Intelligent approval method and system, readable storage medium and computer equipment

Abstract

The embodiment of the invention provides an intelligent examination and approval method and system, a readable storage medium and computer equipment, and relates to the field of credit risk control examination and approval in the financial industry, and the method comprises the steps: decomposing an examination and approval problem into a basic query problem and an index analysis problem through a base large model; a base large language model is called, a basic query SQL is formed according to field assembly contained in the basic query problem, and the basic query SQL comprises a specific query field; calling a base large language model, assembling according to quantitative analysis indexes contained in the index analysis problem to form an index query SQL, and nesting the basic query SQL into the index query SQL to obtain a final query SQL; calling the database to execute the final query SQL, obtaining transaction record detail data according to the basic query SQL, and obtaining a quantitative analysis index statistical result according to the index query SQL. The SQL query statements corresponding to different questions are assembled for quantitative query, and large model illusion and answer instability can be made up.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of credit risk control and approval in the financial industry, and in particular to an intelligent approval method and system, a readable storage medium and computer equipment. BACKGROUND

[0002] When a financial institution conducts approval of a customer's financial transaction log file, it usually manually interprets and analyzes the financial transaction log file line by line, and in the process of line-by-line interpretation and analysis, some general auxiliary tools (such as Excel) or custom-developed special systems (log analysis systems based on fixed procedures, SQL and scripts) are used to reduce workload and improve approval quality.

[0003] In the process of implementing the present application, the applicant found that at least the following problems exist in the prior art: In actual practice, whether using general auxiliary tools (such as Excel) or custom-developed special systems, there are some problems that cannot be handled, such as unknown table patterns of financial transaction log files, only single search methods can be provided, only fixed indicators can be provided, and a large amount of repetitive work is generated for the approval personnel. SUMMARY

[0004] The embodiments of the present application provide an intelligent approval method, system, readable storage medium and computer equipment, which can solve the above technical problems in the prior art.

[0005] To achieve the above purpose, in a first aspect, the embodiments of the present application provide an intelligent approval method, comprising: After extracting and saving the to-be-recognized financial transaction log file provided by the to-be-approved customer with credit needs in the database, receiving an approval question raised by an approval personnel when approving the credit needs of the to-be-approved customer, and decomposing the approval question into a basic query question for querying the basic information of the transaction log and an index analysis question for querying the quantitative analysis index by using a base large language model; Calling the base large language model, assembling a basic query SQL for querying the basic information of the transaction log according to the fields contained in the basic query question, the basic query SQL including specific query fields; Calling the base large language model, assembling an index query SQL for querying the quantitative analysis index of the transaction log according to the quantitative analysis index contained in the index analysis question, and nesting the basic query SQL into the index query SQL to obtain a final query SQL; The database is invoked to execute the final query SQL, transaction record detail data is obtained for the basic query SQL, quantitative analysis index statistical results are obtained for the index query SQL, and quantitative analysis results are formed based on the transaction record detail data and the quantitative analysis index statistical results.

[0006] In a second aspect, an embodiment of the present application provides an intelligent approval system, comprising: An approval problem decomposition unit is configured to, after extracting and saving a to-be-identified financial transaction flow file provided by a to-be-approved client with credit needs in a database, receive an approval problem raised by an approval personnel when the approval personnel is approving the credit needs of the to-be-approved client, and decompose the approval problem into a basic query problem for querying basic information of the transaction flow and an index analysis problem for querying quantitative analysis indexes by using a base large language model. A query SQL assembly unit is configured to invoke the base large language model, assemble a basic query SQL for querying the basic information of the transaction flow according to fields contained in the basic query problem, the basic query SQL including specific query fields, and invoke the base large language model, assemble an index query SQL for querying the quantitative analysis indexes of the transaction flow according to quantitative analysis indexes contained in the index analysis problem, and nest the basic query SQL in the index query SQL to obtain a final query SQL. A quantitative analysis unit is configured to invoke a database, execute the final query SQL, obtain transaction record detail data for the basic query SQL, obtain quantitative analysis index statistical results for the index query SQL, and form quantitative analysis results based on the transaction record detail data and the quantitative analysis index statistical results.

[0007] In a third aspect, an embodiment of the present application provides a computer readable storage medium storing one or more programs, which when executed by a computer device, cause the computer device to perform the intelligent approval method described above.

[0008] In a fourth aspect, an embodiment of the present application provides a computer device, comprising: A processor and a memory arranged to store computer executable instructions, which when executed cause the processor to perform the intelligent approval method described above.

[0009] The technical scheme has the following beneficial effects: the format of a financial transaction flow file issued by any financial institution can be identified, and corresponding data in the flow table can be intelligently queried and statistically analyzed according to the questions of an approval personnel, while complex approval questions are decomposed into multiple simple questions, and the simple questions are assembled into a corresponding SQL query statement for quantitative query, which can make up for the shortcomings of the large model, such as illusion, unstable answers, and length constraints of user query text, compared with directly querying the large model according to the approval questions and the financial transaction flow data. BRIEF DESCRIPTION OF DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0011] Figure 1 is a flowchart of an intelligent approval method according to an embodiment of the present application; Figure 2 is a structural diagram of an intelligent approval system according to an embodiment of the present application; Figure 3 is a complete flowchart of import and analysis of a financial transaction flow file according to an embodiment of the present application. DETAILED DESCRIPTION

[0012] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0013] As shown in Figure 1 , according to an embodiment of the present application, an intelligent approval method is provided, comprising: S101: After extracting and saving a to-be-identified financial transaction flow file provided by a to-be-approved client with credit needs in a database, receiving an approval question raised by an approval personnel when approving the credit needs of the to-be-approved client, and decomposing the approval question into a basic query question for querying basic information of the transaction flow and an index analysis question for querying quantitative analysis indexes by a base large language model; S102: Calling the base large language model, assembling a basic query SQL for querying the basic information of the transaction flow according to the fields contained in the basic query question, the basic query SQL including specific query fields; S103: calling a base large language model, assembling a quantitative analysis index group contained in the problem according to the index to form an index query SQL for querying the quantitative analysis index of the transaction flow, embedding the basic query SQL into the index query SQL to obtain a final query SQL; S104: calling a database to execute the final query SQL, obtaining transaction record detail data for the basic query SQL, obtaining quantitative analysis index statistical results for the index query SQL, and forming a quantitative analysis result based on the transaction record detail data and the quantitative analysis index statistical results.

