Government information consultation system based on large language model

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses a government information consultation system based on a large language model, which comprises a user interaction module, a large language model core engine, a knowledge base integration module, a multi-level authority management module, a feedback optimization mechanism and a risk control module. The comprehensive intelligent government affair service system is constructed through the six core modules, remarkable advantages are shown in government affair service digital transformation, the system innovatively adopts multi-mode interactive design, multiple input and output modes of voice, text and images are supported, an intelligent authority management mechanism is matched, and the intelligent authority management mechanism is matched with the intelligent authority management mechanism. According to the technical scheme, the accessibility and convenience of government affair services are greatly improved, precise services for different user groups are achieved, it is guaranteed that sensitive data are safe and controllable while wide spreading of government affair information is guaranteed, and a large language model of a system core is subjected to professional government affair scene optimization training and is combined with a dynamically-updated knowledge graph technology.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to a government information consultation system based on a large language model; Background Art

[0002] The government information consultation system based on a large language model (LLM) is a next-generation government service platform built on advanced pre-trained models such as GPT and Claude, deeply integrating government knowledge bases with intelligent interaction technologies. This system uses natural language understanding to accurately analyze user inquiries and, in conjunction with a real-time policy database, automatically generates standardized and accurate responses. While existing government information consultation systems based on large language models (LLMs) offer significant advantages in improving service efficiency, they face multiple challenges in practical application. Technically, these systems lag behind in updating specialized domain knowledge, such as the system still recommending outdated processes after policy adjustments. They also lack complex logical reasoning capabilities, making it easy to miss conditions in multi-step inquiries. Support for multimodal interactions is limited, requiring additional integrated modules to process non-text information. Data-wise, there is a significant risk of sensitive information leakage, and data bias and discrimination are prominent issues, such as low response rates for inquiries in minority languages ​​and incomplete responses due to data silos. Management-wise, there is a difficulty balancing manual intervention with model updates, a lack of emergency response mechanisms, and delayed policy updates in emergencies. Furthermore, support for multiple languages ​​and dialects is inadequate. At the user level, answers lack interpretability, easily leading to trust crises; emotional interaction is weak; and the lack of age-friendly and barrier-free design neglects the needs of the elderly and disabled. Responsibility is unclear, and over-reliance on LLM may weaken the capabilities of grassroots government officials, which is detrimental to the health of the service ecosystem. Collaborative improvements are needed across multiple aspects, including technical optimization, data governance, improved management mechanisms, user-centered design, and the establishment of an ethical framework, to unleash the potential of LLM in government consultation. Summary of the Invention

[0003] The purpose of the present invention is to provide a government information consultation system based on a large language model to solve the problems raised in the above background technology; To achieve the above objectives, the present invention provides the following technical solutions: a government information consultation system based on a large language model, the system comprising a user interaction module, a large language model core engine, a knowledge base integration module, a multi-level authority management module, a feedback optimization mechanism, and a risk control module; The user interaction module is used to receive natural language queries input by the user, supports text, voice and multimodal input, and returns a structured response; the large language model core engine is based on a pre-trained large language model, and is fine-tuned to adapt to government affairs scenarios, with policy interpretation, legal query, and service process guidance functions; the knowledge base integration module is used to dynamically link government public databases, policy document libraries and public service knowledge graphs to ensure the authority and timeliness of information; the multi-level authority management module provides differentiated information access rights based on user identity; the feedback optimization mechanism continuously optimizes model output through user satisfaction scores and manual review logs; the risk control module is mainly used for intelligent protection of sensitive information, automatic flow of cross-departmental work orders, and full-process compliance audits to comprehensively ensure data security, improve government efficiency and ensure legality and compliance.

[0004] Preferably, the specific steps of the fine-tuning method of the large language model core engine are as follows: Step 1: Build a training set for government vertical fields To build a training set for the government vertical domain, we first need to collect government Q&A pairs, policy white papers, and historical consultation records from government portals, government transparency platforms, and the 12345 hotline. Key terms are extracted and desensitized. Then, the data is cleaned through a rule engine and manual review to filter out invalid or outdated information. Government experts then classify the intent of the questions and annotate key entities. Finally, the processed data is converted into JSON or TFRecord format and divided into training, validation, and test sets in a 7:2:1 ratio to form structured training data. Step 2: RLHF-based intent recognition optimization In the initial phase, LLaMA-3 / GPT-4 was fine-tuned using a government training set to provide a preliminary understanding of government knowledge. Experts then rated the model's various responses on a scale of 1-5 based on accuracy, completeness, and clarity. A reward model was trained based on these scores. Finally, a PPO reinforcement learning algorithm was employed, using the reward model scores as an optimization signal to iteratively adjust the generation strategy of the large language model. This focused on improving the accuracy of parsing fuzzy queries related to "elderly subsidies," ensuring that responses more closely aligned with government scenarios. Step 3: Lightweight Adaptation Layer Deployment and Policy Update Lightweight adaptation layer technology is used to achieve efficient updates of government models: a low-rank adaptation matrix is ​​inserted next to the attention layer of the basic model, and only 0.1%-1% of the parameters are trained. When a new policy is released, the policy text is converted into LoRA adaptation data for incremental training, which greatly reduces the computing cost. After verifying the effect through A / B testing, it is hot deployed to the production environment, and independent LoRA modules are maintained for different policy versions, supporting switching and rollback in seconds.

