Data insight report generation method and system based on knowledge enhancement and fact verification

Abstract

The application discloses a data insight report generation method and system based on knowledge enhancement and fact verification. A domain index relationship knowledge graph is constructed by training a large language model; the insight is mapped to the graph, a weighted insight graph is constructed, and spectral clustering is performed to divide the insight theme area; a viewpoint unit is generated by an intelligent agent, and consistency verification and correction are performed by a fact verification intelligent agent; user intent is analyzed, candidate themes are screened, a report outline is generated by heuristic search arrangement based on SRCI score, semantic progression and logic entropy calculation; finally, verified viewpoint units are integrated according to the outline, and the final report is output. The application realizes the generation of a fully automatic and high-credibility report, and improves accuracy, logic and explainability.

Description

Technical Field

[0001] This invention relates to the field of data generation technology, and in particular to a method and system for generating data insight reports based on knowledge enhancement and fact verification. Background Technology

[0002] As enterprises deepen their digital transformation, especially in fields like agricultural production where the demand for automated analytical reports is growing, traditional technical solutions face significant challenges. Existing technologies mainly include traditional business intelligence (BI) reporting systems, automated writing software based on fixed templates, and early-generation retrieval-enhanced generation solutions based on large language models. These solutions have obvious shortcomings in complex business scenarios.

[0003] First, these methods typically rely on static domain knowledge graphs or single text similarity methods for insight clustering, failing to effectively collaboratively model the graph structure relationships between semantic information and indicators. This results in weak logical connections and a lack of granularity in the generated analytical viewpoints, making it difficult to form a clear and coherent analytical narrative. Second, report planning is rigid, either limited to "fixed templates + fill-in-the-blank" generation, leading to monotonous content, or relying on large language models for free interpretation, easily resulting in scattered content, lack of focus, and low information entropy. They lack a mechanism to dynamically assess the value of the topic and automatically arrange the logical outline according to user intent. Most importantly, existing methods suffer from the "black box generation" problem; the viewpoints generated by the model lack explicit and traceable binding with underlying data facts, easily leading to numerical errors, logical contradictions, or factual illusions, casting doubt on their authority in data-sensitive decision-making scenarios.

[0004] Therefore, there is an urgent need for an automated report generation method and system that can construct deep logical connections, implement rigorous fact verification, and achieve adaptive outline planning. Summary of the Invention

[0005] To address at least one of the aforementioned technical problems, this invention proposes a data insight report generation method and system based on knowledge enhancement and fact verification.

[0006] The first aspect of this invention provides a method for generating data insight reports based on knowledge enhancement and fact verification, comprising: Based on domain corpus, a large language model is trained for domain knowledge and entity relationship extraction. The trained model is used to automatically extract indicator entities and their relationships from domain text, and a domain indicator relationship knowledge graph is constructed with indicator entities as nodes and indicator relationships as edges. The main indicator of each insight in the set of insights to be analyzed is mapped to the domain indicator relationship knowledge graph to construct a structure-semantic weighted insight graph, and the structure-semantic weighted insight graph is divided into K insight topic regions based on the spectral clustering method; For the aforementioned insight topic region, the writing agent generates opinion units containing viewpoints, conclusion types, and verifiable evidence chains. The fact verification agent performs consistency verification and correction on the opinion units in terms of numerical values, logic, and knowledge graph constraints to obtain an optimized insight topic region. The system parses user input to obtain user intent features, filters candidate topics that are semantically related to the user intent features from K optimized insight topic regions, and assigns chapter positions to each candidate topic through heuristic search based on the SRCI four-dimensional scoring model, semantic progression relationship between topics and logical entropy calculation, and automatically arranges and generates a report outline. Based on the structure of the report outline, the viewpoints corresponding to the themes of each chapter are organized and integrated to output the final data insight report.

[0007] In this solution, the domain knowledge training and entity relation extraction training of the large language model based on domain corpus are performed. The trained model is then used to automatically extract indicator entities and their relations from the domain text. A domain indicator relation knowledge graph is constructed using indicator entities as nodes and indicator relations as edges. Specifically: A domain corpus is constructed by acquiring production management documents, indicator business definitions, and internal analysis reports for the target domain. The domain corpus is then cleaned of duplicate data and structured to form a domain training corpus. Based on the domain training corpus, a large language model is trained to understand the professional concepts and indicator semantics of the target domain, resulting in a domain-specific large language model with domain knowledge understanding capabilities. Based on the imperative supervised fine-tuning method, the domain corpus is used to construct training data with domain text and entity relation extraction instructions as input and corresponding structured entity relations as supervision signals. The domain large language model is trained on the entity relation extraction task to obtain a trained model with the ability to automatically extract indicator entities and their relations. The target domain text is input into the trained model, which performs semantic parsing on the text, identifies and outputs the indicator entities in the text and the relationships between the indicator entities; The index relationship results output by the model are uniformly converted into a structured representation in the form of triples, and attribute information is added to nodes and edges; Using the transformed indicator entities as graph nodes and the relationships between indicator entities as edges in the graph, a domain indicator relationship knowledge graph is constructed.

[0008] In this solution, the main indicator of each insight in the set of insights to be analyzed is mapped to the domain indicator relationship knowledge graph to construct a structure-semantic weighted insight graph. Then, based on the spectral clustering method, the structure-semantic weighted insight graph is divided into K insight topic regions. Specifically: The set of insights to be analyzed is parsed to obtain the main indicators corresponding to each insight object and its associated insight interpretation text. Each insight consists of insight interpretation text, a list of design indicators, insight name, and insight calculation data dictionary. The main indicators of each insight are used as query keys and mapped to the domain indicator relationship knowledge graph. The nodes corresponding to each indicator in the graph are identified, and the discrete insight objects are connected to construct an original insight graph that conforms to the indicator relationship structure based on the relationship between indicators recorded in the graph. Formalize the original insight map as Where V represents the set of nodes and E represents the set of edges, each node In this context, indicator represents the indicator name, and explain represents the corresponding insight interpretation text. Define edge relationship types and assign preset structural weights to different types of edge relationships, including , Calculate the semantic similarity between the insight interpretation texts corresponding to the two nodes connected by each edge, and use this as the semantic weight of the edge. The formula is as follows: , in, This represents a function for calculating semantic similarity. The final weight of each edge is defined as the product of the structural weight and the semantic weight corresponding to that edge, thus constructing a structure-semantic weighted insight graph, which is formally represented as: , Construct the weighted adjacency matrix of the structure-semantic weighted insight graph: , Calculate the degree matrix of the nodes based on the weighted adjacency matrix: , Then, the normalized graph Laplacian matrix is ​​constructed: , in, Given the identity matrix, perform eigenvalue decomposition on the Laplacian matrix L: ,in, For the k-th eigenvalue, sorted in ascending order. For each eigenvector, select the eigenvectors corresponding to the K smallest eigenvalues ​​to form the spectral embedding matrix of the node in the low-dimensional space: , Embed the spectrum in each row of matrix U Consider as a node The low-dimensional representation of the data is obtained, and the k-means clustering algorithm is used to partition the representation space: , Ultimately, each cluster This corresponds to a semantically and structurally consistent area of ​​insight into a topic.

