Knowledge graph fusion method for diabetes health management

Abstract

The invention discloses a knowledge graph fusion method for diabetes health management. The knowledge graph fusion method comprises the following steps: step 1, acquiring two source knowledge graphs and storing the two source knowledge graphs in a Neo4J database; step 2, determining a target Scheme of the diabetes knowledge graph, and performing data preprocessing according to the Scheme; 3, entity alignment operation is achieved through a similarity algorithm and a large language model, the corresponding relation between entities is established, adjustment is conducted continuously according to output of the model, and the most matched corresponding entity is found; step 4, performing entity insertion based on an entity alignment result and a Scheme specification, and adding entities which exist in one knowledge graph but are missing in the other knowledge graph to corresponding positions; and step 5, adjusting the inserted entities and relationships according to Schema, eliminating redundancy and conflicts, and completing the fusion of the knowledge graph. The method provided by the invention solves the problems of long period and high cost of a general fusion process, is not only suitable for the field of diabetes, but also can be applied to mapping knowledge domain fusion in other fields.

Description

Technical Field

[0001] The present invention relates to the technical field of diabetes knowledge graphs and artificial intelligence models, and in particular to a knowledge graph fusion method for diabetes health management. Background Art

[0002] Knowledge graph is a structured knowledge base that contains rich entity, attribute and relationship information. Its application can solve multiple problems such as knowledge management, information retrieval, data integration and intelligent decision-making. By combining knowledge graph with health management system, the system can obtain rich domain knowledge from the graph. Knowledge graph (KG) is crucial for intelligent applications such as question answering and recommendation system. Although existing KGs are available in diabetes research, they usually cannot fully meet the requirements of specific tasks. On the one hand, building a new knowledge graph from scratch can ensure the comprehensiveness and accuracy of information, but such a process is often long and costly. On the other hand, most of the current knowledge graph fusion focuses on how to improve the entity alignment method in the fusion process. There are very few detailed descriptions of the entire process of knowledge fusion, and there is no complete and effective method. Therefore, integrating existing knowledge graphs and customizing and optimizing them according to the needs of specific tasks is an important research content of existing methods and technologies. Summary of the invention

[0003] The purpose of the present invention is to provide a knowledge graph fusion method for diabetes health management, aiming to achieve the fusion of diabetes knowledge graphs and provide a universal method and process for the construction of knowledge graphs in other fields.

[0004] In order to achieve the above object, the technical solution of the present invention is to provide a knowledge graph fusion method for diabetes health management, which is characterized by comprising the following steps:

[0005] S1. Obtain two source knowledge graphs and store them in the Neo4J database;

[0006] S2. Determine the target Scheme of the diabetes knowledge graph and perform data preprocessing according to the Scheme;

[0007] S3. Implement entity alignment operations through similarity algorithms and large language models, establish corresponding relationships between entities, and continuously adjust based on the output of the model to find the most matching corresponding entity;

[0008] S4. Perform entity insertion based on the entity alignment results and Scheme specifications, and add entities that exist in one knowledge graph but are missing in the other to the corresponding positions;

[0009] S5. Adjust the inserted entities and relationships according to the Schema to eliminate redundancy and conflict and complete the fusion of the knowledge graph.

[0010] Preferably, step S1 specifically includes the following steps: searching for a knowledge graph related to a field close to the target schema as a reference, selecting a knowledge graph that is highly relevant to the target schema, large in scale, and authoritative in the medical field, and storing the acquired knowledge graph triples in a Neo4J graph database using Cypher statements.

[0011] Preferably, the step S2 specifically includes the following steps:

[0012] S21. Refer to the Stanford seven-step university construction method, and determine the target scheme in combination with the task requirements of diabetes health management and the standardized domains in the Diachat dialogue dataset.

[0013] S22. Use the large language model Prompt and the data visualization tool Neo4j Browser to query and analyze the two knowledge graphs, find problems, and modify the problems based on the Schema specification and using Cypher statements.

