Method for implementing knowledge graph intelligent question answering system
By combining a large language model and a hierarchical dynamic knowledge graph with reinforcement learning, this question-answering system addresses the shortcomings of existing systems in semantic generalization and real-time updates, achieving flexible interaction and efficient response, and providing interpretable answers.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 陈世恩
- Filing Date
- 2026-03-07
- Publication Date
- 2026-06-09
Smart Images
Figure CN122173605A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence, and in particular to a method for implementing a knowledge graph-based intelligent question-answering system. Background Technology
[0002] Knowledge graph intelligent question answering is a cutting-edge direction in the field of artificial intelligence that deeply integrates structured knowledge (knowledge graph) with natural language processing technology. It transforms users' natural language questions into precise queries on entities, relationships, and attributes in the graph through semantic parsing. This not only returns factual answers but also enables multi-hop reasoning and logical deduction based on the rich semantic associations of the graph.
[0003] Currently, while knowledge graph-based question-answering systems are widely used in fields such as intelligent customer service, they face two major bottlenecks: weak semantic generalization and lagging knowledge updates. First, traditional methods rely excessively on predefined templates, rule matching, or fixed semantic parsers, making it difficult to handle diverse, colloquial, and syntactically complex natural language expressions. Once a question deviates from the preset pattern (e.g., using synonyms or ellipses), the accuracy of intent recognition drops significantly, leading to entity mapping failures and irrelevant answers. Even with the introduction of deep learning, it remains ineffective in niche domains lacking labeled data. Second, knowledge graph construction often employs offline processing, which is cumbersome and time-consuming, failing to capture the latest information in high-frequency, changing fields such as news and finance, resulting in outdated or incorrect answers. Simultaneously, the high cost and complexity of continuous model learning severely restrict the system's adaptability in dynamic scenarios.
[0004] In summary, existing systems urgently need breakthroughs in flexible interaction and real-time performance to resolve the contradiction between static knowledge bases and dynamic real-world needs. Summary of the Invention
[0005] The purpose of this invention is to provide a method for implementing a knowledge graph-based intelligent question answering system, which solves the problems in existing knowledge graph-based question answering systems, such as weak semantic generalization ability and difficulty in coping with diverse natural language expressions due to over-reliance on predefined templates and rules, and the lag in knowledge updates and inability to respond to dynamic information changes in real time due to offline construction mechanisms.
[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:
[0007] The method for implementing a knowledge graph-based intelligent question answering system is characterized by including: receiving natural language questions input by users, performing template-free semantic parsing using a domain-adapted large language model, and directly generating a structured query representation containing intent recognition, entity slots, and logical constraints. The large language model adapts to colloquial and long-tailed sentence expressions through a few-shot prompt learning strategy.
[0008] Based on the structured query representation, entity linking and relationship mapping are performed in the hierarchical dynamic knowledge graph. If the target entity or relationship is detected to be missing in the graph or the confidence level is lower than a preset threshold, the real-time knowledge acquisition mechanism is automatically triggered.
[0009] The real-time knowledge acquisition mechanism retrieves the latest information from external data sources through an asynchronous streaming pipeline, uses an incremental knowledge extraction algorithm to instantly generate new triples with timestamps, and integrates them into the real-time hotspot layer of the hierarchical dynamic knowledge graph.
[0010] Based on the updated hierarchical dynamic knowledge graph, a reinforcement learning-guided graph neural network inference engine is used to perform multi-hop path search to obtain a set of candidate answers containing explicit logical chains;
[0011] The candidate answer set is reconstructed using natural language using a generative response model, and a fact consistency verification module is introduced to verify the logical consistency between the generated content and the graph data, and the final answer is output.
[0012] As an improvement, a dynamic semantic space mapping model is constructed to encode user questions into high-dimensional semantic vectors and calculate their similarity with entity label vectors and relation type vectors in the knowledge graph pattern layer.
