Large language model and knowledge graph fused professional question-answering system and construction method
By integrating knowledge graphs and large language models, the system achieves automatic updates and expansion of professional question-answering systems, solving the problems of generating error messages in LLM and the complexity of knowledge graph construction, and providing efficient, accurate and real-time question-answering services.
Patent Information
- Application Number
- CN202511759817.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-03-06
AI Technical Summary
The problems include the 'illusion' phenomenon of generating incorrect information in professional fields by Large Language Models (LLM), the complexity and high cost of knowledge graph construction, the difficulty in system expansion and deployment, the high user threshold, and the inability of existing technologies to effectively handle real-time dynamic information.
By integrating knowledge graphs and LLM, we designed automatic and manual update mechanisms, combined with named entity recognition, vector similarity calculation, and Internet information retrieval, to reduce system integration complexity, provide an intuitive interface and automated tools, and ensure the real-time updates and accuracy of the knowledge graph.
It enables efficient collaboration between large language models and knowledge graphs, reduces system integration costs, improves the accuracy and flexibility of question answering, lowers the user threshold, and ensures the real-time nature and reliability of information.
Smart Images
Figure CN121614583A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of natural language processing and knowledge graph technology, specifically to a professional question-answering system and its construction method that integrates large language models and knowledge graphs. Background Technology
[0002] In recent years, Large Language Models (LLMs) such as GPT-3 and BERT have performed exceptionally well in natural language processing tasks, excelling at generating natural language and engaging in dialogue. These models rely on large-scale training data and are capable of understanding and generating complex sentences. However, when faced with complex problems in specialized domains, LLMs often fall short and are prone to the so-called 'illusion' phenomenon, where the generated content appears reasonable but is actually incorrect. This is mainly because LLMs rely on training corpora, which have limitations in their knowledge base within specialized domains, making them unable to effectively handle highly specialized or cutting-edge knowledge.
[0003] A knowledge graph is a structured database used to store knowledge, where nodes represent entities and edges represent relationships between entities. It helps computer systems understand and process complex professional knowledge. In question-answering systems, knowledge graphs can provide accurate answers and citation sources, thereby improving the accuracy and professionalism of question answers. Cypher is a query language used for managing and operating graph databases such as Neo4j. Although powerful, its specialized and complex nature makes the construction process of knowledge graphs cumbersome, requiring expert participation and hindering its widespread adoption.
[0004] The existing technology still has the following problems: (1) The coupling of Large Language Models (LLM) and knowledge graphs requires deep customization at the system architecture level, covering multiple tasks such as data interface development, query logic optimization, and response latency control. Such tasks usually rely on interdisciplinary engineering team collaboration, leading to extended project cycles and increased costs. At the same time, the highly customized nature significantly weakens the portability of the solution, thus hindering its large-scale deployment and rapid implementation. In the medical scenario, seamless integration interfaces for complex medical record retrieval require high customization, severely limiting the system's versatility and promotion potential. The same applies to the financial field, where embedding LLM into existing high-level financial data analysis systems still requires heavy integration work, further increasing time and economic costs and significantly increasing the overall system complexity.
[0005] (2) Existing knowledge graph systems generally rely on predefined rules and domain expert knowledge. Their continuous updates and maintenance require continuous expert intervention. As knowledge evolves, the maintenance cost of the rule base increases, which limits the system's flexibility and scalability. When faced with real-time dynamic information, the rule lag is particularly prominent. In the logistics field, emerging models and demands change frequently, and pre-set rules are difficult to adjust synchronously, resulting in the system's response lagging behind market changes and failing to effectively match real-time business scenarios. In the financial field, the market fluctuates drastically. If the knowledge graph cannot be updated in time, the system's decision-making accuracy and timeliness will be significantly impacted.
