A method and apparatus for determining a reply to a question, an electronic device, and a storage medium

Abstract

This application provides a method, apparatus, electronic device, and storage medium for determining responses to questions, belonging to the technical field of question response determination, and applicable to scenarios in financial, medical, health, and elderly care insurance application systems. The method includes: responding to a question input by a user, performing semantic and compliance constraints on the question, and generating a structured control vector corresponding to the question; inputting the structured control vector into multiple perturbation decoding heads, and generating candidate response sequences corresponding to the business semantic distribution of the question and each perturbation decoding head based on the structured control vector; wherein the business semantic distributions of the multiple perturbation decoding heads are different; and determining the target response sequence for the question based on the candidate response sequences. This application improves the diversity and accuracy of responses by constructing a structured control vector through pre-generation semantic and compliance constraints, and by generating candidate responses in parallel using perturbation decoding heads with differentiated business semantic distributions.

Description

Technical Field

[0001] This application relates to the field of problem response determination technology, and in particular to a method, apparatus, electronic device and storage medium for problem response determination. Background Technology

[0002] In management systems and applications within the financial and insurance sector, features such as intelligent question answering, product recommendations, claims processing, and risk control assistance are common. Large Language Models (LLMs) have become a core supporting technology, significantly improving business processing efficiency and service response speed, meeting the industry's demand for intelligent and efficient services. Currently, LLM applications in financial and insurance scenarios cover the entire process from customer consultation to claims processing, becoming a crucial foundation for the industry's digital transformation.

[0003] Existing methods for determining the answer to a problem rely on large-scale sampling or multi-step reasoning, and employ single-generation or beam search strategies, which can easily lead to getting stuck in local optima and result in poor accuracy in answering the problem. Summary of the Invention

[0004] The main objective of this application is to provide a method, apparatus, electronic device, and storage medium for determining the answer to a problem, aiming to solve the problem of poor accuracy in answering existing problems.

[0005] To achieve the above objectives, a first aspect of this application proposes a method for determining the response to a problem, the method comprising:

[0006] In response to receiving a question input by the user, semantic and compliance constraints are applied to the question, and a structured control vector corresponding to the question is generated; The structured control vector is input into multiple perturbation decoding heads, and each perturbation decoding head generates a candidate response sequence for the question corresponding to the business semantic distribution of each perturbation decoding head based on the structured control vector; wherein the business semantic distributions of the multiple perturbation decoding heads are different; Based on the candidate response sequence, the target response sequence for the question is determined.

[0007] In some embodiments, the structured control vector includes task intent, business domain, set of constraint rules, response style preference, and risk level. The step of applying semantic and compliance constraints to the question to generate the structured control vector corresponding to the question includes: Based on the questions to be answered, determine the task intent, the business domain, and the set of constraint rules; The reply style preference is determined based on the user's user information identifier and historical question and answer data; The risk level is determined based on the questions to be answered, the historical Q&A data, and the business area.

[0008] In some embodiments, determining the response style preference based on the user's user information identifier and historical question-and-answer data includes: Based on the user information identifier, adapt the corresponding static profile data; Based on the historical question-and-answer data, the corresponding historical behavior data is calculated; The response style preference is calculated based on the static profile data and the historical behavior data.

[0009] In some embodiments, determining the risk level based on the question to be answered, the historical question-and-answer data, and the business domain includes: Analyze the questions to be answered to obtain the corresponding semantic risk factors; Based on the aforementioned business areas, match the corresponding risk weights for those areas; Extract user complaint rate, historical proportion of high-risk questions, and claims dispute records from the historical Q&A data; Based on the user complaint rate, the historical proportion of high-risk questions, and the claims dispute records, historical risk behavior characteristics are generated. The risk level is calculated based on the semantic risk factor, the risk weight, and the historical risk behavior characteristics.

[0010] In some embodiments, the plurality of perturbation decoding heads are constructed based on the same basic model parameters, and the decoding parameters of each perturbation decoding head are different.

[0011] In some embodiments, determining the target response sequence for the question based on the candidate response sequence includes: Based on the set of constraint rules and the response style preference, calculate the compliance index, information completeness index, and style matching index for each candidate response sequence; The overall score for each candidate response sequence is calculated based on its compliance index, information completeness index, and style matching index. The target response sequence for the question is determined based on the candidate response sequence whose overall score is higher than a preset score.

[0012] In some embodiments, determining the target response sequence for the question based on the candidate response sequence whose overall score is higher than a preset score includes: Candidate response sequences with an overall score higher than a preset score are selected as alternative response sequences. The candidate response sequence is divided into a retention part and a rewritten part; the retention part is the part that remains unchanged; the rewritten part includes at least one of a rule reference part, a risk warning part, and a vague liability judgment part. Based on the set of constraint rules and the response style preference, the rewritten part is optimized to obtain the optimized rewritten part; The target response sequence is obtained by combining the preserved portion and the optimized rewritten portion.

[0013] To achieve the above objectives, a second aspect of this application provides a problem response determining apparatus, the apparatus comprising: The response module is used to respond to a question input by the user, apply semantic and compliance constraints to the question, and generate a structured control vector corresponding to the question; A decoding module is used to input the structured control vector into multiple perturbation decoding heads, and to generate candidate response sequences for the question corresponding to the business semantic distribution of each perturbation decoding head based on the structured control vector; wherein the business semantic distributions of the multiple perturbation decoding heads are different; A determining module is used to determine the target response sequence for the question based on the candidate response sequence.

[0014] To achieve the above objectives, a third aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0015] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0016] The method, apparatus, electronic device, and storage medium for determining responses to questions proposed in this application, in response to receiving user input questions, apply semantic and compliance constraints to the questions to generate a structured control vector corresponding to the questions. The structured control vector is then input into multiple perturbation decoding heads, each of which generates a candidate response sequence corresponding to the business semantic distribution of the question based on the structured control vector. The business semantic distributions of the multiple perturbation decoding heads are different. Based on the candidate response sequences, a target response sequence for the questions is determined. This improves the diversity and accuracy of responses by constructing structured control vectors through pre-generation semantic and compliance constraints and by generating candidate responses in parallel using perturbation decoding heads with differentiated business semantic distributions, thereby enhancing the structured control vectors constructed from the pre-generation semantic and compliance constraints. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for determining the response to the problem provided in the embodiments of this application; Figure 2 yes Figure 1 The flowchart of step S101 in the text; Figure 3 yes Figure 2 The flowchart of step S202 in the text; Figure 4 yes Figure 2 The flowchart of step S203 in the process; Figure 5 yes Figure 1 The flowchart of step S103 in the process; Figure 6 yes Figure 5 The flowchart of step S503 in the process; Figure 7 This is a schematic diagram of the structure of the problem response determination device provided in the embodiments of this application; Figure 8 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0021] In management systems and applications within the financial and insurance sector, features such as intelligent question answering, product recommendations, claims processing, and risk control assistance are common. Large Language Models (LLMs) have become a core supporting technology, significantly improving business processing efficiency and service response speed, meeting the industry's demand for intelligent and efficient services. Currently, LLM applications in financial and insurance scenarios cover the entire process from customer consultation to claims processing, becoming a crucial foundation for the industry's digital transformation.