[0013] An adversarial training mechanism is introduced to construct an augmented dataset containing synonym replacement, sentence restructuring, elliptical sentences, and noise interference. This dataset is then used to continuously fine-tune the domain adaptation layer of the large language model to eliminate the dependence on predefined rule templates and improve semantic generalization ability.
[0014] For out-of-vocabulary (OOV) words, the generative reasoning capabilities of large language models are used to predict their possible graph semantic correspondences, and they are automatically added to the database or manually confirmed through confidence assessment.
[0015] As an improvement, the hierarchical dynamic knowledge graph adopts a three-level storage architecture, including:
[0016] Static base layer: Relatively stable ontology definitions, common-sense knowledge, and historical archived data within the storage domain;
[0017] Dynamic business layer: Stores periodically updated business status data and versioned knowledge snapshots;
[0018] Real-time hotspot layer: Stores frequently changing data that is updated every minute or even every second. This layer is retrieved first during queries and a time decay weighting mechanism is used.
[0019] The incremental knowledge extraction algorithm includes incremental block processing of unstructured text, dynamic event recognition based on event detection model, and generation of triples based on timestamp marking and storage of temporary incremental areas. It also periodically performs entity disambiguation and knowledge fusion operations to merge temporary data into the main graph.
[0020] As an improvement, the operation of the reinforcement learning-guided graph neural network inference engine includes:
[0021] Using structured query representation as the initial state, we simulate multi-step walks of an agent on a knowledge graph subgraph;
[0022] Design a composite reward function to comprehensively evaluate the confidence, logical completeness, and semantic matching degree with the user's intent of each hop, and automatically prune low-confidence paths;
[0023] It supports a hybrid reasoning mode that combines explicit logical deduction with implicit semantic association to generate interpretable multi-hop reasoning paths and solve complex logical jump problems.
[0024] As an improvement, the fact consistency verification module specifically performs the following operations:
[0025] The generated natural language answers are back-mapped into graph query statements or logical forms;
[0026] Perform the reverse query on the knowledge graph to verify whether the returned results are consistent with the factual statements in the generated answer;
[0027] If the verification is inconsistent, a regeneration mechanism is triggered or an uncertainty warning is marked in the final answer. At the same time, the case is recorded for subsequent model optimization.
[0028] As an improvement, this method also includes an adaptive continuous learning mechanism:
[0029] Collect user feedback data on generated answers in real time, including explicit feedback (likes, dislikes, suggestions for improvement) and implicit feedback (stay time, follow-up questions).
[0030] When the proportion of negative feedback exceeds the preset threshold or a new error pattern is detected, the model update process is automatically triggered, and the error cases and corrected data are added to the training set.
[0031] The Parametric Efficient Fine-Tuning (PEFT) technique is used to update only some key parameters of the large language model and graph neural network, thereby reducing the computational cost of continuous learning and preventing catastrophic forgetting.
[0032] As an improvement, the entity linking and relationship mapping adopts a two-stage coarse-fine matching strategy:
[0033] The first stage utilizes a fuzzy matching algorithm based on edit distance and vector retrieval to quickly recall the candidate entity set;
[0034] The second stage utilizes a context-aware reordering model, combining global semantic information of the user's question with graph neighborhood structure information to refine the ranking of candidate entities, thus solving the problems of polysemy and referential resolution.
[0035] For sparse entities and relations, we utilize graph neural networks and their neighborhood information to enhance vector representations, thereby alleviating the cold start problem in mapping long-tailed entities.
[0036] As an improvement, this method also includes a multimodal knowledge fusion step:
[0037] It supports extracting multimodal features from images, tables, and unstructured documents, and constructing multimodal triples containing visual, textual, and structured data;
[0038] When processing user questions, if an intent involving multimodal entities is identified, the corresponding visual parsing or table understanding module is automatically invoked to assist in semantic parsing, and the multimodal retrieval results are integrated into the candidate answer set.
[0039] As an improvement, the generative response model is also responsible for generating explanatory text containing reasoning chains:
[0040] Extract key nodes and relationships from multi-hop reasoning paths and transform them into logical deduction processes described in natural language.
