A method and system for mining key knowledge paths by fusing a knowledge graph and a large language model

Abstract

This invention discloses a method and system for key knowledge path mining that integrates knowledge graphs and large language models, belonging to the field of artificial intelligence application technology. The method includes: extracting knowledge entities from raw data using a large language model, guided by domain knowledge; matching and aligning the extracted entities with standardized, authoritative entities in the domain knowledge graph using a multi-stage approach; using the aligned question entity set as the source node set and the answer entity set as the target node set in the graph network of the domain knowledge graph, traversing all possible source and target node pairs using an algorithm to obtain the shortest candidate knowledge path between them; and filtering and ranking the obtained candidate paths based on the large language model and centrality indicators to mine key knowledge paths. This invention has advantages such as high accuracy, high reliability, high efficiency, strong interpretability, and strong traceability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence application technology, and more specifically, to a method and system for mining key knowledge paths that integrates knowledge graphs and large language models. Background Technology

[0002] In professional fields such as chemical engineering, knowledge graphs, as carriers of knowledge, contain complex semantic relationships between entities. Key Knowledge Paths (KKPs) are high-value semantic chains connecting knowledge entities and carrying causal relationships, which are of great significance for scientific research, intelligent question answering, and technological innovation. Existing key knowledge path mining technologies mainly face two major challenges: Traditional knowledge graph-based mining methods, such as graph traversal or random walk algorithms, while finding paths, are prone to the "path explosion" problem, where the number of candidate paths grows exponentially, causing valuable key paths to be submerged by a massive number of irrelevant paths (knowledge dilution). At the same time, these methods often rely on the graph's topological structure, lacking sufficient understanding of deep semantic information, and their mining efficiency and accuracy need improvement. Furthermore, some methods rely on predefined rule templates from experts, resulting in poor generalization ability and portability. Mining methods based on Large Language Models (LLMs): LLMs possess powerful contextual understanding and reasoning capabilities, and can directly attempt to generate knowledge paths from text. However, this method has significant drawbacks: First, the paths generated by LLM are prone to "illusion," meaning the content does not conform to facts or domain knowledge, lacking credibility; second, its reasoning process is opaque, resulting in poor interpretability and traceability of the mined paths; finally, LLM relies on general knowledge, and its accuracy and depth are insufficient when dealing with vertical domain tasks requiring deep expertise. Therefore, how to combine the accuracy of domain knowledge with the semantic understanding capabilities of large language models to achieve accurate, efficient, and interpretable mining of key knowledge paths is a pressing technical challenge that needs to be addressed in the current technological field. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a key knowledge path mining method and system that integrates knowledge graphs and large language models. It at least solves the problems of path explosion, knowledge dilution, low mining efficiency, unreliable results, lack of interpretability, and insufficient accuracy in vertical domain applications in existing key knowledge path mining technologies. It has the advantages of high accuracy, good reliability, high efficiency, strong interpretability and traceability.

[0004] The objective of this invention is achieved through the following solution: A key knowledge path mining method integrating knowledge graphs and large language models includes the following steps: Using a large language model, knowledge entities are extracted from the raw data to be processed based on domain knowledge guidance; A multi-stage approach is used to match and align the extracted entities with standardized and authoritative entities in the domain knowledge graph; In the graph network of the domain knowledge graph, the aligned question entity set is used as the source node set and the answer entity set is used as the target node set. For all possible source and target node pairs, an algorithm is used to traverse them and obtain the shortest candidate knowledge path between them. The candidate paths are filtered and ranked based on a large language model and centrality metrics to uncover key knowledge paths.

[0005] Furthermore, the step of using a large language model to extract knowledge entities from the raw data to be processed based on domain knowledge guidance specifically includes the following sub-steps: providing prompts to the large language model containing a preset entity type system in the domain knowledge graph, and guiding the large language model to identify and classify entities according to predetermined specifications through task-oriented instructions and examples.