[0022] Existing methods for determining the answer to a problem rely on large-scale sampling or multi-step reasoning, and employ single-generation or beam search strategies, which can easily lead to getting stuck in local optima and result in poor accuracy in answering the problem.

[0023] Based on this, embodiments of this application provide a method, apparatus, electronic device, and storage medium for determining the response to a problem, aiming to improve the accuracy of responses to existing technical problems.

[0024] The problem response determination method, apparatus, electronic device, and storage medium provided in this application are specifically described through the following embodiments. First, the problem response determination method in this application is described.

[0025] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0026] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0027] The method for determining the answer to a problem provided in this application relates to the technical field of problem answer determination and can be applied to scenarios in financial, medical, health, and elderly care insurance application systems. The method for determining the answer to a problem provided in this application can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the method for determining the answer to a problem, but is not limited to the above forms.

[0028] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0029] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0030] Figure 1 This is an optional flowchart of the method for determining the response to the problem provided in the embodiments of this application. Figure 1The method may include, but is not limited to, steps S101 to S103.

[0031] Step S101: In response to receiving a question input by the user, semantic and compliance constraints are applied to the question, and a structured control vector corresponding to the question is generated; Step S102: The structured control vector is input into multiple perturbation decoding heads, and each perturbation decoding head generates a candidate response sequence for the question corresponding to the business semantic distribution of each perturbation decoding head based on the structured control vector; wherein the business semantic distributions of the multiple perturbation decoding heads are different. Step S103: Determine the target response sequence for the question based on the candidate response sequence.

[0032] In this step, upon receiving the user's input question, the unstructured question is transformed into a structured control vector that can guide the subsequent generation process through semantic parsing and compliance constraints. This lays the foundation for the quality and compliance of the entire generation process. The structured control vector includes the task intent, business domain, set of constraint rules, response style preference, and risk level.

[0033] Specifically, upon receiving a user's input question, semantic parsing is first performed to accurately extract the task intent from the question and clarify the user's core need type. Simultaneously, the corresponding business domain is matched to determine the specific insurance category to which the question belongs. Then, a retrieval mechanism is used to obtain a set of constraint rules corresponding to the user's input question and business domain. This set covers relevant business terms and industry regulatory rules, ensuring that the generation process has a clear compliance basis.

[0034] Furthermore, by combining static user profiles and dynamic behavioral data, a style vector adapted to the current user and business scenario is formed to construct response style preferences. In addition, the risk level corresponding to the user's input question is determined by combining an automatic risk assessment model with a rule engine or by user pre-setting, specifically by comprehensively considering the semantic risk factors of the user's question, the preset risk coefficient of the corresponding business domain, and the user's historical risk behavior characteristics, to quantify the prior risk intensity of the question in the dimensions of compliance, responsibility, and financial risk.

[0035] Furthermore, the task intent, business domain, set of constraint rules, response style preferences, and risk level are integrated to form a structured control vector, thereby achieving dual constraints on pre-generation semantics and compliance for subsequent generation processes.

[0036] Furthermore, the structured control vector is input into multiple perturbation decoders. Each perturbation decoder generates a candidate response sequence corresponding to the business semantic distribution of the question based on the structured control vector. The multiple perturbation decoders are based on the original parameters of the same basic large language model. The network structure and pre-training parameters of the model are not modified. The differentiated perturbation decoders are constructed by adjusting the key parameters in the decoding stage. Each decoder corresponds to a unique business semantic distribution, and the decoding parameters of each perturbation decoder are different.

[0037] Specifically, each perturbation decoder generates a complete candidate response sequence in parallel based on its own business semantic distribution. This parallel design of multiple decoders enables the coverage of multiple business-feasible solutions beyond the optimal language probability, thereby improving the diversity and business adaptability of candidate responses.

[0038] Furthermore, based on the candidate response sequence, the target response sequence for the question is determined. First, target responses that meet business requirements are screened and optimized from multiple candidate response sequences. Specifically, for each candidate response sequence, three types of core evaluation indicators are calculated: compliance indicator (statistical percentage of responses that satisfy the set of constraint rules), information completeness indicator (calculation of the proportion of responses that hit the required information slots in the business domain), and style matching indicator (calculation of the degree of fit between the response and the preset style preference through vector fit).

[0039] Furthermore, by using preset weighting coefficients to weight and sum the three types of indicators, a comprehensive business score for each candidate response sequence is obtained. A preset number of candidate response sequences are then selected based on their scores, from highest to lowest, completing the initial narrowing down of business feasibility. Next, while maintaining the original semantic core of the candidate response sequences, targeted optimization operations are performed to address issues such as unclear clause citations, missing risk warnings, ambiguous expressions, and redundant content. The resulting processed responses are the target response sequences. These targeted optimization operations include supplementing clause support, completing risk warnings, eliminating ambiguity, and compressing redundant information, ensuring that the responses are both logically rigorous and concisely accurate.

[0040] For example, in a management system or application software for the financial business sector, taking a car insurance claims consultation scenario as an example, when a user inputs the question: "I have purchased vehicle damage insurance, and my car was flooded in a rainstorm. The repair cost is 80,000 yuan. Can I get a full refund?", the system first responds to the question and extracts the core information through semantic parsing: the task intent is "car insurance claim amount consultation," and the business domain is clearly "commercial car insurance." Then, it searches and matches the "vehicle damage insurance liability" and "flood damage claim conditions" in the "Model Clauses for Commercial Motor Vehicle Insurance" and the regulatory rules of relevant departments regarding property insurance claims. A set of constraint rules is formed; then the initial risk level is calculated: the problem involves a claim amount of 80,000 yuan (high semantic risk factor), the risk coefficient of the auto insurance business domain is preset to 0.6 (medium), and the user has no history of complaints or claims disputes (low historical risk behavior characteristics). Combining the risk discrimination model and the rule engine, the initial risk level is determined to be "medium risk". In addition, in terms of style preference, combining the user's identity (individual car owner, channel is the official APP) and historical interaction data (past preference for concise and clear replies), a style preference vector of "stable and compliant + concise and easy to understand" is generated. Finally, the above information is integrated to generate a structured control vector: {Task Intent: Auto insurance claim amount consultation, Business Domain: Commercial auto insurance, Set of Constraint Rules: Vehicle damage insurance water damage claim clauses + Property insurance regulatory rules, Style Preference: Stable and compliant + concise and easy to understand, Initial Risk Level: Medium Risk}.

[0041] Furthermore, the structured control vector is input into three perturbation decoding heads respectively to generate candidate response sequences in parallel. Among them, the three perturbation decoding heads are differentiated perturbation decoding heads constructed based on the same basic large language model parameters. The decoding parameters and business semantic distribution of each perturbation decoding head are as follows: Perturbation decoding head 1 (compliance priority type), with a temperature of 0.3 (low randomness), a sampling threshold of 0.8 (high probability token screening), a style vector of "enhanced compliance expression", and a risk bias vector of "reducing the probability of absolute statements about full compensation". The business semantic distribution focuses on "compliance priority, clarifying the preconditions for claims". Perturbation decoding head 2 (detailed information type), with a temperature of 0.6 (medium randomness), a sampling threshold of 0.7, a style vector of "complete information coverage", and a risk bias vector of "balancing risk warnings and process descriptions". The business semantic distribution focuses on "detailed breakdown of claims process and material requirements". Perturbation decoding head 3 (concise and efficient type), with a temperature of 0.4, a sampling threshold of 0.9, a style vector of "extremely simple expression", and a risk bias vector of "prioritizing core conclusions". The business semantic distribution focuses on "quickly providing claims conclusions and key precautions".

