Intelligent question answering system based on language large model
By constructing an intelligent question-answering system based on a large language model, the problem of the inability of question-answering systems to update in real time was solved, and dynamic responses driven by knowledge graphs were realized, ensuring the timeliness and accuracy of the response content.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHONGNAN TRANSPORT
- Filing Date
- 2026-01-20
- Publication Date
- 2026-05-15
AI Technical Summary
Existing question-and-answer systems rely on historical response data and cannot be adjusted and updated in real time, which may result in incorrect or outdated responses.
A smart question-answering system based on a large language model is adopted. The system automatically acquires structured and unstructured data through the data acquisition module, constructs a knowledge graph, processes and updates text information in real time using the large language model, and optimizes responses by combining the feedback optimization module.
The system enables real-time updates of the question-and-answer system, with responses automatically adjusted as new or modified files are added, avoiding the lag of historical data and improving the accuracy and efficiency of responses.
Smart Images
Figure CN122045430A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent question answering technology, specifically to an intelligent question answering system based on a large language model. Background Technology
[0002] A management platform is an information management system designed for organizations at all levels. It aims to leverage modern information technology to improve team building efficiency and enhance personnel education, management, and service capabilities. By integrating multiple functional modules, this platform supports teams in areas such as organizational development, personnel management, learning and education, activity execution, and supervision and evaluation, thereby achieving the digitalization, intelligentization, and standardization of team building work.
[0003] However, current management platforms have low levels of informatization and rely on menu-based navigation for inquiries. Users need to be familiar with the operation path to complete the inquiry, which is time-consuming for new users. Therefore, it is not user-friendly for new users. To reduce the impact of menu-based navigation on new users, patent application number 202510194353.8 discloses a corporate knowledge question-and-answer system based on a large language model, involving the field of text data processing and analysis technology. This system establishes user inquiry nodes to collect corresponding corporate knowledge inquiry questions from users, protects the privacy of the user inquiry node collection environment, performs feature analysis on each user's corporate knowledge inquiry questions to obtain several feature keywords for each question, and establishes a feature knowledge graph of these keywords for each question. A large language model is then constructed to analyze the corporate knowledge inquiry questions. The large language model is input into the feature knowledge graph of each corporate knowledge inquiry question to obtain a reference answer set for each question. The reference answer set for each question is then ranked by knowledge value, and the final question answer set required for the corporate knowledge inquiry question is extracted from the reference answer set. However, in this patent, the answers to the consultation questions are provided by an artificial intelligence model trained on historical response data. This means that, firstly, the response content is based on historical data and will not be updated in real time over time; secondly, if the historical response content is incorrect, the output response will also be incorrect. Summary of the Invention
[0004] The present invention aims to provide an intelligent question-answering system based on a large language model, in order to solve the problem that current question-answering systems rely on historical response data and cannot adjust and update the content of the responses in real time.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a smart question-answering system based on a large language model, including a data acquisition module, a knowledge construction module, and an intelligent question-answering module; The data acquisition module is used to collect structured data, unstructured data, and special data to form a text library; the structured data includes personnel information forms, organizational life records, and performance indicators; the unstructured data includes policy documents, meeting minutes, and rectification reports; and the special data includes local team building cases and personalized system documents. The knowledge construction module is used to process text information in the text library and construct a knowledge graph through a large language model; text information processing includes extracting key elements of text data in the text library, eliminating data ambiguity through semantic analysis, and key elements are used to provide a foundation for constructing the knowledge graph. The key element extraction method for text data in the text library is as follows: A1 utilizes NLP techniques from large language models and the semantic logic of policy texts to extract nouns, verbs, and adjectives, while excluding stop words with no actual semantic meaning and low-frequency irrelevant words. A2 uses word frequency and inverse document frequency to calculate weights, filtering words in the text library whose text data exceeds a threshold; the weight calculation formula is:
[0006] Where Value represents the weight, TF(t,d) represents the word frequency of the word sequence t in the information paradigm text d, and IDF(t) represents the inverse document frequency of the word sequence t in the entire corpus; A3, process-oriented policies, extract the three elements of steps, subjects, and time limits; standard-oriented policies, extract indicators, values, and evaluation standards. The intelligent question-answering module is used to retrieve related knowledge through a knowledge graph, then generate standardized text answers from the retrieved related knowledge and display them. At the same time, it displays the navigation path of the index node and provides prompts on the time and popularity of the text on which the answer is based.