[0006] (3) The construction and querying of knowledge graphs usually require users to master professional query languages such as Cypher and SPARQL, which poses a high learning threshold and significant technical barriers for general users or application developers. Users need to invest a lot of cognitive resources to master specific tools, thereby inhibiting the penetration of knowledge graphs in a wider range of fields. In the public health scenario, the construction of medical knowledge graphs requires the use of Cypher to complete a large number of complex queries and pattern definitions. This process not only relies on the deep participation of information technology experts, but also increases the cognitive load of doctors and medical researchers, making it difficult for them to directly intervene in the graph construction process, thus limiting the efficient application and promotion of knowledge graphs in the medical field. The same is true in the education field. Teachers and students expect to achieve curriculum design and learning guidance through educational knowledge graphs, but the professional query languages and operational complexity make this vision difficult to achieve.
[0007] (4) Large language models are prone to "hallucination" when generating natural language responses, that is, outputting content that appears reasonable but is inconsistent with the facts. This phenomenon is particularly fatal in high-risk professional fields, as misinformation can lead to serious consequences. In the medical diagnosis process, if an LLM provides incorrect advice, it will directly harm the patient's treatment outcome; in legal scenarios, erroneous legal opinions may mislead lawyers and clients, leading to unfavorable judgments. Summary of the Invention
[0008] The purpose of this invention is to provide a professional question-answering system and construction method that integrates large language models and knowledge graphs, so as to solve the problems mentioned in the background art.
[0009] To achieve the above objectives, the present invention provides the following technical solution: A professional question-answering system that integrates large language models and knowledge graphs is characterized by including a knowledge graph construction system and an LLM question-answering system; The knowledge graph construction system includes a knowledge graph construction module, a JSON data import module, and a cell tree import module; The knowledge graph construction module uses predefined models and templates to standardize and normalize the knowledge graph structure, and supports automatic and manual updates. Automatic updates import literature data and use named entity recognition and relation extraction technology to capture and integrate the latest information in real time. Manual updates allow users to directly create and delete entities and relations in the knowledge graph after logging into the platform. It also supports the creation and updating of the knowledge graph by importing JSON data or cell trees. The LLM question answering system includes a question answering module, a vector similarity calculation module, an intent recognition module, a named entity recognition module, and a retrieval enhancement generation module; The question-and-answer module is used to receive questions input by users and coordinate with other modules to generate answers; The intent recognition module is used to determine the user's question intent; the named entity recognition module is used to extract keywords from the question; the vector similarity calculation module is used to calculate the vector similarity between the user query and knowledge graph nodes, retrieve information corresponding to the keywords from the knowledge graph, and extract the answer in combination with the intent; the retrieval enhancement generation module is used to perform Internet information retrieval and generate a natural language answer by integrating the retrieval results when the corresponding keywords are not found in the knowledge graph.
[0010] Preferably, the intent recognition module combines context and linguistic context, and uses a deep learning model to analyze and recognize the user's query intent.
[0011] Preferably, the named entity recognition module uses a pre-trained model to identify named entities in the query and match them in the knowledge graph. The named entities include personal names, place names, item names, organization names, and professional domain-specific entities.
[0012] Preferably, the vector similarity calculation module converts natural language queries into vector representations using deep learning algorithms and calculates the cosine similarity between vectors to determine the degree of matching between the user query and the knowledge graph nodes.
[0013] Preferably, after performing an internet information retrieval, the retrieval enhancement generation module will perform structured processing on the retrieval results and verify them against the content in the knowledge graph to generate a reliable answer that meets the user's needs.
[0014] Preferably, the JSON data import module supports automatic format recognition of the input JSON format data and completes the mapping between data fields and knowledge graph rules, transforming the JSON data into nodes and relationships of the knowledge graph; The cell tree import module supports automatic structure parsing of input spreadsheet-type cell tree format data, identifying the data hierarchy, headers, and specific data, and mapping cell elements to nodes and relationships in a knowledge graph.
[0015] The construction method of a professional question-answering system integrating large language models and knowledge graphs includes the following steps: Step 1: Building a knowledge graph. The system is configured with predefined models and templates to standardize the knowledge graph structure. Step 2: Build an intent recognition module for the LLM question-answering system, and use a deep learning model to train and determine the user's question intent; Step 3: System Integration and Deployment. Integrate the knowledge graph building system with the LLM question answering system through the underlying code, and design modules that can automatically process data interfaces, optimize query logic, and control response time. Adopt a configuration-based deployment method, so that developers can achieve collaborative work between LLM and knowledge graph through simple configuration without the need for heavy system reconstruction or custom development.
