Intelligent question and answer method and system, electronic equipment and storage medium

Abstract

The invention relates to the technical field of artificial intelligence, and provides an intelligent question answering method and system, electronic equipment and a storage medium, and the method comprises the steps: obtaining an original financial question input by a user, carrying out the intention recognition of the original financial question, and obtaining a question type corresponding to the original financial question; based on the question type, determining a target thinking model matched with the question type from a preset thinking model library, and according to the target thinking model, expanding the original financial question to generate an expanded financial question; based on the expansion writing financial question and the question type, a retrieval strategy matched with the question type is executed, and a retrieval result is obtained; and based on the retrieval result and the target thinking model, generating response information corresponding to the output template of the target thinking model. According to the method, the thinking model matched with the question type is introduced to guide question expanding and writing and knowledge retrieval, so that targeted and structured processing of different financial questions is realized, and the specialty and accuracy of response results are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent question-answering method, system, electronic device, and storage medium for use in the financial field. Background Technology

[0002] In recent years, artificial intelligence technologies, represented by Large Language Models (LLMs), have been increasingly widely applied in the financial industry. Utilizing AI to build intelligent question-answering systems to assist financial professionals in information acquisition, analysis, and decision-making has become an important trend in the industry.

[0003] However, in practical applications, it has been found that existing financial industry question-answering systems largely rely on traditional retrieval methods or large, single-pattern models, leading to certain shortcomings when dealing with complex financial question-answering scenarios. Specifically, existing technologies typically employ a general, fixed processing model to handle all types of financial questions. This approach ignores the significant differences in the professionalism, complexity, and intent of different financial questions, resulting in system-generated answers lacking depth or precision. Furthermore, existing technologies often directly use users' colloquial questions to retrieve knowledge bases. This method struggles to accurately match specialized knowledge content, leading to poor relevance of the retrieved information and consequently reducing the accuracy of the final generated answer. Summary of the Invention

[0004] This invention provides an intelligent question-answering method, system, electronic device, and storage medium to address the shortcomings of existing technologies in handling financial issues, which suffer from insufficient professionalism and low accuracy due to a single processing mode and inaccurate knowledge retrieval.

[0005] This invention provides an intelligent question-answering method, comprising: Obtain the original financial question input by the user, and perform intent recognition on the original financial question to obtain the question type corresponding to the original financial question; Based on the problem type, a target thinking pattern matching the problem type is determined from a preset thinking pattern library, and the original financial problem is expanded according to the target thinking pattern to generate an expanded financial problem. Based on the expanded financial question and the question type, a search strategy adapted to the question type is executed to obtain search results; Based on the search results and the target thinking pattern, response information corresponding to the output template of the target thinking pattern is generated.

[0006] According to an intelligent question-answering method provided by the present invention, the question types include at least one of factual questions, analytical questions, operational questions, creative questions, and metacognitive questions.

[0007] According to an intelligent question-answering method provided by the present invention, the step of expanding the original financial question according to the target thinking pattern to generate an expanded financial question includes: When the problem type is an analytical problem, the original financial problem is expanded from multiple preset analytical dimensions based on the main thinking mode to generate the first expanded content; Based on the auxiliary thinking mode, the original financial problem is expanded from the dimension of factual elements to generate the second expanded content; Based on the first expanded content and the second expanded content, the expanded financial issue is determined.

[0008] According to an intelligent question-answering method provided by the present invention, the preset analysis dimensions include at least one of background dimension, historical data dimension, case dimension, alternative solution dimension and target dimension.

[0009] According to an intelligent question-answering method provided by the present invention, the step of expanding the original financial question according to the target thinking pattern to generate an expanded financial question includes: When the problem type is an operational problem, the target task contained in the original financial problem is broken down into multiple basic units, and corresponding execution tools are matched for the multiple basic units. The multiple basic units and corresponding execution tools are serialized into step instructions, and the step instructions are used as the expanded financial problem.

[0010] According to an intelligent question-answering method provided by the present invention, the step of executing a retrieval strategy adapted to the question type based on the expanded financial question and the question type to obtain retrieval results includes: Based on the expanded financial question and the question type, a retrieval strategy adapted to the question type is executed to obtain multiple candidate results; The multiple candidate results are scored based on a preset scoring model to obtain the scoring results. The scoring model includes three scoring dimensions: rule matching degree, business compliance, and historical association frequency. Based on the scoring results, the multiple candidate results are sorted, and the search results are determined according to the sorted results.

[0011] According to an intelligent question-answering method provided by the present invention, the step of generating response information corresponding to the output template of the target thinking pattern based on the search results and the target thinking pattern includes: Based on the target thinking pattern, a matching response template is determined from the preset output template library; The search results are filled into the response template to generate the response information.

[0012] The present invention also provides an intelligent question-answering system, comprising: The intent recognition module is used to obtain the original financial question input by the user, and to perform intent recognition on the original financial question to obtain the question type corresponding to the original financial question; The matching and expansion module is used to determine the target thinking pattern that matches the question type from the preset thinking pattern library based on the question type, and expand the original financial question according to the target thinking pattern to generate an expanded financial question. The retrieval execution module is used to execute a retrieval strategy adapted to the question type based on the expanded financial question and the question type, and obtain retrieval results; The response output module is used to generate response information corresponding to the output template of the target thinking pattern based on the search results and the target thinking pattern.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the intelligent question-answering method as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent question-answering method as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the intelligent question-answering method as described above.

