Business question and answer method and device, equipment, storage medium and program product

Abstract

The invention discloses a business question-answering method and device, equipment, a storage medium and a program product, and relates to the technical field of natural language processing, and the business question-answering method comprises the following steps: receiving a business question text; extracting keyword information and slot position information in the business problem text; determining context information corresponding to the business problem text according to the keyword information and the slot position information; and according to the context information, performing service analysis on the service question text through a question and answer large model to obtain a service reply result. Context business question answering based on a large model is realized, the problem that business question answering in the prior art depends on manual business information transmission and is low in accuracy is solved, and the accuracy of business question answering is improved.

Description

Technical Field

[0001] The present application relates to the field of natural language processing technology, and in particular to a business question-answering method, apparatus, device, storage medium, and program product. Background Art

[0002] With the rapid development of banking services, the need for intelligent decision support for deposit and loan pricing analysis is becoming increasingly urgent. Currently, the banking industry generally uses deposit and loan pricing systems based on fixed-format reports, which use pre-set templates to complete basic data statistics and graphical presentations, and are open to certain business managers.

[0003] However, the existing technical business knowledge and system operation procedures are relatively complex. When users encounter problems during use, they often need to ask questions to the head office business and IT staff on duty. They rely on business managers to manually update and convey information, which cannot adapt to the differentiated and accurate needs of multi-level users in the head office and branches.

[0004] The above content is only used to assist in understanding the technical solution of this application and does not constitute an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a business question-and-answer method, apparatus, device, storage medium, and program product, aiming to solve the technical problem that the existing business question-and-answer technology relies on manual transmission of business information with low accuracy.

[0006] To achieve the above objectives, this application proposes a business question-answering method, which includes:

[0007] Receive business question text;

[0008] Extracting keyword information and slot information from the business question text;

[0009] Determining context information corresponding to the business question text based on the keyword information and slot information;

[0010] Based on the context information, the business question text is analyzed using a large question-answer model to obtain a business answer result.

[0011] In one embodiment, before the step of determining the context information corresponding to the business question text based on the keyword information and slot information, the method further includes:

[0012] Using the edit distance algorithm, query the database for similar field information of the slot information;

[0013] Based on the similar field information, standard slot information is generated to determine context information corresponding to the business question text according to the keyword information and the standard slot information.

[0014] In one embodiment, the step of determining the context information corresponding to the business question text based on the keyword information and slot information includes:

[0015] Identify the table information corresponding to the business problem text;

[0016] Generate context query conditions based on the table information and keyword information;

[0017] According to the context query condition, the context information corresponding to the business question text in the preset corpus is queried.

[0018] In one embodiment, the table information includes a table number and a table name, the keyword information includes keywords, key phrases, key attributes and named entities, the corpus stores table structure information and business context, the table structure information includes table description and column description, and the business context includes business keywords, default context logic and unmentioned logic.

[0019] In one embodiment, the step of performing business analysis on the business question text using a question-answering model based on the context information to obtain a business answer result includes:

[0020] Generating business analysis prompt words for the business problem text through a thought chain prompt process according to the context information;

[0021] Based on the business analysis prompt words, reasoning analysis is performed through the question-answering model to obtain business response results.

[0022] In one embodiment, the step of generating business analysis prompt words for the business problem text through a thought chain prompt process according to the context information includes:

[0023] Decomposing the business question text into a sub-question chain with logical dependencies according to the context information, wherein the sub-question chain includes a plurality of interrelated sub-questions;

[0024] Semantically binding the entities in the sub-problem chain with the table structure of the database to generate an enhanced problem statement with table association tags;

[0025] Based on the enhanced problem statement of the table association mark, business analysis prompt words of the business problem text are generated.

[0026] In addition, to achieve the above-mentioned purpose, the present application also proposes a business question-answering device, which includes:

[0027] Receiving module, used to receive business question text;

[0028] An extraction module, used to extract keyword information and slot information from the business question text;

[0029] A determination module, configured to determine context information corresponding to the business question text based on the keyword information and slot information;

[0030] The analysis module is used to perform business analysis on the business question text based on the context information through the question-answering model to obtain a business response result.

[0031] In addition, to achieve the above-mentioned purpose, the present application also proposes a business question-and-answer device, which includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the computer program is configured to implement the steps of the business question-and-answer method described above.

[0032] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium. A computer program is stored on the storage medium, and when the computer program is executed by the processor, the steps of the business question-and-answer method described above are implemented.

[0033] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the business question and answer method described above.

[0034] This application provides a business question-and-answer method. First, it receives business question texts in natural language form, so that different users can raise diverse demands in an intuitive way. Second, by extracting keyword information and slot information, it accurately captures the core elements and dynamic parameters in the question, avoiding the semantic understanding deviation problem caused by a single dimension. Then, based on the keyword and slot information, it generates context information that is highly adapted to the question, breaking the constraints of static analysis logic and realizing flexible integration of multi-source data. Finally, through the question-and-answer big model, it conducts deep reasoning and business logic analysis on the context information, and automatically generates business response results containing data analysis conclusions and decision recommendations. The deep integration of natural language interaction, dynamic semantic understanding and intelligent analysis capabilities can respond to users' complex and multi-dimensional pricing analysis needs in real time, significantly improve the timeliness and accuracy of business decisions, and greatly reduce the cost of manual intervention. It realizes contextual business question-and-answer based on a big model, solves the problem of low accuracy of business question-and-answer in existing technologies that relies on manual transmission of business information, and improves the accuracy of business question-and-answer. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0036] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0037] Figure 1 A flowchart of the first embodiment of the business question-and-answer method of this application is provided;

[0038] Figure 2 A flowchart of the second embodiment of the business question-and-answer method of this application is provided;

[0039] Figure 3 A flowchart of the third embodiment of the business question-and-answer method of this application is provided;

[0040] Figure 4 A flowchart of the table service of the third embodiment of the business question-and-answer method of this application;

[0041] Figure 5 A flowchart of the business context service of the third embodiment of the business question-answering method of this application;

[0042] Figure 6 A flowchart illustrating the slot information filling process of the third embodiment of the business question-and-answer method of this application;

[0043] Figure 7 This is a schematic diagram of the module structure of the business question-and-answer device according to an embodiment of the present application;

[0044] Figure 8 This is a schematic diagram of the device structure of the hardware operating environment involved in the business question-and-answer method in the embodiment of this application.

[0045] The purpose, features and advantages of this application will be further explained with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0046] It should be understood that the specific embodiments described herein are merely used to explain the technical solutions of the present application and are not intended to limit the present application.

[0047] In order to better understand the technical solution of this application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0048] The main solution of the embodiment of this application is: receiving a business question text; extracting keyword information and slot information from the business question text; determining the context information corresponding to the business question text based on the keyword information and slot information; and performing business analysis on the business question text through a question-answering model based on the context information to obtain a business response result.

[0049] As existing technologies evolve with the rapid development of banking services, the need for intelligent decision support for deposit and loan pricing analysis is becoming increasingly urgent. Currently, the banking industry generally uses deposit and loan pricing systems based on fixed-format reports. These systems use pre-set templates to perform basic data statistics and graphical presentations, and are open to selected business managers. These systems typically rely on manual data compilation and report generation, supplemented by static business logic rules to support pricing decisions.

[0050] However, users at all levels of the head office and branches focus on different issues. The system's flaw lies in its lack of flexible, efficient, multi-dimensional, and intelligent data analysis capabilities. Deposit and loan pricing policies are frequently updated, and various users, including account managers and branch finance, often need to understand the latest pricing policies. However, the lack of unified management and access channels easily leads to information asymmetry. Users cannot quickly obtain dynamic analysis results through natural language interaction, and the system's single data dimension and rigid analysis logic make it difficult to adapt to the differentiated needs of multi-level users in the head office and branches. For example, when business personnel need to optimize pricing based on the latest policies or complex business scenarios, the system cannot automatically parse multi-source data and generate customized analysis reports, resulting in inefficient decision-making and heavy reliance on manual experience and repetitive work.

