Financial consultation response method and system

Abstract

The invention relates to the technical field of natural language processing, in particular to a financial consultation response method and system. The method comprises the following steps: receiving a financial consultation question; performing reasoning to form candidate answers; executing a structured reasoning process in parallel through a plurality of reasoning devices, wherein each reasoning device independently executes the following progressive steps: identifying and explaining financial concepts and terms involved in the financial consultation question; based on a result of the concept interpretation, decomposing the financial consultation question into a plurality of logically associated sub-questions; according to the logic sequence of the sub-questions, the current sub-question is answered only according to the answers of the previously answered sub-questions, the answers of all the sub-questions are generated in sequence, and a reasoning chain is formed; reasoning chains generated by the reasoning devices and corresponding final answers are collected, an answer pool is formed, and the answer with the highest occurrence frequency is selected from the answer pool to serve as a candidate answer; evaluating the candidate answers; and outputting candidate answers as financial consultation answers when evaluation is passed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, and in particular to a financial consultation response method and system. BACKGROUND

[0002] With the rapid development of large language model technology, large language models are increasingly widely used in various professional fields.

[0003] However, because financial decisions directly affect economic interests, any analysis errors or misleading suggestions can lead to property losses, and in the field of financial consultation, large models face more stringent accuracy requirements than general application scenarios.

[0004] However, existing large language models are prone to factual cognitive errors such as misunderstanding of financial concepts or incorrect reference to market data when generating answers based on customer financial consultation, which directly affects the reliability of subsequent reasoning. And when processing multi-step financial reasoning, the model is prone to logical jumps in the reasoning chain and confusion in the reasoning relationship, leading to contradictory analysis conclusions, so that the error reasoning is extended and amplified in the process.

[0005] Therefore, the existing technology still has some problems, resulting in inaccurate answers generated by the financial consultation large model. SUMMARY

[0006] In order to solve the above technical problems or at least partially solve the above technical problems, the present application provides a financial consultation response method and system, which can more accurately answer financial consultation questions.

[0007] In a first aspect, the present application provides a financial consultation response method, which comprises the following steps: receiving a financial consultation question; reasoning to form a candidate answer: performing a structured reasoning process in parallel by multiple reasoners, each reasoner independently performing the following progressive steps: identifying and interpreting financial concepts and terms involved in the financial consultation question; based on the results of the concept interpretation, decomposing the financial consultation question into multiple logically associated sub-questions; in accordance with the logical order of the sub-questions, only based on the answers to the previously answered sub-questions, sequentially generating the answers to each sub-question to form a reasoning chain; collecting the reasoning chains and their corresponding final answers generated by each reasoner to form an answer pool, and selecting the answer with the highest frequency of occurrence from the answer pool as the candidate answer; evaluating the candidate answer; The above reasoning and evaluation processes are repeatedly performed, and the candidate answer is output as the financial consulting answer when all evaluation conditions are met.

[0008] Optionally, after receiving the financial consulting question, before the reasoning forms the candidate answer, the following steps are further included: generating a fuzzy noise feature description according to the financial consulting question: analyzing the language features, expression modes and question structures of the financial consulting question to identify the user's implicit cognitive needs; Based on the implicit cognitive needs, a fuzzy noise feature description is generated to introduce controllable fuzziness to modulate the reasoning process; When the reasoning forms the candidate answer, each reasoner receives the fuzzy noise feature description when performing the structured reasoning process, so as to modulate the reasoning process of the reasoner through the fuzzy noise feature description.

[0009] Optionally, the fuzzy noise feature description includes: Attention allocation feature description, used to adjust the attention intensity of the reasoner to different types of information to affect the weight allocation of direct factors and indirect factors in the reasoning process; Speculative depth feature description, used to adjust the decomposition granularity and hierarchical depth of the reasoner to affect the hierarchical depth of sub-problem decomposition; Information dependency feature description, used to adjust the dependency weight of the reasoner on existing knowledge and reasoning information to affect the dependency degree of the reasoning result on existing evidence and information generated based on reasoning; Conclusion concentration feature description, used to adjust the convergence tendency of the reasoner to affect the degree of convergence of the reasoning output from multiple possibilities to a single conclusion.

[0010] Optionally, the financial consulting response method is implemented through a large language model with the following structure: Fuzzy noise identification module, used to perform the step of generating a fuzzy noise feature description according to the financial consulting question, including: Context analysis unit, used to analyze the language features, expression modes and question structures of the financial consulting question; Cognitive needs inference unit, used to identify the user's implicit cognitive needs based on the context analysis results; Noise feature generation unit, used to generate a fuzzy noise feature description based on the implicit cognitive needs; Reasoning module, used to perform the step of reasoning to form a candidate answer, including a plurality of parallel reasoners, each reasoner including: Concept explanation unit, used to identify and explain financial concepts and terms in the financial consulting question; Sub-problem decomposition unit, used to decompose the financial consulting question into logically related sub-problems; a step-by-step solving unit for generating answers to each sub-problem in a logical order of sub-problems; wherein each reasoner receives the fuzzy noise feature description to modulate the reasoning process; an answer pool statistics module for collecting the reasoning chains and final answers of each reasoner, forming an answer pool and selecting candidate answers; an evaluation module for performing the step of evaluating the candidate answers, comprising: a global logical evaluation unit for checking the logical reasonableness and factual accuracy of each reasoning step step-by-step under the premise of known reasoning chains and final answers; a consistency verification unit for re-evaluating the effectiveness of evidence in the reasoning chain and verifying the reasoning consistency by assuming that the candidate answer and the alternative answer are both true at the same time; an iteration control module for controlling the repeated execution of the reasoning and evaluation process until all evaluation steps are met.

[0011] Optionally, the specific structure of the fuzzy noise identification module comprises: a context analysis unit receiving the financial consulting question as input and processing it through the following sub-units to form a context analysis result: a lexical feature extraction sub-unit for identifying key words, emotional color and urgency expression in the financial consulting question; an expression mode analysis sub-unit for analyzing the language style, certainty degree and openness features of the financial consulting question; a question structure analysis sub-unit for identifying the complexity, hierarchy and information demand degree of the financial consulting question; outputting the context analysis result to the cognitive demand inference unit to form the implied cognitive demand; a cognitive demand inference unit receiving the context analysis result and processing it through the following sub-units: a cognitive style identification sub-unit for inferring the user's preferred analysis style and thinking mode based on the context analysis result; a demand depth evaluation sub-unit for judging the user's expectations for analysis depth, detail level and complexity based on the context analysis result; a decision-oriented analysis sub-unit for identifying the user's decision-making time window and certainty demand based on the context analysis result; outputting the implied cognitive demand to the noise feature generation unit; a noise feature generation unit receiving the implied cognitive demand and generating a fuzzy noise feature description through the following sub-units: an attention feature generation sub-unit for generating an attention allocation feature description based on the implied cognitive demand; a depth feature generation subunit configured to generate a speculative depth feature description based on the implicit cognitive requirement; a dependency feature generation subunit configured to generate an information dependency feature description based on the implicit cognitive requirement; a concentration feature generation subunit configured to generate a conclusion concentration feature description based on the implicit cognitive requirement; output the fuzzy noise feature description to the reasoning module.

