Intelligent report generation system and method based on knowledge graph enhancement
By introducing the collaboration between the knowledge graph construction module and the large model module, the issues of accuracy, real-time performance, and interpretability in generating structured reports from large models are resolved, achieving efficient and transparent report generation suitable for scenarios such as enterprise data analysis and industry research.
Patent Information
- Application Number
- CN202511938565.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-22
- Publication Date
- 2026-04-28
AI Technical Summary
In existing technologies, large models rely on their own knowledge when generating structured reports, which can easily lead to inaccurate conclusions, excessive consumption of computational resources, and opaque reasoning processes, failing to meet the requirements for accuracy, real-time performance, and interpretability.
A knowledge graph construction module is introduced, which combines entity, entity attribute and relationship definition with large model module for querying and generation. The efficient query capability of knowledge graph reduces complex relationship reasoning, provides clear traceability basis and query language adaptation, and adopts a single large model to replace the two large models to simplify the system.
It improves the accuracy and reliability of reports, meets real-time requirements, enhances the interpretability of results, and reduces system complexity and hardware resource consumption, while being compatible with various knowledge graph types and system architectures.
Smart Images
Figure CN121936562A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of natural language processing and knowledge graph fusion technology. Specifically, it relates to an intelligent report generation system and method based on knowledge graph enhancement, which can be applied to intelligent generation scenarios of various structured reports such as enterprise data analysis reports, industry research reports, and information summary reports. Background Technology
[0002] With the development of artificial intelligence technology, large models have been applied in the field of structured report generation. In the existing technology, when generating structured reports using large models, information extraction and logical reasoning are mainly carried out by the model's own intrinsic knowledge. When dealing with complex relationships between entities, large models usually make inferences based on existing corpora, lacking the assistance of external structured knowledge.
[0003] The aforementioned existing technologies have three key problems: First, large models rely solely on their own knowledge for reasoning, which can easily lead to inaccurate report conclusions due to incomplete knowledge, information lag, or logical biases, failing to meet the needs of scenarios with high requirements for result reliability; second, reasoning about complex relationships between entities consumes a large amount of computing resources, resulting in slow report generation response speeds and difficulty in adapting to real-time requirements; third, the reasoning process has black-box characteristics, making it impossible for users to trace the source of the report conclusions, reducing the credibility and interpretability of the results.
[0004] In fields where there are clear requirements for the accuracy, timeliness, and interpretability of reports (such as enterprise data analysis and industry research), the shortcomings of existing technologies have become an obstacle to the practical application of intelligent report generation technology. Therefore, there is an urgent need for a structured report generation solution that can make up for the above deficiencies. Summary of the Invention
[0005] In view of this, the purpose of this invention is to overcome the shortcomings of existing related technologies. The purpose of this invention is to provide an intelligent report generation system and method based on knowledge graph enhancement, aiming to solve the core defects of existing technologies that rely solely on the inherent knowledge of large models to generate structured reports: First, by introducing structured knowledge (including entities, entity attributes, and relationships between entities) from knowledge graphs, it aims to compensate for the incomplete knowledge and information lag of large models, improve the accuracy of entity relationship acquisition, and thus ensure the reliability of report conclusions; Second, by leveraging the efficient query capabilities of knowledge graphs, it aims to reduce the deep reasoning consumption of large models on complex relationships, accelerate the response speed of report generation, and meet real-time requirements; Third, by relying on the clear definition and source of entity relationships in knowledge graphs, it aims to solve the problem of opaque reasoning processes in large models, enabling users to trace the basis of report conclusions and enhance the credibility and interpretability of results; At the same time, it also aims to improve the adaptability of the technology to different knowledge graph types and different system architecture requirements by providing flexible solutions such as query language replacement (e.g., SPARQL replacing Cypher) and single large model replacing dual large models, further enhancing the practicality and versatility of the solution.