[0016] The intelligent question-answering method, system, electronic device, and storage medium provided by this invention achieve differentiated and targeted processing of different types of financial questions by identifying the intent of the original financial question and matching it with the corresponding target thinking pattern. This overcomes the shortcomings of traditional processing models, which are limited by their singularity and inability to adapt to complex financial scenarios. By expanding the original financial question using this target thinking pattern, more structured and informative expanded financial questions are generated. Adaptive retrieval strategies are then executed based on the question type. This approach transforms vague and brief user input into precise retrieval instructions, thereby efficiently obtaining highly relevant searches and improving the accuracy and effectiveness of the retrieval. Furthermore, generating structured response information based on this target thinking pattern and precise retrieval results ensures that the final output is accurate, logically clear, and enhances the professionalism and accuracy of financial question answering. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the intelligent question-answering method provided by the present invention; Figure 2 This is one of the flowcharts illustrating the dynamic expansion of the problem provided by this invention; Figure 3 This is the second flowchart illustrating the dynamic expansion of the problem provided by this invention; Figure 4 This is a schematic diagram of the retrieval and execution process based on expanded financial questions provided by the present invention; Figure 5 This is a schematic diagram of the process for generating structured response information provided by the present invention; Figure 6 This is a schematic diagram of the intelligent question-answering system provided by the present invention; Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] With the rapid development of artificial intelligence technology, large language models (LLMs) have demonstrated powerful capabilities in natural language processing tasks and are being gradually applied to various vertical fields, including finance. In the financial industry, using artificial intelligence to build intelligent question-answering systems to assist professionals in market analysis, compliance inquiries, product design, and risk assessment has become an important means of improving work efficiency and decision-making quality.

[0021] Currently, intelligent question-answering technology in the financial field mainly relies on traditional search engines or dialogue systems based on general large-scale language models, which leads to certain shortcomings when dealing with complex financial question-answering scenarios.

[0022] First, the rigid response model fails to meet the professional and diverse needs of the financial sector. Most existing question-and-answer systems employ a universal, single dialogue model to address all types of financial questions. However, the types of questions in the financial field are extremely diverse, including market analysis, compliance consulting, and product design. Using a single response model to handle these vastly different questions often results in answers lacking depth or precision in professional responses.

[0023] Secondly, relying on static prompting engineering cannot dynamically adapt to complex financial scenarios. Existing technologies typically use fixed, pre-set prompt templates to interact with large language models. This static prompting engineering cannot adaptively adjust to the specific intent and dynamic needs of financial questions. For example, when a user needs to analyze credit risk, a high-quality answer requires the system to proactively integrate multi-dimensional information such as macroeconomic data, industry policies, and historical default rates; while when a user needs to design financial products, a creative and inspiring framework is needed to guide them. Fixed templates clearly cannot meet these scenario-based, in-depth analytical needs, thus limiting the model's capabilities and affecting the quality and professionalism of the final generated content.

[0024] Furthermore, the low efficiency and accuracy of retrieval make it difficult to effectively utilize the massive financial expertise base. Decision-making and analysis in the financial field heavily rely on expert knowledge bases, such as regulatory documents, market research reports, customer data, and internal case studies. Current systems typically use short, colloquial questions entered by users as search terms when performing knowledge retrieval. Due to the arbitrariness of user questions and the ambiguity of financial terminology, this approach leads to vague keywords, resulting in poorly relevant search results that often include a large amount of irrelevant historical data or customer cases. This not only reduces retrieval efficiency but also seriously affects the accuracy and reliability of subsequent question-and-answer generation.

[0025] To address these shortcomings, this invention provides an intelligent question-answering method and system applicable to the financial field. It should be noted that the intelligent question-answering method provided by this invention can be executed by one or more electronic devices. This electronic device can be a standalone hardware entity, such as a server, workstation, or personal computer; it can also be a server cluster, or a distributed computing system or cloud computing platform composed of multiple servers. For example, the executing entity can be a backend server deployed with an LLM (Limited Language Management System) that provides intelligent question-answering services to frontend applications through an Application Programming Interface (API). These frontend applications can be banking apps, financial analysis software, intelligent customer service systems, etc.

[0026] Figure 1This is a flowchart illustrating the intelligent question-answering method provided by the present invention, as follows: Figure 1 As shown, the method includes: Step S10: Obtain the original financial question input by the user, and perform intent recognition on the original financial question to obtain the question type corresponding to the original financial question.

[0027] Specifically, a raw financial question refers to a finance-related inquiry entered by the user through an interactive interface in natural language. The input format can vary, including entering a text string through a text box, inputting voice through a microphone, or receiving a request from another program via an API interface. Here, when voice is input through a microphone, the system backend will first perform speech-to-text conversion. It should be understood that raw financial questions entered by users are usually conversational, concise, and sometimes even ambiguous; for example, a user might simply enter "latest LPR interest rate" or "consumer credit risk analysis."

