Big model question and answer method and electronic equipment

Abstract

The present application relates to the technical field of artificial intelligence, in particular to an intelligence large model question and answer method and an electronic device. The method comprises: obtaining a user query, obtaining a structured intent description object about the query through an intelligence large model and an intelligence field knowledge graph; inputting the structured intent description object into the intelligence field knowledge graph for context reasoning to generate a visual generation scheme specification; retrieving search data about the structured intent description object from multiple sources; generating a view driving data package corresponding to each view, injecting linkage configuration information into data packages having linkage relationships according to the linkage relationships between views defined in the visual generation scheme specification; and generating corresponding views based on the view driving data package, and establishing an event response mechanism between the views. The present application realizes accurate problem analysis and intelligent classification based on intelligence scenario structured intent, and multi-dimensional and linkage visual automatic generation and presentation enhanced by the intelligence field knowledge graph.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-related technologies, and in particular to an intelligence large-scale model question-and-answer method, electronic device, storage medium, and computer program product. Background Technology

[0002] With the rapid development of Large Language Models (LLM) and natural language processing technologies, their application potential in various fields is becoming increasingly apparent. However, general-purpose large models typically have the following significant shortcomings when directly addressing complex, specialized domain tasks: (1) The disconnect between intent understanding and visualization response makes it difficult to meet the needs of multidimensional analysis scenarios. Existing intelligence big data model question answering systems usually directly input users' natural language queries into the model, and the output is mostly in the form of long text answers, lacking in-depth recognition and structured decomposition of users' intelligence analysis intent. Especially in professional intelligence analysis scenarios, users often need to simultaneously understand the situation, track events, explore correlations or predict trends, and relying solely on text output cannot intuitively present multidimensional information and complex relationships. Most current systems adopt fixed or single visualization output modes, which cannot dynamically generate suitable visualization combinations according to user intent, resulting in an unintuitive analysis process and inconvenient result exploration.

[0003] (2) The separation of knowledge graphs from visualization generation processes limits the intuitive expression of the depth of intelligence analysis. In existing technical solutions, large intelligence models or traditional intelligence analysis engines are responsible for generating analytical conclusions or structured data, and the results are then handed over to independent backend modules for visualization rendering. Knowledge graphs are usually only used as a backend knowledge source, and their rich structured semantic relationship network is not deeply integrated into the generation logic and linkage definition of visualization content. This separation results in visualization views being merely passive displays of analytical results, rather than an organic extension and intuitive representation of the analytical reasoning process. There is a lack of intelligent linkage capabilities between views based on underlying semantic relationships, making it difficult for users to explore the deep connections between data through interaction. Summary of the Invention

[0004] Therefore, it is necessary to address the technical problem that existing technologies cannot generate visualized information based on large models and knowledge graphs by providing an intelligence large model question answering method, electronic device, storage medium, and computer program product.

[0005] This invention provides a question-answering method for a large intelligence model, comprising: The system obtains a structured intent description object for the query by acquiring the user's query input via natural language and using the intelligence big data model and intelligence domain knowledge graph. The structured intent description object is input into the knowledge graph of the intelligence domain for contextual reasoning to determine the core entity relationship network involved in the structured intent description object. Based on the structured intent description object, a visualization generation scheme specification is generated. The visualization generation scheme specification includes at least the data patterns of multiple views and the linkage relationship between views. By combining the core entity relationship network, retrieval data about the structured intent description object is obtained from multiple sources, resulting in a basic fact dataset; According to the data pattern defined for each view in the visualization generation scheme specification, data is extracted from the basic fact dataset to generate a view-driven data package corresponding to each view. Based on the inter-view linkage relationship defined in the visualization generation scheme specification, linkage configuration information is injected into the data packages with linkage relationships. Based on the view-driven data package, a corresponding view is generated, and an event response mechanism between views is established based on the linkage configuration information in the data package.

[0006] Furthermore, the step of obtaining the user's query input via natural language, and using the intelligence big data model and intelligence domain knowledge graph to obtain a structured intent description object for the query, includes: The system obtains queries input by users through natural language, performs semantic parsing on the queries using an intelligence big data model, extracts task type, time constraints, spatial constraints and a set of attention indicators from the queries, obtains a preliminary structured parsing including task type, time constraints, spatial constraints and a set of attention indicators, and identifies one or more entity mentions in the queries. The entity references are input into the knowledge graph of the intelligence domain, and entity linking based on vector similarity and context matching is performed to obtain a core entity set including one or more core entities; Based on the preliminary analysis results and the core entity set, a confidence score is given for each field to obtain a structured intent description object that includes all fields and the confidence score of each field. The fields include task type, time constraint, spatial constraint, core entity set, and attention indicator set.

[0007] Furthermore, the entity references are input into the knowledge graph of the intelligence domain, and entity linking is performed based on vector similarity and context matching to obtain a core entity set including one or more core entities, including: Calculate the semantic similarity score between each entity mention and each entity node in the intelligence domain knowledge graph. Based on the semantic similarity score, match each entity mention with the entity nodes in the intelligence domain knowledge graph. The semantic similarity score is calculated as follows: ,in, It is the first similarity weight coefficient. It is the second similarity weight coefficient. It is the embedding vector of the entity mentioned in the query. These are pre-trained embedding vectors of entity nodes in a knowledge graph within the intelligence domain. It is the context matching degree, and The calculation formula is: ,in, The first context weight coefficient, For the second context weight coefficient, For the third context weight coefficient, The semantic cosine similarity between entity mentions and the overall query context. To ensure the compatibility between entity type and query task type, The degree of overlap and fit between the spatiotemporal characteristics of the entity and the spatiotemporal constraints of the query.

[0008] Furthermore, the confidence score for each field includes: Cross-validation was performed on each field using the corresponding parsing method, and the confidence score for each field was calculated as follows: ,in Score the confidence level of the field. The first confidence level weighting coefficient is used. This is the second confidence level weighting coefficient. This is the third confidence level weighting coefficient. This is the fourth confidence level weighting coefficient. Generate output probabilities for fields in the intelligence big data model. To score the consistency of cross-validation of the fields, Scoring for logical rules in the intelligence field Score the semantic relevance of the field context.

[0009] Furthermore, the step of inputting the structured intent description object into an intelligence domain knowledge graph for contextual reasoning to determine the core entity relationship network involved in the structured intent description object, and generating a visualization generation scheme specification based on the structured intent description object, includes: The structured intent description object is input into an intelligence domain knowledge graph for contextual reasoning to determine the core entity relationship network involved in the structured intent description object; Get a combination of multiple candidate views; Calculate the total fit score between each candidate view combination and the structured intent description object; The candidate view combination with the highest overall fit score is selected as the optimal view combination, and a visualization generation scheme specification is generated based on the optimal view combination.

[0010] Furthermore, calculating the total fit score between each candidate view combination and the structured intent description object includes: The fit score for the i-th candidate view combination is calculated as follows: , in, For the i-th candidate view combination, Let be the fit score of the i-th candidate view combination. The first fitness score is the weighting coefficient. This is the second fitness weighting coefficient. This is the third fitness weighting coefficient. This is the fourth fitness weighting coefficient. Let i be the task matching degree of the i-th candidate view combination. Let i be the entity coverage of the i-th candidate view combination. Let i be the metric support for the i-th candidate view combination. Let be the complexity of the i-th candidate view combination.

[0011] Furthermore, by combining the core entity relationship network, retrieval data about the structured intent description object is obtained from multiple sources, resulting in a basic fact dataset, including: The intelligence big data model is invoked to generate a general analytical narrative text based on the structured intent description object and the core entity relationship network; Based on the structured intent description object, query conditions are generated, and a dedicated data interface or computing engine corresponding to the data pattern in the visualization generation scheme specification is called to obtain the search data retrieved through the query conditions. Based on the core entity relationship network, entities, attributes and relationships related to the retrieved data are extracted from the knowledge graph of the intelligence domain to form a retrieval knowledge subgraph; The retrieved data is aligned, verified, and fused with the retrieved knowledge subgraph to form a basic fact dataset.

[0012] This invention provides an electronic device, comprising: At least one processor; and, A memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that are executed by at least one of the processors to enable the at least one of the processors to perform the intelligence big data model question-answering method as described above.

[0013] The present invention provides a storage medium that stores computer instructions, which, when executed by a computer, are used to perform all the steps of the intelligence big data model question-answering method as described above.

[0014] This invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the intelligence big data model question-and-answer method as described above.