[0014] In the credit approval process of a financial institution, analyzing the financial transaction flow file provided by the customer is one of the important steps in risk identification. However, line-by-line interpretation and analysis of the financial transaction flow file is a time-consuming and labor-intensive process that is prone to errors and highly dependent on personal experience. In order to minimize repetitive work and improve approval quality, a system is needed that can automatically and intelligently complete related tasks.

[0015] The financial transaction flow file provided by the client with credit needs is extracted and saved in the database, and the problem of unknown table patterns not being able to be identified is solved by extracting and saving data from the financial transaction flow file to be identified.

[0016] The approval questions of the approval personnel are decomposed into different basic query questions and index analysis questions, and the corresponding basic query SQL and index query SQL are formed according to the respective field groups, so that different indexes can be queried according to different questions, achieving the diversity of index query. The base large language model is used to query the database to obtain quantitative analysis results. Decomposing the approval questions can make up for the shortcomings of large model illusion and unstable answers caused by directly inputting the approval questions into the base large language model. In addition, it reduces the workload of the approval personnel and improves the efficiency and accuracy of manual approval.

[0017] Preferably, the intelligent approval method further comprises S105: extracting and saving the financial transaction flow file provided by the client with credit needs in the database, comprising: S1051: receiving the original financial transaction flow file provided by the client with credit needs, and converting the original financial transaction flow file into a financial transaction flow file to be identified using the original expression method; S1052: Based on the base large language model, the metadata of the original financial transaction flow file is obtained by identifying the to-be-identified financial transaction flow file, and the metadata is represented in the original expression mode. The metadata is used to generate a table building SQL statement for storing data in the to-be-identified financial transaction flow file. S1053: Based on the metadata, the data in the to-be-identified financial transaction flow file is disassembled, and the table building SQL statement is used to write the disassembled data into the table in the database where the metadata is located.

[0018] The original financial transaction flow file is converted into a to-be-identified financial transaction flow file in the original expression mode, and the metadata is in the original expression mode in the original financial transaction flow file. This facilitates the to-be-approved client to quickly identify and correspond to the to-be-approved client in the later period. Compared with the prior art in which the fields of the original financial transaction flow files of different financial institutions are all converted into the same field, the reaction time of the to-be-approved client is reduced, and the to-be-approved client is more convenient.

[0019] The original financial transaction flow file of different financial institutions can be automatically identified, and the financial inflow can be automatically imported into the database without the intervention of developers (without manual table building and data import). It also does not need to go through a traditional IT development process (such as designing a table building and manually building a table, developing an extension interface, testing, and online data import). Only the original financial transaction flow file needs to be uploaded to the intelligent approval system to complete the conversion of the original financial transaction flow file to the data in the database table. Especially when the amount of work for connecting the unknown financial institution transaction flow format original financial transaction flow file is very small, almost zero, the work efficiency is greatly improved.

[0020] Preferably, S1051: receiving the original financial transaction flow file provided by the to-be-approved client with credit needs, converting the original financial transaction flow file into a to-be-identified financial transaction flow file in the original expression mode, comprising: The PDF format original financial transaction flow file provided by the to-be-approved client with credit needs is received, and the original financial transaction flow file is decrypted using triple data symmetric encryption (Triple Data Encryption Standard, 3DES). The decrypted original financial transaction flow file is de-noised, and de-noising includes de-watermarking and line folding. Line folding refers to the behavior of automatically moving the part exceeding the set width (such as window boundary, text box limit or page width) to the next line when a line of text exceeds the set width. The data in the original financial transaction file after denoising processing is extracted, and the data in the original financial transaction file is saved in a TXT format file in sequence. Meanwhile, the page break and / or page number in the original financial transaction file are identified by a large language model and a special tool program. According to the identified page break and / or page number, the corresponding page number and page break are set in the TXT format file. The broken words and / or broken lines in the original financial transaction file are supplemented to form complete words and / or complete line contents, and a to-be-identified financial transaction file is obtained.

[0021] In one example, a user A (such as an operation officer) responsible for importing the original financial transaction file uploads a PDF format original financial transaction file provided by a to-be-approved client through an interface of an intelligent approval system. The intelligent approval system reads the PDF format original financial transaction file through a PDF reading special tool, including: PDF format decryption, watermark removal, content reading, page break and page number identification, and broken word / broken line completion processing, and extraction to a TXT format file.

[0022] In the TXT format file, the original expression method in the original financial transaction file is used. Compared with the prior art in which the fields of the original financial transaction files of different financial institutions are all converted into the same field, the reaction time of the to-be-approved client is reduced, and the to-be-approved client is more convenient.

[0023] Preferably, S1052: identifying the to-be-identified financial transaction file based on a base large language model, obtaining metadata of the original financial transaction file, and representing the metadata in an original expression method, generating a table building SQL statement for storing data in the to-be-identified financial transaction file based on the metadata, including: The first page content of the to-be-identified financial transaction file is identified based on a base large language model, and metadata is extracted. The metadata is represented in an original expression method. The metadata at least includes: table header fields (each field in the table header is separately used as a column), original table header and page break (used as a column), amount format category (each field is used as a column), specific query field (each field is used as a column), to-be-approved client name, and financial institution issuing the original financial transaction file. The specific query field is derived from the table header field, and the table header field corresponds one by one to the fields included in the original table header of the to-be-identified financial transaction file.

[0024] The amount format category refers to how the transaction amount field in the original financial transaction file is composed, which is used for accurate judgment of the income and expenditure amount subsequently. The amount format category includes three categories: Class C: only "transaction amount" field, distinguish between income and expenditure by positive and negative signs; Class B: use "transaction amount" and "type of income and expenditure" two fields to represent, the amount is positive, and the type distinguishes between income and expenditure; Class A: use "expenditure amount" and "income amount" two fields to represent, the amount is positive.