[0005] Preferably, the operation process of the knowledge base integration module is: 1. Data Access and Analysis The knowledge base integration module first automatically captures public data from governments at all levels through APIs, including policy documents and service guidelines. It then uses OCR to parse PDFs / scans, combines NLP to extract key information, and cleans and converts unstructured data into a standard JSON format. 2. Knowledge Graph Construction and Update A dynamic knowledge graph is constructed based on extracted government data. Its core consists of three entities: policy clauses, service agencies, and application materials. By establishing "basis" relationships and "process precondition" relationships, a structured knowledge network is formed. The system has a daily automatic detection mechanism to track changes in government data sources at all levels in real time and dynamically update the knowledge graph content to ensure the timeliness and accuracy of policy clauses and service process information, providing reliable knowledge support for intelligent consultation. The entire process realizes the automated organization and continuous evolution of government knowledge. 3. Answer Generation and Tracing When generating answers, the system automatically associates policy entities in the knowledge graph, and marks the standard citation format and original text link at the end of the answer. For complex policy cross-references, it intelligently generates an expandable "Related Policies" folding panel to facilitate users to review the complete basis.

[0006] Preferably, the multi-level authority management module is implemented by adopting an intelligent authority classification mechanism through the government information consultation system to provide precise services for different user groups; public users directly access basic policy consultation and routine service process guidance to meet daily government service needs; corporate users need to pass a strict authentication process for industrial and commercial business license information verification, and after passing the authentication, they can unlock exclusive corporate service modules, including industry tax preferential policy analysis, special subsidy application guidelines, and business license processing green channel in-depth service content; government staff must pass the government intranet dual identity authentication, and after successful authentication, they can obtain an exclusive knowledge base containing internal operating specifications, cross-departmental collaboration processes, and administrative approval standards core government resources. At the same time, the system will record complete operation logs to ensure information security; this refined authority management system not only ensures the widespread dissemination of government information, but the system also supports a dynamic authority adjustment mechanism to update the access scope of each user group in real time according to policy changes.

[0007] Preferably, the feedback optimization mechanism sets up a two-dimensional scoring system of "answer accuracy" and "service attitude" in the user interaction interface, and each answer requires users to rate it with 1-5 stars; when the score is lower than the preset threshold, the system automatically transfers the question to the government expert team for manual review, and the corrected standard answer will be used as high-quality training data to feed back to the model; at the same time, a multi-level error classification system is established, including six categories: "factual errors", "unclear expressions", "incomplete information", "process errors", "timeliness errors" and "irregular formats"; based on the classification statistical results, the system intelligently identifies the knowledge shortcomings and expression defects of the model, and achieves precise optimization through targeted data supplementation and parameter fine-tuning.

[0008] Preferably, the risk control module adopts a triple protection mechanism to ensure service security; first, a real-time filtering system for sensitive information is deployed, which automatically identifies and intercepts confidential information or personal privacy content through keyword matching and semantic analysis technology, and desensitizes necessary but sensitive information; secondly, when a complex cross-departmental business consultation is detected, the system automatically generates a standardized electronic work order, which is dispatched to the relevant responsible unit through an intelligent routing algorithm, and the processing progress is tracked in real time; finally, all interaction processes are fully recorded, including user questions, system replies, and key information of operation time, to form a structured log file, strictly following the storage specifications of the "Regulations on the Management of Electronic Archives of Government Services".

[0009] Preferably, the system supports multimodal intelligent interaction: voice queries are converted into text through high-precision ASR voice recognition, and the reply content is selected for TTS voice broadcast; it supports uploading ID card pictures, automatically identifies ID card number and unified social credit code key information through CV technology, and intelligently fills in spreadsheets; for complex government processes, the system automatically generates visual flowcharts and timelines to intuitively display the processing steps and time limit requirements; multimodal interaction covers multiple input and output methods such as voice, image, and text, making government services smarter and more convenient.

[0010] Preferably, the system deployment architecture adopts a hybrid cloud model for non-sensitive consultations to be processed by the public cloud to ensure response speed in high-concurrency scenarios; core models and confidential data are deployed in a private cloud, and knowledge security synchronization is achieved through federated learning technology; the domestic SM4 encryption algorithm is used throughout the data transmission process to meet the security and compliance requirements of the government system.