[0015] This invention utilizes a large-scale intelligence model and an intelligence domain knowledge graph to obtain a structured intent description object for the query, and generates a visualization generation scheme specification. Based on this specification, a view of the retrieved data is generated. This invention achieves accurate problem analysis and intelligent classification based on structured intent in intelligence scenarios, as well as multi-dimensional, interconnected, and automatic visualization generation and presentation enhanced by an intelligence domain knowledge graph. Through structured intent parsing in intelligence scenarios, collaboration with an intelligence domain knowledge graph, and intelligent visualization view arrangement, this invention constructs a technical framework that systematically enhances the analytical capabilities of large-scale intelligence models. It effectively solves the problems of non-standardized output and superficial analysis in professional scenarios, and further meets the high standards of intelligence professional decision-making by generating standardized multimodal intelligence analysis results, possessing significant technological advancement and commercialization potential. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the workflow of an intelligence big data model question-answering method according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the workflow of a large-scale intelligence model question-answering method according to another embodiment of the present invention. Figure 3 A flowchart illustrating the workflow of a large-scale intelligence model question-answering method, representing the preferred embodiment of the present invention; Figure 4 This is a schematic diagram of the technical link for knowledge graph reasoning in the preferred embodiment of the present invention; Figure 5 This is a schematic diagram of the hardware structure of an electronic device according to the present invention. Detailed Implementation

[0017] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. Unless the context otherwise requires, throughout the specification and claims, the term "comprising" is interpreted as open-ended and encompassing, that is, "including, but not limited to." In the description of the specification, terms such as "one embodiment," "some embodiments," "exemplary embodiment," "exemplarily," or "some examples" are intended to indicate that a particular feature, structure, material, or characteristic associated with that embodiment or example is included in at least one embodiment or example of this application. The illustrative representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics mentioned may be included in any suitable manner in any one or more embodiments or examples.

[0018] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this application, unless otherwise stated, "a plurality of" means two or more.

[0019] Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0020] The intelligence big data model question-answering method provided by this invention is a user intent and knowledge graph-driven approach. The core of this method lies in constructing a full-link intelligent question-answering mechanism through structured intent recognition and knowledge graph-enhanced reasoning. This mechanism extends from user natural language input to the intelligence big data model generating text responses, and finally to a multi-dimensional, interconnected, and visualized analysis interface, achieving an intelligent upgrade from text-based dialogue to intelligence analysis-assisted decision-making.

[0021] like Figure 1 The diagram shown is a flowchart of a question-answering method for a large intelligence model according to an embodiment of the present invention, including: Step S101: Obtain the query input by the user through natural language, and obtain a structured intent description object about the query through the intelligence big data model and the intelligence domain knowledge graph; Step S102: Input the structured intent description object into the knowledge graph of the intelligence domain for contextual reasoning, determine the core entity relationship network involved in the structured intent description object, and generate a visualization generation scheme specification based on the structured intent description object. The visualization generation scheme specification includes at least the data patterns of multiple views and the linkage relationship between views. Step S103: Combining the core entity relationship network, retrieve retrieval data about the structured intent description object from multiple sources to obtain a basic fact dataset; Step S104: According to the data pattern defined for each view in the visualization generation scheme specification, extract data from the basic fact dataset, generate a view-driven data package corresponding to each view, and inject linkage configuration information into the data packages with linkage relationships according to the linkage relationship between views defined in the visualization generation scheme specification. Step S105: Based on the view-driven data packet, generate the corresponding view, and based on the linkage configuration information in the data packet, establish an event response mechanism between views.

[0022] This invention performs automatic user intent recognition and deep reasoning, and generates an intelligent question-answering method and system with a unified format, clear semantics, and support for multi-dimensional visualization and linkage presentation. It can be used in professional question-answering scenarios in the intelligence field, such as intelligence analysis, situation assessment, and clue mining.

[0023] Specifically, the present invention can be applied to electronic devices with processing capabilities, such as computers.

[0024] First, step S101 is executed to obtain the query input by the user through natural language. Then, through the intelligence big data model and the intelligence domain knowledge graph, a structured intent description object about the query is obtained.

[0025] Specifically, step S101 involves structured intent parsing. A user's natural language query is received, and the query is subjected to deep semantic parsing using an intelligence big data model. Preliminary structured parsing is completed, extracting task type, time constraints, spatial constraints, and a set of relevant indicators, and identifying entity mentions in the query. These entity mentions are preferentially input into an intelligence domain knowledge graph, which is a specialized knowledge graph for the intelligence domain, distinct from general knowledge graphs. It focuses on intelligence-specific entities and professional relationships, and includes standardized fields required for intelligence analysis, such as spatiotemporal features and indicator attributes. Entity linking based on vector similarity and context matching is performed to obtain a standardized core entity set. Combining the preliminary parsing results with these standardized core entities, multi-dimensional confidence scores are applied to each field, outputting a structured intent description object that has passed format, confidence, and logical consistency checks.

[0026] The intelligence big model serves as a large language model fine-tuned for domain instructions. This large language model employs a pre-trained language model based on the Transformer architecture, and its parameters are optimized using domain-specific text and intent parsing task data, enhancing its instruction following and structured output capabilities.

[0027] The structured intent description object is a machine-readable data structure, including: task type, time constraint, space constraint, core entity set, set of indicators of interest, and confidence level of each field; among which, the task type is selected from a predefined enumeration set, including situation overview, event tracking, correlation mining, and trend prediction.

[0028] Then, step S102 is executed, in which the structured intent description object is input into the knowledge graph of the intelligence domain for contextual reasoning, the core entity relationship network involved in the structured intent description object is determined, and a visualization generation scheme specification is generated based on the structured intent description object. The visualization generation scheme specification includes at least the data patterns of multiple views and the linkage relationship between views.

[0029] Specifically, step S102 involves planning and deciding on a visualization solution based on a knowledge graph. The structured intent description object output from step S101 is input into the knowledge graph for contextual reasoning to determine the core entity relationship network involved in the analysis. Simultaneously, a view orchestration decision-maker generates a visualization generation solution specification based on the structured intent description object and the reasoning results from the knowledge graph. The decision-maker selects the optimal solution from multiple candidate view combinations using a suitability scoring model, and then generates the visualization generation solution specification based on the optimal solution.

[0030] The visualization generation solution specification clearly defines the data mode, rendering rules, and linkage relationships of each view.

[0031] Then, step S103 is executed, and the core entity relationship network is combined to retrieve retrieval data about the structured intent description object from multiple sources to obtain the basic fact dataset.

[0032] Specifically, step S103 involves collaborative acquisition and standardized fusion of multi-source data, and the following data processing sub-steps are scheduled and executed in parallel: Semantic narrative generation. By calling upon the intelligence big data model and based on the semantic relationships in the knowledge graph, a general analytical narrative text is generated.

[0033] Precise retrieval of structured data. This involves calling dedicated data interfaces or computing engines corresponding to the data patterns described in the solution specification, constructing query conditions based on structured intent description objects, and then retrieving precise data from multiple sources using these query conditions. The retrieved data includes time-series data, spatial data, and attribute data.

[0034] Knowledge subgraph extraction is performed. Based on the core entity relationship network, entities, attributes, and relationships directly related to the current analysis are extracted from the knowledge graph of the intelligence domain to form a standardized knowledge subgraph.

[0035] Data alignment and fusion. Based on unified time stamps, spatial coordinates, and entity identifiers, the acquired retrieval data and extracted retrieval knowledge subgraphs are aligned, verified, and fused to form a standardized basic fact dataset.

[0036] Then, step S104 is executed: according to the data pattern defined for each view in the visualization generation scheme specification, data is extracted from the basic fact dataset to generate a view-driven data package corresponding to each view; and according to the inter-view linkage relationship defined in the visualization generation scheme specification, linkage configuration information is injected into the data packages with linkage relationship.

[0037] Specifically, based on a pre-built intelligence domain visualization specification library, the basic fact dataset output in step S103 is subjected to key element identification and standardization transformation to obtain a standardized dataset. Subsequently, according to the data pattern defined for each view in the visualization generation scheme specification, data is extracted and reorganized from the standardized dataset to generate a view-driven data package with a standardized format for each view. Finally, based on the inter-view linkage relationship defined in the scheme specification, linkage configuration information is injected into the relevant data packages.

[0038] Finally, step S105 is executed, generating a corresponding view based on the view-driven data packet, and establishing an event response mechanism between views based on the linkage configuration information in the data packet.

[0039] Specifically, the front-end visualization rendering engine receives the visualization generation scheme specification generated in step S102 and the view driver data packages generated in step S104, automatically generates the corresponding views, and establishes an event response mechanism between views based on the linkage configuration in the data packages, so that the user's interaction in a single view can automatically trigger the synchronous update of other related views, forming a unified and linked analysis dashboard.

[0040] This invention utilizes a large-scale intelligence model and an intelligence domain knowledge graph to obtain a structured intent description object for the query, and generates a visualization generation scheme specification. Based on this specification, a view of the retrieved data is generated. This invention achieves accurate problem analysis and intelligent classification based on structured intent in intelligence scenarios, as well as multi-dimensional, interconnected, and automatic visualization generation and presentation enhanced by an intelligence domain knowledge graph. Through structured intent parsing in intelligence scenarios, collaboration with an intelligence domain knowledge graph, and intelligent visualization view arrangement, this invention constructs a technical framework that systematically enhances the analytical capabilities of large-scale intelligence models. It effectively solves the problems of non-standardized output and superficial analysis in professional scenarios, and further meets the high standards of intelligence professional decision-making by generating standardized multimodal intelligence analysis results, possessing significant technological advancement and commercialization potential.