[0009] In this solution, for the insight topic region, the writing agent generates opinion units containing viewpoints, conclusion types, and verifiable evidence chains. The fact-verifying agent then performs consistency checks and corrections on the opinion units based on numerical values, logic, and knowledge graph constraints to obtain an optimized insight topic region. Specifically: The writing agent generates multiple viewpoint units from the insight topic region, wherein each viewpoint unit Defined as: , in, The business entities involved in expressing the viewpoint Indicates the type of assertion relation. This indicates the quantitative results contained in the viewpoint. Indicates contextual conditions; For each viewpoint unit Binding a one-to-one data evidence chain The data evidence chain It consists of multiple evidence nodes, defined as: , Each of these evidence nodes Defined as: , in, Indicates the data source. Indicates the query path. Indicates entity identifier, Indicates the return value. The system indicates the time of evidence recording and prohibits the output of viewpoints that are not supported by the aforementioned data evidence chain. Viewpoint Unit and the data evidence chain it is bound to Input is fed into the fact-verifying agent for consistency verification; Based on the aforementioned chain of data evidence Recalculate to obtain the actual calculation result. Calculate its correlation with the quantification results generated in the viewpoint unit. Deviation: , When the deviation Δ is greater than the tolerance threshold set by the system, it is determined that the viewpoint unit has a numerical conflict. The conclusions expressed by the viewpoint units are verified through logical reasoning based on the data evidence chain, and the verification results are... They are categorized as supportive, consistent, neutral, or conflicting. If the viewpoint unit involves business causal relationships, then verify whether the business causal relationships conform to the constraint paths defined in the domain indicator relationship knowledge graph; If it does not conform, the viewpoint unit is corrected for consistency to obtain an optimized insight topic area.

[0010] In this solution, if the above conditions are not met, the viewpoint unit will be corrected for consistency to obtain an optimized insight topic area, specifically as follows: When the fact-verifying agent detects any consistency conflict, it performs a structured model of the conflict, generating a conflict feature vector. : , in, Indicates the type of conflict. Indicates the intensity of the conflict. Indicates the position of the conflict within the viewpoint unit. Indicates the violation of the validation rule; Based on the conflict feature vector Construct explicit negative feedback constraint information and the negative feedback constraint information Return to the writing agent; The writing agent, based on the received negative feedback constraint information... For viewpoint units Make corrections and generate new viewpoint units. : , The revised viewpoint unit The updated data evidence chain is re-inputted into the fact-verifying agent, and the consistency verification, conflict modeling, negative feedback construction and opinion correction steps are repeatedly executed to form a multi-round iterative collaboration between the writing agent and the fact-verifying agent. During multiple iterations, the consistency convergence score of the viewpoint unit is calculated. : , in, Indicates the numerical consistency score. Indicates the logical consistency score. The knowledge graph constraint consistency score is represented by α, β, and γ, which are weight coefficients. When the consistency convergence score is greater than the preset convergence threshold, it is determined that the current viewpoint unit has reached the factual consistency convergence state, and the viewpoint unit is marked as the optimized viewpoint. When the number of iterations exceeds the preset limit or the consistency convergence score fails to reach the convergence threshold, the system triggers a human-machine collaboration mechanism, whereby a human intervenes to correct the viewpoint unit and its data evidence chain, and the correction result is structured into a knowledge patch and written into the domain indicator relationship knowledge graph or fact verification rule base. An iterative verification and correction process is performed on all viewpoint units within the insight topic area, and the final set of all optimized viewpoint units is defined as the optimized insight topic area.

[0011] In this solution, parsing user input to obtain user intent features, filtering candidate topics semantically related to the user intent features from K optimized insight topic regions, and assigning chapter positions to each candidate topic through heuristic search based on the SRCI four-dimensional scoring model, semantic progression relationships between topics, and logical entropy calculation, automatically arranging and generating a report outline, specifically: The system receives natural language input from the user, and the intent parsing agent performs semantic parsing on the input text, extracting user preference features regarding analysis objectives, focus areas, narrative style, and information density, and encodes these features into a unified intent feature vector. Its formal definition is: , in, This indicates the distribution of user attention weight across different business dimensions. This indicates the narrative style characteristics desired by the user. This represents the core topic vector that the user is explicitly or implicitly interested in. This indicates the strength of a user's preference for the level of detail in information; The intent feature vector is mapped to a high-dimensional semantic embedding space to obtain the intent embedding vector. The central topic representations of the K optimized insight topic regions are embedded into the same high-dimensional space to obtain the semantic embedding vectors of each topic region. Calculate the intent embedding vector semantic embedding vectors of each topic region Semantic similarity between them: , Select the top with the highest semantic similarity Each optimized insight topic area forms a candidate topic set. : , For candidate topic set Each candidate topic in The SRCI four-dimensional scoring model is introduced to calculate the comprehensive value of the topic as a component of the report outline. The SRCI score is defined as follows: , Define any two candidate topics and Semantic progression intensity between This is used to measure the logical coherence between the two in the analysis of the narrative, and its calculation method is as follows: , in, This represents the semantic similarity between two topic vectors generated by the same text embedding model. This represents the directional constraint factor of the topic in the analysis structure, when arrive When the order conforms to the preset positive analysis logic, Take the higher value, and vice versa; at the same time, introduce logical entropy. To measure the degree of disorder in the information organization of the generated outline fragments: , in This indicates the distribution ratio of chapters of different logical types in the outline; A hierarchical heuristic search strategy is employed to assign chapter positions to candidate topics, generating the final report outline. For the l-th chapter position to be filled in the outline, each candidate topic is calculated. Overall score for this position: , in, This indicates the topic that has been selected at the l-1 section position. Indicates the topic The change in the logical entropy of the outline when placed in the current position. and To balance the weighting of information compactness and semantic progression, the system selects the topic with the highest score as the topic of the current chapter. , After selection, the logical state of the remaining candidate topics is dynamically updated until all candidate topics are assigned or the preset outline chapter level depth is reached, and a report outline is generated.