[0028] After obtaining the original financial question, the system needs to perform intent recognition to understand which category the user truly wants to solve belongs to. This intent recognition can be performed by a pre-trained classification model. This model can be fine-tuned based on a general text classification model using a large amount of labeled data in the financial field, thereby accurately identifying the specific intent of questions in the financial field.

[0029] The result of intent recognition is the question type. The question type categorizes the essential attributes of a financial question and serves as the basis for subsequent dynamic routing. In specific scenarios of this invention, question types may include, but are not limited to, factual questions, analytical questions, operational questions, creative questions, and metacognitive questions. For example, for the original financial question "2024 personal housing loan interest rate policy," the system will identify it as a factual question. As another example, for the original financial question "Our company wants to launch a pocket money investment product; what are the risks, and how should it be designed?", the system will use natural language processing technology to accurately identify the core intent of the question. It will find that "what are the risks" is an analytical question requiring in-depth and rigorous analysis, while "how should it be designed" is a creative question requiring divergent thinking and specific solutions. Therefore, the system will identify the question type as either an analytical question or a creative question.

[0030] Understandably, the question types include at least one of the following: factual questions, analytical questions, operational questions, creative questions, and metacognitive questions. Factual questions aim to obtain objective and clear factual information and typically have one or more standard answers. For example, when querying specific data or policy provisions, the core requirements for these questions are accuracy and timeliness.

[0031] Analytical problems require systematic and in-depth processing of information, involving complex logical reasoning such as comparison, evaluation, attribution, and prediction. These problems typically do not have a single correct answer; rather, they require a well-reasoned analytical process and conclusions. For example, they may require multi-dimensional analysis of market trends and risks; the core requirements for these types of problems are depth, logical reasoning, and comprehensiveness.

[0032] Operational questions involve performing a specific task or inquiring about the steps to complete a task. For example, asking how to use a financial instrument or conduct a transaction. The core requirement of these questions is practicality and feasibility, and the result may be instructional text or directly calling upon a tool to complete a task.

[0033] Creative problems require the system to engage in divergent thinking and generate novel ideas or solutions. For example, a question asking for the design of a financial product is one that prioritizes innovation and inspiration.

[0034] Metacognitive questions focus on the reasoning process, information sources, or capability boundaries of the question-answering system itself. For example, questions that inquire about the model's own reasoning process have a core requirement of transparency and interpretability.

[0035] This invention, through a refined classification of financial questions, enables a deep understanding of user intent. This allows the system to initiate the most appropriate follow-up processing path for different types of questions, thereby providing more professional and relevant responses. This achieves refined routing and differentiated processing of question-and-answer strategies, improving the quality of responses.

[0036] Step S20: Based on the problem type, determine the target thinking pattern that matches the problem type from the preset thinking pattern library, and expand the original financial problem according to the target thinking pattern to generate an expanded financial problem.

[0037] Specifically, once the question type is determined, the system will initiate a dynamic routing mechanism. The core of this mechanism is that different question types will be handled using different thinking patterns to ensure the professionalism and depth of subsequent answers.

[0038] The Mind Pattern Library is a pre-built collection of various structured thinking frameworks. Each mind pattern includes a set of logical frameworks and expansion rules for analyzing problems and organizing information. For example, the library may include precise retrieval patterns for fact-finding, such as a simplified version of 5W2H; the CHEAP analysis pattern for in-depth analysis; the first principles decomposition pattern for task breakdown; the SCAMPER innovation pattern for sparking creativity; and metacognitive patterns for metacognitive questions. This mind pattern library is configurable and extensible, and can be continuously updated according to business development.

[0039] Based on the question type obtained in the previous step, the system matches and determines the most suitable target thinking mode for handling that type of question within the database. This matching relationship can be a pre-set rule or a dynamically optimized strategy through reinforcement learning or other methods. For example, the system's pre-set rules could be: when the question type is a factual question, match the precise retrieval mode; when the question type is an analytical question, match the CHEAP analysis mode; when the question type is an operational question, match the first principles decomposition mode; when the question type is a creative question, match the SCAMPER innovation mode; and when the question type is a metacognitive question, match the metacognitive mode.

[0040] Once the target thinking pattern is determined, the system will expand the original financial problem according to that pattern. The essence of expansion is to transform the user's vague and brief original problem into a more complete, clearly structured financial problem that is more suitable for efficient machine retrieval and processing. The expansion process will supplement the original problem with necessary context, constraints, and keywords based on the inherent logic of the thinking pattern.

[0041] For example, if the original financial question is "the latest LPR interest rate," its question type is identified as a factual question, and the matched target thinking mode is the precise retrieval mode. This mode requires specific elements such as time and subject. Therefore, the system will expand it to "Query the specific values ​​of the latest 1-year and 5-year loan prime rates (LPR) published by a certain bank in 2024, as well as the most recent effective date." This expanded financial question greatly reduces the ambiguity of subsequent searches. It should be understood that "a certain bank" in the above expanded financial question is merely an example and will be replaced with a specific bank name in actual applications.