[0014] Preferably, step S3 specifically includes the following steps: using the Levenshtein edit distance algorithm as the initial similarity calculation method, treating each entity name as a string, and determining their similarity by calculating the edit distance between different entity names. If the edit distance is less than a set threshold, they are regarded as similar entities; then further using the large language model, using the few-shot prompt generation method to take the entity pair to be evaluated as the input of the model, and if the model finally outputs a similarity score greater than 0.8, it is identified as a high-similarity entity pair.

[0015] Preferably, step S4 specifically includes the following steps: according to the result of entity alignment, use Cypher query statements to find the corresponding node in a knowledge graph. If it already exists, there is no need to insert it again; if it does not exist, perform the entity insertion operation; after determining the insertion position, it is necessary to select an appropriate relationship type to represent the association between entities. According to the structure and semantics of the knowledge graph, during the insertion process, select to create a new relationship type through the Cypher query statement Merge, connect the entity in the knowledge graph with the node label of the insertion position, and form a hierarchical relationship.

[0016] Preferably, step S4 specifically includes the following steps: using the visualization tool Neo4j Browser in combination with Cypher's pattern matching and aggregation functions to retrieve the relationships and attributes in the knowledge graph, and then manually modifying its entity relationships based on the Schema specification; processing the nodes generated after the modification, mainly including deleting duplicate nodes and creating relationships; adding sources and SubClassOf and other relationships to the new nodes in the graph, and iterating in this way until all entity relationships and attributes in the graph are processed.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] The knowledge graph fusion method for diabetes health management proposed in the present invention screens the source knowledge graph based on the determined target Schema of the diabetes knowledge graph, and preprocesses the data according to the Schema specification. Finally, with the support of the large language model, entity alignment, entity insertion, relationship adjustment and other technologies are used to propose a rapid knowledge fusion framework for multi-source knowledge graphs, and finally the fusion of knowledge graphs is completed. This framework is not only applicable to the field of diabetes, but can also be applied to knowledge graphs in other fields, and can provide data support for subsequent health diagnosis and personalized recommendation work. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The overall fusion framework flow chart of the present invention is

[0020] Figure 2 The target Scheme framework diagram of the present invention DETAILED DESCRIPTION

[0021] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. It should be understood that the specific embodiments described herein are part of the embodiments of the present invention, rather than all of the embodiments, and are only used to explain the embodiments of the present invention, and are not used to limit the embodiments of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0022] A knowledge graph fusion method for diabetes health management includes the following steps:

[0023] S1. Obtain two source knowledge graphs and store them in the Neo4J database, including:

[0024] Find knowledge graphs related to fields close to the target schema as references, and select knowledge graphs that are highly relevant to the target schema, large in scale, and authoritative in the medical field. Store the acquired knowledge graph triples in the Neo4J graph database using Cypher statements.

[0025] S2. Determine the target Scheme of the diabetes knowledge graph and perform data preprocessing according to the Scheme, including:

[0026] S21. Refer to the Stanford seven-step university construction method, and determine the target scheme in combination with the task requirements of the diabetes health management dialogue system and the standardized domains in the Diachat dialogue dataset.

[0027] S22. Use the large language model Prompt and the data visualization tool Neo4j Browser to query and analyze the two knowledge graphs, find problems, and modify the problems based on the Schema specification and using Cypher statements.

[0028] S3. Implement entity alignment operations through similarity algorithms and large language models, establish corresponding relationships between entities, and continuously adjust based on the output of the model to find the most matching corresponding entity, including:

[0029] The Levenshtein edit distance algorithm is used as the initial similarity calculation method. Each entity name is regarded as a string. The edit distance between different entity names is calculated to determine their similarity. If the edit distance is less than the set threshold, they are regarded as similar entities. Then, the large language model is further used to generate the entity pairs to be evaluated as the input of the model using the few-shot prompt generation method. If the model finally outputs a similarity score greater than 0.8, it is considered a high-similarity entity pair.

