspondence between the filtering operation and the customer group tag. When the ratio does not reach the set ratio, it puts the filtering operation into the pending record set and sends the pending record set to the customer group tag maintenance terminal. It receives the customer group tags returned by the customer group tag maintenance terminal and adds the returned customer group tags and the filtering operation in the pending record set to the corresponding record set.

[0007] As a further aspect of the present invention, the specific process of S2 is as follows: For filtering operations on the same customer group label in the corresponding record set, read the filter control identifier, filter condition value, and operation order in the filtering operation; when there are multiple filter control identifiers or multiple filter condition values ​​in the filtering operation, determine the filter control identifier according to the matching result of the filter control identifier in the customer group label field, determine the filter condition value according to the matching result of the filter condition value in the filter control in the filtering interface, form a filter condition item by combining the filter control identifier and the filter condition value, arrange the filter condition items according to the operation order to obtain path segments, form a positive sample pair by any two path segments of the same customer group label, and form a negative sample pair by any two path segments of different customer group labels.

[0008] As a further aspect of the present invention, the specific process of S3 is as follows: Input the path segments from the positive and negative sample pairs into the path feature table; the path feature table reads the filter control identifier, filter condition value, and operation order from the path segments, converts the filter control identifier, filter condition value, and operation order into control record, condition value record, and sequence record respectively, and concatenates them according to the operation order to obtain the path feature record; for all path feature records under the same customer group label, read the number of task executions of the filter operation corresponding to each path feature record from the customer manager terminal operation log, determine the number of task executions as the weight of the path feature record, add all path feature records according to their weights and divide by the total weight to obtain the customer group prototype record, and store the customer group prototype record in the customer group label library.

[0009] As a further aspect of the present invention: In step S4, the specific process of calculating the path comparison loss based on the distance between path feature records and the distance between the path feature records and the customer group prototype records is as follows: For each positive sample pair, calculate the positive sample pair distance between the two path feature records and add the positive sample pair distance to the path sample comparison part; for each negative sample pair, calculate the negative sample pair distance between the two path feature records, invert the negative sample pair distance and add it to the path sample comparison part; for each path segment, calculate the prototype distance between the path feature record and the prototype record of the customer group to which it belongs, add the prototype distance to the prototype comparison part; calculate the non-prototype distance between the path feature record and the non-prototype record of the customer group to which it belongs, invert the non-prototype distance and add it to the prototype comparison part; add the path sample comparison part and the prototype comparison part to the customer group according to a set ratio to obtain the path comparison loss.

[0010] As a further aspect of the present invention: the specific process of updating the path feature table and the customer prototype record in the customer tag library in step S4 is as follows: The parameters of the path feature table are updated based on the path comparison loss. The current round customer group prototype record is obtained by adding all path feature records under the same customer group label in the current round according to their weights and dividing by the total weights. This is then added to the previous round customer group prototype record according to a set ratio to obtain the updated current round customer group prototype record. The updated current round customer group prototype record replaces the previous round customer group prototype record in the customer group label library. When calculating the path comparison loss in the next round, the updated path feature table and the updated customer group prototype record are used.

[0011] As a further aspect of the present invention, the specific process of S5 is as follows: The system receives the current filtering operation from the account manager, reads the filter control identifier, filter condition value, and operation order from the current filtering operation, inputs the current filtering operation into the updated path feature table to obtain the current path feature record, calculates the similarity between the current path feature record and each customer group prototype record in the customer group tag library, and sorts them from largest to smallest. When the similarity of the first sorted record is not lower than the set similarity threshold, the customer group tag corresponding to the first sorted similarity is determined as the recognition result. When the similarity of the first sorted record is lower than the set similarity threshold, the current filtering operation is placed in the pending record set and sent to the customer group tag maintenance terminal. The system receives the customer group tag returned by the customer group tag maintenance terminal as the recognition result, and adds the current filtering operation and the returned customer group tag to the corresponding record set.