[0042] The following section uses an analytical problem as an example to introduce the process of problem expansion. Figure 2 This is one of the flowcharts illustrating the dynamic expansion of the problem provided by this invention, such as... Figure 2 As shown, in step S20, expanding the original financial problem according to the target thinking pattern to generate an expanded financial problem includes: Step S21a: When the problem type is an analytical problem, the original financial problem is expanded from multiple preset analytical dimensions based on the main thinking mode to generate the first expanded content. The preset analytical dimensions include at least one of the following: background dimension, historical data dimension, case dimension, alternative solution dimension, and target dimension. Step S22a: Based on the auxiliary thinking mode, the original financial problem is expanded from the dimension of factual elements to generate the second expanded content; Step S23a: Based on the first expanded content and the second expanded content, determine the expanded financial issue.

[0043] Specifically, when the identified problem type is an analytical problem, the expansion process of the original financial problem can be implemented through the following steps: First, the system expands the original financial problem from multiple preset analytical dimensions using a master thinking mode, generating the first expanded content. Here, the master thinking mode is a macro-level, structured framework used to guide the analysis of complex problems. Its role is to ensure the breadth and depth of the analysis and avoid missing key perspectives. For example, the master thinking mode could be the CHEAP thinking mode.

[0044] Pre-defined analytical dimensions, as a specific component of the main thinking mode, provide concrete entry points for analysis. These dimensions may include, but are not limited to, background dimensions, historical data dimensions, case study dimensions, alternative solutions dimensions, and objective dimensions. Among them, the background dimension refers to the macro-environment, industry background, and regulatory policies in which the problem is analyzed. The historical data dimension refers to reviewing data performance and event evolution under similar historical circumstances. The case study dimension refers to finding and citing specific company, market, or event cases as supporting evidence or reference. The alternative solutions dimension refers to considering whether other possibilities, alternative paths, or opportunity costs exist. The objective dimension refers to clarifying the specific goals to be achieved by this analysis, such as for investment decisions, risk warning, or academic research.

[0045] Then, the system expands the original financial problem from the perspective of factual elements using a secondary thinking mode, generating a second expanded version. Here, the secondary thinking mode is a micro-level framework used to clarify the basic factual elements of the problem, such as the 5W2H analysis method. Its role is to ensure that the analysis is based on accurate and unambiguous facts.

[0046] Finally, the system will integrate the first and second expanded content to determine the final expanded financial issue.

[0047] For example, assuming the original financial question is "Analyze the risks of Credit Asset Securitization (CLO)," the system identifies it as an analytical question and initiates a collaborative expansion strategy using primary and secondary thinking modes. Based on the primary thinking mode, the system generates the first expansion content as follows: "Please conduct a systematic analysis of the risks of Credit Asset Securitization (CLO) from the perspectives of the current macroeconomic interest rate environment and corporate default rate levels (background dimension), combined with the performance of CLOs during historical financial crises (historical data dimension), and whether there are any recent notable CLO default or restructuring cases (case dimension)." Based on the secondary thinking mode (such as 5W2H), the system generates the second expansion content, which needs to specify the market and credit rating of the CLO product being analyzed, the current timeframe, and the specific risk types to focus on, such as credit risk, liquidity risk, and model risk. The system combines these two approaches, ultimately generating the expanded financial question as: "Please comprehensively analyze the credit risk, liquidity risk, and model risk faced by high-yield (B-grade) Credit Asset Securitization (CLO) products, taking into account the current macroeconomic interest rate environment and corporate default rate data of a certain market, and referring to the historical performance of financial crises, and citing recent relevant cases for illustration." This invention, through an expansion method combining primary and secondary thinking modes, systematically decomposes a general and complex analytical problem into a query that combines an analytical framework with factual elements. The primary thinking mode ensures the breadth and depth of the analysis, guaranteeing a multidimensional and comprehensive analytical perspective; the secondary thinking mode ensures the accuracy and reliability of the analytical foundation. This method provides clearer and more explicit guidance for subsequent retrieval and generation steps, thereby improving the logic, rigor, and professionalism of the final generated answer.

[0048] The following section uses an operational question as an example to introduce the process of expanding a question. Figure 3 This is the second flowchart illustrating the dynamic expansion of the problem provided by this invention, such as... Figure 3 As shown, in step S20, expanding the original financial problem according to the target thinking pattern to generate an expanded financial problem includes: Step S21b: If the problem type is an operational problem, the target task contained in the original financial problem is broken down into multiple basic units, and corresponding execution tools are matched for the multiple basic units. Step S22b: Serialize the plurality of basic units and corresponding execution tools into step instructions, and use the step instructions as the expanded financial problem.

[0049] Specifically, when the identified problem type is an operational problem, the expansion process of the original financial problem can be achieved through the following steps: First, the system needs to break down the target task contained in the original financial problem into multiple basic units. Here, the target task refers to the core operational intent parsed from the user's natural language instructions.

[0050] The process of breaking down a complex financial task into multiple basic units is essentially a task planning process. The system invokes a planning model or follows preset business process rules to decompose a macro-level, complex financial task into a series of logically independent, smaller-granular, executable subtasks, i.e., basic units. Each basic unit represents an atomic operation required to complete the overall task.