[0042] Furthermore, multiple candidate response sequences were obtained from the outputs of the three perturbation decoding heads, including the following: The first candidate response sequence is "According to Article X of the Model Clauses for Commercial Motor Vehicle Insurance, vehicle damage insurance is liable for compensation for vehicle damage caused by rainstorms, but only if the vehicle was not restarted after being submerged in water. If you did not restart the vehicle and have reported the damage as required, the repair costs can be paid according to the actual loss within the insurance amount. The specific amount is subject to the damage assessment result, and full compensation cannot be directly promised."; The second candidate response sequence is "Hello! Your situation falls within the scope of vehicle damage insurance liability, and..." The following steps are required to file a claim: 1. Confirm that the vehicle was not restarted after being submerged in water; 2. Contact the insurance company to report the incident (within 48 hours); 3. Cooperate with the loss assessor's inspection; 4. Provide repair invoices, loss assessment reports, and other materials. If all conditions are met, the repair cost of 80,000 yuan will be compensated according to the actual loss within the insurance amount. The final compensation amount is subject to the loss assessment conclusion. Please keep relevant vouchers. The third candidate response sequence is: "Vehicle flooding due to heavy rain is within the scope of vehicle damage insurance claims. There is no compensation for restarting. The 80,000 yuan needs to be confirmed by the loss assessment. If the conditions are met, compensation will be paid according to the actual loss. It is recommended to report the incident within 48 hours." Based on multiple candidate response sequences, a multi-dimensional screening and refined closed-loop approach is used to determine the target response.

[0043] Specifically, the core indicators for each candidate response sequence were first calculated, and the results are as follows: Regarding compliance, the first candidate response sequence scored 0.98, the second 0.95, and the third 0.92. This indicates that each candidate response sequence meets the constraint rules, with the first candidate response sequence scoring the highest due to its explicit citation of clauses. Regarding information completeness, the second candidate response sequence scored 0.96, the first 0.85, and the third 0.70. This indicates that the second candidate response sequence covers both the process and materials, providing the most complete information. Regarding style matching, the third candidate response sequence scored 0.90, the first 0.88, and the second 0.82. This indicates that the third candidate response sequence aligns with the preference for "concise and easy to understand."

[0044] Furthermore, by using preset weighting coefficients (compliance 0.4, information completeness 0.3, style matching degree 0.3) to calculate the comprehensive score, the comprehensive score of the first candidate response sequence is 0.92, the comprehensive score of the second candidate response sequence is 0.91, and the comprehensive score of the third candidate response sequence is 0.84. The first candidate response sequence and the second candidate response sequence are then selected to enter the refinement stage.

[0045] Furthermore, during the refinement process, the original semantic core was maintained, and a risk warning regarding "whether the insurance amount is adequately insured" was added. Redundant statements in the comprehensive scoring of the second candidate response sequence were compressed, ultimately generating the target response: "According to Article X of the 'Model Clauses for Commercial Motor Vehicle Insurance,' vehicle flooding caused by heavy rain falls under the liability of vehicle damage insurance. If your vehicle was not restarted after being submerged and has adequately insured for vehicle damage, the repair cost of 80,000 yuan can be compensated based on the actual loss (the final amount is subject to the damage assessment result). Please contact the insurance company to report the incident within 48 hours, cooperate with the damage assessment, and retain repair invoices, damage assessment reports, and other supporting documents to avoid affecting the claim due to late reporting or missing materials."

[0046] For example, in the management system or application software for the medical, health and pension insurance business, taking the critical illness insurance claims consultation scenario as an example, when a user enters the question: "I have purchased critical illness insurance and have been diagnosed with early-stage lung cancer. Can I apply for critical illness compensation? How much will the compensation be?", the task intent is extracted through semantic parsing as "critical illness insurance claim eligibility and amount consultation", and the business field is "commercial health insurance (critical illness insurance)". Then, the definition of "lung cancer" in the "Standardized Use of Disease Definitions for Critical Illness Insurance (2020 Edition)", the terms of the purchased critical illness insurance contract (including critical illness compensation ratio and waiting period requirements) and relevant departmental rules on health insurance claims supervision are retrieved and matched to form a set of constraint rules.

[0047] Furthermore, the initial risk level is calculated. Based on the question involving critical illness claims (high semantic risk factor), the risk coefficient for the critical illness business domain of health insurance is preset to 0.7 (medium-high risk). The user has been insured for at least 2 years (waiting period expired, no history of fraudulent claims). Combining the model and rule engine, the initial risk level is calculated as "medium-high risk." Additionally, regarding style preferences, the user is a middle-aged or elderly group (static profile), and historical interactions repeatedly requested "simple explanations of professional terminology" (dynamic behavioral data), generating a style preference vector of "professional rigor + simple interpretation." Finally, a structured control vector is generated: {Task Intent: Critical Illness Insurance Claim Eligibility and Amount Consultation, Business Domain: Commercial Health Insurance (Critical Illness Insurance), Constraint Rule Set: 2020 Critical Illness Definition + Critical Illness Insurance Contract Terms + Health Insurance Regulatory Rules, Style Preference: Professional Rigor + Simple Interpretation, Initial Risk Level: Medium-High Risk}.

[0048] Furthermore, the structured control vector is input into three perturbation decoding heads respectively to generate candidate response sequences in parallel. Among them, the three perturbation decoding heads are differentiated perturbation decoding heads constructed based on the same basic large language model parameters. The decoding parameters and business semantic distribution of each perturbation decoding head are as follows: Perturbation decoding head 1 (precise terms type), with a temperature of 0.2 (low randomness), a sampling threshold of 0.85, a style vector of "adaptation to the original text of the terms", and a risk bias vector of "strengthening the boundary prompts of critical illness definition", and a business semantic distribution focusing on "strictly adhering to the critical illness definition and contract terms to clarify the eligibility for claims"; Perturbation decoding head 2 (popular interpretation type), with a temperature of 0.5 (medium randomness), a sampling threshold of 0.75, a style vector of "converting terminology into popular terms", and a risk bias vector of "balancing professionalism and ease of understanding", and a business semantic distribution focusing on "converting the critical illness definition and compensation rules into popular language"; Perturbation decoding head 3 (process guidance type), with a temperature of 0.4, a sampling threshold of 0.8, a style vector of "clear claims process", and a risk bias vector of "highlighting material and timeliness requirements", and a business semantic distribution focusing on "clarifying the claims application steps, required materials, and time nodes".