[0041] like Figure 2The diagram shown is a flowchart of a question-answering method for a large intelligence model according to another embodiment of the present invention, including: Step S201: Obtain the query input by the user through natural language, use the big data intelligence model to perform semantic parsing on the query, extract the task type, time constraint, space constraint and set of attention indicators from the query, obtain a preliminary structured parsing including task type, time constraint, space constraint and set of attention indicators, and identify one or more entity mentions in the query. The entity references are input into the knowledge graph of the intelligence domain, and entity linking based on vector similarity and context matching is performed to obtain a core entity set including one or more core entities; Based on the preliminary analysis results and the core entity set, a confidence score is given for each field to obtain a structured intent description object that includes all fields and the confidence score of each field. The fields include task type, time constraint, spatial constraint, core entity set, and attention indicator set.

[0042] Step S202: Input the structured intent description object into the intelligence domain knowledge graph for contextual reasoning to determine the core entity relationship network involved in the structured intent description object; Get a combination of multiple candidate views; Calculate the total fit score between each candidate view combination and the structured intent description object; The candidate view combination with the highest overall fit score is selected as the optimal view combination. A visualization generation scheme specification is generated based on the optimal view combination. The visualization generation scheme specification includes at least the data patterns of multiple views and the linkage relationships between the views.

[0043] Step S203: Invoke the intelligence big model and generate a generalized analytical narrative text based on the structured intent description object and the core entity relationship network; Based on the structured intent description object, query conditions are generated, and a dedicated data interface or computing engine corresponding to the data pattern in the visualization generation scheme specification is called to obtain the search data retrieved through the query conditions. Based on the core entity relationship network, entities, attributes and relationships related to the retrieved data are extracted from the knowledge graph of the intelligence domain to form a retrieval knowledge subgraph; The retrieved data is aligned, verified, and fused with the retrieved knowledge subgraph to form a basic fact dataset.

[0044] Step S204: According to the data pattern defined for each view in the visualization generation scheme specification, extract data from the basic fact dataset, generate a view-driven data package corresponding to each view, and inject linkage configuration information into the data packages with linkage relationships according to the inter-view linkage relationships defined in the visualization generation scheme specification.

[0045] Step S205: Based on the view-driven data packet, generate the corresponding view, and based on the linkage configuration information in the data packet, establish an event response mechanism between views.

[0046] First, step S201 is executed to obtain the query input by the user through natural language. The big data intelligence model is used to perform semantic parsing on the query, extracting task type, time constraint, space constraint and set of attention indicators from the query, and obtaining a preliminary structured parsing including task type, time constraint, space constraint and set of attention indicators, and identifying one or more entity mentions in the query. The entity references are input into the knowledge graph of the intelligence domain, and entity linking based on vector similarity and context matching is performed to obtain a core entity set including one or more core entities; Based on the preliminary analysis results and the core entity set, a confidence score is given for each field to obtain a structured intent description object that includes all fields and the confidence score of each field. The fields include task type, time constraint, spatial constraint, core entity set, and attention indicator set.

[0047] This embodiment illustrates a general intelligent intelligence analysis method. Its core lies in driving a large intelligence model to generate multi-dimensional, standardized, and visual responses that transcend text through intent recognition and knowledge graph reasoning. It aims to address the urgent need of users in professional analysis scenarios to directly obtain intuitive, interconnected, and standardized visual analysis results from single natural language queries.

[0048] Specifically, step S201 is executed first to receive the natural language query text submitted by the user and initiate the structured intent parsing process.

[0049] Specifically, an intelligence structured intent parsing module is constructed. This module, centered on a large intelligence model, integrates various independent parsing methods from the intelligence analysis field, including rule matching, intent classification models, temporal parsing, vector similarity retrieval, and terminology mapping, as well as specialized tools such as regular expression matching tools, domain semantic encoders, knowledge graph entity linking engines, intelligence domain logical rule verification tools, and indicator terminology dictionary projects. These are unified into the intelligence structured intent parsing module system through modular interfaces. Leveraging the semantic understanding capabilities of the large intelligence model and the professional parsing capabilities of the tools, it achieves multi-path independent parsing and cross-validation of results for a single parsing field, ultimately outputting a standardized, machine-readable structured intent description object.

[0050] Then, a predefined structured description template is loaded. This template uses JSON Schema format to define the constraint specifications for multiple dimension fields, including the enumeration value range of the task type field, the format requirements of the time constraint field, and the data type of the core entity field.

[0051] Then, based on prompting engineering techniques, a multi-layered structured input prompt is constructed. The system combines explicit role instructions, complete JSON Schema template descriptions, carefully selected few-sample parsing examples, and the user query text to be processed into a complete sequence of prompt words in a specific order.

[0052] Then, the intelligence big model is invoked, using the constructed prompt words as input. The intelligence big model serves a large language model that has been fine-tuned based on domain instructions. This large language model adopts a pre-trained language model based on the Transformer architecture, and its parameters are optimized through domain-specific text and intent parsing task data, which enhances its instruction following and structured output capabilities.

[0053] Under the strong constraints of structured templates and the guidance of few sample examples, the large language model performs end-to-end semantic understanding and preliminary structured generation tasks: First, it extracts entity mentions, temporal expressions, spatial descriptions, and quantitative dimensions of interest from the query, while initially identifying the task type; then, it inputs the extracted entity mentions into the knowledge graph to perform entity linking, and executes an entity linking algorithm based on vector similarity to accurately match the entities mentioned in the natural language with the standard entity nodes in the knowledge graph.

[0054] In one embodiment, the entity reference is input to an intelligence domain knowledge graph, and entity linking based on vector similarity and context matching is performed to obtain a core entity set including one or more core entities, including: Calculate the semantic similarity score between each entity mention and each entity node in the intelligence domain knowledge graph. Based on the semantic similarity score, match each entity mention with the entity nodes in the intelligence domain knowledge graph. The semantic similarity score is calculated as follows: ,in, It is the first similarity weight coefficient. It is the second similarity weight coefficient. It is the embedding vector of the entity mentioned in the query. These are pre-trained embedding vectors of entity nodes in a knowledge graph within the intelligence domain. It is the context matching degree, and The calculation formula is: ,in, The first context weight coefficient, For the second context weight coefficient, For the third context weight coefficient, The semantic cosine similarity between entity mentions and the overall query context. To ensure the compatibility between entity type and query task type, The degree of overlap and fit between the spatiotemporal characteristics of the entity and the spatiotemporal constraints of the query.

[0055] Specifically, for each matched entity, its semantic similarity score is calculated as follows: ,in: It is the embedding vector of the entity mentioned in the query.

[0056] It is the pre-trained embedding vector of entity nodes in the knowledge graph.

[0057] This refers to the context matching degree. The calculation formula is: , These are weighted coefficients, and their sum is 1. ;in, The semantic cosine similarity between entity mentions and the overall context of the query is calculated by converting both into vectors using a pre-trained domain semantic encoder. The matching degree between entity type and query task type is obtained by looking up a table based on a predefined domain matching degree matrix. The overlap and fit between the entity's spatiotemporal features and the query's spatiotemporal constraints is represented by the weighted value of the spatial area overlap ratio and the temporal duration overlap ratio; weighting coefficients. It can be configured according to specific fields.

[0058] The spatial area overlap ratio refers to the ratio of the area of ​​overlap between the geographical area covered by the inherent spatial features of an entity in the knowledge graph of the intelligence domain and the geographical area limited by the spatial constraints in the user query to the area of ​​their union. The value range is [0,1], and it is used to quantify the degree of matching between the spatial features of the entity and the spatial constraints of the query.

[0059] The time overlap ratio refers to the ratio of the overlap between the inherent time features of an entity in the knowledge graph of the intelligence domain and the time constraints in the user query, to the total time of their union. The value ranges from [0,1] and is used to quantify the degree of matching between the entity's time features and the query time constraints.

[0060] and It is an adjustable similarity weight coefficient that satisfies... .

[0061] This method associates entity surface names with knowledge graph standard identifiers; then, it integrates standardized entities, normalized spatiotemporal constraints, and standardized indicator terms to determine the final task type classification.

[0062] The structured intent description object output by the large language model strictly follows the JSON Schema specification, including accurate task type classification, standardized entity recognition results, normalized spatiotemporal constraints, standardized indicator terminology descriptions, and confidence scores for each of the above fields.

[0063] In one embodiment, the confidence score for each field includes: Cross-validation was performed on each field using the corresponding parsing method, and the confidence score for each field was calculated as follows: ,in Score the confidence level of the field. The first confidence level weighting coefficient. This is the second confidence level weighting coefficient. This is the third confidence level weighting coefficient. This is the fourth confidence level weighting coefficient. Generate output probabilities for fields in the intelligence big data model. To score the consistency of cross-validation of the fields, Scoring for logical rules in the intelligence field Score the semantic relevance of the field context.