[0051] This application leverages the data analysis and reasoning capabilities of a large model to provide a complete end-to-end solution from natural language questions to data analysis results. This addresses the intelligentization of pricing business analysis and post-supervision analysis. Automatically parsing policy documents addresses the issues of delayed and inaccurate policy information updates. Regularly updating business knowledge documents, operating manuals, and Q&A pairs addresses the significant labor consumption associated with on-call Q&A.

[0052] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, mobile phone, etc., or an electronic device or business Q&A device capable of performing the above functions. The following uses a business Q&A device as an example to illustrate this embodiment and the following embodiments.

[0053] Based on this, the embodiment of the present application provides a business question-answering method, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the business question-and-answer method of this application.

[0054] In this embodiment, the business question-answering method includes steps S10 to S40:

[0055] Step S10, receiving a business question text;

[0056] It should be noted that business question text refers to business-related questions or demands raised by users in natural language, such as "What business types are currently available in deposit pricing sheets?"

[0057] It can be understood that business question text refers to business query requests entered by users in natural language. It provides the system with raw input and triggers subsequent analysis processes. By receiving natural language questions from users, the system converts the unstructured business questions raised by users into text data that can be processed by subsequent processes, providing a foundation for keyword extraction and semantic analysis.

[0058] Specifically, the system can receive user business question texts in a variety of ways. For example, within a bank's internal business system, a text input box can be set up in a specific functional module or question-and-answer interface. After the user enters the business question text in this input box, the system receives the text through an interface. Alternatively, the system can integrate with the user's instant messaging tool to receive business question text messages sent by the user. In addition, the system can also read user-submitted business question texts from log files or records of other business systems. These different reception methods can meet the needs of different user scenarios, such as business personnel asking questions directly while operating the business system, or asking questions through communication tools in remote or mobile office scenarios.

[0059] Step S20, extracting keyword information and slot information from the business question text;

[0060] It should be noted that extraction refers to the process of identifying and obtaining specific words or phrases from the received business question text. Keyword information refers to words or phrases that can reflect the core content and focus of the business question, such as "deposit pricing sheet" and "business varieties", which play a key role in determining the subject, logic and scope of the business question. Slot information refers to entity information in the business question text that has specific business meaning or is used to limit query conditions. It usually includes institution name, customer name, time range, etc., such as "Shenzhen Branch" and "December 2024". Slot information is used to accurately locate the specific business objects and scope involved in the business question.

[0061] It is understandable that in order to conduct preliminary semantic analysis and identify key elements of business question text, so as to be able to more accurately determine the context information of the business question and conduct business analysis later, by extracting keyword information, the system can understand the subject area and key content of the business question; by extracting slot information, the system can clarify the specific business entities and conditions involved in the business question, thereby providing key basis for subsequent accurate business answers. It can help the system filter out the most valuable parts from a large amount of text information, avoid the mismatch problem caused by semantic ambiguity in traditional systems, and improve the efficiency and accuracy of business question answering.

[0062] Optionally, keyword and slot information extraction can be achieved through various natural language processing technologies. For example, rule-based matching methods can be used to search for matching keywords and slot information within the business question text using a predefined business vocabulary and slot patterns. Alternatively, machine learning methods can be employed, such as using named entity recognition models and keyword extraction models trained on large amounts of business text data, to automatically identify and extract keyword and slot information from the text. Furthermore, semantic analysis techniques can be combined to identify potential keywords and slot information based on the context and semantic relationships within the business question text.

[0063] For example, for the business question text "How many pricing orders with negative returns in December 2024 at the Shenzhen branch?", the system uses keyword extraction technology to identify "negative returns" and "number of pricing orders" as keyword information, and uses slot information extraction technology to identify "Shenzhen branch" and "December 2024" as slot information. Among them, "negative returns" and "number of pricing orders" clarify that the business question focuses on the revenue situation and quantity statistics of pricing orders, while "Shenzhen branch" and "December 2024" limit the institutional scope and time range of the business question.

[0064] Step S30, determining context information corresponding to the business question text based on the keyword information and slot information;

[0065] It's important to note that contextual information refers to the associated collection of database table structures, business rules, and knowledge documents related to the business problem. This information provides a more comprehensive semantic context and business scenario for the business problem, helping the system more accurately understand the connotation and extension of the business problem. For example, if the keyword contains "deposit pricing sheet," then associated Table 1 (the detailed table of corporate deposit pricing sheets) and its related field descriptions can be retrieved through vector database retrieval and dynamic matching with the rule engine.

[0066] It is understandable that in order to place the business question text in a specific business context and data environment for understanding and analysis, thereby providing accurate background basis for subsequent business analysis through the question-answering big model, by combining keyword information and slot information, the system can associate relevant contextual information from multiple dimensions (such as business domain, business entity, time range, etc.), making the semantics of the business question more clear and complete, avoiding ambiguity and misunderstanding caused by missing context, and improving the accuracy and pertinence of business analysis.

[0067] Optionally, determining context information can be achieved in the following ways: first, the system can query the preset business terminology library and business rule library based on the keyword information to obtain business concepts, business rules, business processes and other information related to the keywords as part of the context; second, based on the entity information such as the organization name and customer name in the slot information, query the business data warehouse or database for relevant business data records and business status information. These data records and status information constitute the business data context of the business problem; third, combined with the time range, business scenario and other conditions in the slot information, extract relevant trend information, related business events, etc. from historical business data and business logs as the time and scenario background of the context information.

[0068] Step S40: Based on the context information, a business analysis is performed on the business question text using a question-answering model to obtain a business answer result.

[0069] It should be noted that a Q&A big model refers to a large language model trained with extensive business and text data. It possesses the ability to semantically understand and analyze natural language text and generate responses based on the input question and context. Business analysis refers to the process of using the Q&A big model to conduct in-depth semantic analysis, logical reasoning, and the application of business logic to business question text to obtain answers or solutions. Business response results refer to the answers or responses to business questions generated by the Q&A big model based on the business question text and context. These can be textual or graphical explanations, data statistics, business action suggestions, etc.

[0070] It's understandable that by leveraging the large model's powerful natural language processing and business understanding capabilities, it can conduct in-depth analysis of business question text based on contextual information, generating accurate responses that align with business logic and user needs. The large Q&A model comprehensively considers factors such as the business background, data environment, and business rules within contextual information, conducting multi-dimensional analysis and reasoning on business questions, uncovering the underlying meaning and business connections behind them. It provides comprehensive, accurate, and useful responses to users, enhancing the intelligence level of business Q&A and the user experience.

[0071] For example, for the above situation containing contextual information (definition of pricing orders with negative profits, pricing order records and business status, profit trends, etc. of Shenzhen Branch in December 2024) and business question text (how many pricing orders with negative profits of Shenzhen Branch in December 2024?), the question-and-answer big model defines "pricing orders with negative profits" according to the business rules in the contextual information, screens and counts all pricing order records of Shenzhen Branch in December 2024, calculates the number of pricing orders that meet the "negative profits" condition, and presents the results to the user in text form (such as "the number of pricing orders with negative profits of Shenzhen Branch in December 2024 is XX"), and can also attach relevant business explanations or suggestions, such as analysis of the reasons for negative profits or suggestions for improvement measures.

[0072] In a feasible implementation manner, before the step of determining the context information corresponding to the business question text based on the keyword information and slot information, the step further includes:

[0073] Step S301, querying similar field information of the slot information in the database using an edit distance algorithm;

[0074] It's important to note that the edit distance algorithm measures the similarity between two strings by calculating the minimum number of edit operations (such as insertion, deletion, and substitution) required to transform one string into the other. Similar field information refers to fields in the database that are similar to the extracted slot information, found using the edit distance algorithm. These similar fields may have business semantic relevance or similarity to the slot information, helping to more accurately understand and standardize slot information.