[0012] Optionally, the specific structure of each reasoner is as follows: a concept explanation unit configured to receive the financial consulting question and the attention allocation feature description, and process the same through the following subunits: a financial term identification subunit configured to extract financial concepts and professional terms from the financial consulting question; a concept definition generation subunit configured to generate accurate definitions and explanations of the financial concepts and terms under the condition that the attention allocation feature description modulates the detail and focus of the concept explanation; output the concept explanation result to the subproblem decomposition unit; a subproblem decomposition unit configured to receive the financial consulting question, the concept explanation result and the speculative depth feature description, and decompose the financial consulting question into a plurality of logically associated subproblems according to the concept explanation result to form a subproblem sequence under the condition that the speculative depth feature description modulates the depth of subproblem decomposition; output the subproblem sequence to the step-by-step solving unit; a step-by-step solving unit configured to receive the subproblem sequence, the information dependency feature description and the conclusion concentration feature description, and process the same through the following subunits: a logical derivation subunit configured to generate answers to the subproblems in sequence according to the logical order of the subproblems under the condition that the information dependency feature description modulates the facts on which the logical derivation of the subproblems is based; a conclusion modulation subunit configured to modulate the openness and certainty degree of the answers according to the conclusion concentration feature description; output the reasoning chain to the answer pool statistics module.

[0013] Optionally, the specific structure of the evaluation module includes: a global logical evaluation unit configured to receive the reasoning chain corresponding to the candidate answer, and process the same through the following subunits: a reasoning step extraction subunit configured to extract each reasoning step from the reasoning chain; a global information construction subunit configured to construct a global information background based on the reasoning chain and the final answer; a step-by-step checking subunit configured to check the logical rationality and factual accuracy of each reasoning step step by step from the starting step under the premise that the global information is known; The output logic evaluates a result to the iteration control module. The consistency verification unit receives the inference chain and the answer pool information corresponding to the candidate answer, and processes the information through the following sub-units: The alternative answer selection sub-unit is configured to select an alternative answer different from the candidate answer from the answer pool. The premise construction sub-unit is configured to construct a hypothetical premise under which the candidate answer and the alternative answer are both true. The evidence effectiveness evaluation sub-unit is configured to reevaluate the effectiveness of each evidence in the original inference chain under the hypothetical premise. The evidence screening sub-unit is configured to screen out the effective evidence. The re-inference sub-unit is configured to perform re-inference based on the effective evidence. The consistency checking sub-unit is configured to check whether the result of the re-inference is consistent with the candidate answer. The consistency verification result is output to the iteration control module.

[0014] In a second aspect, the present application further provides a financial consulting system, which comprises a processor and a memory, and the memory stores at least one instruction, at least one program, a code set or an instruction set, which is loaded and executed by the processor to implement the financial consulting response method according to any one of the first aspect.

[0015] Compared with the prior art, the technical scheme provided by the present application has the following advantages: One of the beneficial effects and the working principle thereof is that: Although the traditional large language model has certain reasoning ability, in the high-risk professional field of financial consulting, the reasoning process is prone to factual errors and reasoning confusion, resulting in insufficient reasoning credibility, which may mislead the user's financial decision and cause economic losses.

[0016] The present application forms an inference chain by independently performing the progressive steps of financial concept interpretation, sub-problem decomposition and step-by-step solving by each reasoner, collects the answers of each reasoner to form an answer pool, and then verifies the candidate answer through evaluation.

[0017] The working principle thereof is that the reasoner design of the present application ensures the logical rigor of financial analysis through structured reasoning, wherein the concept interpretation step guarantees the accuracy of financial term understanding, the sub-problem decomposition step converts complex financial problems into manageable sub-problems, and the step-by-step solving step ensures the coherence of the inference chain based on logical correlation.

[0018] The financial consulting response method provided by this application improves the accuracy and credibility of financial consulting response reasoning, and can more accurately answer financial consulting questions raised by customers.

[0019] The second beneficial effect and its working principle are: Financial consulting scenarios are characterized by diversity and scenario-dependence. Many financial problems are essentially multi-solution problems, requiring multiple reasonable solutions based on different assumptions or risk preferences. For example, asset allocation issues require consideration of allocation strategies under different scenarios such as market optimism, neutrality, and pessimism. Investment timing selection requires providing corresponding suggestions based on different market cycle assumptions.

[0020] However, although the above-mentioned reasoning and evaluation methods can improve the credibility of reasoning, because the step-by-step solution link of the reasoner requires answering the current sub-problem only based on the answers to the previously solved sub-problems, and the concept interpretation step always tends to obtain standard definitions, this composite structure that strongly relies on general prior knowledge and rigid causal dependencies prevents the reasoner from performing hypothetical analysis and scenario modeling.

[0021] The structure of the above reasoners results in each reasoner following the most generally correct steps for concept interpretation, subproblem decomposition, and step-by-step autoregressive solution. Consequently, this cognitive model, which relies heavily on general prior knowledge, lacks the ability to perform contextualized reasoning based on different conditions. Consequently, even with multiple reasoners, the answers they generate always converge on a single case.

[0022] Because this structure inherently excludes multiple solutions and hypothetical analysis, it's unable to handle complex financial consulting questions that require multiple answers or conditional recommendations. It also fails to adjust the depth, focus, and openness of reasoning based on the user's cognitive state and the context of the question, reducing the accuracy and professionalism of financial consulting responses relative to user intent. In other words, this structure lacks the ability to recognize and reason that certain knowledge may be more accurate in certain specific situations.

[0023] This application adds a fuzzy noise identification step before reasoning to form candidate answers. It analyzes the linguistic features and question structure of financial consulting questions to identify users' implicit cognitive needs. This generates a fuzzy noise feature description to introduce a controllable fuzzy noise modulation reasoning process. This allows the reasoner to break through cognitive model limitations and dynamically adjust the focus, depth, information dependency, and conclusion openness of reasoning based on noise features. This approach restores the flexibility and adaptability of reasoning while maintaining logical rigor.

[0024] The working principle is that the fuzzy noise feature description can adaptively modulate the cognitive processing of the inference engine. Among them, the attention allocation feature description can make the inference engine dynamically adjust between direct factor analysis and indirect factor analysis according to the problem characteristics and user needs, the speculation depth feature description can make the inference engine adjust the hierarchical depth of sub-problem decomposition according to the user's question, the information dependence feature can make the inference engine find a balance point between priori fact verification and logical extrapolation, and the conclusion concentration feature fundamentally solves the problem of forced answer convergence, allowing the inference engine to dynamically select between providing deterministic suggestions and maintaining multiple possibilities according to the multi-solution nature of the problem and the decision needs of the user.

[0025] Therefore, by introducing controllable fuzzy noise feature description, the processing logic of the inference engine is modulated to ensure the predictability and reliability of the modulation effect, so that it can support situation analysis and multi-solution coexistence, thereby solving the limitations of the fixed reasoning structure and improving the professionalism in answering complex financial consulting problems and the accuracy of the answer relative to the user's intention. The controllable fuzzy noise on simple single-solution problems can still accurately converge on the unique solution.