[0011] Preferably, the service implementation scenarios include: Smart guidance terminals are deployed in government service halls to provide self-service; integrated into government official websites and apps to implement online intelligent customer service; and the 12345 hotline is enabled to implement intelligent voice response. At the same time, it serves as a mobile policy think tank for grassroots staff, quickly obtaining accurate policy interpretations through natural language queries, significantly improving government service efficiency and public satisfaction.

[0012] The beneficial effects of the present invention are as follows: 1. The present invention has built a comprehensive intelligent government service system through six core modules, showing significant advantages in the digital transformation of government services. The system innovatively adopts multimodal interactive design, supports multiple input and output methods such as voice, text, and images, and cooperates with the intelligent authority management mechanism. It has greatly improved the accessibility and convenience of government services, and achieved precise services for different user groups. While ensuring the widespread dissemination of government information, it also ensures the security and controllability of sensitive data. The large language model at the core of the system has been optimized and trained for professional government scenarios, combined with dynamically updated knowledge graph technology, it can provide authoritative and accurate policy interpretations and service process guidance. Its innovative feedback optimization mechanism forms a virtuous closed loop of continuous improvement through user ratings and expert reviews. In terms of technical architecture, The system adopts a hybrid cloud deployment model, and through multiple security protection measures, it not only ensures the service responsiveness in high-concurrency scenarios, but also strictly meets the government data security and compliance requirements. It is particularly worth mentioning that the system's built-in intelligent work order system and visual process display function have significantly improved the efficiency of cross-departmental collaboration and government transparency. The system has highly flexible deployment capabilities and can be widely used in various service scenarios such as smart terminals in government halls, government portals, 12345 hotlines, etc., to achieve 7×24 hours of uninterrupted government services. At the same time, as an intelligent policy assistant for grassroots staff, it can quickly obtain accurate policy interpretations through natural language interaction, comprehensively improve government service efficiency from multiple dimensions, optimize people's service experience, and provide strong technical support for building a service-oriented government.

[0013] 2. The present invention realizes the intelligent upgrade of government services through a three-stage fine-tuning method, which has significant technical advantages and application value. First, the system constructs a professional government vertical field training set, collects and cleans data from authoritative channels such as government portals and government open platforms, and has government experts perform professional annotations to ensure the quality and professionalism of the training data, providing a solid knowledge foundation for the model. Secondly, the system innovatively adopts RLHF technology, and trains the reward model through multi-dimensional scoring of the model output by government experts, and then combines the PPO algorithm to optimize the generation strategy, so that the model can accurately understand the government scenario needs of fuzzy queries such as "elderly subsidies", which greatly improves the government The accuracy and practicality of policy interpretation. Finally, the system introduces lightweight adaptation layer technology, which only requires training a very small proportion of model parameters to achieve policy updates. It is quickly deployed after A / B test verification, supports seamless switching of different policy versions, and ensures that government services can respond to policy changes in a timely manner. This fine-tuning method not only significantly reduces the computational cost of model updates, but also maintains high availability of services, enabling the system to continue to provide accurate and timely government services. The entire solution deeply integrates government expertise with large language model capabilities, which not only ensures the authority and accuracy of policy interpretation, but also achieves high efficiency and timeliness of services, providing reliable technical support for the construction of digital government. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Schematic diagram of the system module of the present invention; Figure 2 Schematic diagram of the steps of the fine-tuning method of the large language model core engine of the present invention. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by persons of ordinary skill in the art based on the embodiments of the present invention without creative work shall fall within the scope of protection of the present invention. like Figures 1 to 2 As shown, an embodiment of the present invention provides a government information consultation system based on a large language model, which includes a user interaction module, a large language model core engine, a knowledge base integration module, a multi-level authority management module, a feedback optimization mechanism and a risk control module; The user interaction module is used to receive natural language queries input by the user, supports text, voice and multimodal input, and returns a structured response; the large language model core engine is based on a pre-trained large language model (such as GPT-4, Claude or a domestic model), and is fine-tuned to adapt to government affairs scenarios, with policy interpretation, legal query, and service process guidance functions; the knowledge base integration module is used to dynamically link government public databases, policy document libraries and public service knowledge graphs to ensure the authority and timeliness of information; the multi-level authority management module provides differentiated information access rights based on user identity (public, enterprise, government employee); the feedback optimization mechanism continuously optimizes model output through user satisfaction scores and manual review logs; the risk control module is mainly used for intelligent protection of sensitive information, automatic flow of cross-departmental work orders, and full-process compliance audits to comprehensively ensure data security, improve government efficiency and ensure legality and compliance.

