Intelligent question-answering system and method in field of maintenance funds, and electronic equipment

Abstract

The invention relates to the technical field of natural language processing, discloses an intelligent question and answer system and method in the field of maintenance funds and electronic equipment, and aims to solve the problem that an existing intelligent question and answer system is poor in answer accuracy in the field of maintenance funds. The knowledge mapping module is used for constructing and maintaining a structured knowledge graph containing multi-level policy provisions; the mixed retrieval enhancement generation engine is used for receiving user query and retrieving an initial policy knowledge fragment set from the knowledge base by adopting a keyword and vector mixed retrieval mode; the policy conflict processing module is used for carrying out priority ranking and relation identification on the initial set according to a policy level principle and outputting a target policy knowledge fragment sequence subjected to ranking and labeling; and the question-answer generation module is used for generating a final answer based on the target sequence. The question and answer accuracy and compliance in the field of maintenance funds are improved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to an intelligent question-answering system, method, and electronic device in the field of maintenance funds. Background Technology

[0002] As a crucial economic foundation for maintaining the normal operation of communities and ensuring the safety and value of housing, the special maintenance fund for residential buildings is often hailed as the "pension fund" for homes. The compliance and efficiency of its collection, use, management, and supervision directly affect the vital interests of homeowners and social harmony and stability. However, in practice, this sector has long faced a core contradiction: the coexistence of dormant funds and procedural abuse. The essence of this contradiction stems from significant information barriers and policy misunderstandings among various stakeholders (homeowners, property management companies, homeowners' committees, and government departments). Homeowners are unclear about how to apply for and use the fund, property management companies struggle to efficiently streamline the process, and regulatory departments face immense pressure from numerous inquiries. This information asymmetry severely hinders the effectiveness of the maintenance fund policy.

[0003] In the search for solutions, intelligent question-answering systems have been proposed as a potential technological approach. However, existing technological solutions have shown significant limitations in addressing the highly specialized, policy-sensitive, and geographically specific field of maintenance funds, failing to meet actual needs. One issue is the illusion problem of general-purpose large-scale language models in specialized fields. Currently, while general-purpose large-scale models (LLMs) trained on massive amounts of internet data possess powerful language generation capabilities, their internal knowledge has two fatal flaws: First, their training data lacks in-depth, structured professional knowledge in the field of maintenance funds, which makes them prone to illusion when answering specific policy questions. That is, they generate seemingly reasonable but actually fictitious or inaccurate policy clauses and procedural descriptions, which can lead to serious legal risks and misleading information in practical applications.

[0004] Second, traditional policy retrieval systems lack sufficient semantic understanding and correlation capabilities. Traditional policy question-and-answer systems largely rely on keyword matching techniques (such as the BM25 algorithm). This method cannot understand the deep semantics of user queries. For example, a user's question, "Whose money should be used to repair the leaking roof?" involves multiple policy concepts such as "common areas of the building," "emergency use of maintenance funds," and "owner voting." Simply relying on keywords like "roof," "leak," and "money" makes it difficult to accurately and completely retrieve all relevant clauses. While pure vector retrieval technology can perform semantic matching to a certain extent, it may suffer from decreased accuracy due to semantic generalization when dealing with precise proper nouns and accurate policy numbers in policy texts, failing to ensure that core policy evidence is accurately retrieved.

[0005] Third, existing technologies lack the intelligent identification and handling capabilities for multi-level policy conflicts. Maintenance fund management is a typical highly regional matter, with different provinces and cities formulating numerous local implementation rules and regulations within the framework of national higher-level laws. These local policies may supplement higher-level laws in detail, or they may conflict in application under specific circumstances. Existing question-and-answer systems generally lack conflict handling mechanisms based on policy hierarchy theory, resulting in either providing overly general national-level answers lacking practicality when facing cross-regional or complex policy scenarios, or offering local interpretations that contradict the spirit of higher-level laws, thus lacking policy compliance guarantees. Summary of the Invention

[0006] This invention aims to address the problem of poor accuracy in existing intelligent question-answering systems when dealing with maintenance funds, and proposes an intelligent question-answering system, method, and electronic device for the maintenance funds field.

[0007] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides an intelligent question-and-answer system for the field of maintenance funds, the system comprising: The knowledge base management module is used to build and maintain a multi-level policy knowledge base. The multi-level policy knowledge base is a structured knowledge graph built based on the ontology of the maintenance fund field. It integrates policy provisions of multiple different levels, and the policy provisions are parsed into policy entities, policy relationships and policy concept levels and stored in a graph structure. A hybrid retrieval enhancement generation engine, connected to the knowledge base management module, is used to receive user queries and retrieve an initial set of policy knowledge fragments related to the user query from the multi-level policy knowledge base using a hybrid retrieval method that combines keyword retrieval and vector retrieval. The policy conflict handling module, connected to the hybrid retrieval enhancement generation engine, is used to receive the initial set of policy knowledge fragments, and according to the preset policy hierarchy and applicable principle rule set, prioritize, identify conflict and supplementary relationships among the policy knowledge fragments in the initial set of policy knowledge fragments, and output a sorted and conflict-annotated sequence of target policy knowledge fragments. The question-and-answer generation module, connected to the policy conflict handling module, is used to receive the target policy knowledge fragment sequence and generate a final answer based on the target policy knowledge fragment sequence and return it to the user. The user query, the initial set of policy knowledge fragments, the sequence of target policy knowledge fragments, and the final answer flow sequentially to form the main data stream for question and answer processing.

[0008] Furthermore, the execution flow of the hybrid retrieval enhancement generation engine specifically includes: The user query is parsed to generate a corresponding keyword set, and the user query is encoded into a query vector using an Embedding model; Based on the keyword set, a first search is performed in the inverted index associated with the multi-level policy knowledge base to obtain a first search result list; simultaneously, based on the query vector, a similarity search is performed in the vector database associated with the multi-level policy knowledge base to obtain a second search result list. A reciprocal ranking fusion algorithm is used to perform score fusion and reordering on the policy knowledge fragments in the first search result list and the second search result list. The sorted result is then output as the initial set of policy knowledge fragments to the policy conflict processing module.