[0049] Furthermore, multiple candidate response sequences were obtained from the output of the three perturbation decoding heads, including the following: The first candidate response sequence is "According to the 'Standardized Use of Disease Definitions in Critical Illness Insurance (2020 Edition)', early-stage lung cancer (TNM stage I non-small cell lung cancer) is covered by critical illness insurance. Your critical illness insurance contract stipulates that the payout ratio is 100% of the basic sum insured. If you have passed the waiting period and are diagnosed with a condition that meets the definition in the policy, you can apply for full critical illness compensation, the specific amount of which is based on your basic sum insured. Claims require the provision of a pathological diagnosis report, medical records, and other materials."; The second candidate response sequence is "Hello! Early-stage lung cancer (stage I non-small cell lung cancer) is covered by critical illness insurance. Simply put, as long as you have purchased insurance..." If, after the waiting period, the diagnosed lung cancer meets the national definition of critical illness, you will receive full compensation according to the basic sum insured stipulated in the contract. For example, if your basic sum insured is 500,000 yuan, you will receive 500,000 yuan. When applying, you need to prepare the hospital's pathology report, medical records, and other relevant documents. The third candidate response sequence is: "Your diagnosed early-stage lung cancer meets the critical illness insurance coverage conditions (must meet the TNM stage I non-small cell lung cancer standard). After the waiting period, you can apply for compensation, and the compensation amount will be the full basic sum insured. The claims process is: 1. Report the incident to the insurance company (within 3 days); 2. Submit the pathology report, medical records, ID card, bank card, and other materials; 3. The insurance company reviews the claim (10-15 working days); 4. Payment is made after the review is approved." This process involves multi-dimensional screening and refinement of multiple candidate response sequences to determine the target response.

[0050] First, the core indicators for each candidate response sequence were calculated, and the specific results are as follows: Regarding compliance, the first candidate response sequence scored 0.97, the second 0.93, and the third 0.95. This indicates that the first candidate response sequence strictly adheres to the terms and the definition of critical illness, resulting in the highest score. Regarding information completeness, the first candidate response sequence scored 0.90, the second 0.86, and the third 0.94. This indicates that the third candidate response sequence covers the process, materials, and timeliness, providing the most complete information. Regarding style matching, the first candidate response sequence scored 0.80, the second 0.95, and the third 0.88. This indicates that the second candidate response sequence better aligns with the preference for "popular interpretation," resulting in the highest score.

[0051] Furthermore, by using preset weighting coefficients (compliance 0.4, information completeness 0.3, style matching degree 0.3) to calculate the comprehensive score, the comprehensive score of the first candidate response sequence is 0.91, the comprehensive score of the second candidate response sequence is 0.91, and the comprehensive score of the third candidate response sequence is 0.92. The second and third candidate response sequences are then selected to enter the refinement stage.

[0052] Furthermore, during the refinement process, while maintaining the original semantic core, a risk warning was added stating "It is necessary to confirm whether the TNM stage in the pathology report is stage I non-small cell lung cancer." The plain language of the second candidate response sequence and the clarity of the process in the third candidate response sequence were integrated, redundant statements were compressed, and the final target response was generated: "According to the 'Standardized Use of Disease Definitions in Critical Illness Insurance (2020 Edition),' stage I non-small cell lung cancer with a TNM stage falls within the coverage of critical illness insurance. If you have passed the waiting period for your critical illness insurance and the diagnosis meets the above definition, you can apply for full compensation according to the basic sum insured stipulated in the contract (e.g., if the basic sum insured is 500,000 yuan, the compensation will be 500,000 yuan). When applying, please contact the insurance company to report the case within 3 days, and prepare materials such as the pathology report, medical records, ID card, and bank card. The insurance company will complete the review within 10-15 working days, and pay the compensation after the review is approved. Friendly reminder: Please confirm the TNM stage and cell type of the lung cancer in the pathology report to ensure that it meets the claim conditions stipulated in the terms and conditions."

[0053] Steps S101 to S103 of this embodiment involve responding to a user-inputted question by applying semantic and compliance constraints to the question and generating a structured control vector corresponding to the question. The structured control vector is then input into multiple perturbation decoders, each of which generates a candidate response sequence corresponding to the business semantic distribution of the question based on the structured control vector. The business semantic distributions of the multiple perturbation decoders are different. Based on the candidate response sequences, a target response sequence for the question is determined. This process, by constructing a structured control vector through pre-generation semantic and compliance constraints and combining the parallel generation of candidate responses with perturbation decoders exhibiting differentiated business semantic distributions, improves the diversity and accuracy of responses.

[0054] In some implementation methods, an implementation system is constructed using the method for determining the response to the application problem. The system consists of five core modules, which are not parallel and isolated from each other, but rather form a strict time sequence and feedback loop. Specifically, these modules are: the task and preference parsing module (Pre-Control), the broad exploration generation module (BEG, Decoding Control), the multi-objective pruning module (MOC, Post-Selection), the refinement and optimization module (FRO, Local Refinement), and the preference evolution and policy update module (PGEM, Global Feedback).

[0055] Among them, the task and preference parsing module controls the "semantic space before generation", the extensive exploration generation module controls the "sampling distribution during generation", the multi-target pruning module controls the "business feasibility of candidates entering the next stage", the fine-tuning and optimization module performs fine-grained rewriting and optimization of the final solution, and the preference evolution and strategy update module performs long-term evolution and update of global preferences, weights and parameters. The five modules constitute a complete "generation control closed loop".

[0056] Specifically, the task and preference parsing module (pre-generation control) is functionally defined as being used to process the user's original input questions. Transform it into a structured control vector that can be directly called by subsequent generation and pruning modules. This module serves as the pre-control entry point for the entire system to constrain the generation process. The task and preference parsing module takes a user's natural language question as input, such as, "Will I receive full compensation for my car accident?", and outputs: ; in, For the purpose of the task (claims consultation, terms and conditions interpretation, etc.). For business domains (auto insurance, life insurance, health insurance, etc.) For the retrieved set of terms and regulations, These are expression style vectors obtained from user requests or system learning. Initial risk level (low / medium / high).

[0057] Specifically, in In the case of the expression style vector obtained by the system learning, Determined by both static profiling and dynamic behavior learning, its expression is as follows: ; in, The expression style vector requested by the user or organization serves as an important basis for subsequent generation, distribution control, and scoring. Templates for user identity, channel, and organization communication. Data derived from historical interactions, frequency of follow-up inquiries, complaint rates, and other behavioral patterns. This is an adjustable fusion coefficient.

[0058] Specifically, the initial risk level characterizes the prior risk intensity of the current user's issue in terms of compliance, responsibility, and financial risk dimensions, and is jointly determined by the risk assessment model and the rule engine. Specifically, the initial risk level is determined by the issue's semantic risk factor. Business domain risk coefficient Table and user's historical risk behavior characteristics The three types of information are calculated together.

[0059] Among them, the semantic risk factor of the problem (Output from the classification model) is used to characterize whether it involves monetary compensation, whether it involves claim rejection and exemption, and whether it involves regulatory red lines (such as "full compensation" or "bypassing review"); business domain risk coefficient. Table (manually pre-configured), as shown in Table 1; User historical risk behavior characteristics It is used to characterize the complaint rate, the proportion of high-risk questions in a user's historical Q&A, and the record of claims disputes.

[0060] Table 1

[0061] Furthermore, the formula for calculating the initial risk level is as follows: ; in, The initial risk level, The weighting factor for the semantic risk factor of the problem. For semantic risk factors of the problem, As a weighting factor for the risk coefficient of the business domain, For business domain risk coefficient, As a weighting factor for users' historical risk behavior characteristics, This refers to the user's historical risk behavior characteristics.