[0041] Presenting the logical derivation process and the final factual answer together to the user achieves explainable question answering, enhancing the user's trust in the system's response.
[0042] A knowledge graph fusion-based intelligent question-answering system, characterized by comprising:
[0043] The semantic parsing module is used to perform template-free semantic understanding and generate structured query representations;
[0044] The dynamic graph management module is used to perform entity linking, real-time knowledge acquisition, incremental extraction, and hierarchical graph updating.
[0045] The inference engine module is used to perform the reinforcement learning-based multi-hop inference path search.
[0046] The answer generation and verification module is used to perform answer reconstruction, factual consistency verification, and explanatory text generation.
[0047] The continuous learning and optimization module is used to perform feedback collection and efficient parameter fine-tuning.
[0048] The method comprises a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program to implement the method as described in any one of claims 1 to 9.
[0049] The beneficial effects of this invention are: extremely strong semantic generalization ability: abandoning traditional template matching, it utilizes a domain-adapted large language model and few-sample learning to accurately understand colloquialisms, elliptical sentences and long-tail problems, and significantly improves the intent recognition rate.
[0050] Real-time dynamic response: The unique hierarchical dynamic knowledge graph architecture, combined with the asynchronous streaming knowledge acquisition mechanism, realizes the leap from "offline construction" to "real-time update", ensuring the timeliness of the response.
[0051] High credibility and interpretability: By generating explicit logical chains through multi-hop reasoning guided by reinforcement learning and combining them with a fact consistency verification module, the "illusion" problem of large models is effectively solved, enhancing users' trust in the system.
[0052] Low-cost continuous evolution: By introducing an adaptive continuous learning mechanism and efficient parameter fine-tuning technology, the system can iterate itself at low cost based on user feedback and adapt to constantly changing business scenarios. Attached Figure Description
[0053] Figure 1 This is an overall flowchart of the intelligent question-answering method for knowledge graphs in this invention.
[0054] Figure 2 This is an architecture diagram of the knowledge graph intelligent question answering system of this invention.
[0055] Figure 3 This is a diagram of the three-level storage architecture of the hierarchical dynamic knowledge graph in the intelligent question-answering method of the knowledge graph of the present invention. Detailed Implementation
[0056] To make the content of this invention easier to understand, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Identical components are represented by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, while the terms "inner" and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.
[0057] like Figures 1 to 2As shown, the method for implementing a knowledge graph-based intelligent question-answering system is characterized by the following steps: receiving natural language questions input by users; performing template-free semantic parsing using a domain-adapted large language model to directly generate a structured query representation containing intent recognition, entity slots, and logical constraints; wherein the large language model adapts to colloquial and long-tailed sentence expressions through a few-shot prompt learning strategy; and performing entity linking and relationship mapping in a hierarchical dynamic knowledge graph based on the structured query representation. If a target entity or relationship is detected to be missing in the graph or its confidence level is lower than a preset threshold, a real-time knowledge acquisition mechanism is automatically triggered. The real-time knowledge acquisition mechanism captures the latest information from external data sources through an asynchronous streaming pipeline, generates new timestamped triples in real time using an incremental knowledge extraction algorithm, and integrates them into the real-time hotspot layer of the hierarchical dynamic knowledge graph. Based on the updated hierarchical dynamic knowledge graph, a reinforcement learning-guided graph neural network inference engine performs multi-hop path search to obtain a set of candidate answers containing explicit logical chains. The generative answer model reconstructs the candidate answer set using natural language, and a fact consistency verification module is introduced to verify the logical consistency between the generated content and the graph data, outputting the final answer.
[0058] A dynamic semantic space mapping model is constructed to encode user questions into high-dimensional semantic vectors and calculate their similarity with entity label vectors and relation type vectors in the knowledge graph pattern layer. An adversarial training mechanism is introduced to construct an augmented dataset containing synonym replacement, sentence recombination, ellipsis, and noise interference. This dataset is continuously fine-tuned for the domain adaptation layer of the large language model to eliminate the dependence on predefined rule templates and improve semantic generalization ability. For out-of-vocabulary (OOV) words, the generative reasoning ability of the large language model is used to predict their possible graph semantic correspondences, and they are automatically entered into the database or manually confirmed through confidence evaluation.