[0012] In this solution, based on the structure of the report outline, the viewpoint units corresponding to the themes of each chapter are organized and integrated to output the final data insight report, specifically as follows: Match the topic of each chapter in the report outline with the optimized insight topic area to obtain a set of fact-verified and optimized viewpoint units generated for each topic area; Based on the causal, relational, and compositional relationships between indicators recorded in the knowledge graph, the logical order between viewpoint units is organized to form a coherent content flow within the chapter; The writing agent generates coherent chapter text based on the chapter theme, the organized content flow, and the assigned viewpoint units. During the text generation process, the data evidence chain bound to the viewpoint units is integrated with the generated text content in the form of structured footnotes, data cards, or interactive traceable links to form a data insight report file.

[0013] A second aspect of the present invention also provides a data insight report generation system based on knowledge enhancement and fact verification. The system includes a memory and a processor. The memory includes a data insight report generation method program based on knowledge enhancement and fact verification. When the processor executes the data insight report generation method program based on knowledge enhancement and fact verification, it implements the steps of the data insight report generation method based on knowledge enhancement and fact verification as described in any of the preceding claims.

[0014] This invention discloses a method and system for generating data insight reports based on knowledge enhancement and fact verification. It constructs a domain-specific indicator relationship knowledge graph by training a large language model; maps insights to the graph, constructs a weighted insight graph, and performs spectral clustering to divide insight topic regions; generates opinion units through a writing agent, and performs consistency verification and correction by a fact verification agent; analyzes user intent, filters candidate topics, and generates a report outline through heuristic search and arrangement based on SRCI scoring, semantic progression, and logical entropy calculation; finally, it integrates verified opinion units according to the outline and outputs the final report. This invention achieves fully automated, highly reliable report generation, improving accuracy, logical consistency, and interpretability. Attached Figure Description

[0015] Figure 1 A flowchart of a data insight report generation method based on knowledge enhancement and fact verification according to the present invention is shown; Figure 2 This invention illustrates a flowchart of the construction and insight clustering of an indicator relationship knowledge graph in the field of this invention. Figure 3 The flowchart of spectral clustering of the semantic-structure weighted insight graph of the present invention is shown; Figure 4 The flowchart illustrating the implementation process of generating the task viewpoint of this invention is shown. Figure 5 This diagram illustrates the implementation flowchart for the automatic generation of the report outline of this invention. Figure 6A block diagram of a data insight report generation system based on knowledge enhancement and fact verification according to the present invention is shown. Detailed Implementation

[0016] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0017] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0018] Figure 1 The flowchart of a data insight report generation method based on knowledge enhancement and fact verification according to the present invention is shown.

[0019] like Figure 1 As shown, the first aspect of the present invention provides a method for generating data insight reports based on knowledge enhancement and fact verification, comprising: Based on domain corpus, a large language model is trained for domain knowledge and entity relationship extraction. The trained model is used to automatically extract indicator entities and their relationships from domain text, and a domain indicator relationship knowledge graph is constructed with indicator entities as nodes and indicator relationships as edges. The main indicator of each insight in the set of insights to be analyzed is mapped to the domain indicator relationship knowledge graph to construct a structure-semantic weighted insight graph, and the structure-semantic weighted insight graph is divided into K insight topic regions based on the spectral clustering method; For the aforementioned insight topic region, the writing agent generates opinion units containing viewpoints, conclusion types, and verifiable evidence chains. The fact verification agent performs consistency verification and correction on the opinion units in terms of numerical values, logic, and knowledge graph constraints to obtain an optimized insight topic region. The system parses user input to obtain user intent features, filters candidate topics that are semantically related to the user intent features from K optimized insight topic regions, and assigns chapter positions to each candidate topic through heuristic search based on the SRCI four-dimensional scoring model, semantic progression relationship between topics and logical entropy calculation, and automatically arranges and generates a report outline. Based on the structure of the report outline, the viewpoints corresponding to the themes of each chapter are organized and integrated to output the final data insight report.

[0020] It should be noted that by fine-tuning the large language model in two stages and enabling it to automatically extract indicator relationships, this method can construct and dynamically update a domain indicator relationship knowledge graph, providing the system with an evolvable and structured domain knowledge foundation, overcoming the drawbacks of the lagging updates of static knowledge graphs. Next, the set of insights to be analyzed is used to construct a structure-semantic weighted insight graph based on this dynamic knowledge graph, and spectral clustering is performed. This comprehensively considers the semantic features between insights and the structural relationships between indicators, dividing scattered insights into semantically consistent and structurally closely related thematic regions, laying the foundation for the subsequent generation of logically coherent viewpoints. Then, for each thematic region, the writing agent and the fact-verifying agent collaborate to execute an iterative process of "generation-verification-correction." This method, by forcibly binding verifiable evidence chains and performing multi-dimensional consistency verification of numerical, logical, and knowledge graph constraints, effectively eliminates or significantly reduces the risks of numerical errors, logical contradictions, and factual illusions common in traditional large language model applications, ensuring the accuracy and authority of the generated viewpoints. Subsequently, by analyzing user intent and introducing the SRCI four-dimensional scoring model, semantic progression relationships between topics, and logical entropy calculation, this method enables heuristic search and dynamic scoring iteration. This generates a report outline that aligns with user preferences and possesses a clear logical framework, upgrading the report structure from "template filling" to "adaptive arrangement." Finally, by integrating validated and optimized viewpoint units based on this outline, a data insight report with a clear structure, accurate content, prominent key points, and rigorous logic can be output. The entire method achieves a reliable, controllable, and high-quality report generation process without relying on preset templates or external online searches.

[0021] Figure 2 The flowchart illustrates the construction and insight clustering of the knowledge graph of indicator relationships in this invention field.

