Knowledge enhancement based federated search method, system, medium and terminal

Abstract

The application provides a knowledge enhancement-based joint retrieval method, system, medium and terminal, which ensures that the retrieval result is no longer limited to text fragments similar in literal, but forms an organic knowledge whole integrating direct clauses, relevant articles, case gist and internal rules, so that the generated reply has rigorous legal basis and complete business context, reducing the claim settlement disputes and legal risks caused by the lack of key knowledge association. The compliance control is changed from lagging manual inspection to real-time pre-built-in system, reducing the compliance vacuum period caused by policy update lag, reducing the regulatory penalties and reputation risks of enterprises. By retaining historical versions and establishing version replacement chain, the historical context-based explanation is ensured for the query of the policy life cycle across different time points, enhancing the credibility and accuracy of customer service.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and natural language processing technology, and in particular to a knowledge-enhanced joint retrieval method, system, medium and terminal. Background Technology

[0002] In recent years, Retrieval-Augmented Generation (RAG) technology has seen initial applications in intelligent question-answering scenarios in vertical industries such as finance and insurance due to its ability to combine the generative capabilities of Large Language Models (LLMs) with the accuracy of external knowledge bases. However, the knowledge system of the insurance industry is characterized by high complexity, strong specialization, and stringent regulatory requirements. Traditional RAG architectures have revealed many fundamental flaws in practical applications, failing to meet the core needs of the industry's digital transformation.

[0003] (1) Fragmented search results in a lack of relevance of terms.

[0004] Traditional RAG technology relies on similarity calculations between text embedding vectors for semantic retrieval, only recalling text fragments that are literally or semantically similar to the query question. It fails to identify the deep logical connections between insurance clauses, laws and regulations, and judicial precedents. For example, when a user inquires about "the scope of compensation for accidental injuries during a road trip," existing technology only recalls clause fragments containing the literal words "road trip" and "accidental injury," omitting key content such as "exclusion of liability for high-risk sports" in the disclaimer. In major accident claims scenarios, such as "how to claim compensation if a factory collapses due to a typhoon, and the insured property insurance policy did not include natural disaster insurance?", traditional RAG can only recall the "natural disaster exclusion" clause in the property insurance policy, but it overlooks related knowledge such as the principle of adverse interpretation of standard clauses in the Insurance Law, local high court precedents (standards for determining mixed causation), and the "catastrophic liability sharing" mechanism in reinsurance agreements. This lack of knowledge connection can easily lead to claims disputes and legal proceedings, causing direct economic losses and damage to brand reputation for enterprises.

[0005] (2) Disconnection between business rules and lack of compliance control

[0006] Traditional RAG generation processes lack embedded constraints on actuarial rules, underwriting policies, and regulatory requirements. Their outputs often fail to meet professional standards and are prone to misleading sales practices. For example, when an agent asks, "What are the returns on an increasing whole life insurance policy?", the system can easily deviate from the strict terms of the insurance contract, generating statements such as "guaranteed profit"—absolute promises with clear misleading intent—thus violating regulatory red lines. Traditional systems cannot adapt to dynamic changes in regulatory policies. Faced with newly introduced regulations (such as the prohibition of "100% payout" in the "Internet Insurance Sales Behavior Code"), the system's reliance on delayed human intervention creates a significant compliance vacuum after the new regulations take effect. During this period, the system continues to output non-compliant statements, exposing companies to significant regulatory penalties and reputational risks, ultimately leading to penalties.

[0007] (3) The static knowledge system is disconnected from the dynamic business scenarios.

[0008] Insurance policies have lifecycles spanning decades, during which the insured's health status, product terms, regulatory policies, and premium rules may all change. Traditional RAG (Rapid Insurance Group) systems have static knowledge bases that cannot perceive or correlate these temporal changes, leading to ineffective risk assessments and service gaps. For example, when a customer asks, "Why is the critical illness insurance I bought in 2018 more expensive than it is now?", the system can only provide the current premium rate table, unable to correlate it with the 2018 disease incidence rate table, historical pricing rules, and then-current regulatory policies, thus failing to provide an accurate explanation and causing a decline in customer trust. Summary of the Invention

[0009] In view of the shortcomings of the prior art described above, the purpose of this application is to provide a knowledge-enhanced joint retrieval method, system, medium and terminal to solve the technical problems of fragmented retrieval leading to the lack of relevance of terms, disconnection of business rules and lack of compliance control in the prior art.

[0010] To achieve the above and other related objectives, a first aspect of this application provides a knowledge-enhanced joint retrieval method, comprising: acquiring different types of original documents; the original documents including one or more combinations of insurance policy documents, insurance regulatory documents, insurance rule documents, court case insurance documents, and historical question-and-answer record documents; based on a large language model, performing entity information extraction, high-level semantic unification, and intrinsic association mining operations on the various types of original documents to construct a structured knowledge base, a domain tag knowledge base, and a multi-source heterogeneous knowledge graph, and updating the structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph when a new document is detected; acquiring user query data, and performing joint retrieval based on the updated structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph to obtain a candidate knowledge document set; and generating preferred answer data based on the candidate knowledge document set and the large language model.

[0011] In some embodiments of the first aspect of this application, the method of constructing a structured knowledge base, a domain tag knowledge base, and a multi-source heterogeneous knowledge graph by performing entity information extraction, high-level semantic unification, and intrinsic association mining operations on various types of original documents based on a large language model includes: preprocessing the various types of original documents to obtain various types of normalized documents; performing entity information extraction operations on the various types of normalized documents based on a large language model to identify and extract multiple named entities and the attributes of each named entity; performing high-level semantic unification operations on each extracted named entity to generate a series of first standard entities, and determining the attributes of each first standard entity, the tags of each first standard entity, and assigning corresponding identifiers to each first standard entity based on the attributes of each named entity; performing intrinsic association mining operations on each first standard entity to obtain the relationships between each first standard entity; and constructing a structured knowledge base, a domain tag knowledge base, and a multi-source heterogeneous knowledge graph based on the attributes of each first standard entity, the tags of each first standard entity, the identifiers corresponding to each first standard entity, and the relationships between each first standard entity.

[0012] In some embodiments of the first aspect of this application, the method of updating the structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph when a new document is detected includes: acquiring the new document and semantically segmenting the new document to obtain multiple first text blocks; summarizing each first text block to obtain summary text data of each first text block; summarizing the new document based on the summary text data of each first text block to generate summary text data of the new document; performing feature extraction operations on each first text block based on the summary text data of the new document and the summary text data of each first text block to obtain newly added entity relation triples, attributes of newly added entities, and tags of newly added entities in the new document, and determining the identifier of newly added entities; performing standardized matching on the newly added entity relation triples, attributes of newly added entities, identifier of newly added entities, and tags of newly added entities, and collaboratively updating the structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph based on the standardized matching results.