[0006] To achieve the above objectives, a first aspect of the present invention provides an intelligent report generation system based on knowledge graph enhancement, comprising: The knowledge graph construction module is used to construct a knowledge graph using a graph database (such as neo4j). The knowledge graph includes entities, entity attributes, and relationships between entities, and defines the types of relationships between entities and the attribute rules of entity attributes. The first major model module is used to receive the user's original natural language question and the knowledge graph construction logic and query examples output by the knowledge graph construction module, and to convert the natural language question into a structured query statement adapted to the knowledge graph. The knowledge graph query module is used to receive the structured query statement output by the first large model module, perform query operations in the knowledge graph, and obtain structured query results containing relationships between entities. The second major model module is used to receive the user's original natural language questions and the structured query results output by the knowledge graph query module, and generate report content with clear structure and well-defined conclusions. The data storage module is used to store the structured query results obtained by the knowledge graph query module. The data storage module includes a cache layer and a database. The cache layer is used to call the structured query results in real time, and the database is used to store historical structured query results. At the same time, the data storage module is also used to generate collection logs, which include records of the integrity of the collection of query results and records of abnormal retry.
[0007] Furthermore, the knowledge graph construction logic and query examples are in Cypher statement form and / or SPARQL statement form, and a query language adapter is set to perform equivalent conversion of the structured query statements between Cypher and SPARQL.
[0008] Furthermore, the first and second large model modules are replaced by a single large model with multi-stage processing capabilities, which sequentially completes the entire process of converting natural language questions into structured query statements, obtaining knowledge graph query results, and generating report content.
[0009] Furthermore, the collection log also records the query time, query statement type, query time, and the number of entities associated with the query result, and establishes an association with the unique identifier of the structured query result to support traceability.
[0010] Furthermore, it also includes a result alignment and evidence chain module, which is used to standardize the structured query results into evidence units and record node / edge identifiers, data sources and timestamps. The report content references the identifiers of the evidence units to achieve traceability of evidence.
[0011] Another aspect of the present invention provides a knowledge graph-based intelligent report generation method, which is implemented based on the above-mentioned system and includes: S1. Constructing a knowledge graph: Constructing a knowledge graph using a graph database (such as neo4j), the knowledge graph includes entities, entity attributes, and relationships between entities, and defines the types of relationships between entities and the attribute rules of entity attributes; the construction of the knowledge graph can be automatically completed in the graph database (such as neo4j) by generating structured graph construction statements based on the natural language logic input by the user through a large model. S2. Knowledge Embedding and Training: The knowledge graph construction logic and query examples from step S1 are used as the first model of the corpus input, enabling it to learn to convert natural language questions into structured query statements that are adapted to the knowledge graph; S3. Query statement conversion: Receive the natural language question input by the user, and convert the natural language question into a structured query statement using the first major model after fine-tuning the corpus; S4. Obtain query results: Execute the structured query statement generated in step S3 in the knowledge graph constructed in step S1 to obtain structured query results containing relationships between entities; S5. Generation and Storage: Input the user's natural language question and the structured query results from step S4 into the second major model to generate a report with a clear structure and well-defined conclusions; and store the structured query results in the cache layer and the database respectively, wherein the cache layer is used for real-time retrieval and the database is used to store historical structured query results; at the same time, a collection log is generated, which includes collection integrity records and abnormal retry records.
[0012] Furthermore, the construction logic and query examples in steps S1 / S2 are Cypher statements and / or SPARQL statements, and when the knowledge graph type changes, they are equivalently converted to the structured query statements in step S3 through query language adaptation.
[0013] Furthermore, steps S2 to S5 are completed sequentially by a single large model with multi-stage processing capabilities.
[0014] Furthermore, the collection log generated in step S5 also records the query time, query statement type, query time, and the number of entities associated with the query result, and establishes a unique identifier association with the structured query result in step S4 to support traceability.
[0015] Furthermore, before returning the structured query results, step S4 performs deduplication and sorting, with the sorting criteria including at least one of factual confidence, source credibility, and time freshness; and between step S4 and step S5, result alignment and evidence chain construction are performed, and the evidence unit identifier is cited and confidence level is added in the report.