[0062] Furthermore, the initial risk level affects the token-level probability distribution during the BEG decoding probability reconstruction process, thus directly "lowering the generation probability of high-risk token representations," as expressed below: ; in, For the first Each decoder at time The token probability distribution For the first Temperature of each decoder head For the first Style bias vector of each decoder head This is the risk penalty item obtained by mapping the initial risk level.

[0063] Furthermore, extensive exploration of Generative BEG (Generation Process Control, not Post-processing) is conducted to use it in the same generation task. By constructing multiple differentiated decoding distributions, the generation process itself is controlled in parallel and semantically expanded. It does not modify the model weights, but directly affects the probability distribution shape in the decoding stage, which belongs to "Decoding Control".

[0064] Specifically, the system is based on the same basic model parameters. Construct K perturbation decoding heads, where each perturbation decoding head is based on: ; in, For the first Candidate response sequence, For the user's original input problem, For the first Equivalent perturbation control parameters for each perturbation decoder head The perturbation strength is used to control the degree of semantic divergence. It follows a Gaussian distribution.

[0065] Specifically, the engineering equivalent implementation methods include different temperatures. Different top-p / top-k vectors, different semantic style vectors In addition to different risk warning bias vectors, these control terms directly change the probability distribution of the sequence (token) during the decoding stage, thereby constructing multiple complete candidate solutions with significantly different semantic directions during the "generation process". The specific token probability distribution expression during the decoding stage is as follows: ; in, For the first A decoder in The token probability distribution at time step. For the first One token, For the front -1 token, For the original model A decoder in The token probability distribution at time step. For the first A set of disturbance decoding control parameters For the first The temperature of a moment.

[0066] Specifically, basic model parameters The relationships with temperature, top-p, style vector, and risk bias are as follows: ; in, Used to control the randomness of generation. Used to control the quality of sampling truncation probability. Used to control tone and style (conservative / explanatory / marketing). Used to control the intensity of risk warnings.

[0067] Furthermore, the multi-target pruning module MOC (first business constraint contraction) applies each of the BEG outputs. (Complete natural language response) Perform the first business space trimming. Specifically, first calculate three types of basic scoring indicators, where the first scoring indicator is the compliance score, and its calculation formula is as follows: ; in, This refers to the total number of regulatory rules participating in this compliance verification, i.e., the total number of all rule entries that need to be met for the current business scenario. The rule index number indicates the number of the rules. Specific regulatory or business rules, For the first Candidate response sequence The compliance score ranges from [0, 1], with a higher value indicating a higher level of compliance. For the first Specific compliance rules, such as "no promises of full compensation", "no bypassing of manual review", and "disclaimer clauses must be displayed", etc. For indicator functions, when candidate responses Satisfy the first If a rule is met, the value is 1; otherwise, it is 0.

[0068] Therefore, for the first 1 reply, put it in all The compliance status of each business activity with compliance rules is normalized and statistically analyzed to obtain an overall compliance score. If all rules are satisfied, then If only half of the rules are satisfied, then .

[0069] Furthermore, the second scoring indicator is information completeness, and its calculation formula is as follows: ; in, For the first Candidate response sequence Information completeness, with a value range of [0, 1]. Candidate Response The subset of required slots that are actually hit in the middle. This is a set of standard information slots that must be covered in the current business scenario, such as: insurance type, scope of liability, compensation ratio, time limit, disclaimer, etc.

[0070] Therefore, the information completeness index is used to measure whether the current response is complete in terms of "key information coverage". For example, if 4 out of 5 slots are required to be covered, then... It is 0.8, if 2 are hit then It is 0.4.

[0071] Furthermore, the third scoring metric is style fit, and its calculation formula is as follows: ; in, For the first Candidate response sequence The style matching score ranges from 100 to 100. In actual systems, normalization is performed to [0, 1]. A style embedding mapping function is used to map text into style feature vectors. This is a target style preference vector, determined by the current business scenario or risk level, such as "conservative," "prudent," or "marketing-friendly." Therefore, the style matching index is used to measure whether the current candidate responses conform to the system's expected style in dimensions such as "language style, tone, and risk expression." For example, for high-risk cases, its... For the purpose of "conservative and prudent", if If the language uses "can guarantee" and "will definitely compensate", then... It will decrease significantly.

[0072] Furthermore, the comprehensive score of each candidate response sequence based on its three basic scoring indicators is calculated using a comprehensive pruning function. The calculation formula is as follows: ; in, For the first The comprehensive score of the candidate response sequence is used as the ranking criterion for entering the subsequent refinement module, FRO. The weighting coefficient for compliance score. The scoring weight coefficient for information completeness. This is the weighting coefficient for the style matching score.

[0073] Specifically, the complete processing flow of the multi-target pruning module (MOC) is as follows: First, it receives K complete candidate responses generated in parallel by the BEG module. Then for each one Calculate their compliance scores separately. Information completeness score and style matching score Then, the overall score is calculated using a comprehensive cropping function. Finally press Sort from high to low, and retain only the top-N items as the "high-quality candidate set" to enter the next module FRO.

[0074] Furthermore, the FRO (Partial Rewrite and Constraint Enhancement) is refined and optimized to enhance clause citations, complete risk warnings, disambiguate vague expressions, and compress redundant content in the pruned Top-N candidates. Notably, the FRO is not regenerated but rather rewritten while maintaining the original semantic core. This ensures that the final output retains the diversity of BEG (Best Before Government) while further meeting the refined requirements of management systems or application software in the financial business domain. Its expression is as follows: ; in, For the refined version of the first Candidate response sequence, For the first Candidate response sequence, For the retrieved set of terms and regulations, A vector for the expressive style required by the user or organization.

[0075] Furthermore, the Preference Evolution and Policy Update (PGEM) (long-term closed loop) updates the pruning weights based on historical task feedback. Style preference vector weights, risk bias weights, and perturbation strength in BEG. And a Top-N strategy. Among them, historical task feedback includes user satisfaction scores, complaint records, manual review and modification content, MOC scoring trajectories for each indicator, and FRO modification magnitude; the update method can adopt reinforcement learning, multi-armed bandit, and user satisfaction regression to achieve an upgrade from "real-time generation control" to "long-term system self-evolution control".

[0076] Therefore, this implementation method forms a full-process substantive control mechanism for response determination by: before response generation, structural constraints on the semantic space are imposed by module pair; during response generation, the probability distribution of the decoding stage is reconstructed in multiple semantic directions by BEG; after response generation, the business target space is pruned and refined by MOC and FRO; and in the long-term operation after response confirmation, the generation strategy is self-evolved and updated by PGEM.

[0077] Please see Figure 2 In some embodiments, the semantic and compliance constraints applied to the problem in step S101 to generate the structured control vector corresponding to the problem may include, but are not limited to, steps S201 to S203: Step S201: Based on the question to be answered, determine the task intent, business domain, and set of constraint rules; Step S202: Determine the reply style preference based on the user's user information identifier and historical Q&A data; Step S203: Determine the risk level based on the questions to be answered, historical Q&A data, and business area.

[0078] In this step, the user's question is first deeply analyzed using natural language semantic parsing technology to extract its task intent and business domain. The task intent focuses on the user's core needs, while the business domain corresponds to specific sub-scenarios within the business domain. By combining keywords or implicit business scenarios in the question, its specific category is determined, providing direction for subsequent rule matching.