[0022] According to an embodiment of the present invention, the step of training a large language model based on domain corpus for domain knowledge and entity relation extraction, and automatically extracting indicator entities and their relations from domain text using the trained model, and constructing a domain indicator relation knowledge graph with indicator entities as nodes and indicator relations as edges, specifically involves: A domain corpus is constructed by acquiring production management documents, indicator business definitions, and internal analysis reports for the target domain. The domain corpus is then cleaned of duplicate data and structured to form a domain training corpus. Based on the domain training corpus, a large language model is trained to understand the professional concepts and indicator semantics of the target domain, resulting in a domain-specific large language model with domain knowledge understanding capabilities. Based on the imperative supervised fine-tuning method, the domain corpus is used to construct training data with domain text and entity relation extraction instructions as input and corresponding structured entity relations as supervision signals. The domain large language model is trained on the entity relation extraction task to obtain a trained model with the ability to automatically extract indicator entities and their relations. It should be noted that the target field includes the aquaculture sector. The production management documents include various operational records, management standards, process data, and related procedural documents related to the aquaculture production process. The indicator business definition descriptions include the specific definitions, calculation formulas, statistical ranges, and business interpretations of various financial, production, and efficiency indicators. Internal analysis reports include various historical analysis summaries, research reports, and decision support documents generated internally based on business data. The entity relationship extraction task training enables the model to learn the mapping relationship from natural language descriptions to structured outputs, completing entity relationship extraction training; the indicator entities include cost indicators, efficiency indicators, output indicators, and risk indicators, and the relationships include influence relationships, compositional relationships, causal relationships, and dependency relationships.

[0023] The target domain text is input into the trained model, which performs semantic parsing on the text, identifies and outputs the indicator entities in the text and the relationships between the indicator entities; The index relationship results output by the model are uniformly converted into a structured representation in the form of triples, and attribute information is added to nodes and edges; Using the transformed indicator entities as graph nodes and the relationships between indicator entities as edges in the graph, a domain indicator relationship knowledge graph is constructed.

[0024] It should be noted that the triple includes a source indicator entity, a relation type, and a target indicator entity, i.e., <indicator entity A, relation type, indicator entity B>. The attribute information includes the indicator category and the relation confidence score. Constructing a domain indicator relation knowledge graph specifically includes: deduplicating and normalizing the extracted indicator entities, mapping the normalized indicator entities to graph nodes, mapping the relationships between indicator entities to edges in the graph, and writing the attribute information for the nodes and edges to form a complete graph structure. After model training, the model outputs structured indicator relation results based on the entity boundary recognition and relation determination capabilities learned during training. Each output result includes at least: a source indicator entity, a target indicator entity, entity type information, and the relation type between the indicators. The additional attribute information includes the indicator category, indicator definition, and relation confidence score. Figure 3 A flowchart of spectral clustering of the semantic-structure weighted insight graph of the present invention is shown.

[0025] According to an embodiment of the present invention, the step of mapping the main indicator of each insight in the set of insights to be analyzed to the domain indicator relationship knowledge graph, constructing a structure-semantic weighted insight graph, and dividing the structure-semantic weighted insight graph into K insight topic regions based on the spectral clustering method, specifically involves: The set of insights to be analyzed is parsed to obtain the main indicators corresponding to each insight object and its associated insight interpretation text. Each insight consists of insight interpretation text, a list of design indicators, insight name, and insight calculation data dictionary. The main indicators of each insight are used as query keys and mapped to the domain indicator relationship knowledge graph. The nodes corresponding to each indicator in the graph are identified, and the discrete insight objects are connected to construct an original insight graph that conforms to the indicator relationship structure based on the relationship between indicators recorded in the graph. Formalize the original insight map as Where V represents the set of nodes and E represents the set of edges, each node In this context, indicator represents the indicator name, and explain represents the corresponding insight interpretation text. Define edge relationship types and assign preset structural weights to different types of edge relationships, including , Calculate the semantic similarity between the insight interpretation texts corresponding to the two nodes connected by each edge, and use this as the semantic weight of the edge. The formula is as follows: , in, This represents a function for calculating semantic similarity. It should be noted that in defining variable relation types, parent(·) represents its parent node in the domain indicator relation knowledge graph: , The final weight of each edge is defined as the product of the structural weight and the semantic weight corresponding to that edge, thus constructing a structure-semantic weighted insight graph, which is formally represented as: , Construct the weighted adjacency matrix of the structure-semantic weighted insight graph: , Calculate the degree matrix of the nodes based on the weighted adjacency matrix: , Then, the normalized graph Laplacian matrix is ​​constructed: , in, Given the identity matrix, perform eigenvalue decomposition on the Laplacian matrix L: ,in, For the k-th eigenvalue, sorted in ascending order. For each eigenvector, select the eigenvectors corresponding to the K smallest eigenvalues ​​to form the spectral embedding matrix of the node in the low-dimensional space: , Embed the spectrum in each row of matrix U Consider as a node The low-dimensional representation of the data is obtained, and the k-means clustering algorithm is used to partition the representation space: , Ultimately, each cluster This corresponds to a semantically and structurally consistent area of ​​insight into a topic.

[0026] It's important to clarify that an insight refers to the analysis of data within a data space, focusing on a core indicator and its related indicators, including their numerical changes, trend changes, and rankings. An insight comprises fields such as the insight interpretation text, a list of designed indicators, the insight name, and an insight calculation data dictionary. Therefore, each insight involves a primary indicator. This primary indicator is used as a key to map to the constructed indicator relationship knowledge graph, transforming discrete insight objects into an insight graph that conforms to the indicator relationship structure. The primary indicator refers to the core financial, production, efficiency, cost, or risk indicators associated with each insight object. Based on spectral clustering, the structure-semantic weighted insight graph is divided into K insight theme regions. Its core purpose is to automatically and structurally summarize and organize massive amounts of fragmented initial insights. This aims to overcome the limitations of traditional clustering techniques that only consider semantic similarity or rely solely on simple graph structure relationships. By constructing a weighted insight graph that simultaneously encodes the semantic relationships between insights and the graph structure relationships between indicators, it can more accurately capture the inherent connections between insights in terms of analytical logic and business meaning. Based on this weighted graph, spectral clustering algorithms can project nodes into the feature space, thereby identifying node clusters with high semantic consistency and graph structural correlation. Ultimately, each segmented insight theme region represents a set of insights that are closely related in business logic and mutually supportive in analytical perspectives. This provides high-quality, structured input material for subsequently generating logically coherent and thematically clear chapter viewpoints, and is a key technical step in achieving the hierarchical abstraction and summarization from fine-grained data insights to high-level thematic viewpoints.

[0027] Figure 4 A flowchart illustrating the implementation process of generating the task viewpoint of this invention is shown.