[0016] Preferably, it further includes: A knowledge graph construction module was developed, which implements automatic and manual update functions. The automatic update function imports literature data, calls named entity recognition and relation extraction technology to extract entities and relations, and integrates them into the knowledge graph. The manual update function allows users to directly manipulate entities and relations and import JSON data and cell tree data. Develop a JSON data import module and a cell tree import module to convert JSON data and cell tree data into knowledge graph nodes and relationships, respectively.
[0017] Preferably, it further includes: Develop a named entity recognition module to extract keywords and perform graph matching in the problem based on a pre-trained model; Develop a vector similarity calculation module to realize vector conversion of natural language queries and similarity calculation with graph nodes through deep learning algorithms; Develop a search enhancement generation module to enable internet information retrieval, structured processing of search results, and verification with knowledge graph content; Develop a question-and-answer module to coordinate the above modules and complete the question-and-answer process.
[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention directly integrates LLM and knowledge graph through the underlying code during program development, fundamentally simplifying the embedding process of knowledge graph and large language model, and reducing the need for deep customization at the system architecture level. During the integration process, modules are designed to automatically handle data interfaces, optimize query logic, and control response time, so that developers only need to perform simple configurations to achieve a high degree of collaboration between LLM and knowledge graph, without the need for heavy system reconstruction or custom development.
[0019] This invention designs a knowledge graph platform that can be automatically updated and expanded to meet the needs of dynamically changing information and real-time data. By introducing an automated data collection and update mechanism, the knowledge graph platform can dynamically capture and integrate the latest information, ensuring that the knowledge graph is always up-to-date.
[0020] This invention lowers the barrier to knowledge graph construction and querying by using a more intuitive user interface and automated tools, enabling non-professional users to easily build and query knowledge graphs.
[0021] This invention leverages the structure and precision of knowledge graphs to assist large language models in generating more reliable and accurate answers. By verifying and correcting the answers generated by the large language model using the knowledge graph, the incidence of hallucinations can be significantly reduced, ensuring the accuracy and reliability of the answers.
[0022] This invention designs a knowledge graph platform capable of automatic updates and expansion to meet the needs of dynamically changing information and real-time data. By introducing an automated data collection and update mechanism, the knowledge graph platform can dynamically capture and integrate the latest information, ensuring that the knowledge graph is always up-to-date. This platform can be applied to various scenarios. Attached Figure Description
[0023] Figure 1 This is a diagram of the overall system architecture of the present invention; Figure 2 This is a flowchart illustrating the knowledge graph construction process of this invention. Figure 3 This is a flowchart of the JSON data import process for this invention; Figure 4 This is a flowchart of the cell tree import process for this invention; Figure 5 This is a flowchart of the question-and-answer process of this invention; Figure 6 This is a flowchart of the vector similarity calculation process of the present invention; Figure 7 This is a flowchart illustrating the intent recognition process of this invention. Figure 8 This is a flowchart of the named entity recognition process of this invention; Figure 9 This is a flowchart of the retrieval enhancement generation module of the present invention. Detailed Implementation
[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] like Figure 1-9 As shown, a professional question-answering system that integrates large language models and knowledge graphs includes a knowledge graph construction system and an LLM question-answering system; The knowledge graph construction system includes a knowledge graph construction module, a JSON data import module, and a cell tree import module; The knowledge graph construction module uses predefined models and templates to standardize and normalize the knowledge graph structure, and supports both automatic and manual updates. Automatic updates import literature data and use named entity recognition and relation extraction technologies to capture and integrate the latest information in real time. Manual updates allow users to directly create and delete entities and relations in the knowledge graph after logging into the platform. It also supports the creation and updating of the knowledge graph by importing JSON data or cell trees. The JSON data import module automatically recognizes the format of input JSON data and maps data fields to knowledge graph rules, transforming JSON data into nodes and relations in the knowledge graph. The cell tree import module automatically parses the structure of input spreadsheet-like cell tree format data, identifying the data hierarchy, headers, and specific data, mapping cell elements to nodes and relations in the knowledge graph.
