Question answering method and system for fusing dynamic intention recognition and GraphRAG for open domain

Abstract

The invention discloses a dynamic intention recognition and GraphRAG fusion method and system for open domain questions and answers, and relates to the technical field of intelligent and natural language processing. The method comprises the following steps: constructing and dynamically updating a heterogeneous knowledge graph; performing dual-stage intention recognition on user query; subtasks are disassembled based on an intention result, and cooperative execution is carried out through multiple agents; service standardization integration is realized by means of a model context protocol; and closed-loop optimization is carried out in combination with user feedback. The system comprises an image knowledge base, an intention recognition module, an agent collaboration module, an MCP service integration module and a feedback optimization module. The system is a novel open domain question-answering system integrating structured knowledge modeling, dynamic intention understanding, intelligent collaborative reasoning and standardized service interaction, and through fusion of structured knowledge modeling and intelligent collaborative reasoning, question-answering accuracy and complex intention understanding ability are improved, system adaptability and efficiency are enhanced, multi-system collaboration is promoted, and the system has a wide application prospect. The method is suitable for scenes such as knowledge services and intelligent assistants.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, specifically to a question-answering method and system for open-domain fusion dynamic intent recognition and GraphRAG. Background Technology

[0002] With the rapid development of artificial intelligence and natural language processing technologies, open-domain question answering systems are increasingly widely used in knowledge services, intelligent assistants, and enterprise decision support. However, traditional question answering systems still have significant limitations when facing complex user intent and multi-hop reasoning requirements. For example, mainstream retrieval augmented generation (RAG) frameworks rely on vector similarity matching to retrieve information from flattened text blocks, making it difficult to capture semantic connections and structured knowledge relationships between documents. This results in poor performance in relationship-intensive scenarios such as medical diagnosis and legal case analysis. At the same time, limited by the context length of large language models, the system can often only piece together limited text fragments, which can easily lead to the loss of key information and affect the quality of the answers.

[0003] To address these shortcomings, related research has attempted to introduce graph neural networks (GNNs) and heterogeneous knowledge graphs to enhance reasoning capabilities. However, existing solutions still have significant limitations: on the one hand, the construction of heterogeneous knowledge graphs relies on manual annotation, making updates and expansion difficult and unable to adapt to dynamic knowledge changes in open-domain environments; on the other hand, reasoning paths are mostly predefined, lacking flexibility and unable to handle diverse user question logic. Furthermore, existing systems have low levels of automation and intelligence, relying on manual intervention in knowledge updates and answer verification, resulting in low response efficiency and poor result consistency; their multi-dimensional information processing capabilities are weak, failing to effectively integrate features such as structured data and time series data, making it difficult to fully understand the deep semantics of complex problems.

[0004] The development of agent technology and Model Context Protocol (MCP) services has provided new approaches to solving the aforementioned problems. For example, the AutoCoA framework has demonstrated that agents with autonomous decision-making capabilities can significantly improve task completion rates, while Yaochi Database's MCP Server supports efficient communication among multiple agents, laying the foundation for building flexible and collaborative intelligent systems. However, current technologies have not yet deeply integrated dynamic intent recognition, GraphRAG architecture, agent adaptive planning, and MCP services, resulting in significant room for improvement in question-answering capabilities, system interpretability, and ecosystem collaboration in complex scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a question-answering method and system for open-domain fusion dynamic intent recognition and GraphRAG, which can solve the technical problems of misunderstanding bias, inaccurate answers and lack of contextual relevance in the existing technology.

[0006] To achieve the above objectives, a specific embodiment of the present invention provides a question-answering method for open-domain fusion of dynamic intent recognition and GraphRAG, the method comprising the following steps: Heterogeneous knowledge graphs are constructed based on multi-source heterogeneous data. Semantic relationships between entities are identified through entity relation extraction technology, and the heterogeneous knowledge graphs are dynamically updated according to preset trigger conditions. The system performs a two-stage extraction of low-level entity information and high-level relational information from user queries, and combines few-shot prompting learning techniques to generate accurate intent classification results. Based on the intent recognition results, user needs are broken down into a sequence of sub-tasks with dependencies. The optimal processing strategy is dynamically selected through reinforcement learning, and multiple agents are scheduled to execute collaboratively. By standardizing the interfaces of heterogeneous knowledge graphs, intent recognition, intelligent agents, and external tools through the model context protocol, dynamic service scheduling and collaborative reasoning based on context awareness are achieved. Collect user feedback data, analyze the deviation between question-and-answer results and user intent, and iteratively optimize the heterogeneous knowledge graph structure, intent recognition model parameters, and agent decision-making strategies.

[0007] In one or more embodiments of the present invention, the step of constructing a heterogeneous knowledge graph includes: extracting triples from unstructured text using a large language model to generate an entity-relation-entity knowledge representation; vectorizing the heterogeneous knowledge graph using a graph neural network to achieve semantic similarity calculation; and setting a time window or data volume threshold to trigger incremental updates of the heterogeneous knowledge graph.