[0012] As a further aspect of the present invention, the specific process of S6 is as follows: Based on the identification results, query the customer group tag records in the customer management system, read the customer identifier set in the customer group tag records as the initial customer list, read the customer manager's account management relationship record and permission scope record from the customer management system, delete the customer identifiers in the initial customer list that do not belong to the account management relationship record or exceed the permission scope record, and obtain the customer list. Write the customer list, identification results, and current filtering operation into the filtering task work order, send it to the approval terminal, receive the approval result returned by the approval terminal, when the approval result is approved, send the approval result and the filtering task work order to the customer manager terminal, add the current filtering operation and identification results to the corresponding record set, when the approval result is rejected, put the current filtering operation and identification results into the pending record set, and send it to the customer group tag maintenance terminal.

[0013] The beneficial effects of this invention are: This invention generates path segments based on operational elements and sequence during the screening operation. Path segments with the same customer group label form positive sample pairs, and path segments with different customer group labels form negative sample pairs. Path feature records are obtained through a path feature table, and a customer group prototype record is obtained by weighted averaging based on the number of task executions. The path comparison loss is calculated based on the distance between path feature records and the distance between path feature records and customer group prototype records, and the path feature table and customer group prototype records are updated accordingly. This transforms the correspondence between screening operations and customer group labels into a comparison relationship between path feature records and customer group prototype records, replacing the judgment of bank account managers and reducing the number of times screening operations are re-executed. On this basis, when the customer list is generated, customer identifiers that do not belong to the account management relationship or exceed the authorization scope are deleted according to the account management relationship record and the authorization scope record. The customer list, identification results, and current screening operation are written into the screening task work order and sent to the approval terminal. After receiving the approval result, it is sent to the account manager terminal. Thus, the filtering is completed before the customer list enters the approval stage, reducing the repeated flow of the screening task between the account manager terminal and the approval terminal and shortening the customer list generation time. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0016] 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.

[0017] Please see Figure 1 As shown, this invention is a method for recognizing the intent of bank customer screening instructions based on contrastive learning, including the following steps: S1. Collect the filtering operations and customer group tags in the customer management system, compare them with the customer identifier set, and establish the correspondence between the filtering operations and customer group tags when the set ratio is reached, so as to obtain the corresponding record set. S2. Based on the corresponding record set, for the filtering operation of the same customer group label, generate path segments according to the operation elements and order. Path segments of the same customer group label form positive sample pairs, and path segments of different customer group labels form negative sample pairs. S3. Input positive and negative sample pairs into the path feature table to obtain path feature records. Calculate the customer group prototype records by weighted average according to the number of task executions and store them in the customer group label library. S4. Calculate the path comparison loss based on the distance between path feature records and the distance with the customer group prototype records, and update the path feature table and the customer group prototype records in the customer group tag library. S5. Receive the current filtering operation, obtain the current path feature record after updating the path feature table, calculate the similarity with the prototype record of each customer group, and determine the customer group label with the highest similarity as the recognition result. S6. Based on the identification results, filter the customer list according to the relationship between the account manager and the scope of authority and the approval node, write it into the screening task work order, and send it to the account manager terminal after approval. Add the current screening operation and identification results to the corresponding record set.

[0018] In a preferred embodiment of the present invention, the specific process of S1 is as follows: The system reads the customer identifier set from the filtering operation and the customer identifier set from the customer group tag record from the customer management system. A filtering operation is a complete operation performed by a bank account manager on the customer filtering interface, including filter control identifiers, filter condition values, operation order, and the filtered result customer identifier set. The filtered result customer identifier set is the set of customer identifiers obtained after this filtering operation, such as C001, C002, C003 to C100. The customer group tag record is the corresponding record of customer group tags and customer identifier sets stored in the customer management system. For example, the customer group tag "Loan Loss" corresponds to customer identifiers C001, C002, and C005, and the customer group tag "Recent Settlement" corresponds to C003, C004, and C006.

[0019] The process involves comparing the intersection of the filtered customer identifier set with the customer identifier set in the customer group tag records to obtain the customer identifiers commonly included in both sets. For example, if the filtered result set contains C001, C002, C003, and C004, and the customer group tag record set contains C001, C002, C005, and C006, the intersection is C001 and C002. The union comparison yields all customer identifiers included in both sets, for example, C001, C002, C003, C004, C005, and C006. The ratio of the number of customer identifiers in the intersection comparison to the number of customer identifiers in the union comparison is calculated. If the intersection count is 2 and the union count is 6, the ratio is 2 divided by 6. This ratio is preset according to business needs, for example, 0.5. When the ratio reaches the preset ratio, a correspondence is established between the filtering operation and the customer group tags. When the ratio is lower than the preset ratio, the filtering operation is placed in the pending record set. The set of records to be determined is a collection used to store filtering operations for which a direct correspondence cannot be established.