[0064] Specifically, the confidence score for each field is calculated using a weighted fusion formula:

[0065] in, This represents the final confidence score of field f, with a value range of [0,1]. , , , The confidence weight coefficient satisfies . , , , The values ​​are all in the range of [0,1].

[0066] The output probabilities of fields generated for the intelligence big data model are derived from the output probabilities of the intelligence big data model when generating the parsing results. The intelligence big data model is based on a Transformer architecture and employs an autoregressive generation mechanism. With the model's probability output function enabled, when the model generates the content of each field, it simultaneously acquires the raw score (logits) corresponding to each token during the generation process. After normalization using the Softmax function, the generation probability of each token is obtained, forming the probability distribution of tokens corresponding to each field. For discrete fields, the normalized probability of the model's output target value is used; for text-type fields, the average token probability of the generated text sequence is used. This parameter directly reflects the model's certainty regarding the parsing results, and its value range is [insert range here]. .

[0067] The cross-validation consistency score for the parsing results is obtained through cross-validation of multiple independent parsing methods integrated into the intelligence big data model. Preferably, the task type field uses both rule matching and intent classification models; the time field uses both temporal parsers and regular expression matching; the entity field uses both vector similarity retrieval and dictionary-based exact matching; and the indicator field uses both terminology mapping tables and semantic similarity calculation. This parameter represents the proportion of consistency between the parsing results from each method. A proportion of 1 indicates complete consistency between the two parsing methods, while a proportion of 0 indicates inconsistency, reflecting the robustness of the parsing results.

[0068] The scoring for logical rule validation in the intelligence domain is obtained through the validation of logical rules in the intelligence domain. The system's built-in rule base includes type matching rules, value range constraint rules, and association rationality rules. Type matching rules check the compatibility between entity types and indicators; value range constraint rules verify time sequence and numerical range; association rationality rules evaluate the semantic association strength between different fields. This component calculates a penalty value based on the number and severity of rule violations. Type matching rules and value range constraint rules are classified as core rules, and association rationality rules are classified as general rules. Violation of a core rule deducts 0.5 points, violation of a general rule deducts 0.3 points, and no violation results in a penalty. The minimum score after deduction is .

[0069] The semantic relevance score of the parsed results is obtained through semantic similarity calculation. The system uses a pre-trained semantic encoder to convert the parsed field results and the original query into vector representations, and calculates the cosine similarity between the two as the contextual relevance. Simultaneously, the semantic relevance of this field to other parsed fields is calculated. This invention employs a weighted fusion method to calculate the final score. The value, where the weight of contextual relevance to the original query is taken as... The semantic relevance weights of the data to other parsed fields are taken as follows: The sum of the two weights is This component reflects the reasonableness of the parsing results within the overall query context.

[0070] Finally, the output results undergo multi-layered quality checks: First, JSON format syntax is checked to ensure the data structure meets expectations; second, a threshold judgment is made on the confidence scores of core fields. When the confidence scores of key fields such as task type and core entities are lower than the preset threshold, the system triggers a multi-round intent clarification interaction process to obtain clear feedback from users through targeted questions; finally, the logical consistency between fields is checked to eliminate contradictory or conflicting parsing results.

[0071] During the intent clarification interaction, the system will dynamically update the confidence score based on user feedback. When a user confirms a parsing result, the confidence score of the relevant field increases according to the following formula:

[0072] in For the updated confidence level, The confidence level before the update. This is the learning rate parameter, with a default value of 0.3. When the user corrects the parsing results, the system recalculates the confidence score of this field and records the correction pattern for subsequent parsing optimization.

[0073] After all validations pass, the system will pass the complete and standardized structured intent description object to the downstream knowledge graph reasoning and visualization decision-making module, providing accurate machine-readable input for subsequent analysis processes.

[0074] Then, step S202 is executed, in which the structured intent description object is input into the intelligence domain knowledge graph for contextual reasoning to determine the core entity relationship network involved in the structured intent description object; Get a combination of multiple candidate views; Calculate the total fit score between each candidate view combination and the structured intent description object; The candidate view combination with the highest overall fit score is selected as the optimal view combination. A visualization generation scheme specification is generated based on the optimal view combination. The visualization generation scheme specification includes at least the data patterns of multiple views and the linkage relationships between the views.

[0075] Specifically, step S202 involves making a decision on a knowledge graph enhancement and visualization solution.

[0076] like Figure 4 The diagram shown is a schematic representation of the technical link of knowledge graph reasoning in the preferred embodiment of the present invention, including: Step S401, Input: Structured intent description object, extract core entity identifiers; Step S402, entity matching, calculate similarity score based on vector similarity algorithm; Step S403: Knowledge graph standard entity node matching, dynamically adapted based on task type; Step S404, multi-hop relationship reasoning, heuristic search strategy; Step S405: Construction of core entity relationship network, including deduplication, critical path extraction, and standardized encapsulation; Step S406: Extract knowledge subgraphs; Step S407: Visualize the solution decision-making process and integrate multi-source data.

[0077] Specifically, a knowledge graph storage and reasoning engine is built to receive structured intent description objects and extract standardized core entity identifiers from them.

[0078] Based on the linked, standardized entity nodes, the engine executes existing multi-hop relationship reasoning algorithms to retrieve relevant subgraph structures. The algorithm automatically determines the depth and breadth of relationship exploration based on the task type; for example, it uses three-hop relationship retrieval for "association mining" tasks and two-hop relationship retrieval for "situational overview" tasks.

[0079] The multi-hop relation reasoning algorithm executed by the knowledge graph storage and inference engine is a task-adaptive, semantically weighted graph traversal method. The algorithm first dynamically determines the maximum query depth, priority relation types, and result size constraints based on the task type in the structured intent. The result size constraint is a quantitative result control threshold adapted to the task type, corely including three types of indicators: node number constraint, relation edge constraint, and path number constraint, and supports an elastic threshold mechanism: the node number constraint limits the maximum number of entity nodes in the retrieved knowledge subgraph according to the task type, with 20-50 for situation overview, 50-100 for association mining, 30-60 for event tracking, and 20-40 for trend prediction; the relation edge constraint is set to 1.5-3 times the number of nodes and dynamically adjusted in conjunction with relation semantic weights; the path number constraint sets hierarchical thresholds for different hop counts and filters duplicate paths; simultaneously, based on the knowledge graph network density of the core entities, each threshold is elastically adjusted by 20%-30% to avoid subgraph redundancy or incompleteness.

[0080] During the traversal, a heuristic search strategy is employed, specifically a weighted scoring and ranking strategy. This strategy calculates the expansion priority score of candidate objects using a quantification formula, expanding them sequentially from highest to lowest score. The priority is determined by comprehensively considering the semantic weight of the relationship type, the centrality measure of the target node in the network, and the confidence decay caused by path length. First, the three dimensions are normalized and quantified (all values ​​range from [0,1]). The semantic weight of the relationship type is assigned according to domain importance, the centrality measure of the target node is calculated jointly using fusion centrality and betweenness centrality, and the confidence decay of path length is quantified using an exponential decay formula. Then, the expansion priority score is obtained by weighted summation using task-adaptive weight coefficients, which are dynamically adjusted according to the task type. Simultaneously, a preset scoring threshold filters low-priority candidate objects. If candidate objects with the same score have the same number of hops, objects with higher relational semantic weights are prioritized for expansion, followed by objects with higher node centrality. The algorithm achieves efficient traversal by calling the declarative multi-hop query interface of the graph database, and performs deduplication, key path extraction and standardization encapsulation on the returned results. Finally, it outputs a core entity relationship network knowledge subgraph that is closely related to the user's query intent, has a compact structure and is rich in semantics, and serves as the semantic basis for subsequent visualization decision-making and retrieval knowledge subgraph extraction.

[0081] Simultaneously, the structured intent description object is received through the view orchestration decision-maker. Internally, the decision-maker maintains a configurable policy knowledge base, storing recommended candidate view combinations for different task and entity type combinations. The decision-making process employs a strategy combining rule-based initial screening with optimization-based fine-grained ranking: First, based on the task type in the structured intent description object and the core entity type of the core entities identified through the intelligence domain knowledge graph, all relevant candidate view combinations are screened from the policy base. Next, for each candidate view combination, its overall fit score is calculated, and the candidate view combination with the highest fit score is selected as the final choice.

[0082] In one embodiment, calculating the total fit score between each candidate view combination and the structured intent description object includes: The fit score for the i-th candidate view combination is calculated as follows: , in, For the i-th candidate view combination, Let be the fit score of the i-th candidate view combination. The first fitness score is the weighting coefficient. This is the second fitness weighting coefficient. This is the third fitness weighting coefficient. This is the fourth fitness weighting coefficient. Let i be the task matching degree of the i-th candidate view combination. Let i be the entity coverage of the i-th candidate view combination. Let i be the metric support for the i-th candidate view combination. Let be the complexity of the i-th candidate view combination.