[0006] Furthermore, the raw data to be processed includes question-and-answer text from the chemical industry.

[0007] Furthermore, the process of matching and aligning the extracted entities with standardized, authoritative entities in the domain knowledge graph using a multi-stage approach specifically includes the following sub-steps: Phase 1, Precise Matching: Directly compare whether the extracted entity and the knowledge graph entity string are completely identical; if the match is successful, the standardized mapping from the extracted entity to the standard entity of the knowledge graph is achieved; if the match is unsuccessful, proceed to Phase 2. Phase Two, Similarity Matching: If no exact match is found, the text embedding model is used to calculate the vector similarity between the extracted entity and all entities in the knowledge graph. When the highest similarity exceeds a preset threshold, the most similar entity is selected as the matching result. If the match is successful, the standardized mapping from the extracted entity to the standard entity in the knowledge graph is implemented. If the match is unsuccessful, proceed to Phase Three. Phase 3, context selection based on the large language model: If neither Phase 1 nor Phase 2 is successful, the list of candidate entities with the top k similarity rankings and the original data context are input into the large language model. The semantic understanding capability of the large language model is used to select the most relevant entity in the corresponding context as the matching result. Phase 4, No Matching Entity: If Phases 1 through 3 fail, the entity is determined to have no matching entity in the knowledge graph.

[0008] Furthermore, the step of using an algorithm to traverse and obtain the shortest candidate knowledge paths between them specifically includes the following sub-steps: using a shortest path search algorithm to traverse and obtain all the shortest candidate knowledge paths between them.

[0009] Furthermore, the process of filtering and ranking the acquired candidate paths based on a large language model and centrality metrics to uncover key knowledge paths includes the following sub-steps: Step 1, semantic filtering based on key entities: First, use a large language model to analyze the question-answer pair context and generate the core domain concept entities most relevant to the context, i.e., seed entities; then, filter out the candidate paths that contain at least one seed entity. Step 2, Structural ranking based on degree centrality: For the filtered paths, calculate the average degree centrality index of the nodes on the path in the entire knowledge graph. Nodes with higher degree centrality are core or pivotal concepts in the domain. Finally, sort the paths in descending order according to their average degree centrality scores. The paths with the highest rankings are the key knowledge paths that have been finally mined.

[0010] Furthermore, the text embedding model includes MatSciBERT.

[0011] Furthermore, the preset threshold includes 0.85.

[0012] Furthermore, the shortest path search algorithm includes Dijkstra's algorithm or BFS algorithm.

[0013] A key knowledge path mining system integrating knowledge graphs and large language models includes a processor and a memory, wherein the memory stores a computer program that executes the method described in any of the preceding methods when the computer program is loaded by the processor.

[0014] The beneficial effects of this invention include: (1) This invention significantly improves accuracy and reliability. Specifically, by integrating domain knowledge graphs as the "gold standard of facts," it effectively constrains LLM, avoids content illusion, and ensures the correctness of the mined knowledge paths in terms of logic, facts, and domain knowledge.

[0015] (2) This invention significantly improves mining efficiency and quality. Specifically, by combining the powerful semantic understanding capabilities of LLM with the efficient structured retrieval of knowledge graphs, it synergistically overcomes the problems of "path explosion" and "knowledge dilution," enabling it to accurately and efficiently mine high-value, interpretable key knowledge paths from massive amounts of data.

[0016] (3) This invention enhances the interpretability and traceability of the paths. Specifically, each path mined is composed of real entities and relationships in the knowledge graph, with clear logic and traceable origin, providing a reliable reasoning basis for downstream applications (such as intelligent question answering and scientific research assistance).