[0020] The pending record set is sent to the customer group tag maintenance terminal, which is an internal bank device or workbench responsible for maintaining customer group tags. The system receives the customer group tags returned by the customer group tag maintenance terminal for each filtering operation in the pending record set and adds the returned customer group tags and their corresponding filtering operations in the pending record set to the corresponding record set. This corresponding record set stores the correspondence between filtering operations and customer group tags. Subsequently, S2 reads filtering operations for the same customer group tag from this corresponding record set. Through this process, the correspondence between filtering operations and customer group tags is determined based on the intersection-union ratio of the customer identifier sets, providing the corresponding record set for subsequent path segment generation.

[0021] In a preferred embodiment of the present invention, the specific process of S2 is as follows: For filtering operations on the same customer group tag within the corresponding record set, the filter control identifier, filter condition value, and operation sequence are read. The filter control identifier is the name or code of the filter control in the customer filtering interface, such as "Loan Status," "Loan Balance," or "Last Settlement Time." The filter condition value is the value corresponding to the filter control, such as "Settled," "Greater than 100,000," or "Last 30 Days." The operation sequence is the order in which the bank's account manager selects the filter control and enters the filter condition value in the filtering interface; for example, first select "Loan Status," then enter "Settled," then select "Last Settlement Time," and then enter "Last 30 Days."

[0022] When multiple filter control identifiers or multiple filter condition values ​​exist in the filtering operation, the filter control identifier is checked to see if it matches the field name in the customer group label field. If they match, the filter control identifier is determined to be a valid filter control identifier. The filter condition value is checked to see if it falls within the selectable value range of the filter control in the filtering interface. If it does, the filter condition value is determined to be a valid filter condition value. The filter control identifier and the filter condition value are combined to form filter condition items. For example, "Loan Status_Settled" is one filter condition item, and "Recent Settlement Time_Last 30 Days" is another. The filter condition items are arranged in the order of operation to obtain path fragments. For example, the path fragments are "Loan Status_Settled; Recent Settlement Time_Last 30 Days". Any two path fragments of the same customer group label are combined into a positive sample pair, and any two path fragments of different customer group labels are combined into a negative sample pair. A positive sample pair indicates that the two path fragments correspond to the same customer group label, and a negative sample pair indicates that the two path fragments correspond to different customer group labels. Through the above process, the filtering operation is converted into path fragments, providing input for subsequent path feature records and customer group prototype records.

[0023] In a preferred embodiment of the present invention, the specific process of S3 is as follows: The path segments from positive and negative sample pairs are input into the path feature table. The path feature table is a pre-established mapping table that stores conversion rules from filter control identifiers to control records, filter condition values ​​to condition value records, and operation sequences to sequence records. The path feature table reads the filter control identifiers, filter condition values, and operation sequences from the path segments, converting the filter control identifiers to control records (e.g., "Loan Status" to the control record "Loan Status"); converting the filter condition values ​​to condition value records (e.g., "Settled" to the condition value record "Settled"); and converting the operation sequences to sequence records (e.g., "1" to the sequence record "1"). The control records, condition value records, and sequence records are then concatenated according to the operation sequence to obtain the path feature record, for example, "Loan Status_Settled_1; Recent Settlement Time_Last 30 Days_2".

[0024] For all path feature records under the same customer group label, the number of times the filtering operation corresponding to each path feature record is executed is read from the account manager terminal operation log, and the number of task executions is determined as the weight of the path feature record. For example, if the filtering operation corresponding to a certain path feature record is executed 5 times, the weight of the path feature record is 5; if another path feature record is executed 3 times, the weight is 3; and if another is executed 2 times, the weight is 2. For all path feature records under the same customer group label, the weighted occurrence count of each control record and condition value record at each operation sequence position is calculated according to the weight. The weighted occurrence count refers to the value obtained by adding the weight of the path feature record to the corresponding control record and condition value record. For example, in operation sequence 1, the weighted frequency of the control record "Loan Status" is 5, the weighted frequency of the control record "Loan Balance" is 3, and the weighted frequency of the control record "Settlement Time" is 2, with a total weight of 10. Therefore, the control record "Loan Status" with the highest weighted frequency is identified as the control record of the customer prototype record in operation sequence 1. Similarly, the weighted frequency of the conditional value record "Settled" is 5, the weighted frequency of the conditional value record "Greater than 100000" is 3, and the weighted frequency of the conditional value record "Last 30 Days" is 2. Therefore, the conditional value record "Settled" with the highest weighted frequency is identified as the conditional value record of the customer prototype record in operation sequence 1. The same statistics are performed on operation sequence 2 to obtain the control records and conditional value records of the customer prototype record in operation sequence 2.