[0007] Furthermore, the data acquisition module includes an automatic acquisition module and a manual supplementation module. The automatic acquisition module connects to the OA system, document management system and business system through a large language model, and uses API interfaces or web crawling technology to obtain data in batches to achieve automated data acquisition. The manual supplementation module is used to import Excel files and local files.
[0008] Furthermore, the map construction includes the following steps: S1. Define knowledge nodes and extract core entities from the text information processing results. Core entities are words or phrases that reflect the core semantics of the text and have key significance. Combine semantic analysis, domain knowledge rules and core entity feature matching and label their type as policy, process or case. S2. Establish logical connections between core entities, classify core entities by labeled core entity types, analyze the semantics of core entities using a large language model, extract predicate relationships between core entities, and form triples. S3, Attribute Calibration, checks whether the information description of the core entity is consistent with the original text in the text library, and automatically corrects any inconsistent information descriptions; S4, Knowledge Graph Storage and Indexing, uses a graph database to store the knowledge graph, with nodes representing core entities, edges representing relationships, and attribute tables recording detailed information; S5, establish a multi-dimensional index.
[0009] Furthermore, the knowledge construction module is also used to automatically collect and update knowledge content from the automatic collection module, use semantic comparison algorithms to identify newly added or changed policy documents, automatically perform incremental updates to the text library, and insert newly added knowledge content into the corresponding nodes of the knowledge graph in real time.
[0010] Furthermore, it also includes a feedback optimization module, which is used to rate the answers output by the intelligent question-answering module as satisfactory / unsatisfactory and to manually annotate information errors and missing content.
[0011] Working principle and beneficial effects of the present invention: This solution utilizes a data acquisition module that connects to OA and document management systems via a large language model. It employs API interfaces or web crawling to acquire data in batches, and a manual supplementation module imports Excel files and local files to create a text library. This library includes personnel information forms, organizational life records, performance indicators, policy documents, meeting minutes, rectification reports, local team building cases, and personalized policy documents. Next, a knowledge construction module processes the text library using the large language model, extracting key elements and eliminating ambiguity through semantic analysis. Then, it constructs a knowledge graph through steps such as defining knowledge nodes and labeling their types, establishing logical connections between entities to form triples, attribute calibration, storage and indexing, and building multi-dimensional indexes. Finally, an intelligent question-answering module uses the knowledge graph to retrieve related knowledge and generates standardized text answers. Finally, a feedback optimization module collects user satisfaction / dissatisfaction ratings for answers, as well as feedback on manually labeled errors and missing content, to optimize the functionality of each module, forming a closed loop.
[0012] 1. Compared to existing technologies, the current management structure uses a pre-defined tree hierarchy, requiring users to click navigation nodes level by level. The path is fixed and relies on manual pre-classification. If a new policy is added, its position within the hierarchy must be manually determined; otherwise, users cannot find it through navigation. This solution, however, uses knowledge graph-driven management navigation. The management structure is built on a network of knowledge entities and their relationships. There is no need for a pre-defined fixed hierarchy. When a new policy is added, the knowledge construction module periodically scans newly released content across the policy network using a large language model. Semantic comparison algorithms identify newly added or changed policy documents, automatically updating the text library incrementally and inserting the new knowledge content into the corresponding nodes of the knowledge graph in real time. The system automatically associates the new content with relevant entities through entity recognition, allowing users to access the system from any associated entity. The navigation path is dynamically generated based on entity relationships. This not only provides a quick response but also automatically adjusts and updates the response content based on newly added or changed documents.