[0016] The present invention, by adopting the above technical solution, has at least the following beneficial effects: This invention, through its core technical solution of "knowledge graph and large model collaboration," precisely addresses the core shortcomings of existing intelligent report generation methods that rely solely on the inherent knowledge of large models, such as insufficient accuracy, low response efficiency, and opaque reasoning processes. Furthermore, through flexible technical adaptation design, it further enhances the practicality and versatility of the solution. Specific beneficial effects are as follows: This invention addresses the problem in existing technologies where large models rely solely on their own parameterized knowledge to infer entity relationships, leading to "conclusion bias" due to incomplete knowledge and outdated information. The invention constructs a structured knowledge system containing entities, entity attributes, and clearly defined relationship types through a knowledge graph construction module. Combined with a first-level model module that converts user natural language questions into precise structured query statements, the knowledge graph query module directly retrieves objective results of entity relationships—results derived from predefined explicit associations in the knowledge graph (not probabilistic inferences from large models). A second-level model module then generates a report based on the user's question and the structured results. This fundamentally compensates for the knowledge deficiencies of large models, avoids unfounded conclusions or factual biases, and provides stronger factual support for the report's conclusions. This significantly improves the accuracy and reliability of structured reports, making it particularly suitable for scenarios with stringent requirements for result accuracy, such as enterprise data analysis and industry research.
[0017] This invention alleviates the problem of response delays caused by the high computational cost of deep semantic reasoning in traditional large-scale models when dealing with complex relationships between entities. Leveraging the efficient query capabilities of knowledge graphs, it transforms "complex relationship reasoning" into "structured querying of the knowledge graph." The first large-scale model module only needs to convert "natural language and query statements," without performing complex relational calculations. The knowledge graph query module can quickly match and output entity relationship results, significantly reducing the reasoning time of large-scale models. Simultaneously, the data storage module uses a caching layer to access historical query results in real time, further shortening response time in repetitive query scenarios. The overall process significantly improves report generation efficiency compared to existing technologies, meeting real-time report generation requirements and adapting to business scenarios involving multi-turn conversations or high-frequency queries.
[0018] This invention addresses the problem of the "black box" nature of reasoning processes in existing large-scale model-generated reports, which prevents users from tracing the source of conclusions and reduces the credibility of generated reports in professional decision-making scenarios. The invention utilizes a structured knowledge storage and evidence chain construction design based on knowledge graphs to ensure that report conclusions have clear traceability: entities, attributes, and relationships in the knowledge graph are clearly defined and have sources. The structured results output by the knowledge graph query module can form a complete evidence chain through "node / edge identifiers, data sources, and timestamps." When the second large-scale model module generates the report, it directly references the entity relationships associated with this evidence chain. Users can trace back to the corresponding knowledge graph data through the implicit evidence unit identifiers in the report, solving the problem of opaque reasoning processes and significantly enhancing the interpretability and credibility of reports, making it particularly suitable for fields such as finance and government where "conclusion traceability" is required.
[0019] This invention expands the applicability of the solution and reduces implementation costs through multi-dimensional technical adaptation design: On the one hand, by supporting equivalent conversion of various graph query languages such as Cypher and SPARQL through a query language adapter, the appropriate query method can be flexibly selected according to the type of knowledge graph (such as attribute graph, RDF graph), without having to reconstruct the entire system due to changes in the type of knowledge graph, thus improving the versatility of the technical solution; on the other hand, it supports replacing the first and second large model modules with a "single large model with multi-stage processing capabilities", reducing the data transmission and interaction costs between the two models, simplifying the system architecture, reducing the difficulty of model deployment and maintenance and hardware resource consumption, while avoiding compatibility issues in the process of dual-model collaboration, providing a more lightweight implementation option for small and medium-sized application scenarios.