[0017] (4) This invention enhances adaptability in vertical fields. Specifically, the method framework of this invention has good versatility. By replacing the knowledge graphs of different fields, it can be easily applied to other professional fields to solve the knowledge discovery problem in specific fields. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the multi-stage efficient entity alignment steps in an embodiment of the present invention; Figure 2 This is a diagram illustrating the shortest path style for connecting the question and answer entity sets in an embodiment of the present invention. Figure 3 This is a flowchart illustrating the overall process steps of an embodiment of the present invention. Detailed Implementation

[0020] All features disclosed in all embodiments of this specification, or steps in all methods or processes implied in the disclosure, may be combined and / or extended or replaced in any way, except for mutually exclusive features and / or steps.

[0021] The specific implementation process of this invention is as follows: As a first aspect of the present invention, in a preferred embodiment, such as Figures 1-3 As shown, this embodiment of the invention provides a method for mining key knowledge paths that integrates knowledge graphs and large language models, including the following steps: Step S1: Precise entity extraction guided by domain knowledge. Specifically, a Large Language Model (LLM) is used to extract knowledge entities from the raw data to be processed (e.g., question-and-answer text in the chemical industry). The prompts provided to the LLM contain a pre-defined entity type system from the domain knowledge graph (e.g., materials, properties, applications, synthesis methods, characterization methods, etc.). Through task-oriented instructions and a few examples, the LLM is guided to accurately identify and classify entities according to established specifications, ensuring the accuracy and standardization of the extraction results.

[0022] Step S2: Multi-stage Efficient Entity Alignment. Specifically, the entities extracted in Step S1 are matched and aligned with standardized, authoritative entities in the domain knowledge graph. This step employs a four-stage matching strategy, including an exact matching stage, a similarity matching stage, an LLM-based context selection stage, and a no-match matching stage. Specifically: 1) In the exact matching stage, the extracted entity is directly compared to the entity strings in the knowledge graph for complete consistency. 2) In the similarity matching stage, if no exact match is found, a text embedding model (such as MatSciBERT) is used to calculate the vector similarity between the extracted entity and all entities in the knowledge graph. When the highest similarity exceeds a preset threshold (e.g., 0.85), the most similar entity is selected as the matching result. 3) In the LLM-based context selection stage, if the first two stages fail to match, the list of the top k similarity candidate entities and the original data context (e.g., questions and answers) are input into the LLM. Leveraging the semantic understanding capabilities of the LLM, the most relevant entity in the specific context is selected as the matching result. 4) No Matching Stage: If all the above stages fail, the entity is determined to have no corresponding match in the knowledge graph. Through the above four-stage strategy, a standardized mapping from extracted entities to standard entities in the knowledge graph can be achieved, such as... Figure 1 As shown.

[0023] Step S3: Shortest path search based on graph network. Specifically, in the graph network of the domain knowledge graph, the question entity set aligned in Step S2 is used as the source node set, and the answer entity set is used as the target node set. For all possible source and target node pairs, a shortest path search algorithm (such as Dijkstra's algorithm or BFS algorithm) is used to traverse and obtain all shortest candidate knowledge paths between them.

[0024] Step S4: Key Knowledge Path Identification and Ranking Based on LLM and Centrality Indicators. Specifically, the large number of candidate paths obtained in Step S3 are subjected to quality filtering and ranking to extract high-value key knowledge paths. This step integrates information from both semantic and structural dimensions. 1) Semantic Filtering Based on Key Entities: First, LLM is used to analyze the question-answer pair context to generate the core domain concept entities (i.e., seed entities) most relevant to that context. Then, paths containing at least one seed entity are selected from the candidate paths to ensure that the path content is highly relevant to the question context. 2) Structural Ranking Based on Degree Centrality: For the filtered paths, the average degree centrality index of all nodes on the path in the entire knowledge graph is calculated. Nodes with higher degree centrality are usually core or pivotal concepts in the domain. Finally, the paths are sorted in descending order based on their average degree centrality score, and the top-ranked paths are the final key knowledge paths extracted.