[0059] The hierarchical dynamic knowledge graph adopts a three-level storage architecture, including: a static base layer: storing relatively stable ontology definitions, common sense knowledge, and historical archived data within the domain; a dynamic business layer: storing periodically updated business status data and versioned knowledge snapshots; and a real-time hotspot layer: storing frequently changing data updated at the minute or even second level by crawling, with this layer being retrieved first during queries and employing a time decay weight mechanism. The incremental knowledge extraction algorithm includes incremental block processing of unstructured text, dynamic event recognition based on an event detection model, and the generation of triples based on timestamps and storage of temporary incremental areas. Entity disambiguation and knowledge fusion operations are periodically performed to merge temporary data into the main graph.
[0060] The working process of the reinforcement learning-guided graph neural network inference engine includes: using the structured query representation as the initial state, simulating multi-step walks of the agent on the knowledge graph subgraph; designing a composite reward function to comprehensively evaluate the path confidence, logical completeness, and semantic matching degree with the user intent of each walk, and automatically pruning low-confidence paths; supporting a hybrid reasoning mode that combines explicit logical deduction and implicit semantic association to generate interpretable multi-hop reasoning paths and solve complex logical jump problems.
[0061] The fact consistency verification module performs the following operations: it reverse-maps the generated natural language answer to a graph query statement or logical form; it executes the reverse query on the knowledge graph to verify whether the returned result is consistent with the factual statement in the generated answer; if the verification is inconsistent, it triggers a regeneration mechanism or marks an uncertainty prompt in the final answer, and records the case for subsequent model optimization.
[0062] The method also includes an adaptive continuous learning mechanism: real-time collection of user feedback data on generated answers, including explicit feedback (likes, dislikes, correction suggestions) and implicit feedback (stay time, follow-up questions); when the proportion of negative feedback exceeds a preset threshold or a new error pattern is detected, the model update process is automatically triggered, adding error cases and correction data to the training set; and the parameter efficient fine-tuning (PEFT) technique is used to update only some key parameters of the large language model and graph neural network to reduce the computational cost of continuous learning and prevent catastrophic forgetting.
[0063] The entity linking and relation mapping adopts a two-stage coarse-fine matching strategy: the first stage uses a fuzzy matching algorithm based on edit distance and vector retrieval to quickly recall the candidate entity set; the second stage uses a context-aware reordering model, combined with the global semantic information of the user question and the graph neighborhood structure information, to finely rank the candidate entities and solve the problems of polysemy and referential resolution; for sparse entities and relations, the graph neural network and its neighborhood information are used to enhance the vector representation to alleviate the cold start problem of mapping long-tail entities.
[0064] The method also includes a multimodal knowledge fusion step: it supports the extraction of multimodal features from images, tables and unstructured documents, and the construction of multimodal triples containing visual, text and structured data; when processing user questions, if an intent involving multimodal entities is identified, the corresponding visual parsing or table understanding module is automatically invoked to assist semantic parsing, and the multimodal retrieval results are integrated into the candidate answer set.
[0065] The generative answer model is also responsible for generating explanatory text containing reasoning chains: extracting key nodes and relationships in multi-hop reasoning paths and transforming them into logical deduction processes described in natural language; presenting the logical deduction process and the final factual answer to the user to achieve explainable question answering and enhance the user's trust in the system's answers.
[0066] A knowledge graph fusion-based intelligent question-answering system, characterized by comprising: a semantic parsing module for performing template-free semantic understanding and structured query representation generation; a dynamic graph management module for performing entity linking, real-time knowledge acquisition, incremental extraction, and hierarchical graph updating; an inference engine module for performing reinforcement learning-based multi-hop inference path search; an answer generation and verification module for performing answer reconstruction, factual consistency verification, and explanatory text generation; a continuous learning optimization module for performing feedback collection and efficient parameter fine-tuning; and a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1 to 9.