[0075] It is understandable that the slot information in the business question text may contain spelling errors, homophone errors, non-standard expressions, etc., which will affect the subsequent understanding and application accuracy of the slot information. In order to correct and standardize the slot information and ensure that the subsequent business question-and-answer process is based on accurate business entities and conditional information, similar field information is queried through the edit distance algorithm to discover fields that are similar in form to the slot information but may be more accurate or standardized in business meaning, thereby providing candidate options for further generating standard slot information.

[0076] Specifically, the edit distance algorithm query process can be implemented in a variety of ways: one way is to pre-establish an index of all field information in the database, and calculate the edit distance of each field information with other field information to construct an edit distance matrix. When it is necessary to query similar fields of slot information, the fields with a smaller edit distance to the slot information are obtained from the matrix as similar fields. Another way is to use a dynamic programming algorithm to calculate the edit distance between the slot information and each field information in the database in real time, and filter out fields with an edit distance less than a certain threshold as similar fields. In addition, the edit distance algorithm can be supplemented and optimized in combination with other string similarity algorithms (such as cosine similarity, etc.) to improve the accuracy and efficiency of similar field queries.

[0077] Step S302: Generate standard slot information based on the similar field information, so as to determine context information corresponding to the business question text according to the keyword information and the standard slot information.

[0078] It should be noted that standard slot information refers to slot information that has been corrected and standardized. It is more accurate in business semantics, complies with business data standards and specifications, and can be better matched and associated with business data in the database.

[0079] Furthermore, it should be noted that the purpose of generating standard slot information is to convert the original slot information into a standardized form that the system can more accurately recognize and process. This eliminates ambiguity and uncertainty caused by input errors or non-standardized expressions, and ensures the accuracy of subsequent business analysis and data queries based on the slot information. The process of generating standard slot information essentially maps the natural language entity information in the business question text to standardized entity identifiers in the business data model.

[0080] It is understandable that after generating standard slot information, combined with keyword information, the system can more accurately determine the key elements involved in the business question text, thereby more accurately constructing context information. By introducing standard slot information, the entity information in the business question text is more standardized and accurate, avoiding the context information deviation caused by errors or ambiguity in the original slot information, further improving the reliability and business relevance of the context information, and providing a more solid foundation for subsequent business analysis and question-and-answering.

[0081] Specifically, the process of determining contextual information includes the following steps: First, based on keyword information, relevant business concepts, business rules, and business processes are retrieved from the business terminology library and business rule library. Second, standard slot information is used to query the business data warehouse for corresponding business data records and business status information. Third, by combining the keyword information and the time and business scenario conditions in the standard slot information, relevant background information and related business events are extracted from historical business data and business logs. After this information is integrated and correlated, it forms complete contextual information.

[0082] For example, when the slot information is "Shenzhen Branch", first, the edit distance algorithm is used to query similar field information in the institution name field of the database, and it is found that the edit distance between "Shenzhen Branch" and "Shenzhen Branch" is small (only one character difference), then "Shenzhen Branch" is identified as field information similar to the slot information. Then, based on the queried field information "Shenzhen Branch" which is similar to "Shenzhen Branch", the system determines that "Shenzhen Branch" is an institution name that meets the business standards based on the institution name data dictionary and business rules in the database, and therefore uses "Shenzhen Branch" as the standard slot information to replace the original slot information "Shenzhen Branch"; then, for the keyword information "negative income", "number of pricing orders" and the standard slot information "Shenzhen Branch", "December 2024", the system retrieves the relevant business rule definitions and processing procedures based on the keyword "negative income", and queries the business data records and business status of the branch in the database based on the standard slot information "Shenzhen Branch", and at the same time extracts background information such as the branch's business activities and income trends during the time period based on "December 2024"; finally, use this information to jointly construct the context information of the business problem text.

[0083] In this embodiment, by performing an edit distance algorithm query on the slot information and generating standard slot information before determining the context information, the accuracy and reliability of the context information are further improved, thereby enhancing the accuracy and effectiveness of business questions and answers, and better solving the problem of inaccurate business information communication in the existing technology.

[0084] In a feasible implementation manner, the step of determining the context information corresponding to the business question text based on the keyword information and slot information includes:

[0085] Step S303, identifying table information corresponding to the business question text;

[0086] It should be noted that table information refers to the relevant information of the tables in the business database used to store specific business data, usually including identification information such as table number and table name. These tables are organized according to business themes and data structures and are used to store business data of different types and themes.

[0087] It is understandable that since different business problems usually involve business data on different topics, and these data are stored in different data tables, by identifying table information related to the business problem text and establishing an association between the business problem and the specific business data table, subsequent contextual information query and business analysis can be performed based on the correct data table, which can narrow the data search scope, improve query efficiency, and ensure that the queried data is directly relevant to the business problem.

[0088] Optionally, the identification table information can utilize semantic analysis and keyword matching methods in natural language processing technology to compare and match keywords in the business problem text with predefined data table topics and table names; or, use a large model to classify the business problem text and map the problem to the corresponding data table category; it can also determine the corresponding business data table based on business rules and business process definitions, combined with slot information in the business problem text (such as business type, business scenario, etc.).

[0089] Optionally, before identifying table information, you can create a database containing all the system's data table structures (including table numbers, table names, column names, etc.) and table-related summary information, perform semantic analysis on the business problem text, and match it with the semantics of the business problem and the table-related summary information in the database, so that you can identify table information such as tables and fields in the database that are related to the business problem text.

[0090] For example, for the business question text "How many pricing orders with negative profits in December 2024 of the Shenzhen Branch?", through semantic analysis and keyword matching, it is found that the keywords "negative profits" and "pricing order" in the question are related to the "Comprehensive Negative Profit Tracking Detail Table" in the business database. Therefore, the table information corresponding to the table is identified (such as the table number is 4, and the table name is "Comprehensive Negative Profit Tracking Detail Table") as the table information corresponding to the business question text.

[0091] Step S304: Generate context query conditions based on the table information and keyword information;

[0092] It should be noted that the context query condition refers to a combination of conditions used to query context information in the corpus, using table information and keyword information as key elements of the query to limit the scope and direction of the query.

[0093] It can be understood that table information determines the storage location and subject scope of business data, while keyword information clarifies the core focus and semantic direction of business problems. By combining table information and keyword information, an accurate query condition is constructed to quickly and accurately retrieve contextual information related to the business problem text in the preset corpus.

[0094] Step S305 : According to the context query condition, query the preset corpus for context information corresponding to the business question text.

[0095] It should be noted that a corpus refers to a database or data warehouse that stores business data table structure information and business context information (such as business keywords, default context logic, and unmentioned logic). Context information includes the business background, business rules, data relationships, and other information related to the business problem text, providing a comprehensive semantic background and business scenario description for the business problem.

[0096] It can be understood that by converting the identified table information and keyword information into actual query operations, contextual information that can help understand and analyze business problems is extracted from the corpus, and the business background and logic related to the business problems can be understood.

[0097] For example, based on the generated context query conditions, the table structure information of the "Negative Comprehensive Income Tracking Detail Table" (including column descriptions such as "customer name", "pricing order number", "income situation", etc.) and business context information (such as the processing flow corresponding to the business keyword "negative income", the method of calculating income in the default context logic, etc.) are queried in the preset corpus. These information together constitute the context information corresponding to the business problem text, providing detailed business background and data support for subsequent business analysis.

[0098] Specifically, the table information includes table number and table name, the keyword information includes keywords, key phrases, key attributes and named entities, the corpus stores table structure information and business context, the table structure information includes table description and column description, and the business context includes business keywords, default context logic and unmentioned logic.

[0099] The table number is a number or character code that uniquely identifies a data table, and the table name is the name of the data table. It usually intuitively reflects the subject of the business data stored in the table, such as "Corporate Deposit Pricing Details" or "Negative Comprehensive Income Tracking Details." The table number and table name play a key role in identifying the data table associated with the business question text.