[0008] In one or more embodiments of the present invention, the two-stage intent recognition step includes: Core entities and domain keywords in user queries are extracted using named entity recognition technology. Identify the types of relationships between entities through semantic role labeling; By combining semantic paths in heterogeneous knowledge graphs, ambiguity is resolved for multiple candidate intents, and the intent categories are output in probability ranking.

[0009] In one or more embodiments of the present invention, the adaptive task planning step includes: User requirements are broken down into subtasks that conform to the INVEST principle, and a directed acyclic graph is constructed to represent the dependencies. Based on heterogeneous knowledge graph reasoning and reinforcement learning, dynamic selection of processing strategies such as retrieval, generation, and verification is achieved. The process involves scheduling the retrieval agent to acquire relevant knowledge, generating an agent to construct candidate answers, and evaluating the agent to verify the reasonableness of the answers.

[0010] In one or more embodiments of the present invention, the MCP service integration step includes: Based on the unified communication standard defined by the model context protocol, the interfaces of the functional units in the system involving heterogeneous knowledge graph construction, intent recognition, task planning and tool invocation are standardized and encapsulated. The service registration mechanism of the Model Context Protocol combines the encapsulated functional units into complete services that conform to the MCP specification, and defines standardized service discovery, invocation and parameter passing rules. The completed MCP service is uploaded to the open source market for hosting, and version control and permission management mechanisms are configured to form a reusable standardized service package; By leveraging the distribution mechanism of the open-source marketplace, pre-configured MCP service packages are provided to third-party developers, supporting the rapid integration of Q&A functionality through standardized interfaces and enabling plug-and-play deployment of services. Collect integration feedback data from third-party users, and based on the compatibility mechanism of the MCP protocol, iteratively optimize and update the service to ensure that incremental improvements do not affect existing integration solutions.

[0011] In one or more embodiments of the present invention, the closed-loop optimization step includes: Collect user feedback data such as satisfaction ratings for answers and suggestions for improvement; By comparing the discrepancy between the actual intent and the system's recognized intent, performance bottlenecks can be identified. Based on the analysis results, update the weights of the heterogeneous knowledge graph, adjust the parameters of the intent recognition model, and optimize the decision-making rules of the intelligent agent.

[0012] In another aspect of the present invention, a question-answering system for open-domain fusion of dynamic intent recognition and GraphRAG is provided, the system comprising: The graph knowledge base module is used to build, store, and update heterogeneous knowledge graphs, supporting semantic retrieval and reasoning; The intent recognition module is used to perform two-stage intent recognition and generate a structured intent representation; The agent collaboration module is used to realize adaptive task planning and multi-agent collaborative execution; The MCP service integration module is used to standardize component interfaces and implement dynamic service scheduling. The feedback optimization module is used to collect user feedback and drive iterative optimization of the system.

[0013] In one or more embodiments of the present invention, the graph knowledge base module includes a knowledge extraction submodule, a graph storage submodule, and a knowledge fusion submodule, wherein, The knowledge extraction submodule is used to extract entities and relationships from multi-source data based on large language models and natural language processing techniques; The graph storage submodule is used to store heterogeneous knowledge graphs using a distributed graph database. The knowledge fusion submodule is used to handle conflicts between newly added knowledge and existing knowledge graphs.

[0014] In one or more embodiments of the present invention, the intent recognition module includes a two-stage information extraction submodule, a few-shot learning submodule, and an intent decision submodule, wherein, The dual-order information extraction submodule is used to extract low-order entity information and high-order relationship information respectively. The few-shot learning submodule is used to generate intent discrimination rules based on a small amount of labeled data; The intent decision submodule is used to combine user profiles and dialogue history to output the optimal intent classification result.

[0015] In one or more embodiments of the present invention, the agent coordination module includes a task planning submodule, an agent scheduling submodule, and a strategy optimization submodule, wherein, The task planning submodule decomposes user needs into a sequence of sub-tasks based on heterogeneous knowledge graph reasoning. The agent scheduling submodule is used to manage the collaboration of retrieving agents, generating agents, and evaluating agents; The strategy optimization submodule is used to continuously improve the task processing strategy through reinforcement learning.