[0009] Furthermore, the execution flow of the policy conflict handling module specifically includes: The source of each policy knowledge fragment in the initial set of policy knowledge fragments is analyzed, and its policy level is marked according to the policy hierarchy and applicable principle rule set; Based on the definition rules in the policy hierarchy and application principle rule set, the initial set of policy knowledge fragments is sorted, and policy knowledge fragments from the lowest applicable level and with relevant content from the user's location are given the highest priority; the definition rules include: higher-level laws take precedence over lower-level laws, special laws take precedence over general laws, and new laws take precedence over old laws. Compare policy knowledge fragments from different policy levels to determine whether there are supplementary or refining relationships or direct conflicts between them; When the relationship is determined to be supplementary and detailed, the specific provisions of the subordinate law are linked and integrated with the principle provisions of the superior law; when the relationship is determined to be a direct conflict, the conflict points are explicitly marked in the output sequence of target policy knowledge fragments, and prompts are generated to follow the higher-level policy.

[0010] Furthermore, it also includes a self-optimization module, which includes: The data acquisition unit is used to continuously collect user interaction data from the question-and-answer generation module. The user interaction data includes user queries that have not been effectively answered by the system and their session context. Semantic clustering units are used to perform semantic-level clustering analysis on collected user queries using unsupervised clustering algorithms, and to identify high-frequency question clusters that represent blind spots in the system's knowledge. The knowledge base update driving unit is communicatively connected to the knowledge base management module. It is used to generate optimization instructions for the multi-level policy knowledge base based on the high-frequency question cluster, and drive the knowledge base management module to supplement the multi-level policy knowledge base with targeted content.

[0011] Furthermore, the question-and-answer generation module includes: The multi-path generation unit is used to generate multiple candidate answers by calling the underlying large language model multiple times based on the same sequence of policy knowledge fragments; The rule evaluation unit is used to perform rule-based consistency evaluation on the multiple candidate answers, specifically including: Based on the knowledge graph of the multi-level policy knowledge base, key policy entities and question-answering processes are extracted from each candidate answer; Comparative analysis was conducted to assess the similarity of key policy entities and question-answering processes among multiple candidate answers. When the similarity of the key policy entities and question-answering processes of all candidate answers is lower than the first preset threshold, the question is determined to be a complex question, triggering the expert review process. When the similarity of the key policy entities and question-answering processes of multiple candidate answers is higher than the second preset threshold, the multiple similar candidate answers will be output to the answer evaluation unit. The answer evaluation unit is used to score the quality of multiple candidate answers using a pre-trained evaluation model. The optimal aggregation unit is used to select candidate answers whose quality scores are higher than a preset scoring threshold, and calls an aggregation model to integrate and refine the information of the candidate answers to generate the final answer.

[0012] Furthermore, the process of constructing a multi-level policy knowledge base by the knowledge base management module includes: The policy logic definition based on the maintenance fund field includes the policy ontology and the policy relationship ontology. The policy ontology includes: owners, property service companies, owners' committees, and residential special maintenance funds. The policy relationship includes: ownership, management rights, usage rights, and deposit obligations. The collected multi-level policy texts are subjected to automated preprocessing, which includes word segmentation, part-of-speech tagging, and syntactic analysis. Guided by the domain ontology, policy entities are extracted from the preprocessed text using a named entity recognition model, and policy relationships between entities are extracted using a relation extraction model to form policy element triples. The policy element triples are stored in a graph database to construct a structured knowledge graph, and connections are established with the inverted index and vector database to complete the construction of a multi-level policy knowledge base.

[0013] Furthermore, it also includes a case association module, which is used for: When the question-and-answer generation module generates the final answer, it retrieves typical policy dispute cases that are semantically related to the current user's query from the case library linked to the multi-level policy knowledge base. The retrieved typical cases of policy disputes, along with the final answer, will be output to the user.

[0014] Furthermore, it also includes a policy comparison and analysis module, which is used for: Respond to user input indicating the selection of multiple target areas; Extract the differentiated policy provisions for the multiple target regions on the preset key items of maintenance fund management from the multi-level policy knowledge base; Generate and output a policy differentiation comparison report.

[0015] Secondly, the present invention provides an intelligent question-and-answer method in the field of maintenance funds, applied to the intelligent question-and-answer system in the field of maintenance funds described in the first aspect, the method comprising: Construct and maintain a multi-level policy knowledge base, which is a structured knowledge graph based on the ontology of the maintenance fund field. It integrates policy provisions at multiple different levels, and the policy provisions are parsed into policy entities, policy relationships and policy concept levels and stored in a graph structure. The system receives user queries and uses a hybrid retrieval method that combines keyword retrieval and vector retrieval to retrieve an initial set of policy knowledge fragments related to the user query from the multi-level policy knowledge base. Based on a pre-set set of policy hierarchy and applicable principle rules, the policy knowledge fragments in the initial set of policy knowledge fragments are prioritized, and conflict and supplementary relationships are identified to generate a sequence of target policy knowledge fragments that have been sorted and marked with conflict. The final answer is generated based on the sequence of knowledge fragments related to the target policy and returned to the user.

[0016] Thirdly, the present invention provides an electronic device, the electronic device comprising a processor, a memory, and a communication bus; The communication bus is used to enable communication between the processor and the memory; The processor is used to execute one or more programs in the memory to implement the steps of the intelligent question-and-answer method in the field of maintenance funds as described in the second aspect.