[0026] The LLM question answering system includes a question answering module, a vector similarity calculation module, an intent recognition module, a named entity recognition module, and a retrieval enhancement generation module; The question-and-answer module receives user-input questions and coordinates with other modules to generate answers. The intent recognition module determines the user's query intent. It combines context and language, using a deep learning model to analyze and identify the user's query intent. The named entity recognition module extracts keywords from the question. Using a pre-trained model, it identifies named entities in the query and matches them in the knowledge graph. Named entities include names of people, places, items, organizations, and domain-specific entities. The vector similarity calculation module calculates the vector similarity between the user query and knowledge graph nodes, retrieving information corresponding to keywords from the knowledge graph and extracting the answer based on the intent. The retrieval enhancement and generation module performs internet information retrieval when no corresponding keywords are found in the knowledge graph, generating a natural language answer by synthesizing the retrieval results. The vector similarity calculation module uses a deep learning algorithm to convert the natural language query into a vector representation and calculates the cosine similarity between vectors to determine the degree of matching between the user query and knowledge graph nodes. After performing internet information retrieval, the retrieval enhancement and generation module structures the retrieval results and verifies them against the content in the knowledge graph to generate a reliable answer that meets the user's needs.
[0027] The construction method of a professional question-answering system integrating large language models and knowledge graphs includes the following steps: Step 1: Construct a knowledge graph construction system. Configure predefined models and templates to standardize the knowledge graph structure. Develop a knowledge graph construction module to implement automatic and manual update functions. The automatic update function imports literature data, calls named entity recognition and relation extraction technologies to extract entities and relations, and integrates them into the knowledge graph. The manual update function allows users to directly manipulate entities and relations and import JSON data and cell tree data. Develop a JSON data import module and a cell tree import module to respectively convert JSON data and cell tree data into knowledge graph nodes and relations. The specific process for automatic updating of knowledge graphs is as follows: Step 11: Import literature data; Step 12: Perform named entity recognition; Step 13: Perform relation extraction; Step 14: Generate entity relationships; Step 15: Integrate into the knowledge graph.
[0028] Step 2: Develop an intent recognition module for the LLM question-answering system, using a deep learning model to determine user question intent; develop a named entity recognition module, using a pre-trained model to extract keywords from questions and perform graph matching; develop a vector similarity calculation module, using deep learning algorithms to perform vector conversion of natural language queries and calculate similarity with graph nodes; develop a retrieval enhancement generation module, enabling internet information retrieval, structured processing of retrieval results, and verification with knowledge graph content; and develop a question-answering module, coordinating the above modules to complete the process of "question reception → intent recognition → keyword extraction → graph retrieval → retrieval enhancement generation (on demand) → answer generation". Step 21: User inputs question; Step 22: The intent recognition module determines the intent of the question; Step 23: The named entity recognition module extracts the problem keywords; Step 24: The vector similarity calculation module retrieves matching nodes and relationships in the knowledge graph; Step 25: If a match is successful, extract relevant information from the knowledge graph to generate an answer; if a match fails, call the retrieval enhancement generation module to perform an internet search and combine it with the knowledge graph content to generate an answer. Step 3: System Integration and Deployment. Integrate the knowledge graph building system with the LLM question answering system through the underlying code, and design modules that can automatically process data interfaces, optimize query logic, and control response time. Adopt a configuration-based deployment method, so that developers can achieve collaborative work between LLM and knowledge graph through simple configuration without the need for heavy system reconstruction or custom development.
[0029] like Figure 1-9 As shown, the professional question-answering system that integrates large language models and knowledge graphs includes: The knowledge graph construction system consists of three modules: (1) Knowledge Graph Construction Module (2) JSON data import module (3) Cell Tree Import Module Knowledge Graph Construction System Description: (1) Knowledge graph construction module: The system flexibly constructs knowledge graphs through automation and user interaction. It uses predefined models and templates to ensure the standardization and normalization of the graph structure.
