Financial field data intelligent dialogue type analysis system based on AI large model

Abstract

The invention relates to a financial field data intelligent dialogue analysis system based on an AI large model, which relates to the field of artificial intelligence and financial data analysis, and comprises a database preposed preprocessing module and a financial data intelligent dialogue analysis module. The financial data analysis module is used for normalizing financial data, and the financial data analysis module is used for realizing natural language question analysis, data query and analysis, dialogue answer generation and visual display based on a Qwen-72B model and a tool set. The system improves the accuracy through professional knowledge injection, supports data attribution and user feedback optimization, and is suitable for complex financial data analysis scenes.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and financial data analysis, and in particular to an intelligent conversational analysis system for financial data based on an AI big model. Background Art

[0002] In today's digital age, AI (artificial intelligence) technology has achieved groundbreaking progress, particularly with the emergence of AI big models, which have brought revolutionary changes to numerous fields. Based on deep learning architectures, AI big models are able to learn common language patterns, semantic understanding, and knowledge representation through unsupervised pre-training on massive amounts of data. These models have demonstrated outstanding performance in natural language processing (NLP) tasks such as text generation, machine translation, and sentiment analysis, enabling more natural and fluid interaction between humans and computers. At the same time, data volumes are exploding, and various industries are accumulating vast amounts of data resources. Efficiently extracting valuable information from this data has become a key challenge. This technological development trend is prompting various industries to actively explore the deep integration of AI big models with their business scenarios to improve efficiency, optimize decision-making, and create new value.

[0003] In the financial sector, data complexity and diversity continue to increase. Enterprises not only possess traditional structured financial data, such as financial statements and detailed account statements, but also generate a vast amount of unstructured data, such as financial reports, meeting minutes, and policy documents. Financial personnel, business managers, investors, and other users need to quickly obtain accurate and insightful information from this complex data in their daily work and decision-making. However, traditional financial data processing methods have many limitations. Traditional query tools primarily rely on Structured Query Language (SQL), requiring users to have certain database knowledge and operational skills, making them a high barrier to entry for non-technical financial professionals. Furthermore, when faced with complex financial problems, such as analyzing corporate financial risks in light of the macroeconomic environment or assessing the long-term impact of new policies on a company's financial health, traditional tools are unable to effectively understand the semantics of the problem and provide comprehensive and in-depth analysis. Summary of the Invention

[0004] Purpose of the invention: To provide an intelligent conversational analysis system for financial data based on an AI big model to solve the above-mentioned problems existing in the prior art.

[0005] Technical solution: A financial data intelligent conversational analysis system based on an AI big model, including a database pre-processing module and a financial data intelligent conversational analysis module; the database pre-processing module is used to format financial data, clean up erroneous and missing values, semantically annotate, add metadata, perform desensitizing filtering, and build data indexes; the financial data intelligent conversational analysis module is used to receive user natural language questions, understand and parse the questions, call the AI ​​big model and tool set to query and analyze data, generate conversational answers, and support data attribution and graphical display; wherein, the AI ​​big model is the Qwen-72B model, and the tool set includes PlanningAgent, text2sql, and LLM summary tools.

[0006] In a further embodiment, the database pre-processing module includes six components: a data table configuration unit, a metadata management unit, a field configuration unit, a table association configuration unit, an indicator management unit, and a dimension management unit. The data table configuration unit determines the number of data tables to be created, the name of each table, and the overall layout based on business requirements. The metadata management unit comprehensively coordinates metadata, recording detailed background information such as the data source, meaning, owner, and update frequency. The field configuration unit meticulously defines and configures each field in the data table, specifying each field's name, data type, length limit, and nullability attributes. The table association configuration unit deeply analyzes the inherent connections between data and establishes one-to-one, one-to-many, or many-to-many associations between different data tables based on business rules. The indicator management unit systematically plans and controls key indicators, clarifying their definitions, calculation methods, data sources, and applicable scenarios. The dimension management unit rationally plans and manages data dimensions, determining data classification and analysis from perspectives such as time, region, and product.

