RAG question and answer self-optimization method based on spatio-temporal knowledge graph reasoning

Abstract

The invention discloses an RAG question and answer self-optimization method based on spatio-temporal knowledge graph reasoning, which comprises the following steps: 1) acquiring a data source, processing the data source, and constructing a domain knowledge graph; 2) obtaining a user question text, and performing multi-hop logical reasoning and semantic association analysis on the user question text by using a graph neural network based on the space-time knowledge graph to generate a structured reasoning path; 3) collecting explicit feedback and implicit feedback of a user, and generating an optimization signal; and 4) automatically updating the domain knowledge graph and / or adjusting parameters of the graph neural network according to the optimization signal. According to the method, knowledge boundaries can be dynamically expanded, manual annotation is reduced, the multi-hop reasoning capability of complex problems is improved, and synchronous optimization of system response speed and accuracy is realized.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, specifically a self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning. Background Technology

[0002] Current intelligent question-answering systems based on Retrieval-Augmented Generation (RAG) primarily achieve intelligent answering of domain-specific questions through the collaborative work of external knowledge base retrieval and Large Language Models (LLMs). The general process is: knowledge base construction → user question parsing → retrieval of relevant knowledge base fragments → synthesis of the final answer using the LLM model. However, this architecture suffers from the following significant limitations:

[0003] (1) Knowledge retrieval loses relevant segments, resulting in insufficient semantic understanding.

[0004] Traditional text similarity-based retrieval methods are susceptible to keyword interference and lack the ability to reason about the deep semantics and contextual logic of questions. They may only return isolated fragments, leading to a break in the logical chain of generating answers and thus causing semantic bias in the retrieval.

[0005] (2) Collaboration defects of the retrieval and generation modules

[0006] Knowledge-answer consistency vulnerability: Generative models may over-rely on prior knowledge, ignoring the constraints of retrieval results and producing factual errors. When faced with low-frequency or highly specialized questions, sparse retrieval struggles to find relevant documents, and generative models, due to training data bias, tend to output generic but substantive answers.

[0007] (3) Lack of system self-evolution ability

[0008] First, manual maintenance is costly, as knowledge base updates rely on manual annotation by experts, resulting in significant operational costs. Second, feedback utilization is inefficient; while existing RAG systems can collect user feedback, they lack automated mechanisms to transform feedback into incremental optimizations of the knowledge base or model, leading to long iteration cycles for retrieval methods. Third, reasoning and retrieval are disconnected; most methods only use the graph as a retrieval source, failing to embed graph reasoning depth into the generation process, thus hindering the self-optimization capabilities of RAG systems. Summary of the Invention

[0009] The purpose of this invention is to provide a self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning, comprising the following steps:

[0010] 1) Obtain the data source, process the data source, and construct a domain knowledge graph;

[0011] 2) Obtain the user's question text, and based on the spatiotemporal knowledge graph, use a graph neural network to perform multi-hop logical reasoning and semantic association analysis on the user's question text to generate a structured reasoning path;

[0012] 3) Collect explicit and implicit user feedback to generate optimization signals;

[0013] 4) Automatically update the domain knowledge graph and / or adjust the parameters of the graph neural network based on the optimized signal, including the weight matrix of the aggregation function, node update weights, attention metric weights, and global pooling weights.

[0014] Furthermore, the data sources include structured databases, unstructured text, and domain knowledge bases.

[0015] Furthermore, in step 1), the steps for constructing the domain knowledge graph include:

[0016] 1.1) For structured data, entity relationships in relational databases are automatically resolved using predefined mapping rules;

[0017] 1.2) Perform cross-source alignment on the data and use graph embedding to calculate the semantic similarity between entities from different sources to achieve a unified representation of heterogeneous data;

[0018] 1.3) Identify contradictory knowledge based on statistical models and perform semantic disambiguation;

[0019] 1.4) Construct a domain knowledge graph that includes entity-relationship-attribute triples.

[0020] Furthermore, the reasoning path is output in the form of a subgraph or a logic chain.