[0013] In some embodiments of the first aspect of this application, obtaining user query data and performing joint retrieval based on the updated structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph to obtain a candidate knowledge document set includes: parsing and feature extraction operations on the user query data to obtain target retrieval data; obtaining an enhanced entity set and an enhanced tag set based on the target retrieval data and the updated multi-source heterogeneous knowledge graph; performing joint retrieval on the updated structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph based on the target retrieval data and the enhanced entity set and the enhanced tag set, and fusing the joint retrieval results to generate a candidate knowledge document set.

[0014] In some embodiments of the first aspect of this application, the method of parsing and extracting features from the user query data to obtain target retrieval data includes: parsing the user query data, and performing entity feature extraction and tag feature extraction on the parsed user query data to obtain an initial entity set and an initial tag set; and performing a standardization operation on the initial entity set to generate target retrieval data consisting of a second standard entity set and the initial tag set.

[0015] In some embodiments of the first aspect of this application, obtaining an enhanced entity set and an enhanced tag set based on the target retrieval data and utilizing the updated multi-source heterogeneous knowledge graph includes: performing a multi-level traversal operation on the updated multi-source heterogeneous knowledge graph based on the target retrieval data to obtain an expanded entity set; calculating the global semantic relevance score between each candidate entity in the expanded entity set and the second standard entity set, and recalling the most relevant target entities from the expanded entity set based on the ranking result of the global semantic relevance score to obtain an enhanced entity set; and selecting relevant tags from the initial tag set based on the attribution relationship between each target entity in the enhanced entity set and the initial tag set to form an enhanced tag set.

[0016] In some embodiments of the first aspect of this application, the candidate knowledge document set includes multiple candidate documents; the method of generating preferred answer data based on the candidate knowledge document set and the large language model includes: determining whether the total length of the candidate knowledge document set exceeds a preset length threshold; if the total length of the candidate knowledge document set does not exceed the preset length threshold, directly inputting the candidate knowledge document set into the large language model to generate preferred answer data; if the total length of the candidate knowledge document set exceeds the preset length threshold, filtering the second text blocks in each candidate document based on the enhanced entity set and the enhanced tag set to obtain a first-level candidate text block set; determining whether the total length of the first-level candidate text block set exceeds a preset length threshold; if the total length of the first-level candidate text block set does not exceed the preset length threshold, directly inputting the first-level candidate text block set into the large language model to generate preferred answer data; The data is input into the large language model to generate preferred answer data. If the total length of the first-level candidate text block set exceeds a preset length threshold, the first-level candidate text block set is summarized to generate a second-level summary text set. It is then determined whether the total length of the second-level summary text set exceeds the preset length threshold. If the total length of the second-level summary text set does not exceed the preset length threshold, the second-level summary text set is directly input into the large language model to generate preferred answer data. If the total length of the second-level summary text set exceeds the preset length threshold, the second-level summary text set is divided into multiple third-level summary text sets that meet the preset length threshold. Each of the third-level summary text sets is input into the large language model in batches to obtain the sub-answer data corresponding to each of the third-level summary text sets. The sub-answer data corresponding to each of the third-level summary text sets is then fused to generate preferred answer data.

[0017] To achieve the above and other related objectives, a second aspect of this application provides a knowledge-enhanced joint retrieval system, comprising: a data acquisition module for acquiring different types of original documents; the original documents include one or more combinations of insurance policy documents, insurance regulatory documents, insurance rule documents, court case insurance documents, and historical question-and-answer record documents; a knowledge system construction and updating module for performing entity information extraction, high-level semantic unification, and intrinsic association mining operations on the various types of original documents based on a large language model to construct a structured knowledge base, a domain tag knowledge base, and a multi-source heterogeneous knowledge graph, and updating the structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph when a new document is detected; a retrieval module for acquiring user query data and performing joint retrieval based on the updated structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph to obtain a set of candidate knowledge documents; and an answer generation module for generating preferred answer data based on the set of candidate knowledge documents and the large language model.

[0018] To achieve the above and other related objectives, a third aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the knowledge-enhanced joint retrieval method described above.

[0019] To achieve the above and other related objectives, a fourth aspect of this application provides an electronic terminal, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the knowledge-enhanced joint retrieval method described above.

[0020] As described above, the knowledge-enhanced joint retrieval method, system, medium, and terminal of this application have the following beneficial effects:

[0021] (1) By introducing a structured knowledge base, a domain-labeled knowledge base, and a multi-source heterogeneous knowledge graph, it is possible to identify and traverse the deep semantic relationships and business logic between insurance clauses, laws and regulations, and judicial precedents. This ensures that the search results are no longer limited to literal text fragments, but form an organic knowledge whole that integrates direct clauses, relevant legal provisions, the essence of precedents, and internal rules. The resulting responses have rigorous legal basis and complete business context, reducing claims disputes and legal risks caused by the lack of key knowledge connections.

[0022] (2) Through the synergistic effect of structured knowledge bases, domain-labeled knowledge bases, and multi-source heterogeneous knowledge graphs, actuarial rules, underwriting policies, and regulatory requirements can be transformed into computable knowledge units and constraints. During the generation process, these knowledge units constitute hard constraints, which can proactively trigger compliance verification processes to ensure that the output content meets professional standards and regulatory requirements. This mechanism transforms compliance control from delayed manual inspections to real-time, pre-implemented system updates, reducing compliance vacuum periods caused by policy update delays and lowering regulatory penalties and reputational risks for enterprises.

[0023] (3) By parsing, summarizing, extracting features, and standardizing matching new documents, the structured knowledge base, domain tag knowledge base, and multi-source heterogeneous knowledge graph are updated. During the update, not only the currently effective knowledge content is stored, but also historical version knowledge and its valid timestamps are retained, and a clear version replacement relationship chain is established in the knowledge graph. This enables the system to accurately understand and associate product terms, rate rules, and regulatory policies at different points in time, thereby providing accurate interpretations of queries across the policy lifecycle based on historical context, enhancing the accuracy and credibility of customer service, and effectively maintaining customer relationships and brand trust. Attached Figure Description

[0024] Figure 1 The diagram shown is a flowchart of a knowledge-enhanced joint retrieval method according to an embodiment of this application.

[0025] Figure 2 The diagram shown is a flowchart illustrating the process of constructing a knowledge system in one embodiment of this application.

[0026] Figure 3 The diagram shown is a schematic representation of the joint search process in one embodiment of this application.

[0027] Figure 4 The diagram shown is a flowchart illustrating the process of generating preferred answer data in one embodiment of this application.