[0007] In a further embodiment, the vector database storage submodule specifically performs operations including mapping each data record into a vector based on the characteristics of the structured data. For example, in financial data modeling, the amount, time, transaction type, and other information of each transaction record are converted into a vector using a specific algorithm. The converted vector data is stored in the vector database Milvus. During storage, it is configured based on the characteristics of the vector database, supporting multiple index types, such as IVF-FLAT and HNSW, to optimize storage and query performance. Relevant metadata is added to each vector data, such as the source and creation time of the data, to facilitate subsequent data management and query screening.

[0008] In a further embodiment, the PlanningAgent generates a task plan by breaking down the user question into subtasks and defining dependencies between the subtasks. Subtask breaking must adhere to the following requirements: No hypothetical tools are provided; the tools must actually exist. If the broken down tasks share the same metrics and can be placed into a single tool, there is no need to break them into multiple subtasks. If the user's question is a data query for different entities, it should be broken down into multiple subtasks. If the user's question contains multiple conditions, it should be placed in a single subtask and not split into multiple subtasks. If the user's question contains multiple unrelated subtasks, it should be broken down into multiple subtasks. If the user's question asks for the sum, average, maximum, or count of a metric, it should be placed in a single subtask and not split into multiple subtasks. If the user's question is a trend or ranking question, it should be broken down into a subtask that queries detailed data and a summary task. If the user's original question does not contain time terms such as dates, do not arbitrarily add date or other time information when breaking down the subtasks. If the user's question is about recent years, such as the last three or five years, it should not be broken down into multiple subtasks based on time. Except for analytical questions, do not arbitrarily rewrite user questions and do not perform extended queries on user questions. LLM tools cannot be used alone to answer user questions. They must be based on text2sql tools. If a subtask needs to use data results that depend on other subtasks, use<sequentially depends on1,2,3…> Dependencies are marked. If a subtask does not require dependencies, it must be an independent task. When establishing dependencies between subtasks, carefully consider the problem and the above requirements. Do not arbitrarily establish subtask dependencies. Do not output any thought process to ensure that the subtask breakdown is correct. Use the text2sql tool to convert natural language into SQL queries and execute database queries. Use the LLM summary tool to perform semantic analysis on the query results, generating a structured summary and visual charts.

[0009] In a further embodiment, the text2sql tool includes three components: a natural language parsing submodule, an SQL generation submodule, and an execution and optimization submodule. The natural language parsing submodule, based on large-scale natural language processing technology, performs a comprehensive and in-depth analysis of the input natural language. It breaks sentences into their smallest semantic units through word segmentation, uses part-of-speech tagging to clarify the grammatical properties of each word, extracts time, entity, query field, enumeration conditions, filter conditions, grouping fields, group aggregation filter conditions, sorting operations, and item limit elements, and ensures that the extracted elements or functions support MySQL version 5.7 syntax. Based on the parsing results, the SQL generation submodule uses a neural network model and a Transformer architecture model to train data on large-scale natural language-SQL, automatically learning the complex mapping relationship between natural language and SQL statements, and generating SQL query statements that meet the user's intent. The execution and optimization submodule sends the generated SQL statements to the corresponding database for execution and obtains the execution results in a timely manner. For complex queries, it can optimize the execution plan of the SQL statement and utilize the database's indexing engine to improve query efficiency. At the same time, during the data transmission process, it ensures the integrity and accuracy of the data and avoids data loss or errors.

[0010] In a further embodiment, the LLM summary tool uses the powerful semantic understanding ability of LLM to perform in-depth analysis of the retrieved data support to extract key information, covering the extreme values, means, and sums of numerical data and high-frequency words and important categories in text data. The extracted key information is then organized into a logically coherent and fluent summary. The summary contains the core data and a brief explanation of the trends and relationships behind the data. During the generation process, technologies such as reinforcement learning may be used to continuously optimize the quality and accuracy of the summary based on user feedback.