[0025] The control records, condition value records, and sequence records at each operation sequence position are arranged by operation sequence number to obtain the customer group prototype record. This customer group prototype record is then stored in the customer group tag library. The customer group tag library is a collection used to store the customer group prototype records corresponding to each customer group tag. Through the above process, multiple path feature records under the same customer group tag are summarized into a single customer group prototype record, which is used to calculate similarity after receiving the current filtering operation.

[0026] In a preferred embodiment of the present invention, the specific process of calculating the path comparison loss based on the distance between path feature records and the distance with the customer group prototype records in step S4 is as follows: Path feature records consist of control records, condition value records, and sequence records arranged in the order of operation. The distance between path feature records measures the degree of difference between two path feature records in their operation sequence positions. The distance between path feature records is calculated as follows: Control records and condition value records at the same operation sequence position are compared one by one, and the number of inconsistent operation sequence positions in control records and condition value records is counted. When the number of inconsistent operation sequence positions is inconsistent between two path feature records, the one with the larger number of positions is used as the benchmark, and the control records and condition value records at the missing positions are recorded as inconsistent and included in the difference count. The number of inconsistent control records and the number of inconsistent condition values ​​are added together to obtain the number of differences between the two path feature records, and this number is determined as the distance between the two path feature records. For example, path feature record A has "Loan Status_Settled" in operation sequence 1 and "Recent Settlement Time_Last 30 Days" in operation sequence 2; path feature record B has "Loan Status_Settled" in operation sequence 1 and "Loan Balance_Greater than 100,000" in operation sequence 2. Both path feature records have the same control records and condition value records in operation sequence 1, but inconsistent control records and condition value records in operation sequence 2. The number of differences is 2, and the distance between path feature record A and path feature record B is 2. The smaller the distance between path feature records, the closer the contents of the two path feature records are in their operation sequence positions.

[0027] For each positive sample pair, calculate the positive sample pair distance between the two path feature records. A positive sample pair consists of any two path segments with the same customer group label, and the two path feature records in the positive sample pair correspond to the same customer group label. Add the positive sample pair distance to the path sample comparison part. The path sample comparison part is used to accumulate the distances between the path feature records of positive and negative sample pairs; it is a numerical accumulation part. For each negative sample pair, calculate the negative sample pair distance between the two path feature records. A negative sample pair consists of any two path segments with different customer group labels, and the two path feature records in the negative sample pair correspond to different customer group labels. Invert the negative sample pair distance and add it to the path sample comparison part. Inverting means adding a negative sign before the negative sample pair distance value; for example, a negative sample pair distance of 3 becomes -3 after inversion. Adding the inverted negative sample pair distance to the path sample comparison part ensures that the path sample comparison part has a smaller value when the positive sample pair distance is small and a smaller value when the negative sample pair distance is large.

[0028] For each path segment, calculate the prototype distance between the path feature record and its corresponding customer group prototype record. The corresponding customer group prototype record is the customer group prototype record in the customer group tag library for the path segment. The prototype distance is calculated in the same way as the distance between path feature records, i.e., comparing the control records and condition value records in the same operation sequence position of the path feature record and the corresponding customer group prototype record one by one, counting the number of inconsistencies, and determining the number of inconsistencies as the prototype distance. The prototype distance is added to the customer group prototype comparison part. Calculate the non-prototype distance between the path feature record and non-customer group prototype records. Non-customer group prototype records are the customer group prototype records of other customer group tags in the customer group tag library besides the customer group tag corresponding to the path segment. The non-prototype distance is calculated in the same way as the distance between path feature records. The non-prototype distance is inverted and added to the customer group prototype comparison part. The meaning of inversion is the same as the inversion method of the non-prototype distance, i.e., adding a negative sign before the value of the non-prototype distance. The path sample comparison part and the customer group prototype comparison part are added according to a set ratio to obtain the path comparison loss. The set ratio is preset according to business needs, for example, 0.5. The path comparison loss is used to update the parameters of the path feature table.