[0026] Therefore, the financial consulting response method provided by the present application can more accurately answer the user's financial consulting questions. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 The flowchart of the financial consulting response method provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0028] The technical solutions in the present application will be described below with reference to the accompanying drawings.

[0029] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the present application, not all embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0030] It should be noted that in the embodiments of the present application, the financial consulting response method is actually implemented by a pre-trained large language model with specific structural arrangement and adjustment.

[0031] The pre-trained large language model can be, but is not limited to, a GPT series model, a BERT series model, a T5 model, a PaLM model, an LLaMA model, a ChatGLM model, a Baidu Wenxin Yiyan, and an Ali Tongyi Qianwen, etc. deep learning model with text understanding and generation capability. In this embodiment, the GPT series model is taken as an example for illustration.

[0032] Therefore, in the embodiments of the present application, the corresponding specific structural part of the large language model is introduced in the description of each step to facilitate understanding.

[0033] As shown in the Figure 1 The financial consulting response method proposed in the embodiments of the present application comprises the following steps: S1: receiving a financial consulting question.

[0034] S2: generating a fuzzy noise feature description according to the financial consulting question: analyzing the language features, expression methods and question structures of the financial consulting question, and identifying the implicit cognitive needs of the user; Based on the implicit cognitive needs, a fuzzy noise feature description is generated to introduce controllable fuzziness to modulate the reasoning process; When the reasoning forms a candidate answer, each reasoner receives the fuzzy noise feature description when performing a structured reasoning process, so as to modulate the reasoning process of the reasoner through the fuzzy noise feature description.

[0035] Specifically, the fuzzy noise feature description comprises: attention allocation feature description, for adjusting the attention intensity of the reasoner to different types of information to affect the weight allocation of direct factors and indirect factors in the reasoning process; speculative depth feature description, for adjusting the decomposition granularity and hierarchical depth of the reasoner to affect the hierarchical depth of sub-problem decomposition; information dependency feature description, for adjusting the dependency weight of the reasoner on existing knowledge and reasoning information to affect the dependency degree of the reasoning result on existing evidence and information generated based on reasoning; conclusion concentration feature description, for adjusting the convergence tendency of the reasoner to affect the degree of convergence of the reasoning output from multiple possibilities to a single conclusion.

[0036] Specifically, the structure for implementing this step in the pre-set large language model comprises: a fuzzy noise recognition module for performing the step of generating a fuzzy noise feature description according to the financial consulting question, comprising: a context analysis unit for analyzing the language features, expression methods and question structures of the financial consulting question.

[0037] Specifically, the context analysis unit receives a financial consultation question as input and processes it through the following sub-units to form a context analysis result. Taking the financial consultation question "The central bank just announced a 50BP interest rate cut. I want to understand what impact this will have on the real estate market?" as an example, the specific implementation of each sub-unit is as follows: The lexical feature extraction sub-unit is used to identify key words, emotional color, and urgency expression in the financial consultation question.

[0038] The lexical feature extraction sub-unit is implemented through the following prompt words: Please analyze the key lexical features in the following financial consultation question: Question: {Financial consultation question}; Please identify: Key words: extract financial professional terms, industry words, and core concepts; Emotional color: analyze the emotional inclination (positive / neutral / negative) reflected in the question; Urgency expression: identify time-related words and urgency level indicators; Output format: Key words: [list of words]; Emotional color: [emotional judgment]; Urgency: [urgency level]; For the above example, the output of the current sub-unit processing the example question is: Key words: [central bank, interest rate cut, 50BP, real estate market, impact]; Emotional color: [neutral exploration]; Urgency: [immediate attention - just released].

[0039] The expression analysis sub-unit is used to analyze the language style, certainty level, and openness features of the financial consultation question.

[0040] Specifically, the expression analysis sub-unit is implemented through the following prompt words: Please analyze the expression features of the following financial consultation question: Question: {Financial consultation question}; Please analyze: Language style: judge the formal level and professional level of expression; Certainty level: identify the strength of certainty demand reflected in the question; Openness feature: judge the openness and exploration willingness of the question; Output format: Language style: [style description]; Certainty level: [certainty demand]; Openness feature: [openness level]; For the above example, the output of the current sub-unit processing the example question is: Language style: [Formal but calm consulting tone]; Degree of certainty: [Seeks clear analysis but accepts complexity]; Openness: [Open to inquiry, willing to accept comprehensive analysis].

[0041] The problem structure analysis sub-unit is used to identify the complexity, hierarchy and information requirement of financial consulting problems; The question structure analysis subunit is implemented through the following prompt words: Please analyze the structural characteristics of the following financial consulting questions: Question: {Financial Counseling Question}; Please identify: Complexity: Determine the number of elements involved in the problem and the complexity of the associations; Hierarchy: whether the analysis problem contains multiple levels of analysis requirements; Information needs: assess the user's expectations for the level of information detail; Output format: Complexity: [complexity]; Hierarchy: [hierarchical analysis]; Information requirements: [information expectations]; For the above example, the output of the current subunit processing the example problem is: Complexity: [Medium complexity - involving analysis of the relationship between monetary policy and the real estate market]; Hierarchy: [Multi-level needs - from policy mechanisms to market impact]; Information needs: [Relatively comprehensive analysis needs].

[0042] Output the context analysis results, that is, the output results of the above three units to the cognitive demand inference unit to form implicit cognitive needs.

[0043] The cognitive needs inference unit is used to identify the user's implicit cognitive needs based on the context analysis results.

[0044] Specifically, the cognitive demand inference unit receives the context analysis result and processes it through the following subunits: The cognitive style identification subunit is used to infer the user's preferred analysis style and thinking mode based on the context analysis results.

[0045] Specifically, the cognitive style identification subunit is implemented through the following prompt words: Please infer the user's cognitive style preference based on the following context analysis results: Context analysis results: {lexical features, expression methods, question structure} Please infer: Analysis style preference: Determine whether the user prefers detailed analysis or a macro perspective; Thinking mode: Identify the user's thinking bias (rational analysis / intuitive judgment / balanced synthesis); Expertise: Assess the user's level of financial expertise and understanding; Output format: Analysis style: [Style assessment]; Thought pattern: [Thinking characteristics]; Professional level: [Professional level]; For the above example, the output of the current sub-unit is: Analysis style: [Balanced - Combination of macro policy analysis and specific market impact]; Thought pattern: [Rational inquiry - Logical-based systematic analysis]; Professional level: [Moderate professional level - Understanding of basic concepts but need in-depth explanation].

[0046] The demand depth assessment sub-unit is used to determine the user's expectations for analysis depth, detail level, and complexity based on the context analysis results.