[0011] In a further embodiment, the professional business knowledge injection unit improves the accuracy of analysis by injecting financial term definitions, injecting business rules, and injecting a synonym library. Injecting financial term definitions, such as the "two funds" in the fields of social insurance and housing provident fund usually refer to housing provident fund and enterprise annuity, and the "two funds" in the field of corporate financial management usually refer to various types of receivables and inventory. It is clear that the scope of receivables may include accounts receivable, other receivables, and prepaid accounts. Injecting business rules, such as financial indicator calculation formulas (asset-liability ratio = total liabilities / total assets × 100%), risk assessment logic, etc. Injecting a synonym library, such as the correspondence between "net profit" and "net income".

[0012] In a further embodiment, the AI ​​large model uses the Qwen-72B model to give it a multi-domain knowledge reserve through pre-training data of more than 3 trillion tokens, which can understand and handle complex financial problems and cope with multi-language scenarios in global financial business. The AI ​​large model uses the Qwen-72B model to accurately interpret the logical relationship between financial policies and reports in Chinese and English downstream evaluation tasks. The 32k context length is conducive to long text analysis and avoids information omissions. In terms of application convenience, it is easy to call and can be loaded through simple Python code. It supports a variety of hardware configurations and precision settings, and can adapt to the diverse hardware environments of financial departments, improve resource utilization and reduce costs.

[0013] In a further embodiment, the result presentation and feedback unit can automatically complete the dimension attribution, indicator attribution and change attribution of core data indicators. The large model can accurately identify indicator-related dimensions, such as time, space, business and other dimensions, clarify their impact on indicators, and deeply explore the internal connections between indicators, analyze the contribution of related indicators to core indicators, and quickly dig out key driving factors when indicators change. To ensure the accuracy of the results, the large model needs to be manually reviewed after automatic matching; for the displayed answers, users can easily trace back to their original data sources and operation logic in the database to ensure the accuracy and reliability of the data and enhance the user's trust in the analysis results. The result presentation and feedback unit cooperates to build a complete user satisfaction feedback mechanism. After obtaining the query results and analysis, users can evaluate and feedback the accuracy, completeness, practicality and other aspects of the answers based on their own experience, and collect user opinions in a timely manner, gain in-depth insights into changes in user needs, and optimize algorithms and models in a targeted manner to continuously improve user experience while continuously improving the accuracy of analysis results.

[0014] Beneficial effects: The present invention relates to an intelligent conversational analysis system for financial data based on an AI large model, which has the following beneficial effects:

[0015] 1. Improve data analysis efficiency: Through automated data preprocessing and intelligent conversational analysis, the time and workload of manual operations are reduced, and users' query needs can be quickly responded to, thereby improving the efficiency of financial data analysis.

[0016] 2. Enhanced analysis accuracy: By leveraging large AI models and injecting professional business knowledge, it can accurately understand users' natural language questions and generate accurate analysis results. At the same time, through result attribution and manual review mechanisms, the accuracy of analysis results is further guaranteed.

[0017] 3. Lower the threshold for use: It supports users to query in natural language without the need for professional financial knowledge and complex operating skills, making it easy for non-professionals to perform financial data analysis.

[0018] 4. Improve user experience: Provide conversational answers and graphical displays to make analysis results more intuitive and easy to understand. At the same time, a comprehensive user satisfaction feedback mechanism can respond to user needs in a timely manner, continuously optimize system performance, and improve user experience.

[0019] 5. Adaptability to multiple scenarios: The Qwen-72B model's multi-domain knowledge reserves and multi-language processing capabilities can meet the various scenario requirements in global financial business. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a system structure diagram of the present invention. DETAILED DESCRIPTION

[0021] In the following description, numerous specific details are provided to provide a more thorough understanding of the present invention. However, it will be apparent to those skilled in the art that the present invention may be practiced without one or more of these details. In other instances, certain technical features well known in the art are not described to avoid confusion with the present invention.