[0067] In implementation, the system first receives natural language questions input by users. It then performs template-free semantic parsing using a domain-adapted large language model, directly generating a structured query representation containing intent, entities, and logical constraints. Next, entity links and relationship mappings are performed within a hierarchical dynamic knowledge graph (comprising a static base layer, a dynamic business layer, and a real-time hotspot layer). If knowledge gaps are detected, a real-time knowledge acquisition mechanism is automatically triggered, fetching the latest information from external data sources and immediately integrating it into the real-time hotspot layer. Subsequently, a reinforcement learning-guided graph neural network inference engine performs multi-hop path search to obtain candidate answers containing logical chains. Finally, a generative answer model reconstructs the candidate answers using natural language, verifies the accuracy of the content through a fact consistency verification module, and outputs the final interpretable answer. Throughout the process, user feedback is collected through an adaptive continuous learning mechanism, and the model is continuously optimized using efficient parameter fine-tuning techniques, effectively addressing the problems of weak semantic generalization and lagging knowledge updates in traditional question-answering systems.
[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for implementing a knowledge graph-based intelligent question-answering system, characterized in that: include: The system receives natural language questions input by users, performs template-free semantic parsing using a domain-adapted large language model, and directly generates a structured query representation that includes intent recognition, entity slots, and logical constraints. The large language model adapts to colloquial and long-tailed sentence expressions through a few-shot prompt learning strategy. Based on the structured query representation, entity linking and relationship mapping are performed in the hierarchical dynamic knowledge graph. If the target entity or relationship is detected to be missing in the graph or the confidence level is lower than a preset threshold, the real-time knowledge acquisition mechanism is automatically triggered. The real-time knowledge acquisition mechanism retrieves the latest information from external data sources through an asynchronous streaming pipeline, uses an incremental knowledge extraction algorithm to instantly generate new triples with timestamps, and integrates them into the real-time hotspot layer of the hierarchical dynamic knowledge graph. Based on the updated hierarchical dynamic knowledge graph, a reinforcement learning-guided graph neural network inference engine is used to perform multi-hop path search to obtain a set of candidate answers containing explicit logical chains; The candidate answer set is reconstructed using natural language using a generative response model, and a fact consistency verification module is introduced to verify the logical consistency between the generated content and the graph data, and the final answer is output.
2. The method for implementing a knowledge graph-based intelligent question-answering system according to claim 1, characterized in that, A dynamic semantic space mapping model is constructed to encode user questions into high-dimensional semantic vectors and calculate their similarity with entity label vectors and relation type vectors in the knowledge graph pattern layer. An adversarial training mechanism is introduced to construct an augmented dataset containing synonym replacement, sentence restructuring, elliptical sentences, and noise interference. This dataset is then used to continuously fine-tune the domain adaptation layer of the large language model to eliminate the dependence on predefined rule templates and improve semantic generalization ability. For out-of-vocabulary (OOV) words, the generative reasoning capabilities of large language models are used to predict their possible graph semantic correspondences, and they are automatically added to the database or manually confirmed through confidence assessment.
3. The method for implementing a knowledge graph-based intelligent question-answering system according to claim 1, characterized in that, The hierarchical dynamic knowledge graph adopts a three-level storage architecture, including: Static base layer: Relatively stable ontology definitions, common-sense knowledge, and historical archived data within the storage domain; Dynamic business layer: Stores periodically updated business status data and versioned knowledge snapshots; Real-time hotspot layer: Stores frequently changing data that is updated every minute or even every second. This layer is retrieved first during queries and a time decay weighting mechanism is used. The incremental knowledge extraction algorithm includes incremental block processing of unstructured text, dynamic event recognition based on event detection model, and generation of triples based on timestamp marking and storage of temporary incremental areas. It also periodically performs entity disambiguation and knowledge fusion operations to merge temporary data into the main graph.