[0029] In a preferred embodiment of the present invention, the specific process of updating the path feature table and the customer prototype records in the customer tag library in step S4 is as follows: The parameters of the path feature table are adjustable values ​​from the conversion rules stored in the table: conversion rules from filter control identifiers to control records, conversion rules from filter condition values ​​to condition value records, and conversion rules from operation order to sequence records. The parameters of the path feature table are updated based on the path comparison loss. The update is performed as follows: when the path comparison loss value is greater than the previous round's value, the values ​​of each conversion rule in the path feature table are adjusted to decrease the distance between positive sample pairs and increase the distance between negative sample pairs; when the path comparison loss value is less than or equal to the previous round's value, the values ​​of each conversion rule in the path feature table remain unchanged. After the path feature table parameters are updated, the updated path feature table is used in the next round of path comparison loss calculation.

[0030] The current round customer group prototype record is obtained by adding all path feature records under the same customer group tag in the current round according to their weights and dividing by the total weights. This sum is then added to the previous round customer group prototype record according to a set ratio to obtain the updated current round customer group prototype record. The current round customer group prototype record is obtained in the same way as in S3: for all path feature records under the same customer group tag, the weighted occurrence count of each control record and condition value record at each operation sequence position is calculated according to their weights. The control record and condition value record with the highest weighted occurrence count are determined as the content of the current round customer group prototype record at that operation sequence position. The control records, condition value records, and sequence records at each operation sequence position are arranged according to their operation sequence number to obtain the current round customer group prototype record. The previous round customer group prototype record is the customer group prototype record stored in the customer group tag library from the previous round. The set ratio is preset according to business needs, for example, 0.5.

[0031] The meaning of adding according to the set ratio is as follows: when the control record or condition value record of the current round customer group prototype record is consistent with the control record or condition value record of the previous round customer group prototype record in the same operation sequence position, the record is retained; when they are inconsistent, the weighted occurrence counts of the two records in the current round and the previous round are compared, and the record with the larger weighted occurrence count is determined as the content of the updated current round customer group prototype record in that operation sequence position; when the weighted occurrence counts are equal, the content of the previous round customer group prototype record in that operation sequence position is retained. For example, if the current round customer group prototype record's control record at position 1 of the operation sequence is "Loan Status," and the previous round customer group prototype record's control record at position 1 of the operation sequence is also "Loan Status," then the updated control record for the current round customer group prototype record at position 1 of the operation sequence will be "Loan Status." If the current round customer group prototype record's control record at position 2 of the operation sequence is "Recent Settlement Time," and the previous round customer group prototype record's control record at position 2 of the operation sequence is "Loan Balance," then the two records are inconsistent. The weighted occurrence count for the current round customer group prototype record at position 2 of the operation sequence is 6, and the weighted occurrence count for the previous round customer group prototype record at position 2 of the operation sequence is 4. Therefore, the control record with the higher weighted occurrence count, "Recent Settlement Time," will be determined as the updated control record for the current round customer group prototype record at position 2 of the operation sequence. The updated current round customer group prototype record will then replace the previous round customer group prototype record in the customer group tag library. In the next round of calculating the path comparison loss, path segments from positive and negative sample pairs are input into the updated path feature table to obtain new path feature records. Records of belonging to a specific customer group and records of non-belonging customer groups are read from the updated customer group label library. Through this process, the path feature table and the customer group prototype records in the customer group label library are updated progressively according to the round, and are used to calculate similarity after receiving the current filtering operation.

[0032] In a preferred embodiment of the present invention, the specific process of S5 is as follows: Receive the current filtering operation from the account manager. The current filtering operation is the action performed by the bank account manager in the customer filtering interface, including the filter control identifier, filter condition values, and operation order. The filter control identifier is the name or code of the filter control in the customer filtering interface, such as "Loan Status," "Loan Balance," or "Last Settlement Time." The filter condition values ​​are the values ​​corresponding to the filter controls, such as "Settled," "Greater than 100,000," or "Last 30 Days." The operation order is the order in which the bank account manager selects the filter control and enters the filter condition values ​​in the filtering interface; for example, first selecting "Loan Status," then entering "Settled," then selecting "Last Settlement Time," and then entering "Last 30 Days." Read the filter control identifier, filter condition values, and operation order from the current filtering operation.