[0025] As a second aspect of the present invention, a corresponding system for implementing the above-described method is also provided. The system includes a processor and a memory, wherein a computer program is stored in the memory. When the computer program is loaded by the processor, it executes the method as described above. For example, it includes an entity extraction module, an entity alignment module, a path search module, and a path recognition module, etc. The specific implementation process of each module corresponds to the corresponding steps in the above-described method, and will not be elaborated further here.

[0026] In other embodiments, a question-and-answer pair from the chemical industry is provided as an example to illustrate the execution flow of the method described in the above embodiments of the present invention in more detail. Let the original data be: Question: "What is a heterogeneous liquid system? What are the causes of its formation and the purpose of its separation?" Answer: "A heterogeneous liquid-phase system, also known as a liquid-phase suspension system, is a system in which an insoluble substance is suspended in a liquid. The formation of a heterogeneous liquid-phase system is due to the concentration of the solution and the precipitation of crystals, the occurrence of chemical reactions in the liquid phase, or the presence of impurities in the liquid phase. The purpose is to obtain a pure product through separation." Using the method of the above embodiments of the present invention, the following steps are performed on the above-mentioned raw data: Step 1, Entity Extraction: Input the above question-and-answer text into an LLM that has preset prompts for entity types (materials, properties, etc.) in the chemical industry. The LLM will extract entities such as "liquid-phase heterogeneous system", "separation", and "solution concentration" and label their types.

[0027] Step 2, Entity Alignment: The entities extracted in Step 1 are "separated." First, exact matching is performed to find identical entities in the knowledge graph. For "liquid-phase heterogeneous systems," exact matching may not be possible. In this case, similarity matching is initiated, calculating the vector similarity between the entity and entities such as "heterogeneous composite materials" in the knowledge graph. If the similarity is higher than 0.85, alignment is performed. For "insoluble substances," if the similarity is low, LLM selection is initiated. Candidate entities (such as hβ molecular sieves) and context are input into LLM for evaluation. LLM selects the best match based on the context.

[0028] Step 3, Path Search: Use entities aligned with the question (e.g., "membrane separation") as source nodes and entities aligned with the answer (e.g., "crystallization process," "concentration") as target nodes. Perform a shortest path search within the knowledge graph network. For example, a candidate path might be obtained: Membrane separation - [synthesis] → Carbon dioxide - [property is] → Crystal - [determine] → Crystallization process. Figure 2 As shown.

[0029] Step 4, Path Identification and Ranking, specifically includes screening and ranking processes. During screening, LLM analysis is first used to analyze question-answer pairs, generating key seed entities such as "heterogeneous," "separation," and "crystallization process." Then, paths containing these seed entities are retained from the candidate paths in Step 3. During ranking, the degree centrality of all nodes (membrane separation, carbon dioxide, crystal, crystallization process) in the path "membrane separation → ... → crystallization process" is calculated, and their average value is taken. This score is calculated for all screened paths, and they are ranked from highest to lowest score. The top 10 paths are output as the final result. For example, "membrane separation – [synthesis] → carbon dioxide – [attribute is] → crystal – [determine] → crystallization process" is a high-value path.

[0030] As can be seen, through steps 1 to 4 above, the embodiment of the present invention successfully extracted a set of high-quality key knowledge paths from a chemical engineering question and answer pair, which can explain "why separation is needed" and "the relationship between separation and system formation".

[0031] It should be noted that, within the scope of protection defined in the claims of this invention, the following embodiments can be combined and / or extended or replaced in any logical manner from the above specific embodiments, such as the disclosed technical principles, disclosed technical features or implicitly disclosed technical features.

[0032] Example 1 A key knowledge path mining method integrating knowledge graphs and large language models includes the following steps: Using a large language model, knowledge entities are extracted from the raw data to be processed based on domain knowledge guidance; A multi-stage approach is used to match and align the extracted entities with standardized and authoritative entities in the domain knowledge graph; In the graph network of the domain knowledge graph, the aligned question entity set is used as the source node set and the answer entity set is used as the target node set. For all possible source and target node pairs, an algorithm is used to traverse them and obtain the shortest candidate knowledge path between them. The candidate paths are filtered and ranked based on a large language model and centrality metrics to uncover key knowledge paths.