[0022] The AI ​​big model-based intelligent conversational analysis system for financial data involved in the present invention includes two major modules: database pre-processing process and financial data intelligent conversational analysis process, as follows.

[0023] 1. Database preprocessing process

[0024] When interacting with large models, database preprocessing is crucial. It improves the accuracy of large model understanding by standardizing data formats, cleaning erroneous and missing values, and adding semantic annotations and metadata. It enhances the analytical capabilities of large models through data conversion, feature engineering, aggregation, and hierarchical construction. It improves interaction efficiency by eliminating noise and redundancy, optimizing layout, and preparing indexes. Furthermore, it desensitizes sensitive data and filters data based on user permissions to ensure the security of large model applications. Database preprocessing primarily includes the following key steps:

[0025] Step 1: When connecting to a financial database, the system provides flexible database connection configuration options. Select the appropriate database type and then accurately enter the database's IP address, port number, username, password, URL, and database name. Once this information is entered, click "Test Connection." If the connection is successful, the system will automatically trigger database data synchronization.

[0026] At the same time, leveraging powerful big model technology, the system intelligently and automatically generates detailed table descriptions based on synchronized data, providing foundational information for subsequent automated analysis processes. Furthermore, when analyzing financial data, analysts can use the table descriptions generated by the big model to quickly determine which tables are relevant to the analysis objective and the potential impact of each field on the analysis, thereby improving the speed and accuracy of analysis. Furthermore, this module allows for convenient viewing of field details for each table and intuitive browsing of data instances under corresponding fields, providing comprehensive convenience for in-depth understanding and analysis of financial data.

[0027] Step 2: Data modeling is like building a blueprint for a data building, establishing a clear and orderly structure for massive and complex data. By defining data tables, fields, and their relationships, scattered data can be integrated into a reasonable framework, making data storage and management more efficient and facilitating quick location and retrieval of required information. At the same time, ensure the consistency and standardization of the data. During the modeling process, clarify the data type, format, value range, and other standards of each field to avoid data entry errors and inconsistencies. For example, the date field uniformly adopts the "YYYY-MM-DD" format, and the gender field value is limited to "male" and "female", laying the foundation for the accuracy and reliability of the data. In the data processing process, data modeling is a key link, which covers multiple core steps.

[0028] 1. Data table configuration. Carefully plan and set various parameters for the data table, and determine the table structure framework based on business needs, including determining the number of data tables to be created, the name of each table, and the overall layout, to ensure that the data table can accurately carry and organize relevant data.

[0029] 2. Metadata management. Comprehensively manage metadata and record all background information about the data, such as its source, meaning, owner, and update frequency. Effective metadata management provides a clear "archive" for the data, facilitating subsequent understanding, maintenance, and use.

[0030] 3. Field configuration. Carefully define and configure each field in the data table, specifying each field's name, data type (e.g., integer, string, date), length limit, whether it can be left empty, and other attributes. Field configuration must closely align with business logic to ensure that each field accurately reflects the corresponding data characteristics.

[0031] 4. Table association configuration. Thoroughly analyze the inherent connections between data and establish associations between different data tables based on business rules. Whether it's a one-to-one, one-to-many, or many-to-many relationship, precise configuration is required to ensure accurate association and integration between different tables, laying the foundation for comprehensive data analysis.

[0032] 5. Indicator Management. Systematically plan and control key indicators, clarifying their definitions, calculation methods, data sources, and applicable scenarios. Indicator management aims to extract data metrics that reflect the core business status and provide strong data support for business decision-making.

[0033] 6. Dimension Management. Rationally plan and manage data dimensions, determining the perspectives from which to classify and analyze data, such as time, region, and product. Through scientific dimension management, data can be observed and analyzed from multiple perspectives, unlocking richer information value.

[0034] Step 3: After data modeling, the data is stored in the vector database, and a data index module is further constructed. This is of great significance for improving data query efficiency and application value.