[0028] According to an embodiment of the present invention, for the insight topic region, the writing agent generates opinion units containing viewpoints, conclusion types, and verifiable evidence chains, and the fact-verifying agent performs consistency checks and corrections on the opinion units in terms of numerical values, logic, and knowledge graph constraints to obtain an optimized insight topic region, specifically as follows: The writing agent generates multiple viewpoint units from the insight topic region, wherein each viewpoint unit Defined as: , in, This indicates the business entity involved in the viewpoint, such as "cost per piglet". Indicates the type of assertion relationship, such as "year-on-year increase" or "significantly higher than"; This indicates the quantitative result included in the viewpoint, such as 3.3%; Indicates contextual conditions, including time interval, statistical scope, or business scope; It should be noted that the writing agent is a functional module built upon a domain-specific large language model. After dividing the insight topic region into K regions, the writing agent generates opinion units containing viewpoints, conclusion types, and verifiable evidence chains. Its core task is to receive the insight topic region as input and automatically generate structured analytical opinion units, which can be integrated with the large language model to function as an agent.

[0029] For each viewpoint unit Binding a one-to-one data evidence chain The data evidence chain It consists of multiple evidence nodes, defined as: , Each of these evidence nodes Defined as: , in, Indicates the data source. Indicates the query path. Indicates entity identifier, Indicates the return value. The system indicates the time of evidence recording and prohibits the output of viewpoint units that are not supported by the data evidence chain. For example, an evidence node can correspond to "the actual value recorded by a company in the cost table of a single piglet in December 2025". Viewpoint Unit and the data evidence chain it is bound to Input is fed into the fact-verifying agent for consistency verification; Based on the aforementioned chain of data evidence Recalculate to obtain the actual calculation result. Calculate its correlation with the quantification results generated in the viewpoint unit. Deviation: , When the deviation Δ is greater than the tolerance threshold set by the system, it is determined that the viewpoint unit has a numerical conflict. The conclusions expressed by the viewpoint units are verified through logical reasoning based on the data evidence chain, and the verification results are... They are categorized as supportive, consistent, neutral, or conflicting. If the viewpoint unit involves business causal relationships, then verify whether the business causal relationships conform to the constraint paths defined in the domain indicator relationship knowledge graph. For example, if If it means "caused", then the system needs to verify. Does the relevant indicator have a legitimate causal edge in the knowledge graph? If it does not conform, the viewpoint unit is corrected for consistency to obtain an optimized insight topic area.

[0030] It should be noted that after receiving the opinion unit and its chain of evidence, the fact-verifying agent performs consistency verification. This verification is not a single rule judgment, but is completed jointly by multiple interrelated verification modules. At the numerical consistency level, the verification agent re-executes the calculation based on the chain of evidence to obtain the factual calculation result. and the generated values ​​given in the viewpoint In comparison, among which, This indicates the quantitative results generated by the intelligent agent. This indicates the result obtained by recalculating based on the chain of evidence. For example, The figure was "costs increased by 3.3% year-on-year," while It could be "a year-on-year increase of 4.1% calculated based on raw data". When there is a deviation... When the percentage exceeds the system's set tolerance threshold (e.g., 0.01% or 1%), the system determines that the viewpoint is inconsistent at the numerical level. By forcibly introducing a traceable chain of data evidence during the viewpoint generation process, each viewpoint unit establishes an explicit binding relationship with its underlying data facts. The fact-verifying agent then uses this evidence chain to conduct multi-dimensional and quantifiable rigorous verification of the consistency of the generated numerical values, logic, and knowledge graph. Once a conflict is detected, the system models the conflict as a structured negative feedback, driving the writing agent to perform targeted iterative corrections until the viewpoint meets the preset factual consistency convergence criteria. This process not only ensures the accuracy and authority of individual viewpoint units, but its built-in human-machine collaborative knowledge patching mechanism also continuously enhances the system's factual constraint capabilities, thereby ultimately ensuring that all viewpoints within the optimized insight topic area achieve high-quality output at both the logical and factual levels, providing a solid guarantee for generating credible and verifiable report content.

[0031] According to an embodiment of the present invention, if the condition is not met, consistency correction is performed on the viewpoint unit to obtain an optimized insight topic area, specifically as follows: It should be noted that when the fact-verifying agent detects a consistency conflict, it does not immediately terminate the process, but instead performs a structured model of the conflict. Each conflict is represented as a conflict feature vector.

[0032] When the fact-verifying agent detects any consistency conflict, it performs a structured model of the conflict, generating a conflict feature vector. : , in, Indicates the type of conflict (numerical conflict, logical conflict, or causal conflict). Indicates the intensity of the conflict. Indicates the position of the conflict within the viewpoint unit. Indicates the violation of the validation rule; Based on the conflict feature vector Construct explicit negative feedback constraint information and the negative feedback constraint information Return to the writing agent; The writing agent, based on the received negative feedback constraint information... For viewpoint units Make corrections and generate new viewpoint units. : , The revised viewpoint unit The updated data evidence chain is re-inputted into the fact-verifying agent, and the consistency verification, conflict modeling, negative feedback construction and opinion correction steps are repeatedly executed to form a multi-round iterative collaboration between the writing agent and the fact-verifying agent. During multiple iterations, the consistency convergence score of the viewpoint unit is calculated. : , in, Indicates the numerical consistency score. Indicates the logical consistency score. The knowledge graph constraint consistency score is represented by α, β, and γ, which are weight coefficients. When the consistency convergence score is greater than the preset convergence threshold, it is determined that the current viewpoint unit has reached the factual consistency convergence state, and the viewpoint unit is marked as the optimized viewpoint. When the number of iterations exceeds the preset limit or the consistency convergence score fails to reach the convergence threshold, the system triggers a human-machine collaboration mechanism, whereby a human intervenes to correct the viewpoint unit and its data evidence chain, and the correction result is structured into a knowledge patch and written into the domain indicator relationship knowledge graph or fact verification rule base. An iterative verification and correction process is performed on all viewpoint units within the insight topic area, and the final set of all optimized viewpoint units is defined as the optimized insight topic area.

[0033] It should be noted that when an opinion fails to meet the convergence condition within a limited number of iterations, or when it involves high-impact business decisions, the system triggers a human-machine collaboration mechanism. Human experts intervene in the opinion and evidence chain, and their corrections are structured as knowledge patches. The results are then written into a knowledge graph or a verification rule base. Through this mechanism, the system can correct individual generation results and continuously improve its factual constraint capabilities. Using the above method, this invention constructs a generation architecture driven collaboratively by a writing agent and a fact-verifying agent, based on a factual consistency perspective, enabling the generation process to achieve verifiable, correctable, and sustainably evolving technical effects.

