Expert aid decision-making system and method based on knowledge graph construction

Abstract

The invention discloses an expert auxiliary decision-making system and method constructed based on a knowledge graph. The system comprises a management module used for accessing a large model and a third-party tool; the intelligent interaction sub-module is used for receiving a service request input by a user and performing semantic analysis to obtain a service demand; the expert knowledge construction module is used for constructing a domain ontology model and knowledge assets; the task planning sub-module is used for generating a structured task execution flow table by calling pre-modeling knowledge in the domain ontology model, an online retrieval result and large model reasoning, and configuring a conflict detection and knowledge completion mechanism; and the business application module is used for sequentially calling corresponding tools, models or modules to execute each task according to the task execution flow table and outputting a business processing result. According to the method, transparency, interpretability and controllability of decision making can be assisted, automatic generation, scenarized and personalized services are provided, and efficient business knowledge processing of man-machine collaboration is realized.

Description

Technical Field

[0001] This application relates to the field of decision support and knowledge processing technology, specifically to an expert-assisted decision-making system and method based on knowledge graph construction. Background Technology

[0002] In today's digital age, with the explosive growth of data, the demand for efficient and intelligent decision support systems is becoming increasingly urgent across various sectors. Expert systems, as a technology capable of simulating the decision-making process of human experts, have demonstrated enormous potential in solving complex problems. They can help people quickly obtain accurate information and advice, making more informed decisions. Widely used in fields such as healthcare, finance, and manufacturing, they have played a crucial role in driving development and innovation across industries.

[0003] Traditional expert systems primarily employ symbolic representation and reasoning, heavily relying on domain knowledge and empirical rules. This requires a significant investment of time and effort to collect, organize, and encode domain knowledge, building knowledge bases and rule sets, resulting in extremely high costs. Furthermore, these rules are often fixed and unchanging, lacking flexibility and adaptability in complex and dynamically changing situations, easily leading to irrational decisions. Traditional expert systems are often developed for specific domains, and their versatility and transferability are poor when applied across domains due to the lack of universal knowledge representation and reasoning mechanisms. Currently, the emergence of large-scale models offers new ideas and methods to address the challenges that traditional expert systems struggle with, becoming a highly promising development path. Internationally, breakthroughs have been achieved in combining large-scale models with rule-based expert systems, with systems like PalantirAIP and IBM Watson demonstrating advantages in their respective application scenarios. Some domestic enterprises and research institutions have also begun exploring the combination of these two approaches, with companies like Wenxin Yiyan, Coze, and Dify making attempts in different areas.

[0004] However, existing methods of combining large models with expert systems still have some problems. On the one hand, the controllability and interpretability of knowledge are insufficient, making it difficult to guarantee the transparency and reliability of the decision-making process. On the other hand, the lack of effective modeling of business logic and data makes it impossible to provide automatically generated, scenario-based, and personalized services, failing to meet the needs of different users in different scenarios. Moreover, in the process of processing business knowledge, the efficiency of human-machine collaboration is low, making it difficult to fully integrate expert experience and generative AI capabilities. Summary of the Invention

[0005] To achieve transparency, interpretability, and controllability in decision support, and to provide automatically generated, scenario-based, and personalized services, thereby enabling efficient business knowledge processing through human-machine collaboration, this application provides an expert-assisted decision support system and method based on knowledge graph construction.

[0006] In a first aspect, this application provides an expert-assisted decision-making system based on knowledge graph construction, comprising: a management module, an expert knowledge construction module, an expert decision-making module, and a business application module, wherein the modules are interconnected; the expert decision-making module includes: an intelligent interaction submodule and a task planning submodule; The management module is used to access large models, third-party domain decision-making tools, and domain knowledge reading and writing tools; The intelligent interaction submodule is used to receive user input business requests and perform semantic analysis to obtain business requirements; The expert knowledge construction module is used to configure the processing nodes of business rules, business terms and business operators through visual drag and drop, and to access domain knowledge reading and writing tools to extract knowledge elements from the domain knowledge base and business document library to construct domain ontology models and knowledge assets. The task planning submodule is used to call pre-modeled knowledge in the domain ontology model based on business needs, and connect to the industry knowledge base and real-time business data platform through external interfaces to retrieve domain knowledge data related to business needs in real time; inject the called pre-modeled knowledge and real-time retrieved knowledge data into the large model inference to generate a structured task execution process table, and configure conflict detection and knowledge completion mechanisms; the conflict detection and knowledge completion mechanism includes performing knowledge completion and conflict detection on the generated task execution process table through the configured ontology-driven inference engine; The business application module is used to sequentially call the corresponding tools, models or modules to execute each task according to the task execution flow table, and output the business processing results.

[0007] By adopting the above scheme, an expert knowledge construction module is set up to support the dynamic construction of domain ontology models and knowledge assets, providing automatically generated, scenario-based, and personalized services to achieve the structuring and unified semantics of business knowledge; a task planning sub-module is set up to generate task execution flow tables by combining pre-modeled knowledge and real-time retrieved data, integrating expert experience and generative AI capabilities, and configuring conflict detection and knowledge completion mechanisms to achieve transparency, interpretability, and controllability in assisted decision-making, as well as efficient business knowledge processing through human-machine collaboration; combined with other modules, the various modules of the expert assisted decision-making system can work collaboratively to accurately understand user business needs, effectively integrate domain knowledge, and generate reasonable task execution flows.

[0008] Preferably, the expert decision-making module further includes: a verification and optimization submodule; The verification optimization submodule also includes: a result integration and multi-dimensional verification unit and a self-learning unit; The result integration and multi-dimensional verification unit is used to integrate the results according to a preset output structure to form a unified business processing result; and to perform multi-dimensional verification on the integrated business processing result; the multi-dimensional verification includes: illusion verification, factual verification and logical verification. The self-learning unit is used to adaptively learn and optimize the construction of the domain ontology model and adjust the parameters of the large model by collecting user feedback and system operation data.