[0017] The beneficial effects of this invention are as follows: The intelligent question-answering system, method, and electronic device in the field of maintenance funds provided by this invention, by constructing a structured multi-level policy knowledge base as the source of facts, and by using a hybrid retrieval enhancement generation engine, combines the accuracy of keyword retrieval with the semantic understanding capability of vector retrieval, ensuring that relevant policy basis can be comprehensively and accurately retrieved from the knowledge base, effectively eliminating the illusion phenomenon; furthermore, the policy conflict handling module automatically identifies and resolves the differences and contradictions between policies at different levels, so that the final answer is not only accurate, but also fully conforms to the policy hierarchy requirements, possessing strong practical guidance value, and improving the modernization and intelligence level of maintenance fund management. Attached Figure Description

[0018] Figure 1 A schematic diagram of the structure of an intelligent question-and-answer system in the field of maintenance funds provided in this embodiment; Figure 2 A flowchart illustrating an intelligent question-and-answer method in the field of maintenance funds provided for an embodiment; Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment. Detailed Implementation

[0019] When dealing with maintenance funds, a field characterized by strong regionality and multi-level policies, general intelligent question-answering systems suffer from policy illusions. Traditional retrieval technologies struggle to understand deep semantics, and existing systems completely lack the ability to handle multi-level policy conflicts, resulting in an inability to guarantee the accuracy and practicality of the answers.

[0020] Based on this, the technical solution of this invention is proposed. This invention combines the powerful generation capabilities of a general large language model with structured professional knowledge in the field of maintenance funds, and ensures the accuracy, compliance, practicality, and reliability of the final answer by constructing a conflict resolution mechanism based on policy hierarchy theory.

[0021] Specifically, firstly, the knowledge base management module, based on domain ontology, constructs policies and regulations into a structured knowledge graph—a multi-level policy knowledge base—transforming unstructured text into a semantic network that machines can understand and reason about. Then, the hybrid retrieval enhancement generation engine employs a dual-path parallel retrieval strategy—keyword retrieval and vector retrieval—to overcome the limitations of a single retrieval method. Subsequently, a reciprocal ranking fusion algorithm intelligently merges and reorders the two retrieval results, generating a comprehensive and accurate initial set of policy knowledge fragments. Next, the policy conflict processing module receives the initial set of policy knowledge fragments output by the hybrid retrieval engine and automatically performs hierarchical identification, priority sorting, and relational judgment, outputting a sorted, integrated, and labeled sequence of target policy knowledge fragments. Finally, this sequence of target policy knowledge fragments, used as the sole context and factual basis, is input into the underlying large language model, instructing it to generate the final natural language answer based on these fragments.

[0022] The technical solutions in this embodiment will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0023] Figure 1 A schematic diagram of an intelligent question-and-answer system in the field of maintenance funds is shown below. Figure 1The system includes: a knowledge base management module, a hybrid retrieval enhanced generation engine, a policy conflict handling module, and a question-and-answer generation module. The hybrid retrieval enhanced generation engine is connected to the knowledge base management module, the policy conflict handling module is connected to the hybrid retrieval enhanced generation engine, and the question-and-answer generation module is connected to the policy conflict handling module.

[0024] The knowledge base management module is used to build and maintain a multi-level policy knowledge base. The multi-level policy knowledge base is a structured knowledge graph built based on the ontology of the maintenance fund field. It integrates policy provisions at multiple different levels, and the policy provisions are parsed into policy entities, policy relationships and policy concept levels and stored in a graph structure.

[0025] In practical applications, it is first necessary to collect knowledge sources corresponding to policy provisions at multiple different levels, including: National-level policies and regulations: "Civil Code of the People's Republic of China" and "Administrative Measures for Special Housing Maintenance Funds"; Provincial policies: For example, the "Management Measures for Special Housing Maintenance Funds" formulated by Guangdong Province, Jiangsu Province, Sichuan Province and other places; City-level detailed rules and regulations: For example, the implementation details of maintenance funds in cities such as Chongqing, Beijing, and Guangzhou.

[0026] Based on the collected knowledge sources, the process for constructing a multi-level policy knowledge base in this embodiment includes: Step 101: The policy logic definition based on the maintenance fund field includes the policy ontology and the policy relationship ontology. The policy ontology includes: owners, property service companies, owners' committees and residential special maintenance funds. The policy relationships include: ownership, management rights, usage rights and deposit obligations.

[0027] In the field of maintenance funds, policy reasoning revolves around specific "facts" (such as "roof leaks") and "conditions" (such as "belonging to a common area" and "meeting the conditions for emergency repairs"). The purpose of defining a domain ontology is to abstract the complex policy texts within the domain into a machine-understandable, structured conceptual system (policy ontology) and the logical relationships between them (policy relationships), providing a unified data model for subsequent automated information extraction.

[0028] In practical application, policy experts and knowledge engineers collaborate to extract key concepts based on core regulations such as the Civil Code and the Measures for the Administration of Special Maintenance Funds for Residential Buildings. The core policy entities are clearly defined as "owners" (the subjects of rights and obligations), "property service enterprises" (the entities providing management services), "owners' committees" (the entities responsible for decision-making and implementation), and "special maintenance funds for residential buildings" (the core object). The static logical relationships between these entities are also clarified, such as: "owners" - [ownership] - "maintenance funds"; "owners" - [obligation to deposit] - "maintenance funds"; "property service enterprises" - [management rights] - "maintenance funds" (under specific conditions); "owners' committees" - [right of use] (after decision-making) - "maintenance funds".

[0029] By transforming unstructured policy texts into a structured knowledge framework, the problem of loose structure in traditional knowledge bases is solved. This endows the knowledge graph with preliminary logical reasoning capabilities. For example, the system can deduce "the maintenance fund may be used for rooftop repairs" based on the triples "the owner owns the maintenance fund" and "the rooftop belongs to all owners".

[0030] Step 102: Perform automated preprocessing on the collected multi-level policy texts, including word segmentation, part-of-speech tagging, and syntactic analysis.

[0031] This step utilizes fundamental natural language processing techniques to transform continuous, human-readable policy text into structured language units that are easier for computers to process, and analyzes its grammatical structure, laying the foundation for identifying entities and relationships.

[0032] In practical applications, firstly, Chinese word segmentation tools (such as Jieba and HanLP) are used to divide the policy text into independent lexical units; then, each segmentation result is labeled with its part of speech (such as noun, verb, etc.) to identify potential entities and relationships; finally, dependency parsing models (such as BERT-based models) are used to identify the modification, subject-verb, and verb-object dependencies between words in the sentence. This provides accurate underlying language features for subsequent named entity recognition and relation extraction models.