[0016] The government information consultation system, based on a large language model, constructs an intelligent government service platform through six core modules. The system utilizes a multimodal user interface that supports natural language queries and structured response output. The core engine, based on a pre-trained large language model and optimized specifically for government scenarios, provides precise policy interpretation and process guidance. The knowledge base module connects to government data sources in real time, constructing a dynamically updated domain knowledge graph to ensure authoritative and accurate information. The intelligent rights management system provides differentiated service content for different user groups, enabling secure and controllable information sharing. The system innovatively introduces a dual-loop optimization mechanism to continuously improve service quality through user feedback and manual review. The risk control system integrates sensitive information protection, cross-departmental collaboration, and full-process auditing capabilities to comprehensively ensure service security and compliance. This architecture enables intelligent government service throughout the entire process, from information query to business processing, significantly improving the accessibility, accuracy, and security of government services and providing reliable technical support for the development of a digital government.

[0017] The specific steps of the fine-tuning method of the large language model core engine are as follows: Step 1: Build a training set for government vertical fields To construct a training set for the government vertical domain, we first need to collect government Q&A pairs, policy white papers, and historical consultation records from channels such as government portals, government transparency platforms, and the 12345 hotline. Key terms are extracted and desensitized. The data is then cleaned using a rule engine and manual review to filter out invalid or outdated information. Government experts then categorize the question intent (e.g., "policy inquiry" or "process handling") and annotate key entities (such as time and location). Finally, the processed data is converted into JSON or TFRecord format and divided into training, validation, and test sets in a 7:2:1 ratio to form structured training data. Step 2: RLHF-based intent recognition optimization In the initial stage, supervised fine-tuning (SFT) of LLaMA-3 / GPT-4 was performed using a government training set to enable it to initially master government knowledge. Subsequently, experts scored the various answers generated by the model (such as different versions of "applying for social security materials") on a scale of 1-5 based on accuracy, completeness, and clarity. Based on these scores, a reward model was trained. Finally, the PPO reinforcement learning algorithm was adopted, with the reward model score as the optimization signal, to iteratively adjust the generation strategy of the large language model, focusing on improving the accurate parsing ability of fuzzy queries such as "elderly subsidies" to make the answers more in line with the needs of government scenarios.

[0018] Step 3: Lightweight Adaptation Layer (LoRA) deployment and policy update The lightweight adaptation layer (LoRA) technology is used to achieve efficient updates of government models: a low-rank adaptation matrix is ​​inserted next to the attention layer of the basic model, and only 0.1%-1% of the parameters are trained. When a new policy is released, the policy text is converted into LoRA adaptation data for incremental training, which greatly reduces the computing cost. After verifying the effect through A / B testing, it is hot deployed to the production environment, and independent LoRA modules are maintained for different policy versions, supporting switching and rollback in seconds. This method reduces the time required for policy updates from the traditional several days to hours, while ensuring service stability.

[0019] This large language model fine-tuning method achieves precise government services through a three-stage optimization process: 1) constructing a high-quality government training dataset with expert annotation to ensure data expertise; 2) using reinforcement learning to optimize the model's intent understanding, significantly improving fuzzy query processing; and 3) using LoRa technology to rapidly update policies, significantly reducing model adjustment costs. This approach empowers the government Q&A system with deep policy understanding, enabling it to accurately interpret complex regulations and promptly respond to policy changes. This provides the public with authoritative and reliable intelligent consulting services, effectively improving the efficiency and quality of government services.

[0020] The operation process of the knowledge base integration module is as follows: Data access and analysis The knowledge base integration module first automatically captures public data from governments at all levels through APIs, including policy documents and service guidelines. It then uses OCR to parse PDFs / scans and, combined with NLP, extracts key information (such as terms and dates). It then cleans and converts unstructured data into a standard JSON format, providing structured data support for subsequent knowledge graph construction. This process ensures the real-time and accuracy of government information. Knowledge graph construction and update A dynamic knowledge graph is constructed based on extracted government data. Its core entities include three types: policy clauses (such as Article 12 of the Social Security Law), service agencies (such as the Social Security Bureau), and application materials (such as ID card copies). By establishing "basis" relationships (linking materials and clauses) and "process preconditions" relationships (the order in which business is processed), a structured knowledge network is formed. The system has a daily automatic detection mechanism to track changes in government data sources at all levels in real time, dynamically updating the knowledge graph content to ensure the timeliness and accuracy of information such as policy clauses and service processes, providing reliable knowledge support for intelligent consulting. The entire process realizes the automated organization and continuous evolution of government knowledge. Answer generation and tracing When generating answers, the system automatically associates policy entities in the knowledge graph and marks the standard citation format (such as "

Based on

[0021] The knowledge base integration module ensures the authority and reliability of government information through intelligent processes: 1) Automatically collects public data from governments at all levels, intelligently analyzes unstructured documents, and achieves standardized processing of government information; 2) Build a dynamic knowledge graph that accurately links policy terms, service agencies, and application materials, and updates it in real time to ensure the timeliness of information; 3) Answers are automatically labeled with policy references and provided with links to the original text. Complex questions display a dashboard of related policies. This system significantly improves the accuracy of policy interpretation, enabling efficient organization and intelligent application of government knowledge, and providing the public with authoritative, transparent, and traceable government services.