[0079] Furthermore, based on the identified business area and task intent, official clauses strongly related to that business area are retrieved, as well as regulatory rules issued by relevant departments for the corresponding area. These two types of rules are integrated to form a set of constraint rules, setting inviolable compliance red lines for the subsequent generation process, ensuring that responses always conform to business norms and regulatory requirements.

[0080] Furthermore, user static profiles and historical Q&A data are retrieved through user information identifiers (such as user ID and account-linked identity information). These user information identifiers include user identity attributes (individual / corporate user, age group, occupation type), business processing channels (official app, offline branches, customer service hotline), and the required script templates from the affiliated institution or partner (e.g., banks selling insurance products must adhere to the bank's service script standards). Historical Q&A data covers past interactions with the system (e.g., preference for brief, concise responses or detailed, process-oriented responses), follow-up question frequency (frequent inquiries about details indicate a need for comprehensive information), and feedback data (e.g., satisfaction scores for past responses, and whether complaints arose due to the user's writing style).

[0081] Furthermore, a basic style benchmark is established using static profiles (e.g., the default style for middle-aged and elderly users is "popular and rigorous", and the default style for enterprise users is "professional and standardized"). The style bias is adjusted using dynamic historical Q&A data (e.g., if users repeatedly ask for details, the "detailed information" attribute is strengthened; if users repeatedly request simplified expression, the "concise and efficient" attribute is strengthened). Ultimately, a unified response style preference is formed to ensure that the subsequent generated responses not only meet the personalized needs of users but also conform to the expression norms of business scenarios.

[0082] Furthermore, the semantic risk factors of the questions to be answered are extracted first. Specifically, this is determined by analyzing whether the question involves monetary compensation (especially large compensation), whether it mentions situations of refusal to pay / exemption, and whether it contains regulatory red lines such as "full compensation" or "bypassing review," thus determining the semantic risk intensity of the question itself. Simultaneously, historical risk behavior characteristics of users are extracted. Specifically, historical Q&A data is used to extract users' past risk-related records, including complaint rates, the historical proportion of high-risk questions (such as inquiries involving violations), dispute records in past claims or inquiries, and records of false information declarations, thus characterizing potential risks at the user level. Additionally, based on pre-set benchmark values ​​from industry data, risk coefficients for the business domain are extracted.

[0083] Furthermore, the three types of risk factors are input into the risk discrimination model for preliminary scoring, and then the scoring results are verified and corrected by the rule engine (if there are multiple complaint records, the risk level is directly increased), and finally the three-level risk level of "low, medium and high" is output, which provides a basis for the subsequent adjustment of the decoding parameters of the disturbance decoder.

[0084] Please see Figure 3 In some embodiments, determining the response style preference based on the user's user information identifier and historical question-and-answer data in step S202 may include, but is not limited to, steps S301 to S303: Step S301: Adapt the corresponding static profile data according to the user information identifier; Step S302: Calculate the corresponding historical behavior data based on historical question-and-answer data; Step S303: Calculate the response style preference based on static profile data and historical behavior data.

[0085] In this step, the user's unique information identifier is used to retrieve and adapt static profile data strongly bound to the user, establishing a basic benchmark for response style preferences. The user information identifier, as the core key for data association, accurately locates the user's identity profile within the system, extracting stable and long-term static attribute information. The static profile data covers user identity attributes, business channels, and the script template requirements of the institution or partner, providing a stable foundation for subsequent style preference calculations.

[0086] Furthermore, dynamic and variable historical behavioral data is extracted from users' past interactions with the system to capture users' personalized needs and changing trends in communication preferences. This historical Q&A data includes all interaction information such as users' past inquiries, system responses, follow-up questions, interaction duration, satisfaction ratings, and complaint or feedback records.

[0087] Furthermore, static profile data is combined with historical behavior data to generate response style preferences that are both general and targeted. Specifically, firstly, adjustable fusion coefficients are assigned to both static profile data and historical behavior data (these coefficients can be optimized based on business scenario requirements or system iteration data). The coefficients corresponding to the static profile data determine the basic direction of the style preference, while the coefficients corresponding to the historical behavior data determine the personalized bias of the style. Then, the two types of data are transformed into feature vectors of a unified dimension. Subsequently, the two types of feature vectors are fused into a unified response style preference vector through a weighted summation of the vectors. This vector is used to clarify the quantitative standards for specific dimensions such as the intensity of the response tone, the proportion of professional terminology used, the information presentation structure, the risk warning method, and the level of detail.

[0088] Please see Figure 4 In some embodiments, determining the risk level in step S203 based on the question to be answered, historical Q&A data, and business domain may include, but is not limited to, steps S401 to S405: Step S401: Analyze the questions to be answered and obtain the corresponding semantic risk factors; Step S402: Based on the business domain, match the risk weight of the corresponding domain; Step S403: Extract user complaint rate, historical proportion of high-risk questions, and claims dispute records from historical Q&A data; Step S404: Generate historical risk behavior characteristics based on user complaint rate, historical proportion of high-risk questions, and claims dispute records; Step S405: Calculate the risk level based on semantic risk factors, risk weights, and historical risk behavior characteristics.

[0089] In this step, intent recognition, keyword extraction, and semantic matching technologies from Natural Language Processing (NLP) are used to analyze the content, core demands, and implicit intent of the question, obtaining quantifiable semantic risk factors. This provides fundamental data support for the subsequent comprehensive calculation of risk levels, reflecting the inherent risk intensity of the question. Specifically, this includes question analysis and classification. Question analysis focuses on whether it directly involves high-risk core elements, whether it involves sensitive business scenarios such as claim denial, exemption, or policy cancellation, and whether it contains expressions that cross regulatory red lines, such as "full compensation," "bypassing review," or "absolute approval." Classification, through semantic intent classification, determines whether the question contains potential risks such as inducing illegal responses or blurring the boundaries of responsibility.

[0090] Furthermore, the risk weight configuration library is first determined according to the business domain to retrieve the risk weights that match the domain. In this way, the influence of semantic risk factors and users' historical risks is corrected through risk weights to ensure that the risk assessment conforms to the actual risk distribution pattern of the industry.

[0091] Furthermore, historical Q&A data (including past consultations, claims applications, feedback complaints, and other full-process interaction records) corresponding to user information identifiers are extracted for targeted extraction and statistics on three core indicators: First, the user complaint rate, which calculates the proportion of historical user complaints to total interactions, directly reflecting the degree of controversy users have had regarding the quality, compliance, or handling results of responses; second, the historical proportion of high-risk questions, which statistically analyzes the proportion of high-risk questions in past user consultations, such as those involving regulatory red lines, consultations on violations, and absolute demands, to characterize whether users have a tendency to frequently test the boundaries of risk; and third, claims dispute records, which retrieve disputed cases from past user claims applications to clarify the user's risk behavior trajectory in actual business scenarios, reflecting the potential risk characteristics of individual users.

[0092] Furthermore, the three types of extracted indicators are standardized to eliminate differences in the dimensions and numerical ranges of different indicators. Based on the weight of each indicator's impact on the overall risk, a weighted fusion algorithm is used to comprehensively calculate the standardized indicators. Finally, a multi-dimensional historical risk behavior feature vector is obtained, which not only includes the quantitative results of each individual indicator but also reflects the synergistic impact between indicators through fusion calculation. This provides core risk basis at the user level for risk level calculation.