[0009] By adopting the above scheme, the result integration and multi-dimensional verification unit is set up to integrate and verify the business processing results, ensuring the accuracy and reliability of the results; the self-learning unit is set up to adaptively learn and optimize based on user feedback and system operation data, thereby improving the system's performance and decision quality.

[0010] Preferably, the task planning submodule is further used to analyze business requirements and construct search keywords during the process of calling pre-modeling knowledge in the domain ontology model based on business requirements, perform structured search in the pre-modeling knowledge, and determine the execution framework and constraints of the task. The constraints include: domain scope constraints, knowledge source constraints, and output structure constraints. The task planning submodule is also used to generate extended search keywords by semantic expansion using a large model based on the semantic and ontological concepts of business needs during the real-time retrieval of domain knowledge data related to business needs. Based on the extended search keywords, the module performs online real-time retrieval of the industry knowledge base and the real-time business data platform to obtain domain knowledge data related to business needs. The task planning submodule is also used to construct a domain ontology-driven prompt template and limit the output format requirements; inject the pre-modeled knowledge and real-time retrieved knowledge data into the large model, and generate a structured task execution flow table that meets the format requirements by reasoning according to the constructed prompt template.

[0011] By adopting the above approach, the task planning submodule is set up to construct keywords based on business needs. When calling pre-modeled knowledge in the domain ontology model, the task framework and constraints can be accurately determined. Furthermore, the large model is used to semantically expand the business needs to generate extended search keywords, which can more comprehensively and accurately obtain relevant domain knowledge from the industry knowledge base and real-time business data platform. At the same time, the domain ontology is used to construct prompt templates to generate structured task execution process tables that meet the format requirements, thereby improving the accuracy and standardization of task planning.

[0012] Preferably, the task planning submodule is further used to inject the pre-modeled knowledge and real-time retrieved knowledge data into the large model, initiate a multi-layer planning strategy, and infer and generate a structured task execution process table; the multi-layer planning strategy includes: a strategic planning strategy based on the decomposition of large goals based on domain ontology, a tactical planning strategy based on the generation of candidate solutions based on the large model, and an execution planning strategy based on the execution process generated by constraints. The strategic planning strategy includes: defining the strategic objectives of the domain ontology, mapping business requirements to the strategic objectives in the domain ontology, formulating strategic objective decomposition rules, decomposing the strategic objectives into several sub-objectives, and obtaining the relationships between the sub-objectives; the tactical planning strategy includes: generating several tactical solutions using a large model for each sub-objective; each tactical solution includes a series of task groups; evaluating, screening, and analyzing task dependencies for each tactical solution to determine the selected tactical solution; the execution planning strategy includes: for the selected tactical solution, splitting the task groups into several sub-tasks using a large model, clarifying the execution order, dependencies, tools invoked, and expected results of each sub-task; integrating the sub-tasks corresponding to the sub-objectives according to the relationships between the sub-objectives, and outputting a task execution flow table.

[0013] By adopting the above approach, a task planning sub-module is set up in conjunction with a multi-layered planning strategy. When generating a structured task execution process table, business requirements are broken down and planned in a hierarchical and systematic manner, which improves the rationality, comprehensiveness, and adaptability of task planning to complex business, and ensures the efficiency and accuracy of task execution.

[0014] Preferably, the task planning submodule is further configured to establish a conflict classification mechanism based on the detected conflict results and match differentiated conflict handling strategies during the conflict detection process via the ontology-driven inference engine; the conflict classification mechanism divides conflict levels by predefining conflict types and conflict characteristics, and pre-sets conflict handling strategies to match different conflict levels. The task planning submodule is also used to transform the task execution plan into a simulation environment, define random events and uncertainties in the environment, simulate the task execution process using a large model, predict execution conflicts and assess the probability and impact of the conflicts, and generate corresponding conflict response strategies.

[0015] By adopting the above scheme, the task planning submodule establishes a conflict classification mechanism based on the detected conflict results and matches differentiated conflict handling strategies. Appropriate solutions are adopted for different levels of conflict, thereby improving the pertinence and effectiveness of conflict handling. At the same time, the task planning submodule uses a large model to simulate the task execution process, predict conflicts and assess their probability and impact, and take measures in advance to avoid or mitigate the harm of conflicts, thereby improving the stability and reliability of task execution.

[0016] Preferably, the result integration and multi-dimensional verification unit is further used to adaptively match different dimension verification weights according to the business scenario in the corresponding domain of the business processing result during the multi-dimensional comprehensive verification process for the integrated business processing result, and each business scenario in each domain is adaptively configured with different dimension verification weights that match it, and add the matching different dimension verification weights to complete the multi-dimensional verification comprehensive evaluation.

[0017] By adopting the above scheme, the result integration and multi-dimensional verification unit is set up to adaptively match different dimension verification weights according to the business scenario in the corresponding domain of the business processing result. This allows for a more accurate multi-dimensional comprehensive evaluation of the integrated business processing result, making the multi-dimensional verification more in line with the actual needs of different domains and improving the accuracy and effectiveness of the verification results.

[0018] Preferably, the intelligent interaction submodule is further configured to identify core business requirements based on a domain ontology model and a preset intent library during the process of acquiring business requirements, and to determine whether the core business requirements are clear. If the core business requirements are unclear, a multi-turn dialogue function is initiated to supplement the acquisition of business requirements. The preset intent library stores intent categories, intent expressions, entity slots, dialogue strategies, and intent context associations. The criteria for determining whether the core business requirements are clear include: whether the intent confidence level is less than the intent confidence level threshold and whether entity slot information is missing.