[0083] Specifically, the fit score for the i-th candidate view combination is calculated as follows: , in, For the i-th candidate view combination, The fit score is the score for the i-th candidate view combination. Let i be the task matching degree of the i-th candidate view combination. The task matching function, with values ​​ranging from [0,1], is based on a predefined [task type - view type] adaptation matrix (manually calibrated by intelligence experts and embedded in the system). It weights and sums the adaptation scores of each view in the candidate view combination, using the following formula: ,in, Combining candidate views The number of views in the data. Combining candidate views The Middle The weight of each view ( (Core views have higher weight) Combining candidate views The Middle Each view and structured intent describes the current task within the object. The fit score assesses how well the candidate view combination matches the current task type. Let i be the entity coverage of the i-th candidate view combination. Let be the entity coverage function, with a value range of [0,1]. Let be the ratio of the number of core entities that can be displayed in a candidate view combination to the total number of core entities. Its formula is: ,in Combining candidate views The number of core entities that can be displayed The total number of core entities in the core entity set of the structured intent description object. Evaluate the display coverage of the core entity set by the candidate view combination; Let i be the metric support for the i-th candidate view combination. The indicator support function takes values ​​in the range [0,1]. Based on a predefined [indicator type - view type] support matrix (calibrated manually by intelligence experts and embedded in the system), it sums the highest support scores of each relevant indicator in the candidate view combination and then normalizes the result. The formula is as follows: ,in The total number of metrics for a set of metrics of interest within a structured intent description object. Indicates candidate view combination A single view in Combining candidate views A collection of views For the first Individual metrics in candidate view combinations The highest support score among them. Evaluate the ability of candidate view combinations to visually support a set of metrics of interest; Let be the complexity of the i-th candidate view combination. This is a complexity penalty function, with values ​​ranging from [0,1], combined with the i-th candidate view. The number of views, rendering, and interaction complexity are directly proportional, as shown in the formula: ,in , , Preset weighting coefficients (which can be determined manually) to satisfy , The view complexity is calculated based on the number of views in the view composition and the system's preset maximum number of views. The specific formula is as follows: ,in Combining candidate views The number of views, Set a maximum allowed number of views for the system; The rendering complexity is determined according to the system's predefined hierarchical rules of [data volume level - rendering complexity], such as data volume. When, Data volume When, When multiple views exist in the same combination, the one with the highest rendering complexity is taken as the maximum. ; Interaction complexity is determined according to the system's predefined hierarchical rules of [interaction function set - interaction complexity], such as static display only, no interaction: It supports filtering, sorting, and viewing: 3; , , , This is the fitness weighting coefficient, used to balance the contributions of different objectives.

[0084] The decision-maker calculates the total fit score of all candidate view combinations using the above formula, selects the candidate view combination with the highest score as the optimal view combination, and systematically integrates the core decision results of the optimal view combination, single-view generation configuration, multi-view linkage rules, and downstream data retrieval and rendering interface specifications based on a pre-defined domain visualization specification library (which is a set of standardized visualization rules predefined for the needs of the vertical domain and covering the entire process of view generation, linkage, and rendering). This is then standardized and encapsulated according to a preset structured data format (JSON) to generate a detailed visualization generation scheme specification. This specification clarifies the full-dimensional executable configuration information for visualization generation and linkage presentation, providing a unified parsing and execution basis for downstream multi-source data collaborative scheduling, visualization-driven data construction, and front-end rendering.

[0085] Here is a simple example of a visualization-based solution specification: { "viewCombination": { "mainView": "Entity Relationship Network Diagram", "auxiliaryView": "Time-series statistical bar chart", "layoutMode": "Primary and secondary partition layout" "dataMode": [{ "viewType": "Entity Relationship Network Diagram", "requiredFields": ["Source Entity Identifier", "Target Entity Identifier", "Relationship Type", "Relationship Weight"], "dataRequirements": "Completely matches the core entity relationship network, retains only high-confidence related data, and normalizes relationship weights to the [0,1] interval." {"viewType": "Time Series Statistical Bar Chart", "requiredFields": ["Statistical Time", "Association Frequency", "Entity Identifier"], "dataRequirements": "Complies with the visualization data call interface specifications, with daily statistics and removal of redundant data" "linkageRule": { "linkageKey": "Entity Identifier", "linkageLogic": "Using the entity relationship network diagram as the interaction trigger source, when the user selects any entity node, the time series statistics bar chart automatically filters and renders the time series association frequency data of the corresponding entity; after deselecting, the auxiliary view restores the full data display."

[0086] Then, step S203 is executed, invoking the intelligence big model to generate a generalized analytical narrative text based on the structured intent description object and the core entity relationship network; Based on the structured intent description object, query conditions are generated, and a dedicated data interface or computing engine corresponding to the data pattern in the visualization generation scheme specification is called to obtain the search data retrieved through the query conditions. Based on the core entity relationship network, entities, attributes and relationships related to the retrieved data are extracted from the knowledge graph of the intelligence domain to form a retrieval knowledge subgraph; The retrieved data is aligned, verified, and fused with the retrieved knowledge subgraph to form a basic fact dataset.

[0087] Specifically, in step S203, multi-source data collaborative acquisition and fusion processing are performed.

[0088] First, the visualization of the solution specification is analyzed and extracted using the multi-source data collaborative scheduling module.

[0089] The multi-source data collaborative scheduling module schedules multiple data processing tasks in parallel based on the type of data requirement (time-series data, spatial data, attribute data) and timeliness requirements. It retrieves data based on structured intent description objects and the core entity relationship network. For time-series data requirements, it calls the time-series database query interface to construct a query statement containing time range, aggregation granularity, and filtering conditions, retrieving time-series quantitative data of core entities and key indicators under specified spatiotemporal constraints. For spatial data requirements, it calls the geographic information system interface or spatial database, submitting a spatial query containing geographic range, layer type, and accuracy requirements, retrieving the geospatial attributes, spatial distribution characteristics, and regional association data of core entities. For attribute data requirements, it calls the Representational State Transfer Application Programming Interface (REST API) of the business database to construct an entity attribute retrieval request, retrieving the basic business attributes, static characteristics, and inherent association information of core entities. Simultaneously, it calls a domain-adapted large language model, generating a summary analytical narrative text based on the structured information of the knowledge subgraph of the core entity relationship network.

[0090] The generation of the summary analysis text first inputs three core pieces of information into the domain-adaptive large language model: basic constraint information such as task type, spatiotemporal constraints, core entities, and attention indicators from the user's structured intent; structured semantic information such as core entity relationships, attribute features, and semantic logic extracted from the knowledge subgraph of the core entity relationship network; and real-valued information such as quantitative data, distribution characteristics, and trends obtained from precise multi-source retrieval of time, space, and attributes, achieving the fusion of multi-dimensional information input. Secondly, through domain-based prompting engineering, generation constraints are injected into the large language model, clarifying the professional style requirements, structural logic requirements, domain expression standards, and content boundary requirements of the text. The professional style requirements match the analytical expression habits of the vertical domain; the structural logic requirements organize the text according to a general-specific-general logic; and the domain expression standards require the unified use of domain-standard terminology, indicator names, and expression forms. The content boundary requires analysis and generation based solely on input multi-source information, without unfounded reasoning or extension. Finally, the domain-adapted large language model, based on the Transformer architecture's deep semantic understanding and generation capabilities, deeply mines and logically integrates the fused structured semantic information and quantified real-value information. End-to-end natural language generation is performed according to the constraints of the prompting engineering. After generation, the system performs data consistency and semantic fit checks on the text. Data consistency checks ensure that the entity names, indicator values, spatiotemporal ranges, and other information in the text are completely consistent with the multi-source retrieved data. Semantic fit checks ensure that the text's analytical logic is consistent with the entity relationships and semantic logic in the core entity relationship network knowledge subgraph. If the checks pass, it becomes the final summary analysis and description text; if the checks fail, the prompting engineering constraints are fine-tuned and information is re-entered for generation.

[0091] After each data processing task is completed, based on the core entity relationship network, entities, attributes, and relationships related to the retrieved data are extracted from the intelligence domain knowledge graph to form a retrieval knowledge subgraph. Then, the data alignment and fusion engine receives the returned raw data. The engine performs the following key processes: Entity Unification Alignment: Based on standard entity identifiers obtained from the retrieval knowledge subgraph, data pointing to the same entity from different data sources are associated and merged. Spatiotemporal Consistency Verification: The continuity of timestamps in time-series data and the consistency of geographic coordinates in spatial data are checked, and outliers are marked or corrected. Multi-Source Evidence Fusion: For multiple data sources of the same fact, a weighted fusion algorithm based on source credibility is used.

[0092] Finally, the data alignment and fusion engine outputs a standardized dataset of basic facts. All data items in this dataset have uniform entity identifiers, timestamp formats, coordinate reference systems, and units of measurement.

[0093] Then, step S204 is executed: according to the data pattern defined for each view in the visualization generation scheme specification, data is extracted from the basic fact dataset to generate a view-driven data package corresponding to each view; and according to the inter-view linkage relationship defined in the visualization generation scheme specification, linkage configuration information is injected into the data packages with linkage relationship.