[0033] The current filtering operation is input into the updated path feature table to obtain the current path feature record. The updated path feature table is the path feature table updated after the path comparison loss in S4. The path feature table reads the filter control identifier, filter condition value, and operation order from the current filtering operation, converts the filter control identifier into a control record, the filter condition value into a condition value record, and the operation order into a sequence record. The control record, condition value record, and sequence record are arranged according to the operation order number to obtain the current path feature record. For example, if the current filtering operation is "Loan Status_Cleared" at position 1 of the operation order and "Recent Clearance Time_Last 30 Days" at position 2 of the operation order, then the current path feature record is "Loan Status_Cleared_1; Recent Clearance Time_Last 30 Days_2".

[0034] Calculate the similarity between the current path feature record and each customer group prototype record in the customer group tag library, and sort them in descending order. The customer group tag library is a collection used to store the customer group prototype records corresponding to each customer group tag. The similarity is calculated as follows: Compare the control records and condition value records at the same operation sequence positions of the current path feature record and the customer group prototype record one by one, and count the number of operation sequence positions where both control records and condition value records are consistent. When the number of operation sequence positions of the current path feature record and the customer group prototype record is inconsistent, the one with the larger number of operation sequence positions is used as the benchmark, and the control records and condition value records at the missing positions are recorded as inconsistent and included in the similarity calculation. Divide the number of consistent operation sequence positions by the total number of operation sequence positions to obtain the similarity between the current path feature record and the customer group prototype record. For example, if the current path feature record has 2 operation sequence positions, and it is consistent with a certain customer group prototype record in both control records and condition value records at operation sequence position 1, but inconsistent in both control records and condition value records at operation sequence position 2, the number of consistent operation sequence positions is 1, the total number of operation sequence positions is 2, and the similarity is 1 divided by 2. Similarity scores range from 0 to 1. The similarity scores are then sorted from highest to lowest.

[0035] When the similarity score of the top-ranked item is not lower than the set similarity threshold, the customer tag corresponding to the top-ranked similarity score is determined as the recognition result. The set similarity threshold is preset according to business needs, for example, 0.6. When the similarity score of the top-ranked item is lower than the set similarity threshold, the current filtering operation is placed in the pending record set and sent to the customer tag maintenance terminal. The pending record set is used to store filtering operations whose recognition results cannot be directly determined. The customer tag maintenance terminal is a terminal device or workbench within the bank responsible for maintaining customer tags. The customer tag maintenance terminal receives the customer tags returned by the customer tag maintenance terminal as the recognition result. The current filtering operation and the returned customer tags are added to the corresponding record set. The corresponding record set is the set defined in S1 used to store the correspondence between filtering operations and customer tags. Through the above process, the customer tags corresponding to the current filtering operation are determined according to the similarity sorting and the set similarity threshold. When the similarity is lower than the set similarity threshold, the customer tags are returned by the customer tag maintenance terminal for subsequent generation of customer lists based on the recognition results.

[0036] In a preferred embodiment of the present invention, the specific process of S6 is as follows: The customer group tag records in the customer management system are queried based on the recognition results. The customer management system is a data processing device used by banks to manage customer information and customer group tags. The recognition results are the customer group tags determined in S5. The customer group tag records are the corresponding records of customer group tags and customer identifier sets stored in the customer management system. The customer identifier set in the customer group tag records is read as the initial customer list. The initial customer list is the set of all customer identifiers under the customer group tag corresponding to the recognition result. For example, if the recognition result is "loan loss", and the customer group tag record for "loan loss" corresponds to customer identifiers C001, C002, C005, C007, and C008, then the initial customer list is C001, C002, C005, C007, and C008.