[0019] By adopting the above solution, the intelligent interaction submodule is set up to accurately identify business requirements based on the domain ontology model and the preset intent library, effectively determine whether the requirements are clear, and promptly initiate multi-round dialogues to supplement information when the requirements are unclear, thereby ensuring that the obtained business requirements are accurate and comprehensive and avoiding subsequent processing deviations caused by ambiguous requirements.

[0020] Secondly, this application provides an expert-assisted decision-making method based on knowledge graph construction, including: Access to large-scale models, third-party domain decision-making tools, and domain knowledge reading and writing tools; Receive user input business requests and perform semantic analysis to obtain business requirements; The processing nodes of business rules, business terms and business operators are configured through visual drag and drop, and knowledge elements are extracted from the domain knowledge base and business document library by connecting to the domain knowledge reading and writing tool to build the domain ontology model and knowledge assets. Based on business needs, pre-modeled knowledge in the domain ontology model is invoked, and industry knowledge bases and real-time business data platforms are connected through external interfaces to retrieve domain knowledge data related to business needs in real time. The invoked pre-modeled knowledge and real-time retrieved knowledge data are injected into the large model for reasoning to generate a structured task execution process table, and a conflict detection and knowledge completion mechanism is configured. The conflict detection and knowledge completion mechanism includes performing knowledge completion and conflict detection on the generated task execution process table through the configured ontology-driven reasoning engine. According to the task execution flow chart, the corresponding tools, models or modules are called in sequence to execute each task and output the business processing results.

[0021] By adopting the above solution, we can achieve end-to-end decision support from business requirements to output results, ensuring that the acquisition, integration and utilization of knowledge during business processing are controllable within the scope of domain knowledge. At the same time, we can supplement knowledge and detect conflicts in real time, thereby improving the accuracy, interpretability and traceability of decisions.

[0022] Thirdly, this application provides a computer-readable storage medium including a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to perform the method described above.

[0023] Fourthly, this application provides a computer device, the computer device including a memory, a processor and a program stored in the memory and executable thereon, the program being executed by the processor to implement the steps of the method described above.

[0024] In summary, this application has the following beneficial effects: 1. By connecting to large models and tools through the management module, the task planning submodule combines pre-modeled knowledge with real-time retrieved data to generate task execution flowcharts, achieving adaptive knowledge-controlled task planning and ensuring the transparency, interpretability, and controllability of the decision-making process; by constructing a domain ontology model through the expert knowledge construction module, knowledge completion and conflict detection are performed through an ontology-driven inference engine during task execution, realizing ontology-based business knowledge representation and retrieval, meeting the needs of automatic generation, scenario-based, and personalized services, and providing semantic support; by setting up collaborative work among various modules, the semantic understanding and language generation capabilities of the large model are utilized to assist business personnel in completing knowledge processing, realizing human-machine collaborative business knowledge processing based on the large model; 2. By setting up a verification and optimization submodule, the system can execute tasks according to the process table and output business processing results. It can also perform result integration and verification, self-learning optimization, improve the model's ability to understand, decompose and schedule tasks, realize the semantic unity and structural expression of knowledge, ensure the quality and controllable evolution of processed content, and realize the continuous evolution of system capabilities. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the structure of the expert-assisted decision-making system based on knowledge graphs as described in a specific embodiment; Figure 2 This is a schematic diagram of the principle architecture of the expert-assisted decision-making system based on knowledge graphs as described in a specific embodiment; Figure 3 This is a functional structure diagram of the expert-assisted decision-making system based on knowledge graphs as described in a specific embodiment; Figure 4 This is a flowchart of the expert-assisted decision-making method based on knowledge graph construction as described in a specific embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] like Figure 1 As shown in the figure, this application discloses an expert-assisted decision-making system based on knowledge graph construction, specifically including: a management module 100, an expert knowledge construction module 200, an expert decision-making module 300, and a business application module 400. The modules interact and call functions through preset internal interfaces, forming a whole to jointly complete complex business-related auxiliary decision-making tasks. The functional settings of each module will be described in detail below.

[0028] First, the management module 100 is used to access large models, third-party domain decision-making tools, and domain knowledge reading and writing tools. Specifically, the management module 100 includes: a large model access unit 110, a third-party domain decision-making tool access unit 120, and a domain knowledge reading and writing tool access unit 130.

[0029] like Figure 2As shown, the large model access unit 110 is used to import large model services. By filling in the large model name and calling address, the large model is registered and accessed, enabling various modules to utilize the large model for intelligent reasoning and content generation. The large model can be set to a general-purpose large model from the GPT series, BERT, etc., or it can be a specialized large model trained for specific domains. The third-party domain decision-making tool access unit 120 is responsible for accessing third-party domain decision-making tools, including machine learning models, rule models, service interfaces, workflows, etc., and performing connectivity tests while configuring tool parameters. The domain knowledge reading and writing tool access unit 130 is used to access domain knowledge reading and writing tools, including knowledge base reading, file reading, and processing result writing tools, and also performs connectivity tests and parameter configurations. Furthermore, the management module 100 is also used to set parameters for the accessed large models, such as adjusting model inference parameters according to different business scenario requirements, and providing editing, modification, and deletion functions for the accessed tools to meet the needs of different business scenarios.

[0030] Second, the expert knowledge construction module 200 is used to configure the processing flow nodes of business rules, business terms, and business operators through a visual drag-and-drop method, and to extract knowledge elements from the domain knowledge base and business document library by connecting to domain knowledge reading and writing tools to construct a domain ontology model and knowledge assets. Specifically, the expert knowledge construction module 200 includes: a processing flow node configuration unit 210 and a knowledge element extraction unit 220.

[0031] like Figure 2 As shown, the processing flow node configuration unit 210 is used to configure processing flow nodes of business rules, business terms, business operators and business logic in a visual drag-and-drop manner based on the knowledge processing orchestration management function; these nodes include start, end, conditional branch, code node, variable node, etc.