[0016] Compared with existing technologies, this invention has the following advantages: By integrating an agent mechanism, a graph-enhanced retrieval generation (GraphRAG) architecture, and a model context protocol (MCP) technology, this invention constructs an automated and intelligent question-answering process. The system employs a two-stage information extraction and intent discrimination strategy, combined with few-shot learning, enabling it to quickly understand the core semantics of user questions and accurately locate relevant knowledge sources. Compared with traditional retrieval methods based on vector similarity matching, this invention significantly improves the accuracy and recall of information matching by introducing graph structure modeling and multi-hop path reasoning mechanisms, especially demonstrating stronger adaptability when facing complex reasoning problems. This invention designs a dynamic update mechanism and an efficient retrieval system. The system incorporates an internet information integration module, combined with agent technology and external search engine interfaces, enabling real-time acquisition of the latest data and incremental graph updates. Simultaneously, based on a fast traversal algorithm and reranking mechanism for graph databases, the system ensures that it can complete the retrieval and filtering of large-scale heterogeneous knowledge graphs within milliseconds, thereby achieving immediate response to user requests and high-quality output. This invention emphasizes the system's self-evolution capability. By integrating a feedback optimization module, the system can automatically adjust its retrieval strategy and generation logic based on user behavior and feedback. This closed-loop intelligent evolution mechanism not only enhances the system's adaptability but also improves user experience and satisfaction. Particularly in professional fields such as law and medicine, continuous system optimization helps to better align with industry scenarios and improve the professionalism and credibility of services. This invention places particular emphasis on the deep integration of industry, academia, and research, and improving the efficiency of knowledge transformation. By constructing a unified information sharing platform and supply-demand matching mechanism, the system breaks down information barriers between research institutions, universities, and enterprises, promoting the rapid dissemination and market application of scientific and technological achievements. Furthermore, the system can be widely applied to high-value scenarios such as legal case analysis, medical diagnostic assistance, and policy consultation, driving the digital transformation and intelligent upgrading of various industries. This invention also introduces the GraphRAG framework and Graph Neural Networks (GNNs) to achieve effective modeling of multi-hop paths in heterogeneous knowledge graphs. The system can automatically identify entity links and reasoning paths related to user questions in the graph structure, and perform deep analysis in combination with contextual semantics, thereby effectively supporting the answering of questions that require multi-step reasoning, such as "the derivative development of a certain technology" or "the potential etiological association of a certain disease", greatly expanding the system's capability boundaries.

[0017] In the answer generation stage, this invention not only relies on the natural language generation capabilities of large-scale language models (LLMs), but also integrates structured knowledge from the graph into the generation process, forming a "graph-aware" answer mechanism. For example, when recommending a scientific and technological achievement, the system can not only provide its basic introduction, but also automatically generate its technological evolution path, its correlation and impact with other technologies, and applicable industries, making the output more logical, coherent, and explanatory, and meeting users' needs for in-depth information.

[0018] This invention, leveraging Agent technology and the MCP model context protocol mechanism, successfully achieves seamless integration between private data and publicly available internet data. The system can flexibly access internal enterprise knowledge bases, industry databases, and internet information sources while protecting data privacy, breaking down the "data silos" phenomenon present in traditional systems and providing users with richer, more authoritative, and real-time knowledge support. (See attached figures.) To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a question-answering method for open-domain fusion dynamic intent recognition and GraphRAG according to one embodiment of the present invention; Figure 2 This is a schematic diagram of the framework of a question-answering system for open-domain fusion dynamic intent recognition and GraphRAG according to one embodiment of the present invention; Figure 3 This is a schematic diagram of the main process of a question-answering method for open-domain fusion dynamic intent recognition and GraphRAG in one embodiment of the present invention; Figure 4 This is a block diagram of a question-answering device for open-domain fusion dynamic intent recognition and GraphRAG according to one embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0021] As described in the background section, existing methods for evaluating tobacco quality suffer from problems such as complex data processing, highly subjective weighting of evaluation indicators, and inability to adjust them dynamically in real time.

[0022] To address the aforementioned technical issues, such as Figure 1 As shown, this invention introduces a question-answering method for open-domain fusion dynamic intent recognition and GraphRAG, including the following steps: S1. Construct a heterogeneous knowledge graph based on multi-source heterogeneous data, identify semantic relationships between entities through entity relationship extraction technology, and dynamically update the heterogeneous knowledge graph according to preset trigger conditions. S2. Perform a two-stage extraction of low-order entity information and high-order relation information from user queries, and combine few-shot prompting learning techniques to generate accurate intent classification results. S3. Based on the intent recognition results, user needs are broken down into a sequence of sub-tasks with dependencies. The optimal processing strategy is dynamically selected through reinforcement learning, and multiple agents are scheduled to execute collaboratively. S4. By standardizing the interfaces of heterogeneous knowledge graphs, intent recognition, intelligent agents and external tools through model context protocols, dynamic service scheduling and collaborative reasoning based on context awareness are realized. S5. Collect user feedback data, analyze the deviation between the question-and-answer results and the user's intent, and iteratively optimize the heterogeneous knowledge graph structure, intent recognition model parameters, and agent decision-making strategies.