[0034] According to an embodiment of the present invention, the step of parsing user input to obtain user intent features, filtering candidate topics semantically related to the user intent features from K optimized insight topic regions, and assigning chapter positions to each candidate topic through heuristic search based on the SRCI four-dimensional scoring model, the semantic progression relationship between topics, and logical entropy calculation, and automatically arranging and generating a report outline, specifically: The system receives natural language input from the user, and the intent parsing agent performs semantic parsing on the input text, extracting user preference features regarding analysis objectives, focus areas, narrative style, and information density, and encodes these features into a unified intent feature vector. Its formal definition is: , in, This indicates the weight distribution of user attention across different business dimensions, such as cost, risk, and efficiency. This indicates the narrative style characteristics that users expect, such as overview-oriented, problem-oriented, or attribution-oriented. This represents the core topic vector that the user is explicitly or implicitly interested in. This indicates the strength of the user's preference for the level of detail in the information. This intention feature vector serves as a global constraint input for subsequent outline arrangement and is used throughout the entire topic selection and sorting process. It should be noted that the system pre-maintains a dynamically updated pool of candidate topics. , is represented as: , Each theme Each is associated with a semantic embedding vector generated by the Qwen3-embedding model. This is used to characterize the high-dimensional business semantic features of the topic. The outline orchestration agent calls the Qwen3-embedding model to extract the intent feature vector. Mapping to the same high-dimensional embedding space yields the intention embedding vector. .

[0035] The intent feature vector is mapped to a high-dimensional semantic embedding space to obtain the intent embedding vector. The central topic representations of the K optimized insight topic regions are embedded into the same high-dimensional space to obtain the semantic embedding vectors of each topic region. Calculate the intent embedding vector semantic embedding vectors of each topic region Semantic similarity between them: , Select the top with the highest semantic similarity Each optimized insight topic area forms a candidate topic set. : , For candidate topic set Each candidate topic in The SRCI four-dimensional scoring model is introduced to calculate the comprehensive value of the topic as a component of the report outline. The SRCI score is defined as follows: , It should be noted that, among them, Indicate topic With user narrative style preferences The matching score is directly calculated by the intent parsing agent; Indicate topic User core focus vector The degree of support is calculated through semantic similarity: The formula for calculating the case or factual support capacity of a topic is as follows: ,in This indicates the number of metrics associated with this topic. This indicates the number of cases corresponding to this topic; The score represents the information contribution of the topic, and the calculation formula is as follows: ,in Indicate topic Does it cover the indicators? , Indicators Information weight; , , , These are the weighting coefficients for each score item; Define any two candidate topics and Semantic progression intensity between This is used to measure the logical coherence between the two in the analysis of the narrative, and its calculation method is as follows: , in, This represents the semantic similarity between two topic vectors generated by the same text embedding model. This represents the directional constraint factor of the topic in the analysis structure, when arrive When the order conforms to the preset positive analysis logic, Take the higher value, and vice versa; at the same time, introduce logical entropy. To measure the degree of disorder in the information organization of the generated outline fragments: , in This indicates the distribution ratio of chapters of different logical types in the outline; A hierarchical heuristic search strategy is employed to assign chapter positions to candidate topics, generating the final report outline. For the l-th chapter position to be filled in the outline, each candidate topic is calculated. Overall score for this position: , in, This indicates the topic that has been selected at the l-1 section position. Indicates the topic The change in the logical entropy of the outline when placed in the current position. and To balance the weighting of information compactness and semantic progression, the system selects the topic with the highest score as the topic of the current chapter. , After selection, the logical state of the remaining candidate topics is dynamically updated until all candidate topics are assigned or the preset outline chapter level depth is reached, and a report outline is generated.

[0036] It should be noted that this invention proposes a heuristic automatic outline arrangement method based on intent parsing and a four-dimensional SRCI scoring model. This method is used to automatically generate a logically coherent and focused report outline structure based on the user's unstructured input preferences without any preset template constraints. Through the collaborative work of an intent parsing agent and an outline arrangement agent, this method integrates user intent, topic semantics, and chapter logical relationships into a single computational framework, achieving adaptive generation of the outline structure.

[0037] In this method, the system includes at least two functional modules: an intent parsing agent and an outline arrangement agent. The intent parsing agent is responsible for mapping user input into a computable intent representation, while the outline arrangement agent, based on this intent representation, scores, ranks, and combines candidate topics to generate a report outline step by step.

[0038] According to an embodiment of the present invention, the step of organizing and integrating the viewpoint units corresponding to the themes of each chapter based on the structure of the report outline, and outputting the final data insight report, specifically includes: Match the topic of each chapter in the report outline with the optimized insight topic area to obtain a set of fact-verified and optimized viewpoint units generated for each topic area; Based on the causal, relational, and compositional relationships between indicators recorded in the knowledge graph, the logical order between viewpoint units is organized to form a coherent content flow within the chapter; The writing agent generates coherent chapter text based on the chapter theme, the organized content flow, and the assigned viewpoint units. During the text generation process, the data evidence chain bound to the viewpoint units is integrated with the generated text content in the form of structured footnotes, data cards, or interactive traceable links to form a data insight report file.

[0039] Figure 6 A block diagram of a data insight report generation system based on knowledge enhancement and fact verification according to the present invention is shown.

[0040] A second aspect of the present invention also provides a data insight report generation system based on knowledge enhancement and fact verification. The system includes: a memory 601, a processor 602, and a communication interface 603. The memory includes a data insight report generation method program based on knowledge enhancement and fact verification. The communication interface is used for data connection and communication between the memory and the processor. When the data insight report generation method program based on knowledge enhancement and fact verification is executed by the processor, it implements the steps of the data insight report generation method based on knowledge enhancement and fact verification as described in any of the above claims.

[0041] This invention discloses a method and system for generating data insight reports based on knowledge enhancement and fact verification. It constructs a domain-specific indicator relationship knowledge graph by training a large language model; maps insights to the graph, constructs a weighted insight graph, and performs spectral clustering to divide insight topic regions; generates opinion units through a writing agent, and performs consistency verification and correction by a fact verification agent; analyzes user intent, filters candidate topics, and generates a report outline through heuristic search and arrangement based on SRCI scoring, semantic progression, and logical entropy calculation; finally, it integrates verified opinion units according to the outline and outputs the final report. This invention achieves fully automated, highly reliable report generation, improving accuracy, logical consistency, and interpretability.