[0047] Specifically, the demand depth assessment sub-unit is implemented through the following prompt words: Please assess the user's analysis depth demand based on the following context analysis results: Context analysis results: {Vocabulary features, expression methods, question structure}; Please assess: Analysis depth expectation: Determine the user's desired analysis depth (shallow / medium / deep); Detail attention: Evaluate the user's attention to specific details and data; Complexity acceptance: Determine the user's acceptance of complex analysis and multi-factor consideration; Output format: Depth expectation: [Depth level]; Detail attention: [Attention level]; Complexity acceptance: [Acceptance level]; For the above example, the output of the current sub-unit is: Depth expectation: [Medium depth - Need mechanism analysis but avoid excessive technicality]; Detail attention: [Moderate attention - Attention to key data and impact path]; Complexity acceptance: [Medium acceptance - Can understand multi-level analysis but need structured presentation].

[0048] The decision-oriented analysis sub-unit is used to identify the user's decision time window and certainty demand based on the context analysis results. Specifically, the decision-oriented analysis sub-unit is implemented through the following prompt words: Please analyze the user's decision-oriented characteristics based on the following context analysis results: Context analysis results: {Vocabulary features, expression methods, question structure}; Please analyze: Decision time window: Determine the user's decision urgency (immediate / short-term / medium and long-term); Certainty demand: Evaluate the user's demand for clear conclusions; Action orientation: Determine whether the user needs specific action suggestions; Output format: Time window: [temporal focus]; Certainty demand: [certainty level]; Action orientation: [action demand]. For the above example, the output of the current subunit is: Time window: [short-term focus - immediate understanding of recent policy impact demand]; Certainty demand: [moderate certainty - hope to obtain relatively clear impact judgment]; Action orientation: [information-oriented - mainly for understanding impact rather than immediate action].

[0049] The output implicit cognitive demand to the noise feature generation unit.

[0050] The noise feature generation unit is configured to generate a fuzzy noise feature description based on the implicit cognitive demand.

[0051] Specifically, the noise feature generation unit receives the implicit cognitive demand and generates a fuzzy noise feature description through the following subunits: The attention feature generation subunit is configured to generate an attention allocation feature description based on the implicit cognitive demand. Specifically, the attention feature generation subunit is implemented through the following prompt words: Please generate an attention allocation feature description based on the following implicit cognitive demand: Implicit cognitive demand: {cognitive style, demand depth, decision orientation}; Please generate a feature description of the attention allocation of the modulation reasoner, indicating how to adjust the attention intensity to different information; Output format: Attention allocation feature description: [specific modulation guidance description]; For the above example, the output of the current subunit is: Attention allocation feature description: Balance macro policy impact and specific market performance, focus on direct impact and indirect effects, and attention should be moderately allocated among various aspects of policy transmission.

[0052] The depth feature generation subunit is configured to generate a speculation depth feature description based on the implicit cognitive demand.

[0053] Specifically, the depth feature generation subunit is implemented through the following prompt words: Please generate a speculation depth feature description based on the following implicit cognitive demand: Implicit cognitive demand: {cognitive style, demand depth, decision orientation}; Please generate a feature description of the speculation depth of the modulation reasoner, indicating how to adjust the hierarchical depth of sub-problem decomposition; Output format: Speculation depth feature description: [specific modulation guidance description]; For the above example, the output of the current subunit is: Speculative depth feature description: based on policy logic and historical experience, moderate extrapolation can be combined with medium-depth causal analysis of transmission mechanisms to avoid excessive speculative prediction; Dependency feature generation subunit, for generating information dependency feature description based on implicit cognitive needs.

[0054] Specifically, the dependency feature generation subunit is implemented through the following prompt words: Please generate information dependency feature description based on the following implicit cognitive needs: Implicit cognitive needs: {cognitive style, depth of demand, decision orientation}; Please generate a feature description of the modulation of the reasoning information dependency, indicating how to adjust the dependency weight of existing evidence and reasoning information; Output format: Information dependency feature description: [specific modulation guidance description]; For the above example, the output of the current subunit is: Information dependency feature description: mainly rely on policy mechanisms and market rules, can moderately combine historical patterns for reasoning, maintain the balance of factual basis and logical reasoning.

[0055] Concentration feature generation subunit, for generating conclusion concentration feature description based on implicit cognitive needs.

[0056] Specifically, the concentration feature generation subunit is implemented through the following prompt words: Please generate a conclusion concentration feature description based on the following implicit cognitive needs: Implicit cognitive needs: {cognitive style, depth of demand, decision orientation}; Please generate a feature description of the modulation of the reasoning conclusion concentration, indicating how to adjust the openness and convergence of the reasoning output; Output format: Conclusion concentration feature description: [specific modulation guidance description]; For the above example, the output of the current subunit is: Conclusion concentration feature description: tend to provide structured impact analysis, can contain multiple levels of impact, but should form a relatively clear overall judgment.

[0057] Output fuzzy noise feature description to reasoning module.

[0058] S3: Reasoning forms candidate answers: Through the parallel execution of multiple reasoners, each reasoner independently executes the following progressive steps: Identify and interpret the financial concepts and terms involved in the financial consulting problem; Based on the results of the concept interpretation, decompose the financial consulting problem into multiple logically related sub-problems; According to the logical order of sub-problems, the answers to each sub-problem are generated in turn only according to the answers to the previously answered sub-problems, forming a reasoning chain; Collecting the reasoning chains generated by each reasoner and their corresponding final answers forms an answer pool, from which the answer with the highest frequency of occurrence is selected as the candidate answer.

[0059] Specifically, the structure for implementing this step in a pre-set large language model includes: A reasoning module for performing the step of reasoning to form a candidate answer, including a plurality of parallel reasoners, each of which includes: A concept interpretation unit for identifying and interpreting financial concepts and terms in the financial consulting question; Specifically, the concept interpretation unit receives the financial consulting question and the attention allocation feature description and processes it through the following sub-units: A financial term identification sub-unit for extracting financial concepts and professional terms from the financial consulting question; The financial term identification sub-unit is implemented through the following prompt words: Please identify and extract financial concepts and professional terms from the following financial consulting question: Question: {Financial consulting question}; Please extract: Financial professional terms: Identify professional financial vocabulary in the question; Industry concepts: Extract related industry and market concepts; Technical indicators: Identify specific numerical values and indicators involved; Output format: Financial terms: [list of terms]; Industry concepts: [list of concepts]; Technical indicators: [list of indicators]; For the recently announced 50BP interest rate cut by the central bank, I want to know what impact it will have on the real estate market? The output of the current sub-unit is: Financial terms: [central bank, interest rate cut, basis points BP]; Industry concepts: [real estate market, monetary policy]; Technical indicators: [50BP, interest rate changes].

[0060] A concept definition generation sub-unit for generating accurate definitions and explanations of the financial concepts and terms under the condition that the attention allocation feature description modulates the level of detail and focus of the concept interpretation.

[0061] The concept definition generation sub-unit is implemented through the following prompt words: Please generate definitions and explanations for the following financial concepts based on the attention allocation feature description: Financial concepts: {extracted financial terms and concepts}; Attention allocation features: {attention allocation feature description}; Please generate: Basic definition: Provide a precise definition of the concept; Mechanism of association: Explain the interrelation between concepts; Influencing factors: Highlight relevant influencing factors according to attention characteristics; Output format: [Concept]: Definition + Association + Influencing factors.