[0032] The knowledge element extraction unit 220 is used to access domain knowledge reading and writing tools, extract knowledge elements such as entities, attributes, relationships, and rules from the domain knowledge base and business document library, automatically generate structured knowledge drafts (knowledge assets), and integrate them to form a domain ontology model with unified semantics. Alternatively, it can use the accessed large model to automatically complete the extraction of core domain concepts, sort out the hierarchical relationship of concepts, define attributes, and output an initial ontology framework; then, combined with OWL / RDF ontology syntax rules, the large model verifies the logical consistency of the initial ontology, and finally, after manual review by experts, construct the domain ontology model.

[0033] In addition, the expert knowledge construction module 200 is also used to support business personnel in classifying, adjusting, reviewing and confirming structured knowledge drafts (knowledge assets), recording revision and feedback information, optimizing subsequent parsing strategies, forming a domain ontology model with unified semantics, and storing it in the knowledge base.

[0034] Third, the expert decision-making module 300 specifically includes: an intelligent interaction submodule 310 and a task planning submodule 320.

[0035] Firstly, the intelligent interaction submodule 310 is used to receive user input business requests and perform semantic analysis to obtain business requirements. Specifically, the intelligent interaction submodule 310 includes a business request receiving unit 311 and a semantic analysis unit 312.

[0036] The business request receiving unit 311 is used to receive business requests input by users through web browsers or other means; wherein, business requests include various types such as business Q&A, in-depth research, and report preparation.

[0037] The semantic analysis unit 312 is used to perform semantic analysis on the received business requests, identifying business objectives (e.g., obtaining question-and-answer answers, generating research reports, formulating solutions, etc.), the type of domain (e.g., equipment, high-end manufacturing processes, etc.), and business intents. Specifically, through the intent understanding function, it identifies business requirements, including core business requirements, based on the domain ontology model and a preset intent library; based on the identified core business requirements, it determines whether the core business requirements are clear; if the core business requirements are unclear, it initiates a multi-turn dialogue function, and based on the continuity of the context topic, it asks supplementary questions to the user until the core business requirements are clear.

[0038] The preset intent library stores: intent categories (such as "disease diagnosis", "treatment plan query", "drug information", "symptom consultation" etc.), intent expressions (each intent corresponds to a series of natural language expressions (templates), which describe the typical way users express the intent, such as "What should I do if I have a headache?" and "What disease causes frequent coughing?" for the intent "disease diagnosis"), entity slots (each intent needs to be filled with some key information (slots), which correspond to concepts in the domain ontology, such as "symptoms, duration, age, etc." in the intent "disease diagnosis"), dialogue strategies (for each intent, the next action or question that the system may take is preset to guide the user to complete the request, such as "if the user provides symptoms, the duration of symptoms will be asked"), and intent context association information (the relationship between intents to handle the switching or continuation of intents in multi-turn dialogues).

[0039] Since the domain ontology model defines the concepts, attributes, relationships, and constraints within the domain, it provides a semantic foundation for identifying core business needs. Core business needs typically correspond to key tasks or problems within the corresponding domain; for example, in the medical field, core needs might include disease diagnosis, treatment recommendations, and drug searches. Therefore, the process of identifying business needs through intent understanding, based on the domain ontology model and a pre-defined intent library, specifically includes: identifying core needs by matching user input with the domain ontology and the pre-defined intent library; matching the parsed intent with the intents in the pre-defined intent library based on the entities (domain type) and intents identified from the user input to determine the intent category; and extracting relevant information from the user input according to the corresponding domain ontology and filling it into the entity slots of the intent to determine the final core business needs.

[0040] The criteria for determining whether the core business requirement is clear include: whether the intent confidence level is less than the intent confidence level threshold, whether entity slot information is missing, and whether there is confusion due to multiple intents. Specifically, the intent confidence level is calculated through a classification model or rule matching. If the intent confidence level is lower than the intent confidence level threshold, it indicates that the user's intent cannot be determined, and the core business requirement is considered unclear. Similarly, if entity slot information is missing, it indicates that the core business requirement cannot be accurately determined, and the core business requirement is considered unclear. Furthermore, if there is confusion due to multiple intents, such as a user asking about both symptoms and medication information without clearly defining the focus, the core business requirement is considered unclear.

[0041] Secondly, the task planning submodule 320 is used to generate a structured task execution flow table based on business needs by calling pre-modeling knowledge, online retrieval results, and large model reasoning, and to configure conflict detection and knowledge completion mechanisms. Specifically, the task planning submodule of the expert decision-making module includes a pre-modeling knowledge calling unit 321, a real-time knowledge retrieval unit 322, a task execution flow table generation unit 323, and a conflict detection and knowledge completion unit 324.

[0042] The pre-modeling knowledge invocation unit 321 is used to parse business requirements and construct search keywords. It performs structured retrieval within the pre-modeling knowledge of the domain ontology model to determine the task's execution framework and constraints, including domain scope constraints, knowledge source constraints, and output structure constraints. Specifically, based on business requirements, the domain ontology is locked, and domain concepts, entities, attributes, and relationships are extracted from the business requirements by the ontology inference engine to serve as query instructions for subsequent bidirectional enhanced retrieval. Using the domain concepts and relationships parsed from the business requirements as search keywords, a structured retrieval is performed within the pre-modeling knowledge to obtain the task's basic execution framework (e.g., standard process) and constraints. Furthermore, the pre-modeling knowledge invocation unit 321 is also used to split the core rule ontology (e.g., general decision-making process) and the scenario-based ontology (e.g., differentiated rules for emergency / routine projects), dynamically invoking them according to the corresponding scenario of the business requirements.