[0035] Each data record is mapped into a vector based on the characteristics of structured data. For example, in financial data modeling, information such as the amount, time, and transaction type of each transaction record is converted into a vector using a specific algorithm. The converted vector data is then stored in the Milvus vector database. During storage, configuration is required based on the characteristics of the vector database. Multiple index types are supported, and appropriate index types, such as IVF-FLAT and HNSW, can be selected based on data scale and query requirements to optimize storage and query performance. Furthermore, relevant metadata, such as the data source and creation time, is added to each vector data to facilitate subsequent data management and query filtering.

[0036] 2. Intelligent Conversational Analysis Process of Financial Data

[0037] Step 1: User question input. Users ask financial questions using natural language, such as "What was the company's net profit growth rate over the past three years?" and "Which department has the highest expenses?"

[0038] Step 2: Question Understanding and Parsing. The big model understands and parses the user's natural language question, identifying key information such as timeframe, financial indicators, and analysis dimensions. For example, for the question "What was the company's net profit growth rate over the past three years?" the big model extracts key information such as "company," "past three years," and "net profit growth rate."

[0039] Step 3: Based on the key information from the big model, invoke big model tools to query data, generate optimal answers, and present them graphically. The big model tools involved primarily include Planning Agent, text2sql, and LLM Summary.

[0040] (1) The Planning Agent is a programmatic LLM prompt structure that represents robot plans as Python programs. LLM generates contextual robot task plans by leveraging programming language constructs and LLM training on a code corpus. Prompts are constructed using import statements, a list of goals, and example tasks. Assertions are also included as an environmental feedback mechanism to ensure preconditions are met and to recover from errors.

[0041] (2) text2sqlTool, a general database query tool that supports simple queries, aggregation, grouping, sorting, multi-table joins, nested queries and other complex query scenarios.

[0042] Natural language processing capabilities: Its large-scale model-based natural language processing technology enables comprehensive and in-depth analysis of input natural language. Sentences are broken down into their smallest semantic units through word segmentation, while part-of-speech tagging clarifies the grammatical attributes of each word, such as noun, verb, or adjective, laying the foundation for subsequent semantic analysis. For example, for the sentence "Statistics of sales of electronic products in North China in 2023," the tool can accurately identify "statistics" as a verb, and "2023," "North China," "electronic products," and "sales" as different types of nouns. Natural language optimization can be performed through prompt word engineering.

[0043] (3) The LLM summary tool leverages LLM's powerful semantic understanding capabilities to perform in-depth analysis of retrieved data to extract key information. This includes extreme values, means, and sums for numerical data, as well as high-frequency words and important categories in textual data. For example, it can quickly identify the products with the highest sales and the regions with the fastest sales growth from sales data. This extracted key information is then organized into a logically coherent and fluent summary. The summary not only contains the core data but also briefly explains the underlying trends and relationships. For example, "In 2023, the electronics department contributed the most to sales, accounting for 40% of total sales, followed by the home furnishings department at 30%. Compared to the previous year, the electronics department's performance increased by 15%, mainly due to new product launches." Specifically, the retrieved data returned in structured form is further processed by LLM. Key information is identified through statistical analysis and semantic understanding, and summaries are generated based on natural language generation rules to align with human reading habits. During the generation process, techniques such as reinforcement learning may be employed to continuously optimize the quality and accuracy of the summary based on user feedback.

[0044] Step 4: Inject professional business knowledge to improve analysis accuracy. Although large models demonstrate certain analytical capabilities in many analytical scenarios thanks to their powerful language processing capabilities and the advantage of learning from massive amounts of data, in practice, when faced with complex and ever-changing business scenarios in the financial field, the system's answers may still be inaccurate. In this case, injecting professional business knowledge becomes the key to improving analysis accuracy and optimizing system performance.