[0023] In this embodiment, the graph knowledge base construction steps include: S11. Knowledge Extraction and Entity Recognition: Based on Large Language Modeling (LLM), this system automatically extracts key entities (such as technology names, R&D personnel, R&D institutions, and patent abstracts) from unstructured texts like scientific literature, patents, and project reports. By designing structured prompt word templates, the system guides the model to output standardized JSON-formatted entities and triple relationships, achieving semantic parsing and knowledge structuring of complex text content. To enhance the system's continuous adaptability, an incremental update mechanism is integrated, supporting two triggering methods: a time window mechanism that automatically updates the knowledge graph every 7 days; and a data volume threshold mechanism that immediately triggers the update process when the number of newly added documents reaches or exceeds 500. This ensures system stability while improving the timeliness and completeness of knowledge discovery, continuously maintaining high-quality knowledge discovery capabilities.

[0024] S12. Knowledge Graph Construction and Relationship Modeling: Extracted scientific and technological achievements and their attributes are constructed into a heterogeneous knowledge graph. Nodes represent entities, and edges represent various types of relationships between entities (such as "applied to," "developed by," "belonging to," etc.). Through knowledge representation learning methods such as graph neural networks (GNNs), the nodes and relationships in the graph are mapped to a low-dimensional vector space, enabling semantic similarity measurement and reasoning capabilities.

[0025] S13. Multi-source data fusion supports unified modeling and fusion of data from different sources and formats, eliminating information redundancy and conflicts. Heterogeneous knowledge graphs are stored and managed through graph databases (such as Neo4j and JanusGraph), improving data query efficiency and scalability.

[0026] S14. Establish a dynamic update mechanism: Design an incremental graph update strategy, combined with event detection and version control mechanisms, to ensure that the heterogeneous knowledge graph can reflect the evolution trend and technological dynamics of the latest scientific and technological achievements in real time, and to ensure the timeliness and accuracy of the recommendation system.

[0027] S15. Enhanced Graph Retrieval: Building upon traditional vector retrieval, graph structure information is introduced. Graph traversal algorithms (such as breadth-first search and random walk) are used to uncover potential related paths, enhancing the diversity and contextual relevance of retrieval results. Graph embedding technology is used to assist the ranking model, improving the intelligence level of cross-domain result matching.

[0028] In a further embodiment, the step of constructing a heterogeneous knowledge graph includes: S21. Fine-grained requirement extraction: Combining named entity recognition (NER) and keyword extraction technologies, key elements are extracted from user queries, including target fields (such as artificial intelligence, new materials), technical indicators (such as accuracy, efficiency), application scenarios (such as industrial testing, medical assistance), and cooperation preferences (such as R&D institution type, maturity requirements), as shown in Table 1 below. This provides structured input for subsequent retrieval and recommendation.

[0029] Table 1. Structured Definition Table of Key Elements

[0030] S21. Semantic parsing and intent classification, including deep semantic parsing of natural language input by users based on a large language model to identify the core intent of their technical needs. Through multi-level classifiers or sequence labeling models, user queries are divided into different intent categories, such as "technology search," "application scenario matching," "results comparison," and "feasibility analysis," as shown in Table 2.

[0031] Table 2. Definition of Intent Category System

[0032] S23. Context Modeling and Interaction Understanding: In multi-turn dialogue scenarios, a Dialogue State Tracking (DST) mechanism is introduced to maintain user historical interaction information, understand the logical relationship between the current query and previous dialogues, and improve the system's ability to understand complex technical requirements and its responsiveness. Specifically, the DST implementation first extracts and identifies key entities and information elements in the text using a large model. Then, the intent recognition module determines the semantic intent type of the current query. Based on this, the dialogue state update module merges the current input with historical interaction information, dynamically maintaining and updating the global dialogue state. Next, the context consistency verification module compares the current state with the historical state to detect conflicts or missing information. If an anomaly is found, a clarification mechanism is triggered to ask the user guiding questions to ensure accurate understanding of the requirements. Finally, the response generation module generates a natural language response based on the complete context state, achieving coherent understanding and intelligent response to user technical requirements in multi-turn dialogues.

[0033] S24. Intent Enhancement and Ambiguity Resolution: For ambiguous or incomplete user input, proactive questioning strategies or reasoning methods based on heterogeneous knowledge graphs are used to complete and clarify intent. For example, when a user mentions "I'm looking for an image recognition technology," the system can automatically identify potential concerns (such as recognition accuracy, deployment environment, whether it supports edge computing, etc.) and guide the user to provide more specific details of their needs.

[0034] S25. Personalized Intent Modeling: Combining user profiles and historical behavior data, construct a personalized intent model to identify the differences in the expression of technical needs among different user groups (such as enterprise technical leaders, researchers, and investors), and achieve more targeted knowledge services and results recommendations.

[0035] In a further embodiment, the two-stage information extraction and intent discrimination includes understanding user input using a two-stage strategy. The first stage is based on a language model to extract keywords, identify entities, and perform preliminary intent classification. The second stage mines entity relationship overviews and contextual information to achieve deep semantic modeling and intent refinement. It also combines few-shot prompting learning technology to accurately capture user needs and improve the accuracy and relevance of retrieval and generation.