[0025] Table header / page separator, refers to the original table header row or page separator in the original financial transaction flow file, used to inform the subsequent data warehousing program how to distinguish between file data and field name, to ensure the generality of the flow data warehousing processing program; the page separator is used to judge what is the page separator, and then the specified symbol is used as the page separator to realize the paging processing; wherein, the original table header is the original table header in the transaction flow file, containing the information field, all contents in the original table header are treated as a field in the metadata.

[0026] Specific query field name refers to the set of DB (database) field names required for subsequent feature analysis and query of transaction flow, such as transaction purpose (salary, pawn, etc.), which can also be transaction counterpart, transaction summary, addendum and remarks, etc.

[0027] Generate a table-building SQL statement to store the data in the to-be-identified financial transaction flow file, and carry all the fields included in the metadata in the table-building SQL statement, that is, the fields included in the table-building SQL statement correspond one by one with the metadata.

[0028] According to the metadata obtained above, first judge whether it is a new style of financial transaction flow file, that is, judge according to the three conditions of financial institution name, original table header and table-building SQL statement, when the metadata involved in the three conditions is completely the same as the metadata in the existing table, when the metadata involved in the three conditions does not match the metadata in the existing table, it is a new style, if it is a new style, first create a database table according to all the fields carried by the table-building SQL statement, then import the data in the to-be-identified financial transaction flow file in the original expression, otherwise, directly import the data in the to-be-identified financial transaction flow file into the existing table in the database.

[0029] Preferably, in S101, the approval question is decomposed into a basic query question for querying the basic information of the transaction flow and an index analysis question for querying the quantitative analysis index by the base large language model, including: The base large language model is called to perform word segmentation on the approval problem by the base large language model, extract a specific keyword representing a specific query field, and extract a table header field related to the specific keyword as a field contained in a basic query problem. The field contained in the basic query problem at least includes: the specific keyword, a transaction time range, and an amount range. The keyword of the quantitative analysis index is extracted to form an index analysis problem for querying the quantitative analysis index. The keyword of the quantitative analysis index at least includes the following fields: a transaction counterparty, a transaction month, a transaction total amount, a cumulative number of transaction pens, a transaction average amount, and an amount standard deviation.

[0030] The user B responsible for analysis (such as an approval personnel) can ask questions to the intelligent approval system (intelligent agent) (or generated in advance by the intelligent approval system), and the base large language model such as LLM (Large Language Model) analyzes and decomposes the approval personnel's question to obtain a basic query problem and an index analysis problem, especially extracts a table header field related to the specific keyword. The approval problem is decomposed, which can make up for the large model illusion and unstable answers caused by directly inputting the approval problem into the base large language model.

[0031] The basic query problem and the index analysis problem perform quantitative analysis on the original financial transaction flow file as needed. There are four types of quantitative analysis dimensions: transaction amount summary, transaction frequency, amount fluctuation range, and transaction time period. Based on these dimensions and in cooperation with specific keywords (specific to detailed specific query fields), the financial transaction flow file is queried and sorted, so that from thousands of financial transaction flow files, the transaction flow record meeting the specified specific query field can be quickly located, such as a transaction amount that is too high, or a transaction that is too frequent, or a transaction cycle that is stable, or a transaction record that occurs in a specific time period.

[0032] Preferably, S102: calling the base large language model, assembling a basic query SQL for querying the basic information of the transaction flow according to the fields contained in the basic query problem, the basic query SQL including the specific query field, including: The base large language model is called to perform word segmentation on the approval problem by the base large language model, extract a specific keyword representing a specific query field, and extract a table header field related to the specific keyword as a field contained in a basic query problem. The field contained in the basic query problem at least includes: the specific keyword, a transaction time range, and an amount range.

[0033] The base large language model is called, the quantitative analysis index set contained in the problem is analyzed according to the index, the index query SQL for querying the quantitative analysis index of the transaction flow is assembled, the index query SQL is composed of different indexes such as index 1, index 2, etc., the basic query SQL is nested in the index query SQL, and the final query SQL is obtained; the integration of the basic query SQL and the statistical query SQL is realized based on the SQL nesting mode according to the characteristics that the second SQL includes the first SQL is the SQL itself.

[0034] Preferably, S103: calling a database, executing the final query SQL, obtaining transaction record detail data for the basic query SQL, and obtaining quantitative analysis index statistical results for the index query SQL, the quantitative analysis index statistical results including four-dimensional data: transaction amount summary, transaction frequency, amount fluctuation range, and transaction time period, and being returned in a JSON set format, wherein KEY is a field name, and VALUE is a specific value.

[0035] Intelligent quantitative analysis can automatically perform quantitative analysis auxiliary analysis on the original financial transaction flow file provided by the to-be-approved client according to the demand specified index, for example, from the aspects of amount size, frequency level, fluctuation range, and time range, and the like, so that the approval personnel can quickly analyze positive information elements and negative information elements from the transaction flow, and the workload of the approval personnel in checking and verifying thousands of flows one by one is greatly reduced.