[0033] Step 103: Under the guidance of the domain ontology, policy entities are extracted from the preprocessed text using a named entity recognition model, and policy relationships between entities are extracted using a relation extraction model to form policy element triples.

[0034] This step maps the specific information in the preprocessed text to the concepts and relationships defined in the ontology, forming standardized (entity-relationship-entity) triples to construct a knowledge graph.

[0035] In practical applications, firstly, a named entity recognition model is used to identify and classify policy entities defined in the policy ontology from the preprocessed text. For example, from the sentence "The Beijing Municipal Detailed Rules stipulate that property service companies should regularly publish their accounts," "Beijing Municipal Detailed Rules" is accurately identified as a policy document entity, and "property service company" is identified as a property service company entity. Then, a relation extraction model is used to determine, based on the syntactic analysis results, what predefined relationship exists between the two identified entities, thereby forming policy element triples. For example, from the sentence "The owners' meeting decides to use the maintenance fund," a triple (owners' meeting, right of use, maintenance fund) is extracted.

[0036] Step 104: Store the policy element triples into a graph database to construct a structured knowledge graph, and establish connections with the inverted index and vector database to complete the construction of a multi-level policy knowledge base.

[0037] In practical applications, triples are stored in a graph database, with each entity being a node and each relation being an edge; the original text or key information of policy knowledge fragments are stored in an inverted index engine; and the policy knowledge fragments are converted into high-dimensional vectors and stored in a vector database, thus completing the construction of a multi-level policy knowledge base.

[0038] Through the above steps, the original, unstructured, and heterogeneous policy texts are transformed into a multi-level policy knowledge base that is clearly structured, machine-understandable, capable of deep reasoning, and supports efficient multi-modal retrieval. This is the foundation for the entire intelligent question-answering system to provide accurate and compliant answers.

[0039] The hybrid retrieval enhancement generation engine receives user queries and uses a hybrid retrieval method that combines keyword retrieval and vector retrieval to retrieve an initial set of policy knowledge fragments related to the user query from the multi-level policy knowledge base.

[0040] In this embodiment, the execution flow of the hybrid retrieval enhancement generation engine specifically includes: Step 201: Parse the user query, generate a corresponding keyword set, and encode the user query into a query vector using the Embedding model.

[0041] User queries contain both literal and semantic logic. By parsing user queries in two different ways, two complementary retrieval paths can be established for subsequent queries.

[0042] In practical applications, on the one hand, keywords are extracted and a sparse vector representation is constructed. For example, for the query "The roof is leaking, whose money should be used to repair it?", the generated keyword set may be: "roof", "leaking", "money", "repair".

[0043] On the other hand, specialized policy domain embedding models (such as M3E) can be used to transform the problem into a high-dimensional dense vector. This model, trained on text related to maintenance funds, can map semantically similar sentences to close positions in the vector space. For example, the vectors generated for "How to use housing pension funds?" and "The usage process of residential special maintenance funds" will be very similar.

[0044] Step 202: Based on the keyword set, perform a first search in the inverted index associated with the multi-level policy knowledge base to obtain a first search result list; simultaneously, based on the query vector, perform a similarity search in the vector database associated with the multi-level policy knowledge base to obtain a second search result list.

[0045] This step employs a dual-path parallel retrieval system, running two retrieval methods simultaneously to balance precision and recall, ensuring no potentially relevant policy knowledge fragments are missed. On one hand, keyword retrieval, based on Boolean models and term frequency statistics, yields a first list of search results. This method accurately matches proper nouns and policy numbers, resulting in highly interpretable results. On the other hand, query vector retrieval, based on semantic similarity, recalls semantically related but literally mismatched fragments, enabling the understanding of complex intents. Keyword retrieval guarantees precise matching of policy terms, while vector retrieval overcomes the lexical gap, understanding the user's true intent and further improving question-answering accuracy.

[0046] In practical applications, the implementation method of keyword retrieval is as follows: the set of related words is queried in the inverted index, and the retrieval system uses algorithms such as BM25 to calculate based on the frequency and distribution of keywords in the document, and returns a list of first search results sorted by relevance score.

[0047] The specific implementation method of vector similarity retrieval is as follows: the query vector is searched in a vector database (such as Milvus, FAISS) using k-NN (k nearest neighbor) search. The system calculates the cosine similarity between the query vector and all policy knowledge fragment vectors in the database, and returns the top k results with the highest similarity to form a second retrieval result list.

[0048] When constructing a multi-level policy knowledge base, each policy knowledge fragment is accompanied by rich metadata, including: issuing authority, policy name, effective date, and applicable region. During the actual search process, the system first identifies the user's city (e.g., through user profile or IP address), then prioritizes relevant policy fragments from the lowest applicable level (i.e., the most basic city / county level) of the user's location, using city-level policies as the highest priority starting point for retrieval. Next, following the priority order of city, province, country, and Civil Code, the system sequentially performs RAG searches in the corresponding knowledge bases. If a city-level policy provides a complete and valid answer, it is used first. If no match is found, the system automatically backtracks to the next higher level (province) to continue searching until the most relevant answer is found. This mechanism ensures that when dealing with regional policy issues, the system always adheres to policy norms that are closest to the user's actual situation.

[0049] Step 203: Using the inverse ranking fusion algorithm, the policy knowledge fragments in the first search result list and the second search result list are fused and reordered by score fusion, and the sorted result is output as the initial set of policy knowledge fragments to the policy conflict processing module.

[0050] Since the BM25 score and cosine similarity score have different dimensions and distributions, they cannot be directly compared or weighted averaged. Therefore, this embodiment employs the Reciprocal Ranking Fusion Algorithm (RRF), which discards absolute scores and relies only on the relative ranking of the results. Fair fusion is achieved by converting the ranking into a new standardized score.

[0051] In practical applications, for each policy knowledge fragment that appears in the first or second search results list, its RRF score is calculated. All unique policy knowledge fragments are sorted from high to low according to their RRF scores, and the final sorted list is output as the initial set of policy knowledge fragments.

[0052] Even if a fragment doesn't rank highly in vector retrieval, as long as it contains all the keywords in the user's query (ranking first in keyword retrieval), its RRF score will still be high, thus boosting it to the top. This ensures a perfect balance between factual accuracy and semantic relevance in the final result.