[0036] In a further embodiment, the adaptive task planning includes the following steps: S31. Task Decomposition and Path Modeling: Based on user intent recognition results and semantic relationships in the graph knowledge base, the user's high-level technical requirements are decomposed into a sequence of executable sub-tasks. In the task decomposition process, we first use a planning strategy to enable the agent to decompose user requirements into phased goals and determine execution paths based on the top-level design capabilities of the large model. Then, a divide-and-conquer strategy is adopted to decompose each phased goal into atomic sub-tasks conforming to the INVEST principle. Simultaneously, execution tools such as APIs, graph knowledge bases, LangChain, or native code are matched according to task characteristics. Throughout the decomposition process, a chain-of-thought strategy is continuously used to explicitly deduce the logical dependencies and resource constraints between sub-tasks, forming an interpretable task topology. The network; then, through a multi-agent collaboration strategy, a dedicated agent (such as PlanAgent) is responsible for dynamically adjusting the task sequence, and ActionAgent calls the MCP plugin to execute, and a third-party agent is introduced for real-time verification; finally, relying on a self-reflection strategy, the execution results are automatically tested and logically backtracked, and the feedback data is injected into the heterogeneous knowledge graph to optimize the next decomposition process, forming a closed-loop system of "planning-decomposition-execution-reflection" (see supplementary materials for detailed examples), automatically planning the technical docking path from demand to result matching.

[0037] S32. Multi-Strategy Decision Engine: Introduces a hybrid decision-making mechanism combining reinforcement learning (RL) and rule-based reasoning. By defining a structured state space (including user intent, domain features, data completeness, and historical feedback) and action space (such as graph structure reasoning, vector retrieval, ranking, generation, and clarification), the system achieves intelligent selection of task processing strategies. The system first makes a preliminary judgment on the task type based on preset rules, directly calling the corresponding module when the conditions are clear. Simultaneously, it introduces a reinforcement learning model (such as DQN or Q-learning) to dynamically optimize strategy selection based on real-time feedback, enabling the agent to dynamically select the most suitable processing strategy based on different types of task characteristics (such as technology maturity assessment, cross-domain adaptation, and collaborative resource recommendation). For example, when facing queries about emerging technologies, graph structure reasoning capabilities are prioritized; under existing clear indicator constraints, the system switches to vector retrieval and ranking models.

[0038] S33. Context Awareness and Environmental Feedback Adjustment: The agent possesses the ability to respond in real time to user feedback, interaction history, and changes in external data. When a user modifies query conditions or evaluates recommendation results, the system updates the current task status through an online learning mechanism and dynamically adjusts the execution order and methods of subsequent steps, improving the flexibility and accuracy of the recommendation process.

[0039] S34. Multi-Agent Collaboration Mechanism: In complex technical scenarios, multiple functional agents (such as retrieval agents, generation agents, and evaluation agents) work collaboratively, autonomously assigning roles and exchanging intermediate results based on task requirements (see supplementary materials for relevant JSON schemas). By sharing global state information and local decision-making mechanisms, efficient knowledge retrieval and logical reasoning are achieved, enhancing the system's ability to support multi-dimensional scientific and technological achievement matching tasks.

[0040] S35. Self-optimization and experience accumulation: The agent has the ability to self-evaluate and optimize its behavior. By recording the results of each task execution (such as user satisfaction, conversion rate, and feedback rating), it continuously optimizes its internal decision-making model and execution process, forming a continuously evolving knowledge service capability.

[0041] In a further embodiment, Model Context Protocol (MCP) service integration includes the following steps: S41. Service Standardization and Protocol Alignment: To achieve efficient collaboration across multiple models, tools, and platforms, the system introduces the Model Context Protocol (MCP) as a unified communication standard. This protocol defines core interaction mechanisms such as context management, prompt word exchange, and tool invocation, ensuring the compatibility and scalability of various AI service capabilities in different environments.

[0042] S42. Modular Service Encapsulation and Registration: The Model Context Protocol (MCP) Server provides unified registration and metadata management for various services, including Large Language Models (LLM), vector databases, retrieval services, generation services, and ranking models. Each service is encapsulated in a modular manner, providing standardized interfaces for agents to dynamically invoke. For example, the retrieval service can be registered as `retrieval-service`, and the generation service as `generation-service`, supporting on-demand loading and combined execution.

[0043] S43. Context Awareness and Task-Driven Scheduling: The MCP client is embedded in the Agent runtime, responsible for receiving user input, maintaining context state, and automatically selecting the required service based on the current task type. For example, in a technology achievement matching scenario, the Agent will sequentially call the graph knowledge base retrieval service, intent understanding service, ranking service, and generation service to form an end-to-end service chain to respond to user needs.