[0051] Next, the system needs to match corresponding execution tools for multiple basic units. Here, execution tools refer to predefined software modules, functions, or API interfaces within the system that can perform specific functionalities. These tools are encapsulated and registered in a tool library, each with a clear functional description, input parameter specification, and output format definition. For example, the tool library might contain user authentication APIs, account balance query APIs, and financial product list query APIs. The matching process involves the system iterating through all the basic units decomposed in the previous step and, based on the intent description of each basic unit, searching the tool library for the most functionally matching execution tool.

[0052] Finally, the system serializes multiple basic units and their corresponding execution tools into step instructions, and uses these step instructions as the final expanded financial problem. Here, serialization into step instructions means organizing the results of task breakdown and tool matching into an ordered, machine-readable structured data format. This structured data is a complete execution plan. It can be a JSON array, an XML file, or another form of instruction chain. The instructions clearly define which tool should be invoked at each step (basic unit) and the parameters required for that invocation.

[0053] The generated step instructions constitute the final expanded financial problem. It is no longer natural language text for information retrieval, but rather a task plan to be executed. This plan is then passed to step S30, where the retrieval strategy transforms into an execution strategy. A tool execution engine then calls the tools specified in the instructions one by one to complete the entire task.

[0054] For example, assuming the original financial problem is "Please generate a financial risk assessment report on listed company XX Technology", the system identifies it as an operational problem, initiates a task planning process, and breaks down the target task into the following basic units: First, obtain basic information about XX Technology, such as its stock code; second, call the data interface to obtain XX Technology's financial statements for the past three years; third, based on the financial statement data, calculate key financial ratios related to solvency, profitability, and operational efficiency; fourth, obtain the average financial ratios of the industry in which XX Technology operates as a benchmark; fifth, comprehensively compare the company's data with the industry benchmark, conduct a risk assessment, and generate structured text. Subsequently, the system matches an execution tool to each basic unit and serializes the above basic units and matched execution tools into step instructions, such as a JSON object, which becomes the expanded financial problem for this instance.

[0055] This invention transforms fuzzy natural language tasks into precise, machine-executable step instructions and combines the reasoning and planning capabilities of large language models with the stable execution capabilities of external specialized tools. This achieves a leap from knowledge engine to task agent, expanding the application boundaries of financial intelligent question-answering systems. It enables these systems to reliably and automatically complete complex financial business processes, improving business processing efficiency and accuracy.

[0056] Step S30: Based on the expanded financial question and the question type, execute a retrieval strategy adapted to the question type to obtain retrieval results.

[0057] Specifically, unlike traditional methods that directly use the original question for retrieval, this embodiment of the invention uses expanded financial questions that have been modified by thought patterns for retrieval, which helps to improve the accuracy of the retrieval.

[0058] Furthermore, the system will execute a retrieval strategy adapted to the question type. This means that the system will not use a single retrieval method, but will intelligently distribute searches based on the question type. If the question type is a factual question, the system will execute a precise matching strategy, calling the precise query interfaces of the internal knowledge base, such as the query interfaces of the policy and regulation database, product parameter database, etc.

[0059] If the problem type is analytical, the system will execute a multi-source fusion retrieval strategy, simultaneously obtaining information from multiple heterogeneous data sources. For example, it will call the Wind database to obtain market data, retrieve relevant analysis reports from the industry report library, and find similar cases from the internal case library.

[0060] If the problem type is an operational problem, the retrieval strategy can be transformed into a tool routing strategy, which directly routes the expanded problem (which may have been transformed into an API call instruction) to the corresponding tool or business system for execution.

[0061] After executing the retrieval strategy, the system will obtain the retrieval results. These results are the original materials used to answer the questions, and they can take various forms, such as structured data returned by a database, text paragraphs from a knowledge base, JSON data returned by an external API, or a collection of information from multiple sources. This embodiment of the invention does not specifically limit these forms.

[0062] Furthermore, Figure 4 This is a schematic diagram of the retrieval and execution process based on expanded financial questions provided by the present invention, such as... Figure 4 As shown, step S30 specifically includes: Step S31: Based on the expanded financial question and the question type, execute a retrieval strategy adapted to the question type to obtain multiple candidate results.

[0063] Specifically, based on the expanded question and question type, the system retrieves all potentially relevant original information fragments from one or more data sources (such as internal knowledge bases, external databases, the internet, etc.). These information fragments constitute a set of multiple candidate results. At this point, the candidate results may contain information that is not highly relevant to the question, outdated information, or even information that does not conform to business regulations.

[0064] Step S32: The multiple candidate results are scored based on a preset scoring model to obtain the scoring results. The scoring model includes three scoring dimensions: rule matching degree, business compliance, and historical correlation frequency.

[0065] Specifically, to select the best from numerous candidate results, the system employs a scoring model to quantitatively evaluate each candidate. This model is not a single-dimensional one, but rather integrates multiple key indicators to ensure the comprehensiveness and professionalism of the evaluation. Specifically, the scoring model includes at least three scoring dimensions: rule matching degree, business compliance, and historical correlation frequency.