[0036] Preferably, the intelligent approval method further comprises: S106: calling a base large language model, analyzing the transaction record detail data and the quantitative analysis index statistical results to obtain qualitative analysis results of the to-be-identified financial transaction flow file; Specifically, step S106: calling a base large language model, analyzing the transaction record detail data and the quantitative analysis index statistical results to obtain qualitative analysis results of the to-be-identified financial transaction flow file, comprising: S1061: submitting the transaction record detail data and the quantitative analysis index statistical results to the base large language model, performing qualitative analysis through the base large language model to obtain a comprehensive analysis text of the financial transaction of the to-be-approved client; S1062: According to the set qualitative analysis indicators (items of risk control checks) and the risk control types (including income stability, bad spending, transaction behavior anomaly, and implicit liabilities) corresponding to the qualitative analysis indicators, the vector query method is used to obtain content matching the approval problem from the knowledge base as a judgment rule and basis. The content matching the approval problem includes one of the following: regulations and systems, typical cases, and credit product introductions (for example, the regulations stipulate that stable income is monthly income volatility exceeding S%). The content matching the approval problem and the transaction record detail data are submitted to the base large language model, and the base large language model is used to qualitatively analyze the risk control types of the qualitative analysis indicators in the approval problem to obtain qualitative indicator analysis results, i.e., the qualitative analysis results of each type of qualitative analysis indicator, including income stability, bad spending, transaction behavior anomaly, and implicit liabilities. S1063: The comprehensive analysis text of the financial transaction of the to-be-approved client and the qualitative indicator analysis results are submitted to the base large language model, and the base large language model is used for analysis to obtain the overall qualitative analysis results of the to-be-identified financial transaction flow file.

[0037] Qualitative analysis is intelligent auxiliary approval, which is essentially used for qualitative analysis of transaction flow, and combines existing knowledge of the base large language model, local knowledge base data, database data, and external data to comprehensively give corresponding risk prompts and related approval suggestions, which can be used to assist approval personnel in issuing approval opinions and suggestions.

[0038] Preferably, the intelligent approval method can further include: S107: The metadata is stored in the database as process trace data for subsequent manual correction of the quantitative analysis results and the qualitative analysis results.

[0039] S108: In the process of converting the original financial transaction flow file into the to-be-identified financial transaction flow file using the original expression method, each transaction flow in the original financial transaction flow file is associated with the page number and saved in the database, which is used for directly locating the corresponding flow transaction in the original financial transaction flow file when the approval personnel reviews and traces the quantitative analysis results. Subsequently, from the application end UI (User Interface), the transaction flow can be easily located to the specific page of the original financial transaction flow file.

[0040] S109: Save the basic query SQL in the database, which is directly used by the approval personnel when reviewing and tracing the transaction record detail data, and is input into the base large language model to output the corresponding transaction record detail data. The SQL statement returned by the detail query is applied to the application, and the application end can query the database to obtain transaction detail record data and directly show it to the user.

[0041] S1010: Save the index query SQL in the database, which is directly used by the approval personnel when reviewing and tracing the quantitative analysis index statistical results, and is input into the base large language model to output the corresponding quantitative analysis index statistical results. The SQL statement returned by the detail query is applied to the application, and the application end can query the database to obtain specific transaction detail record data (details that meet the current index characteristics), and directly show it to user B. The specific detail record data obtained through the above search corresponds to the index details, and the page number, transaction counterpart, and transaction time can be used to locate the specific rows of the table in the original PDF file. User B can check each piece of original data in the details.

[0042] The basic query SQL and the corresponding transaction record detail data, the index query SQL statement and the corresponding quantitative analysis index statistical results are returned to the business application end and saved in the database for subsequent traceability processing. This avoids repeated operations and repeated decomposition problems when user B queries in the future, improves the operation experience, and reduces the consumption of computing power.

[0043] In summary, the complete flowchart of importing the original financial transaction flow file and intelligent approval is as shown in Figure 3

[0044] Import the original financial transaction flow file; read the original financial transaction flow file to form a to-be-recognized financial transaction flow file; extract metadata; store the metadata in the database; and disassemble the data of the to-be-recognized financial transaction flow file and write it to the database.

[0045] The approval personnel raises an approval question; the approval question is decomposed; the metadata of the database is read to assemble a basic query SQL; an index query SQL is assembled, which may include index 1 and index 2, etc.; the database is called to form quantitative analysis results, which are written to the database and displayed to the approval personnel; and the database is called to form qualitative analysis results, which are displayed to the approval personnel.

[0046] As shown in Figure 2 According to the embodiments of the present application, an intelligent approval system is provided, which comprises: ​The approval problem decomposition unit 21 is configured to, after extracting and storing the to-be-identified financial transaction flow file provided by the to-be-approved client with credit needs in the database, receive an approval question raised by an approval personnel when the approval personnel approves the credit needs of the to-be-approved client, decompose the approval question into a basic query question for querying the basic information of the transaction flow and an index analysis question for querying the quantitative analysis index by using the base large language model; The query SQL assembly unit 22 is configured to call the base large language model, assemble a basic query SQL for querying the basic information of the transaction flow according to the fields contained in the basic query question, and the basic query SQL includes specific query fields; and The base large language model is called to assemble an index query SQL for querying the quantitative analysis index of the transaction flow according to the quantitative analysis index contained in the index analysis question, and the basic query SQL is nested in the index query SQL to obtain a final query SQL. The quantitative analysis unit 23 is configured to call the database to execute the final query SQL, obtain transaction record detail data according to the basic query SQL, obtain quantitative analysis index statistical results according to the index query SQL, and form a quantitative analysis result based on the transaction record detail data and the quantitative analysis index statistical results.

[0047] In the credit approval process of a financial institution, analyzing the financial transaction flow file provided by a client is one of the important steps of risk identification. However, line-by-line interpretation and analysis of the financial transaction flow file is a time-consuming and labor-intensive work that is prone to errors and highly dependent on personal experience. In order to reduce the amount of repetitive work as much as possible and improve the approval quality, the system needs to automatically and intelligently complete the related work.

[0048] The to-be-identified financial transaction flow file provided by the to-be-approved client with credit needs is extracted and stored in the database. By extracting and storing the to-be-identified financial transaction flow file, the problem of unknown table patterns not being able to be identified is solved.

[0049] The approval questions of the approval personnel are decomposed into different basic query questions and index analysis questions, and the basic query SQL and the index query SQL corresponding to each question are formed according to the respective fields. Therefore, different indexes can be queried according to different questions, and the diversity of index query is realized. The quantitative analysis result is obtained by querying in the database through the base large language model. The workload of the approval personnel is reduced, and the efficiency and accuracy of manual approval are improved.