[0044] S44. Heterogeneous Platform Compatibility and Cross-System Collaboration: The MCP protocol boasts excellent cross-platform adaptability, supporting various development frameworks (such as LangChain and LlamaIndex) and deployment environments (such as local servers, cloud platforms, and edge devices). Through MCP integration, multiple external systems can be connected under a unified interface, such as enterprise internal knowledge management systems, research institution results publishing platforms, or third-party open APIs, breaking down information silos and improving collaboration efficiency.

[0045] S45. Service Monitoring and Dynamic Optimization: The system has a built-in MCP service monitoring module that collects real-time metrics such as call frequency, response time, and success rate of each service node, and dynamically adjusts service combination strategies based on feedback mechanisms. For example, when the response latency of a certain search service is too high, the system can automatically switch to a backup service instance to ensure the stability and efficiency of the overall inference process.

[0046] In a further embodiment, the closed-loop optimization steps include: collecting user satisfaction ratings for answers, correction suggestions, and other feedback data; comparing the deviation between the actual intent and the system's recognized intent to pinpoint performance bottlenecks; and updating the weights of the heterogeneous knowledge graph, adjusting the parameters of the intent recognition model, and optimizing the agent's decision-making rules based on the analysis results.

[0047] like Figure 4 As shown, a system for implementing a question-answering method for open-domain fusion dynamic intent recognition and GraphRAG, according to a specific embodiment of the present invention, is described. The system includes: The graph knowledge base module is used to build, store, and update heterogeneous knowledge graphs, supporting semantic retrieval and reasoning. The graph knowledge base module includes a knowledge extraction submodule, a graph storage submodule, and a knowledge fusion submodule. The knowledge extraction submodule is used to extract entities and relations from multi-source data based on large language models and natural language processing techniques. The graph storage submodule is used to store heterogeneous knowledge graphs using a distributed graph database. The knowledge fusion submodule is used to handle conflicts between newly added knowledge and existing graphs.

[0048] The intent recognition module performs two-stage intent recognition and generates a structured intent representation. It includes a two-stage information extraction submodule, a few-shot learning submodule, and an intent decision submodule. The two-stage information extraction submodule extracts low-level entity information and high-level relational information respectively. The few-shot learning submodule generates intent discrimination rules based on a small amount of labeled data. The intent decision submodule combines user profiles and dialogue history to output the optimal intent classification result.

[0049] The agent coordination module is used to implement adaptive task planning and multi-agent collaborative execution. The agent coordination module includes a task planning submodule, an agent scheduling submodule, and a strategy optimization submodule. The task planning submodule decomposes user needs into a sequence of subtasks based on heterogeneous knowledge graph reasoning; the agent scheduling submodule manages the retrieval agent, generates agents, and evaluates agent collaboration; and the policy optimization submodule continuously improves task processing strategies through reinforcement learning.

[0050] The MCP service integration module is used to standardize component interfaces and implement dynamic service scheduling; the MCP service integration module includes: The protocol adaptation and interface standardization submodule is used to define the interfaces of the functional units of the question-answering system, such as knowledge processing, intent parsing, and task execution, according to the communication standard of the Model Context Protocol (MCP), unify the input and output formats, service call specifications, and error handling mechanisms, and ensure that each functional unit conforms to the interaction requirements of the MCP protocol.

[0051] The service encapsulation and composition submodule is used to modularly encapsulate standardized functional units. It completes the registration of metadata, such as service name, function description, and call parameters, through the MCP service registration mechanism. Based on the service logic, it combines the encapsulated units into a complete question-and-answer service chain, and supports the dynamic loading of different service units according to task requirements.

[0052] The open-source hosting and distribution submodule is used to package MCP services, including version identification, dependency list and deployment configuration; it connects to the hosting interface of the open-source marketplace to complete service upload, permission configuration and version management; it provides standardized acquisition and deployment guidelines, and supports third parties to directly obtain and integrate services through the open-source marketplace.

[0053] The monitoring and iterative optimization submodule is used to collect service call data (such as call frequency, response time, error rate) and user feedback in the open source market; based on the compatibility mechanism of the MCP protocol, it optimizes the performance or iterates the function of the service, generates a new service version and synchronizes it to the open source market, ensuring that the existing integration solution is not affected by incremental updates.

[0054] The feedback optimization module is used to collect user feedback and drive iterative system optimization. It includes a feedback collection submodule, which gathers qualitative and quantitative user evaluations of the question-and-answer results; a performance analysis submodule, which compares the actual intent with the system output to pinpoint the source of errors; and a parameter tuning submodule, which automatically adjusts the key parameters of each module based on the analysis results.