[0045] Professional business knowledge encompasses multiple key aspects in the financial field. First, fundamental and core financial concepts and terminology not only have precise definitions, but also strict requirements for their application and interpretation in different scenarios. For example, the concept of "two funds" ("repos") is interpreted in multiple ways by the big model. In the social insurance and housing provident fund sectors, these typically refer to housing provident funds and enterprise annuities. In corporate financial management, these typically refer to various types of receivables and inventory. The scope of receivables in this context is more fluid and may include accounts receivable, other receivables, and prepaid expenses. This can lead to discrepancies in the big model's understanding. To ensure that the system accurately understands the "two funds" when processing relevant questions and provides correct answers that align with actual business logic, it is necessary to pre-inject relevant professional terminology definitions into the big model to help it accurately understand user intent and find precise answers.

[0046] Step 5: Leveraging the large model, the system automatically completes dimension attribution, indicator attribution, and change attribution for core data indicators. Data attribution configuration is crucial in data management and analysis.

[0047] The big model accurately identifies dimensions related to indicators, such as time, space, and business, and clearly defines their impact on them. It also deeply explores the inherent connections between indicators and analyzes the contribution of related indicators to core indicators. When indicators fluctuate, it quickly identifies the key drivers. To ensure accurate results, the big model automatically matches data and requires manual review. Professionals draw on their experience to review the results and correct any discrepancies. When used by users, the system not only supports querying indicators and dimensions but also displays the reasons for data fluctuations, providing strong data support for user decision-making.

[0048] Step 6: In digital decision-making scenarios, users can initiate data queries and in-depth analysis requests at any time based on real-time needs. The system will respond quickly, presenting accurate query results instantly and providing detailed attribution analysis. This attribution analysis not only deeply analyzes the causes behind the data but also clearly outlines the connections between various factors, helping users gain comprehensive insight into the essence of the data.

[0049] At the same time, the system provides a powerful database traceability function. For the displayed answers, users can easily trace back to the original data source and calculation logic in the database, ensuring the accuracy and reliability of the data and enhancing users' trust in the analysis results.

[0050] To further enhance service quality, the system has built a comprehensive user satisfaction feedback mechanism. After obtaining query results and undergoing analysis, users can provide feedback on the accuracy, completeness, and practicality of the answers based on their own experience. This closed-loop feedback mechanism enables the system to promptly collect user feedback and gain in-depth insights into evolving user needs, enabling targeted algorithm and model optimization. This continuously improves the user experience while also continuously enhancing the accuracy of analysis results, providing users with higher-quality and more efficient data services.

[0051] The advantages of the present invention are as follows:

[0052] In the financial data analysis field, AI-based big model solutions offer significant advantages over rule-based and simple machine learning model-based fusion solutions, as shown below:

[0053] In terms of flexibility, rule-based systems and simple machine learning models have limited adaptability and struggle to adapt to complex and ever-changing financial scenarios. Large models, on the other hand, leverage their powerful capabilities to quickly understand and adapt to various scenarios, providing comprehensive support.

[0054] In terms of cost, the former has high development and maintenance costs, requiring significant manpower and time for rule development, data processing, and model training and adjustment. The large-scale model solution only requires quick data configuration and tuning prompts, reducing initial work and costs.

[0055] In terms of development cycles, fusion solutions based on rules and simple machine learning models are cumbersome and prone to extended development cycles due to process issues. Large-scale model solutions, once connected to data sources, can quickly build systems through data configuration and prompt word optimization, significantly shortening development cycles and helping companies quickly realize the value of data analysis.

[0056] In a further preferred embodiment, the specific implementation of each module in the present invention is as follows:

[0057] (1) Implementation of database pre-processing module

[0058] Data table configuration: Based on the specific needs of financial business, determine the data tables that need to be created, such as income statement, cost statement, balance sheet, etc., name each table, and plan the overall layout of the table.

[0059] Metadata management: Record in detail the data source, meaning, owner, and update frequency of each table and each field. For example, the data in the income table comes from the sales system and is updated daily.

[0060] Field configuration: Define each field in the data table in detail. For example, for the "income amount" field in the income table, define its data type as numeric, with a length limit of 10 decimal places and no empty space allowed.

[0061] Table association configuration: Analyze the relationships between different data tables and establish associations. For example, a one-to-many association can be established between the income table and the customer table using the customer ID.