[0033] Example 2 Based on Example 1, the step of using a large language model to extract knowledge entities from the raw data to be processed based on domain knowledge guidance specifically includes the following sub-steps: providing prompts to the large language model containing a preset entity type system in the domain knowledge graph, and guiding the large language model to identify and classify entities according to predetermined specifications through task-oriented instructions and examples.

[0034] Example 3 Based on Example 2, the raw data to be processed includes question-and-answer text from the chemical industry.

[0035] Example 4 Based on Example 1, the multi-stage approach to matching and aligning the extracted entities with standardized, authoritative entities in the domain knowledge graph specifically includes the following sub-steps: Phase 1, Precise Matching: Directly compare whether the extracted entity and the knowledge graph entity string are completely identical; if the match is successful, the standardized mapping from the extracted entity to the standard entity of the knowledge graph is achieved; if the match is unsuccessful, proceed to Phase 2. Phase Two, Similarity Matching: If no exact match is found, the text embedding model is used to calculate the vector similarity between the extracted entity and all entities in the knowledge graph. When the highest similarity exceeds a preset threshold, the most similar entity is selected as the matching result. If the match is successful, the standardized mapping from the extracted entity to the standard entity in the knowledge graph is implemented. If the match is unsuccessful, proceed to Phase Three. Phase 3, context selection based on the large language model: If neither Phase 1 nor Phase 2 is successful, the list of candidate entities with the top k similarity rankings and the original data context are input into the large language model. The semantic understanding capability of the large language model is used to select the most relevant entity in the corresponding context as the matching result. Phase 4, No Matching Entity: If Phases 1 through 3 fail, the entity is determined to have no matching entity in the knowledge graph.

[0036] Example 5 Based on Example 1, the step of using an algorithm to traverse and obtain the shortest candidate knowledge paths between them specifically includes the following sub-steps: using a shortest path search algorithm to traverse and obtain all the shortest candidate knowledge paths between them.

[0037] Example 6 Based on Example 1, the process of quality filtering and ranking of the acquired candidate paths based on a large language model and centrality indicators to mine key knowledge paths specifically includes the following sub-steps: Step 1, semantic filtering based on key entities: First, use a large language model to analyze the question-answer pair context and generate the core domain concept entities most relevant to the context, i.e., seed entities; then, filter out the candidate paths that contain at least one seed entity. Step 2, Structural ranking based on degree centrality: For the filtered paths, calculate the average degree centrality index of the nodes on the path in the entire knowledge graph. Nodes with higher degree centrality are core or pivotal concepts in the domain. Finally, sort the paths in descending order according to their average degree centrality scores. The paths with the highest rankings are the key knowledge paths that have been finally mined.

[0038] Example 7 Based on Example 4, the text embedding model includes MatSciBERT.

[0039] Example 8 Based on Example 4, the preset threshold includes 0.85.

[0040] Example 9 Based on Example 5, the shortest path search algorithm includes Dijkstra's algorithm or BFS algorithm.

[0041] Example 10 A key knowledge path mining system integrating knowledge graphs and large language models includes a processor and a memory, wherein the memory stores a computer program that executes the method described in any of the preceding methods when the computer program is loaded by the processor.

[0042] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0043] According to one aspect of the present invention, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0044] In another aspect, embodiments of the present invention also provide a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods described in the above embodiments.

Claims

1. A key knowledge path mining method integrating knowledge graphs and large language models, characterized in that, Includes the following steps: Using a large language model, knowledge entities are extracted from the raw data to be processed based on domain knowledge guidance; A multi-stage approach is used to match and align the extracted entities with standardized and authoritative entities in the domain knowledge graph; In the graph network of the domain knowledge graph, the aligned question entity set is used as the source node set and the answer entity set is used as the target node set. For all possible source and target node pairs, an algorithm is used to traverse them and obtain the shortest candidate knowledge path between them. The candidate paths are filtered and ranked based on a large language model and centrality metrics to uncover key knowledge paths.