[0037] The system retrieves the account manager's account relationship records and access permission records from the customer management system. The account relationship record records the attribution relationships between bank account managers and customer identifiers. For example, account manager Zhang Ming corresponds to customer identifiers C001, C002, and C005, while account manager Li Hua corresponds to customer identifiers C003, C004, and C006. The access permission record records the range of customer identifiers that the bank account manager can access. For example, account manager Zhang Ming's access permission record records customer identifiers with the prefix C00, while account manager Li Hua's access permission record records customer identifiers with the prefix C01. Customer identifiers that do not belong to the account relationship records or exceed the access permission records in the initial customer list are deleted, resulting in the final customer list. For example, if the initial customer list is C001, C002, C005, C007, and C008, and the account manager Zhang Ming's account management relationship records are C001, C002, and C005, and the permission scope record is prefixed with C00, then C007 and C008 do not belong to the account management relationship records and are deleted from the initial customer list, and the customer list becomes C001, C002, and C005.

[0038] The customer list, identification results, and current filtering operation are written into the filtering task work order and sent to the approval terminal. The filtering task work order is a document used to record the processing of this filtering task, including the customer list, identification results, and current filtering operation. The approval terminal is the internal terminal device or workbench of the bank responsible for approving filtering tasks. The system receives the approval results returned by the approval terminal. Approval results include two states: approved and rejected. When the approval result is approved, the approval result and the filtering task work order are sent to the customer manager terminal, and the current filtering operation and identification results are added to the corresponding record set. The customer manager terminal is the terminal device used by the bank's customer managers. When the approval result is rejected, the current filtering operation and identification results are deleted from the corresponding record set, placed in the pending record set, and sent to the customer group label maintenance terminal. The pending record set is used to store filtering operations for which the identification result cannot be directly determined. The customer group label maintenance terminal is the internal terminal device or workbench of the bank responsible for maintaining customer group labels. Through the above process, the customer list is filtered according to the records of management relationship and scope of authority before entering the approval stage. After approval, it is sent to the customer manager terminal. When the approval is rejected, it is processed by the customer group tag maintenance terminal to reduce the repeated flow of screening tasks between the customer manager terminal and the approval terminal.

[0039] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A bank customer group screening instruction intent recognition method based on contrastive learning, characterized in that, Includes the following steps: S1. Collect the filtering operations and customer group tags in the customer management system, compare them with the customer identifier set, and establish the correspondence between the filtering operations and customer group tags when the set ratio is reached, so as to obtain the corresponding record set. S2. Based on the corresponding record set, for the filtering operation of the same customer group label, generate path segments according to the operation elements and order. Path segments of the same customer group label form positive sample pairs, and path segments of different customer group labels form negative sample pairs. S3. Input positive and negative sample pairs into the path feature table to obtain path feature records. Calculate the customer group prototype records by weighted average according to the number of task executions and store them in the customer group tag library. S4. Calculate the path comparison loss based on the distance between path feature records and the distance with the customer group prototype records, and update the path feature table and the customer group prototype records in the customer group tag library. S5. Receive the current filtering operation, obtain the current path feature record after updating the path feature table, calculate the similarity with the prototype record of each customer group, and determine the customer group label with the highest similarity as the recognition result. S6. Based on the identification results, filter the customer list according to the relationship between the account manager and the scope of authority and the approval node, write it into the screening task work order, and send it to the account manager terminal after approval. Add the current screening operation and identification results to the corresponding record set.

2. The method for recognizing bank customer screening instruction intent based on contrastive learning according to claim 1, characterized in that, The specific process of S1 is as follows: The system reads the customer identifier set from the filtering operation and the customer identifier set from the customer group tag record from the customer management system. It then compares the intersection of the filtering result customer identifier set and the customer identifier set, and calculates the ratio of the number of customer identifiers in the intersection comparison result to the number of customer identifiers in the union comparison result. When the ratio reaches a set ratio, it establishes a correspondence between the filtering operation and the customer group tag. When the ratio does not reach the set ratio, it puts the filtering operation into the pending record set and sends the pending record set to the customer group tag maintenance terminal. It receives the customer group tags returned by the customer group tag maintenance terminal and adds the returned customer group tags and the filtering operation in the pending record set to the corresponding record set.

3. The method for recognizing bank customer screening instruction intent based on contrastive learning according to claim 1, characterized in that, The specific process of S2 is as follows: For filtering operations on the same customer group label in the corresponding record set, read the filter control identifier, filter condition value, and operation order in the filtering operation; when there are multiple filter control identifiers or multiple filter condition values ​​in the filtering operation, determine the filter control identifier according to the matching result of the filter control identifier in the customer group label field, determine the filter condition value according to the matching result of the filter condition value in the filter control in the filtering interface, form a filter condition item by combining the filter control identifier and the filter condition value, arrange the filter condition items according to the operation order to obtain path segments, form a positive sample pair by any two path segments of the same customer group label, and form a negative sample pair by any two path segments of different customer group labels.