[0050] Preferably, the intelligent approval system further comprises: The data extraction unit is configured to extract and store a to-be-identified financial transaction flow file provided by a to-be-approved client with a credit demand in a database. The data extraction unit comprises: The data format conversion subunit is configured to receive an original financial transaction flow file provided by a to-be-approved client with a credit demand, and convert the original financial transaction flow file into a to-be-identified financial transaction flow file in an original expression manner. The metadata extraction subunit is configured to identify the to-be-identified financial transaction flow file based on a base large language model, obtain metadata of the original financial transaction flow file, represent the metadata in the original expression manner, and generate a table-building SQL statement for storing data in the to-be-identified financial transaction flow file based on the metadata. The data extraction and storage subunit is configured to disassemble the data in the to-be-identified financial transaction flow file based on the metadata, and the table-building SQL statement is configured to write the disassembled data into a table in the database in the original expression manner.

[0051] The original financial transaction flow file is converted into a to-be-identified financial transaction flow file in an original expression manner, and the metadata is represented in the original expression manner in the original financial transaction flow file, which facilitates the to-be-approved client to quickly identify and correspond to the to-be-approved client in the later verification, and reduces the reaction time of the to-be-approved client compared with the prior art in which the fields of the original financial transaction flow files of different financial institutions are all converted into the same field.

[0052] The original financial transaction flow files of different financial institutions can be automatically identified and imported into the database, without the intervention of developers (without manual table building and data import), and without the traditional IT development process (such as manual table building, development of an extension interface, testing, and online data import), so that the original financial transaction flow file can be converted into data in the database table only by uploading the original financial transaction flow file to the intelligent approval system. Especially when the amount of work for connecting the original financial transaction flow file of unknown financial institutions is very small, almost zero, the work efficiency is greatly improved.

[0053] Preferably, the data format conversion subunit is specifically configured to: The data format conversion subunit is configured to receive an original financial transaction flow file provided by a to-be-approved client with a credit demand, and convert the original financial transaction flow file into a to-be-identified financial transaction flow file in an original expression manner. Extract the data in the original financial transaction file after denoising, save the data in the original financial transaction file to the TXT format file in turn, identify the page symbol and / or page number in the original financial transaction file, set the corresponding page number and / or page symbol in the TXT format file according to the identified page symbol and / or page number, and supplement the broken words and / or broken lines of the original financial transaction file to form complete words and / or complete line contents, and obtain the to-be-recognized financial transaction file.

[0054] The user A (such as an operation clerk) responsible for importing the original financial transaction file uploads the PDF format original financial transaction file provided by the to-be-approved customer through the interface of the intelligent approval system. The intelligent approval system reads the PDF format original financial transaction file through the PDF reading special tool developed by itself, including: PDF format decryption, watermark removal, content reading, page division and page number recognition, and broken word / broken line completion processing, and extraction to the TXT format file.

[0055] In the TXT format file, the original expression method in the original financial transaction file is adopted, which will reduce the reaction time of the to-be-approved customer compared with the prior art of converting the fields of the original financial transaction files of various financial institutions into the same field, and the to-be-approved customer is more convenient.

[0056] Preferably, the metadata extraction subunit is specifically used for: The first page content of the to-be-recognized financial transaction file is recognized based on the base large language model, and metadata is extracted and represented in the original expression method. The metadata at least includes: header field (each field contained in the header is individually as a column), original header and page symbol (as a column), amount format category (each field is individually as a column), specific query field (each field is individually as a column), to-be-approved customer name and financial institution issuing the original financial transaction file, wherein the specific query field is derived from the header field, and the header field one-to-one corresponds to the fields included in the original header of the to-be-recognized financial transaction file.

[0057] The amount format category refers to how the transaction amount field in the original financial transaction file is composed, which is used for accurate judgment of the income and expenditure amount subsequently. The amount format category includes three categories: Class C: only the "transaction amount" field, which distinguishes income and expenditure by positive and negative signs; Class B: represented by "transaction amount" and "type of income and expenditure" two fields, the amount is always positive, and the type of income and expenditure distinguishes income and expenditure; Class A: respectively with "expenditure amount", "income amount" two fields to represent, its amount is positive.

[0058] Table header / page separator, refers to the original table header row or page separator in the original financial transaction log file, used to inform the subsequent data storage program how to distinguish the data in the file and the field name, to ensure the generality of the log data storage processing program; the page separator is used to judge what is the page separator, and then the specified symbol is used as the page separator to realize the paging processing; wherein, the original table header is the original table header in the transaction log file, containing the information field, all the contents in the "original table header" are treated as a field in the metadata.

[0059] Specific query field name refers to the set of DB (database) field names required for subsequent feature analysis and query of transaction log, such as transaction purpose (salary, pawn, etc.), which can also be transaction counterparty, transaction summary, addendum and remarks, etc.

[0060] Generate a table-building SQL statement to store the data in the to-be-identified financial transaction log file, and carry all the fields included in the metadata in the table-building SQL statement, i.e. the fields included in the table-building SQL statement correspond one-to-one with the metadata.

[0061] According to the metadata obtained above, it is first judged whether it is a new style of financial transaction log file, i.e. the three conditions of financial institution name, original table header and table-building SQL statement are judged simultaneously, when the metadata involved in the three conditions are completely the same as the metadata in the existing table, when the metadata involved in the three conditions do not match the metadata in the existing table, it is indicated that it is a new style, if it is a new style, a database table is created according to all the fields carried by the table-building SQL statement, and then the data in the to-be-identified financial transaction log file is imported in the original expression, otherwise the data in the to-be-identified financial transaction log file is directly imported into the existing table in the database.

[0062] Preferably, the approval problem decomposition unit 21 comprises: The first decomposition subunit is configured to call a base large language model, perform word segmentation on the approval problem by the base large language model, extract a specific keyword representing a specific query field, and extract a table header field related to the specific keyword as a field included in a basic query problem, the field included in the basic query problem at least including: the specific keyword, a transaction time range and an amount range. The second decomposition subunit is configured to extract a keyword of a quantitative analysis index to form an index analysis problem for querying the quantitative analysis index, wherein the keyword of the quantitative analysis index at least includes the following fields: a transaction counterparty, a transaction month, a transaction total amount, a cumulative number of transaction pens, a transaction average amount and an amount standard deviation.