[0021] Furthermore, in step 2), the steps for generating the structured reasoning path include:

[0022] 2.1) Define domain logic rules and use an improved A* algorithm to search for a set of candidate reasoning paths that satisfy the constraints in the knowledge graph;

[0023] 2.2) The GraphSAGE model is used to aggregate neighbor node information, identify implicit relationships, dynamically adjust the inference path weights using reinforcement learning, and rank the path confidence.

[0024] 2.3) Perform a weighted vote on the path confidence of the candidate inference path set generated in step 2.1) and the candidate inference path set generated in step 2.2), and generate a structured inference path according to the path confidence after weighted voting.

[0025] Furthermore, before performing multi-hop logical reasoning and semantic association analysis on the user question text, the user question text is converted into a graph through a semantic parser.

[0026] Furthermore, the tool for performing multi-hop queries is the graph database Neo4j.

[0027] Furthermore, the graph neural network is the GraphSAGE model.

[0028] Furthermore, in step 3), the steps of collecting explicit and implicit user feedback and generating optimization signals include:

[0029] 3.1) Collect explicit user feedback, including ratings and error corrections; collect implicit user feedback, including clicks and dwell time.

[0030] 3.2) Analyze user explicit and implicit feedback using natural language processing models, extract error types and / or analyze potential problems based on behavior logs, thereby generating an optimized instruction set.

[0031] The optimization instruction set includes optimization instructions related to user style preferences, format preferences, information depth, interaction patterns, expression tendencies, content optimization tendencies, and interactive experience.

[0032] Furthermore, according to the optimization of the signal, the automatic updating of the domain knowledge graph refers to automatically adding new nodes or correcting attribute values ​​based on feedback data.

[0033] The technical effectiveness of this invention is undeniable. This invention proposes an intelligent question-answering system architecture based on spatiotemporal knowledge graph reasoning and closed-loop self-optimization, aiming to address the limitations of traditional Retrieval Augmentation (RAG) technology in terms of insufficient semantic understanding, retrieval-generation coordination defects, and lack of system self-evolution capabilities. By leveraging the structured logical expression and multi-hop reasoning technology of knowledge graphs (KG), the traditional text fragment retrieval of RAG is upgraded to reasoning path retrieval, and a hybrid mechanism of symbolic reasoning and graph neural networks (GraphSAGE) is employed to optimize logical coherence. Simultaneously, the system introduces a dynamic optimization module, using user feedback-driven reinforcement learning to achieve incremental updates of the knowledge graph and adjustment of generation model parameters, forming a closed-loop self-optimization mechanism of "data-model co-evolution." Core innovations include deep coupling of knowledge graphs and RAG, hybrid reasoning strategies, and automated feedback utilization technology. This method significantly improves the multi-hop reasoning capability for complex problems, reduces manual maintenance costs, and achieves simultaneous optimization of dynamic knowledge boundary expansion and system response efficiency, making it suitable for high-precision intelligent question-answering scenarios in professional fields.

[0034] In summary, this invention can dynamically expand knowledge boundaries, reduce manual annotation, improve multi-hop reasoning capabilities for complex problems, and achieve simultaneous optimization of system response speed and accuracy. Attached Figure Description

[0035] Figure 1 This is a flowchart of the method. Detailed Implementation

[0036] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0037] Example 1:

[0038] See Figure 1 A self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning includes the following steps:

[0039] 1) Obtain the data source, process the data source, and construct a domain knowledge graph;

[0040] 2) Obtain the user's question text, and based on the spatiotemporal knowledge graph, use a graph neural network to perform multi-hop logical reasoning and semantic association analysis on the user's question text to generate a structured reasoning path;

[0041] 3) Collect explicit and implicit user feedback to generate optimization signals;

[0042] 4) Automatically update the domain knowledge graph and / or adjust the parameters of the graph neural network based on the optimized signal, including the weight matrix of the aggregation function, node update weights, attention metric weights, and global pooling weights.