[0020] The data storage module employs a layered storage design of "caching layer and database," enabling real-time retrieval of query results (caching layer) while ensuring persistent storage of historical data (database), facilitating subsequent data review and result reuse across multiple sessions. Simultaneously, the collection logs not only include records of query result collection integrity and exception retry records but also supplement details such as query time, statement type, and number of associated entities. These logs are uniquely linked to the query results, allowing for complete traceability of the entire "input-query-output" data chain for each report generation. This facilitates troubleshooting and business auditing, ensuring the continuity and manageability of the intelligent report generation business.
[0021] Based on the core objectives of improving the accuracy, efficiency, and interpretability of reports, this invention further enhances the practicality and feasibility of the solution through flexible technology adaptation and a comprehensive data management design. It can be widely adapted to various structured report generation scenarios, providing reliable support for the industrial application of intelligent reporting technology. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 This is a schematic diagram of the intelligent report generation system based on knowledge graph enhancement according to the present invention. Detailed Implementation
[0024] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.
[0025] Example 1 like Figure 1 As shown, this embodiment provides an intelligent report generation system based on knowledge graph enhancement, including: The knowledge graph construction module is used to construct a knowledge graph using a graph database (such as neo4j). The knowledge graph includes entities, entity attributes, and relationships between entities, and defines the types of relationships between entities and the attribute rules of entity attributes. The first major model module is used to receive the user's original natural language questions and the knowledge graph construction logic and query examples output by the knowledge graph construction module, and learn to convert the natural language questions into structured query statements that are adapted to the knowledge graph. The knowledge graph query module is used to receive the structured query statement output by the first large model module, perform query operations in the knowledge graph, and obtain structured query results containing relationships between entities. The second major model module is used to receive the user's original natural language questions and the structured query results output by the knowledge graph query module, and generate report content with clear structure and well-defined conclusions. The data storage module is used to store the structured query results obtained by the knowledge graph query module. The data storage module includes a cache layer and a database. The cache layer is used to call the structured query results in real time, and the database is used to store historical structured query results. At the same time, the data storage module is also used to generate collection logs, which include records of the integrity of the collection of query results and records of abnormal retry.
[0026] The user interaction module is used to receive user input of natural language questions and report feedback information, serving as the interaction entry point between the system and the user.
[0027] As one implementation method, the knowledge graph construction logic and query examples described in this embodiment are in Cypher statement form and / or SPARQL statement form, and a query language adapter is set to perform equivalent conversion between Cypher and SPARQL for the structured query statements.
[0028] As one implementation method, in this embodiment, the first large model module and the second large model module are replaced by a single large model with multi-stage processing capabilities, so as to sequentially complete the entire process of converting natural language questions into structured query statements, obtaining knowledge graph query results, and generating report content.
[0029] As one implementation method, the collection log in this embodiment also records the query time, query statement type, query time, and the number of entities associated with the query result, and establishes an association with the unique identifier of the structured query result to support traceability.
[0030] As one implementation method, this embodiment also includes a result alignment and evidence chain module, which is used to standardize the structured query results into evidence units and record node / edge identifiers, data sources and timestamps. The report content references the identifiers of the evidence units to achieve traceability of evidence.
[0031] like Figure 1 As shown in the figure, this invention illustrates the core process logic of the knowledge graph-based intelligent report generation system. The specific interaction process is as follows: Knowledge Embedding Stage: The knowledge graph construction logic (such as Cypher statement form) is input into the large model A (query) through the "knowledge embedding" process, so that the large model A learns the conversion rules of "natural language problems and knowledge graph structured query statements" and completes the fine-tuning of the model corpus.
[0032] User query and processing stage: After the user inputs a "query question" through the user interaction module, the question is transferred to the big model A (query); the big model A generates a structured query statement adapted to the knowledge graph based on the trained transformation rules, and sends the statement to the knowledge graph.
[0033] Knowledge graph query and result feedback stage: After receiving the structured query statement, the knowledge graph executes the query operation, extracts and outputs the structured result of "relationship between entities", and sends it back to the large model B (generated).