[0022] Among them, the implementation method of the multi-level authority management module adopts an intelligent authority classification mechanism through the government information consultation system to provide precise services for different user groups; public users directly access basic policy consultation and routine service process guidance to meet daily government service needs; corporate users need to go through strict authentication processes such as industrial and commercial business license information verification. After passing the authentication, they can unlock exclusive corporate service modules, including industry tax preferential policy analysis, special subsidy application guidelines, business license application green channel and other in-depth service content; government staff must pass the government intranet dual identity authentication (account password + dynamic token). After successful authentication, they can obtain exclusive knowledge base containing core government resources such as internal operating specifications, cross-departmental collaboration processes, administrative approval standards, etc. At the same time, the system will record complete operation logs to ensure information security; this refined authority management system not only ensures the widespread dissemination of government information, but also ensures the security and controllability of sensitive data, achieving a perfect balance between the accessibility of public services and information security; the system also supports a dynamic authority adjustment mechanism, which can update the access scope of each user group in real time according to policy changes.

[0023] The rights management system enables precise service and security control: 1) The public can easily access basic government information, and businesses can unlock exclusive policy interpretations and green channels through authentication, significantly improving service relevance; 2) Government personnel must use dual authentication to access core resources, and complete operation logs ensure the security of sensitive information; 3) A dynamic permissions adjustment mechanism responds to policy changes in real time. This system ensures the inclusive sharing of government information while also protecting confidential data. This effectively improves service satisfaction and reduces complaints, building a "wide coverage, high security" intelligent service system.

[0024] Among them, the feedback optimization mechanism sets up a two-dimensional scoring system of "answer accuracy" and "service attitude" in the user interaction interface, and each answer requires users to rate it with 1-5 stars; when the score is lower than the preset threshold, the system automatically transfers the question to the government expert team for manual review, and the corrected standard answer will be used as high-quality training data to feed back to the model; at the same time, a multi-level error classification system is established, including six categories: "factual errors", "unclear expressions", "incomplete information", "process errors", "timeliness errors" and "irregular formats"; based on the classification statistical results, the system can intelligently identify the model's knowledge shortcomings and expression defects, and achieve precise optimization through targeted data supplementation, parameter fine-tuning and other methods.

[0025] The feedback optimization mechanism continuously improves service quality through a closed-loop design: 1) A two-dimensional scoring system collects user feedback in real time, and manual review ensures the quality of corrections, forming a closed loop of high-quality training data; 2) Accurately classify six types of errors, intelligently locate knowledge blind spots, and significantly reduce errors such as policy timeliness; 3) Targeted optimization significantly improved model iteration efficiency and steadily increased response accuracy. Since the system went live, the number of complaints has significantly decreased, response times have been significantly shortened, and user satisfaction has remained high, achieving the healthy development of government services that become more intelligent with use.

[0026] Among them, the risk control module adopts a triple protection mechanism to ensure service security; first, a real-time filtering system for sensitive information is deployed, which automatically identifies and intercepts confidential information or personal privacy content through keyword matching and semantic analysis technology, and desensitizes necessary but sensitive information (such as hiding the key field of the ID card number); secondly, when a complex cross-departmental business consultation is detected, the system automatically generates a standardized electronic work order, which is dispatched to the relevant responsible units through an intelligent routing algorithm, and the processing progress is tracked in real time; finally, all interaction processes are fully recorded, including key information such as user questions, system replies, and operation time, to form a structured log file, strictly following the storage specifications of the "Regulations on the Management of Electronic Archives of Government Services" (retention period of not less than 3 years) to ensure that the entire process is traceable and auditable.

[0027] The risk control module comprehensively safeguards government service security through a three-tiered protection mechanism: 1) An intelligent filtering system intercepts confidential information in real time, employing semantic analysis and field desensitization technology to ensure zero leakage of sensitive data; 2) Standardized work orders are automatically generated for cross-departmental tasks, with intelligent assignment and progress tracking, improving processing efficiency by 60%; and 3) Full operational traceability is maintained, with structured log storage meeting a three-year archiving requirement, enabling complete auditability. This mechanism has reduced the incidence of security incidents to zero and shortened work order processing time by 50%. It also fully complies with regulations such as the "Measures for the Management of Electronic Archives for Government Services," mitigating data risks while improving cross-departmental collaboration efficiency.