[0093] Furthermore, semantic risk factors, business domain risk weights, and historical risk behavior feature vectors are quantified into risk scores under a unified scoring system. Then, through a pre-set comprehensive calculation model, the three types of scores are merged into a preliminary overall risk score, which is then used to activate the rule engine to verify and correct the preliminary score. The final score after verification is then mapped to three risk levels: "low," "medium," and "high," to adjust the probability distribution in the decoding stage (e.g., high-risk levels reduce the probability of absolute statements and the generation probability of illegal statements), thereby achieving proactive risk control.

[0094] Among them, low risk level corresponds to scenarios with no substantial liability / financial risk and no risky user history; medium risk level corresponds to scenarios involving general liability determination and few risky user history; and high risk level corresponds to scenarios involving large-scale financial compensation, regulatory red lines, and frequent high-risk user history.

[0095] Please see Figure 5 In some embodiments, determining the target response sequence for the question based on the candidate response sequence in step S103 may include, but is not limited to, steps S501 to S503: Step S501: Based on the set of constraint rules and response style preferences, calculate the compliance index, information completeness index, and style matching index for each candidate response sequence; Step S502: Calculate the overall score for each candidate response sequence based on its compliance index, information completeness index, and style matching index. Step S503: Determine the target response sequence for the question based on the candidate response sequence whose overall score is higher than the preset score.

[0096] In this step, for compliance indicators, the set of constraint rules is used as the evaluation benchmark to verify whether each candidate response meets the requirements of each rule; for information completeness indicators, the focus is on the set of standard information slots corresponding to the business domain, and the ratio of the number of required slots actually hit in the candidate responses to the total number of slots is calculated; for style matching indicators, with response style preference as the target, the candidate responses and style preferences are converted into feature vectors respectively, and the degree of fit between the vectors is calculated.

[0097] Furthermore, based on the weight coefficients of the three preset indicators, a weighted summation algorithm is used to calculate the overall score, namely, "the overall score equals (the product of compliance indicator and compliance weight plus the product of information completeness indicator and completeness weight plus the product of style matching indicator and style weight), thereby ensuring that the overall score can accurately reflect the comprehensive performance of candidate responses in the three key dimensions of compliance, information, and style, and provide a unified evaluation standard for subsequent screening."

[0098] Furthermore, based on a preset score threshold, candidate responses with an overall score higher than the threshold are selected to form a "high-quality candidate pool." If there is only one response in the high-quality candidate pool, it directly enters the refinement stage; if there are multiple responses, they can be further sorted by overall score, and the Top-N responses (N is set according to business needs) are selected for subsequent processing. The preset score threshold is determined based on industry compliance standards, business quality benchmarks, and historical optimization data.

[0099] Furthermore, during the refinement and optimization process, while maintaining the original semantic core of the candidate responses, targeted optimization operations were performed to address issues such as unclear clause citations, missing risk warnings, ambiguous wording, and redundant content. These operations included enhancing clause citations, supplementing risk warnings, disambiguating ambiguous wording, and compressing redundant content. Ultimately, the responses that have undergone screening and refinement constitute the target response sequence. This sequence ensures basic quality through scoring and screening while addressing minor deficiencies through refinement and optimization, guaranteeing that the output fully aligns with the compliance requirements, business needs, and user preferences of management systems or application software in the financial and insurance sector.

[0100] Please see Figure 6 In some embodiments, determining the target response sequence for the question based on the candidate response sequence with an overall score higher than a preset score in step S503 may include, but is not limited to, steps S601 to S604: Step S601: Select candidate response sequences with an overall score higher than a preset score as alternative response sequences; Step S602: The candidate response sequence is split into a retainable part and a rewritten part; the retainable part is the part that remains unchanged; the rewritten part includes at least one of the following: rule reference part, risk warning part, and ambiguous liability judgment part. Step S603: Based on the set of constraint rules and response style preferences, optimize the rewritten part to obtain the optimized rewritten part; Step S604: Combine the preserved part and the optimized rewritten part to obtain the target response sequence.

[0101] In this step, by comparing the overall score of each candidate response sequence with a preset score, candidate responses with scores higher than a threshold are directly identified as candidate response sequences. Then, for the key areas to be optimized in the candidate response sequences, "local controlled rewriting" is implemented to avoid semantic deviations and efficiency losses caused by full modification. The splitting logic follows the principle of "keeping the semantic core unchanged and precisely optimizing risk areas".

[0102] Specifically, the content in the candidate response sequence that is semantically clear, compliant, stylistically appropriate, and requires no adjustment is designated as the "retained" section. This includes core conclusions, basic process descriptions, and unambiguous statements. This section retains its original wording to ensure the core meaning of the response remains unchanged. Then, the key risks and compliance-related areas requiring optimization are designated as the "rewritten" section, ensuring optimization directly addresses the core issues. The rewritten section focuses on rule citations (such as unclear clause citations or content lacking specific rule basis), risk warnings (such as missing risk warnings, vague wording, or content not covering key risk points), and ambiguous liability determinations (such as ambiguous, unresolved, or potentially misleading statements regarding liability determination).

[0103] Furthermore, regarding the rule citation section, based on the set of constraint rules, specific clause numbers and regulatory bases are added to clarify the applicable scenarios of the rules and avoid vague expressions such as "based on relevant clauses." For the risk warning section, based on the risk weights of the business domain and semantic risk factors, any missing risk points are supplemented, and the tone of the warnings is optimized to ensure that the risk warnings are comprehensive and targeted. For the section on ambiguous liability judgments, combining constraint rules and business logic, ambiguity in expression is eliminated, and the boundaries of liability are clarified. At the same time, the optimization process strictly follows response style preferences, retaining standardized expressions and strengthening logical coherence to ensure that the rewritten parts are both compliant and accurate, and consistent with the overall style.

[0104] Furthermore, semantic coherence is verified between the retained and optimized rewritten parts to ensure that the two parts are connected naturally without logical gaps, forming a complete target response sequence. Ultimately, the target response sequence not only meets the strong compliance requirements of management systems or application software in the financial and insurance field, but also conforms to the user's style preferences, while maintaining semantic integrity and coherence, thus achieving the response optimization goal of "high quality, low risk, and personalization".

[0105] Figure 7 This is a schematic diagram of the structure of the problem response determination device provided in the embodiments of this application. Please refer to it. Figure 7 This application embodiment also provides a problem response determination device 700, which can implement the above-mentioned problem response determination method. The problem response determination device 700 includes: The response module 701 is used to respond to the user input question, perform semantic and compliance constraints on the question, and generate the structured control vector corresponding to the question; The decoding module 702 is used to input the structured control vector into multiple perturbation decoding heads, and to generate candidate response sequences for the question corresponding to the business semantic distribution of each perturbation decoding head based on the structured control vector; wherein the business semantic distributions of the multiple perturbation decoding heads are different; The determination module 703 is used to determine the target response sequence for the question based on the candidate response sequence.

[0106] In some embodiments, the response module 701 includes: The first module is used to determine the task intent, business domain, and set of constraint rules based on the question to be answered; The second module is used to determine the user's response style preference based on the user's user information identifier and historical Q&A data; The third module is used to determine the risk level based on the questions to be answered, historical Q&A data, and business areas.