[0100] Keyword information elements can fully reflect the core content and semantic details of the business question text. Keywords are single words that best reflect the subject of the business question, such as "revenue" and "pricing list." Key phrases are phrases composed of multiple words that express a complete business concept, such as "negative revenue" and "number of pricing lists." Key adjectives are words or phrases used to modify or qualify nouns, further clarifying the scope and conditions of the business question, such as "December 2024" and "Shenzhen Branch." Named entities are business entities with specific names, such as the institutional name "Shenzhen Branch" and the name of a business system.

[0101] A table description is a brief explanation of the overall content and purpose of a data table, such as "Corporate Deposit Pricing Sheet Details: Mainly involves inquiries on issues related to corporate deposits, including information such as the pricing sheet amount." A column description is a detailed description of each column in the data table, including the column's Chinese name, English name, data type, and column description, such as "Column Chinese name: Pricing Date, Table Number: 1, English Name: prc_dt, Type: date, Description: XX, Whether Empty: FALSE, Whether Enumerated: FALSE." Business keywords are words closely related to business concepts or operations, such as "over the upper limit" and "EVA commitment." Default context logic refers to the default business rules or logic used when certain conditions are not explicitly mentioned in a business question, such as "default to this month when the time range is not mentioned." Unmentioned logic describes business logic or rules that are not mentioned in the business question but may be relevant, such as "default includes all customer types when the customer type is not mentioned."

[0102] As you can see, table numbers and names accurately identify data tables, enabling the system to quickly locate and access business data. Comprehensive extraction of keywords, key phrases, key attributives, and named entities can deeply explore the semantic details of business problem text, providing richer semantic clues for subsequent contextual information queries and business analysis. The table and column descriptions in the table structure information help the system understand the content of the data table and the meaning of the data items. The business keywords, default context logic, and unmentioned logic in the business context provide the business background knowledge and default rules required for business problem analysis.

[0103] In this embodiment, by identifying table information, generating contextual query conditions, and querying contextual information in the corpus, the process of determining contextual information is further optimized, the relevance and accuracy of contextual information to business questions are improved, thereby improving the quality and efficiency of business questions and answers, and effectively solving the problem of inaccurate business information communication in the existing technology.

[0104] This embodiment provides a business question-and-answer method. First, it receives business question text in natural language format, allowing different users to raise diverse demands in an intuitive manner. Second, by extracting keyword information and slot information, it accurately captures the core elements and dynamic parameters in the question, avoiding the semantic understanding deviation problem caused by a single dimension. Then, based on the keyword and slot information, it generates contextual information that is highly adapted to the question, breaking the constraints of static analysis logic and realizing flexible integration of multi-source data. Finally, through the question-and-answer big model, it conducts deep reasoning and business logic analysis on the contextual information, and automatically generates business response results containing data analysis conclusions and decision recommendations. The deep integration of natural language interaction, dynamic semantic understanding and intelligent analysis capabilities can respond to users' complex and multi-dimensional pricing analysis needs in real time, significantly improve the timeliness and accuracy of business decisions, and greatly reduce the cost of manual intervention. It realizes contextual business question-and-answer based on a big model, solves the problem of low accuracy of business question-and-answer in existing technologies that relies on manual transmission of business information, and improves the accuracy of business question-and-answer.

[0105] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the first embodiment can be referred to the above introduction and will not be described in detail later. Figure 2 , Figure 2 This is a flow chart of the second embodiment of the business question-and-answer method of this application.

[0106] In this embodiment, the step of performing business analysis on the business question text using a question-answering model based on the context information to obtain a business answer result includes:

[0107] Step S401, generating business analysis prompt words for the business problem text through a thought chain prompt process according to the context information;

[0108] It's important to note that the thought chain prompting process guides the large model through step-by-step reasoning and analysis. By breaking down a problem into multiple logically connected sub-problems and providing gradual prompts and guidance, the model can gain a deeper understanding and analysis of the problem. Business analysis prompts are keywords, phrases, or question guides generated during the thought chain prompting process to guide the Q&A large model in business analysis. These help the large model focus on the key points of the business problem and conduct analysis and reasoning in a logical order.

[0109] It is understandable that through the thought chain prompting process, complex business problems are decomposed into multiple sub-problems that are easy to understand and analyze, and the big model is gradually guided to focus on the key elements and logical relationships of each sub-problem, so that the big model can more effectively use contextual information to conduct in-depth analysis of business problems, thereby improving the big model's understanding depth and analysis accuracy of business problems, and can give full play to the big model's reasoning ability, explore the deep business logic and associations behind business problems, and generate high-quality business response results.

[0110] Specifically, the thought chain prompting process can be implemented based on pre-defined thought chain templates for business scenarios. Based on the type of business problem and contextual information, the variables and conditions in the template can be filled in to generate specific business analysis prompts. For example, for business problems involving data statistics and analysis, a thought chain template can be used that follows the principle of "first determining the statistical object, then clarifying the statistical scope, then selecting the statistical method, and finally analyzing the results."

[0111] Step S402: Based on the business analysis prompt words, reasoning analysis is performed through the question-answering model to obtain a business response result.

[0112] It should be noted that reasoning analysis refers to the deep thinking and analysis of business problems by the big model using its internal knowledge and logic, including processes such as data query, logical reasoning, and the application of business rules. Business response results refer to the answers or responses to business questions generated by the big model based on reasoning analysis. These can be text or graphic explanations, data statistics, business operation suggestions, etc.

[0113] It is understandable that guiding the question-answering big model to perform reasoning analysis through business analysis prompts provides the big model with a clear analysis idea and logical framework, enabling the big model to more effectively utilize contextual information and the logical guidance in the business analysis prompts, so that the big model can avoid blind exploration and conduct a systematic and in-depth analysis of business problems along the direction guided by the prompts, thereby generating accurate, comprehensive and logical business response results.

[0114] Specifically, after receiving the business analysis prompt word, the question-answering big model first performs semantic understanding and logical analysis on the prompt word to identify the various sub-questions and logical relationships therein; then, combined with the business data and business rules in the context information, it gradually analyzes and answers each sub-question; during the analysis process, the big model may need to query relevant information in the internal knowledge base and data model, perform logical reasoning and verification; finally, integrate the answers to each sub-question to form a complete business response result, and present it to the user in an appropriate manner.

[0115] For example, for the business question text "How many pricing orders with negative profits in December 2024 of the Shenzhen Branch?" and its corresponding context information (including table information, business rules, data range, etc.), the business analysis prompt words generated through the thinking chain prompt process may be: "First, determine which pricing orders belong to the Shenzhen Branch; then filter out the pricing order records for December 2024; then determine which of these pricing orders meet the conditions for negative profits; finally, count the number of pricing orders that meet the conditions, and analyze the possible reasons and influencing factors for the negative profits." Based on the business analysis prompts above, the Q&A model first queries the Shenzhen branch's institutional identifier in the business database to identify pricing order records belonging to the Shenzhen branch. It then filters out pricing orders within the time period of "December 2024" based on the time condition. It then applies the definition of "negative returns" in the business rules to determine the profitability of the filtered pricing orders, filtering out those with negative returns. Finally, it counts the number of these pricing orders and analyzes the possible reasons for the negative returns (such as market interest rate changes, customer credit risk, etc.) and their impact on branch operations based on business data and rules, generating a business response, such as "The Shenzhen branch had XX pricing orders with negative returns in December 2024. The main reason is that market interest rate fluctuations that month led to a decrease in the returns of some pricing orders, while the credit risk of some customers increased. We recommend paying attention to market interest rate trends and customer credit status to optimize pricing strategies."

[0116] In a feasible implementation manner, the step of generating business analysis prompt words for the business problem text through a thought chain prompt process according to the context information includes:

[0117] Step S4011: Decompose the business question text into a sub-question chain with logical dependencies based on the context information, wherein the sub-question chain includes a plurality of interrelated sub-questions;

[0118] It should be noted that business problem text decomposition refers to the process of breaking down a complex business problem text into multiple logically interrelated simple sub-problems. A sub-problem chain refers to a chain of sub-problems arranged in a certain logical order, with logical dependencies between sub-problems. For example, first determine the business scope, then filter data for specific conditions, and finally perform statistical analysis. Sub-problems are relatively independent parts of the original business problem, and each sub-problem focuses on a specific aspect or step of the business problem, such as "determine the branch to which the pricing order belongs," "filter pricing orders for a specified month," "count the number of pricing orders with negative returns," etc.