[0094] Specifically, step S204 involves visually driven data construction and configuration injection.

[0095] The system loads a pre-built domain visualization specification library, which contains data format conversion rules, visual coding specifications, and interactive behavior definitions.

[0096] The domain visualization specification library predefines multiple view types, each associated with a set of metadata descriptions, including applicable task types, recommended entity types, supported data patterns, and visual complexity levels.

[0097] To support intelligent linkage between views, its specification library predefines a set of configurable interactive events and response behavior templates for each view type. For example, the selection highlight behavior means that when a user selects an entity, that entity is highlighted in all views; the filter linkage behavior means that after setting filter conditions in one view, other views will automatically filter synchronously.

[0098] The system first performs a normalization transformation on each data item in the basic fact dataset: a. Convert time data to ISO 8601 standard format.

[0099] b. Unify geographic coordinates to a specified coordinate system (such as WGS84).

[0100] c. Entity names are mapped to standard terms.

[0101] d. Perform necessary normalization or standardization on numerical indicators.

[0102] Based on the data schemas for each view defined in the visualization generation solution specification, data is extracted and organized from the standardized dataset to generate a driving data package for each view. The structure of each data package strictly follows the predefined schema for its view type.

[0103] The driving data package is a standardized data carrier that fits the rendering requirements of view types. Its structure strictly follows the schema predefined for each view type in the domain visualization specification library. The core includes four major modules: basic metadata, core rendering data, linkage configuration related data, and rendering control rules.

[0104] Basic metadata is the foundational data used for view identification and tracing, including the view's unique ID, view type, associated structured intent ID, data source identifier, data generation time, and data version number, ensuring that data packets can be accurately identified and associated by the front-end engine.

[0105] The core rendering data consists of the core quantitative and semantic data for the visualization of the view. It is organized differently according to the view type. For example, the geographic map view consists of latitude and longitude coordinates and entity attribute values, the time series chart view consists of timestamps and indicator values, and the knowledge graph view consists of entity nodes, relationship edges and attribute features, providing core data support for view rendering.

[0106] Linkage configuration data is the key identifier data for realizing intelligent linkage between views. It includes core entity identifiers, indicator codes, spatiotemporal unified identifiers, etc., which are matched one by one with the core identifiers in the linkage configuration information to ensure that the source view interaction events can accurately trigger the state update of the target view.

[0107] Rendering control rules are rule data that guides the standardized rendering of front-end components. They include data display granularity, visual encoding mapping rules (such as the correspondence between numerical values ​​and colors / sizes), legend configuration data, coordinate axis display rules, and abnormal data display styles, ensuring that the rendering results conform to the domain visualization specifications.

[0108] The process of generating the data package involves first matching the predefined schema corresponding to the view type in the domain visualization specification library according to the view type specified in Step 2 of the visualization generation scheme specification, and clarifying the field structure, data type and field constraints of the data package; secondly, according to the field requirements of the schema, extracting the corresponding data content from the standardized basic fact dataset as needed, and organizing and reorganizing it in a structured manner by module; finally, supplementing each module with complete field information to complete the construction of the data package, ensuring that the data package can be directly parsed and called by the front-end visualization rendering engine.

[0109] Based on the view linkage rules defined in the visualization generation solution specification, linkage configuration information is injected into the relevant driver data package. The linkage configuration uses a declarative syntax to define the mapping relationship between source view events and target view state updates. For example, the configuration can specify that when a data item in view A is selected, its associated identifier will be passed to view B as a filter condition.

[0110] Finally, step S205 is executed, generating a corresponding view based on the view-driven data packet, and establishing an event response mechanism between views based on the linkage configuration information in the data packet.

[0111] Specifically, step S205 involves dynamic generation and intelligent interactive presentation of multiple views. The front-end multi-view collaborative rendering engine receives the visualization generation solution specification and the data packages for each view. First, it parses the layout definition in the solution specification, dynamically creates the corresponding view container components according to the layout dimensions, view hierarchy, and display proportion, and establishes the layout relationships between containers according to preset layout rules.

[0112] For each view container, the engine dynamically loads adapted visualization components (such as chart components, map components, knowledge graph display components, statistical report components, etc.) from the domain visualization component library according to its preset view type, and injects the corresponding driver data package completely into each visualization component. Based on the injected driver data package, the visualization component parses the core rendering data, rendering control rules, and basic metadata, and completes visualization rendering in a data-driven manner, automatically adapting to data dimensions and display requirements, generating standardized views that meet the requirements of the domain visualization specification library, and realizing the dynamic generation and display of single views.

[0113] The rendering engine simultaneously parses the pre-defined linkage configuration information in each driver data package, and builds an event listening and response mechanism based on the publish-subscribe pattern among all visualization components. Each component is configured with a dedicated event listener, and a global event publisher is deployed uniformly. When a user triggers an interactive event such as selection, filtering, hovering, or dimension switching in any source view, the event publisher of the source view will extract the event data and core associated identifiers (entity ID, indicator code, spatiotemporal identifier) ​​in real time and broadcast it globally. After receiving the broadcast information, the event listeners of each target view update their own driver data packages in real time and filter the data according to the triggering rules and response logic defined in the linkage configuration, thereby triggering the local dynamic re-rendering of the visualization components and realizing the synchronous update of the target view's state.

[0114] Ultimately, users gain a complete multi-view analytics dashboard. The views are not only visually consistent but also logically interconnected through pre-defined linkage rules, enabling users to gain in-depth analytical insights through interactive exploration.

[0115] This embodiment achieves accurate question analysis and intelligent classification based on structured intent in intelligence scenarios. By parsing the natural language of user intelligence queries into structured intent description objects containing multi-dimensional fields such as intelligence analysis task type, core entities, and spatiotemporal constraints, the system can deeply understand and accurately classify complex and ambiguous user intelligence requests. This allows the system to direct questions to appropriate intelligence professional analysis channels, overcoming the problems of vague intent understanding and lack of targeted response in traditional large-scale intelligence question answering models. This significantly improves the accuracy and professionalism of intelligent question answering in the intelligence field.

[0116] Meanwhile, this embodiment achieves multi-dimensional, interconnected, and automated generation and presentation of knowledge graphs in the intelligence domain. Through the collaboration between the intelligence visualization view orchestration decision-maker and the intelligence domain knowledge graph, the system can dynamically plan and generate the most suitable multi-view combination based on the structured intent of the intelligence scenario, and ensure intelligent linkage between views based on intelligence semantic relationships. This method breaks through the limitations of existing visualization systems that can only provide fixed view menus or isolated charts, upgrading the text output of the large intelligence model into an intuitive and deeply interconnected intelligence situation dashboard, greatly enhancing the intuitiveness and operability of intelligence analysis results, and adapting to the needs of clue mining, situation reconstruction, and correlation verification in intelligence work.

[0117] Finally, this embodiment provides a systematic enhancement solution for the in-depth analysis and decision-making applications of large intelligence models. Through contextualized and structured intent parsing of intelligence scenarios, collaboration with intelligence domain knowledge graphs, and intelligent visualization of intelligence views, this embodiment constructs a technical framework that can systematically improve the analytical capabilities of large intelligence models. It effectively solves the problems of non-standardized output and superficial analysis in professional scenarios, and further meets the high standards of professional intelligence decision-making by generating standardized multimodal intelligence analysis results, demonstrating significant technological advancement and commercialization potential.

[0118] like Figure 3 The diagram shown is a flowchart of a question-answering method for a large intelligence model according to the preferred embodiment of the present invention. Using the enterprise business field as an application scenario, it fully demonstrates the entire execution process of the method and system of the present invention. The method includes: Users submit natural language queries to the system: Analyze the product sales trends and channel layout data of a competitor brand in East China in the third quarter of 2025 to uncover the core logic of the competitor's market strategy.

[0119] Step S301: Structured Intent Parsing: After receiving the above query, the intelligence system processes it through the structured intent parsing module. It first extracts entity mentions from the query, performs knowledge graph entity linking based on vector similarity and context matching on the entity mentions, calculates semantic similarity scores, and obtains a standardized core entity set.

[0120]

[0121] It is the embedding vector mentioned by the entity in the query; It is the pre-trained embedding vector of entity nodes in the knowledge graph; This refers to the context matching degree. The calculation formula is: , These are weighted coefficients, and their sum is 1. ;in, The semantic cosine similarity between entity mentions and the overall context of the query is calculated by converting both into vectors using a pre-trained domain semantic encoder. The matching degree between entity type and query task type is obtained by looking up a table based on a predefined domain matching degree matrix. The overlap and fit between the entity's spatiotemporal features and the query's spatiotemporal constraints is represented by the weighted value of the spatial area overlap ratio and the temporal duration overlap ratio; weighting coefficients. Configuration can be done by business intelligence area, with weighting coefficients set accordingly. .

[0122] and It is an adjustable similarity weight coefficient that satisfies... You can take it here. , .