[0107] In some embodiments, the response module 701 further includes any one of the following: The adaptation module is used to adapt the corresponding static profile data based on the user information identifier; The fourth module is used to calculate the corresponding historical behavior data based on historical question and answer data; The fifth module is used to calculate the response style preference based on static profile data and historical behavior data.

[0108] In some embodiments, the response module 701 further includes any one of the following: The analysis module is used to analyze the questions to be answered and obtain the corresponding semantic risk factors; The matching module is used to match risk weights for corresponding business domains. The extraction module is used to extract user complaint rates, the historical proportion of high-risk questions, and claims dispute records from historical Q&A data; The generation module is used to generate historical risk behavior characteristics based on user complaint rates, the historical proportion of high-risk questions, and claims dispute records. The sixth module is used to calculate the risk level based on semantic risk factors, risk weights, and historical risk behavior characteristics.

[0109] In some embodiments, the determining module 703 includes: The seventh module is used to calculate the compliance index, information completeness index, and style matching index for each candidate response sequence based on the set of constraint rules and response style preferences. The eighth module is used to calculate the overall score of each candidate response sequence based on the compliance index, information completeness index, and style matching index of each candidate response sequence. The ninth module is used to determine the target response sequence for a question based on the candidate response sequence whose overall score is higher than a preset score.

[0110] In some embodiments, the determining module 703 further includes any one of the following: The tenth module is used to select candidate response sequences with an overall score higher than a preset score as alternative response sequences. The splitting module is used to split the candidate response sequence into a retainable part and a rewritten part; the retainable part is the part that remains unchanged; the rewritten part includes at least one of the following: a rule reference part, a risk warning part, and a vague liability judgment part. The optimization module is used to optimize the rewritten part based on the set of constraint rules and response style preferences, so as to obtain the optimized rewritten part; The combination module is used to combine the preserved part and the optimized rewritten part to obtain the target response sequence.

[0111] The specific implementation of the response determination device 700 for this problem is basically the same as the specific embodiment of the response determination method for the above problem, and will not be described again here.

[0112] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method for determining the response to the aforementioned problem. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0113] Please see Figure 8 , Figure 8 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 801 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 802 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 802 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 802, and the processor 801 calls and executes the problem response determination method of the embodiments of this application. The 803 input / output interface is used to implement information input and output. The communication interface 804 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 805 transmits information between various components of the device (e.g., processor 801, memory 802, input / output interface 803, and communication interface 804); The processor 801, memory 802, input / output interface 803, and communication interface 804 are connected to each other within the device via bus 805.

[0114] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for determining the response to the aforementioned problem.

[0115] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0116] The problem response determination method, device, electronic device, and storage medium provided in this application embodiment respond to a user-inputted problem by applying semantic and compliance constraints to the problem and generating a structured control vector corresponding to the problem. The structured control vector is then input into multiple perturbation decoding heads, each generating a candidate response sequence corresponding to the business semantic distribution of the problem and its respective perturbation decoding head based on the structured control vector. The business semantic distributions of the multiple perturbation decoding heads are different. Based on the candidate response sequences, a target response sequence for the problem is determined. This improves the diversity and accuracy of responses by constructing a structured control vector through pre-generation semantic and compliance constraints and by generating candidate responses in parallel using perturbation decoding heads with differentiated business semantic distributions, thereby enhancing the structured control vector structure and the ability to generate candidate responses in parallel.

[0117] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0118] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0119] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; 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.

[0120] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0121] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0122] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0123] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0124] The units described above 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 units can be selected to achieve the purpose of this embodiment according to actual needs.

[0125] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0126] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part 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 multiple 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 of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0127] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for determining the response to a problem, characterized in that, The method includes: In response to receiving a question input by the user, semantic and compliance constraints are applied to the question, and a structured control vector corresponding to the question is generated; The structured control vector is input into multiple perturbation decoding heads, and each perturbation decoding head generates a candidate response sequence for the question corresponding to the business semantic distribution of each perturbation decoding head based on the structured control vector; wherein the business semantic distributions of the multiple perturbation decoding heads are different; Based on the candidate response sequence, the target response sequence for the question is determined.

2. The method for determining the response to the problem according to claim 1, characterized in that, The structured control vector includes task intent, business domain, set of constraint rules, response style preference, and risk level. The process of applying semantic and compliance constraints to the question to generate the corresponding structured control vector includes: Based on the questions to be answered, determine the task intent, the business domain, and the set of constraint rules; The reply style preference is determined based on the user's user information identifier and historical question and answer data; The risk level is determined based on the questions to be answered, the historical Q&A data, and the business area.

3. The method for determining the response to the problem according to claim 2, characterized in that, The step of determining the response style preference based on the user's user information identifier and historical question-and-answer data includes: Based on the user information identifier, adapt the corresponding static profile data; Based on the historical question-and-answer data, the corresponding historical behavior data is calculated; The response style preference is calculated based on the static profile data and the historical behavior data.

4. The method for determining the response to the problem according to claim 2, characterized in that, The step of determining the risk level based on the question to be answered, the historical Q&A data, and the business area includes: Analyze the questions to be answered to obtain the corresponding semantic risk factors; Based on the aforementioned business areas, match the corresponding risk weights for those areas; Extract user complaint rate, historical proportion of high-risk questions, and claims dispute records from the historical Q&A data; Based on the user complaint rate, the historical proportion of high-risk questions, and the claims dispute records, historical risk behavior characteristics are generated. The risk level is calculated based on the semantic risk factor, the risk weight, and the historical risk behavior characteristics.

5. The method for determining the response to the problem according to claim 1, characterized in that, The multiple perturbation decoding heads are constructed based on the same basic model parameters, and the decoding parameters of each perturbation decoding head are different.

6. The method for determining the response to the problem according to claim 1, characterized in that, Determining the target response sequence for the question based on the candidate response sequence includes: Based on the set of constraint rules and the response style preference, calculate the compliance index, information completeness index, and style matching index for each candidate response sequence; The overall score for each candidate response sequence is calculated based on its compliance index, information completeness index, and style matching index. The target response sequence for the question is determined based on the candidate response sequence whose overall score is higher than a preset score.

7. The method for determining the response to the problem according to claim 6, characterized in that, The step of determining the target response sequence for the question based on the candidate response sequence whose overall score is higher than a preset score includes: Candidate response sequences with an overall score higher than a preset score are selected as alternative response sequences. The candidate response sequence is divided into a retention part and a rewritten part; the retention part is the part that remains unchanged; the rewritten part includes at least one of a rule reference part, a risk warning part, and a vague liability judgment part. Based on the set of constraint rules and the response style preference, the rewritten part is optimized to obtain the optimized rewritten part; The target response sequence is obtained by combining the preserved portion and the optimized rewritten portion.

8. A device for determining the answer to a problem, characterized in that, The device includes: The response module is used to respond to a question input by the user, apply semantic and compliance constraints to the question, and generate a structured control vector corresponding to the question. A decoding module is used to input the structured control vector into multiple perturbation decoding heads, and to generate candidate response sequences for the question corresponding to the business semantic distribution of each perturbation decoding head based on the structured control vector; wherein the business semantic distributions of the multiple perturbation decoding heads are different; A determining module is used to determine the target response sequence for the question based on the candidate response sequence.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method for determining the response to the problem as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method for determining the response to the problem as described in any one of claims 1 to 7.

Images