[0043] The real-time knowledge retrieval unit 322 is used to connect to industry knowledge bases, real-time business data platforms, and topic knowledge graphs through external interfaces to retrieve domain knowledge data related to business needs in real time. Specifically, it uses the domain concepts and relationships parsed from business needs as search keywords to perform searches in the connected industry knowledge base, real-time business data platform, and topic knowledge graph (referred to as forward search), and filters, sorts, and structures the search results. In addition, to further ensure that knowledge fragments with "high relevance and high timeliness" to business needs are retrieved first, the real-time knowledge retrieval unit 322 is also used to perform semantic expansion (such as synonym, near-synonym, and related concept expansion) based on the semantics and ontology concepts of business needs using a large model to generate extended search keywords. Based on the extended search keywords, it performs online real-time searches in the industry knowledge base and real-time business data platform (referred to as reverse search) to obtain domain knowledge data related to business needs. Therefore, the static knowledge retrieved from the ontology's forward retrieval and the dynamic knowledge retrieved from the large model's reverse retrieval are integrated. The large model completes knowledge deduplication and relevance ranking, forming a high-quality knowledge material library according to the principle of prioritizing authoritative ontology knowledge and supplementing it with real-time dynamic knowledge.

[0044] The task execution flow table generation unit 323 is used to inject the invoked pre-modeling knowledge and real-time retrieved knowledge data into the large model inference to generate a structured task execution flow table. Specifically, the task planning submodule utilizes the large model to generate a task planning scheme based on the input invoked pre-modeling knowledge and real-time retrieved knowledge data, breaking down complex business tasks into N sub-tasks (N≥1), clarifying the execution order, dependencies, invoked tools, and expected results of each sub-task; and supporting manual intervention and modification of the generated task planning scheme by business personnel, including adjusting the order of sub-tasks, adding or removing sub-tasks, and modifying constraints. The system records modification traces for subsequent optimization. After the task planning scheme is confirmed, a structured task execution flow table is generated, containing information such as sub-task identifier, execution priority, input parameters, invoked tools / models, and output format.

[0045] Furthermore, to further ensure the accuracy and standardization of the generated structured task execution flowchart, a domain-standardized prompt template is constructed using the domain ontology as the core framework and combined with fused knowledge materials. The task execution flowchart generation unit 323 is also used to construct a domain ontology-driven prompt template based on knowledge about task planning in the domain ontology (such as dependencies between tasks, conditions for task execution, etc.), and to define output format requirements. Specifically, the prompt template strictly follows the structural specifications of the domain ontology, including: the task planning objective (the defined execution framework of the task), relevant domain knowledge (invoked pre-modeling knowledge and real-time retrieved knowledge data), task planning conditions (defined constraints), and output format requirements (such as requiring the generation of a table including step numbers, task descriptions, execution conditions, required resources, estimated time, etc.). The task planning submodule 320 is also used to inject the invoked pre-modeling knowledge and real-time retrieved knowledge data into the large model, and to generate a structured task execution flowchart that meets the format requirements by reasoning according to the constructed prompt template. Specifically, the pre-modeled knowledge and real-time retrieved knowledge data are precisely injected into the prompt template as context to form complete and directly executable prompt words. The large model reads the ontology constraints, objectives, and knowledge in the prompt words, and performs inference calculations within the limited semantic and format range to generate a structured task execution flow table that meets the format requirements.

[0046] The conflict detection and knowledge completion unit 324 is equipped with a conflict detection and knowledge completion mechanism, specifically configured with an ontology-driven inference engine. For the generated task execution flow table, the ontology-driven inference engine performs knowledge completion and conflict detection (e.g., checking for time conflicts, resource conflicts, or compliance with business rules between subtasks). Specifically, based on the relational rules of the domain ontology, it automatically infers and completes missing key attributes or related information; it detects attribute conflicts, rule exclusivity, and temporal logic conflicts of the same entity. If a conflict is found, the conflict information is fed back to the larger model, requiring the task execution flow table to be regenerated or adjusted.

[0047] Furthermore, to ensure smooth task execution and complete the final verification of the generated task execution flow table, the task planning submodule 320 is also used to establish a conflict grading mechanism based on the detected conflict results during the conflict detection process using the ontology-driven inference engine, and to match differentiated conflict handling strategies. Specifically, the conflict grading mechanism predefines conflict types (e.g., fatal conflict / minor conflict) and conflict characteristics (violation of compliance requirements / local node deviation, small-scale knowledge gap), classifies conflict levels (first level / second level), and pre-sets matching conflict handling strategies for different conflict levels (the first conflict handling strategy matching the first level: freeze the current task execution flow table, prohibit execution and immediately issue an early warning, triggering replanning; the second conflict handling strategy matching the second level: use the ontology-driven inference engine to complete local node correction, and use the large model to supplement missing knowledge through retrieval).

[0048] Furthermore, to better ensure task execution, the task planning submodule 320 is also used to transform the task execution plan into a simulation environment, including task status, resource status, and timeline, define random events and uncertainties in the environment (such as task delays and resource failures), simulate the task execution process using a large model, predict execution conflicts, assess the probability and impact of conflicts, and generate corresponding conflict response strategies. The large model makes predictions based on historical data (existence of task execution conflicts, probability of occurrence of task execution conflicts, and impact) and domain knowledge. The simulation results are analyzed to identify conflict types, including time conflicts, resource conflicts, and logical conflicts, and corresponding conflict response strategies are generated, such as increasing resource buffers, adjusting task order, and preparing backup plans.

[0049] Fourth, the business application module 400 is used to sequentially call the corresponding tools, models or modules to execute each task according to the task execution flow table, and output the business processing results.