[0119] It's understandable that the purpose of breaking down a business question text into a chain of sub-questions is to decompose a complex business question into multiple, easily understood and manageable sub-questions. By solving these sub-questions one by one, a complete business question answer can ultimately be constructed. Breaking down business questions reduces the complexity of the problem, allowing each sub-question to more accurately map to specific business data and rules. The logical dependencies of the sub-question chain ensure the integrity and coherence of the business analysis process, providing a clear framework for subsequent semantic binding and business analysis prompt generation, and ultimately improving the logic and accuracy of the entire business question-and-answer process.

[0120] Optionally, there are several ways to break down a business problem text into a chain of sub-problems. One method is to decompose the business problem into multiple steps based on business processes and business rules, in the order of business processing, with each step corresponding to a sub-problem. For example, for data query business problems, the problem can be broken down in the order of "determining the business object - clarifying the time range - filtering conditions - statistical analysis." Another method is to use dependency syntactic analysis in natural language processing technology to identify the logical relationships between words in the business problem text and construct a chain of sub-problems based on these logical relationships. In addition, business knowledge can be combined to perform templated decomposition of common business problems, forming a standard sub-problem chain model to improve decomposition efficiency.

[0121] Step S4012: semantically bind the entities in the sub-question chain to the table structure of the database to generate an enhanced question statement with table association tags;

[0122] It should be noted that entities refer to the business objects, business concepts, data items, etc. involved in the sub-problem chain, such as "Shenzhen Branch", "December 2024", "negative income", "number of pricing orders", etc. The table structure of the database refers to the structural information of the data table in the business database, including table name, column name, data type, etc. Semantic binding refers to establishing a semantic association relationship between the entities in the sub-problem chain and the corresponding data elements in the database table structure (such as tables, columns, etc.), so that the system can understand the specific storage location and representation of the entity in the database. Table association marking refers to the identification of which tables and columns are associated in the enhanced problem statement, which is used to guide subsequent data query and business analysis. Enhanced problem statement refers to a sub-problem chain statement that contains table association markings after semantic binding, which is more operational and database-related.

[0123] It can be understood that by semantically binding the entities in the sub-question chain to the database table structure, business entities in natural language are mapped to specific data elements in the database, enabling a direct connection between business questions and business data. Through semantic binding, the system can clearly identify the data source and operation object of each sub-question in the database, providing a foundation for subsequent generation of accurate database query statements and data operations. The table association markup in the enhanced question statement makes the sub-question chain more operational, directly guiding the database query and analysis process, and improving the automation and accuracy of business problem processing.

[0124] Step S4013: generating business analysis prompt words for the business problem text based on the enhanced problem statement of the table association mark.

[0125] It should be noted that business analysis prompts refer to keywords, phrases or question guides generated based on enhanced question statements, which are used to guide the question-answering big model to perform business analysis. The prompts combine the semantics of business problems and database table structure information, and can more accurately guide the big model to perform data query, logical reasoning and business analysis.

[0126] It's understandable that generating business analysis prompts based on enhanced question formulations further refines and clarifies business analysis guidance, tightly integrating sub-question chains with database operations. This allows the large model to more clearly understand the database operation objects and goals of each sub-question, thereby generating more accurate and actionable business responses. Business analysis prompts play a key role in guiding the chain of thought prompts. By translating natural language business questions into specific prompts related to database operations, the large model can better utilize database resources for business analysis.

[0127] Specifically, the process of generating business analysis prompts can include the following steps: first, extracting table-related tags and key semantic information for each sub-problem from the enhanced problem statement; then, constructing corresponding database query operation descriptions or logical reasoning steps based on the table-related tags and business rules; and finally, combining these descriptions and steps into a coherent prompt sequence to form a complete business analysis prompt. Furthermore, during the generation process, prompts can be optimized and adjusted based on the complexity of the business problem and the requirements of the business rules to ensure their rationality and effectiveness.

[0128] For example, first, for the business question text "How many pricing orders with negative revenue in December 2024 for the Shenzhen branch?", with the support of contextual information (including business rules, data table structure, etc.), it is broken down into the following sub-question chain: "1. Determine which pricing orders belong to the Shenzhen branch? 2. Among these pricing orders, filter out pricing order records for December 2024? 3. Determine which of the filtered pricing orders meet the negative revenue condition? 4. Count the number of pricing orders that meet the condition?"

[0129] Then, for the first sub-problem in the above sub-problem chain, "Determine which pricing orders belong to the Shenzhen Branch?", the entity "Shenzhen Branch" is associated with the "Branch Name" column of the "Negative Comprehensive Income Tracking Detail Table" in the database through semantic binding, and the enhanced question statement is generated as "[Table: Negative Comprehensive Income Tracking Detail Table, Column: Branch Name] Determine which pricing orders have a branch name of Shenzhen Branch?"; the second sub-problem, "Among these pricing orders, filter out the pricing order records for December 2024?", the entity "December 2024" is associated with the "Tracking Month" column of the "Negative Comprehensive Income Tracking Detail Table", and the enhanced question statement is "[Table: Negative Comprehensive Income Tracking Detail Table, Column: Tracking Month] Among the pricing orders whose branch name is Shenzhen Branch, filter out the pricing order records whose tracking month is December 2024?".

[0130] Finally, based on the enhanced problem statement above, the generated business analysis prompt might be: "Query pricing order records for the Shenzhen branch based on the branch name column in the negative comprehensive income tracking details table; in the query results, filter pricing orders for December 2024 based on the tracking month column; apply the negative income condition to count the number of pricing orders that meet the condition; and analyze the distribution and possible causes of pricing orders with negative income."

[0131] In this implementation, by breaking down the business question text into sub-question chains, performing semantic binding to generate enhanced question statements, and generating business analysis prompts based on this, the business analysis guidance process is further refined, enabling the question-and-answer model to more accurately combine database operations for business analysis, thereby improving the accuracy of business response results and effectively solving the problem of insufficiently refined business question analysis affecting question-and-answer accuracy and the timeliness of database updates.

[0132] In a feasible implementation, the step of performing reasoning analysis based on the business analysis prompt word through the question-answering model to obtain a business response result includes:

[0133] The business analysis prompt words are input into the question-answering model and processed as follows:

[0134] Step S4021: Generate several candidate business query statements in parallel through a multi-path reasoning mechanism based on the business analysis prompt words;

[0135] It should be noted that the multi-path reasoning mechanism refers to the process by which a large model simultaneously attempts multiple different reasoning paths and methods to generate candidate business query statements. These paths may be based on different assumptions, business rules, or data associations. Candidate business query statements are possible business data query statements generated based on different reasoning paths. They are potential solutions to business problems and require subsequent verification and repair processes to determine the optimal query statement.

[0136] It's understandable that by using a multi-path reasoning mechanism to generate multiple candidate business queries, exploring different possible solutions, and leveraging the parallel processing capabilities of large models to rapidly generate multiple query solutions, the probability of finding the correct and efficient query statement is increased. Different reasoning paths may be based on different business understandings, data association methods, or query strategies. By generating multiple candidate queries in parallel, we can cover more possibilities, reduce the risk of errors caused by deviations from a single reasoning path, and provide a richer range of options for subsequent selection of the optimal query statement.

[0137] Specifically, the multi-path reasoning mechanism can be implemented based on the internal architecture and algorithms of the large model. For example, the large model can generate corresponding query fragments based on the different sub-questions or logical steps in the business analysis prompt words, and then combine these fragments to form different candidate query statements.