[0028] Among them, the system supports multimodal intelligent interaction: voice queries are converted into text through high-precision ASR voice recognition, and the reply content can be selected as TTS voice broadcast; it supports uploading ID card pictures, automatically identifying key information such as ID number and unified social credit code through CV technology, and intelligently filling in spreadsheets; for complex government processes, the system automatically generates visual flowcharts and timelines to intuitively display the processing steps and time limit requirements; multimodal interaction covers multiple input and output methods such as voice, image, and text, making government services smarter and more convenient.

[0029] Significantly improve the government service experience through multimodal interaction: 1) Automatic speech recognition (ASR) and text-to-speech (TTS) technologies allow people who are not good at typing to easily handle business through voice, expanding the service coverage; 2) Intelligent ID image recognition automatically extracts key fields, improving form filling efficiency and reducing error rates; 3) Visualization of complex processes makes the steps clear at a glance, reducing the cost of understanding for the public by 70%. Multimodal support has steadily improved service response speed, particularly benefiting digitally disadvantaged groups such as the elderly, increasing public satisfaction and truly achieving the goal of intelligent services that are "understandable at a glance, masterable at a glance."

[0030] Among them, the system deployment architecture adopts a hybrid cloud model for non-sensitive consultations to be processed by the public cloud, ensuring the response speed in high-concurrency scenarios; the core model and confidential data are deployed in the private cloud, and knowledge security synchronization is achieved through federated learning technology; the data transmission process adopts the domestic SM4 encryption algorithm to meet the security and compliance requirements of the government system. This architecture not only guarantees service efficiency, but also ensures that sensitive data does not leave the intranet, and the overall security protection meets the level 3 security standard.

[0031] Non-sensitive inquiries are processed through the public cloud, easily handling high-concurrency requests and ensuring rapid service response. Core models and confidential data are exclusively deployed on the private cloud, and federated learning is used to achieve secure knowledge synchronization, strictly ensuring that sensitive information does not leave the intranet. Transmission is encrypted throughout using the SM4 national secret algorithm, meeting the security level 3 of the Information Security Protection Level 3 standard. This solution significantly improves government service efficiency (concurrent processing capacity increases by three times), achieves data security control throughout the entire lifecycle (zero security incidents), and reduces operating costs by 40%. It is a new infrastructure architecture that meets the requirements of the digital transformation of government.

[0032] Among them, its service landing scenarios include: Smart guidance terminals are deployed in government service halls to provide self-service; integrated into government official websites and apps to implement online intelligent customer service; and the 12345 hotline is enabled to implement intelligent voice response. At the same time, it serves as a mobile policy think tank for grassroots staff, quickly obtaining accurate policy interpretations through natural language queries, significantly improving government service efficiency and public satisfaction.

[0033] Smart service terminals enable 24-hour self-service, reducing waiting times. Intelligent customer service on the official website and app provides instant online answers, improving service coverage. Voice recognition significantly improves the efficiency of the 12345 hotline. Grassroots personnel can access precise policy interpretations anytime through mobile devices, ensuring high accuracy in handling affairs. Multi-scenario applications have increased public satisfaction and shortened government processing times, truly achieving the goal of "more data processing, less running around for the public."