[0050] like Figure 3As shown, the system executes each subtask sequentially by calling the corresponding tools, models, or modules according to the task execution flowchart. For example: if the subtask is business Q&A, the system calls the business Q&A assistant from the business application module, selects either the knowledge Q&A assistant or the SOP Q&A assistant based on intent recognition results, and generates a mixed text and image answer based on domain knowledge and the large model's thought process chain. If the subtask is in-depth research, the system calls the in-depth research assistant from the business application module, autonomously plans the research path, retrieves and integrates domain knowledge in multiple steps, and generates intermediate research materials and reasoning basis. If the subtask is report compilation, the system calls the report compilation assistant from the business application module, generates multimodal report content including formats such as text, images, and tables based on outline templates and domain data, and optimizes fluency and chapter coherence. If the subtask is complex calculation or professional analysis, a large model-small model mode is adopted, with the large model responsible for parameter extraction and format conversion, and small models such as professional calculation models and performance indicator models responsible for specific calculations and result output. Furthermore, during subtask execution, real-time knowledge completion and conflict detection are also supported through an ontology-driven inference engine.

[0051] In a specific embodiment, to ensure the accuracy and reliability of business processing results and further improve the system's performance and decision quality, the system further includes: the expert decision module 300 further includes: a verification and optimization submodule 330, and the verification and optimization submodule 330 further includes: a result integration and multi-dimensional verification unit 331 and a self-learning unit 332.

[0052] The result integration and multi-dimensional verification unit 331 is used to integrate results according to a preset output structure to form a unified business processing result, such as question-and-answer answers, research reports, and handling plans; and to perform multi-dimensional verification on the integrated business processing result. The multi-dimensional verification includes: illusion verification, fact verification, and logic verification. Specifically, illusion verification includes comparing with the domain knowledge base to ensure that there is no false information in the result, and setting the illusion verification accuracy rate to be no less than 85%, with a matching evaluation score set for the preset illusion verification accuracy rate range; fact verification includes verifying that the data and conclusions in the result are consistent with authoritative sources, and setting the fact verification accuracy rate to be no less than 80%, with a matching evaluation score set for the preset fact verification accuracy rate range; logic verification includes checking whether the reasoning logic of the result is rigorous and without logical contradictions, with a logic verification accuracy rate of no less than 80%, with a matching evaluation score set for the preset logic verification accuracy rate range. Specifically, a large model can be used for verification, and multi-dimensional verification evaluation scores are output.

[0053] Furthermore, this system supports manual review and optimization of the verified results by business personnel (some standards for manual review are shown in Table 1). The system records review comments and feeds them back to the expert knowledge construction module 200 and the expert decision-making module 300 to optimize the domain ontology model and task execution flowchart generation strategy. The system also provides result output and download functions, supporting export in formats such as PDF and Markdown. Simultaneously, it records the reasoning path, knowledge source, and tools used during the result generation process, ensuring traceability of the decision-making process. Table 1 below shows the requirements settings for some task generation and execution processes and verification processes.

[0054] Table 1

[0055] Furthermore, to better adapt the system to different business scenarios and user needs, the result integration and multi-dimensional verification unit 331 is also used to adaptively match different dimension verification weights according to the business scenario in the corresponding domain of the integrated business processing result during the multi-dimensional comprehensive verification process. Each business scenario in each domain is adaptively configured with different dimension verification weights to complete the multi-dimensional verification comprehensive evaluation, i.e., weighted calculation to obtain the final multi-dimensional verification score, and compare it with a preset verification score threshold to determine whether the verification passes. If the verification fails, a prompt message indicating that the current business processing result is abnormal is generated. For example, in the financial field, the matching weights for high-risk emergency scenarios are: logic verification 40%, fact verification 40%, and illusion verification 20%; in the financial field, the matching weights for low-risk transaction scenarios are: fact verification 35%, logic verification 30%, and illusion verification 35%.

[0056] The self-learning unit 332 is used to adaptively learn and optimize the construction of the domain ontology model, adjust the parameters of the large model, and optimize the knowledge retrieval algorithm by collecting user feedback and system operation data. Specifically, the adaptive learning and optimization of the domain ontology model construction includes: collecting the associated content of the initial domain ontology model from the operation data of the system application's initial domain ontology model and user feedback data; using machine learning technology, automatically learning the domain ontology update pattern from the corresponding operation data and feedback data; and completing the dynamic update and version management of the ontology.

[0057] Similarly, by collecting operational data and user feedback data related to the task execution process tables generated by the system's large model, and utilizing machine learning techniques, the large model's update parameters are learned from the corresponding operational and feedback data to finalize the large model parameters. Furthermore, adjusting the large model parameters can also be used to optimize knowledge retrieval. This involves collecting semantic expansion data from the system's large model to generate extended search keywords. Based on these extended search keywords, the large model performs semantic expansion on the operational data and user feedback data related to online real-time retrieval of industry knowledge bases and real-time business data platforms. Using machine learning techniques, the large model's update parameters are learned from the corresponding operational and feedback data to finalize the large model parameters, thereby optimizing the knowledge retrieval algorithm.

[0058] In a specific embodiment, to improve the rationality and comprehensiveness of task planning and its adaptability to complex business processes, and to ensure the efficiency and accuracy of task execution, a multi-layered planning strategy is designed to decompose and plan business requirements in a hierarchical and systematic manner; the system also includes: The task planning submodule 320 is further used to inject the pre-modeled knowledge and real-time retrieved knowledge data into the large model, initiate a multi-layer planning strategy, and infer and generate a structured task execution flow table. The multi-layer planning strategy includes: a strategic planning strategy based on the decomposition of large goals based on the domain ontology, a tactical planning strategy based on the generation of candidate solutions based on the large model, and an execution planning strategy based on the constraint-based execution flow generation.

[0059] The strategic planning approach involves using domain ontology to decompose business needs at a high level, resulting in several strategic directions or objectives. For example, if the business need is "to develop a treatment plan for a patient with hypertension," the corresponding strategic planning strategies would include: Strategic Objective 1: Assess the severity of the patient's current condition; Strategic Objective 2: Develop a drug treatment plan; and Strategic Objective 3: Develop a non-drug treatment plan.