[0042] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0043] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, 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 methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0044] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention 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 the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for generating data insight reports based on knowledge enhancement and fact verification, characterized in that, Includes the following steps: Based on domain corpus, a large language model is trained for domain knowledge and entity relationship extraction. The trained model is used to automatically extract indicator entities and their relationships from domain text, and a domain indicator relationship knowledge graph is constructed with indicator entities as nodes and indicator relationships as edges. The main indicator of each insight in the set of insights to be analyzed is mapped to the domain indicator relationship knowledge graph to construct a structure-semantic weighted insight graph, and the structure-semantic weighted insight graph is divided into K insight topic regions based on the spectral clustering method; For the aforementioned insight topic region, the writing agent generates opinion units containing viewpoints, conclusion types, and verifiable evidence chains. The fact verification agent performs consistency verification and correction on the opinion units in terms of numerical values, logic, and knowledge graph constraints to obtain an optimized insight topic region. The system parses user input to obtain user intent features, filters candidate topics that are semantically related to the user intent features from K optimized insight topic regions, and assigns chapter positions to each candidate topic through heuristic search based on the SRCI four-dimensional scoring model, semantic progression relationship between topics and logical entropy calculation, and automatically arranges and generates a report outline. Based on the structure of the report outline, the viewpoints corresponding to the themes of each chapter are organized and integrated to output the final data insight report.

2. The data insight report generation method based on knowledge enhancement and fact verification according to claim 1, characterized in that, The method involves training a large language model based on domain corpus for domain knowledge and entity relation extraction. The trained model automatically extracts indicator entities and their relationships from domain text, and constructs a domain indicator relation knowledge graph using indicator entities as nodes and indicator relations as edges. Specifically: A domain corpus is constructed by acquiring production management documents, indicator business definitions, and internal analysis reports for the target domain. The domain corpus is then cleaned of duplicate data and structured to form a domain training corpus. Based on the domain training corpus, a large language model is trained to understand the professional concepts and indicator semantics of the target domain, resulting in a domain-specific large language model with domain knowledge understanding capabilities. Based on the imperative supervised fine-tuning method, the domain corpus is used to construct training data with domain text and entity relation extraction instructions as input and corresponding structured entity relations as supervision signals. The domain large language model is trained on the entity relation extraction task to obtain a trained model with the ability to automatically extract indicator entities and their relations. The target domain text is input into the trained model, which performs semantic parsing on the text, identifies and outputs the indicator entities in the text and the relationships between the indicator entities; The index relationship results output by the model are uniformly converted into a structured representation in the form of triples, and attribute information is added to nodes and edges; Using the transformed indicator entities as graph nodes and the relationships between indicator entities as edges in the graph, a domain indicator relationship knowledge graph is constructed.

3. The data insight report generation method based on knowledge enhancement and fact verification according to claim 1, characterized in that, The process involves mapping the main indicator of each insight in the set of insights to be analyzed to the domain indicator relationship knowledge graph, constructing a structure-semantic weighted insight graph, and dividing the structure-semantic weighted insight graph into K insight topic regions based on the spectral clustering method, specifically: The set of insights to be analyzed is parsed to obtain the main indicators corresponding to each insight object and its associated insight interpretation text. Each insight consists of insight interpretation text, a list of design indicators, insight name, and insight calculation data dictionary. The main indicators of each insight are used as query keys and mapped to the domain indicator relationship knowledge graph. The nodes corresponding to each indicator in the graph are identified, and the discrete insight objects are connected to construct an original insight graph that conforms to the indicator relationship structure based on the relationship between indicators recorded in the graph. Formalize the original insight map as Where V represents the set of nodes and E represents the set of edges, each node In this context, indicator represents the indicator name, and explain represents the corresponding insight interpretation text. Define edge relationship types and assign preset structural weights to different types of edge relationships, including , Calculate the semantic similarity between the insight interpretation texts corresponding to the two nodes connected by each edge, and use this as the semantic weight of the edge. The formula is as follows: ， in, This represents a function for calculating semantic similarity. The final weight of each edge is defined as the product of the structural weight and the semantic weight corresponding to that edge, thus constructing a structure-semantic weighted insight graph, which is formally represented as: ， Construct the weighted adjacency matrix of the structure-semantic weighted insight graph: ， Calculate the degree matrix of the nodes based on the weighted adjacency matrix: ， Then, the normalized graph Laplacian matrix is ​​constructed: ， in, Given the identity matrix, perform eigenvalue decomposition on the Laplacian matrix L: ,in, For the k-th eigenvalue, sorted in ascending order. For each eigenvector, select the eigenvectors corresponding to the K smallest eigenvalues ​​to form the spectral embedding matrix of the node in the low-dimensional space: ， Embed the spectrum in each row of matrix U Consider as a node The low-dimensional representation of the data is obtained, and the k-means clustering algorithm is used to partition the representation space: ， Ultimately, each cluster This corresponds to a semantically and structurally consistent area of ​​insight into a topic.

4. The data insight report generation method based on knowledge enhancement and fact verification according to claim 1, characterized in that, For the aforementioned insight topic region, the writing agent generates opinion units containing viewpoints, conclusion types, and verifiable evidence chains. The fact-verifying agent then performs consistency checks and corrections on these opinion units based on numerical values, logic, and knowledge graph constraints to obtain an optimized insight topic region. Specifically: The writing agent generates multiple viewpoint units from the insight topic region, wherein each viewpoint unit Defined as: ， in, The business entities involved in expressing the viewpoint Indicates the type of assertion relation. This indicates the quantitative results contained in the viewpoint. Indicates contextual conditions; For each viewpoint unit Binding a one-to-one data evidence chain The data evidence chain It consists of multiple evidence nodes, defined as: ， Each of these evidence nodes Defined as: ， in, Indicates the data source. Indicates the query path. Indicates entity identifier, Indicates the return value. The system indicates the time of evidence recording and prohibits the output of viewpoints that are not supported by the aforementioned data evidence chain. Viewpoint Unit and the data evidence chain it is bound to Input is fed into the fact-verifying agent for consistency verification; Based on the aforementioned chain of data evidence Recalculate to obtain the actual calculation result. Calculate its correlation with the quantification results generated in the viewpoint unit. Deviation: ， When the deviation Δ is greater than the tolerance threshold set by the system, it is determined that the viewpoint unit has a numerical conflict. The conclusions expressed by the viewpoint units are based on the data evidence chain. Perform logical reasoning verification, and the verification results. They are categorized as supportive, consistent, neutral, or conflicting. If the viewpoint unit involves business causal relationships, then verify whether the business causal relationships conform to the constraint paths defined in the domain indicator relationship knowledge graph; If it does not conform, the viewpoint unit is corrected for consistency to obtain an optimized insight topic area.