4. The method for recognizing bank customer screening instruction intent based on contrastive learning according to claim 1, characterized in that, The specific process of S3 is as follows: Input the path segments from the positive and negative sample pairs into the path feature table; the path feature table reads the filter control identifier, filter condition value, and operation order from the path segments, converts the filter control identifier, filter condition value, and operation order into control record, condition value record, and sequence record respectively, and concatenates them according to the operation order to obtain the path feature record; for all path feature records under the same customer group label, read the number of task executions of the filter operation corresponding to each path feature record from the customer manager terminal operation log, determine the number of task executions as the weight of the path feature record, add all path feature records according to their weights and divide by the total weight to obtain the customer group prototype record, and store the customer group prototype record in the customer group label library.

5. The bank customer group screening instruction intent recognition method based on contrastive learning according to claim 1, characterized in that, In step S4, the specific process of calculating the path comparison loss based on the distance between path feature records and the distance with the customer group prototype records is as follows: For each positive sample pair, calculate the positive sample pair distance between the two path feature records and add the positive sample pair distance to the path sample comparison part; for each negative sample pair, calculate the negative sample pair distance between the two path feature records, invert the negative sample pair distance and add it to the path sample comparison part; for each path segment, calculate the prototype distance between the path feature record and the prototype record of the customer group to which it belongs, add the prototype distance to the prototype comparison part; calculate the non-prototype distance between the path feature record and the non-prototype record of the customer group to which it belongs, invert the non-prototype distance and add it to the prototype comparison part; add the path sample comparison part and the prototype comparison part to the customer group according to a set ratio to obtain the path comparison loss.

6. The bank customer group screening instruction intent recognition method based on contrastive learning according to claim 5, characterized in that, In step S4, the specific process of updating the path feature table and the customer prototype records in the customer tag library is as follows: The parameters of the path feature table are updated based on the path comparison loss. The current round customer group prototype record is obtained by adding all path feature records under the same customer group label in the current round according to their weights and dividing by the total weights. This is then added to the previous round customer group prototype record according to a set ratio to obtain the updated current round customer group prototype record. The updated current round customer group prototype record replaces the previous round customer group prototype record in the customer group label library. When calculating the path comparison loss in the next round, the updated path feature table and the updated customer group prototype record are used.

7. The bank customer group screening instruction intent recognition method based on contrastive learning according to claim 1, characterized in that, The specific process of S5 is as follows: The system receives the current filtering operation from the account manager, reads the filter control identifier, filter condition value, and operation order from the current filtering operation, inputs the current filtering operation into the updated path feature table to obtain the current path feature record, calculates the similarity between the current path feature record and each customer group prototype record in the customer group tag library, and sorts them from largest to smallest. When the similarity of the first sorted record is not lower than the set similarity threshold, the customer group tag corresponding to the first sorted similarity is determined as the recognition result. When the similarity of the first sorted record is lower than the set similarity threshold, the current filtering operation is placed in the pending record set and sent to the customer group tag maintenance terminal. The system receives the customer group tag returned by the customer group tag maintenance terminal as the recognition result, and adds the current filtering operation and the returned customer group tag to the corresponding record set.

8. The bank customer group screening instruction intent recognition method based on contrastive learning according to claim 1, characterized in that, The specific process of S6 is as follows: Based on the identification results, query the customer group tag records in the customer management system, read the customer identifier set in the customer group tag records as the initial customer list, read the customer manager's account management relationship record and permission scope record from the customer management system, delete the customer identifiers in the initial customer list that do not belong to the account management relationship record or exceed the permission scope record, and obtain the customer list. Write the customer list, identification results, and current filtering operation into the filtering task work order, send it to the approval terminal, receive the approval result returned by the approval terminal, when the approval result is approved, send the approval result and the filtering task work order to the customer manager terminal, add the current filtering operation and identification results to the corresponding record set, when the approval result is rejected, put the current filtering operation and identification results into the pending record set, and send it to the customer group tag maintenance terminal.