[0043] The data sources include structured databases, unstructured text, and domain knowledge bases.

[0044] In step 1), the steps for constructing the domain knowledge graph include:

[0045] 1.1) For structured data, entity relationships in relational databases are automatically resolved using predefined mapping rules;

[0046] 1.2) Perform cross-source alignment on the data and use graph embedding to calculate the semantic similarity between entities from different sources to achieve a unified representation of heterogeneous data;

[0047] 1.3) Identify contradictory knowledge based on statistical models and perform semantic disambiguation;

[0048] 1.4) Construct a domain knowledge graph that includes entity-relationship-attribute triples.

[0049] The reasoning path is output in the form of a subgraph or a logic chain.

[0050] Step 2) involves generating a structured reasoning path, including:

[0051] 2.1) Define domain logic rules and use an improved A* algorithm to search for a set of candidate reasoning paths that satisfy the constraints in the knowledge graph;

[0052] 2.2) The GraphSAGE model is used to aggregate neighbor node information, identify implicit relationships, dynamically adjust the inference path weights using reinforcement learning, and rank the path confidence.

[0053] 2.3) Perform a weighted vote on the path confidence of the candidate inference path set generated in step 2.1) and the candidate inference path set generated in step 2.2), and generate a structured inference path according to the path confidence after weighted voting.

[0054] Before performing multi-hop logical reasoning and semantic association analysis on the user question text, the user question text is converted into a graph by a semantic parser.

[0055] The tool for performing multi-hop queries is the graph database Neo4j.

[0056] The graph neural network is the GraphSAGE model.

[0057] Step 3), which involves collecting explicit and implicit user feedback and generating optimization signals, includes the following steps:

[0058] 3.1) Collect explicit user feedback, including ratings and error corrections; collect implicit user feedback, including clicks and dwell time.

[0059] 3.2) Analyze user explicit and implicit feedback using natural language processing models, extract error types and / or analyze potential problems based on behavior logs, thereby generating an optimized instruction set.

[0060] User feedback includes: user-initiated likes, rating tags, dwell time, skip time, number of repeated searches, etc., which correspond to optimization instruction sets such as user style preferences, format preferences, information depth, interaction mode, expression tendency, content optimization tendency, and interactive experience.

[0061] The concept of automatically updating the domain knowledge graph based on optimized signals refers to automatically adding new nodes or correcting attribute values ​​based on feedback data.

[0062] Example 2:

[0063] A self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning includes the following steps:

[0064] 1) Obtain the data source, process the data source, and construct a domain knowledge graph;

[0065] 2) Obtain the user's question text, and based on the spatiotemporal knowledge graph, use a graph neural network to perform multi-hop logical reasoning and semantic association analysis on the user's question text to generate a structured reasoning path;

[0066] 3) Collect explicit and implicit user feedback to generate optimization signals;

[0067] 4) Automatically update the domain knowledge graph and / or adjust the parameters of the graph neural network based on the optimized signal, including the weight matrix of the aggregation function, node update weights, attention metric weights, and global pooling weights;

[0068] 5) Update the large language model using an optimized knowledge graph, and then use the large language model to achieve intelligent answering tasks related to professional domain knowledge. These tasks include user-uploaded questions, such as whether there are any undeveloped urban construction land parcels in a certain area within the land and resources management field. The large language model, updated using the knowledge graph of the land and resources management field, parses the user-uploaded questions, generates answers (yes / no, specific location of undeveloped urban construction land, etc.), and sends them to the user.

[0069] Example 3:

[0070] A self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning, with the same technical content as in Embodiment 2, further wherein the data source includes structured databases, unstructured text, and domain knowledge bases.

[0071] Example 4:

[0072] A self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning, with the same technical content as any one of embodiments 2-3, further comprising the following steps in step 1):

[0073] 1.1) For structured data, entity relationships in relational databases are automatically resolved using predefined mapping rules;

[0074] 1.2) Perform cross-source alignment on the data and use graph embedding to calculate the semantic similarity between entities from different sources to achieve a unified representation of heterogeneous data;

[0075] 1.3) Identify contradictory knowledge based on statistical models and perform semantic disambiguation;

[0076] 1.4) Construct a domain knowledge graph that includes entity-relationship-attribute triples.