2. The key knowledge path mining method integrating knowledge graphs and large language models according to claim 1, characterized in that, The method of extracting knowledge entities from the raw data to be processed using a large language model based on domain knowledge guidance includes the following sub-steps: providing prompts to the large language model containing a pre-defined entity type system in the domain knowledge graph, and guiding the large language model to identify and classify entities according to predetermined specifications through task-oriented instructions and examples.

3. The key knowledge path mining method integrating knowledge graphs and large language models according to claim 2, characterized in that, The raw data to be processed includes question-and-answer text pairs from the chemical industry.

4. The key knowledge path mining method integrating knowledge graphs and large language models according to claim 1, characterized in that, The process employs a multi-stage approach to match and align the extracted entities with standardized, authoritative entities in the domain knowledge graph, specifically including the following sub-steps: Phase 1, Precise Matching: Directly compare whether the extracted entity and the knowledge graph entity string are completely identical; If the match is successful, the extracted entity is mapped to the standard entity of the knowledge graph; if the match is unsuccessful, proceed to stage two. Phase 2, Similarity Matching: If no exact match is found, the text embedding model is used to calculate the vector similarity between the extracted entity and all entities in the knowledge graph. When the highest similarity exceeds a preset threshold, the most similar entity is selected as the matching result. If the match is successful, the extracted entity is mapped to the standard entity of the knowledge graph. If the match is successful, the extracted entity is mapped to the standard entity of the knowledge graph; if the match is unsuccessful, proceed to stage three. Phase 3, context selection based on the large language model: If neither Phase 1 nor Phase 2 is successful, the list of candidate entities with the top k similarity rankings and the original data context are input into the large language model. The semantic understanding capability of the large language model is used to select the most relevant entity in the corresponding context as the matching result. Phase 4, No Matching Entity: If Phases 1 through 3 fail, the entity is determined to have no matching entity in the knowledge graph.

5. The key knowledge path mining method integrating knowledge graphs and large language models according to claim 1, characterized in that, The step of using an algorithm to traverse and obtain the shortest candidate knowledge paths between them specifically includes the following sub-steps: using a shortest path search algorithm to traverse and obtain all the shortest candidate knowledge paths between them.

6. The key knowledge path mining method integrating knowledge graphs and large language models according to claim 1, characterized in that, The process of obtaining candidate paths, based on a large language model and centrality metrics, involves quality filtering and ranking to uncover key knowledge paths. This includes the following sub-steps: Step 1, semantic filtering based on key entities: First, use a large language model to analyze the question-answer pair context and generate the core domain concept entities most relevant to the context, i.e., seed entities; then, filter out the candidate paths that contain at least one seed entity. Step 2, Structural ranking based on degree centrality: For the filtered paths, calculate the average degree centrality index of the nodes on the path in the entire knowledge graph. Nodes with higher degree centrality are core or pivotal concepts in the domain. Finally, sort the paths in descending order according to their average degree centrality scores. The paths with the highest rankings are the key knowledge paths that have been finally mined.

7. The key knowledge path mining method integrating knowledge graphs and large language models according to claim 4, characterized in that, The text embedding model includes MatSciBERT.

8. The key knowledge path mining method integrating knowledge graphs and large language models according to claim 4, characterized in that, The preset threshold includes 0.

85.

9. The key knowledge path mining method integrating knowledge graphs and large language models according to claim 5, characterized in that, The shortest path search algorithm includes Dijkstra's algorithm or BFS algorithm.

10. A key knowledge path mining system integrating knowledge graphs and large language models, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when loaded by the processor, executes the method as described in any one of claims 1 to 9.

Images