[0062] Indicator management: Clarify the definition, calculation method, data source, and applicable scenarios of key financial indicators. For example, net profit, defined as revenue minus costs and expenses, is derived from the income statement and cost statement and is suitable for assessing a company's profitability.

[0063] Dimension management: Classify and analyze data based on time, region, product, and other dimensions. For example, using the time dimension, you can analyze monthly, quarterly, and annual changes in financial data.

[0064] (2) Implementation of the Intelligent Conversational Analysis Module for Financial Data

[0065] User question reception and understanding: The system receives natural language questions input by users, such as "What was the sales revenue of each region in the first quarter of this year?"

[0066] Planning Agent task planning: Decompose the user problem into subtasks. According to the above rules, the problem can be decomposed into a subtask of querying detailed data. Call the text2sql tool to convert the problem into an SQL query statement.

[0067] The text2sql tool processes:

[0068] The natural language parsing submodule parses user questions and extracts elements such as time (the first quarter of this year), entities (various regions), and query fields (sales revenue).

[0069] The SQL generation submodule generates SQL query statements that meet the user's intent based on the parsing results, such as "SELECT region, SUM (sales_revenue) FROM sales_table WHERE quarter = 'Q1' AND year = 'this year' GROUP BY region;"

[0070] The execution and optimization submodule sends the generated SQL statements to the database for execution and optimizes the query plan to improve query efficiency.

[0071] LLM summary tool processing: Perform semantic analysis on query results, extract key information such as the total sales revenue of each region and the region with the highest sales revenue, and generate structured summaries and visual charts, such as bar charts showing sales revenue comparisons across regions.

[0072] (3) Implementation of other modules

[0073] Vector database storage submodule: converts the amount, time, transaction type and other information of financial transaction records into vectors, stores them in the vector database Milvus, and adds relevant metadata, such as the data source is the sales system and the creation time is the current time.

[0074] Professional business knowledge injection unit: inject financial term definitions, business rules and synonym libraries, such as clarifying the meaning of "two funds" in different fields in the system, adding the calculation formula for the debt-to-asset ratio, etc.

[0075] Results Presentation and Feedback Unit: The system automatically performs attribution analysis of core data indicators, such as analyzing the reasons for regional sales revenue differences. Users can trace the original data source and operational logic of the answer. Users can also provide feedback on the query results, and the system optimizes algorithms and models based on user feedback.

[0076] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes may be made to it in form and detail without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. An intelligent conversational analysis system for financial data based on an AI big model, characterized by include: Database pre-processing module and financial data intelligent conversational analysis module; The database pre-processing module is used to format financial data, clean up erroneous and missing values, perform semantic annotation, add metadata, perform desensitization filtering, and build data indexes; the financial data intelligent conversational analysis module is used to receive users' natural language questions, understand and parse the questions, call the AI ​​big model and tool set to query and analyze data, generate conversational answers, and support data attribution and graphical display; among them, the AI ​​big model is the Qwen-72B model, and the tool set includes PlanningAgent, text2sql, and LLM summary tools.

2. The intelligent conversational analysis system for financial data based on AI big model according to claim 1 is characterized in that: The database pre-processing module includes: The data table configuration unit determines the number of data tables to be created, the name of each table, and the overall layout based on business needs; The metadata management unit comprehensively coordinates metadata and records in detail background information such as the source, meaning, owner, and update frequency of the data; The field configuration unit carefully defines and sets each field in the data table, clarifying the name, data type, length limit, and whether it can be empty or not. The table association configuration unit deeply analyzes the inherent connections between data and establishes one-to-one, one-to-many, or many-to-many associations between different data tables based on business rules; The indicator management unit systematically plans and controls key indicators, clarifying the definition, calculation method, data source, and applicable scenarios of the indicators; The dimension management unit rationally plans and manages the dimensions of data, and determines how to classify and analyze data from perspectives such as time dimension, regional dimension, and product dimension.