[0034] Report generation and feedback phase: The large model B (generating) receives the "original query question from the user" and the "entity relationship results returned by the knowledge graph". After processing, it generates a report with a clear structure and well-defined basis. After the report is output to the user through the user interaction module, the user can send their opinions on the report back to the system through the "feedback" path for subsequent model optimization or process iteration.
[0035] This embodiment provides an intelligent report generation system and method based on knowledge graph enhancement. Its core lies in combining the structured knowledge of knowledge graphs with the natural language processing capabilities of large models to improve the accuracy, efficiency, and interpretability of report generation.
[0036] The technical solution includes the following steps: Construct a knowledge graph, including entities, entity attributes, and relationships between entities, and define relationship types and attributes; The knowledge graph construction logic (e.g., in the form of Cypher statements) and query examples are input into the large model A as corpus, enabling it to learn how to convert natural language questions into structured query statements; After the user inputs a question, the large model A converts it into a corresponding query statement, executes the query in the knowledge graph, and obtains the structured results of entity relationships; The query results, along with the user's original question, are input into the large model B, which generates a well-structured and clearly defined report.
[0037] The report content is formatted using a pre-set template library, and includes chapter structure, data charts, and evidence citations.
[0038] Cypher can be replaced by other query languages such as SPARQL to adapt to different types of knowledge graphs; A single large model with multi-stage processing capabilities can be used to replace A and B, completing the entire process from problem analysis to report generation, thus simplifying the system structure.
[0039] This addresses the issue of insufficient accuracy in entity relationship reasoning relying solely on the knowledge of large models in existing technologies. By introducing knowledge graphs, leveraging their structured knowledge storage and precise query capabilities, and combining them with the natural language processing capabilities of large models, the acquisition of relationships between entities becomes more accurate, thereby improving the reliability of the reports.
[0040] This alleviates the problem of low reasoning efficiency. Knowledge graph queries are highly efficient. After the large model A transforms the question into a query statement, it can quickly retrieve the required relationships from the knowledge graph, reducing the time for deep reasoning in the large model and improving the overall response speed.
[0041] This solves the problem of the lack of interpretability in the reasoning process. The relationships between entities in the knowledge graph have clear definitions and origins. The large model B generates answers based on these clear relationships, enabling users to understand the basis of the conclusions and enhancing the credibility of the results.
[0042] Using different query statement replacement schemes in the knowledge embedding stage increases the flexibility and adaptability of the technical solution. Appropriate query statement forms can be selected according to the type of knowledge graph and application scenario, thereby improving the versatility of the technology.
[0043] The approach of replacing two large models with a single large model reduces the cost of data transmission and interaction between models, simplifies the system architecture, and lowers the system's maintenance difficulty and operating costs.
[0044] Example 2 This embodiment provides a knowledge graph-enhanced intelligent report generation method, which is implemented based on the above-described system and includes: S1. Knowledge Graph Construction: Data is collected from multiple heterogeneous data sources (such as enterprise databases, public APIs, text files, etc.), and a knowledge graph is constructed using a graph database (such as Neo4j). The knowledge graph includes entities, entity attributes, and relationships between entities, and defines the types of relationships between entities and the attribute rules of entity attributes. The construction of the knowledge graph can be automated by generating structured graph construction statements based on the natural language logic input by the user through a large model, and completing the graph construction in the graph database automatically. During the construction process, the validity of entity attributes is validated, invalid attributes and duplicate entities are removed, and the rationality of relationships between entities is manually or automatically reviewed. S2. Knowledge Embedding and Training: Input the knowledge graph construction logic (such as Cypher / SPARQL statements) and multiple sets of query examples (including mapping pairs between user natural language questions and corresponding structured query statements) from step S1 into the first large model, and use corpus fine-tuning to enable it to learn to convert natural language questions into structured query statements that are adapted to the knowledge graph. S3. Query statement conversion: Receive the natural language question input by the user, and convert the natural language question into a structured query statement through the first major model after corpus fine-tuning; the generated query statement needs to undergo syntax and semantic rationality checks to ensure that it can be executed in the knowledge graph; S4. Obtain query results: Execute the structured query statement generated in step S3 in the knowledge graph constructed in step S1. Concurrent queries and timeout control mechanisms are supported to obtain structured query results containing relationships between entities. Before executing the query, the statement can be adapted to the language through the query language adapter. After the query, the results are deduplicated and sorted (based on fact confidence, source credibility, and time freshness). S5. Generation and Storage: Input the user's natural language question and the structured query results from step S4 into the second main model. Based on the preset report template library (including chapter structure, chart style, and evidence citation rules), generate a report with a clear structure and well-defined conclusions. Store the structured query results in a cache layer (with cache expiration time and hit rate monitoring) and a database (supporting data backup and recovery). The cache layer is used for real-time retrieval, and the database is used to store historical structured query results. At the same time, a collection log is generated, which includes collection integrity records, abnormal retry records, query time, statement type, time consumption, and number of associated entities, and is associated with a unique identifier of the query result.