5. The data insight report generation method based on knowledge enhancement and fact verification according to claim 4, characterized in that, If the above conditions are not met, the viewpoint unit will be corrected for consistency to obtain an optimized insight topic area, specifically: When the fact-verifying agent detects any consistency conflict, it performs a structured model of the conflict, generating a conflict feature vector. : ， in, Indicates the type of conflict. Indicates the intensity of the conflict. Indicates the position of the conflict within the viewpoint unit. Indicates the violation of the validation rule; Based on the conflict feature vector Construct explicit negative feedback constraint information and the negative feedback constraint information Return to the writing agent; The writing agent, based on the received negative feedback constraint information... For viewpoint units Make corrections and generate new viewpoint units. : ， The revised viewpoint unit The updated data evidence chain is re-inputted into the fact-verifying agent, and the consistency verification, conflict modeling, negative feedback construction and opinion correction steps are repeatedly executed to form a multi-round iterative collaboration between the writing agent and the fact-verifying agent. During multiple iterations, the consistency convergence score of the viewpoint unit is calculated. : ， in, Indicates the numerical consistency score. Indicates the logical consistency score. The knowledge graph constraint consistency score is represented by α, β, and γ, which are weight coefficients. When the consistency convergence score is greater than the preset convergence threshold, it is determined that the current viewpoint unit has reached the factual consistency convergence state, and the viewpoint unit is marked as the optimized viewpoint. When the number of iterations exceeds the preset limit or the consistency convergence score fails to reach the convergence threshold, the system triggers a human-machine collaboration mechanism, whereby a human intervenes to correct the viewpoint unit and its data evidence chain, and the correction result is structured into a knowledge patch and written into the domain indicator relationship knowledge graph or fact verification rule base. An iterative verification and correction process is performed on all viewpoint units within the insight topic area, and the final set of all optimized viewpoint units is defined as the optimized insight topic area.

6. The data insight report generation method based on knowledge enhancement and fact verification according to claim 1, characterized in that, The process involves parsing user input to obtain user intent features, filtering candidate topics semantically related to these features from K optimized insight topic regions, and assigning chapter positions to each candidate topic using a heuristic search based on the SRCI four-dimensional scoring model, semantic progression relationships between topics, and logical entropy calculation. This process automatically generates a report outline, specifically: The system receives natural language input from the user, and the intent parsing agent performs semantic parsing on the input text, extracting user preference features regarding analysis objectives, focus areas, narrative style, and information density, and encodes these features into a unified intent feature vector. Its formal definition is: ， in, This indicates the distribution of user attention weight across different business dimensions. This indicates the narrative style characteristics desired by the user. This represents the core topic vector that the user is explicitly or implicitly interested in. This indicates the strength of a user's preference for the level of detail in information; The intent feature vector is mapped to a high-dimensional semantic embedding space to obtain the intent embedding vector. The central topic representations of the K optimized insight topic regions are embedded into the same high-dimensional space to obtain the semantic embedding vectors of each topic region. Calculate the intent embedding vector semantic embedding vectors of each topic region Semantic similarity between them: ， Select the top with the highest semantic similarity Each optimized insight topic area forms a candidate topic set. : ， For candidate topic set Each candidate topic in The SRCI four-dimensional scoring model is introduced to calculate the comprehensive value of the topic as a component of the report outline. The SRCI score is defined as follows: ， in, Indicates the topic With user narrative style preferences Match score; Indicates the topic User core focus vector The degree of support; The capacity of the case studies or facts to support the topic; A score indicating the information contribution of the topic; , , , These are the weighting coefficients for each score item; Define any two candidate topics and The semantic progression intensity G( , The ), used to measure the logical coherence between the two in the analysis of the narrative, is calculated as follows: ， in, This represents the semantic similarity between two topic vectors generated by the same text embedding model. This represents the directional constraint factor of the topic in the analysis structure, when arrive When the order conforms to the preset positive analysis logic, Take the higher value, and vice versa; at the same time, introduce logical entropy. To measure the degree of disorder in the information organization of the generated outline fragments: ， in This indicates the distribution ratio of chapters of different logical types in the outline; A hierarchical heuristic search strategy is employed to assign chapter positions to candidate topics, generating the final report outline. For the l-th chapter position to be filled in the outline, each candidate topic is calculated. Overall score for this position: ， in, This indicates the topic that has been selected at the l-1 chapter position. Indicates the topic The change in the logical entropy of the outline when placed in the current position. and To balance the weighting of information compactness and semantic progression, the system selects the topic with the highest score as the topic of the current chapter. ， After selection, the logical state of the remaining candidate topics is dynamically updated until all candidate topics are assigned or the preset outline chapter level depth is reached, and a report outline is generated.

7. The data insight report generation method based on knowledge enhancement and fact verification according to claim 1, characterized in that, Based on the structure of the report outline, the viewpoints corresponding to the themes of each chapter are organized and integrated to output the final data insight report, specifically as follows: Match the topic of each chapter in the report outline with the optimized insight topic area to obtain a set of fact-verified and optimized viewpoint units generated for each topic area; Based on the causal, relational, and compositional relationships between indicators recorded in the knowledge graph, the logical order between viewpoint units is organized to form a coherent content flow within the chapter; The writing agent generates coherent chapter text based on the chapter theme, the organized content flow, and the assigned viewpoint units. During the text generation process, the data evidence chain bound to the viewpoint units is integrated with the generated text content in the form of structured footnotes, data cards, or interactive traceable links to form a data insight report file.

8. A data insight report generation system based on knowledge enhancement and fact verification, characterized in that, The data insight report generation system based on knowledge enhancement and fact verification includes a storage unit and a processor. The storage unit includes a data insight report generation method program based on knowledge enhancement and fact verification. When the data insight report generation method program based on knowledge enhancement and fact verification is executed by the processor, it implements the steps of the data insight report generation method based on knowledge enhancement and fact verification as described in any one of claims 1 to 7.

Images