[0063] The user B responsible for analysis (such as an approval person) can ask questions to the intelligent approval system (agent) (or generated in advance by the intelligent approval system), and the base large language model such as the LLM large model analyzes and decomposes the approval person's questions to obtain basic query questions and index analysis questions, especially extracts the table header fields related to the specific keywords. Decomposing the approval question can make up for the large model illusion and unstable answers caused by directly inputting the approval question into the base large language model.

[0064] The basic query question and the index analysis question perform quantitative analysis on the original financial transaction flow file as needed. There are four types of quantitative analysis dimensions: transaction amount summary, transaction frequency, amount fluctuation range, and transaction time period. Based on these dimensions and in conjunction with specific keywords (specific to detailed specific query fields), the financial transaction flow file is queried and sorted, which can quickly locate the transaction flow record that meets the specified specific query field from thousands of financial transaction flow records, such as high transaction amount or excessive transaction frequency or stable transaction period or transaction record in a specific time period.

[0065] The query SQL assembly unit 22 is specifically used for: The base large language model is called to assemble the corresponding query condition according to the specific query field corresponding to the specific keyword in the basic query question through the base large language model, and the basic query SQL is obtained.

[0066] The base large language model is called to assemble the index query SQL for querying the quantitative analysis index of the transaction flow according to the quantitative analysis index contained in the index analysis question: the index query SQL is composed of different indexes such as index 1, index 2, etc., the basic query SQL is nested in the index query SQL, and the final query SQL is obtained; the integration of the basic query SQL and the statistical query SQL is realized based on the SQL nesting mode in which the second SQL includes the first SQL is the SQL itself.

[0067] Preferably, S103: calling a database to execute a final query SQL, obtaining transaction record detail data for the basic query SQL, and obtaining quantitative analysis index statistical results for the index query SQL, the quantitative analysis index statistical results including four dimensions of data: transaction amount summary, transaction frequency, amount fluctuation range, and transaction time period, and being returned in a JSON set format, wherein KEY is a field name, and VALUE is a specific value. Intelligent quantitative analysis can automatically perform quantitative analysis on the original financial transaction flow file provided by the to-be-approved client according to the demand specified indexes, such as analyzing from the aspects of amount size, frequency level, fluctuation range, and time range, so that the approval personnel can quickly analyze positive information elements and negative information elements from the transaction flow, and the workload of the approval personnel in checking and analyzing thousands of flow files one by one is greatly reduced.

[0068] Preferably, the intelligent approval system further comprises: a quantitative analysis unit configured to call a base large language model, and analyze the transaction record detail data and the quantitative analysis index statistical results to obtain a qualitative analysis result of the to-be-identified financial transaction flow file. The quantitative analysis unit is specifically configured to: submit the transaction record detail data and the quantitative analysis index statistical results to the base large language model for qualitative analysis, and obtain a comprehensive analysis text of the financial transaction of the to-be-approved client. According to the set qualitative analysis indexes (items of risk control check) and the risk control types (for example, income stability, undesirable expenditure, etc.) corresponding to the qualitative analysis indexes, the vector query method is used to obtain content matched with the approval question from a knowledge base, the content matched with the approval question including one of the following: regulations and systems, typical cases, and credit product introductions, the content matched with the approval question and the transaction record detail data are submitted to the base large language model, the base large model is used to qualitatively analyze the risk control types of the qualitative analysis indexes in the approval question, and a qualitative index analysis result is obtained, that is, a qualitative analysis result of each type of qualitative analysis index is obtained, such as income stability or undesirable expenditure, etc. The comprehensive analysis text of the financial transaction of the to-be-approved client and the qualitative index analysis result are submitted to the base large language model for analysis, and a qualitative analysis result of the to-be-identified financial transaction flow file as a whole is obtained.

[0069] Qualitative analysis is intelligent auxiliary approval, and is essentially used for qualitative analysis on the transaction flow. By combining the existing knowledge of the base large model, the local knowledge base data, the database data, and the external data, corresponding risk prompts and related approval suggestions are comprehensively given, which can be used to assist the approval personnel in issuing approval opinions and suggestions.

[0070] Preferably, the intelligent approval method further comprises a traceability unit for storing metadata as process trace data into a database for subsequent manual correction of quantitative analysis results and qualitative analysis results.

[0071] In the process of converting the original financial transaction log file into a to-be-identified financial transaction log file using the original expression method, each transaction log is associated with the page number of the original financial transaction log file and saved in the database, for directly locating the corresponding log transaction in the original financial transaction log file when the approval personnel reviews and traces the quantitative analysis results; so that the transaction log can be easily located to the specific page of the original financial transaction log file on the application end UI subsequently.

[0072] The basic query SQL is saved in the database, for directly using the basic query SQL and inputting into the base large language model when the approval personnel reviews and traces the transaction record detail data, and outputting the corresponding transaction record detail data; the SQL statement returned by the detail query is sent to the application, and the application end can obtain the transaction detail record data by querying the database and directly show it to the user.

[0073] The index query SQL is saved in the database, for directly using the index query SQL and inputting into the base large language model when the approval personnel reviews and traces the quantitative analysis index statistical result, and outputting the corresponding quantitative analysis index statistical result. The SQL statement returned by the detail query is sent to the application, and the application end can obtain the specific transaction detail record data (details conforming to the current index characteristics) by querying the database and directly show it to the user B. The specific detail record data obtained by the above-mentioned searching corresponds to the index details, the page number, the transaction counterparty and the transaction time, so that the specific rows of the table in the original PDF file can be located, and the user B can check each original data in the details.

[0074] The basic query SQL and the corresponding transaction record detail data, and the index query SQL statement and the corresponding quantitative analysis index statistical result are returned to the business application end and saved in the database, for subsequent traceability processing. The repeated operation and the process of generating the SQL statement are avoided, the operation experience is improved, and the computing power consumption is reduced.