[0030] Supports both automatic and manual updates: Automatic updates: By importing literature data and utilizing named entity recognition and relation extraction technologies, the latest information is captured and integrated in real time to achieve automatic updates of the knowledge graph.
[0031] Manual Updates: After logging into the platform, users can directly create and delete entities and relationships in the knowledge graph, allowing for flexible management. Users can also import JSON data or cell trees to create and update the knowledge graph.
[0032] (2) JSON data import module: It supports direct import of data from JSON format into the knowledge graph. This includes automatic data format recognition and mapping, quickly transforming JSON data into nodes and relationships in the knowledge graph.
[0033] (2) Importing the cell tree module: It supports importing data from cell trees (such as spreadsheets). It automatically parses the cell tree structure and generates corresponding knowledge graph nodes and relationships.
[0034] The LLM question-and-answer system consists of five modules: (1) Question and Answer Module (2) Vector similarity calculation module (3) Intent recognition module (4) Named Entity Recognition Module (5) Search Enhancement Generation Module (6) Description of LLM Question-Answer System: LLM Q&A System Description: (1) Question and Answer Module: The user inputs a question, and the intent recognition module determines the user's intent.
[0035] Use the named entity recognition module to extract keywords from the question.
[0036] We use vector similarity calculation to retrieve keywords from the knowledge graph and combine them with intent to extract the corresponding answers from the knowledge graph.
[0037] If the corresponding keywords cannot be found in the knowledge graph, the search enhancement generation module will be used to retrieve information from the Internet to complete the answer.
[0038] (2) Vector similarity calculation module: Calculate the vector similarity between user queries and knowledge graph nodes to ensure accurate matching of relevant nodes.
[0039] Deep learning algorithms are used to convert natural language queries into vector representations and perform similarity matching.
[0040] (3) Intent recognition module: Analyze user query intent to ensure the system accurately understands user needs and provides appropriate answers. Intent recognition is performed using a deep learning model, taking into account context and language.
[0041] (4) Named Entity Recognition Module: Identify named entities (such as person names, place names, item names, organization names, etc.) in the query and match them in the knowledge graph. Use a pre-trained model to accurately identify and parse named entities, improving the accuracy of query matching.
[0042] (5) Search Enhancement Generation Module: When the knowledge graph fails to retrieve the corresponding answer, the search enhancement generation module is used to retrieve information from the Internet.
[0043] The final natural language answer is generated by combining the results of internet searches.
[0044] Example 1: Medical Diagnosis and Treatment Recommendation System background: In the medical field, doctors need to provide accurate diagnoses and treatment recommendations based on patients' conditions and medical knowledge bases. Existing medical AI systems often struggle to handle complex medical knowledge and constantly updated medical literature.
[0045] Implementation process: 1. Knowledge Graph Construction: Data sources: collected from hospital electronic medical records, medical literature, drug databases, etc.
[0046] Automatic Updates: Using named entity recognition and relation extraction techniques, key entities and relations are extracted from the latest medical literature to automatically update the medical knowledge graph.
[0047] Manual updates: Doctors can directly add, delete, or modify medical entities and relationships in the knowledge graph through the platform, and can also import JSON data or spreadsheet data.
[0048] 2. Question-and-answer system integration: User input: Doctors input information such as the patient's symptoms and laboratory test results.
[0049] Intent recognition: The system uses the intent recognition module to understand the doctor's query intent, such as diagnosing the cause of the disease or recommending a treatment plan.
[0050] Named entity recognition: Extract key medical entities (such as disease names, symptoms, drugs, etc.) from the query.
[0051] Vector similarity calculation: Retrieve nodes and relationships related to the query from the medical knowledge graph, and extract the corresponding answer based on the intent.
[0052] Search Enhancement Generation: If a complete answer cannot be found in the atlas, the search enhancement generation module is enabled to retrieve the latest medical internet data and generate comprehensive diagnostic suggestions and treatment plans.