[0030] S4. Perform entity insertion based on the entity alignment results and Scheme specifications, and add entities that exist in one knowledge graph but are missing in the other to the corresponding positions, including:

[0031] According to the result of entity alignment, use Cypher query statements to find the corresponding node in a knowledge graph. If it already exists, there is no need to insert it again; if it does not exist, perform entity insertion operation; after determining the insertion position, you need to select the appropriate relationship type to represent the association between entities. According to the structure and semantics of the knowledge graph, choose to create a new relationship type through the Cypher query statement Merge during the insertion process, connect the entity in the knowledge graph with the node label of the insertion position, and form a hierarchical relationship.

[0032] S5. Adjust the inserted entities and relationships according to the Schema, eliminate redundancy and conflict, and complete the fusion of the knowledge graph, including:

[0033] The relationships and attributes in the knowledge graph are retrieved through the visualization tool Neo4j Browser combined with Cypher's pattern matching and aggregation functions, and then the entity relationships are manually modified based on the Schema specification; the nodes generated after the modification are processed, mainly including deleting duplicate nodes and creating relationships; adding sources and SubClassOf and other relationships to the new nodes in the graph, and iterating in this way until all entity relationships and attributes in the graph are processed.

[0034] Embodiment 1

[0035] Combined with Figure 1-2 The embodiment of the present invention provides a knowledge graph fusion method for a diabetes health management dialogue system, the method comprising the following steps:

[0036] S1. Obtain two source knowledge graphs and store them in the Neo4J database;

[0037] S2. Refer to the Stanford seven-step university construction method, and combine the task requirements of the diabetes health management dialogue system and the standardized domains in the Diachat dialogue dataset to determine the target Scheme, such as Figure 2 ;

[0038] S3. Use the large language model Prompt and data visualization tool Neo4j Browser to query and analyze the two knowledge graphs, find problems, and modify them based on the Schema specification and using Cypher statements;

[0039] S4. Implement entity alignment operations through similarity algorithms and large language models, establish corresponding relationships between entities, and continuously adjust based on the output of the model to find the most matching corresponding entity;

[0040] S5. Perform entity insertion based on the entity alignment results and Scheme specifications, and add entities that exist in one knowledge graph but are missing in the other to the corresponding positions;

[0041] S6. Adjust the inserted entities and relationships according to the Schema to eliminate redundancy and conflict and complete the fusion of the knowledge graph.

[0042] Specifically, in Figure 2In this paper, the target Schema construction method is as follows: ① Determine the professional field and scope of the knowledge graph. Based on the task requirements of the diabetes health management dialogue system that this study focuses on and the standardized fields in the Diachat dialogue dataset; ② Consider reusing existing resources. Based on the ontology research in the diabetes field, this study considers reusing the definition of slots in the Diachat dialogue dataset. ③ List the important terms in the ontology field, which are the basis for building the Schema. Extract relevant conceptual terms by observing and analyzing the Diachat dialogue data; ④ Define classes and class levels. There are three methods for class hierarchy construction: top-down, bottom-up, and hybrid. This example adopts the top-down method. Refer to the hierarchies in the two knowledge graphs of Diakg and MKN and combine them with the domain slots of the Diachat dialogue dataset to obtain the final hierarchical classification; ⑤ Define class attributes. The attributes between classes include object attributes and data attributes. Object attributes refer to the semantic relationship between semantic types; ⑥ Create an instance of the ontology; ⑦ Consistency check and visualization. After the ontology construction is completed, the reasoning engine that comes with Protégé5.5.0 is used to check the consistency of the ontology, including the consistency of class hierarchy and semantic relations. In order to more intuitively visualize the internal connections between ontologies, Protégé's own visualization tool (Onto Graf) can be used to visualize the ontology.