[0053] The policy conflict handling module receives the initial set of policy knowledge fragments, and, based on a preset set of policy hierarchy and applicable principle rules, prioritizes, identifies conflict and supplementary relationships among the policy knowledge fragments in the initial set of policy knowledge fragments, and outputs a sequence of target policy knowledge fragments that has been sorted and marked with conflict.

[0054] In this embodiment, the execution flow of the policy conflict handling module specifically includes: Step 301: Analyze the source of each policy knowledge fragment in the initial policy knowledge fragment set, and mark its policy level according to the policy hierarchy and applicable principle rule set.

[0055] The policy hierarchy and applicable principle rule set is a mapping table or rule engine that automatically labels policy knowledge fragments with a standardized policy level tag (such as [national level], [provincial level - Jiangsu Province], [municipal level - Chongqing Municipality]) based on the metadata of the policy knowledge fragments.

[0056] Step 302: Based on the definition rules in the policy hierarchy and applicable principle rule set, sort the initial set of policy knowledge fragments and place the policy knowledge fragments from the lowest applicable level of the user's location and those that are relevant to the content in the highest priority.

[0057] The definition rules in the policy hierarchy and applicable principles rule set include: Higher-level laws prevail over lower-level laws: When policies and regulations of different levels have different provisions on the same matter, the policy with higher legal force shall take precedence. Special laws prevail over general laws: When there are both general and special provisions for the same matter, the special provisions take precedence over the general provisions. New law prevails over old law: When new law and old law have different provisions on the same matter, the new law has the effect of the old law.

[0058] Step 303: Compare the content of policy knowledge fragments from different policy levels to determine whether there is a relationship of supplementation or refinement or a direct conflict between them.

[0059] In practical applications, the system performs in-depth semantic comparisons of policy fragments from different levels concerning the same issue. If the provisions of a lower-level law are within the framework of a higher-level law, specifying and quantifying procedures, time limits, standards, etc., and do not violate the basic principles and mandatory provisions of the higher-level law, then it is determined to be a supplementary refinement. For example, the higher-level law, the Civil Code, stipulates "regular publication," while the lower-level law, the Chongqing Municipal Property Management Regulations, stipulates "publication every six months." The latter is a legitimate refinement of the former.

[0060] If the provisions of a subordinate law directly contradict the explicit provisions or core spirit of a superior law, it is deemed a direct conflict. For example, if a superior law clearly stipulates that maintenance funds belong to all owners, but a local regulation claims that they belong to the developer, this constitutes a direct conflict.

[0061] Step 304: When the relationship is determined to be supplementary and detailed, the specific provisions of the subordinate law are linked and integrated with the principle provisions of the superior law; when the relationship is determined to be direct conflict, the conflict points are explicitly marked in the output sequence of target policy knowledge fragments, and prompts are generated to follow the higher-level policy.

[0062] In practical applications, to supplement and refine relationships, the specific provisions of subordinate laws are linked to the general principles of superior laws within a sequence of target policy knowledge fragments. For example, in the sequence, the specific provisions of Chongqing Municipality are arranged adjacent to the general principles of the Civil Code, or a link is established in the metadata to indicate that the latter is a refinement and implementation of the former.

[0063] For direct conflicts, the system automatically adds a conflict warning label to the conflicting clauses and generates clear prompts, such as: "A discrepancy has been detected between a local regulation and the relevant provisions of the Civil Code. Based on the principle that higher-level laws prevail over lower-level laws, it is recommended that the provisions of the Civil Code prevail."

[0064] The above mechanism not only avoids providing incorrect information, but also enhances the professionalism and credibility of the answers by showcasing the complex relationships between policies at different levels, enabling users to understand policy issues comprehensively and dialectically.

[0065] The question-and-answer generation module is used to receive the sequence of target policy knowledge fragments and generate a final answer based on the sequence of target policy knowledge fragments, which is then returned to the user.

[0066] In this embodiment, the question-and-answer generation module includes: a multi-path generation unit, a rule evaluation unit, an answer evaluation unit, and a selection and summarization unit.

[0067] Among them, the multi-path generation unit is used to generate multiple candidate answers by calling the underlying large language model multiple times based on the same target policy knowledge fragment sequence.

[0068] In practical applications, the multi-path generation unit uses a sequence of target policy knowledge fragments as an invariant context and sends them to the same underlying large language model multiple times (e.g., 3-5 times). Each time it is called, the randomness can be controlled by fine-tuning the model's parameters, or different random seeds can be used to ensure the diversity of the generated results each time. The final output is a set of multiple candidate answers.

[0069] The rule evaluation unit is used to perform rule-based consistency evaluation on the multiple candidate answers, specifically including: Based on the knowledge graph of the multi-level policy knowledge base, key policy entities and question-answering processes are extracted from each candidate answer; the similarity of key policy entities and question-answering processes in multiple candidate answers is compared and analyzed; when the similarity of key policy entities and question-answering processes in all candidate answers is lower than a first preset threshold, the question is determined to be a complex question, triggering the expert review process; when the similarity of key policy entities and question-answering processes in multiple candidate answers is higher than a second preset threshold, the multiple similar candidate answers are output to the answer evaluation unit.

[0070] Understandably, when the answer to a question is clear and based on facts, the results generated multiple times by the large language model should exhibit a high degree of consistency in core factual elements and logical reasoning chains. Conversely, if the answers show significant discrepancies in key entities and processes, it exposes the complexity of the question or the ambiguity of the knowledge base.

[0071] Based on this, this embodiment utilizes named entity recognition technology similar to that used in knowledge base construction to extract key policy entities from each candidate answer. By analyzing logical connectors and verb sequences in the answers, or using sequence labeling models, the solutions are refined into standardized question-answering processes. Each answer is then transformed into two sets: a set of key entities and a sequence of answering processes. Set similarity algorithms (such as the Jaccard index) are used to calculate the similarity between the entity sets of all candidate answers, and sequence alignment algorithms are used to calculate the similarity between the answering process sequences. When the similarity of both key policy entities and question-answering processes reaches a first preset threshold, the model is deemed to have a high grasp of the question, and these highly similar answers are directly sent to the subsequent answer evaluation unit for fine-grained scoring. When the similarity of both key policy entities and question-answering processes is below a second preset threshold, the question is determined to be complex or have knowledge gaps, immediately triggering an expert review process, pausing automated output, and submitting the question and all dissenting answers to manual processing.