[0055] To verify the effectiveness of the present invention during implementation, the following specific embodiments are provided: Case 1: When a user enters keywords such as "smart grid" into the system, the system will automatically determine whether to perform operations such as internet search, patent search, or paper search based on natural language understanding and intent recognition models. By introducing few-shot learning and semantic modeling techniques, the system can accurately identify user intent and conduct multi-dimensional positive reasoning around the keywords, including: extracting the core characteristics of "smart grid," identifying relevant supporting technologies (such as the Internet of Things and big data analysis), reasoning about its application scenarios (such as urban energy management and industrial automation), matching relevant experts and research results, and predicting future development trends. The entire process realizes the automatic output of structured, multi-dimensional knowledge results starting from simple keywords, improving the efficiency and accuracy of matching scientific and technological achievements. Case 2: When a user inputs a brief description of their research achievement, such as a technical description of the synthesis of siloxane resins using silane coupling agents, the system automatically generates matching demand information through a two-stage information extraction mechanism and heterogeneous knowledge graph reasoning capabilities. The system first identifies key technical features of the achievement (such as high-temperature resistance and good adhesion), and based on this, recommends relevant companies, infers potential application areas (such as aerospace and electronic packaging), matches suitable cooperation methods (such as technology transfer and industry-academia-research collaboration), and provides corresponding contact information. This process integrates text parsing, entity relationship modeling, and knowledge graph path reasoning to ensure accurate extraction of key information from unstructured achievement descriptions and mapping it to market demand, thus facilitating the commercialization of scientific and technological achievements.

[0056] Case 3: When a user inputs a legal case description into the system, such as, "An employee of a company was sued for leaking trade secrets, and the court ultimately ruled that the employee was liable for compensation and paid liquidated damages," the system will automatically identify the case type, key elements, and points of contention based on natural language understanding and a heterogeneous legal knowledge graph, and perform structured reasoning. Through a two-stage information extraction mechanism, the system first extracts core entities and relationships such as the case subjects (e.g., "employee," "company"), actions (e.g., "leaking trade secrets"), and results (e.g., "liability for compensation," "liquidated damages"). It then further combines legal provisions, case law databases, and industry standards to automatically match corresponding legal clauses, similar precedents, applicable regulations, and information on relevant legal experts or law firms.

[0057] This invention uses data extraction, vectorization processing, and index creation as the foundation for rapid matching of scientific and technological achievements. Among these, the vectorization processing technology converts text data into machine-understandable vector representations, providing support for subsequent matching algorithms. This invention improves the understanding of complex questions and the effectiveness of multi-hop reasoning by introducing a Graph-Enhanced Retrieval Generation (GraphRAG) architecture, combining heterogeneous knowledge graphs with large-scale document retrieval capabilities. By modeling entity relationships through graph neural networks (GNNs) and optimizing retrieval paths, the system can integrate information from multiple sources and generate logically related answers, thereby improving the accuracy and robustness of the question-answering system. The task planning and autonomous decision-making capabilities of the intelligent agent in this invention enable the deployment of this invention to possess task decomposition, path planning, and dynamic scheduling capabilities. This allows it to identify the type of user question, automatically determine whether multi-step reasoning or multi-hop retrieval is required, and coordinate different modules to collaboratively complete complex tasks. This mechanism significantly improves the system's intelligence level and adaptability, making it closer to a human-like question-and-answer understanding and reasoning process. This invention, based on credibility assessment and explainability mechanisms, is used for credibility assessment and reasoning path recording. It performs source analysis on the generated answers and outputs a credibility score by combining factors such as the authority of the knowledge source and the completeness of the reasoning. Simultaneously, the system supports the visualization of the reasoning process, improving the transparency of the question-and-answer results and user trust, making it particularly suitable for professional fields such as medicine and law, where explainability requirements are high. The full-cycle service projects and information sharing and supply-demand matching platform provided by this invention promote the rapid transformation and application of scientific and technological achievements, strengthen the deep integration of industry, academia and research, and improve the efficiency of marketization and commercialization of scientific and technological achievements.

[0058] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0059] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0060] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0061] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0062] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0063] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A question-answering method for open-domain fusion of dynamic intent recognition and GraphRAG, characterized in that, The method includes the following steps: Heterogeneous knowledge graphs are constructed based on multi-source heterogeneous data. Semantic relationships between entities are identified through entity relation extraction technology, and the heterogeneous knowledge graphs are dynamically updated according to preset triggering conditions. The system performs a two-stage extraction of low-level entity information and high-level relational information from user queries, and combines few-shot prompting learning techniques to generate accurate intent classification results. Based on the intent recognition results, user needs are broken down into a sequence of sub-tasks with dependencies. The optimal processing strategy is dynamically selected through reinforcement learning, and multiple agents are scheduled to execute collaboratively. By standardizing the interfaces of heterogeneous knowledge graphs, intent recognition, intelligent agents, and external tools through the model context protocol, dynamic service scheduling and collaborative reasoning based on context awareness are achieved. Collect user feedback data, analyze the deviation between question-and-answer results and user intent, and iteratively optimize the heterogeneous knowledge graph structure, intent recognition model parameters, and agent decision-making strategies.