4. The method for implementing a knowledge graph-based intelligent question-answering system according to claim 1, characterized in that, The operation of the reinforcement learning-guided graph neural network inference engine includes: Using structured query representation as the initial state, we simulate multi-step walks of an agent on a knowledge graph subgraph; Design a composite reward function to comprehensively evaluate the confidence, logical completeness, and semantic matching degree with the user's intent of each hop, and automatically prune low-confidence paths; It supports a hybrid reasoning mode that combines explicit logical deduction with implicit semantic association to generate interpretable multi-hop reasoning paths and solve complex logical jump problems.
5. The method for implementing a knowledge graph-based intelligent question-answering system according to claim 1, characterized in that, The fact consistency verification module specifically performs the following operations: The generated natural language answers are back-mapped into graph query statements or logical forms; Perform the reverse query on the knowledge graph to verify whether the returned results are consistent with the factual statements in the generated answer; If the verification is inconsistent, a regeneration mechanism is triggered or an uncertainty warning is marked in the final answer. At the same time, the case is recorded for subsequent model optimization.
6. The method for implementing a knowledge graph intelligent question-answering system according to claim 5, characterized in that, This method also includes an adaptive continuous learning mechanism: Collect user feedback data on generated answers in real time, including explicit feedback (likes, dislikes, suggestions for improvement) and implicit feedback (stay time, follow-up questions). When the proportion of negative feedback exceeds the preset threshold or a new error pattern is detected, the model update process is automatically triggered, and the error cases and corrected data are added to the training set. The Parametric Efficient Fine-Tuning (PEFT) technique is used to update only some key parameters of large language models and graph neural networks, thereby reducing the computational cost of continuous learning and preventing catastrophic forgetting.
7. The method for implementing a knowledge graph-based intelligent question-answering system according to claim 2, characterized in that, The entity linking and relation mapping adopts a two-stage coarse-fine matching strategy: The first stage utilizes a fuzzy matching algorithm based on edit distance and vector retrieval to quickly recall the candidate entity set; The second stage utilizes a context-aware reordering model, combining global semantic information of the user's question with graph neighborhood structure information to refine the ranking of candidate entities, thus solving the problems of polysemy and referential resolution. For sparse entities and relations, we utilize graph neural networks and their neighborhood information to enhance vector representations, thereby alleviating the cold start problem in mapping long-tailed entities.
8. The method for implementing a knowledge graph-based intelligent question-answering system according to claim 1, characterized in that, This method also includes a multimodal knowledge fusion step: It supports extracting multimodal features from images, tables, and unstructured documents, and constructing multimodal triples containing visual, textual, and structured data; When processing user questions, if an intent involving multimodal entities is identified, the corresponding visual parsing or table understanding module is automatically invoked to assist in semantic parsing, and the multimodal retrieval results are integrated into the candidate answer set.
9. The method for implementing a knowledge graph intelligent question-answering system according to claim 4, characterized in that, The generative response model is also responsible for generating explanatory text containing reasoning chains: Extract key nodes and relationships from multi-hop reasoning paths and transform them into logical deduction processes described in natural language. Presenting the logical derivation process and the final factual answer together to the user achieves explainable question answering, enhancing the user's trust in the system's response.
10. An intelligent question-answering system based on knowledge graph fusion, characterized in that, include: The semantic parsing module is used to perform template-free semantic understanding and generate structured query representations; The dynamic graph management module is used to perform entity linking, real-time knowledge acquisition, incremental extraction, and hierarchical graph updating. The inference engine module is used to perform the reinforcement learning-based multi-hop inference path search. The answer generation and verification module is used to perform answer reconstruction, factual consistency verification, and explanatory text generation. The continuous learning and optimization module is used to perform feedback collection and efficient parameter fine-tuning. The method comprises a memory and a processor, wherein the memory stores a computer program and the processor executes the computer program to implement the method as described in any one of claims 1 to 9.