[0034] Example 1: System Architecture Example Component modules: User interaction module (supporting multimodal input / output), large language model core engine (fine-tuning for government scenarios), knowledge base integration module (dynamically linking to government databases), multi-level permission management module (differentiated access control), feedback optimization mechanism (two-dimensional scoring optimization), risk control module (triple security protection) Technical features: Through modular design, the entire government consultation process is intelligent, and each module works together to ensure service authority, security, and availability. Example 2: Model fine-tuning example Step 1: Construct a government training set (collect data from the government platform → clean and annotate → divide into training set / validation set / test set) Step 2: RLHF optimization (SFT fine-tuning → expert scoring training reward model → PPO reinforcement learning to improve fuzzy query parsing) Step 3: LoRA dynamic update (insert low-rank matrix → incremental training new policy → hot deployment after A / B testing) Technical Effect: Policy analysis accuracy increased to 89%, and model update time was shortened by 90%. Example 3: Knowledge base construction example Process 1: Data Access (API captures government public data → OCR / NLP parses unstructured files → standardizes to JSON) Process 2: Knowledge Graph Construction (Define three entities: Policy Terms / Offices / Materials → Establish "Based on" and "Process Precedence" relationships → Automatically update daily) Process 3: Answer tracing (automatically annotating policy basis → generating an expandable related policy panel → providing original text links) Technical advantage: realizing automatic association and dynamic tracking of policy terms Example 4: Rights Management Example Public users: Open basic policy consultation and process guidance Enterprise users: Unlock special services such as tax incentives / subsidy declaration after industrial and commercial certification Government personnel: Intranet two-factor authentication to access internal operating manuals / cross-departmental processes Innovation: Supports dynamic adjustment of permissions and real-time synchronization of access scope after policy changes Example 5: Feedback Optimization Example Rating mechanism: Users rate "accuracy" and "service attitude" with a 1-5 star rating Optimization process: Low-scoring answers trigger manual review → Corrected results for incremental training → Targeted optimization based on 6 types of errors Example 6: Risk Control Example Sensitive information protection: keyword matching + semantic analysis to achieve real-time desensitization Cross-department collaboration: Automatically generate work orders → Intelligent routing → Progress tracking Compliance audit: Complete interaction logs → structured storage (in compliance with the "Administrative Measures for Electronic Archives of Government Services") Security indicators: zero leakage of sensitive information, improved work order processing efficiency Example 7: Multimodal interaction example Voice interaction: ASR to text + TTS voice broadcast Image recognition: CV technology automatically extracts key fields from documents Visual output: Generate dynamic flowcharts / timelines for complex processes User experience: Service satisfaction increased by 35 percentage points Example 8: Deployment Architecture Example Hybrid cloud architecture: Public cloud handles non-sensitive queries (high concurrency), private cloud deploys core model (federated learning synchronization) Encryption scheme: SM4 algorithm ensures data transmission security Compliance: Meet the requirements of Level 3 security protection Example 9: Application scenario example Smart terminal: self-service guidance in government affairs hall Online customer service: government official website / APP integration Hotline enhancement: 12345 intelligent voice answering Policy Think Tank: Mobile Query for Grassroots Staff Implementation effect: Consultation response time shortened by 70% It should be noted that, in this document, relational terms such as first and second, etc., are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations; moreover, the terms "include", "comprise" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or apparatus; While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A government information consultation system based on a large language model, characterized by: The system includes a user interaction module, a large language model core engine, a knowledge base integration module, a multi-level authority management module, a feedback optimization mechanism and a risk control module; The user interaction module is used to receive natural language queries input by the user, supports text, voice and multimodal input, and returns a structured response; the large language model core engine is based on a pre-trained large language model, and is fine-tuned to adapt to government affairs scenarios, with policy interpretation, legal query, and service process guidance functions; the knowledge base integration module is used to dynamically link government public databases, policy document libraries and public service knowledge graphs to ensure the authority and timeliness of information; the multi-level authority management module provides differentiated information access rights based on user identity; the feedback optimization mechanism continuously optimizes model output through user satisfaction scores and manual review logs; the risk control module is mainly used for intelligent protection of sensitive information, automatic flow of cross-departmental work orders, and full-process compliance audits to comprehensively ensure data security, improve government efficiency and ensure legality and compliance.

2. The government information consultation system based on a large language model according to claim 1 is characterized by: The specific steps of the fine-tuning method of the large language model core engine are as follows: Step 1: Build a training set for government vertical fields To build a training set for the government vertical domain, we first need to collect government Q&A pairs, policy white papers, and historical consultation records from government portals, government transparency platforms, and the 12345 hotline. Key terms are extracted and desensitized. Then, the data is cleaned through a rule engine and manual review to filter out invalid or outdated information. Government experts then classify the intent of the questions and annotate key entities. Finally, the processed data is converted into JSON or TFRecord format and divided into training, validation, and test sets in a 7:2:1 ratio to form structured training data. Step 2: RLHF-based intent recognition optimization In the initial phase, LLaMA-3 / GPT-4 was fine-tuned using a government training set to provide a preliminary understanding of government knowledge. Experts then rated the model's various responses on a scale of 1-5 based on accuracy, completeness, and clarity. A reward model was trained based on these scores. Finally, a PPO reinforcement learning algorithm was employed, using the reward model scores as an optimization signal to iteratively adjust the generation strategy of the large language model. This focused on improving the accuracy of parsing fuzzy queries related to "elderly subsidies," ensuring that responses more closely aligned with government scenarios. Step 3: Lightweight Adaptation Layer Deployment and Policy Update Lightweight adaptation layer technology is used to achieve efficient updates of government models: a low-rank adaptation matrix is ​​inserted next to the attention layer of the basic model, and only 0.1%-1% of the parameters are trained. When a new policy is released, the policy text is converted into LoRA adaptation data for incremental training, which greatly reduces the computing cost. After verifying the effect through A / B testing, it is hot deployed to the production environment, and independent LoRA modules are maintained for different policy versions, supporting switching and rollback in seconds.