[0075] In combination with the embodiments of the present application, a computer readable storage medium is provided, which stores one or more programs, and the one or more programs, when executed by a computer device, cause the computer device to execute any one of the intelligent approval methods described above.

[0076] In combination with the embodiments of the present application, a computer device is provided, comprising: a processor; and a memory arranged to store computer executable instructions which, when executed, cause the processor to perform the intelligent approval method of any of the foregoing.

[0077] The beneficial technical effects achieved by the embodiments of the present application are as follows: 1. The large model intelligently identifies the original financial transaction flow file of any format, intelligently builds a table and imports data, solving the technical problem of data extraction.

[0078] 2. The financial transaction flow file of any financial institution can be intelligently identified, and quantitative analysis and qualitative analysis can be automatically performed. On the technical level, no code modification is required, saving the adaptation and development time of the flow file of each financial institution (from 3-7 person-days to 0), and on the business level, the work efficiency and quality of the approval personnel are improved (from 2-3 hours to 10-20 minutes).

[0079] 3. Based on the knowledge base, the continuous collection, storage and analysis of the relevant business experience of the financial suggestion flow file analysis are supported, and the accumulation and inheritance of business knowledge are realized.

[0080] 4. According to the approval problems of the approval personnel, the problems can be disassembled to form SQL statements, making up for the shortcomings of large model illusion and unstable answers. The SQL statements can be left with traces in a manner, which can greatly reduce the workload of review.

[0081] It should be understood that the specific order or hierarchy of steps in the disclosed processes is an example of exemplary methods. Based on design preferences, it is understood that the specific order or hierarchy of steps in the processes can be rearranged without departing from the scope of the present disclosure. The accompanying method claims present elements of the various steps in exemplary order, and are not intended to be limited to the specific order or hierarchy presented.

[0082] In the above detailed description, various features are combined together in a single embodiment in order to simplify the present disclosure. Such disclosed methods should not be interpreted as reflecting an intention that the claimed subject matter requires more features than are explicitly stated in each claim. On the contrary, as reflected by the attached claims, the present application is directed to embodiments substantially less than all of the features of the disclosed embodiments. Accordingly, the attached claims are hereby expressly incorporated into this detailed description, with each claim acting as a separate embodiment of the present application.

[0083] The foregoing description of the exemplary embodiments of this application has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the application to the precise form disclosed. Many modifications and variations are possible in light of the above teaching. It is intended that the scope of the application be limited not with this detailed description, but rather by the claims appended hereto.

[0084] The above description includes exemplary embodiments. Of course, not all possible combinations of components or method steps are described above, but one of ordinary skill in the art will recognize that further combinations are possible. Persons of ordinary skill in the art will also recognize that the various embodiments described above can be further modified than described, and thus all modifications and further combinations are believed to be encompassed in the scope of the claims.

[0085] Those of skill would further appreciate that the various illustrative logical blocks, modules, and steps described in connection with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans can implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present embodiments.

[0086] The various illustrative logical blocks, modules, and steps described in connection with the embodiments disclosed herein can be implemented or performed with a general purpose processor, a digital signal processor, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor can be a microprocessor, but in the alternative, the general purpose processor can be any conventional processor, controller, microcontroller, or state machine. A processor can also be implemented as a combination of computing devices, e.g., a combination of a digital signal processor and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a digital signal processor core, or any other such configuration.

[0087] The steps of a method or algorithm described in connection with the embodiments disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium can be integral to the processor. The processor and the storage medium can reside in an ASIC. The ASIC can reside in a user terminal. In the alternative, the processor and the storage medium can reside as discrete components in a user terminal.

[0088] In one or more exemplary designs, the functions described can be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the functions can be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media include both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. Storage media can be any available media that can be accessed by a general purpose or special purpose computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code means in the form of instructions or data structures and that can be accessed by a general-purpose or special-purpose computer, or a general-purpose or special-purpose processor. Also, any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or data

[0089] The above detailed description describes the purpose, technical solutions and advantages of the present application. It should be understood that the above detailed description is only a specific embodiment of the present application and is not intended to limit the scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the scope of the present application.

Claims

1. An intelligent approval method, characterized in that, The method comprises the steps of: After extracting and saving the to-be-identified financial transaction flow file provided by the to-be-approved client with credit demand in the database, receiving an approval question raised by an approval personnel when approving the credit demand of the to-be-approved client, decomposing the approval question into a basic query question for querying the basic information of the transaction flow and an index analysis question for querying the quantitative analysis index by using the base large language model; Calling the base large language model, assembling the basic query SQL for querying the basic information of the transaction flow according to the fields contained in the basic query question, wherein the basic query SQL includes specific query fields; Calling the base large language model, assembling the index query SQL for querying the quantitative analysis index of the transaction flow according to the quantitative analysis index contained in the index analysis question, nesting the basic query SQL into the index query SQL to obtain the final query SQL; Calling the database to execute the final query SQL, obtaining transaction record detail data for the basic query SQL and quantitative analysis index statistical results for the index query SQL, and forming a quantitative analysis result based on the transaction record detail data and the quantitative analysis index statistical results.

2. The intelligent approval method of claim 1, wherein, The step of extracting and saving the to-be-identified financial transaction flow file provided by the to-be-approved client with credit demand in the database comprises: Receiving the original financial transaction flow file provided by the to-be-approved client with credit demand, and converting the original financial transaction flow file into a to-be-identified financial transaction flow file in the original expression manner; Based on the base large language model, identifying the to-be-identified financial transaction flow file to obtain metadata of the original financial transaction flow file, representing the metadata in the original expression manner, and generating a build table SQL statement for storing data in the to-be-identified financial transaction flow file based on the metadata; Based on the metadata, the data in the to-be-identified financial transaction flow file is disassembled, and the build table SQL statement is used to write the disassembled data into the table in the original expression manner.