[0060] Specifically, the strategic planning strategy implementation steps include: defining the strategic objectives of the domain ontology (e.g., based on the constructed domain ontology, clarifying core concepts and relationships such as business objectives, constraints, and resources, and defining strategic objectives 1-5); mapping business requirements to the strategic objectives in the domain ontology (strategic objectives 1-3); utilizing the relationships in the ontology to formulate strategic objective decomposition rules, decomposing the strategic objectives into several sub-objectives and obtaining the relationships between several sub-objectives; displaying the decomposed sub-objectives and their relationships (such as dependencies, temporal relationships, etc.) in a diagrammatic form; in addition, a large model can be used to parse the strategic objectives described in natural language, generate a preliminary decomposition structure, and align and correct the output of the large model with the decomposition rules in the domain ontology.

[0061] The tactical planning strategy involves generating multiple possible tactical options for each strategic objective using a large model. For example, for strategic objective 2, possible tactical options include: Option A: using single-drug treatment; Option B: using combination drug treatment.

[0062] Specifically, the tactical planning strategy includes: generating several tactical plans for each sub-objective using a large model, which can be driven by historical cases; each tactical plan includes a series of task groups and parameters such as expected benefits, main risks, and resource requirements, used to complete a sub-objective; establishing an evaluation model based on expected benefits, main risks, and resource requirements, evaluating each tactical plan, selecting the top N tactical plans based on the scoring results, and analyzing the dependencies between tasks in each selected tactical plan based on task relationship knowledge in the ontology to determine the selected tactical plan, and providing a detailed description of the selected tactical plan, including the task list, resource allocation, and time estimation.

[0063] The execution planning strategy, for the selected tactical plan, utilizes pre-modeling knowledge and real-time retrieved knowledge to generate a specific task execution flow table through large model reasoning.

[0064] Specifically, the execution planning strategy includes: for the selected tactical plan, the task group is divided into several sub-tasks through a large model, and the execution order, dependencies, tools called and expected results of each sub-task are clarified; the sub-tasks corresponding to the sub-goals are integrated according to the relationship between the sub-goals, and the task execution flow table is output.

[0065] like Figure 4 As shown, this application provides an expert-assisted decision-making method based on knowledge graph construction that applies the above-mentioned system. The specific steps include: S1. Access to large models, third-party domain decision-making tools, and domain knowledge reading and writing tools.

[0066] Specifically, using the management module, following the method described in the previous embodiments, import the large model service, fill in the large model name and calling address to complete the registration and access, access the third-party domain decision-making tool and domain knowledge reading and writing tool, and perform connectivity testing and parameter configuration.

[0067] S2. Receive user input business requests and perform semantic analysis to obtain business requirements.

[0068] Specifically, the intelligent interaction submodule receives business requests input by users, performs semantic analysis on the business requests using the domain ontology model and preset intent library, determines whether the core business requirements are clear, and if not, initiates a multi-turn dialogue function to supplement and obtain the business requirements.

[0069] S3. Configure business rules, business terms, and business operator processing nodes through visual drag-and-drop, and access domain knowledge reading and writing tools to extract knowledge elements from the domain knowledge base and business document library to build domain ontology models and knowledge assets.

[0070] Specifically, the expert knowledge building module is used to configure the processing flow nodes through a visual drag-and-drop method, and the domain knowledge reading and writing tools are connected to extract knowledge elements to form a domain ontology model and store it in the knowledge base.

[0071] S4. Based on business needs, call the pre-modeled knowledge in the domain ontology model, and connect to the industry knowledge base and real-time business data platform through external interfaces to retrieve domain knowledge data related to business needs in real time; inject the called pre-modeled knowledge and real-time retrieved knowledge data into the large model reasoning to generate a structured task execution process table, and configure conflict detection and knowledge completion mechanisms.

[0072] S5. According to the task execution flow table, call the corresponding tools, models or modules in sequence to execute each task and output the business processing results.

[0073] In addition, the method also includes: S6. Integrate according to the preset output structure to form a unified business processing result; perform multi-dimensional verification on the integrated business processing result and output the verification result.

[0074] This application also discloses a computer-readable storage medium.

[0075] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the expert-assisted decision-making method based on knowledge graph construction described above. The computer-readable storage medium includes, for example, various media that can store program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0076] This application also discloses a computer device.

[0077] Specifically, the computer device includes a memory and a processor, and the memory stores computer programs that can be loaded and executed by the processor to implement the aforementioned expert-assisted decision-making method based on knowledge graph construction.

[0078] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. An expert-assisted decision-making system based on knowledge graph construction, characterized in that, include: The system includes a management module, an expert knowledge construction module, an expert decision-making module, and a business application module, with all modules interconnected. The expert decision-making module includes: an intelligent interaction submodule and a task planning submodule; The management module is used to access large models, third-party domain decision-making tools, and domain knowledge reading and writing tools; The intelligent interaction submodule is used to receive user input business requests and perform semantic analysis to obtain business requirements; The expert knowledge construction module is used to configure the processing nodes of business rules, business terms and business operators through visual drag and drop, and to access domain knowledge reading and writing tools to extract knowledge elements from the domain knowledge base and business document library to construct domain ontology models and knowledge assets. The task planning submodule is used to call pre-modeled knowledge in the domain ontology model based on business needs, and connect to the industry knowledge base and real-time business data platform through external interfaces to retrieve domain knowledge data related to business needs in real time; inject the called pre-modeled knowledge and real-time retrieved knowledge data into the large model inference to generate a structured task execution process table, and configure conflict detection and knowledge completion mechanisms; the conflict detection and knowledge completion mechanism includes performing knowledge completion and conflict detection on the generated task execution process table through the configured ontology-driven inference engine; The business application module is used to sequentially call the corresponding tools, models or modules to execute each task according to the task execution flow table, and output the business processing results.