[0066] Here, rule matching refers to the degree to which candidate results match the expanded financial question in terms of semantics and key elements. Business compliance is a unique and crucial scoring criterion in the financial field, assessing whether the content of candidate results complies with regulatory requirements and the company's internal risk control system. Historical correlation frequency refers to the system's statistical analysis of historical user behavior data, which identifies which sources or types of candidate results are more frequently adopted or receive positive user feedback when handling similar questions.

[0067] Understandably, these scoring dimensions can be assigned different weights. For example, rule matching score could have a weight of 40%, business compliance a weight of 35%, and historical association frequency a weight of 25%. After obtaining the score for each dimension, a weighted sum can be calculated based on the weight of each dimension to arrive at the final score. Furthermore, an LSTM network can be introduced to learn historical association sequences and dynamically adjust the weights of each dimension. For instance, during compliance-sensitive periods, the compliance weight can be automatically increased to 50%, thereby improving the hit rate of high-confidence results.

[0068] Step S33: Based on the scoring results, sort the multiple candidate results and determine the search results according to the sorted results.

[0069] Specifically, after calculating a comprehensive score (which can be a weighted average of three dimensions) for each candidate result, the system sorts all candidate results in descending order. After sorting, the system can adopt different strategies to determine the final search results. For example, it can directly select the result with the highest score; it can select the top-K results with the highest scores as a set; or, for analytical questions, it can merge and refine the multiple results with the highest scores to form a comprehensive search result.

[0070] In this embodiment of the invention, a multi-dimensional scoring model incorporating rule matching degree, business compliance, and historical correlation frequency is introduced to refine the reordering of candidate results from coarse recall, establishing a quality filtering and optimization mechanism. This mechanism can effectively filter out irrelevant, non-compliant, and unreliable information, and prioritize the most effective results, thereby ensuring the accuracy and compliance of the final response content.

[0071] Further, in step S31, the step of executing a retrieval strategy adapted to the question type based on the expanded financial question and the question type includes: If the question type is a factual question, the matching interface is invoked to retrieve information from a preset professional knowledge base; When the problem type is an analytical problem, a fusion search is performed on multiple heterogeneous data sources; When the problem type is an operational problem, the expanded financial problem is routed to the corresponding tool application from the program interface.

[0072] Specifically, when the system determines that the user's intent is to query a certain fact, the most efficient and reliable method is to query the internal knowledge base. This knowledge base is a pre-built and maintained, highly structured or semi-structured database. The retrieval strategy translates into a single call to a matching interface, such as a precise database query or an API call targeting a specific field. The advantages of this strategy are speed, high accuracy, and reliable source.

[0073] When users require in-depth analysis, a single data source often cannot provide a comprehensive perspective. Therefore, the retrieval strategy shifts to fusing searches across multiple heterogeneous data sources. This means the system will initiate query requests to different data sources in parallel or sequentially. Through fusion retrieval, the system can aggregate various information such as market data, macroeconomic policies, industry trends, and company announcements, providing comprehensive materials for subsequent in-depth analysis.

[0074] When a user's intent is to perform a task, the search goal is no longer to obtain information, but to complete an action. At this point, expanding the financial question itself becomes a sequence of steps. Therefore, the search strategy here is actually a routing and execution strategy. The system routes this step instruction to a tool execution engine, which parses the instruction and calls the corresponding tool application's API one by one.

[0075] In this embodiment of the invention, an intelligent resource scheduling and task distribution mechanism is achieved by customizing specific retrieval (or execution) strategies for different problem types. This avoids the inefficient practice of using a single, general retrieval method to handle all problems. This differentiated strategy design enables the system to acquire the necessary resources efficiently and appropriately, improving the processing efficiency and response quality for various types of problems.

[0076] Step S40: Based on the search results and the target thinking pattern, generate response information corresponding to the output template of the target thinking pattern.

[0077] Specifically, the output templates here are strongly linked to the target thinking patterns identified in step S20. Each thinking pattern corresponds to one or more fixed output structures. For example, the output template for the precise retrieval mode can be composed of core conclusions and data sources; the output template for the CHEAP analysis mode can be a structured report, such as presenting a brief background description, core data, case studies, conclusions, and recommendations in chapters; while the output template for the first principles decomposition mode can consist of step flowcharts, code, and tool screenshots.

[0078] The system invokes a generative model, such as a large language model, using the search results obtained in the previous step as input. The model is then instructed to organize and generate the final response information according to an output template that matches the target thought pattern. The model refines, summarizes, and polishes the scattered retrieved information, filling it into the corresponding positions in the template, ultimately forming a logically clear and structurally sound response.

[0079] For example, continuing with the LPR example above, the system retrieves an LPR value of 3.45%, sourced from a bank's announcement. During the generation process, a template for the core conclusion and data source will be applied, and the final generated response information could be: "According to Announcement No. 3 of 2024 issued by a certain bank, the 1-year Loan Prime Rate (LPR) is 3.45%." Furthermore, Figure 5 This is a schematic diagram of the process for generating structured response information provided by the present invention, such as... Figure 5 As shown, step S40 specifically includes: Step S41: Based on the target thinking pattern, determine the matching response template from the preset output template library; Step S42: Fill the response template with the search results to generate the response information.