[0045] As one implementation method, in this embodiment, the construction logic and query examples in steps S1 / S2 are Cypher statements and / or SPARQL statements, and when the knowledge graph type changes, the query language is adapted to perform an equivalent conversion to the structured query statement in step S3.
[0046] In one implementation, steps S2 to S5 in this embodiment are completed sequentially by a single large model with multi-stage processing capabilities. This large model achieves the entire process of "problem analysis, query generation, result association, and report generation" through multi-task learning.
[0047] As one implementation method, the collection log generated in step S5 in this embodiment also records the query time, query statement type, query time, and the number of entities associated with the query result, and establishes a unique identifier association with the structured query result in step S4 to support traceability. Users can use the identifier to reverse query the source of the original data in the knowledge graph.
[0048] In one implementation, step S4 in this embodiment performs deduplication and sorting before returning the structured query results. The sorting criteria include at least one of factual confidence, source credibility, and time freshness. Between step S4 and step S5, result alignment and evidence chain construction are performed, mapping the structured results into evidence units in triple or JSON format, recording node / edge identifiers, data sources, and timestamps. The report references the evidence unit identifiers and adds confidence labels (confidence is generated based on the confidence scores of relations in the knowledge graph, ranging from 0 to 100).
[0049] As one implementation method, this method also includes an exception handling process: if a query failure or model generation error occurs in steps S3-S5, an automatic retry mechanism is triggered (up to 3 retries), and an exception log is recorded for subsequent analysis; at the same time, user feedback on the report is received, and the large model training data and report template library are iteratively optimized based on the feedback.
[0050] The knowledge graph-based intelligent report generation method described in this embodiment, with knowledge graph construction, model training, query transformation, result acquisition, report generation, and storage as its core processes, achieves three key breakthroughs through the synergy of the structured knowledge of the knowledge graph and the natural language processing capabilities of large-scale models: First, by leveraging the precise entity relationship storage and querying capabilities of the knowledge graph, it solves the problem of inaccurate conclusions caused by traditional large-scale models relying solely on intrinsic knowledge, ensuring that report conclusions have clear factual support; second, by replacing deep reasoning of large-scale models with efficient querying of the knowledge graph, it significantly improves report generation efficiency and meets the needs of real-time scenarios; third, by utilizing the traceability characteristics of the knowledge graph, it constructs a full-link evidence chain from "user question" to "report conclusion," making the reasoning process transparent and enhancing the credibility of the report. Simultaneously, the method improves its adaptability to different knowledge graph types and system architecture requirements through flexible design such as query language adaptation and replacement of single large-scale models, lowering the technical implementation threshold and making it widely applicable to various structured report generation scenarios such as enterprise data analysis and industry research.