[0013] 2. This solution integrates with the enterprise's OA system, document management system, and business systems. It uses message queues to synchronize changes to structured data such as personnel information and performance indicators, avoiding historical data lag and thus preventing delays in responding to questions.
[0014] 3. This solution forms a real-time update capability covering the entire process of knowledge acquisition, knowledge update, and response optimization through a real-time policy scanning and data synchronization mechanism in the knowledge construction module, and an instant response to user feedback and dynamic fine-tuning mechanism in the feedback optimization module. This completely solves the problem that traditional question-answering systems rely on historical data and cannot be dynamically adjusted over time. Attached Figure Description
[0015] Figure 1 This is a system architecture diagram of the intelligent question-answering system based on a large language model according to the present invention. Detailed Implementation
[0016] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 As shown: The intelligent question-answering system based on a large language model includes a data acquisition module, a knowledge construction module, an intelligent question-answering module, and a feedback optimization module.
[0017] The data acquisition module is used to collect structured data, unstructured data, and special data to form a text library.
[0018] Structured data includes personnel information forms, organizational life records, and performance indicators.
[0019] Unstructured data includes policy documents, meeting minutes, and rectification reports.
[0020] Specialized data includes local team building case studies and personalized institutional documents.
[0021] The data acquisition module includes an automatic acquisition module and a manual supplementation module. The automatic acquisition module connects to the OA system, document management system and business system through a large language model, and uses API interfaces or web crawling technology to obtain data in batches to achieve automated data acquisition. The manual supplementation module supports uploading Excel files and local files, adapting to the data entry needs of grassroots units for non-system data.
[0022] The knowledge construction module is used to process text information from a text database and construct a knowledge graph using a large language model; text information processing includes the following steps: A1 utilizes NLP technology based on large language models to extract key elements from policy documents, such as nouns, verbs, and adjectives, while excluding stop words and low-frequency irrelevant words that have no actual semantic meaning. It also automatically extracts standardized fields such as meeting type, frequency of meetings, and participants from the documents.
[0023] A2 uses word frequency and inverse document frequency to calculate weights, filtering words in the text library whose text data exceeds a threshold; the weight calculation formula is:
[0024] Where Value represents the weight, TF(t,d) represents the word frequency of the word sequence t in the information paradigm text d, and IDF(t) represents the inverse document frequency of the word sequence t in the entire corpus.
[0025] A3, process-oriented policies, extract the three elements of steps, subjects, and time limits; standard-oriented policies, extract indicators, values, and evaluation standards. A4 eliminates textual ambiguity through semantic analysis.
[0026] A5 adds semantic tags to text database data and associates them with policy references.
[0027] Map construction includes the following steps: S1. Define knowledge nodes and extract core entities from the text information processing results. Core entities are words or phrases that reflect the core semantics of the text and have key significance. Combine semantic analysis, domain knowledge rules and core entity feature matching, such as personnel development and inspection and supervision, and label their types as policies, processes or cases.
[0028] S2. Establish logical connections between core entities, classify entities by labeled core entity types, analyze the semantics of core entities using a large language model, extract predicate relationships between core entities, and form triples.
[0029] S3, Attribute Calibration, checks whether the information description of the core entity is consistent with the original text in the text library, and automatically corrects any inconsistent information descriptions.
[0030] S4, Knowledge Graph Storage and Indexing, uses a graph database to store the knowledge graph, with nodes representing core entities, edges representing relationships, and attribute tables recording detailed information.
[0031] S5 uses the graph database Neo4j to store graphs and establish a multi-dimensional index of policies, processes, and cases.
[0032] The intelligent question-answering module is used to retrieve related knowledge through a knowledge graph, then generate standardized text answers from the retrieved related knowledge and display them. It also displays the navigation path of the index node and provides prompts regarding the time and popularity of the text upon which the answer is based. Specific steps include: B1, based on the Transformer architecture, captures contextual information and breaks down complex problems.