[0043] In the entire fusion process, the various steps are not carried out linearly, but there is a relationship of mutual influence and feedback, and multiple iterations are required to optimize the results. Specifically, when performing entity alignment, if it is found that some entities cannot be accurately matched, it is necessary to return to the data preprocessing stage for further cleaning and standardization. Similarly, during the process of entity insertion and relationship adjustment, if some conflicts or inconsistent relationships are found, it is necessary to return to the entity alignment or data preprocessing stage for adjustment. In the above iterative process, Neo4j's interactive graphical interface Neo4j Browser is used to operate the data in the graph database by writing and running Cypher query statements, making adjustments and optimizations during the iterative process, and viewing and analyzing the results in real time.

[0044] Embodiment 2

[0045] The method provided by the present invention realizes how to quickly fuse multi-source knowledge graphs. This method is not limited to the field of diabetes, and also has broad application prospects for the fusion of knowledge graphs in other fields.

[0046] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. For those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and design ideas of the present invention. The scope of the present invention is defined by the attached claims and their equivalents.

Claims

1. A knowledge graph fusion method for diabetes health management, characterized in that: The steps include: S1. Obtain two source knowledge graphs and store them in the Neo4J database; S2. Determine the target Scheme of the diabetes knowledge graph and perform data preprocessing according to the Scheme; S3. Implement entity alignment operations through similarity algorithms and large language models, establish corresponding relationships between entities, and continuously adjust based on the output of the model to find the most matching corresponding entity; S4. Perform entity insertion based on the entity alignment results and Scheme specifications, and add entities that exist in one knowledge graph but are missing in the other to the corresponding positions; S5. Adjust the inserted entities and relationships according to the Schema to eliminate redundancy and conflict and complete the fusion of the knowledge graph.

2. According to the knowledge graph fusion method for diabetes health management according to claim 1, the step S1 comprises: Find knowledge graphs related to fields close to the target schema as references, and select knowledge graphs that are highly relevant to the target schema, large in scale, and authoritative in the medical field. Store the acquired knowledge graph triples in the Neo4J graph database using Cypher statements.

3. According to the knowledge graph fusion method for diabetes health management according to claim 1, the step S2 comprises: S21. Refer to the Stanford seven-step university construction method, and determine the target scheme in combination with the task requirements of diabetes health management and the standardized domains in the Diachat dialogue dataset. S22. Use the large language model Prompt and the data visualization tool Neo4j Browser to query and analyze the two knowledge graphs, find problems, and modify the problems based on the Schema specification and using Cypher statements.

4. According to the knowledge graph fusion method for diabetes health management according to claim 1, step S3 comprises: The Levenshtein edit distance algorithm is used as the initial similarity calculation method. Each entity name is regarded as a string. The edit distance between different entity names is calculated to determine their similarity. If the edit distance is less than the set threshold, they are regarded as similar entities. Then, the large language model is further used to use the few-shot prompt generation method to take the entity pair to be evaluated as the input of the model. If the model finally outputs a similarity score greater than 0.8, it is considered a high-similarity entity pair.

5. According to the knowledge graph fusion method for diabetes health management according to claim 1, the step S4 comprises: According to the result of entity alignment, use Cypher query statements to find the corresponding node in a knowledge graph. If it already exists, there is no need to insert it again; if it does not exist, perform entity insertion; after determining the insertion position, you need to select the appropriate relationship type to represent the association between entities. According to the structure and semantics of the knowledge graph, during the insertion process, a new relationship type is created through the Cypher query statement Merge, which connects the entities in the knowledge graph with the node labels at the insertion position to form a hierarchical relationship.

6. According to the knowledge graph fusion method for diabetes health management according to claim 1, the step S5 comprises: The relationships and attributes in the knowledge graph are retrieved through the visualization tool Neo4j Browser combined with Cypher's pattern matching and aggregation functions, and then the entity relationships are manually modified based on the Schema specification; the nodes generated after the modification are processed, mainly including deleting duplicate nodes and creating relationships; adding sources and SubClassOf and other relationships to the new nodes in the graph, and iterating in this way until all entity relationships and attributes in the graph are processed.

Images