[0062] For the above example, under the modulation of the attention characteristic "need to balance the impact of macro policy and specific market performance", the output of the current sub-unit is: Central bank rate cut: Central bank monetary policy tool to release liquidity and reduce market funding costs by lowering the benchmark interest rate, its effect is transmitted to various economic sectors through the banking system; Real estate market: A comprehensive market including residential and commercial real estate, involving development, sales, financing, leasing and other aspects, with high sensitivity to interest rate changes; Base point (BP): Unit of measurement of interest rates, 50BP means a 0.5 percentage point decrease in interest rates, which is a relatively significant policy adjustment; Interest rate transmission mechanism: Multi-level transmission process from central bank policy interest rate to bank lending interest rate and then to various sectors of the real economy, with time lag and transmission efficiency differences.

[0063] Sub-problem decomposition unit, used to decompose financial consulting problems into logically related sub-problems; Specifically, the sub-problem decomposition unit receives the financial consulting problem, the concept interpretation result and the speculation depth characteristic description, and under the modulation of the speculation depth characteristic description decomposition depth, it decomposes the financial consulting problem into multiple logically related sub-problems according to the concept interpretation result, forming a sub-problem sequence.

[0064] Specifically, the sub-problem decomposition unit receives the financial consulting problem, the concept interpretation result and the speculation depth characteristic description, and under the modulation of the speculation depth characteristic description decomposition depth, it decomposes the financial consulting problem into multiple logically related sub-problems according to the concept interpretation result, forming a sub-problem sequence; The sub-problem decomposition unit is implemented through the following prompt words: Please decompose the financial consulting problem into logically related sub-problems based on the concept interpretation result and the speculation depth characteristic: Original problem: {Financial consulting problem}; Concept interpretation: {Concept interpretation result}; Speculation depth characteristic: {Speculation depth characteristic description}; Please generate a logically related sub-problem sequence according to the speculation depth characteristic modulation of the decomposition level; Output format: Sub-problem 1: [Specific question]; Sub-problem 2: [Specific question];... For the above example, under the speculation depth feature modulation of "moderate extrapolation based on policy logic and historical experience, combined with medium-depth causal analysis of transmission mechanisms", the output of the current unit is: Sub-problem 1: How does a 50BP interest rate cut affect housing finance costs through what path? What is the transmission efficiency? Sub-problem 2: How do changes in mortgage costs affect the decisions of different types of homebuyers? Sub-problem 3: How do changes in demand interact with supply factors to affect market performance? Sub-problem 4: How does the effect of the policy manifest in different cities and different time periods? Sub-problem 5: What impact does it have on indicators such as transaction volume and inventory, in addition to price? Output the sub-problem sequence to the step-by-step solving unit.

[0065] The step-by-step solving unit is used to generate answers to each sub-problem in logical order; Specifically, the step-by-step solving unit receives the sub-problem sequence, information dependency feature description, and conclusion concentration feature description, and processes them through the following sub-units: Logical deduction sub-unit, for generating answers to each sub-problem in logical order under the condition that the facts and information relied upon in the logical deduction of the information dependency feature description are modulated.

[0066] Specifically, the logical deduction sub-unit is implemented through the following prompt words: Please answer the sub-problems in logical order based on the information dependency feature: Sub-problem: {Current sub-problem}; Answer to the previous question: {Previous question answer}; Information dependency feature: {Information dependency feature description}; Please answer the current question based on the information relied upon in the reasoning and information sources modulated by the information dependency feature, only based on the previous answer; Output format: Answer: [Specific analysis and conclusion].

[0067] Conclusion modulation sub-unit, for modulating the openness and certainty of the answer according to the conclusion concentration feature description.

[0068] Specifically, the conclusion modulation sub-unit is implemented through the following prompt words: Please modulate the expression of the following answer according to the conclusion concentration feature: Original answer: {Logical deduction result}; Conclusion concentration feature: {Conclusion concentration feature description}; Please adjust the degree of openness and certainty of the answer according to the characteristics of the question; Output format: Answer after modulation: [modulated expression].

[0069] Output the reasoning chain and the final answer to the answer pool statistics module.

[0070] Specifically, the structure for implementing this step in a pre-set large language model includes: An answer pool statistics module for collecting the reasoning chains and final answers of each reasoner, forming an answer pool and selecting a candidate answer.

[0071] The answer pool statistics module collects the reasoning chains and corresponding final answers generated by each reasoner, counts the frequency of each answer, and selects the answer with the highest frequency as the candidate answer, while retaining the reasoning chain corresponding to the candidate answer for subsequent evaluation.

[0072] S4: Evaluate the candidate answer: Obtain the reasoning chain corresponding to the candidate answer, and under the premise of knowing the reasoning chain and the final answer, check the logical rationality and factual accuracy of each reasoning step under the condition of global information.

[0073] Select an alternative answer different from the candidate answer from the answer pool, re-evaluate the effectiveness of each evidence in the original reasoning chain under the premise that the candidate answer and the alternative answer are both true, filter out the still effective evidence, and only re-reason based on the effective evidence to test whether the result of re-reasoning is consistent with the candidate answer.

[0074] Specifically, the structure for implementing this step in a pre-set large language model includes: An evaluation module for performing the steps of evaluating the candidate answer, including: A global logical evaluation unit for checking the logical rationality and factual accuracy of each reasoning step under the premise of knowing the reasoning chain and the final answer.

[0075] Specifically, the global logical evaluation unit receives the reasoning chain corresponding to the candidate answer and processes it through the following subunits: A reasoning step extraction subunit for extracting each reasoning step from the reasoning chain.

[0076] Specifically, the reasoning step extraction subunit is implemented through the following prompt words: Please divide the following reasoning chain into multiple reasoning steps in order: Reasoning chain: {complete reasoning chain}; Output format: Step 1: [Sub-question 1 and its answer]; Step 2: [Sub-question 2 and its answer]; Step 3: [Sub-question 3 and its answer]...

[0077] A global information construction sub-unit is configured to construct a global information background based on the inference chain and the final answer.

[0078] Specifically, the global information construction sub-unit is implemented through the following prompt: Please construct a global information background based on the following inference chain and final answer: Inference chain: {Complete inference chain}; Final answer: {Candidate answer}; Please construct a global information background containing the following: Inference context: The logical main line and causal relationship of the entire reasoning; Key nodes: Important turning points and key conclusions in the reasoning process; Conclusion basis: Main evidence and reasoning steps relied on by the final answer; Output format: Global background: [Comprehensive information description].

[0079] A step-by-step checking sub-unit is configured to check the logical rationality and factual accuracy of each reasoning step from the starting step under the premise of known global information.

[0080] A step-by-step checking sub-unit is configured to check the logical rationality and factual accuracy of each reasoning step from the starting step under the premise of known global information. Specifically, the step-by-step checking sub-unit is implemented through the following prompt: Please check the following reasoning steps under the premise of known global information: Current checking step: {Specific reasoning step}; Global information background: {Global information construction result}; Known final answer: {Candidate answer}; Please check: Logical rationality: Is the logical derivation of this step reasonable? Factual accuracy: Are the facts involved in this step accurate? Consistency: Is this step consistent with the overall reasoning direction? Sufficiency: Considering the final answer, is this step sufficient and necessary? Output format: Check result: Pass / Fail.