2. The expert-assisted decision-making system based on knowledge graph construction according to claim 1, characterized in that, The expert decision-making module also includes: a verification and optimization submodule; The verification optimization submodule also includes: a result integration and multi-dimensional verification unit and a self-learning unit; The result integration and multi-dimensional verification unit is used to integrate the results according to a preset output structure to form a unified business processing result; and to perform multi-dimensional verification on the integrated business processing result; the multi-dimensional verification includes: illusion verification, factual verification and logical verification. The self-learning unit is used to adaptively learn and optimize the construction of the domain ontology model and adjust the parameters of the large model by collecting user feedback and system operation data.

3. The expert-assisted decision-making system based on knowledge graph construction according to claim 1, characterized in that, The task planning submodule is also used to analyze business requirements and construct search keywords during the process of calling pre-modeling knowledge in the domain ontology model based on business requirements, perform structured retrieval in the pre-modeling knowledge, and determine the execution framework and constraints of the task. The constraints include: domain scope constraints, knowledge source constraints, and output structure constraints. The task planning submodule is also used to generate extended search keywords by semantic expansion using a large model based on the semantic and ontological concepts of business needs during the real-time retrieval of domain knowledge data related to business needs. Based on the extended search keywords, the module performs online real-time retrieval in the industry knowledge base and the real-time business data platform to obtain domain knowledge data related to business needs. The task planning submodule is also used to construct a domain ontology-driven prompt template and limit the output format requirements; inject the pre-modeled knowledge and real-time retrieved knowledge data into the large model, and generate a structured task execution flow table that meets the format requirements by reasoning according to the constructed prompt template.

4. The expert-assisted decision-making system based on knowledge graph construction according to claim 3, characterized in that, The task planning submodule is also used to inject the pre-modeling knowledge and real-time retrieved knowledge data into the large model, start a multi-level planning strategy, and infer to generate a structured task execution process table. The multi-layered planning strategy includes: a strategic planning strategy based on the decomposition of large goals based on domain ontology, a tactical planning strategy based on the generation of candidate solutions based on large models, and an execution planning strategy based on the generation of execution processes based on constraints. The strategic planning strategy includes: defining the strategic objectives of the domain ontology, mapping business requirements to the strategic objectives in the domain ontology, formulating strategic objective decomposition rules, decomposing the strategic objectives into several sub-objectives, and obtaining the relationships between the sub-objectives; the tactical planning strategy includes: generating several tactical solutions using a large model for each sub-objective; each tactical solution includes a series of task groups; evaluating, screening, and analyzing task dependencies for each tactical solution to determine the selected tactical solution; the execution planning strategy includes: for the selected tactical solution, splitting the task groups into several sub-tasks using a large model, clarifying the execution order, dependencies, tools invoked, and expected results of each sub-task; integrating the sub-tasks corresponding to the sub-objectives according to the relationships between the sub-objectives, and outputting a task execution flow table.

5. The expert-assisted decision-making system based on knowledge graph construction according to claim 1, characterized in that, The task planning submodule is also used to establish a conflict classification mechanism based on the detected conflict results during the conflict detection process through the ontology-driven inference engine, and to match differentiated conflict handling strategies. The conflict classification mechanism divides conflict levels by predefining conflict types and conflict characteristics, and pre-sets conflict handling strategies to match different conflict levels. The task planning submodule is also used to transform the task execution plan into a simulation environment, define random events and uncertainties in the environment, simulate the task execution process using a large model, predict execution conflicts and assess the probability and impact of the conflicts, and generate corresponding conflict response strategies.

6. The expert-assisted decision-making system based on knowledge graph construction according to claim 2, characterized in that, The result integration and multi-dimensional verification unit is also used to adaptively match different dimension verification weights according to the business scenario in the corresponding domain of the business processing result during the multi-dimensional comprehensive verification process for the integrated business processing result. Each business scenario in each domain is adaptively configured with different dimension verification weights that match it. The matching different dimension verification weights are added to complete the multi-dimensional verification comprehensive evaluation.

7. The expert-assisted decision-making system based on knowledge graph construction according to claim 1, characterized in that, The intelligent interaction submodule is also used to identify core business requirements based on the domain ontology model and the preset intent library during the process of acquiring business requirements, and to determine whether the core business requirements are clear. If the core requirements are unclear, a multi-turn dialogue function is initiated to supplement the acquisition of business requirements. The preset intent library stores intent categories, intent expressions, entity slots, dialogue strategies, and intent context associations. The criteria for determining whether the core business requirements are clear include: whether the intent confidence level is less than the intent confidence level threshold and whether entity slot information is missing.

8. An expert-assisted decision-making method based on knowledge graph construction, applying the system described in any one of claims 1 to 7, characterized in that, include: Access to large-scale models, third-party domain decision-making tools, and domain knowledge reading and writing tools; Receive user input business requests and perform semantic analysis to obtain business requirements; The processing nodes of business rules, business terms and business operators are configured through visual drag and drop, and knowledge elements are extracted from the domain knowledge base and business document library by connecting to the domain knowledge reading and writing tool to build the domain ontology model and knowledge assets. Based on business needs, pre-modeled knowledge in the domain ontology model is invoked, and industry knowledge bases and real-time business data platforms are connected through external interfaces to retrieve domain knowledge data related to business needs in real time. The invoked pre-modeled knowledge and real-time retrieved knowledge data are injected into the large model for reasoning to generate a structured task execution process table, and a conflict detection and knowledge completion mechanism is configured. The conflict detection and knowledge completion mechanism includes performing knowledge completion and conflict detection on the generated task execution process table through the configured ontology-driven reasoning engine. According to the task execution flow chart, the corresponding tools, models or modules are called in sequence to execute each task and output the business processing results.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the method as described in any one of claims 1 to 7.

10. A computer device, characterized in that, The computer device includes a memory, a processor, and a program stored in and executable on the memory, the program being executed by the processor to implement the steps of the method as described in any one of claims 1 to 7.

Images