3. The intelligent conversational analysis system for financial data based on AI big model according to claim 1 is characterized in that: The vector database storage submodule specifically performs the following operations: Map each data record into a vector based on the characteristics of the structured data; The converted vector data is stored in the vector database Milvus. During the storage process, it is configured according to the characteristics of the vector database. It supports multiple index types and the appropriate index type is selected according to the data scale and query requirements. Add relevant metadata to each vector data to facilitate subsequent data management and query screening.

4. The intelligent conversational analysis system for financial data based on an AI big model according to claim 1 is characterized by: The PlanningAgent generates a task plan through the following steps: Decompose user questions into subtasks and define the dependencies between subtasks.

5. The intelligent conversational analysis system for financial data based on AI big model according to claim 1 is characterized in that: The text2sql tool includes: The natural language parsing submodule, based on large-scale natural language processing technology, conducts a comprehensive and in-depth analysis of the input natural language. It breaks sentences into their smallest semantic units through word segmentation and uses part-of-speech tagging to clarify the grammatical properties of each word. It then extracts elements such as time, entities, query fields, enumeration conditions, filter conditions, grouping fields, group aggregation filter conditions, sorting operations, and item limit. It also ensures that the extracted elements or functions support the MySQL 5.7 syntax. The SQL generation submodule uses a neural network model based on the Transformer architecture to train large-scale natural language-SQL data based on the parsing results. It automatically learns the complex mapping relationship between natural language and SQL statements and generates SQL query statements that meet user intent. The execution and optimization submodule sends the generated SQL statements to the corresponding database for execution and obtains the execution results in a timely manner. For complex queries, it can optimize the execution plan of the SQL statements and utilize the database's indexing features to improve query efficiency. At the same time, during the data transmission process, it ensures the integrity and accuracy of the data to avoid data loss or errors.

6. The intelligent conversational analysis system for financial data based on an AI big model according to claim 1 is characterized by: Leveraging LLM's powerful semantic understanding capabilities, the LLM summary tool conducts in-depth analysis of retrieved data to extract key information, including extreme values, means, and sums for numerical data, as well as high-frequency words and important categories for textual data. It then organizes this extracted key information into a logically coherent, fluent summary that includes the core data and a brief explanation of the trends and relationships behind the data.

7. The intelligent conversational analysis system for financial data based on an AI big model according to claim 1 is characterized by: The professional business knowledge injection unit improves analysis accuracy by injecting financial term definitions, injecting business rules, and injecting a synonym library.

8. The intelligent conversational analysis system for financial data based on an AI big model according to claim 1 is characterized by: The AI ​​model uses the Qwen-72B model, which is pre-trained with over 3 trillion tokens of data and endowed with multi-domain knowledge. This allows it to understand and process complex financial problems and cope with multilingual scenarios in global financial operations. The AI ​​large model uses the Qwen-72B model to accurately interpret financial policies and report logical relationships in Chinese and English downstream evaluation tasks. The 32k context length is conducive to long text analysis and avoids information omissions.

9. The intelligent conversational analysis system for financial data based on an AI big model according to claim 1 is characterized by: The result presentation and feedback unit can automatically complete the dimension attribution, indicator attribution and change attribution of core data indicators. The large model can accurately identify the indicator-related dimensions, deeply explore the internal connections between indicators, analyze the contribution of related indicators to core indicators, and quickly discover the key driving factors when indicator changes occur. To ensure the accuracy of the results, manual review is required after the large model automatically matches. For the displayed answers, users can easily trace back to the original data source and operation logic in the database, ensuring the accuracy and reliability of the data and enhancing users' trust in the analysis results. The result presentation and feedback unit cooperate to build a complete user satisfaction feedback mechanism. After obtaining query results and analysis, users can evaluate and feedback on the accuracy, completeness, practicality and other aspects of the answers based on their own experience, collect user opinions in a timely manner, gain in-depth insights into changes in user needs, and optimize algorithms and models in a targeted manner to continuously improve user experience and the accuracy of analysis results.

Images