[0081] Output the logical evaluation result to the iteration control module.

[0082] The consistency verification unit is configured to reevaluate the validity of the evidence in the reasoning chain and verify the reasoning consistency by assuming that the candidate answer and the alternative answer are both correct.

[0083] Specifically, the consistency verification unit receives the reasoning chain corresponding to the candidate answer and the answer pool information, and processes the same by the following subunits: The alternative answer selection subunit is configured to randomly select an alternative answer different from the candidate answer from the answer pool.

[0084] The premise construction subunit is configured to construct a hypothetical premise that the candidate answer and the alternative answer are both correct.

[0085] That is, directly assuming that the candidate answer and the alternative answer are both correct, and constructing a hypothetical premise that the two answers are both correct as a basis for subsequent evidence evaluation.

[0086] The evidence validity evaluation subunit is configured to reevaluate the validity of each evidence in the original reasoning chain under the hypothetical premise.

[0087] Specifically, the evidence validity evaluation subunit is implemented by the following prompt: Please reevaluate the validity of the following evidence under the assumption that both the candidate answer and the alternative answer are correct: Original reasoning chain evidence: {evidence in the reasoning chain}; Candidate answer: {candidate answer}; Alternative answer: {alternative answer}; Hypothetical premise: both the candidate answer and the alternative answer are correct; Please evaluate each evidence: Under the assumption that both answers are correct, is the evidence still valid? Does the evidence have a logical conflict with any answer? Does the evidence need to be excluded? Output format: evidence 1: valid / invalid, reason: [explanation]; evidence 2: valid / invalid.

[0088] The evidence screening subunit is configured to screen out the evidence that is still valid.

[0089] The re-reasoning subunit is configured to re-reason only based on the valid evidence.

[0090] Specifically, the re-reasoning subunit is implemented by the following prompt: {This is the prompt for the above logical derivation subunit}, and please re-reason only based on the following valid evidence: Valid evidence: {screened valid evidence}; Original question: {financial consulting question}.

[0091] Output format: Re-reasoning conclusion: [Based on valid evidence-based conclusion].

[0092] Consistency checking sub-unit, for checking whether the re-reasoning result is consistent with the candidate answer.

[0093] Output consistency verification result to iteration control module.

[0094] S5: Repeat the above reasoning and evaluation process, and output the candidate answer as the financial consulting answer when all evaluation conditions are met.

[0095] Specifically, the structure of this step is implemented in a pre-set large language model, which includes: Iteration control module, for controlling the repeated execution of the reasoning and evaluation process, and outputting the passed candidate answer as the financial consulting answer after meeting the condition that all evaluation steps are passed.

[0096] At the same time, the embodiments of the present application respectively put forward two kinds of example results: Example one (without adding noise fuzzy feature description for modulation): User question: The central bank just announced a 50BP interest rate cut. I want to know what impact it has on the real estate market? Reasoner workflow: Concept explanation: Central bank interest rate cut: Central bank monetary policy operation of lowering the benchmark interest rate; Real estate market: Housing buying and selling and leasing transaction market; Base point (BP): Interest rate measurement unit, 1 base point = 0.01%; Interest rate transmission mechanism: The process of monetary policy affecting the real economy through the banking system; Sub-question decomposition: Sub-question 1: What is the direct impact of a 50BP interest rate cut on mortgage interest rates? Sub-question 2: How does the decline in mortgage costs affect housing demand? Sub-question 3: What impact does demand change have on housing prices? Step-by-step solution: e1: A 50BP interest rate cut usually reduces mortgage interest rates by 30-40BP; e2: Lower mortgage costs improve the affordability of homebuyers and stimulate demand growth; e3: Demand increases drive up housing prices in the case of relatively stable supply; Final answer through evaluation: A 50BP interest rate cut is generally good for the real estate market and is expected to drive up housing prices.

[0097] Example Two (Joining Fuzzy Noise Feature Description for Modulation): First Step: Obtain Fuzzy Noise Identification Description: User Question: The central bank just announced a 50BP interest rate cut. I want to understand what impact this has on the real estate market? Fuzzy Noise Feature Description Generation: Attention Allocation Feature: Balance macro policy impact and specific market performance, focus on direct impact and indirect effects, attention should be moderately distributed among the various links of policy transmission.

[0098] Speculation Depth Feature: Based on policy logic and historical experience, moderate extrapolation can be performed, combining transmission mechanisms for medium-depth causal analysis, avoiding excessive speculative predictions.

[0099] Information Dependency Feature: Mainly rely on policy mechanisms and market rules, can moderately combine historical patterns for reasoning, maintaining a balance between factual basis and logical reasoning.

[0100] Conclusion Concentration Feature: Tend to provide structured impact analysis, can include multiple levels of impact, but should form a relatively clear overall judgment.

[0101] Conceptual Explanation: Central Bank Rate Cut: Central bank releases liquidity and reduces market funding costs through the monetary policy tool of lowering the benchmark interest rate, which is transmitted through the banking system to various economic sectors; Real Estate Market: A comprehensive market including residential and commercial real estate, involving development, sales, financing, leasing, and other aspects, with high sensitivity to interest rate changes; Base Point (BP): Unit of measurement for interest rates, 50BP means a 0.5 percentage point decrease in interest rates, which is a relatively significant policy adjustment; Interest Rate Transmission Mechanism: From the central bank's policy interest rate to bank lending rates, and then to multi-level transmission in various sectors of the real economy, with time lags and transmission efficiency differences; Sub-question Decomposition: Sub-question 1: How does a 50BP interest rate cut affect real estate financing costs through what path? What is the transmission efficiency? Sub-question 2: How do changes in mortgage costs affect the decisions of different types of homebuyers? Sub-question 3: How do changes in demand interact with supply factors to affect market performance? Sub-question 4: How does the policy effect manifest in different cities and time periods? Sub-question 5: What impact does it have on transaction volume, inventory, and other indicators besides price? Step-by-step Solution: e1: Based on historical transmission efficiency, a 50BP central bank rate cut usually leads to a 30-40BP decline in mortgage rates within 2-3 months, but the specific magnitude depends on bank funding costs and risk appetite; e2: A 0.3-0.4% decline in mortgage rates has different impacts on different homebuying groups: first-time homebuyers benefit significantly, and improvement demand is released, but investment demand is still subject to policy constraints; e3: The market effect of demand release depends on supply elasticity: supply is relatively rigid in first- and second-tier cities, and price response is more pronounced; supply is sufficient in third- and fourth-tier cities, and more reflected in increased transaction volume; e4: From a time perspective, policy effects usually emerge gradually within 3-6 months, primarily affecting market expectations and sentiment in the early stage, and only in actual transactions in the later stage; e5: Considering the current policy environment of not speculating on housing, the effect of rate cuts is more reflected in improved market sentiment and the release of reasonable demand, rather than speculative speculation.