[0028] Figure 5 The diagram shown is a schematic block diagram of a knowledge-enhanced joint retrieval system according to an embodiment of this application.

[0029] Figure 6 The diagram shown is a structural schematic of an electronic terminal according to an embodiment of this application. Detailed Implementation

[0030] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0031] Before providing a further detailed description of the present invention, the nouns and terms used in the embodiments of the present invention are explained, and the nouns and terms used in the embodiments of the present invention are subject to the following interpretations:

[0032] <1> Large Language Model: Based on massive amounts of text data, it uses deep learning technology to learn the statistical patterns and regularities of human language.

[0033] To facilitate understanding of the embodiments of this application, firstly, in conjunction with Figure 1 Detailed explanation. Figure 1 This document illustrates a flowchart of the knowledge-enhanced joint retrieval method according to an embodiment of the present invention. The knowledge-enhanced joint retrieval method in this embodiment mainly includes the following steps:

[0034] S101: Obtain original documents of different types; the original documents include one or more of the following: insurance terms documents, insurance regulatory documents, insurance rules documents, court case insurance documents, and historical Q&A records.

[0035] In this embodiment, various types of original documents are obtained. Insurance policy documents include, but are not limited to, different formats of policy terms for critical illness insurance, life insurance, and property insurance. Insurance regulatory documents include, but are not limited to, documents such as the *Insurance Law*, the *Administrative Measures for Health Insurance*, and the *Administrative Measures for Internet Insurance Business*. Insurance rule documents include, but are not limited to, documents such as insurance underwriting rules, insurance claims rules, insurance product rules, actuarial rules, and underwriting policies. Court case law documents include, but are not limited to, civil judgments involving insurance disputes, judgments on claims disputes, and judgments on subrogation. Historical Q&A records include, but are not limited to, customer service dialogue records, online consultation logs, user feedback, and standard Q&A databases.

[0036] S102: Based on the large language model, entity information extraction, high-level semantic unification, and intrinsic association mining operations are performed on the original documents of various types to construct a structured knowledge base, a domain tag knowledge base, and a multi-source heterogeneous knowledge graph. When a new document is detected, the structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph are updated.

[0037] In this embodiment, as Figure 2 The diagram illustrates the process of constructing a knowledge system in an embodiment of the present invention. Based on a large language model, entity information extraction, high-level semantic unification, and intrinsic association mining operations are performed on various types of original documents to construct a structured knowledge base, a domain-labeled knowledge base, and a multi-source heterogeneous knowledge graph.

[0038] S1021: Preprocess the original documents of various types to obtain normalized documents of various types.

[0039] In this embodiment, various types of original documents are preprocessed. Preprocessing methods include, but are not limited to, text cleaning preprocessing, format standardization preprocessing, and semantic segmentation preprocessing to obtain standardized documents of various types. Text cleaning preprocessing removes all "noise" characters and content from the document that are irrelevant to the semantic content, reducing the risk of large language models learning incorrect patterns or being interfered with by irrelevant characters, and improving the accuracy of tasks such as entity extraction and relation mining. Format standardization preprocessing converts documents from different sources and in different formats into one or more unified, machine-processable intermediate formats. Semantic segmentation preprocessing divides the document into coherent and meaningful paragraphs or blocks based on its semantic and logical structure, providing accurate context for large language models.

[0040] S1022: Based on a large language model, entity information extraction is performed on various types of standardized documents to identify and extract multiple named entities and the attributes of each named entity.

[0041] In this embodiment, a large language model is used to identify and extract multiple named entities and their attributes from various types of normalized documents. For example, a named entity named "Critical Illness Insurance" may have attributes such as "Life Insurance", "Policy Number", "Insurance Company", and "Waiting Period". A named entity named "Insurance Period" may have attributes such as "Lifetime", "Start Time" (e.g., a specific year, month, and day) and "End Condition" (e.g., the death of the insured).

[0042] S1023: Perform a high-level semantic unification operation on each of the extracted named entities to generate a series of first standard entities, and determine the attributes of each first standard entity, the label of each first standard entity, and assign a corresponding identifier to each first standard entity based on the attributes of each named entity.

[0043] In this embodiment, when performing high-level semantic unification on the extracted named entities, the semantic similarity between the named entities is calculated to merge multiple expressions pointing to the same named entity into a unified standard form, thereby generating a series of first standard entities. Subsequently, based on the attributes of all merged original named entities, the attributes of these first standard entities are fused, deduplicated, and completed to determine the attributes of each first standard entity. Simultaneously, a label is assigned to each first standard entity according to its unified semantic category. Finally, a unique identifier is assigned to each first standard entity to ensure its identifiability and traceability within the system.

[0044] S1024: Perform an intrinsic association mining operation on each of the first standard entities to obtain the relationship between each of the first standard entities.

[0045] In this embodiment, when performing intrinsic association mining on each first standard entity, potential connections between entities are detected from structured and unstructured data sources based on the constructed first standard entities and their attributes. By analyzing the co-occurrence, semantic compatibility, and domain-specific logic of entity attributes, multiple relationship types such as membership, cooperation, causality, and hierarchy between the first standard entities are mined.

[0046] S1025: Based on the attributes of each first standard entity, the tags of each first standard entity, the identifiers corresponding to each first standard entity, and the relationships between each first standard entity, construct a structured knowledge base, a domain tag knowledge base, and a multi-source heterogeneous knowledge graph.

[0047] In this embodiment, based on the unique identifier and attributes of each first standard entity, entity information is stored and managed in a structured form to construct a structured knowledge base. This structured knowledge base is typically constructed using a relational database or a tabular structure, where each first standard entity is treated as an independent record, and its various normalized attributes are fields within that record. Relationships between first standard entities are mapped and stored through association tables, and the detailed text content associated with them can also be queried through these relationships.

[0048] In this embodiment, based on the tags assigned to the first standard entity, all tags are systematically extracted and organized hierarchically according to the domain ontology to form an independent domain tag knowledge base. This domain tag knowledge base not only includes all tag items and their labeled knowledge objects (such as knowledge modules, clause families, document categories, etc.), but also defines the hierarchical relationships (such as "parent class-child class"), synonym relationships, and correlation relationships between tags. Simultaneously, the knowledge base stores the corresponding authoritative explanations and defining texts for these knowledge objects. This structured tag system serves upper-layer applications, supporting tag-based rapid retrieval, category navigation, and semantic recommendation.