[0051] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
Claims
1. A knowledge graph-based intelligent report generation system, characterized in that, include: The knowledge graph construction module is used to construct a knowledge graph through a graph database. The knowledge graph includes entities, entity attributes, and relationships between entities, and defines the types of relationships between entities and the attribute rules of entity attributes. The first major model module is used to receive the user's original natural language question and the knowledge graph construction logic and query examples output by the knowledge graph construction module, and to convert the natural language question into a structured query statement adapted to the knowledge graph. The knowledge graph query module is used to receive the structured query statement output by the first large model module, perform query operations in the knowledge graph, and obtain structured query results containing relationships between entities. The second major model module is used to receive the user's original natural language questions and the structured query results output by the knowledge graph query module, and generate report content with clear structure and well-defined conclusions. The data storage module is used to store the structured query results obtained by the knowledge graph query module. The data storage module includes a cache layer and a database. The cache layer is used to call the structured query results in real time, and the database is used to store historical structured query results. At the same time, the data storage module is also used to generate collection logs, which include records of the integrity of the collection of query results and records of abnormal retry.
2. The system as described in claim 1, characterized in that, The knowledge graph construction logic and query examples are in Cypher and / or SPARQL statement form, and a query language adapter is set to perform equivalent conversion of the structured query statements between Cypher and SPARQL.
3. The system as described in claim 1, characterized in that, The first and second major model modules are replaced by a single major model with multi-stage processing capabilities, which sequentially completes the entire process of converting natural language questions into structured query statements, obtaining knowledge graph query results, and generating report content.
4. The system as described in claim 1, characterized in that, The collection log also records the query time, query statement type, query time, and the number of entities associated with the query result, and associates it with the unique identifier of the structured query result to support traceability.
5. The system as described in claim 1, characterized in that, It also includes a result alignment and evidence chain module, which is used to standardize the structured query results into evidence units and record node / edge identifiers, data sources and timestamps. The report content references the identifiers of the evidence units to achieve traceability of evidence.
6. A method for generating intelligent reports based on knowledge graph enhancement, characterized in that, include: S1. Constructing a knowledge graph: Constructing a knowledge graph through a graph database [1], the knowledge graph includes entities, entity attributes and relationships between entities, and defines the type of relationships between entities and the attribute rules of entity attributes; the construction of the knowledge graph can be automatically completed in the graph database by generating structured graph construction statements based on the natural language logic input by the user through a large model; S2. Knowledge Embedding and Training: The knowledge graph construction logic and query examples from step S1 are used as the first model of the corpus input, enabling it to learn to convert natural language questions into structured query statements that are adapted to the knowledge graph; S3. Convert query statement: Receive the natural language question input by the user, and convert the natural language question into a structured query statement through the first major model after corpus fine-tuning [2]; S4. Obtain query results: Execute the structured query statement generated in step S3 in the knowledge graph constructed in step S1 to obtain structured query results containing relationships between entities; S5. Generation and Storage: Input the user's natural language question and the structured query results from step S4 into the second major model to generate a report with a clear structure and well-defined conclusions; and store the structured query results in the cache layer and the database respectively, wherein the cache layer is used for real-time retrieval and the database is used to store historical structured query results; at the same time, a collection log is generated, which includes collection integrity records and abnormal retry records.
7. The method as described in claim 6, characterized in that, The construction logic and query examples in steps S1 / S2 are Cypher statements and / or SPARQL statements, and when the knowledge graph type changes, the query language is adapted to perform an equivalent conversion to the structured query statement in step S3.
8. The method as described in claim 7, characterized in that, Steps S2 to S5 are completed sequentially by a single large model with multi-stage processing capabilities.
9. The method as described in claim 8, characterized in that, The collection log generated in step S5 also records the query time, query statement type, query time, and the number of entities associated with the query results, and establishes a unique identifier association with the structured query results in step S4 to support traceability.
10. The method as described in claim 9, characterized in that, Step S4 performs deduplication and sorting before returning the structured query results. The sorting criteria include at least one of factual confidence, source credibility, and time freshness. Between steps S4 and S5, result alignment and evidence chain construction are performed, and evidence unit identifiers are cited and confidence level labels are added to the report.