[0077] Example 5:

[0078] A self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning, with the same technical content as any one of embodiments 2-4, further wherein the reasoning path is output in the form of a subgraph or a logic chain.

[0079] Example 6:

[0080] A self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning, with the same technical content as any one of embodiments 2-5, further comprising the following steps in step 2) for generating structured reasoning paths:

[0081] 2.1) Define domain logic rules and use an improved A* algorithm to search for a set of candidate reasoning paths that satisfy the constraints in the knowledge graph;

[0082] 2.2) The GraphSAGE model is used to aggregate neighbor node information, identify implicit relationships, dynamically adjust the inference path weights using reinforcement learning, and rank the path confidence.

[0083] 2.3) Perform a weighted vote on the path confidence of the candidate inference path set generated in step 2.1) and the candidate inference path set generated in step 2.2), and generate a structured inference path according to the path confidence after weighted voting.

[0084] Example 7:

[0085] A self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning, with the same technical content as any one of embodiments 2-6, further wherein, before performing multi-hop logical reasoning and semantic association analysis on the user question text, the user question text is converted into a graph by a semantic parser.

[0086] Example 8:

[0087] A self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning, with the same technical content as any one of embodiments 2-7, further wherein the tool for performing multi-hop queries is the graph database Neo4j.

[0088] Example 9:

[0089] A self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning, with the same technical content as any one of embodiments 2-8, further wherein the graph neural network is a GraphSAGE model.

[0090] Example 10:

[0091] A self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning, with the same technical content as any one of embodiments 2-9, further comprising the step of collecting explicit and implicit user feedback and generating optimization signals in step 3), including:

[0092] 3.1) Collect explicit user feedback, including ratings and error corrections; collect implicit user feedback, including clicks and dwell time.

[0093] 3.2) Analyze user explicit and implicit feedback using natural language processing models to extract error types and / or analyze potential problems based on behavior logs, thereby generating an optimization instruction set. The optimization instruction set includes optimization instructions related to user style preferences, format preferences, information depth, interaction patterns, expression tendencies, content optimization tendencies, and interactive experience.

[0094] Example 11:

[0095] A self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning, with the same technical content as any one of embodiments 2-10, further, automatically updating the domain knowledge graph according to the optimization signal means: automatically adding new nodes or correcting attribute values ​​based on feedback data.

[0096] Example 12:

[0097] A self-optimizing method for RAG question answering based on spatiotemporal knowledge graph reasoning is proposed. This method consists of five parts: knowledge graph construction module, reasoning engine module, RAG question answering module, feedback analysis module, and dynamic optimization module. Through the structured reasoning and closed-loop self-optimization mechanism of knowledge graph, it solves the problems of insufficient knowledge dynamism, logical coherence, and self-evolution ability of traditional RAG systems.

[0098] 1. Knowledge Graph Construction Module

[0099] Function definition:

[0100] Responsible for dynamically constructing domain knowledge graphs from multi-source heterogeneous data (structured databases, unstructured text), and continuously maintaining their integrity and timeliness.

[0101] Technical Implementation:

[0102] (1) Multimodal knowledge extraction: For structured data: Entity relationships in relational databases are automatically parsed using predefined mapping rules. For unstructured text: Entity recognition, relationship extraction, and attribute filling are performed based on a deep learning model.

[0103] (2) Knowledge fusion and conflict resolution: First, cross-source alignment is performed by using graph embedding to calculate the semantic similarity between entities from different sources, thereby achieving a unified representation of heterogeneous data. Second, conflict detection is performed by identifying contradictory knowledge based on logical rules and statistical models such as Bayesian networks.