[0049] In this embodiment, a first standard entity with a unique identifier is used as a node. Each node stores all its attributes and has one or more corresponding tags attached to it. Relationships between the first standard entities are used as edges, and these edges are assigned types and attributes to construct a multi-source heterogeneous knowledge graph. This multi-source heterogeneous knowledge graph integrates data from different sources and structures, forming a semantically rich networked knowledge system based on "entity-relationship-entity" triples. This graph-based structured representation supports applications such as deep relational reasoning and intelligent question answering. Thus, a structured knowledge base, a domain-labeled knowledge base, and a multi-source heterogeneous knowledge graph form a rich knowledge system.

[0050] In this embodiment, the method of updating the structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph when a new document is detected includes:

[0051] (1) Obtain the new document and perform semantic segmentation on the new document to obtain multiple first text blocks.

[0052] (2) Summarize each of the first text blocks to obtain summary text data of each of the first text blocks.

[0053] (3) Summarize the new document based on the summary text data of each of the first text blocks to generate summary text data of the new document.

[0054] (4) Based on the summary text data of the new document and the summary text data of each of the first text blocks, perform feature extraction operation on each of the first text blocks to obtain the new entity relation triplet, the attributes of the new entity and the label of the new entity in the new document, and determine the identifier of the new entity.

[0055] (5) Standardize the matching of the newly added entity relation triples, the attributes of the newly added entities, the identifiers of the newly added entities, and the tags of the newly added entities, and based on the standardized matching results, collaboratively update the structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph.

[0056] In this embodiment, new documents include, but are not limited to: new insurance product terms and conditions, the latest regulatory rules, new underwriting rules manuals, new insurance claims practices, new court precedents, and new health disclosure documents.

[0057] In this embodiment, a document parsing technique based on natural language processing is used to segment the new document into multiple semantically complete first text blocks. Each first text block contains a relatively independent semantic unit, such as a complete clause, a definition, or a provision. Each first text block is summarized to generate summary text data for each first text block. This summary text data encapsulates the core claims, constraints, or key data of the first text block, discarding redundant embellishments and explanatory statements. Based on the summary text data of all first text blocks, the new document is summarized to generate summary text data for the new document, providing semantic context for subsequent feature extraction.

[0058] In this embodiment, based on the summary text data of the new document and the summary text data of each first text block, new entities are identified from the first text blocks, and the new entity relationship triples (such as [new product X, bound by, new regulation Y]) and various attributes of the new entities (such as the cooling-off period days of the new product) are extracted. According to the entity type and context, one or more labels for the new entities are assigned, and a unique identifier is assigned to each new entity to ensure its traceability in the knowledge system.

[0059] In this embodiment, newly added entities and tags are compared with standards in the existing structured knowledge base and domain tag knowledge base. If synonyms or inclusion relationships exist, they are merged and normalized; if it is entirely new knowledge, it is prepared for addition. The newly added entity and its attributes are written into the structured knowledge base as new records, or as supplementary attributes to existing records, to update the structured knowledge base. The new tag is incorporated into the tag system, establishing its hierarchical relationship with its parent tag, to update the domain tag knowledge base. The newly added entity relationship triples are transformed into new nodes and edges in the graph, and connected to existing nodes to update the multi-source heterogeneous knowledge graph.

[0060] It's important to note that when new regulatory documents are detected, they are parsed to extract core rules such as "prohibition of absolute promises." These rules are then updated as new entities (e.g., "Article X of XX Regulation"), attributes (prohibited statements, such as "guaranteed profit"), and tags (regulatory prohibitions) in the structured knowledge base, domain tag knowledge base, and multi-source heterogeneous knowledge graph. In subsequent Q&A sessions, when a query related to product revenue is received, these new knowledge base units are prioritized. The regulatory prohibition tags in the domain tag knowledge base trigger a compliance verification process, and the newly added "prohibited" relationship edges in the multi-source heterogeneous knowledge graph constitute hard constraints, prohibiting the inclusion of prohibited statements in candidate answers. This transforms compliance control from "post-event manual inspection" to "pre-event system integration," ensuring that the output answers comply with the latest regulatory requirements. This avoids compliance vacuums caused by system update delays and reduces the probability of sales misrepresentation and regulatory penalties.

[0061] Furthermore, traditional static knowledge bases cannot cope with the dynamic scenarios that change over time in insurance business. This application updates the structured knowledge base, domain-labeled knowledge base, and multi-source heterogeneous knowledge graph by parsing, summarizing, extracting features, and standardizing matching new documents. During the update, the new document does not overwrite the old one; instead, it is stored as a new version of the knowledge, along with attributes such as a valid timestamp, in the structured knowledge base. The multi-source heterogeneous knowledge graph then establishes a "substitution" or "update" relationship chain between the old and new versions of knowledge. For example, when a customer asks, "Why is the critical illness insurance I bought in a certain year more expensive than it is now?", this application can use the timestamp attribute and the version relationship in the graph to link to historical knowledge states such as the premium rules, disease incidence tables, and regulatory policies of that year. This allows for an explanation based on historical context, rather than a one-sided answer based solely on the current state, enhancing the accuracy and credibility of customer service and reducing the decline in customer trust caused by information asymmetry.

[0062] S103: Obtain user query data and perform joint retrieval based on the updated structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph to obtain a set of candidate knowledge documents.

[0063] In this embodiment, as Figure 3 The diagram illustrates the process of joint retrieval in an embodiment of the present invention. The method for obtaining a candidate knowledge document set by acquiring user query data and performing joint retrieval based on the updated structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph includes:

[0064] S1031: Perform parsing and feature extraction operations on the user query data to obtain target retrieval data. The methods include:

[0065] (1) Perform parsing operations on the user query data, and perform entity feature extraction and tag feature extraction operations on the parsed user query data to obtain an initial entity set and an initial tag set.

[0066] (2) Perform a standardization operation on the initial entity set to generate target retrieval data consisting of a second standard entity set and the initial tag set.

[0067] In this embodiment, user query data is parsed to extract the core question intent, laying the foundation for subsequent retrieval and answer generation. Entity feature extraction is then performed on the parsed user query data to identify and extract key entities from the query. These entities include, but are not limited to: insured objects (e.g., factory buildings), insurance products (e.g., critical illness insurance), risks and accidents (e.g., typhoons), time and amount, behaviors and processes (e.g., claims), etc. All extracted entities are aggregated into an initial entity set.

[0068] In this embodiment, tag feature extraction is performed on the user query data after parsing. The parsed query semantics are mapped to a standardized tag system in the domain tag knowledge base. By calculating the similarity between the query semantic vector and the tag vector, one or more tags are automatically assigned to form an initial tag set. These tags include, but are not limited to: claims consultation, liability determination, disclaimers, timeliness, product comparison, etc.