[0102] The final answer through evaluation: A 50BP rate cut has multiple levels of impact on the real estate market: Direct effect: reduces mortgage costs, improves homebuyer affordability, and is expected to reduce mortgage rates by 30-40BP within 2-3 months; Market response: improves market expectations, releases some of the expected demand, especially reasonable first-time homebuyers and improvement demand; Regional differences: due to relatively tight supply in first- and second-tier cities, it may be more reflected in price support; in third- and fourth-tier cities, it is more reflected in the recovery of transaction volume; Time rhythm: policy effects are expected to emerge gradually within 3-6 months, affecting expectations in the early stage, transaction volume in the middle stage, and price in the later stage; Overall assessment: under the current policy environment, rate cuts have a generally positive impact on the real estate market, but the impact is relatively mild, mainly supporting reasonable housing demand rather than driving significant price increases.

[0103] In summary, the technical scheme provided by the embodiments of the present application has the following advantages compared with the prior art: One of its beneficial effects and working principles is: Although traditional large language models have certain reasoning capabilities, in the high-risk professional field of financial consulting, their reasoning process is prone to factual errors and reasoning confusion, leading to insufficient reasoning credibility, which may mislead users' financial decisions and cause economic losses.

[0104] This application forms an answer pool by having each reasoner independently perform the progressive steps of financial concept interpretation, sub-problem decomposition and step-by-step solution. After forming a reasoning chain, the answers of each reasoner are collected to form an answer pool, and then the candidate answers are verified through evaluation.

[0105] Its working principle is that the reasoner design of this application ensures the logical rigor of financial analysis through structured reasoning, in which the concept explanation step ensures the accuracy of understanding of financial terms, the sub-problem decomposition step converts complex financial problems into manageable sub-problems, and the step-by-step solution step ensures the coherence of the reasoning chain based on logical association relationships.

[0106] The financial consulting response method provided by this application improves the accuracy and credibility of financial consulting response reasoning, and can more accurately answer financial consulting questions raised by customers.

[0107] The second beneficial effect and its working principle are: Financial consulting scenarios are characterized by diversity and scenario-dependence. Many financial problems are essentially multi-solution problems, requiring multiple reasonable solutions based on different assumptions or risk preferences. For example, asset allocation issues require consideration of allocation strategies under different scenarios such as market optimism, neutrality, and pessimism. Investment timing selection requires providing corresponding suggestions based on different market cycle assumptions.

[0108] However, although the above-mentioned reasoning and evaluation methods can improve the credibility of reasoning, because the step-by-step solution link of the reasoner requires answering the current sub-problem only based on the answers to the previously solved sub-problems, and the concept interpretation step always tends to obtain standard definitions, this composite structure that strongly relies on general prior knowledge and rigid causal dependencies prevents the reasoner from performing hypothetical analysis and scenario modeling.

[0109] The structure of the above reasoners results in each reasoner following the most generally correct steps for concept interpretation, subproblem decomposition, and step-by-step autoregressive solution. Consequently, this cognitive model, which relies heavily on general prior knowledge, lacks the ability to perform contextualized reasoning based on different conditions. Consequently, even with multiple reasoners, the answers they generate always converge on a single case.

[0110] Because this structure inherently excludes multiple solutions and hypothetical analysis, it's unable to handle complex financial consulting questions that require multiple answers or conditional recommendations. It also fails to adjust the depth, focus, and openness of reasoning based on the user's cognitive state and the context of the question, reducing the accuracy and professionalism of financial consulting responses relative to user intent. In other words, this structure lacks the ability to recognize and reason that certain knowledge may be more accurate in certain specific situations.

[0111] The application can identify the implicit cognitive needs of the user by analyzing the language features and problem structure of the financial consulting question before forming a candidate answer through reasoning, generate a fuzzy noise feature description to introduce controllable fuzzy noise modulation reasoning process, enable the reasoner to break through the cognitive mode limit, and dynamically adjust the focus, depth, information dependence and conclusion openness of reasoning according to the noise features. The flexibility and adaptability of reasoning are restored under the premise of maintaining logical rigor.

[0112] The working principle is that the fuzzy noise feature description can adaptively modulate the cognitive processing of the reasoner. Among them, the attention allocation feature description can enable the reasoner to dynamically adjust between direct factor analysis and indirect factor analysis according to the problem features and user needs, the speculation depth feature description can enable the reasoner to adjust the hierarchical depth of sub-problem decomposition according to the user's question, the information dependence feature can enable the reasoner to find a balance point between priori fact verification and logical extrapolation, and the conclusion concentration feature fundamentally solves the problem of forced answer convergence, allowing the reasoner to dynamically select between providing deterministic suggestions and maintaining multiple possibilities according to the multi-solution nature of the question and the decision needs of the user.

[0113] Therefore, the application introduces controllable fuzzy noise feature description to modulate the processing logic of the reasoner to support situation analysis and multi-solution coexistence, thereby solving the limitations of fixed reasoning structure and improving the accuracy of the answer relative to the user's intention. The controllable fuzzy noise on the simple single-solution problem can still accurately converge on the unique solution.

[0114] Therefore, the financial consulting response method provided by the application can more accurately answer the financial consulting questions raised by the user.

[0115] The embodiment of the application also provides a financial consulting system, which comprises a processor and a memory. The memory stores at least one instruction, at least one program, a code set or an instruction set. The at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to realize the financial consulting response method as described in the above embodiment.

[0116] It should be noted that, in the present document, the terms such as "first" and "second" and the like are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual such relationship or order between such entities or operations. In addition, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, an element defined by the statement "comprising a" does not exclude the existence of additional identical elements in the process, method, article, or apparatus that includes the stated element. Also, in the description of the embodiments of the present application, unless otherwise specified, " / " means or, for example, A / B can mean A or B; "and / or" in the present document is only a description of the associated relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, A and B exist together, and B exists alone. Also, in the description of the embodiments of the present application, "a plurality of" means two or more than two.

[0117] The above description is only a specific implementation of the present application, which enables those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments described herein, but will conform to the widest range consistent with the principles and novel features disclosed herein.

Claims

1. A financial consultation response method, characterized in that: The financial consultation response method comprises the following steps: Receive financial counseling questions; Reasoning to form candidate answers: The structured reasoning process is performed in parallel by multiple reasoners, each of which independently performs the following progressive steps: Identify and explain financial concepts and terminology involved in stated financial consulting questions; Based on the results of the concept explanation, the financial consulting problem is decomposed into multiple logically related sub-problems; Following the logical order of the sub-questions, answer the current sub-question based only on the answers to the previously answered sub-questions, generating answers to each sub-question in turn to form a reasoning chain; Collect the reasoning chains generated by each reasoner and their corresponding final answers to form an answer pool, and select the answer with the highest frequency from the answer pool as the candidate answer; evaluating the candidate answers; The above reasoning and evaluation process is repeated, and when all evaluation conditions are met, the candidate answer is output as the financial consulting answer.