[0072] By using rule-based evaluation units for quality screening, high-risk issues that the model cannot handle can be identified early on and then handed over to experts for processing. This effectively prevents the risk of seriously divergent or potentially erroneous answers being passed down and ultimately output, thus improving the system's credibility.

[0073] The answer evaluation unit is used to score the quality of multiple candidate answers using a pre-trained evaluation model.

[0074] The evaluation model is trained under supervision using collected historical data (including manually annotated high-quality answers and answer pairs with user feedback). This model learns to comprehensively score candidate answers from multiple dimensions, primarily including: Accuracy: Does the answer strictly adhere to the provided policy knowledge fragments, and are there any factual errors or illusions? Completeness: Does it cover core information such as policy highlights, key procedures, and applicable conditions? Clarity and Structure: Is the expression clear, organized, and easy to understand?

[0075] In practical applications, pre-trained evaluation models are used to score the quality of multiple candidate answers, enabling rapid, batch, and objective automated evaluation of answer quality, replacing inefficient manual screening.

[0076] The optimal aggregation unit is used to select candidate answers whose quality scores are higher than a preset scoring threshold, and calls an aggregation model to integrate and refine the information of the candidate answers to generate the final answer.

[0077] In practical applications, firstly, based on the scores provided by the answer evaluation unit, all candidate answers with scores higher than a preset score threshold are selected. This ensures that the input materials for the aggregation model are of high quality. Then, an aggregation model is invoked and given explicit instructions, such as: "Based on the following excellent answers to the same question, integrate their information, remove redundancy, address their shortcomings, and generate a final version of the most accurate, comprehensive, and fluent answer." The aggregation model analyzes all high-quality candidate answers, extracts common facts, supplements details omitted by individual answers, optimizes language organization, and generates the final answer.

[0078] The question-and-answer generation module places the powerful generation capabilities of the large language model under a rigorous, multi-step quality control system. Through three processes—diversified generation, quantitative evaluation, and intelligent aggregation—it systematically addresses the core pain points of applying general-purpose large models in professional fields, ultimately outputting professional policy answers that are accurate, comprehensive, stable, and highly readable.

[0079] Please see Figure 1 In this embodiment, the system further includes a self-optimization module, which includes a data acquisition unit, a semantic clustering unit, and a knowledge base update driving unit.

[0080] The data acquisition unit is used to continuously collect user interaction data from the question-and-answer generation module. The user interaction data includes user queries that have not been effectively answered by the system and their session context.

[0081] In practical applications, the data acquisition unit determines whether an interaction has been "not effectively answered" based on various signals, such as: the user explicitly clicks the "Dissatisfied" or "Answer Useless" button; the user repeatedly asks the same question or changes the way they ask it in the same session; the system itself outputs an "Unable to answer" or low-confidence answer; the user subsequently provides negative feedback or remains inactive for an extended period. The data collected by the data acquisition unit includes not only the original user query but also the session context.

[0082] Semantic clustering units are used to perform semantic-level clustering analysis on collected user queries using unsupervised clustering algorithms, and to identify high-frequency question clusters that represent blind spots in the system's knowledge.

[0083] In practical applications, the semantic clustering unit first uses an embedding model to convert all collected user queries into high-dimensional vectors. This brings semantically similar sentences closer together in the vector space. Then, unsupervised clustering algorithms (such as DBSCAN and K-means) are used to group these vectors. Each generated cluster is a high-frequency question cluster, representing a "knowledge blind spot" that the system has not yet covered or has poor coverage of. For example, thousands of questions about "What to do when the elevator breaks down?", "Elevator outage for repair", and "Elevator malfunction application for funding" will be clustered into the same high-frequency question cluster called "Elevator emergency repair procedure".

[0084] The knowledge base update driving unit is communicatively connected to the knowledge base management module. It is used to generate optimization instructions for the multi-level policy knowledge base based on the high-frequency question cluster, and drive the knowledge base management module to supplement the multi-level policy knowledge base with targeted content.

[0085] In practical applications, the knowledge base update-driven unit generates structured optimization instructions based on the content of high-frequency question clusters. For example, for the question cluster "elevator emergency repair process," instructions might include: "Add elevator malfunction as a key entity in emergency repair scenarios to the knowledge base," "Add clauses regarding emergency elevator repair from the 'Special Equipment Safety Supervision Regulations,'" and "Create a detailed Q&A pair explaining the application steps for elevator emergency repair funds." These instructions are sent to the knowledge base management module to drive updates. For example, a pending work order might be generated, prompting the operations or expert team to manually review and supplement the knowledge according to the instructions, or, within the limits allowed by the rules, relevant knowledge might be automatically retrieved from trusted sources and directly injected into the multi-level policy knowledge base after quality verification.

[0086] Through the above mechanism, the multi-level policy knowledge base can keep up with the frequent revisions and dynamic releases of maintenance fund policies from the national to the provincial and municipal levels, so as to avoid the answers it provides being outdated or invalid, and thus dynamically adapt to changes in the policy environment and the development of user needs.

[0087] Please see Figure 1 In this embodiment, the system further includes a case association module, which is used for: When the question-and-answer generation module generates the final answer, it retrieves typical policy dispute cases related to the current user's query semantics from the case library linked to the multi-level policy knowledge base; and outputs the retrieved typical policy dispute cases and the final answer to the user.

[0088] It's understandable that simple policy clauses can be dry and difficult for ordinary users to fully comprehend. By linking them to real-life cases with similar factual backgrounds and shared legal disputes, abstract legal provisions can be transformed into vivid stories, significantly lowering the barrier to understanding for users and enhancing the credibility of the answers. Cases involving disputes and judgments can clearly demonstrate to users the potential legal consequences of non-compliance with policy regulations (such as "individual embezzlement resulting in imprisonment") or show the correct path to successfully resolving the problem (such as "elevator company suing owners' committee for repair fees"). This provides users with clear behavioral expectations and risk warnings.