[0053] Effect: The system can integrate the latest medical knowledge in a timely manner, provide accurate diagnoses and personalized treatment suggestions, and help doctors improve diagnostic efficiency and accuracy.
[0054] It greatly reduces the time doctors spend searching for information in complex literature and improves clinical work efficiency.
[0055] Through the deep integration of knowledge graphs and LLM, basic medical knowledge and the latest research findings can be fully utilized to ensure the reliability and timeliness of the answers.
[0056] Example 2: Intelligent Logistics Scheduling and Management System background: In the logistics sector, businesses need to track transportation routes, warehousing information, and other data in real time for efficient logistics scheduling and management. Existing logistics systems lack flexibility and struggle to cope with rapidly changing demands.
[0057] Implementation process: 1. Knowledge Graph Construction: Data sources: Data collected from logistics companies, including transportation routes, warehousing information, and traffic conditions.
[0058] Automatic updates: Utilizing automated data collection tools and relation extraction technology, the knowledge graph is updated in real time, containing the latest transportation routes and logistics information.
[0059] Manual update: Logistics managers can manually update transportation route information through the platform, or import JSON data or cell tree data for batch updates.
[0060] 2. Question-and-answer system integration: User input: Logistics administrators query information such as the best transportation routes and warehouse scheduling.
[0061] Intent recognition: The system understands the administrator's query needs through the intent recognition module, such as finding the fastest route or scheduling the optimal warehouse.
[0062] Named entity recognition: Extract key logistics entities (such as warehouse location, transport vehicle, etc.) from the query.
[0063] Vector similarity calculation: Retrieve relevant nodes and relationships from the logistics knowledge graph, and extract the corresponding answer based on the intent.
[0064] Search Enhancement Generation: When there is no definite answer in the map, the search enhancement generation module retrieves the latest traffic and logistics information from the Internet to supplement the scheduling plan.
[0065] Effect: It provides real-time, data-driven logistics solutions, significantly improving the scheduling efficiency and response speed of logistics companies.
[0066] The system automatically captures and integrates the latest logistics information to ensure the accuracy and effectiveness of scheduling decisions.
[0067] This greatly reduces the time logistics administrators spend manually searching and updating information, thus improving overall management efficiency.
[0068] Example 3: Financial Investment Analysis and Recommendation System background: In the financial sector, investment advisors and institutions need to provide clients with accurate investment analysis and advice based on massive amounts of market data and financial information. Existing systems struggle to integrate and update data in real time, lacking flexibility and accuracy.
[0069] Implementation process: 1. Knowledge Graph Construction: Data sources: financial data, market analysis reports, real-time stock market data, etc.
[0070] Automatic Updates: Utilizing named entity recognition and relation extraction technologies, key entities and relationships are extracted from the latest financial news and market reports to automatically update the financial knowledge graph.
[0071] Manual updates: Users can manually add, delete, or modify financial entities and relationships through the platform, and can also import JSON data or cell tree data.
[0072] 2. Question-and-answer system integration: User input: Investment advisors input clients' investment needs or market analysis requirements.
[0073] Intent recognition: The system uses the intent recognition module to understand the purpose of the query, such as predicting market trends or suggesting investment portfolios.
[0074] Named entity recognition: Extract key financial entities (such as stock names, market indicators, etc.) from queries.
[0075] Vector similarity calculation: Retrieve nodes and relationships related to the query from the financial knowledge graph, and extract corresponding analysis results based on the intent.
[0076] Search Enhancement Generation: When there is insufficient information in the map, the search enhancement generation module can retrieve the latest market data and analysis from internet channels to generate comprehensive recommendations.
[0077] Effect: The system can integrate the latest market data in real time, helping investment advisors provide accurate investment analysis and recommendations. This improves the efficiency and accuracy of market analysis and investment decisions, enhancing client confidence and satisfaction. Through deep integration of knowledge graphs and LLM (Local Management Model), it better utilizes professional financial data and market dynamics, ensuring the scientific rigor and timeliness of the analysis results.