2. The question-answering method for open-domain fusion dynamic intent recognition and GraphRAG as described in claim 1, characterized in that, The steps for constructing the heterogeneous knowledge graph include: We use a large language model to extract triples from unstructured text and generate an entity-relation-entity knowledge representation. The heterogeneous knowledge graph is vectorized using a graph neural network to achieve semantic similarity calculation. Set a time window or data volume threshold to trigger incremental updates of the heterogeneous knowledge graph.

3. The question-answering method for open-domain fusion dynamic intent recognition and GraphRAG as described in claim 1, characterized in that, The two-stage intent recognition steps include: Core entities and domain keywords in user queries are extracted using named entity recognition technology. Identify the relationship types between entities through semantic role labeling; By combining semantic paths in heterogeneous knowledge graphs, ambiguity is resolved for multiple candidate intents, and the intent categories are output in probability ranking.

4. The question-answering method for open-domain fusion dynamic intent recognition and GraphRAG as described in claim 1, characterized in that, The adaptive task planning steps include: User requirements are broken down into subtasks that conform to the INVEST principle, and a directed acyclic graph is constructed to represent the dependencies. Based on heterogeneous knowledge graph reasoning and reinforcement learning, dynamic selection of processing strategies such as retrieval, generation, and verification is achieved. The process involves scheduling the retrieval agent to acquire relevant knowledge, generating an agent to construct candidate answers, and evaluating the agent to verify the reasonableness of the answers.

5. The question-answering method for open-domain fusion dynamic intent recognition and GraphRAG as described in claim 1, characterized in that, The MCP service integration steps include: Based on the unified communication standard defined by the model context protocol, the interfaces of the functional units in the system involving heterogeneous knowledge graph construction, intent recognition, task planning and tool invocation are standardized and encapsulated. The service registration mechanism of the Model Context Protocol combines the encapsulated functional units into complete services that conform to the MCP specification, and defines standardized service discovery, invocation and parameter passing rules. The completed MCP service is uploaded to the open source market for hosting, and version control and permission management mechanisms are configured to form a reusable standardized service package; By leveraging the distribution mechanism of the open-source marketplace, pre-configured MCP service packages are provided to third-party developers, supporting the rapid integration of Q&A functionality through standardized interfaces and enabling plug-and-play deployment of services. Collect integration feedback data from third-party users, and based on the compatibility mechanism of the MCP protocol, iteratively optimize and update the service to ensure that incremental improvements do not affect existing integration solutions.

6. The question-answering method for open-domain fusion dynamic intent recognition and GraphRAG as described in claim 1, characterized in that, The closed-loop optimization steps include: Collect user feedback data such as satisfaction ratings for answers and suggestions for improvement; By comparing the discrepancy between the actual intent and the system's recognized intent, performance bottlenecks can be identified. Based on the analysis results, update the weights of the heterogeneous knowledge graph, adjust the parameters of the intent recognition model, and optimize the decision-making rules of the intelligent agent.

7. A question-answering system that implements the method of any one of claims 1 to 6, characterized in that, include: The graph knowledge base module is used to build, store, and update heterogeneous knowledge graphs, supporting semantic retrieval and reasoning; The intent recognition module is used to perform two-stage intent recognition and generate a structured intent representation; The agent collaboration module is used to realize adaptive task planning and multi-agent collaborative execution; The MCP service integration module is used to standardize component interfaces and implement dynamic service scheduling. The feedback optimization module is used to collect user feedback and drive iterative optimization of the system.

8. The question-answering system based on dynamic intent recognition and GraphRAG fusion for open-domain question answering according to claim 7, characterized in that, The graph knowledge base module includes a knowledge extraction submodule, a graph storage submodule, and a knowledge fusion submodule, wherein... The knowledge extraction submodule is used to extract entities and relationships from multi-source data based on large language models and natural language processing techniques; The graph storage submodule is used to store heterogeneous knowledge graphs using a distributed graph database. The knowledge fusion submodule is used to handle conflicts between newly added knowledge and existing knowledge graphs.

9. The question-answering system based on dynamic intent recognition and GraphRAG fusion for open-domain question answering according to claim 7, characterized in that, The intent recognition module includes a two-stage information extraction submodule, a few-shot learning submodule, and an intent decision submodule, wherein... The dual-order information extraction submodule is used to extract low-order entity information and high-order relationship information respectively. The few-shot learning submodule is used to generate intent discrimination rules based on a small amount of labeled data; The intent decision submodule is used to combine user profiles and dialogue history to output the optimal intent classification result.

10. The question-answering system based on dynamic intent recognition and GraphRAG fusion for open-domain question answering according to claim 7, characterized in that, The agent coordination module includes a task planning submodule, an agent scheduling submodule, and a strategy optimization submodule, wherein... The task planning submodule decomposes user needs into a sequence of sub-tasks based on heterogeneous knowledge graph reasoning. The agent scheduling submodule is used to manage the collaboration of retrieving agents, generating agents, and evaluating agents; The strategy optimization submodule is used to continuously improve the task processing strategy through reinforcement learning.

Images