[0089] Please see Figure 1 In this embodiment, the system further includes a policy comparison and analysis module, which is used for: The system responds to user input indicating the selection of multiple target regions; extracts differentiated policy provisions for the multiple target regions on preset key items of maintenance fund management from the multi-level policy knowledge base; and generates and outputs a policy differentiation comparison report.

[0090] In practical applications, the policy comparison and analysis module provides an interactive interface that allows users to select two or more target regions (e.g., selecting "Beijing," "Chongqing," and "Guangzhou" from a drop-down menu). The module extracts pre-defined key policy provisions (such as deposit standards and profit distribution) from a knowledge graph, analyzes and reveals the structured differences in key elements across different regions, achieving precise comparative analysis and automatically generating visual comparison reports. This completely eliminates cross-regional policy information barriers, providing accurate decision support for government departments, enterprises, and property owners, while highlighting the system's value leap from a question-and-answer tool to a policy analysis platform.

[0091] The following is a table of the core data involved in this embodiment: Table 1 illustrates how this system transforms abstract policy hierarchy and applicable principles into concrete search and response logic.

[0092] Table 1 Policy Hierarchy and Applicable Principles Table 2 provides a comparative analysis of the key provisions of maintenance fund management under different policy levels, demonstrating the necessity of this system in handling multi-level policies.

[0093] Table 2 Comparison of Major Policies on the Management of Residential Special Maintenance Funds Table 3 compares the advantages and disadvantages of different search technologies in the field of policy Q&A, demonstrating the advantages of the hybrid RAG search provided in this embodiment.

[0094] Table 3 Comparison of Search Enhancement Technologies In summary, the intelligent question-and-answer system for the maintenance fund field provided in this embodiment ensures the accuracy and comprehensiveness of policy-related questions and answers by constructing a multi-level policy knowledge graph and a hybrid retrieval-enhanced generation engine. It innovatively introduces policy conflict handling and self-optimization mechanisms, systematically solving the problems of answer compliance, regional adaptability, and continuous evolution capability. Finally, with extended functions such as multiple-answer selection, case association, and policy comparison, it achieves a leap from passive question-and-answer to proactive analysis and decision-making, completely overcoming the long-standing problems of policy information asymmetry and insufficient service intelligence in the maintenance fund field.

[0095] Based on the above system, Figure 2 A flowchart illustrating an intelligent question-answering method in the field of maintenance funds is shown below. Please refer to [link / reference]. Figure 2 The method includes the following steps: Step 401: Construct and maintain a multi-level policy knowledge base. The multi-level policy knowledge base is a structured knowledge graph based on the ontology of the maintenance fund field. It integrates policy provisions of multiple different levels, and the policy provisions are parsed into policy entities, policy relationships and policy concept levels and stored in a graph structure. Step 402: Receive user query and use a hybrid retrieval method that combines keyword retrieval and vector retrieval to retrieve an initial set of policy knowledge fragments related to the user query from the multi-level policy knowledge base; Step 403: Based on the preset policy hierarchy and applicable principle rule set, prioritize the policy knowledge fragments in the initial policy knowledge fragment set, identify conflict and supplementary relationships, and generate a sequence of target policy knowledge fragments that has been sorted and marked with conflict. Step 404: Generate the final answer based on the sequence of target policy knowledge fragments and return it to the user.

[0096] It is understood that since the intelligent question-and-answer method in the field of maintenance funds described in this embodiment is based on the intelligent question-and-answer system in the field of maintenance funds described in the embodiment, the method disclosed in the embodiment is relatively simple to describe because it corresponds to the system disclosed in the embodiment. For relevant parts, please refer to the description of the system.

[0097] Based on the above methods, Figure 3 A schematic diagram of an electronic device is shown below. Please refer to [link / reference]. Figure 3 The device includes a processor, memory, and a communication bus; The communication bus is used to enable communication between the processor and the memory; The processor is used to execute one or more programs in the memory to implement the steps of the intelligent question-and-answer method in the field of maintenance funds described in the embodiments.

[0098] It is understood that since the electronic device described in this embodiment is a device for implementing the intelligent question-and-answer method in the field of maintenance funds as described in the embodiment, the device disclosed in the embodiment is described in a relatively simple way because it corresponds to the method disclosed in the embodiment. For relevant parts, please refer to the description of the method.

Claims

1. An intelligent question answering system in the field of maintenance funds, characterized by, The system comprises: a knowledge base management module for constructing and maintaining a multi-level policy knowledge base, which is a structured knowledge graph constructed based on a maintenance fund domain ontology, wherein a plurality of policy provisions of different levels are integrated, and the policy provisions are parsed into policy entities, policy relationships and policy concept levels and stored in a graph structure; a hybrid retrieval enhancement generation engine connected with the knowledge base management module, configured to receive a user query, and retrieve an initial policy knowledge fragment set related to the user query from the multi-level policy knowledge base by using a hybrid retrieval method combining keyword retrieval and vector retrieval; a policy conflict processing module connected with the hybrid retrieval enhancement generation engine, configured to receive the initial policy knowledge fragment set, and perform priority sorting, conflict relationship and supplementary relationship identification on the policy knowledge fragments in the initial policy knowledge fragment set according to a preset policy level and applicable principle rule set, and output a target policy knowledge fragment sequence sorted and conflict-labeled; a question and answer generation module connected with the policy conflict processing module, configured to receive the target policy knowledge fragment sequence, and generate a final answer based on the target policy knowledge fragment sequence and return the final answer to the user; wherein the user query, the initial policy knowledge fragment set, the target policy knowledge fragment sequence and the final answer sequentially flow to form a question and answer processing main data flow.