[0104] 2. Inference Engine Module

[0105] Function definition:

[0106] Based on spatiotemporal knowledge graphs, multi-hop logical reasoning and semantic association analysis are performed to provide structured reasoning paths for the RAG question answering module. The reasoning paths are output in the form of subgraphs or logical chains.

[0107] Technical Implementation:

[0108] (1) Symbolic Reasoning: Rule Engine: Defines domain logic rules. Path Search Algorithm: Employs an improved A* algorithm to search for the optimal reasoning path that satisfies the constraints in the knowledge graph.

[0109] (2) Neural Reasoning: Graph Neural Network: Aggregating neighbor node information through the GraphSAGE model to identify implicit relationships. Multi-hop Reasoning Reinforcement: Dynamically adjusting the weights of reasoning paths using reinforcement learning.

[0110] (3) Hybrid reasoning mechanism: Symbolic reasoning outputs a set of candidate paths, graph neural network sorts the confidence of the paths, and finally generates the reasoning result through weighted voting.

[0111] 3. RAG Q&A Module

[0112] Function definition:

[0113] The traditional RAG "text fragment retrieval" is upgraded to "knowledge graph reasoning path retrieval", and answers that conform to logical constraints are generated based on the reasoning results.

[0114] Technical Implementation:

[0115] (1) Knowledge-enhanced retrieval:

[0116] User queries are converted into graph queries by a semantic parser and then executed using the Neo4j graph database. The queries return relevant subgraphs and inference paths. Path confidence scores are included with the search results.

[0117] (2) Logical constraint generation:

[0118] Injecting the topological structure prior of the reasoning path into the attention layer of the large language model ensures that the generated content follows the logical chain, and outputs supporting evidence synchronously when the answer is generated, while also annotating the key nodes in the reasoning path.

[0119] 4. Dynamic Optimization Module

[0120] Function definition:

[0121] Collect explicit user feedback (ratings, error corrections) and implicit feedback (clicks, dwell time), analyze system defects, and generate optimization signals. Based on the feedback analysis results, automatically update the knowledge graph and adjust the parameters of the generation model to achieve system self-evolution.

[0122] Technical Implementation:

[0123] (1) Feedback Classification:

[0124] By parsing user error correction text through natural language processing, error types can be extracted (such as factual errors, logical errors, and unclear expressions), or potential problems can be analyzed based on behavior logs (such as users repeatedly asking the same question → low confidence in the answer).

[0125] (2) Output signal:

[0126] Generate an optimized instruction set (such as "the knowledge graph needs to add edges"), triggering the dynamic optimization module.

[0127] (3) Incremental updates to the knowledge graph:

[0128] The system automatically adds new nodes or corrects attribute values ​​based on feedback data, and ensures consistency through graph structure verification. It automatically collects frequently asked questions daily and dynamically updates the knowledge base, enabling the system to dynamically optimize the handling process for high-frequency matters through real-time learning.

[0129] In summary, the technical means adopted in this invention include:

[0130] (1) Deep coupling of knowledge graph and RAG: Enhance the semantic understanding of retrieval by using graph entity relationships.

[0131] (2) Reasoning optimization strategy: a hybrid reasoning method based on logical rules and graph neural networks.

[0132] (3) Self-optimization mechanism: Automatically identify knowledge gaps and update the knowledge map and generation model through user feedback.

[0133] Example 13:

[0134] A self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning, comprising the following steps:

[0135] Step 1: Knowledge Graph Construction and Embedding

[0136] (1) Data sources: structured databases, unstructured text, domain knowledge bases.

[0137] (2) Graph Coding: Entity / Relation Embedded Representation (e.g., GraphSAGE).

[0138] Step 2: RAG Q&A Process

[0139] The retrieval phase employs a hybrid retrieval strategy based on question semantics and graph subgraph matching, while the generation phase integrates reasoning paths and optimizes the prompt word engineering.

[0140] Step 3: Self-optimization implementation

[0141] (1) Feedback signal collection: user satisfaction rating and answer confidence analysis.

[0142] (2) Optimize triggering conditions: set confidence thresholds or periodic update strategies.