[0069] In this embodiment, the entities in the initial entity set may be original user expressions, which may contain issues such as colloquialisms and ambiguity. The initial entity set is mapped to first standard entities in a structured knowledge base and a multi-source heterogeneous knowledge graph. For example, expressions such as "critical illness insurance" and "major illness insurance" are uniformly mapped to standard product entities, such as "critical illness insurance (version number)," and their identifiers are obtained. After standardization, a second standard entity set is formed, and the second standard entity set and the initial tag set are encapsulated to form the final target retrieval data.

[0070] S1032: Retrieve data based on the target and utilize the updated multi-source heterogeneous knowledge graph to obtain an enhanced entity set and an enhanced tag set. This includes the following methods:

[0071] (1) Based on the target retrieval data, perform a multi-level traversal operation of related entities in the updated multi-source heterogeneous knowledge graph to obtain an extended entity set.

[0072] (2) Calculate the global semantic relevance score between each candidate entity in the extended entity set and the second standard entity set, and based on the ranking result of the global semantic relevance score, recall the most relevant target entities from the extended entity set to obtain the enhanced entity set.

[0073] (3) Based on the attribution relationship between each target entity in the enhanced entity set and the initial tag set, relevant tags are selected from the initial tag set to form an enhanced tag set.

[0074] In this embodiment, starting from the second standard entity set in the target retrieval data, a multi-level traversal operation of related entities is performed in the updated multi-source heterogeneous knowledge graph. The traversal process expands outward along various predefined relation edges (such as membership, reference, constraint, synonym, etc.), exploring other entity nodes that are directly or indirectly related to the starting entity layer by layer. After the traversal is completed, all visited entity nodes are summarized to form an extended entity set containing multiple candidate entities.

[0075] In this embodiment, the overall semantic relevance between each candidate entity in the extended entity set and the second standard entity set is calculated, generating a global semantic relevance score for each candidate entity. Subsequently, all candidate entities are ranked according to this score, and based on a preset score threshold or a Top-K strategy, the most relevant subset of target entities is recalled from the ranked list, thereby forming the enhanced entity set. The Top-K strategy returns only the top K results after ranking all candidate objects, ignoring all objects ranked lower.

[0076] In this embodiment, the association between each target entity in the enhanced entity set and its domain tags in the multi-source heterogeneous knowledge graph is analyzed. Specifically, all tags attached to each target entity are checked, and tags that also exist in the initial tag set are selected. These selected tags, because they are highly relevant to both the original query intent (represented by the initial tag set) and the expanded knowledge network (represented by the enhanced entity set), are identified as valuable subsets and are ultimately aggregated to form the enhanced tag set.

[0077] S1033: Based on the target retrieval data, and based on the enhanced entity set and the enhanced tag set, perform joint retrieval on the updated structured knowledge base, the domain tag knowledge base and the multi-source heterogeneous knowledge graph, and fuse the joint retrieval results to generate a candidate knowledge document set.

[0078] In this embodiment, the attribute retrieval operation of the structured knowledge base is as follows: A query is executed using the unique identifier of the target entity in the enhanced entity set as the core key. This query retrieves all normalized attribute records (i.e., field values ​​in the records) of these target entities in the database, and obtains their associated detailed text content through association table mapping.

[0079] In this embodiment, the semantic retrieval operation of the domain tag knowledge base is as follows: using the enhanced tag set as the filtering condition, after semantic expansion using the defined hierarchy and synonym relationships, all knowledge objects (including knowledge modules, clause clusters or document categories) marked with such tags are located and retrieved in the knowledge base, and the authoritative explanations and definitional texts pre-stored by the knowledge base for these objects are obtained simultaneously.

[0080] In this embodiment, the association retrieval operation on the multi-source heterogeneous knowledge graph is performed as follows: starting from the target entity in the enhanced entity set, traversal is performed along the relational edges connected to it. The core purpose of this traversal is to utilize the inherent semantic network structure of the graph to obtain knowledge fragments rich in semantic relationships that are directly connected to these entities. Specifically, the retrieval process extracts all attribute values ​​stored internally by the target node, traverses and collects all relational edges connected to it, along with their types and attributes, visits its neighboring nodes, and obtains their descriptive text. Finally, the above triplet information (node-edge-neighboring node) is integrated to form a contextual knowledge block centered on relationships. The retrieval results from the three independent knowledge sources are deduplicated, weighted, and sorted to generate a candidate knowledge document set, which includes multiple alternative documents.

[0081] It is worth noting that this application employs a multi-knowledge base joint retrieval method, utilizing enhanced entity sets to traverse predefined business and legal relationship chains within a knowledge graph. This systematically uncovers legal provisions, precedents, and internal rules that are related to the core clauses but not literally similar. Simultaneously, it constrains the semantic boundaries of the retrieval through a domain-labeled knowledge base, combined with attribute queries from a structured knowledge base, thus forming a comprehensive and multi-dimensional retrieval system. This mechanism ensures that the final generated candidate knowledge document set is an organic whole integrating clauses, legal basis, judicial practice, and business rules, rather than fragmented pieces. This provides a solid guarantee for generating professional responses with rigorous logic and complete context, reducing claims disputes and legal risks caused by missing key knowledge connections, and overcoming the shortcomings of traditional fragmented RAG retrieval.

[0082] For example, when a user queries, "A typhoon caused a factory to collapse. I have property insurance but no additional natural disaster insurance. Can I still claim compensation?", the traditional RAG (Revenue Advisory) has a limitation: it can only recall the exclusion clause in the "Basic Terms of Property Insurance" that states, "The insurer is not liable for losses caused by natural disasters." The user receives an absolute rejection conclusion, which may be inaccurate and omit crucial legal and business grounds.

[0083] This application first identifies core entities (such as "property insurance," "natural disaster insurance," and "typhoon") through semantic parsing and standardizes them. Then, starting with these standardized entities, it traverses predefined relationship chains within a multi-source heterogeneous knowledge graph: linking the "property insurance" node to its "exclusion clauses" sub-item, thereby discovering the rules for interpreting standard clauses in relevant legal provisions; simultaneously, it retrieves the key points of judgments regarding the determination of "mixed causation" in similar cases through "judicial precedents," and extracts the internal catastrophe risk-sharing mechanism clauses of insurance companies through "reinsurance agreements." These deeply related but superficially dissimilar key knowledge points, along with specific policy clauses retrieved from a structured knowledge base and "claims dispute" related guidelines obtained from a domain-labeled knowledge base, are systematically integrated into an organic whole. Finally, based on this complete knowledge set, a response is generated, clearly stating that "although exclusion clauses apply, a comprehensive judgment must be made in conjunction with relevant legal interpretation rules, relevant precedents, and the specific cause of loss," thus providing a professional solution and reducing the misunderstandings and legal risks that may arise from traditional RAGs only returning partial exclusion clauses.