[0033] B2: The model associates knowledge based on the semantic association of the question, and then generates standardized text answers from the retrieved related knowledge and displays them. At the same time, it displays the navigation path of the index node and provides prompts on the time and popularity of the text on which the knowledge graph is used to generate the answer.
[0034] The feedback optimization module is used to rate the answers output by the intelligent question-answering module as satisfactory / unsatisfactory and to manually annotate information errors and missing content.
[0035] The feedback optimization module collects user satisfaction ratings for responses, such as incorrect policy years in the answer. Through incremental learning, it adjusts the parameters of the large language model in real time to correct the response content without retraining the entire model; it only fine-tunes the parameters for the knowledge segment corresponding to the feedback question. The specific steps are as follows: C1. User satisfaction / dissatisfaction ratings for the answers output by the intelligent question-answering module, as well as specific feedback such as manually marked errors and missing content.
[0036] C2. Capture user click behavior through front-end tracking technology and transmit evaluation data to the log server of the feedback optimization module in real time.
[0037] C3 performs interaction log analysis, using Nginx logs to parse user interaction patterns based on behavioral data such as user question history, answer click preferences, and frequency of repeated question submissions. For example, if a statistician asks questions about development materials 100 times per week and the number of clicks on the answers is below average, it is judged as a knowledge gap.
[0038] C4. Utilize a large language model to perform word frequency statistics and cluster analysis on interaction logs to identify high-frequency error correction scenarios, such as formatted answers in organizational life records being corrected 30 times per month. Generate a list of high-frequency issues to optimize, sorted by impact scope (e.g., affecting more than 100 users) and urgency.
[0039] C5. Perform semantic analysis on user feedback text, such as the policy year in the answer should be 2023, to locate erroneous nodes in the knowledge base, such as the policy year field; generate knowledge correction instructions by comparing the standard answer with the feedback.
[0040] C6. The feedback optimization module will synchronize correction instructions, such as policy year updates, to the knowledge construction module, triggering knowledge graph node updates.
[0041] The specific implementation process is as follows: This solution utilizes a data acquisition module that connects to OA and document management systems via a large language model. It employs API interfaces or web crawling to acquire data in batches, and a manual supplementation module imports Excel files and local files to create a text library. This library includes personnel information forms, organizational life records, performance indicators, policy documents, meeting minutes, rectification reports, local team building cases, and personalized policy documents. Next, a knowledge construction module processes the text library using the large language model, extracting key elements and eliminating ambiguity through semantic analysis. Then, it constructs a knowledge graph through steps such as defining knowledge nodes and labeling their types, establishing logical connections between entities to form triples, attribute calibration, storage and indexing, and building multi-dimensional indexes. Finally, an intelligent question-answering module uses the knowledge graph to retrieve related knowledge and generates standardized text answers. Finally, a feedback optimization module collects user satisfaction / dissatisfaction ratings for answers, as well as feedback on manually labeled errors and missing content, to optimize the functionality of each module, forming a closed loop.
[0042] Example 2 The difference between Example 2 and Example 1 is that the data acquisition module is also used to acquire user facial images, and the knowledge construction module further includes an information push module. The information push module is used to identify the user based on the facial image compared to the content of the personnel information table, convert the user's occupation, location, work history, and enterprise or unit type into identity tags, and then push information based on the degree of matching between the personnel identity tags and the file features in the text library. The specific method is as follows: D1. Parse the personnel information table and extract key information. For example, for company personnel, obtain information such as company type, industry, and size; for individuals, extract information such as occupation, location, and work history. Then, convert this information into tags, such as "manufacturing company," "micro-enterprise," "teacher," and "resident." These tags will serve as the basis for recommendations.
[0043] D2. Classify policies and documents according to dimensions such as theme, target audience, and industry, and use natural language processing technology to extract key features of the documents, including keywords, core content summary, and applicable conditions.