[0123] After completing confidence scoring and logical validation, the final output is a structured intent description object conforming to the JSON Schema specification. For example, the JSON format of a label might look like this: { "task_type": ["Overview Category", "Relationship Mining Category"], "time_constraint": {"start": "2025-07-01", "end": "2025-09-30", "format": "ISO 8601"}, "space_constraint": {"area": ​​"East China", "coord": ["115°E-123°E", "28°N-35°N"], "coord_system": "WGS84"}, "core_entity_set": ["Competitor Brand A", "Competitor Brand B", "Competitor Brand C"], "attention_indicator_set": ["Monthly Sales", "Single Channel Sales Revenue", "Number of Offline Stores", "Online Conversion Rate", "Regional Channel Coverage"] "field": "business", "confidence": {"task_type": 0.95, "core_entity": 0.96, "time_constraint": 0.98, "space_constraint": 0.97} } The object needs to undergo format validation, confidence threshold judgment (confidence of core fields is ≥0.9) and logical consistency verification. After no abnormalities are found, it is passed to the downstream module.

[0124] Step S302: Decision-making on knowledge graph enhancement and visualization solutions: The knowledge graph storage and reasoning engine extracts standardized core entities (competitor brands) that have completed entity linking in step one, performs three-hop relationship reasoning (association mining task), and sets result size constraints: 50-100 entity nodes and relationship edges twice the number of nodes. Through heuristic search strategies, it filters high semantic weight relationships of brand affiliation, channel cooperation, and regional layout, and finally extracts a standardized knowledge subgraph that covers the core semantic information of brand architecture, product system, channel layout, and regional market.

[0125] The view orchestration decision-maker may be configured with the following weighting coefficients. =0.3、 =0.3、 =0.4、 =0.2. Based on the fit scoring model, the optimal view combination is selected from the candidate view combinations in the business domain. The possible result is a geospatial map view + time series trend chart view + knowledge graph association view + statistical report view, with a layout ratio of 40%:25%:20%:15%.

[0126] The decision-maker generates a JSON-formatted visualization generation solution specification based on a commercial visualization standard library, which clarifies the data patterns, rendering rules, and linkage relationships of each view. It defines the geospatial map view as the core source view, and when a selection / filtering event is triggered, the other views are updated synchronously, providing a unified basis for downstream data processing.

[0127] Step S303: Collaborative Acquisition and Fusion Processing of Multi-Source Data The multi-source data collaborative scheduling module parses and visualizes the generation scheme specification, performs parallel data retrieval and analysis, generates narrative text for three types of data, and finally outputs a standardized basic fact dataset. Its time-series data calls upon a commercial time-series database to retrieve quantitative data on monthly sales, single-channel sales revenue, and online conversion rates for each competing brand from July to September 2025. Its spatial data utilizes commercial GIS spatial databases to retrieve spatial attribute data on geographic coordinates in East China, distribution of competitor offline stores, and spatial location of regional warehousing centers. Its attribute data calls the REST API of the commercial product attribute database to retrieve static attribute data of each competing brand's main product types, pricing ranges, and channel cooperation models.

[0128] It invokes a business-domain-adapted intelligence model, integrates input structured intent, knowledge subgraph semantic information, and multi-source retrieval quantitative data, and generates analytical text based on business intelligence domain prompts and engineering constraints (professional business terminology, overall structure).

[0129] The core content that may be generated is as follows: In the third quarter of 2025, competing brands A / B / C in East China all maintained stable sales growth, with total sales of 120,000 units, 90,000 units, and 75,000 units in July, September, and August, respectively, and online sales accounting for 65%, 58%, and 72% respectively; the offline store coverage areas of brands A and B overlapped by 80%, both mainly through supermarket channels, while the coverage area of ​​brand C overlapped by 40% with A / B, mainly through community convenience stores. The three brands formed a channel layout pattern of "core supermarkets + community supplements" in East China.

[0130] The text has been verified for both data consistency and semantic fit, and has been confirmed as valid.

[0131] The data alignment and fusion engine aligns the retrieved data with the retrieved knowledge subgraph based on the unified business entity identifier, ISO 8601 time format, and WGS84 geographic coordinates. It marks and corrects outliers in the data and integrates multi-source data using a source credibility weighted fusion algorithm to output a standardized basic fact dataset with unified fields, standardized format, and semantic relevance.

[0132] Step S304: Visualization-driven data construction and configuration injection: The system loads a commercial domain visualization standard library and performs normalization transformations on the basic fact dataset (time to ISO 8601, coordinates to WGS84, entity names mapped to commercial standard terminology). It generates driving data packages for four types of views. Each data package contains four main modules: basic metadata, core rendering data, linked configuration data, and rendering control rules, strictly adhering to a predefined commercial domain schema. Simultaneously, it injects linked configuration information into all data packages, defined using a declarative syntax: when a competitor's brand entity ID / regional coordinates are selected in the geospatial map view, other views filter and highlight data based on this identifier.

[0133] Taking a geospatial map view-driven data package as an example, its core framework may be as follows: { "base_meta": {"view_id": "V001", "view_type": "geo_map","intent_id":"SI20251001001", "data_source": "Commercial GIS Spatial Database", "create_time": "2025-10-01T10:20:00Z", "version": "V1.0"}, "core_render_data": [ {"entity": "Competitor Brand A", "entity_code": "B001", "coord": ["120°E", "31°N"], "store_area": ​​["118°E-122°E", "29°N-33°N"], "sales_q3": 120000, "online_sales_ratio": 0.65, "main_channel": "Supermarkets"}, {"entity": "Competitor Brand B", "entity_code": "B002", "coord": ["119°E", "32°N"], "store_area": ​​["117°E-121°E", "30°N-34°N"], "sales_q3": 90000, "online_sales_ratio": 0.58, "main_channel": "Supermarkets"}, {"entity": "Competitor Brand C", "entity_code": "B003", "coord": ["121°E", "30°N"], "store_area": ​​["119°E-123°E", "28°N-32°N"], "sales_q3": 75000, "online_sales_ratio": 0.72, "main_channel": "Community Convenience Store"} ], "linkage_data": {"core_id": ["entity_code", "space_coord"], "source_view": true, "target_view": ["V002", "V003", "V004"]}, "render_rule": {"coord_system": "WGS84", "color_map": {"Competitor Brand A":"#FF4757", "Competitor Brand B": "#1E90FF", "Competitor Brand C": "#2ED573"}, "layer": "Commercial Area Channel Layer", "legend_show": true, "anomaly_style": "Red Flashing"} } Step S305: Dynamic Generation and Intelligent Linkage of Multiple Views: The front-end multi-view collaborative rendering engine receives the visualization generation scheme specification and four types of view-driven data packages, and completes the construction of the multi-view dynamic generation and intelligent linkage mechanism according to the process: Dynamically create view containers: Create 4 view containers dynamically according to the layout ratio of 40%:25%:20%:15%, and establish a side-by-side layout relationship with the geospatial map view as the main view and the rest as sub-views. Dynamically load and render components: Load geographic map, time series chart, knowledge graph and statistical report components from the business domain visualization component library, inject the corresponding driver data package into the component, and each component completes rendering in a data-driven manner to generate standardized views that conform to business domain specifications, such as geographic map view marking competitor offline stores / coverage area, and time series chart view displaying monthly sales / online conversion rate trends; Build a publish-subscribe pattern event listening and response mechanism: Configure dedicated event listeners for the four components, deploy a global event publisher, and realize global broadcasting of source view interaction events and real-time response of target view.

[0134] The system provides users with a multi-view intelligent linkage analysis dashboard of the competitive market situation in East China. This visual interactive interface includes a summary analysis and description text display area plus four core visual views, and has the core capabilities of independent viewing of a single view and intelligent linkage of multiple views.

[0135] Summary Analysis Text Display Area: Displays verified business analysis texts, which can be viewed and exported; Geospatial Map View (Main View): Displays the geographical outline of East China based on the WGS84 coordinate system, accurately marks competitors' offline stores / coverage areas, distinguishes brands with different colors, displays sales density with a heat map, and supports brand / region selection interaction; Time-series trend chart view: The line chart shows the monthly sales and online conversion rate trends of the three brands, and the bar chart shows the sales revenue of a single channel. Time dimension filtering is supported. Knowledge graph association view: Visually displays entity nodes and relationship edges, highlighting core associations such as channel cooperation, product competition, and regional layout between brands. The size of the node corresponds to the brand's market share, and the thickness of the edge corresponds to the semantic weight of the relationship. Statistical Report View: Displays summary statistics of brand basic attributes, sales data, and channel layout data in a standardized business report format, supporting data sorting and filtering.

[0136] Its core linkage effect, for example, when a user selects competitor brand A in the geospatial map view, the system triggers a global linkage: the time-series trend chart view automatically highlights all time-series data of brand A, the knowledge graph association view automatically highlights the associated nodes and relationship edges of brand A, and the statistical report view automatically filters and displays the summary data of brand A. This allows users to explore the relationship between the sales trend and channel layout of competitor brands in a multi-dimensional and in-depth manner, providing intuitive, accurate and interactive decision support for enterprise market decisions.