[0138] Step S4022: executing the candidate query statements to obtain candidate query results, and verifying and repairing the candidate query statements based on the candidate query results to obtain a target business query statement;

[0139] It should be noted that candidate query results refer to the data obtained after executing these query statements. Verification and repair refers to the process of verifying the correctness of candidate query statements and correcting their syntax or logic to ensure that the query statements can be executed correctly and return accurate business data. The target business query statement is the optimal query statement obtained after verification and repair, which can accurately meet the query requirements of the business problem and return correct business data.

[0140] It is understandable that subsequent query results obtained by executing candidate query statements are verified and repaired to obtain the target business query statement, ensuring that the query statement ultimately executed is correct and can accurately retrieve the business data required for the business question from the database. Because candidate query statements are generated through a multi-path reasoning mechanism, they may contain syntax errors, logical deviations, or mismatches with the actual database data structure. By executing candidate query statements and analyzing the candidate query results, these problems can be discovered and repaired in a targeted manner, thereby improving the quality and reliability of the query statements, effectively avoiding business response deviations caused by incorrect query statements, and ensuring the accuracy and credibility of business questions and answers.

[0141] Specifically, the verification and repair process can include the following aspects:

[0142] 1. Syntax Verification: This checks whether the candidate query statement conforms to the syntax rules of the query language used by the database (such as SQL). For example, the statement's grammatical structure is correct, keywords are spelled correctly, and brackets are matched. For query statements with grammatical errors, corrections are made according to the error message.

[0143] 2. Semantic Verification: Verify that the logic of the candidate query statement complies with business rules and the requirements of the business problem. For example, check whether the query conditions accurately express the constraints in the business problem, whether aggregate functions are used correctly, and whether the associations between data tables are reasonable. Semantic verification can be performed by comparing the candidate query results with the expected business results and performing logical analysis based on business rules.

[0144] 3. Data Verification: Analyze the data integrity and correctness of candidate query results. For example, check whether the query results are empty (when the business question expects a result), whether the data is within a reasonable range, and whether there are duplicate or abnormal data. If the query results contain data issues, you may need to adjust the query conditions, data tables, or query methods.

[0145] 4. Performance Verification: Evaluate the execution efficiency of candidate query statements. For query statements that take too long to execute or consume too many resources, optimize them by adding indexes or refining query conditions.

[0146] During the validation process, we can perform targeted repairs on candidate query statements based on the feedback from the candidate query results. For example, if a query statement is found to have syntax errors due to incorrect column names, the column names can be corrected; if the data range of the query results does not meet business expectations, the query conditions can be adjusted. After multiple rounds of validation and repair, a target business query statement is ultimately determined that accurately and efficiently returns business data.

[0147] Step S4023: execute the target business query statement to obtain business structured data, and fuse the business structured data with the unstructured data of the preset knowledge base to obtain a business response result.

[0148] It should be noted that business structured data refers to business data obtained from a query in a database and organized in a structured form (such as tables, JSON, etc.). These data come directly from the business database and have a clear structure and data type. The unstructured data of the preset knowledge base refers to the unstructured information pre-stored in the knowledge base, such as business documents, policy documents, FAQs (Frequently Asked Questions), etc. These data are stored in text form and contain rich business background knowledge and experience summaries. Fusion refers to the comprehensive processing and integration of structured data and unstructured data, and through association, supplementation, etc., the two complement each other to form a more complete and in-depth business response result. The business response result is the final output after fusion, which combines the data query results and business knowledge, and can answer business problems more comprehensively and accurately.

[0149] It's understandable that by integrating structured business data with unstructured data, leveraging the precision of structured data and the richness of unstructured data, we can provide users with comprehensive responses that include both specific data results and detailed business explanations and context. Structured data can intuitively present the quantitative results and objective facts of business issues, while unstructured data provides deeper information such as the background, cause analysis, and solution recommendations. The integration of the two can meet users' diverse needs for business problem resolution, enhancing the practicality and value of business responses.

[0150] Specifically, the fusion process can match and associate key data indicators in structured data with relevant business knowledge in unstructured data based on the type and needs of the business question. For data query business questions, you can first present the query results of structured data, such as statistics and data lists, and then cite business explanations, policy basis, or operational suggestions in the unstructured data for supplementary explanations. For business consulting questions, you can first extract relevant business knowledge and answers from the unstructured data, and then use examples or data from the structured data to verify and support them. Leveraging natural language generation technology, structured and unstructured data can be integrated into a coherent text response. At the same time, data visualization technology can be combined to present structured data graphically, and structured data can be interpreted and explained in conjunction with unstructured data.

[0151] For example, the business structured data obtained after executing the target business query statement is "The number of pricing orders with negative profits in December 2024 of the Shenzhen Branch is 58." The information obtained from the unstructured data in the pre-set knowledge base includes the following: "According to the provisions for handling negative-return pricing orders in the XX Bank Pricing Policy Manual, negative returns may be caused by factors such as market interest rate fluctuations and increased customer credit risk. Business departments are advised to monitor the subsequent returns of these pricing orders and, if necessary, negotiate with customers to adjust pricing plans. Furthermore, based on case studies of similar business scenarios, it is recommended to categorize and analyze pricing orders with negative returns, identify key risk factors, and develop targeted risk control measures." The final business response was: "The Shenzhen Branch had 58 pricing orders with negative returns in December 2024. According to our pricing policy, negative returns may be caused by factors such as market interest rate fluctuations and increased customer credit risk. It is recommended to monitor the subsequent returns of these pricing orders and consider negotiating with customers to adjust pricing plans. Furthermore, pricing orders with negative returns should be categorized and analyzed to identify key risk factors and develop targeted risk control measures to optimize pricing strategies and improve business returns."

[0152] In this embodiment, candidate query statements are generated through multi-path reasoning, and the target query statements are obtained through verification and repair. The business response results are generated by integrating structured and unstructured data, which further optimizes the query and response process of business questions and answers, improves the accuracy and richness of business responses, and effectively solves the problems of inaccurate business data query and single response content in the existing technology.

[0153] In this embodiment, business analysis prompt words are generated through the thought chain prompt process, and the question-answering model is guided to perform reasoning analysis, which further improves the intelligence level and response quality of business questions and answers, enabling the business question-answering system to analyze business problems more deeply and accurately, and provide users with the most relevant answers and suggestions, effectively solving the problems of shallow business analysis and inaccurate responses in the existing technology.

[0154] Based on the first embodiment and / or the second embodiment of the present application, in the third embodiment of the present application, the same or similar contents as those of the first and second embodiments mentioned above can be referred to the above introduction and will not be repeated later.

[0155] For example, to help understand the implementation process of the business question-answering method obtained by combining the first embodiment and the second embodiment, please refer to Figure 3 , Figure 3 This is a schematic diagram of the architecture of the third embodiment of the business question-and-answer method of this application. Specifically, this embodiment mainly includes:

[0156] 1. Infrastructure module, which defines some data analysis-related table information and business context corpora, as well as the vector database that stores these corpora.

[0157] 2. The general service module defines a series of table DDL information queries, table description information queries, business context acquisition, user permission control, and third-party platform calls (intent recognition, policy question and answer, Q&A question and answer, fuzzy search, and vector retrieval).

[0158] 3. The key information retrieval module extracts key feature information from user questions to identify the context-related information contained in the questions, extracts and standardizes the slot information in user questions (such as organization, customer name, account manager name, group name), and corrects the erroneous information in user questions.

[0159] 4. The summary connection module combines the table structure, table description information and problem context information to filter out the table and field information most relevant to the problem, and adds description information and SQL example information to the prompt.

[0160] 5. The SQL generation module calls the large model to complete the generation of candidate SQL, performs syntax verification on the SQL and self-corrects in case of errors, and finally returns the SQL execution result and adds description and summary information.

[0161] 6. Visual display module: Visual SQL displays data analysis or query result information, supporting both text and graphic formats.

[0162] Each module specifically includes:

[0163] 1. Infrastructure module: The infrastructure module includes a corpus and a Gaussian vector database. The corpus can be an Excel document. The corpus contains the original system table structure information (table and column descriptions) and business context (business keywords and default logic, unmentioned default logic, and business terms). Some sample formats of the corpus are as follows:

[0164] Table description format example:

[0165]

[0166] Example of column description format:

[0167]

[0168] Example business term format:

[0169]

[0170] A portion of the corpus content is loaded into the Gaussian vector library through the API interface to facilitate retrieval by other modules.