[0044] D3. Establish association rules between identity and recommended content, and at the same time use collaborative filtering algorithms to analyze the policy documents that people with similar identity characteristics have browsed and followed in the past, and recommend similar content to the current user.
Claims
1. A smart question-answering system based on a large language model, characterized by: It includes a data acquisition module, a knowledge construction module, and an intelligent question answering module; The data acquisition module is used to collect structured data, unstructured data, and special data to form a text library; the structured data includes personnel information forms, organizational life records, and performance indicators; the unstructured data includes policy documents, meeting minutes, and rectification reports; and the special data includes local team building cases and personalized system documents. The knowledge construction module is used to process text information in the text library and construct a knowledge graph through a large language model; text information processing includes extracting key elements of text data in the text library, eliminating data ambiguity through semantic analysis, and key elements are used to provide a foundation for constructing the knowledge graph. The key element extraction method for text data in the text library is as follows: A1 utilizes NLP techniques from large language models and the semantic logic of policy texts to extract nouns, verbs, and adjectives, while excluding stop words with no actual semantic meaning and low-frequency irrelevant words. A2 uses word frequency and inverse document frequency to calculate weights, filtering words in the text library whose text data exceeds a threshold; the weight calculation formula is: Where Value represents the weight, TF(t,d) represents the word frequency of the word sequence t in the information paradigm text d, and IDF(t) represents the inverse document frequency of the word sequence t in the entire corpus; A3, process-oriented policies, extract the three elements of steps, subjects, and time limits; standard-oriented policies, extract indicators, values, and evaluation standards. The intelligent question-answering module is used to retrieve related knowledge through a knowledge graph, then generate standardized text answers from the retrieved related knowledge and display them. At the same time, it displays the navigation path of the index node and provides prompts on the time and popularity of the text on which the knowledge graph is used to generate the answer. The data acquisition module is also used to acquire user facial images. The knowledge construction module also includes an information push module. The information push module is used to identify the user based on the facial image and the contents of the personnel information table. It converts the user's occupation, location, work history, enterprise type or unit type into identity tags. Then, it pushes information based on the degree of matching between the personnel identity tags and the features of the documents in the text library, and pushes relevant information based on the historical question popularity of people with the same identity tags.
2. The intelligent question-answering system based on a large language model according to claim 1, characterized in that: The data acquisition module includes an automatic acquisition module and a manual supplementation module. The automatic acquisition module connects to the OA system, document management system and business system through a large language model, and uses API interfaces or web crawling technology to obtain data in batches to achieve automated data acquisition. The manual supplementation module is used to import Excel files and local files.
3. The intelligent question-answering system based on a large language model according to claim 2, characterized in that: The map construction includes the following steps: S1. Define knowledge nodes and extract core entities from the text information processing results. Core entities are words or phrases that reflect the core semantics of the text and have key significance. Combine semantic analysis, domain knowledge rules and core entity feature matching and label their type as policy, process or case. S2. Establish logical connections between core entities, classify core entities by labeled core entity types, analyze the semantics of core entities using a large language model, extract predicate relationships between core entities, and form triples. S3, Attribute Calibration, checks whether the information description of the core entity is consistent with the original text in the text library, and automatically corrects any inconsistent information descriptions; S4, Knowledge Graph Storage and Indexing, uses a graph database to store the knowledge graph, with nodes representing core entities, edges representing relationships, and attribute tables recording detailed information; S5, establish a multi-dimensional index.
4. The intelligent question-answering system based on a large language model according to claim 3, characterized in that: The knowledge construction module is also used to automatically collect and update knowledge content from the automatic collection module, use semantic comparison algorithms to identify newly added or changed policy documents, automatically perform incremental updates to the text library, and insert newly added knowledge content into the corresponding nodes of the knowledge graph in real time.
5. The intelligent question-answering system based on a large language model according to claim 4, characterized in that: It also includes a feedback optimization module, which is used to rate the answers output by the intelligent question-answering module as satisfactory / unsatisfactory and to manually mark feedback information such as errors and missing content.