[0080] Specifically, in step S20, the system has already matched the most suitable target mindset for the current problem. The system will then use this matching relationship to find a response template that is strongly bound to that mindset from a pre-built output template library.

[0081] Here, the output template library is a collection that stores various response format definitions. Each response template is a structured framework that not only defines the layout of the response but also includes best practices for information presentation in specific scenarios.

[0082] After obtaining the response template through matching, the system needs to populate the response template with the search results to generate the final response information. The search results here refer to the optimized and sorted search results obtained in step S30. These are high-quality information materials acquired in the preceding steps. For operational questions, the search results can also be the data or status returned after calling the tool.

[0083] The filling process is not a simple string replacement, but is accomplished by a generative model (such as a Large Language Model, LLM). The system takes the search results and the matched response template as input and issues a structured generation instruction to the generative model. This instruction requires the model to understand, refine, summarize, and generalize the search results, and then logically and clearly fill the highly summarized and refined content into the corresponding positions in the response template. During this process, the model also needs to polish the language to ensure that the final output text is fluent, professional, and conforms to the expression habits of the financial field.

[0084] The final output is a well-structured, detailed, and standardized response that can be directly presented to the user.

[0085] The method provided in this invention identifies the intent of the original financial question and matches it with a corresponding target thinking pattern, enabling differentiated and targeted processing of different types of financial questions. This overcomes the shortcomings of traditional processing models, which are limited by their singularity and inability to adapt to complex financial scenarios. By expanding the original financial question using this target thinking pattern, a more structured and informative expanded financial question is generated. Combined with the question type, an appropriate retrieval strategy is executed. This approach transforms vague and brief user input into precise retrieval instructions, thereby efficiently obtaining highly relevant searches and improving the accuracy and effectiveness of the retrieval. Furthermore, generating structured response information based on this target thinking pattern and precise retrieval results ensures that the final output is accurate, logically clear, and enhances the professionalism and accuracy of financial Q&A.

[0086] The intelligent question-answering system provided by the present invention is described below. The intelligent question-answering system described below can be referred to in correspondence with the intelligent question-answering method described above.

[0087] Based on any of the above embodiments Figure 6 This is a schematic diagram of the intelligent question-answering system provided by the present invention, as shown below. Figure 6 As shown, the system includes: The intent recognition module 610 is used to obtain the original financial question input by the user, and to perform intent recognition on the original financial question to obtain the question type corresponding to the original financial question. The matching expansion module 620 is used to determine the target thinking pattern that matches the problem type from the preset thinking pattern library based on the problem type, and expand the original financial problem according to the target thinking pattern to generate an expanded financial problem. The retrieval execution module 630 is used to execute a retrieval strategy adapted to the question type based on the expanded financial question and the question type, and obtain retrieval results; The response output module 640 is used to generate response information corresponding to the output template of the target thinking pattern based on the search results and the target thinking pattern.

[0088] The apparatus provided in this invention identifies the intent of the original financial question and matches it with a corresponding target thinking pattern, enabling differentiated and targeted processing of different types of financial questions. This overcomes the shortcomings of traditional processing models, which are limited by their singularity and inability to adapt to complex financial scenarios. By expanding the original financial question using this target thinking pattern, a more structured and informative expanded financial question is generated. Combined with the question type, an appropriate retrieval strategy is executed. This approach transforms vague and brief user input into precise retrieval instructions, thereby efficiently obtaining highly relevant searches and improving the accuracy and effectiveness of the retrieval. Furthermore, generating structured response information based on this target thinking pattern and precise retrieval results ensures that the final output content is accurate and logically clear, enhancing the professionalism and accuracy of financial Q&A.

[0089] Based on any of the above embodiments, the question type includes at least one of factual questions, analytical questions, operational questions, creative questions, and metacognitive questions.

[0090] Based on any of the above embodiments, the matching expansion module 620 is specifically used for: When the problem type is an analytical problem, the original financial problem is expanded from multiple preset analytical dimensions based on the main thinking mode to generate the first expanded content; Based on the auxiliary thinking mode, the original financial problem is expanded from the dimension of factual elements to generate the second expanded content; Based on the first expanded content and the second expanded content, the expanded financial issue is determined.

[0091] Based on any of the above embodiments, the preset analysis dimensions include at least one of the following: background dimension, historical data dimension, case dimension, alternative solution dimension, and target dimension.

[0092] Based on any of the above embodiments, the matching expansion module 620 is specifically used for: When the problem type is an operational problem, the target task contained in the original financial problem is broken down into multiple basic units, and corresponding execution tools are matched for the multiple basic units. The multiple basic units and corresponding execution tools are serialized into step instructions, and the step instructions are used as the expanded financial problem.

[0093] Based on any of the above embodiments, the retrieval execution module 630 is specifically used for: Based on the expanded financial question and the question type, a retrieval strategy adapted to the question type is executed to obtain multiple candidate results; The multiple candidate results are scored based on a preset scoring model to obtain the scoring results. The scoring model includes three scoring dimensions: rule matching degree, business compliance, and historical association frequency. Based on the scoring results, the multiple candidate results are sorted, and the search results are determined according to the sorted results.

[0094] Based on any of the above embodiments, the response output module 640 is specifically used for: Based on the target thinking pattern, a matching response template is determined from the preset output template library; The search results are filled into the response template to generate the response information.