[0137] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0138] like Figure 5 The diagram shown is a hardware structure schematic of an electronic device according to the present invention, comprising: At least one processor 501; and, A memory 502 is communicatively connected to at least one of the processors 501; wherein, The memory 502 stores instructions that are executed by at least one of the processors to enable the at least one of the processors to perform the intelligence big data model question-and-answer method as described above.

[0139] Figure 5 Take a processor 501 as an example.

[0140] The electronic device may also include an input device 503 and a display device 504.

[0141] The processor 501, memory 502, input device 503 and display device 504 can be connected by a bus or other means. The figure shows an example of connection by bus.

[0142] The memory 502, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the intelligence big data model question-and-answer method in the embodiments of this application, for example, Figure 1 , Figure 2The method flow is shown. The processor 501 executes various functional applications and data processing by running non-volatile software programs, instructions, and modules stored in the memory 502, thereby realizing the intelligence big data model question-and-answer method in the above embodiments.

[0143] Memory 502 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the big data model question-and-answer method. Furthermore, memory 502 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, memory 502 may optionally include memory remotely located relative to processor 501, and these remote memories may be connected via a network to the apparatus performing the big data model question-and-answer method. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0144] Input device 503 can receive user clicks and generate signal inputs related to user settings and function control of the big data model question-and-answer method. Display device 504 may include display devices such as a display screen.

[0145] When one or more modules are stored in the memory 502, and are run by one or more processors 501, the big data model question-and-answer method described in any of the above method embodiments is executed.

[0146] This invention utilizes a large-scale intelligence model and an intelligence domain knowledge graph to obtain a structured intent description object for the query, and generates a visualization generation scheme specification. Based on this specification, a view of the retrieved data is generated. This invention achieves accurate problem analysis and intelligent classification based on structured intent in intelligence scenarios, as well as multi-dimensional, interconnected, and automatic visualization generation and presentation enhanced by an intelligence domain knowledge graph. Through structured intent parsing in intelligence scenarios, collaboration with an intelligence domain knowledge graph, and intelligent visualization view arrangement, this invention constructs a technical framework that systematically enhances the analytical capabilities of large-scale intelligence models. It effectively solves the problems of non-standardized output and superficial analysis in professional scenarios, and further meets the high standards of intelligence professional decision-making by generating standardized multimodal intelligence analysis results, possessing significant technological advancement and commercialization potential.

[0147] One embodiment of the present invention provides a storage medium that stores computer instructions, which, when executed by a computer, are used to perform all the steps of the intelligence big data model question-answering method described above.

[0148] In the context of this disclosure, a storage medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. The storage medium can be a machine-readable signal medium or a machine-readable storage medium. Optionally, the storage medium can be a non-transitory computer-readable storage medium, such as a ROM, random access memory (RAM), compact disc ROM (CD-ROM), magnetic tape, floppy disk, and optical data storage device.

[0149] One embodiment of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the intelligence big data model question-answering method as described above.

[0150] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A question-and-answer method for a large intelligence model, characterized in that, include: The system obtains a structured intent description object for the query by acquiring the user's query input via natural language and using the intelligence big data model and intelligence domain knowledge graph. The structured intent description object is input into the knowledge graph of the intelligence domain for contextual reasoning to determine the core entity relationship network involved in the structured intent description object. Based on the structured intent description object, a visualization generation scheme specification is generated. The visualization generation scheme specification includes at least the data patterns of multiple views and the linkage relationship between views. By combining the core entity relationship network, retrieval data about the structured intent description object is obtained from multiple sources, resulting in a basic fact dataset; According to the data pattern defined for each view in the visualization generation scheme specification, data is extracted from the basic fact dataset to generate a view-driven data package corresponding to each view. Based on the inter-view linkage relationship defined in the visualization generation scheme specification, linkage configuration information is injected into the data packages with linkage relationships. Based on the view-driven data package, a corresponding view is generated, and an event response mechanism between views is established based on the linkage configuration information in the data package.

2. The intelligence big data model question-answering method according to claim 1, characterized in that, The process of obtaining a user's query input via natural language, and then using a large intelligence model and an intelligence domain knowledge graph to obtain a structured intent description object for the query, includes: The system obtains queries input by users through natural language, performs semantic parsing on the queries using an intelligence big data model, extracts task type, time constraints, spatial constraints and a set of attention indicators from the queries, obtains a preliminary structured parsing including task type, time constraints, spatial constraints and a set of attention indicators, and identifies one or more entity mentions in the queries. The entity references are input into the knowledge graph of the intelligence domain, and entity linking based on vector similarity and context matching is performed to obtain a core entity set including one or more core entities; Based on the preliminary analysis results and the core entity set, a confidence score is given for each field to obtain a structured intent description object that includes all fields and the confidence score of each field. The fields include task type, time constraint, spatial constraint, core entity set, and attention indicator set.

3. The intelligence big data model question-answering method according to claim 2, characterized in that, The entity references are input into the knowledge graph of the intelligence domain, and entity linking is performed based on vector similarity and context matching to obtain a core entity set including one or more core entities, including: Calculate the semantic similarity score between each entity mention and each entity node in the intelligence domain knowledge graph. Based on the semantic similarity score, match each entity mention with the entity node in the intelligence domain knowledge graph with the highest similarity score. The semantic similarity score is calculated as follows: ,in, It is the first similarity weight coefficient. It is the second similarity weight coefficient. It is the embedding vector of the entity mentioned in the query. These are pre-trained embedding vectors of entity nodes in a knowledge graph within the intelligence domain. It is the context matching degree, and The calculation formula is: ,in, The first context weight coefficient, For the second context weight coefficient, For the third context weight coefficient, The semantic cosine similarity between entity mentions and the overall query context. To ensure the compatibility between entity type and query task type, The degree of overlap and fit between the spatiotemporal characteristics of the entity and the spatiotemporal constraints of the query.

4. The intelligence big data model question-answering method according to claim 2, characterized in that, The confidence score for each field includes: Cross-validation was performed on each field using the corresponding parsing method, and the confidence score for each field was calculated as follows: ,in Score the confidence level of the field. The first confidence level weighting coefficient. This is the second confidence level weighting coefficient. This is the third confidence level weighting coefficient. This is the fourth confidence level weighting coefficient. Generate output probabilities for fields in the intelligence big data model. To score the consistency of cross-validation of the fields, Scoring for logical rules in the intelligence field Score the semantic relevance of the field context.

5. The intelligence big data model question-answering method according to claim 1, characterized in that, The process involves inputting the structured intent description object into an intelligence domain knowledge graph for contextual reasoning to determine the core entity relationship network involved in the structured intent description object, and generating a visualization generation scheme specification based on the structured intent description object, including: The structured intent description object is input into an intelligence domain knowledge graph for contextual reasoning to determine the core entity relationship network involved in the structured intent description object; Get a combination of multiple candidate views; Calculate the total fit score between each candidate view combination and the structured intent description object; The candidate view combination with the highest overall fit score is selected as the optimal view combination, and a visualization generation scheme specification is generated based on the optimal view combination.

6. The intelligence big data model question-answering method according to claim 5, characterized in that, The calculation of the total fit score between each candidate view combination and the structured intent description object includes: The fit score for the i-th candidate view combination is calculated as follows: ， in, For the i-th candidate view combination, Let be the fit score of the i-th candidate view combination. The first fitness score is the weighting coefficient. This is the second fitness weighting coefficient. This is the third fitness weighting coefficient. This is the fourth fitness weighting coefficient. Let i be the task matching degree of the i-th candidate view combination. Let i be the entity coverage of the i-th candidate view combination. Let i be the metric support for the i-th candidate view combination. Let be the complexity of the i-th candidate view combination.

7. The intelligence big data model question-answering method according to claim 1, characterized in that, The process of combining the core entity relationship network to retrieve search data about the structured intent description object from multiple sources yields a basic fact dataset, including: The intelligence big data model is invoked to generate a general analytical narrative text based on the structured intent description object and the core entity relationship network; Based on the structured intent description object, query conditions are generated, and a dedicated data interface or computing engine corresponding to the data pattern in the visualization generation scheme specification is called to obtain the search data retrieved through the query conditions. Based on the core entity relationship network, entities, attributes and relationships related to the retrieved data are extracted from the knowledge graph of the intelligence domain to form a retrieval knowledge subgraph; The retrieved data is aligned, verified, and fused with the retrieved knowledge subgraph to form a basic fact dataset.

8. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that are executed by at least one of the processors to enable the at least one of the processors to perform the big data intelligence question-answering method as described in any one of claims 1 to 7.

9. A storage medium, characterized in that, The storage medium stores computer instructions, which, when executed by the computer, are used to perform all the steps of the intelligence big data model question-and-answer method as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the big data intelligence question-answering method as described in any one of claims 1 to 7.

Images