[0171] 2. General service module: This module includes four sub-modules: table service, context service, permission control, and third-party call. It provides basic services for converting natural language questions into SQL.

[0172] (1) Table service: Figure 4 As shown, Figure 4 This is a flow chart of the table service of the third embodiment of the business question-and-answer method of this application. The main process is as follows:

[0173] 1) Prepare a table definition file in CSV format, containing information such as the table name, field name, description, data type, and whether it is empty. Each table corresponds to one file.

[0174] 2) Prepare a table configuration file in JSON format, including the table definition file path and default field properties.

[0175] 3) Construct table warehouse services to provide table file data reading and table and field information query functions.

[0176] 4) Construct table service, providing the following methods:

[0177] (1) Get table DDL information by table number or table name

[0178] (2) Get field information by table name (normal field or slot filling field)

[0179] (3) Get the table number by table name

[0180] (4) Batch obtain table DDL information by table name

[0181] (2) Business Context Service: This service primarily provides contextual search related to the question. Search methods include keyword search, table name search, and column name search. The goal is to obtain relevant contextual information and improve the accuracy of SQL generated by large models.

[0182] There are two retrieval channels:

[0183] 1) Vector database retrieval: A portion of the corpus is loaded into the Gaussian vector database and retrieved through the API interface provided by a third party.

[0184] 2) Excel search: A corpus is mounted in an Excel file and searched through an Excel reader.

[0185] like Figure 5 As shown, Figure 5This is a flow chart of the business context service of the third embodiment of the business question-and-answer method of this application. The main process is as follows:

[0186] 1) First, for a data analysis question, identify and list all relevant keywords based on the question content and optional context. At the same time, identify the table information (table number, table name, etc.) that is most relevant to the question content based on the question.

[0187] 2) The Vector Recall Context Service uses table ID, table name, and keyword information as query conditions to retrieve the most similar context information (table description, column description, and business terms) from the Gaussian vector database (cosine similarity). The Excel Corpus Context Service uses table ID as query conditions to retrieve context (professional keywords and default logic, and default logic not mentioned) from the Excel corpus.

[0188] (3) Permission control: The core capability of this invention is to use natural language to perform data analysis. The data analysis capability mainly relies on the generated SQL. Therefore, the core of permission control is to control the scope of SQL data access. Users of the banking system usually include information in dimensions such as system roles and institutions. We divide users into five types of roles, and then divide roles into three types of permissions. Different SQL filtering conditions are generated according to the permission type.

[0189] (4) Third-party service calls: Third-party service calls include:

[0190] 1) Call the embedding model (text embedding model) of the inline model platform to complete the vectorized storage of the corpus.

[0191] 2) Call the AI ​​smart factory to develop intent recognition components, open API services, and complete user question intent recognition.

[0192] 3) Call the GPT platform API to complete knowledge base training (QA question and answer pairs, system business policy documents, etc.), maintenance, and knowledge retrieval. Provide a retrieval source for business policy and operational process Q&A.

[0193] 4) Call the AI ​​smart factory interface to extract problem slot information (customer name, account manager, organization, and group name).

[0194] 5) Call the pricing middle platform AIGC application interface to complete the slot information standardization processing.

[0195] 3. Functional Module: This module handles the entire process from submitting a data analysis question to outputting an analysis summary report. It includes three submodules: key information retrieval, summary connection, and SQL generation. Each module is described in detail below:

[0196] (1) Key Information Retrieval: Responsible for extracting question keywords and filling in slot information (extraction and correction), providing input information for the Schema Linking module to build prompts.

[0197] 1) Keyword extraction: Keywords refer to key phrases, key attributes, and named entities. For example, for the question "What are the current business types of deposit pricing sheets?", the corresponding keywords are ['deposit', 'pricing sheet', 'business types'].

[0198] 2) Slot information filling: Slot information refers to some characteristic information in the question, such as the organization, customer name, account manager name, group name, etc. The purpose of slot information filling is to correct the incorrect information in the question, thereby obtaining the correct dimension query conditions and helping the large model generate correct SQL. Slot information filling includes two steps: slot information extraction and standardization. Take a question as an example:

[0199] Original question: How many pricing orders did Shenzhen Branch have with negative returns in December 2024?

[0200] like Figure 6 As shown, Figure 6 This is a flow chart of slot information filling in the third embodiment of the business question-and-answer method of this application. The flow of slot information filling is roughly as follows.

[0201] (1) Extract slot information: Use large model prompt technology to extract slot information:

[0202] [{"slotName":"Branch Name","slotContent":"Shen Shen Branch","model":"11m"}]

[0203] (2) Standardization: Call the pricing platform ES interface fuzzy matching and use the edit distance algorithm to find the name most similar to the slot information to ensure the correctness of the slot information:

[0204] [{"slotName":"Branch Name","slotContent":"Shenzhen Branch","model":"1lm"}]

[0205] (2) Summary connection: This module selects the most relevant tables, fields, and context information for the corrected problem. The main algorithm flow is as follows:

[0206] 1) Table selection: Based on the questions input by the user, the big model prompt technology is used to let the big model select the most appropriate table.

[0207] 2) Build a complete context: Using the existing information (keywords, table indexes), call the business context service submodule in the general module to complete the construction of context-related information.

[0208] 3) Column selection: After completing table selection and full context retrieval related to the question, use the large model prompt technology to find the table fields most relevant to the question and output them in JSON format.

[0209] (3) SQL generation: A complete SQL generation process includes candidate SQL generation, SQL verification, SQL self-repair, result execution and report output (including description + summary).

[0210] 1) Candidate SQL Generation: The CoT prompt module is introduced to allow the model to gradually demonstrate its reasoning process as required when generating answers, thereby improving the accuracy and explainability of the answers.

[0211] 2) SQL validation and self-repair: SQL validation is introduced to ensure that the output SQL is executable standard SQL and connect to the database for SQL execution. If a SQL syntax error is reported, the syntax error information will be used as the prompt content, allowing the model to repair the SQL and retry up to 3 times before returning the final result.

[0212] 3) SQL result execution and report output: Use prompt technology and Python drawing library to summarize and graphically display SQL execution results.

[0213] 4. Visual display module: Visually render the output results.

[0214] This embodiment leverages the data analysis capabilities of a large model to provide a complete end-to-end solution from natural language questions to data analysis results. This addresses the challenges of intelligent pricing business analysis and post-supervision analysis. Automatically parsing policy documents addresses the issues of delayed and inaccurate policy information updates. Regularly updating business knowledge documents, operating manuals, and Q&A sessions addresses the significant labor consumption associated with on-call Q&A.

[0215] It should be noted that the above examples are only used to understand this application and do not constitute a limitation on the business question-and-answer method of this application. More simple transformations based on this technical concept are all within the scope of protection of this application.

[0216] This application also provides a business question and answer device, please refer to Figure 7 , the business question-answering device includes:

[0217] Receiving module 10, for receiving a business question text;

[0218] Extraction module 20, used to extract keyword information and slot information from the business question text;

[0219] A determination module 30, configured to determine context information corresponding to the business question text based on the keyword information and slot information;

[0220] The analysis module 40 is used to perform business analysis on the business question text based on the context information through the question-answering model to obtain a business answer result.

[0221] Optionally, the determining module 30 is further configured to:

[0222] Using the edit distance algorithm, query the database for similar field information of the slot information;

[0223] Based on the similar field information, standard slot information is generated to determine context information corresponding to the business question text according to the keyword information and the standard slot information.

[0224] Optionally, the determining module 30 is further configured to:

[0225] Identify the table information corresponding to the business problem text;

[0226] Generate context query conditions based on the table information and keyword information;

[0227] According to the context query condition, the context information corresponding to the business question text in the preset corpus is queried.