3. The government information consultation system based on a large language model according to claim 1 is characterized by: The operation process of the knowledge base integration module is as follows:

1. Data Access and Analysis The knowledge base integration module first automatically captures public data from governments at all levels through APIs, including policy documents and service guidelines. It then uses OCR to parse PDFs / scans, combines NLP to extract key information, and cleans and converts unstructured data into a standard JSON format.

2. Knowledge Graph Construction and Update A dynamic knowledge graph is constructed based on extracted government data. Its core consists of three entities: policy clauses, service agencies, and application materials. By establishing "basis" relationships and "process precondition" relationships, a structured knowledge network is formed. The system has a daily automatic detection mechanism to track changes in government data sources at all levels in real time and dynamically update the knowledge graph content to ensure the timeliness and accuracy of policy clauses and service process information, providing reliable knowledge support for intelligent consultation. The entire process realizes the automated organization and continuous evolution of government knowledge.

3. Answer Generation and Tracing When generating answers, the system automatically associates policy entities in the knowledge graph, and marks the standard citation format and original text link at the end of the answer. For complex policy cross-references, it intelligently generates an expandable "Related Policies" folding panel to facilitate users to review the complete basis.

4. The government information consultation system based on a large language model according to claim 1 is characterized in that: The multi-level authority management module is implemented by using an intelligent authority classification mechanism through the government information consultation system to provide precise services for different user groups; Public users can directly access basic policy consultation and routine service process guidance to meet daily government service needs; corporate users need to pass a strict authentication process for industrial and commercial business license information verification. After passing the authentication, they can unlock the exclusive corporate service module, including industry tax preferential policy analysis, special subsidy application guidelines, and in-depth service content of the business license application green channel; government staff must pass the government intranet dual identity authentication. After successful authentication, they can obtain an exclusive knowledge base containing internal operating specifications, cross-departmental collaboration processes, and administrative approval standards. Core government resources, at the same time, the system will record complete operation logs to ensure information security; this refined permission management system not only ensures the widespread dissemination of government information, but the system also supports a dynamic permission adjustment mechanism to update the access scope of each user group in real time according to policy changes.

5. The government information consultation system based on a large language model according to claim 1 is characterized in that: The feedback optimization mechanism sets up a two-dimensional rating system of "answer accuracy" and "service attitude" in the user interaction interface, and each answer requires users to rate it with 1-5 stars; When the score falls below a preset threshold, the system automatically forwards the question to a team of government experts for manual review. The revised standard answer will be fed back into the model as high-quality training data. A multi-level error classification system will also be established, including six categories: "factual errors," "unclear statements," "incomplete information," "process errors," "timeliness errors," and "irregular formatting." Based on the classification statistical results, the system intelligently identifies the knowledge gaps and expression defects of the model, and achieves precise optimization through targeted data supplementation and parameter fine-tuning.

6. The government information consultation system based on a large language model according to claim 1 is characterized by: The risk control module uses a triple protection mechanism to ensure service security. First, it deploys a real-time sensitive information filtering system that automatically identifies and intercepts confidential information or personal privacy content through keyword matching and semantic analysis technology, and desensitizes necessary but sensitive information. Secondly, when a complex cross-departmental business inquiry is detected, the system automatically generates a standardized electronic work order, assigns it to the relevant responsible unit through an intelligent routing algorithm, and tracks the processing progress in real time; finally, all interaction processes are fully recorded, including user questions, system responses, and key information on operation time, forming a structured log file.

7. The government information consultation system based on a large language model according to claim 1 is characterized in that: The system supports multimodal intelligent interaction: voice queries are converted into text through high-precision ASR voice recognition, and replies are broadcast using TTS voice. It supports uploading ID images, automatically identifying key information such as ID numbers and unified social credit codes through CV technology, and intelligently filling in spreadsheets. For complex government processes, the system automatically generates visual flowcharts and timelines to intuitively display the processing steps and time requirements. Multimodal interaction covers multiple input and output methods such as voice, images, and text, making government services smarter and more convenient.

8. The government information consultation system based on a large language model according to claim 1 is characterized by: The system deployment architecture adopts a hybrid cloud model for non-sensitive consultations to be processed by the public cloud, ensuring response speed in high-concurrency scenarios; core models and confidential data are deployed in the private cloud, and knowledge security synchronization is achieved through federated learning technology; The domestic SM4 encryption algorithm is used throughout the data transmission process to meet the security and compliance requirements of government systems.

9. The system according to any one of claims 1 to 8, characterized in that: Its service implementation scenarios include: Smart guidance terminals are deployed in government service halls to provide self-service; integrated into government official websites and apps to implement online intelligent customer service; and the 12345 hotline is enabled to implement intelligent voice response. At the same time, it serves as a mobile policy think tank for grassroots staff, quickly obtaining accurate policy interpretations through natural language queries, significantly improving government service efficiency and public satisfaction.

Images