[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A professional question answering system fusing a large language model and a knowledge graph, characterized in that, The system comprises a knowledge graph construction system and an LLM question answering system. The knowledge graph construction system comprises a knowledge graph construction module, a JSON data import module, and a cell tree import module. The knowledge graph construction module adopts predefined models and templates to realize the standardization and normalization of the knowledge graph structure, and supports automatic updating and manual updating. The LLM question answering system comprises a question answering module, a vector similarity calculation module, an intent recognition module, a named entity recognition module, and a retrieval enhancement generation module. The question answering module is used to receive user input questions and coordinate other modules to complete answer generation. The intent recognition module is used to determine the user question intent, the named entity recognition module is used to extract keywords in the question, and the vector similarity calculation module is used to calculate the vector similarity between the user query and the knowledge graph node, retrieve information corresponding to the keywords from the knowledge graph, and extract answers combined with the intent. The retrieval enhancement generation module is used to perform internet information retrieval and generate natural language answers by integrating retrieval results when the knowledge graph does not retrieve corresponding keywords.
2. The professional question answering system integrating a large language model and a knowledge graph according to claim 1, characterized in that, The intent recognition module combines context and context to realize the analysis and recognition of user query intent through a deep learning model.
3. The professional question answering system of claim 1, wherein, The named entity recognition module uses a pre-trained model to identify named entities in the query and match them in the knowledge graph, including names, place names, object names, organization names, and professional field-specific entities.
4. The professional question answering system integrating a large language model and a knowledge graph according to claim 1, characterized in that, The vector similarity calculation module converts natural language queries into vector representations through a deep learning algorithm and calculates the cosine similarity between vectors to determine the matching degree of user queries and knowledge graph nodes.
5. The professional question answering system of claim 1, wherein, The retrieval enhancement generation module performs structured processing on the retrieval results after internet information retrieval, and checks them with the content in the knowledge graph to generate reliable answers that meet user needs.
6. The professional question answering system of claim 1, wherein, The JSON data import module supports automatic format recognition of input JSON format data and completes the mapping of data fields and knowledge graph rules to convert JSON data into knowledge graph nodes and relationships. The cell tree import module supports automatic structure parsing of input spreadsheet type cell tree format data, identifies data levels, table headers, and specific data, and maps cell elements to knowledge graph nodes and relationships.
7. The method according to any one of claims 1-6, wherein the method comprises the following steps: obtaining a knowledge graph; obtaining a large language model; obtaining a professional question and answer system; and fusing the large language model and the knowledge graph to obtain a professional question and answer system. The steps comprise: Step 1: Configure predefined models and templates to standardize the knowledge graph structure in the knowledge graph construction system. Step 2: Develop the intent recognition module in the LLM question answering system to determine user question intent through deep learning model training. Step 3: System integration and deployment The knowledge graph construction system and the LLM question answering system are integrated through underlying code, and modules for automatically processing data interfaces, query logic optimization, and response time control are designed. Configuration-based deployment is used, allowing developers to easily configure the collaboration between LLM and knowledge graph without the need for heavy system reconstruction or custom development.
8. The method for constructing a professional question-answering system integrating a large language model and a knowledge graph according to claim 7, characterized in that, Also includes: Develop a knowledge graph construction module to implement automatic and manual update functions. The automatic update function imports literature data, calls named entity recognition and relationship extraction techniques to extract entities and relationships, and integrates them into the knowledge graph. The manual update function supports direct manipulation of entities and relationships, as well as importing JSON data and cell tree data. Develop JSON data import modules and cell tree import modules to convert JSON data and cell tree data into knowledge graph nodes and relationships.
9. The method for constructing a professional question-answering system integrating a large language model and a knowledge graph according to claim 7, characterized in that, Also includes: Develop a named entity recognition module based on a pre-trained model to extract keywords from questions and match them with the graph. Develop a vector similarity calculation module to convert natural language queries into vectors and calculate their similarity to graph nodes using deep learning algorithms. Develop a search enhancement generation module to implement internet information retrieval, structured processing of search results, and verification with knowledge graph content. Develop a question and answer module to coordinate the above modules to complete the question and answer process.