[0095] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute an intelligent question-answering method. This method includes: acquiring an original financial question input by a user, and performing intent recognition on the original financial question to obtain the question type corresponding to the original financial question; based on the question type, determining a target thinking pattern matching the question type from a preset thinking pattern library, and expanding the original financial question according to the target thinking pattern to generate an expanded financial question; based on the expanded financial question and the question type, executing a retrieval strategy adapted to the question type to obtain retrieval results; and based on the retrieval results and the target thinking pattern, generating response information corresponding to the output template of the target thinking pattern.

[0096] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to related technologies, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0097] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the intelligent question-answering method provided by the above methods. The method includes: acquiring an original financial question input by a user, and performing intent recognition on the original financial question to obtain a question type corresponding to the original financial question; determining a target thinking pattern matching the question type from a preset thinking pattern library based on the question type, and expanding the original financial question according to the target thinking pattern to generate an expanded financial question; executing a retrieval strategy adapted to the question type based on the expanded financial question and the question type to obtain retrieval results; and generating response information corresponding to the output template of the target thinking pattern based on the retrieval results and the target thinking pattern.

[0098] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the intelligent question-answering method provided by the above methods. This method includes: acquiring an original financial question input by a user, and performing intent recognition on the original financial question to obtain a question type corresponding to the original financial question; based on the question type, determining a target thinking pattern matching the question type from a preset thinking pattern library, and expanding the original financial question according to the target thinking pattern to generate an expanded financial question; based on the expanded financial question and the question type, executing a retrieval strategy adapted to the question type to obtain retrieval results; and based on the retrieval results and the target thinking pattern, generating response information corresponding to the output template of the target thinking pattern.

[0099] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0100] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent question answering method, characterized by, The method comprises the following steps: acquiring a user-input original financial question, and performing intent recognition on the original financial question to obtain a question type corresponding to the original financial question; determining a target thinking mode matching the question type from a preset thinking mode library, and performing expansion writing on the original financial question according to the target thinking mode to generate an expanded financial question; performing a retrieval strategy adapted to the question type based on the expanded financial question and the question type to obtain a retrieval result; generating response information corresponding to an output template of the target thinking mode based on the retrieval result and the target thinking mode. 2.The intelligent question answering method of claim 1, wherein, The question type comprises at least one of a factual question, an analytical question, an operational question, a creative question and a meta-cognitive question. 3.The intelligent question answering method of claim 2, wherein, The expansion writing on the original financial question according to the target thinking mode to generate an expanded financial question comprises the following steps: in a case where the question type is an analytical question, performing expansion writing on the original financial question from a plurality of preset analytical dimensions based on a main thinking mode to generate first expansion writing content; performing expansion writing on the original financial question from a factual element dimension based on a secondary thinking mode to generate second expansion writing content; determining the expanded financial question based on the first expansion writing content and the second expansion writing content. 4.The intelligent question answering method of claim 3, wherein, The preset analytical dimensions comprise at least one of a background dimension, a historical data dimension, a case dimension, an alternative solution dimension and a target dimension. 5.The intelligent question answering method of claim 2, wherein, The expansion writing on the original financial question according to the target thinking mode to generate an expanded financial question comprises the following steps: in a case where the question type is an operational question, decomposing a target task contained in the original financial question into a plurality of basic units, and matching corresponding execution tools for the plurality of basic units; serializing the plurality of basic units and the corresponding execution tools into step instructions, and taking the step instructions as the expanded financial question. 6.The intelligent question-answering method of any one of claims 1-5, wherein, The performing of the retrieval strategy adapted to the question type based on the expanded financial question and the question type to obtain a retrieval result comprises the following steps: performing a retrieval strategy adapted to the question type based on the expanded financial question and the question type to obtain a plurality of candidate results; performing scoring on the plurality of candidate results based on a preset scoring model to obtain a scoring result, the scoring model comprising three scoring dimensions of rule matching degree, business compliance and historical association frequency; performing sorting on the plurality of candidate results based on the scoring result, and determining the retrieval result according to the sorted result. 7.The intelligent question-answering method of any one of claims 1-5, wherein, The generating of response information corresponding to an output template of the target thinking mode based on the retrieval result and the target thinking mode comprises the following steps: determining a matching response template from a preset output template library according to the target thinking mode; filling the retrieval result into the response template to generate the response information.

8. An intelligent question answering system characterized by, The method comprises the following steps: an intent recognition module is configured to acquire a user-input original financial question, and perform intent recognition on the original financial question to obtain a question type corresponding to the original financial question; The matching expansion module is configured to determine a target thinking mode matched with the question type from a preset thinking mode library based on the question type, and expand the original financial question based on the target thinking mode to generate an expanded financial question. The retrieval execution module is configured to execute a retrieval strategy adapted to the question type based on the expanded financial question and the question type to obtain a retrieval result. The answer output module is configured to generate answer information corresponding to an output template of the target thinking mode based on the retrieval result and the target thinking mode.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, The processor executes the computer program to implement the intelligent question answering method of any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the intelligent question answering method of any one of claims 1 to 7.

Images