2. The financial consultation response method according to claim 1, characterized in that: After receiving the financial consulting question and before forming a candidate answer through reasoning, the following steps are also included: Generate fuzzy noise feature description based on the financial consulting question: Analyze the language characteristics, expression methods and question structure of the financial consulting questions to identify the user's implicit cognitive needs; Based on implicit cognitive needs, fuzzy noise feature descriptions are generated to introduce controllable fuzziness to modulate the reasoning process; When forming candidate answers through reasoning, each reasoner receives the fuzzy noise feature description when executing a structured reasoning process, so as to modulate the reasoning process of the reasoner through the fuzzy noise feature description.

3. The financial consultation response method according to claim 2, characterized in that: The fuzzy noise feature description includes: Attention allocation feature description, used to adjust the reasoner's attention intensity to different types of information, so as to affect the weight distribution of direct and indirect factors in the reasoning process; The inferred depth feature description is used to adjust the decomposition granularity and hierarchical depth of the reasoner to affect the hierarchical depth of sub-problem decomposition; Information dependency feature description is used to adjust the reasoner's reliance on existing knowledge and reasoning information, thereby affecting the degree to which the reasoning results rely on existing evidence and information generated based on reasoning; The conclusion concentration feature description is used to adjust the convergence tendency of the reasoner to affect the degree to which the reasoning output converges from multiple possibilities to a single conclusion.

4. The financial consultation response method according to claim 2, characterized in that: The financial consultation response method is implemented through a large language model with the following structure: The fuzzy noise recognition module is used to execute the step of generating a fuzzy noise feature description based on the financial consulting question, including: Contextual Analysis Unit, which is used to analyze the language features, expressions, and question structures of financial consulting questions; Cognitive needs inference unit, used to identify users' implicit cognitive needs based on context analysis results; A noise feature generation unit, configured to generate a fuzzy noise feature description based on implicit cognitive requirements; The reasoning module is used to perform reasoning to form candidate answers. It includes multiple parallel reasoners, each of which includes: Conceptual explanation unit, which is used to identify and explain financial concepts and terms used in financial consulting problems; Sub-problem decomposition unit, used to decompose financial consulting problems into logically related sub-problems; A step-by-step solution unit is used to generate answers to each sub-problem in the logical order of the sub-problems; wherein each reasoner receives the fuzzy noise feature description to modulate the reasoning process; The answer pool statistics module is used to collect the reasoning chains and final answers of each reasoner, form an answer pool, and select candidate answers; An evaluation module, configured to execute the step of evaluating candidate answers, comprising: A global logic evaluation unit, which checks the logical rationality and factual accuracy of each reasoning step by step, given that the reasoning chain and final answer are known; The consistency verification unit is used to re-evaluate the validity of evidence in the reasoning chain and test the consistency of reasoning by assuming that the candidate answer and the alternative answer are both true; The iterative control module is used to control the repeated execution of the reasoning and evaluation process until all evaluation steps are met.

5. The financial consultation response method according to claim 4, characterized in that: The specific structure of the fuzzy noise recognition module includes: The context analysis unit receives financial consulting questions as input and processes them through the following subunits to form context analysis results: Lexical feature extraction subunit, used to identify key words, sentiment and urgency expressions in financial consulting questions; The expression analysis subunit is used to analyze the language style, degree of certainty and openness characteristics of financial consulting questions; The problem structure analysis sub-unit is used to identify the complexity, hierarchy and information requirement of financial consulting problems; Output the context analysis results to the cognitive demand inference unit to form implicit cognitive demands; The cognitive demand inference unit receives the context analysis result and processes it through the following subunits: Cognitive style identification subunit, used to infer the user's preferred analytical style and thinking mode based on contextual analysis results; Depth of Requirement Assessment subunit, which is used to determine the user's expectations for analysis depth, detail, and complexity based on the results of the contextual analysis; The decision-oriented analysis subunit is used to identify the user's decision time window and deterministic needs based on the context analysis results; Output implicit cognitive requirements to the noise feature generation unit; The noise feature generation unit receives the implicit cognitive requirement and generates a fuzzy noise feature description through the following subunits: Attention feature generation subunit, used to generate attention allocation feature description based on implicit cognitive needs; A deep feature generation subunit, used to generate inferred deep feature descriptions based on implicit cognitive needs; A dependency feature generation subunit, used to generate information dependency feature descriptions based on implicit cognitive needs; The concentration feature generation subunit is used to generate a description of the conclusion concentration feature based on implicit cognitive needs; Output the fuzzy noise feature description to the inference module.

6. The financial consultation response method according to claim 5, characterized in that: The specific structure of each reasoner is: The concept interpretation unit receives financial consulting questions and attention allocation feature descriptions and processes them through the following subunits: Financial terminology identification subunit, used to extract financial concepts and professional terms from financial consulting questions; A concept definition generation subunit, configured to generate accurate definitions and explanations of the financial concepts and terms while the attention allocation feature description modulates the level of detail and emphasis of the concept explanations; Output the concept explanation results to the sub-problem decomposition unit; a sub-problem decomposition unit, receiving a financial consulting question, a concept interpretation result, and an inferred depth feature description, and configured to decompose the financial consulting question into a plurality of logically related sub-problems according to the concept interpretation result, forming a sub-problem sequence, while modulating the sub-problem decomposition depth by the inferred depth feature description; Output subproblem sequence to step-by-step solution unit; The step-by-step solution unit receives the sub-problem sequence, information dependency feature description, and conclusion concentration feature description, and processes them through the following sub-units: A logic deduction subunit is used to generate answers to the sub-questions in sequence according to the logical order of the sub-questions when the information dependency feature describes the facts relied upon when the sub-questions are logically deduced; The conclusion modulation subunit is used to describe the degree of openness and certainty of the modulated answer based on the conclusion concentration characteristics; Output the reasoning chain to the answer pool statistics module.

7. The financial consultation response method according to claim 4, characterized in that: The specific structure of the assessment module includes: The global logic evaluation unit receives the reasoning chain corresponding to the candidate answer and processes it through the following sub-units: An inference step extraction subunit, used to extract each inference step from the inference chain; The global information construction subunit is used to construct the global information context based on the reasoning chain and the final answer; The step-by-step checking sub-unit is used to check the logical rationality and factual accuracy of each reasoning step starting from the initial step under the premise that the global information is known; Output logic evaluation results to the iterative control module; The consistency verification unit receives the reasoning chain and answer pool information corresponding to the candidate answer and processes it through the following subunits: an alternative answer selection subunit, configured to select an alternative answer different from a candidate answer from an answer pool; The premise construction subunit is used to construct the hypothetical premise that both the candidate answer and the alternative answer are valid; An evidence validity evaluation subunit, used to re-evaluate the validity of each piece of evidence in the original reasoning chain under the assumptions; The evidence screening subunit is used to screen out evidence that is still valid; A re-reasoning subunit, used to perform re-reasoning based only on valid evidence; The consistency check subunit is used to check whether the result of the re-reasoning is consistent with the candidate answer; Output the consistency verification results to the iteration control module.

8. Financial consulting system, characterized in that, The system includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to implement the financial consultation response method as described in any one of claims 1-7.

Images