[0228] Optionally, the table information includes table number and table name, the keyword information includes keywords, key phrases, key attributes and named entities, the corpus stores table structure information and business context, the table structure information includes table description and column description, and the business context includes business keywords, default context logic and unmentioned logic.

[0229] Optionally, the analysis module 40 is further configured to:

[0230] Generating business analysis prompt words for the business problem text through a thought chain prompt process according to the context information;

[0231] Based on the business analysis prompt words, reasoning analysis is performed through the question-answering model to obtain business response results.

[0232] Optionally, the analysis module 40 is further configured to:

[0233] Decomposing the business question text into a sub-question chain with logical dependencies according to the context information, wherein the sub-question chain includes a plurality of interrelated sub-questions;

[0234] Semantically binding the entities in the sub-problem chain with the table structure of the database to generate an enhanced problem statement with table association tags;

[0235] Based on the enhanced problem statement of the table association mark, business analysis prompt words of the business problem text are generated.

[0236] Optionally, the analysis module 40 is further configured to:

[0237] The business analysis prompt words are input into the question-answering model and processed as follows:

[0238] Based on the business analysis prompt words, a plurality of candidate business query statements are generated in parallel through a multi-path reasoning mechanism;

[0239] Executing the candidate query statements to obtain candidate query results, and verifying and repairing the candidate query statements based on the candidate query results to obtain a target business query statement;

[0240] The target business query statement is executed to obtain business structured data, and the business structured data is integrated with the unstructured data of the preset knowledge base to obtain a business response result.

[0241] The business Q&A device provided in this application utilizes the business Q&A method described in the aforementioned embodiments, resolving the technical issue of prior art business Q&A, which relies on manual transmission of business information and suffers from low accuracy. Compared to the prior art, the beneficial effects of the business Q&A device provided in this application are the same as those of the business Q&A method described in the aforementioned embodiments. Other technical features of the business Q&A device are the same as those disclosed in the aforementioned embodiments and are not further elaborated upon here.

[0242] The present application provides a business Q&A device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the business Q&A method in the first embodiment described above.

[0243] Reference below Figure 8, which shows a schematic diagram of the structure of a business question-and-answer device suitable for implementing an embodiment of the present application. The business question-and-answer device in the embodiment of the present application may include, but is not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), in-vehicle terminals (such as in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 8 The business question-and-answer device shown is merely an example and should not limit the functions and scope of use of the embodiments of the present application.

[0244] like Figure 8 As shown, the business Q&A device may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the business Q&A device. The processing device 1001, the read-only memory 1002, and the random access memory 1004 are connected to each other via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: an input device 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the business Q&A device to communicate with other devices wirelessly or wired to exchange data. Although the figure shows a business Q&A device with various systems, it should be understood that it is not required to implement or have all of the systems shown. More or fewer systems may be implemented or provided instead.

[0245] In particular, according to the embodiments disclosed in the present application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the embodiments disclosed in the present application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program comprising program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device 1003, or installed from a read-only memory 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are performed.

[0246] The business Q&A device provided in this application utilizes the business Q&A method in the above-described embodiment, resolving the technical issue of prior art business Q&A, which relies on manual transmission of business information and suffers from low accuracy. Compared to the prior art, the business Q&A device provided in this application offers the same beneficial effects as the business Q&A method provided in the above-described embodiment. Other technical features of this device are the same as those disclosed in the above-described embodiment and are not further elaborated here.

[0247] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0248] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0249] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer programs) stored thereon, and the computer-readable program instructions are used to execute the business question-and-answer method in the above-mentioned embodiment.

[0250] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared or semiconductor systems, systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0251] The computer-readable storage medium may be included in the business question-and-answer device, or may exist independently without being incorporated into the business question-and-answer device.

[0252] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by the business question and answer device, the business question and answer device: receives a business question text; extracts keyword information and slot information from the business question text; determines the context information corresponding to the business question text based on the keyword information and slot information; and performs business analysis on the business question text through the question and answer big model based on the context information to obtain a business response result.

[0253] Computer program code for performing the operations of the present application may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on the remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0254] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, program segment or part of code, and the module, program segment or part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, as well as the combination of boxes in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or can be implemented using a combination of dedicated hardware and computer instructions.

[0255] The modules described in the embodiments of the present application may be implemented in software or hardware, wherein the name of a module does not necessarily limit the unit itself.

[0256] The computer-readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the aforementioned business Q&A method. This computer-readable storage medium can address the technical issue of prior art business Q&A, which relies on manual communication of business information and suffers from low accuracy. Compared to the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the business Q&A method provided in the aforementioned embodiment, and are not further elaborated here.

[0257] The present application also provides a computer program product, including a computer program, which implements the steps of the above-mentioned business question-and-answer method when executed by a processor.

[0258] The computer program product provided in this application can solve the technical problem of the low accuracy of existing business Q&A techniques, which rely on manual communication of business information. Compared with the existing technology, the beneficial effects of the computer program product provided in this application are the same as those of the business Q&A methods provided in the above embodiments, and will not be elaborated here.

[0259] The above description is only part of the embodiments of the present application and does not limit the patent scope of the present application. All equivalent structural transformations made by using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect application in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A business question-answering method, characterized in that: The business question-answering method includes: Receive business question text; Extracting keyword information and slot information from the business question text; Determining context information corresponding to the business question text based on the keyword information and slot information; Based on the context information, the business question text is analyzed using a large question-answer model to obtain a business answer result.

2. The business question-answering method according to claim 1, wherein: Before the step of determining the context information corresponding to the business question text based on the keyword information and the slot information, the method further includes: Using the edit distance algorithm, query the database for similar field information of the slot information; Based on the similar field information, standard slot information is generated to determine context information corresponding to the business question text according to the keyword information and the standard slot information.

3. The business question-answering method according to claim 1, wherein: The step of determining context information corresponding to the business question text based on the keyword information and slot information includes: Identify the table information corresponding to the business problem text; Generate context query conditions based on the table information and keyword information; According to the context query condition, the context information corresponding to the business question text in the preset corpus is queried.

4. The business question-answering method according to claim 3, wherein: The table information includes table number and table name, the keyword information includes keywords, key phrases, key attributes and named entities, the corpus stores table structure information and business context, the table structure information includes table description and column description, and the business context includes business keywords, default context logic and unmentioned logic.

5. The business question-answering method according to claim 1, wherein: The step of performing business analysis on the business question text using a question-answering model based on the context information to obtain a business answer result includes: Generating business analysis prompt words for the business problem text through a thought chain prompt process according to the context information; Based on the business analysis prompt words, reasoning analysis is performed through the question-answering model to obtain business response results.

6. The business question-answering method according to claim 5, wherein: The step of generating business analysis prompt words for the business problem text through a thought chain prompt process according to the context information includes: Decomposing the business question text into a sub-question chain with logical dependencies according to the context information, wherein the sub-question chain includes a plurality of interrelated sub-questions; Semantically binding the entities in the sub-problem chain with the table structure of the database to generate an enhanced problem statement with table association tags; Based on the enhanced problem statement of the table association mark, business analysis prompt words of the business problem text are generated.

7. The business question-answering method according to claim 5, wherein: The step of performing reasoning analysis based on the business analysis prompt word through the question-answering big model to obtain a business response result includes: The business analysis prompt words are input into the question-answering model and processed as follows: Based on the business analysis prompt words, a plurality of candidate business query statements are generated in parallel through a multi-path reasoning mechanism; Executing the candidate query statements to obtain candidate query results, and verifying and repairing the candidate query statements based on the candidate query results to obtain a target business query statement; The target business query statement is executed to obtain business structured data, and the business structured data is integrated with the unstructured data of the preset knowledge base to obtain a business response result.

8. A business question-answering device, characterized in that: The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the business question-answering method according to any one of claims 1 to 7.

9. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the business question-and-answer method according to any one of claims 1 to 7 are implemented.

10. A computer program product, characterized in that The computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the business question-and-answer method according to any one of claims 1 to 7 are implemented.

Images