[0084] Furthermore, the structured knowledge base, domain-labeled knowledge base, and multi-source heterogeneous knowledge graph in this application can be dynamically updated. By instantly parsing new regulatory rules, compliance control is transformed from "post-event manual inspection" to "pre-event system integration," avoiding compliance vacuums caused by system update delays and reducing the risk of violations. Secondly, during knowledge updates, historical states and relationships are preserved, giving it spatiotemporal cognitive capabilities and enabling it to provide credible explanations based on historical context, enhancing customer trust.

[0085] S104: Generate preferred answer data based on the candidate knowledge document set and the large language model.

[0086] In this embodiment, as Figure 4 The diagram illustrates the process of generating preferred answer data in an embodiment of the present invention. The candidate knowledge document set includes multiple alternative documents; the method for generating preferred answer data based on the candidate knowledge document set and the large language model includes:

[0087] S1041: Determine whether the total length of the candidate knowledge document set exceeds a preset length threshold.

[0088] S1042: If the total length of the candidate knowledge document set does not exceed the preset length threshold, the candidate knowledge document set is directly input into the large language model to generate preferred answer data.

[0089] S1043: If the total length of the candidate knowledge document set exceeds a preset length threshold, then based on the enhanced entity set and the enhanced tag set, the second text blocks in each of the candidate documents are filtered to obtain a first-level candidate text block set.

[0090] S1044: Determine whether the total length of the first-level candidate text block set exceeds a preset length threshold.

[0091] S1045: If the total length of the first-level candidate text block set does not exceed the preset length threshold, the first-level candidate text block set is directly input into the large language model to generate the preferred answer data.

[0092] S1046: If the total length of the first-level candidate text block set exceeds a preset length threshold, then the first-level candidate text block set is summarized to generate a second-level summary text set.

[0093] S1047: Determine whether the total length of the secondary summary text set exceeds a preset length threshold.

[0094] S1048: If the total length of the secondary summary text set does not exceed the preset length threshold, the secondary summary text set is directly input into the large language model to generate the preferred answer data.

[0095] S1049: If the total length of the secondary summary text set exceeds a preset length threshold, the secondary summary text set is divided into multiple tertiary summary text sets that meet the preset length threshold; each of the tertiary summary text sets is input into the large language model in batches to obtain the sub-answer data corresponding to each of the tertiary summary text sets; the sub-answer data corresponding to each of the tertiary summary text sets is fused to generate the preferred answer data.

[0096] In this embodiment, the length of the retrieved candidate knowledge document set is evaluated. If it does not exceed the context window limit of the large language model, it is directly input into the model to generate the answer. If the length exceeds the limit, a multi-level processing flow is initiated: First, the second text block is semantically filtered using the enhanced entity set and the enhanced tag set to obtain a first-level candidate text block set with high relevance; if this set is still too long, it is summarized and compressed to generate a second-level summary text set; if the second-level summary text set still exceeds the limit, the final scheme of segmentation-batch processing-answer fusion is executed. Through this adaptive processing mechanism, it flexibly adapts to knowledge sets of different sizes and generates optimal answers for knowledge sets of different sizes.

[0097] In this embodiment, the knowledge-enhanced joint retrieval method further includes:

[0098] (1) Collect multiple negative feedback answer data from the output of the large language model. The negative feedback answer data indicates that the answer output by the large language model does not conform to user preferences or contains errors.

[0099] (2) Construct a failure case dataset based on the negative feedback answer data. Each negative feedback answer data point is associated with its corresponding historical user query, the historical candidate knowledge document set used to generate the answer, and the relevant dialogue history to form a failure case dataset.

[0100] (3) Based on the failure case dataset, construct a question-and-answer dataset and a preference dataset, wherein: the question-and-answer dataset includes correct answer data regenerated by expert annotators based on the historical user queries and the historical candidate knowledge document set; the preference dataset includes multiple comparison data pairs, each comparison data pair consisting of the correct answer data as a positive sample and the corresponding negative feedback answer data as a negative sample.

[0101] (4) Use the question-and-answer dataset and the preference dataset to fine-tune the large language model, and deploy the fine-tuned large language model to process subsequent user queries and generate new answers, thereby forming a continuously optimized closed-loop system.

[0102] It is worth noting that the knowledge-enhanced joint retrieval method of this application has the following advantages:

[0103] (1) By introducing a structured knowledge base, a domain-labeled knowledge base, and a multi-source heterogeneous knowledge graph, it is possible to identify and traverse the deep semantic relationships and business logic between insurance clauses, laws and regulations, and judicial precedents. This ensures that the search results are no longer limited to literal text fragments, but form an organic knowledge whole that integrates direct clauses, relevant legal provisions, the essence of precedents, and internal rules. The resulting responses have rigorous legal basis and complete business context, reducing claims disputes and legal risks caused by the lack of key knowledge connections.

[0104] (2) Through the synergistic effect of structured knowledge bases, domain-labeled knowledge bases, and multi-source heterogeneous knowledge graphs, actuarial rules, underwriting policies, and regulatory requirements can be transformed into computable knowledge units and constraints. During the generation process, these knowledge units constitute hard constraints, which can proactively trigger compliance verification processes to ensure that the output content meets professional standards and regulatory requirements. This mechanism transforms compliance control from delayed manual inspections to real-time, pre-implemented system updates, reducing compliance vacuum periods caused by policy update delays and lowering regulatory penalties and reputational risks for enterprises.

[0105] (3) By parsing, summarizing, extracting features, and standardizing matching new documents, the structured knowledge base, domain tag knowledge base, and multi-source heterogeneous knowledge graph are updated. During the update, not only the currently effective knowledge content is stored, but also historical version knowledge and its valid timestamps are retained, and a clear version replacement relationship chain is established in the knowledge graph. This enables the system to accurately understand and associate product terms, rate rules, and regulatory policies at different points in time, thereby providing accurate interpretations of queries across the policy lifecycle based on historical context, enhancing the accuracy and credibility of customer service, and effectively maintaining customer relationships and brand trust.

[0106] In the embodiments of this application, terms such as "first" and "second" are used to distinguish identical or similar items with essentially the same function and purpose. For example, "first text block" and "second text block" are used only to distinguish different text blocks and do not limit their order. Those skilled in the art will understand that terms such as "first" and "second" do not limit the quantity or execution order, and that terms such as "first" and "second" do not necessarily imply that they are different.

[0107] It should be noted that, in the embodiments of this application, the words "exemplary" or "for example" indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0108] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0109] Figure 5 This is a schematic block diagram of a knowledge-enhanced federated retrieval system provided in an embodiment of this application. For example... Figure 5 As shown, the knowledge-enhanced federated retrieval system 500 includes:

[0110] The data acquisition module 501 is used to acquire different types of original documents; the original documents include one or more of the following: insurance terms documents, insurance regulatory documents, insurance rules documents, court case insurance documents, and historical Q&A records documents.