3. The intelligent approval method of claim 2, wherein, The step of receiving the original financial transaction flow file provided by the to-be-approved client with credit demand and converting the original financial transaction flow file into a to-be-identified financial transaction flow file in the original expression manner comprises: Receiving the original financial transaction flow file provided by the to-be-approved client with credit demand in PDF format, decrypting the data of the original financial transaction flow file, and performing denoising processing on the decrypted original financial transaction flow file; The data in the original financial transaction file after denoising processing is extracted, and the data in the original financial transaction file is saved in a TXT format file in sequence, while the page symbol and / or page number in the original financial transaction file are identified, the corresponding page number and / or page symbol are set in the TXT format file according to the identified page symbol and / or page number, and the broken words and / or broken lines of the original financial transaction file are supplemented to form complete words and / or complete line contents, thereby obtaining a to-be-identified financial transaction file.

4. The intelligent approval method of claim 2, wherein, The base large language model is used to identify the to-be-identified financial transaction file, obtain metadata of the original financial transaction file, and represent the metadata in the original expression manner, and a table-building SQL statement for storing data in the to-be-identified financial transaction file is generated based on the metadata, including: The first page content of the to-be-identified financial transaction file is identified based on the base large language model, and metadata is extracted, and the metadata is represented in the original expression manner, and the metadata at least includes: table header field, original table header and page symbol, amount format category, specific query field, to-be-approved client name, and financial institution issuing the original financial transaction file, wherein the specific query field is derived from the table header field, and the table header field one-to-one corresponds to the fields included in the original table header of the to-be-identified financial transaction file; A table-building SQL statement for storing data in the to-be-identified financial transaction file is generated, and all fields included in the metadata are carried in the table-building SQL statement.

5. The intelligent approval method of claim 4, wherein, The base large language model is used to decompose the approval problem into a basic query problem for querying basic information of the transaction file and an index analysis problem for querying quantitative analysis indexes, including: The base large language model is called to perform word segmentation on the approval problem through the base large language model, extract a specific keyword representing a specific query field, and extract a table header field related to the specific keyword as a field included in the basic query problem, and the field included in the basic query problem at least includes: specific keyword, transaction time range, and amount range; The keyword of the quantitative analysis index is extracted to form an index analysis problem for querying the quantitative analysis index, wherein the keyword of the quantitative analysis index at least includes the following fields: amount, cumulative number of transactions, average transaction amount, and amount standard deviation.

6. The intelligent approval method of claim 5, wherein, The base large language model is called to assemble a basic query SQL for querying basic information of the transaction file according to the fields included in the basic query problem, and the basic query SQL includes a specific query field, including: The base large language model is called to assemble a basic query SQL for querying basic information of the transaction file according to the fields included in the basic query problem, and the basic query SQL includes a specific query field, including:

7. The intelligent approval method of claim 1, wherein, Further comprising: The base large language model is called to analyze the transaction record detail data and the quantitative analysis index statistical result, and a qualitative analysis result of the to-be-identified financial transaction flow file is obtained. The base large language model is called to analyze the transaction record detail data and the quantitative analysis index statistical result, and a qualitative analysis result of the to-be-identified financial transaction flow file is obtained. The transaction record detail data and the quantitative analysis index statistical result are submitted to the base large language model, and a comprehensive analysis text of the financial transaction of the to-be-approved client is obtained through qualitative analysis of the base large language model. According to the set qualitative analysis index and the risk control type corresponding to the qualitative analysis index, the content matched with the approval problem is obtained from the knowledge base based on the vector query method, and the content matched with the approval problem includes one of the following: rules and regulations, typical cases, and credit product introduction. The content matched with the approval problem and the transaction record detail data are submitted to the base large language model, and the risk control type of the qualitative analysis index in the approval problem is qualitatively analyzed through the base large language model to obtain a qualitative index analysis result. The comprehensive analysis text of the financial transaction of the to-be-approved client and the qualitative index analysis result are submitted to the base large language model, and a qualitative analysis result of the whole to-be-identified financial transaction flow file is obtained through analysis of the base large language model.

8. The intelligent approval method of claim 2, wherein, Further comprising: In the process of converting the original financial transaction flow file into a to-be-identified financial transaction flow file using the original expression method, each transaction flow in the original financial transaction flow file is associated with the page number and saved in the database, which is used for directly locating the corresponding flow transaction in the original financial transaction flow file when the approval personnel reviews and traces the quantitative analysis result; The basic query SQL is saved in the database, which is used for directly using the basic query SQL and inputting into the base large language model when the approval personnel reviews and traces the transaction record detail data, and outputting the corresponding transaction record detail data; The index query SQL is saved in the database, which is used for directly using the index query SQL and inputting into the base large language model when the approval personnel reviews and traces the quantitative analysis index statistical result, and outputting the corresponding quantitative analysis index statistical result.

9. An intelligent approval system characterized by, Comprising: An approval problem decomposition unit is configured to, after extracting and saving the to-be-identified financial transaction flow file provided by the to-be-approved client with credit demand in the database, receive an approval problem raised by the approval personnel when approving the credit demand of the to-be-approved client, and decompose the approval problem into a basic query problem for querying the basic information of the transaction flow and an index analysis problem for querying the quantitative analysis index through the base large language model. The query SQL assembly unit is configured to invoke the base large language model, assemble a basic query SQL for querying basic information of the transaction flow according to fields contained in the basic query question, and the basic query SQL includes specific query fields. The base large language model is invoked to assemble an index query SQL for querying quantitative analysis indexes of the transaction flow according to quantitative analysis indexes contained in the index analysis question, the basic query SQL is nested in the index query SQL, and a final query SQL is obtained. The quantitative analysis unit is configured to invoke a database, execute the final query SQL, obtain transaction record detail data according to the basic query SQL, obtain quantitative analysis index statistical results according to the index query SQL, and form quantitative analysis results based on the transaction record detail data and the quantitative analysis index statistical results.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores one or more programs, which when executed by a computer device, cause the computer device to perform the intelligent approval method of any one of claims 1-8.

11. A computer device, comprising: Comprise: A processor; And a memory arranged to store computer executable instructions which when executed cause the processor to perform the intelligent approval method of any one of claims 1-8.

Images