2. The intelligent question answering system in the field of repair funds according to claim 1, characterized in that, The execution process of the hybrid retrieval enhancement generation engine specifically comprises: parsing the user query to generate a corresponding keyword set, and encoding the user query into a query vector by using an Embedding model; based on the keyword set, performing first retrieval in an inverted index associated with the multi-level policy knowledge base to obtain a first retrieval result list; at the same time, based on the query vector, performing similarity retrieval in a vector database associated with the multi-level policy knowledge base to obtain a second retrieval result list; using a reverse ranking fusion algorithm to perform score fusion and reordering on the policy knowledge fragments in the first retrieval result list and the second retrieval result list, and outputting the reordered results as the initial policy knowledge fragment set to the policy conflict processing module.

3. The intelligent question answering system in the field of repair funds according to claim 1, characterized in that, The execution process of the policy conflict processing module specifically comprises: parsing the source of each policy knowledge fragment in the initial policy knowledge fragment set, and labeling the policy level according to the policy level and applicable principle rule set; sorting the initial policy knowledge fragment set according to the definition rules in the policy level and applicable principle rule set, and placing the policy knowledge fragments from the lowest applicable level of the user's location and related in content at the highest priority; the definition rules include: higher-level law is superior to lower-level law, special law is superior to general law, and new law is superior to old law; comparing the contents of the policy knowledge fragments from different policy levels to determine whether there is a supplementary refinement relationship or a direct conflict relationship between them; When the supplementary refinement relationship is determined, the specific provisions of the lower law are associated and integrated with the principle provisions of the higher law; when the direct conflict relationship is determined, the conflict points are explicitly marked in the output target policy knowledge fragment sequence, and prompt information following the high-level policy is generated.

4. The intelligent question answering system in the field of service fund according to claim 1, characterized in that, Further comprising a self-optimization module, the self-optimization module comprising: a data acquisition unit for continuously acquiring user interaction data from the question and answer generation module, the user interaction data including user queries that are not effectively answered by the system and their conversation context; a semantic clustering unit for performing semantic clustering analysis on the acquired user queries using an unsupervised clustering algorithm to identify high-frequency problem clusters representing knowledge gaps of the system; a knowledge base update driving unit in communication connection with the knowledge base management module, configured to generate optimization instructions for the multi-level policy knowledge base according to the high-frequency problem clusters, and drive the knowledge base management module to make targeted content supplements to the multi-level policy knowledge base.

5. The intelligent question answering system in the field of repair funds according to claim 1, characterized in that, The question and answer generation module comprises: a multi-path generation unit for calling a bottom-layer large language model multiple times based on the same target policy knowledge fragment sequence to generate multiple candidate answers; a rule evaluation unit for performing rule-based consistency evaluation on the multiple candidate answers, specifically including: extracting key policy entities and problem solving processes from each candidate answer based on the knowledge graph of the multi-level policy knowledge base; comparing and analyzing the similarity of key policy entities and problem solving processes in multiple candidate answers; when the similarity of key policy entities and problem solving processes of all candidate answers is lower than a first preset threshold, determining that the problem is a complex problem and triggering an expert review process; when the similarity of key policy entities and problem solving processes of multiple candidate answers is higher than a second preset threshold, outputting similar multiple candidate answers to an answer evaluation unit; an answer evaluation unit for quality scoring of multiple candidate answers by a pre-trained evaluation model; a best selection and summarization unit for selecting candidate answers with quality scores higher than a preset score threshold, and calling a summarization model to integrate and refine the candidate answers to generate a final answer.

6. The intelligent question answering system in the field of service fund according to claim 1, wherein, The process of constructing a multi-level policy knowledge base by the knowledge base management module comprises: defining a domain ontology including a policy ontology and a policy relationship based on the policy logic of the maintenance fund field, the policy ontology including: owners, property service enterprises, owner committees, and residential special maintenance funds, and the policy relationship including: ownership, management rights, use rights, and deposit obligations; automatically preprocessing the collected multi-level policy texts, the preprocessing including word segmentation, part-of-speech tagging, and syntax analysis; under the guidance of the domain ontology, extracting policy entities from the preprocessed texts using a named entity recognition model, and extracting policy relationships between entities using a relationship extraction model to form policy element triples; storing the policy element triples into a graph database to construct a structured knowledge graph, and establishing an association with an inverted index and a vector database to complete the construction of the multi-level policy knowledge base.

7. The intelligent question answering system in the field of repair funds according to claim 1, characterized in that, Further comprising a case association module, the case association module being configured to: When the question and answer generation module generates the final answer, a typical case of policy dispute related to the semantic of the current user query is retrieved from a case base linked with the multi-level policy knowledge base; The retrieved typical case of policy dispute and the final answer are output to the user.

8. The intelligent question answering system in the field of service fund according to claim 1, characterized in that, Further comprising a policy comparison and analysis module, configured to: In response to a selection instruction of a plurality of target regions input by the user; Extract the differentiated policy provisions of the plurality of target regions on the preset key items of maintenance fund management from the multi-level policy knowledge base; Generate and output a policy differentiated comparison report.

9. An intelligent question-answering method in the field of maintenance funds, characterized by, The intelligent question and answer system applied to the field of maintenance fund as claimed in any one of claims 1 to 8, the method comprising: Building and maintaining a multi-level policy knowledge base, which is a structured knowledge graph based on the ontology of the field of maintenance fund, wherein a plurality of policy provisions of different levels are integrated, and the policy provisions are parsed into policy entities, policy relationships and policy concept levels and stored in a graph structure; Receiving a user query, and using a hybrid retrieval method combining keyword retrieval and vector retrieval to retrieve an initial policy knowledge fragment set related to the user query from the multi-level policy knowledge base; According to the pre-set policy rank and application principle rule set, the policy knowledge fragments in the initial policy knowledge fragment set are prioritized, the conflict relationship and the supplementary relationship are identified, and a target policy knowledge fragment sequence sorted and conflict annotated is generated; Based on the target policy knowledge fragment sequence, a final answer is generated and returned to the user.

10. An electronic device, comprising: The electronic device comprises a processor, a memory and a communication bus; The communication bus is used to realize the connection communication between the processor and the memory; The processor is used to execute one or more programs in the memory to realize the steps of the intelligent question and answer method in the field of maintenance fund as claimed in claim 9.

Images