[0111] The knowledge system construction and update module 502 is used to perform entity information extraction, high-level semantic unification, and intrinsic association mining operations on various types of original documents based on a large language model, so as to construct a structured knowledge base, a domain tag knowledge base, and a multi-source heterogeneous knowledge graph, and update the structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph when a new document is detected.

[0112] The retrieval module 503 is used to obtain user query data and perform joint retrieval based on the updated structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph to obtain a set of candidate knowledge documents.

[0113] The answer generation module 504 is used to generate preferred answer data based on the candidate knowledge document set and the large language model.

[0114] It should be understood that the specific process of each module performing the above-mentioned steps has been described in detail in the above method embodiments, and will not be repeated here for the sake of brevity.

[0115] It should also be understood that the module division in the embodiments of this application is illustrative and only represents a logical functional division; in actual implementation, there may be other division methods. Furthermore, the functional modules in the various embodiments of this application can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0116] Figure 6 This is a schematic block diagram of an electronic terminal provided in an embodiment of this application. The electronic terminal includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the knowledge-enhanced joint retrieval method described above. Figure 6 As shown, the electronic terminal includes at least one processor 601, a memory 602, at least one network interface 603, and a user interface 605. The various components in the device are coupled together via a bus system 604. It is understood that the bus system 604 is used to implement communication between these components. In addition to a data bus, the bus system 604 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in… Figure 6 The general will label all buses as bus systems.

[0117] The user interface 605 may include a monitor, keyboard, mouse, trackball, clicker, button, touchpad, or touch screen.

[0118] It is understood that memory 602 can be volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM) or programmable read-only memory (PROM), used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM) and synchronous static random access memory (SSRAM). The memories described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable categories of memory.

[0119] In this embodiment of the invention, the memory 602 is used to store various categories of data to support the operation of the electronic terminal 600. Examples of this data include: any executable program for operation on the electronic terminal 600, such as the operating system 6021 and application programs 6022; the operating system 6021 contains various system programs, such as the framework layer, core library layer, driver layer, etc., for implementing various basic services and handling hardware-based tasks. The application program 6022 may contain various applications, such as media players, browsers, etc., for implementing various application services. The knowledge-enhanced federated retrieval method provided in this embodiment of the invention can be included in the application program 6022.

[0120] The methods disclosed in the above embodiments of the present invention can be applied to processor 601, or implemented by processor 601. Processor 601 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in processor 601 or by instructions in the form of software. The processor 601 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 601 can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. General-purpose processor 601 may be a microprocessor or any conventional processor, etc. The steps of the accessory optimization method provided in the embodiments of the present invention can be directly reflected as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software module may be located in a storage medium, which is located in memory. The processor reads the information in the memory and combines it with its hardware to complete the steps of the aforementioned method.

[0121] In an exemplary embodiment, the electronic terminal 600 may be used by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), or complex programmable logic devices (CPLDs) to perform the aforementioned method.

[0122] According to the method provided in the embodiments of this application, this application also provides a computer-readable storage medium storing program code, which, when executed on a computer, causes the computer to perform... Figures 1 to 5 The method of any of the embodiments shown.

[0123] As used in this specification, the terms "component," "module," "system," etc., are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a component can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, an execution thread, a program, and / or a computer. As illustrated, applications running on computing devices and computing devices can both be components. One or more components may reside in a process and / or an execution thread, and components may be located on a single computer and / or distributed among two or more computers. Furthermore, these components can be executed from various computer-readable media on which various data structures are stored. Components can communicate, for example, via local and / or remote processes based on signals having one or more data packets (e.g., data from two components interacting with another component between a local system, a distributed system, and / or a network, such as the Internet interacting with other systems via signals).

[0124] Those skilled in the art will recognize that the various illustrative logical blocks and steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0125] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0126] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0127] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0128] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0129] In the above embodiments, the functions of each functional unit can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. A computer program product includes one or more computer instructions (programs). When the computer program instructions (programs) are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs, DVDs), or semiconductor media (e.g., solid-state disks, SSDs, etc.).

[0130] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0131] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0132] In summary, this application provides a knowledge-enhanced joint retrieval method, system, medium, and terminal. Through a structured knowledge base, a domain-labeled knowledge base, and a multi-source heterogeneous knowledge graph, it ensures that retrieval results are no longer limited to literally similar text fragments, but rather form an organic knowledge whole integrating direct clauses, relevant legal provisions, case law, and internal rules. The resulting responses possess rigorous legal basis and complete business context, reducing claims disputes and legal risks caused by missing key knowledge connections. It transforms compliance control from delayed manual checks to real-time, pre-installed system checks, reducing compliance vacuums caused by policy update delays and lowering regulatory penalties and reputational risks for enterprises. Furthermore, by retaining historical versions and establishing a version replacement chain, it ensures that policy lifecycle queries across different time points provide interpretations based on historical context, enhancing the credibility and accuracy of customer service. Therefore, this application effectively overcomes the various shortcomings of existing technologies and has high industrial applicability.

[0133] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A knowledge-enhanced joint retrieval method, characterized in that, include: Obtain different types of original documents; the original documents include one or a combination of the following: insurance terms documents, insurance regulatory documents, insurance rules documents, court case insurance documents, and historical Q&A records; Based on a large language model, entity information extraction, high-level semantic unification, and intrinsic association mining operations are performed on various types of original documents to construct a structured knowledge base, a domain tag knowledge base, and a multi-source heterogeneous knowledge graph. When a new document is detected, the structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph are updated. During the high-level semantic unification operation on the extracted named entities, the semantic similarity between the named entities is calculated, and multiple expressions pointing to the same named entity are merged into a unified standard form, thereby generating a series of first standard entities. During the intrinsic association mining operation on each first standard entity, potential connections between entities are detected from structured and unstructured data sources based on the constructed first standard entities and their attributes. Obtain user query data and perform joint retrieval based on the updated structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph to obtain a set of candidate knowledge documents; Based on the candidate knowledge document set and the large language model, optimal answer data is generated. The method of updating the structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph when a new document is detected includes: The new document is obtained, and semantic segmentation is performed on the new document to obtain multiple first text blocks; Each of the first text blocks is summarized to obtain summary text data for each of the first text blocks; The new document is summarized based on the summary text data of each of the first text blocks to generate summary text data of the new document. Based on the summary text data of the new document and the summary text data of each of the first text blocks, feature extraction is performed on each of the first text blocks to obtain the new entity relation triples, the attributes of the new entities and the labels of the new entities in the new document, and to determine the identifier of the new entities. The newly added entity relation triples, the attributes of the newly added entities, the identifiers of the newly added entities, and the tags of the newly added entities are standardized and matched. Based on the standardized matching results, the structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph are collaboratively updated.