[0143] (3) Iterative update: Generate models incrementally based on user feedback, high-frequency issues, etc., and dynamically add or delete graph nodes / edges.

Claims

1. A self-optimizing RAG question-answering method based on spatiotemporal knowledge graph reasoning, characterized in that, Includes the following steps: 1) Obtain the data source, process the data source, and construct a domain knowledge graph; 2) Obtain the user's question text, and based on the spatiotemporal knowledge graph, use a graph neural network to perform multi-hop logical reasoning and semantic association analysis on the user's question text to generate a structured reasoning path; 3) Collect explicit and implicit user feedback to generate optimization signals; 4) Automatically update the domain knowledge graph and / or adjust the parameters of the graph neural network based on the optimized signal, including the weight matrix of the aggregation function, node update weights, attention metric weights, and global pooling weights.

2. The RAG question-answering self-optimization method based on spatiotemporal knowledge graph reasoning according to claim 1, characterized in that, The data sources include structured databases, unstructured text, and domain knowledge bases.

3. The RAG question-answering self-optimization method based on spatiotemporal knowledge graph reasoning according to claim 1, characterized in that, In step 1), the steps for constructing the domain knowledge graph include: 1.1) For structured data, entity relationships in relational databases are automatically resolved using predefined mapping rules; 1.2) Perform cross-source alignment on the data and use graph embedding to calculate the semantic similarity between entities from different sources to achieve a unified representation of heterogeneous data; 1.3) Identify contradictory knowledge based on statistical models and perform semantic disambiguation; 1.4) Construct a domain knowledge graph that includes entity-relationship-attribute triples.

4. The RAG question-answering self-optimization method based on spatiotemporal knowledge graph reasoning according to claim 1, characterized in that, The reasoning path is output in the form of a subgraph or a logic chain.

5. The RAG question-answering self-optimization method based on spatiotemporal knowledge graph reasoning according to claim 1, characterized in that, Step 2) involves generating a structured reasoning path, including: 2.1) Define domain logic rules and use an improved A* algorithm to search for a set of candidate reasoning paths that satisfy the constraints in the knowledge graph; 2.2) The GraphSAGE model is used to aggregate neighbor node information, identify implicit relationships, dynamically adjust the weights of candidate inference paths using reinforcement learning, and rank the paths by confidence. 2.3) Perform a weighted vote on the path confidence of the candidate inference path set generated in step 2.1) and the candidate inference path set generated in step 2.2), and generate a structured inference path according to the path confidence after weighted voting.

6. The RAG question-answering self-optimization method based on spatiotemporal knowledge graph reasoning according to claim 1, characterized in that, Before performing multi-hop logical reasoning and semantic association analysis on the user question text, the user question text is converted into a graph by a semantic parser.

7. The RAG question-answering self-optimization method based on spatiotemporal knowledge graph reasoning according to claim 1, characterized in that, The tool for performing multi-hop queries is the graph database Neo4j.

8. The RAG question-answering self-optimization method based on spatiotemporal knowledge graph reasoning according to claim 1, characterized in that, The graph neural network is the GraphSAGE model.

9. The RAG question-answering self-optimization method based on spatiotemporal knowledge graph reasoning according to claim 1, characterized in that, Step 3), which involves collecting explicit and implicit user feedback and generating optimization signals, includes the following steps: 3.1) Collect explicit user feedback, including ratings and error corrections; collect implicit user feedback, including clicks and dwell time. 3.2) Analyze user explicit and implicit feedback using natural language processing models, extract error types and / or analyze potential problems based on behavior logs, thereby generating an optimization instruction set; the optimization instruction set includes optimization instructions related to user style preferences, format preferences, information depth, interaction patterns, expression tendencies, content optimization tendencies, and interactive experience.

10. The RAG question-answering self-optimization method based on spatiotemporal knowledge graph reasoning according to claim 1, characterized in that, The concept of automatically updating the domain knowledge graph based on optimized signals refers to automatically adding new nodes or correcting attribute values ​​based on feedback data.

Images