2. The knowledge-enhanced joint retrieval method according to claim 1, characterized in that, Based on a large language model, entity information extraction, high-level semantic unification, and intrinsic association mining are performed on various types of original documents to construct a structured knowledge base, a domain-labeled knowledge base, and a multi-source heterogeneous knowledge graph. The original documents of various types are preprocessed to obtain normalized documents of various types; Based on a large language model, entity information extraction is performed on various types of standardized documents to identify and extract multiple named entities and the attributes of each named entity. A high-level semantic unification operation is performed on each of the extracted named entities to generate a series of first standard entities. Based on the attributes of each of the named entities, the attributes of each first standard entity, the tags of each first standard entity, and the corresponding identifiers are assigned to each first standard entity. Perform an intrinsic association mining operation on each of the first standard entities to obtain the relationships between them; Based on the attributes of each first standard entity, the tags of each first standard entity, the identifiers corresponding to each first standard entity, and the relationships between each first standard entity, a structured knowledge base, a domain tag knowledge base, and a multi-source heterogeneous knowledge graph are constructed.

3. The knowledge-enhanced joint retrieval method according to claim 1, characterized in that, The methods for obtaining a candidate knowledge document set include: acquiring user query data and performing joint retrieval based on the updated structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph. The user query data is parsed and feature extracted to obtain the target retrieval data; Data is retrieved based on the target, and the updated multi-source heterogeneous knowledge graph is used to obtain an enhanced entity set and an enhanced tag set. Based on the target retrieval data, and based on the enhanced entity set and the enhanced tag set, a joint retrieval is performed on the updated structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph, and the joint retrieval results are fused to generate a candidate knowledge document set.

4. The knowledge-enhanced joint retrieval method according to claim 3, characterized in that, The methods for parsing and extracting features from the user query data to obtain the target retrieval data include: The user query data is parsed, and entity feature extraction and tag feature extraction are performed on the parsed user query data to obtain an initial entity set and an initial tag set. The initial entity set is standardized to generate target retrieval data consisting of a second standard entity set and the initial tag set.

5. The knowledge-enhanced joint retrieval method according to claim 4, characterized in that, The methods for obtaining an enhanced entity set and an enhanced tag set based on the target data retrieval and the updated multi-source heterogeneous knowledge graph include: Based on the target retrieval data, a multi-level traversal operation of related entities is performed in the updated multi-source heterogeneous knowledge graph to obtain an extended entity set. Calculate the global semantic relevance score between each candidate entity in the extended entity set and the second standard entity set, and based on the ranking result of the global semantic relevance score, recall the most relevant target entities from the extended entity set to obtain the enhanced entity set; Based on the affiliation relationship between each target entity in the enhanced entity set and the initial tag set, relevant tags are selected from the initial tag set to form an enhanced tag set.

6. The knowledge-enhanced joint retrieval method according to claim 3, characterized in that, The candidate knowledge document set includes multiple alternative documents; the method for generating preferred answer data based on the candidate knowledge document set and the large language model includes: Determine whether the total length of the candidate knowledge document set exceeds a preset length threshold; If the total length of the candidate knowledge document set does not exceed the preset length threshold, the candidate knowledge document set is directly input into the large language model to generate preferred answer data. If the total length of the candidate knowledge document set exceeds a preset length threshold, then based on the enhanced entity set and the enhanced tag set, the second text blocks in each of the candidate documents are filtered to obtain a first-level candidate text block set. Determine whether the total length of the first-level candidate text block set exceeds a preset length threshold; If the total length of the first-level candidate text block set does not exceed the preset length threshold, the first-level candidate text block set is directly input into the large language model to generate the preferred answer data; If the total length of the first-level candidate text block set exceeds a preset length threshold, then the first-level candidate text block set is summarized to generate a second-level summary text set. Determine whether the total length of the secondary summary text set exceeds a preset length threshold; If the total length of the secondary summary text set does not exceed the preset length threshold, the secondary summary text set is directly input into the large language model to generate the optimal answer data; If the total length of the secondary summary text set exceeds a preset length threshold, the secondary summary text set is divided into multiple tertiary summary text sets that meet the preset length threshold; each of the tertiary summary text sets is input into the large language model in batches to obtain the sub-answer data corresponding to each of the tertiary summary text sets; the sub-answer data corresponding to each of the tertiary summary text sets is fused to generate the preferred answer data.

7. A knowledge-enhanced joint retrieval system, characterized in that, include: The data acquisition module is used to acquire different types of original documents; the original documents include one or a combination of the following: insurance terms documents, insurance regulatory documents, insurance rules documents, court case insurance documents, and historical Q&A records documents; The knowledge system construction and update module is used to perform entity information extraction, high-level semantic unification, and intrinsic association mining operations on various types of original documents based on a large language model, in order to construct a structured knowledge base, a domain tag knowledge base, and a multi-source heterogeneous knowledge graph. When a new document is detected, the module updates the structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph. The method of updating the structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph when a new document is detected includes: acquiring the new document and performing semantic segmentation on the new document to obtain multiple first text blocks; summarizing each first text block to obtain summary text data for each first text block; summarizing the new document based on the summary text data of each first text block to generate summary text data for the new document; and summarizing the summary text data of the new document and the summarization data of each first text block. The abstract text data is used to extract features from each of the first text blocks to obtain the new entity relation triples, attributes, and tags of the new entities in the new document, and to determine the identifier of the new entities. Standardized matching is then performed on the new entity relation triples, attributes, identifiers, and tags, and the structured knowledge base, domain tag knowledge base, and multi-source heterogeneous knowledge graph are collaboratively updated based on the standardized matching results. When performing high-level semantic unification on the extracted named entities, the semantic similarity between the named entities is calculated, and multiple expressions pointing to the same named entity are merged into a unified standard form, thereby generating a series of first standard entities. When performing intrinsic association mining on each first standard entity, potential connections between entities are detected from structured and unstructured data sources based on the constructed first standard entities and their attributes. The retrieval module is used to obtain user query data and perform joint retrieval based on the updated structured knowledge base, the domain tag knowledge base, and the multi-source heterogeneous knowledge graph to obtain a set of candidate knowledge documents. The answer generation module is used to generate preferred answer data based on the candidate knowledge document set and the large language model.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the knowledge-enhanced joint retrieval method as described in any one of claims 1 to 6.

9. An electronic terminal, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the knowledge-enhanced joint retrieval method as described in any one of claims 1 to 6.

Images