Knowledge graph optimization method and device, and electronic device

Abstract

This application provides a knowledge graph optimization method, apparatus, and electronic device. In this application, by analyzing the entity embedding vectors corresponding to the knowledge graph, outliers are filtered out from the knowledge graph at the semantic level; furthermore, by analyzing the structural features of each entity in the knowledge graph, meaningless entities are filtered out from the structural feature level; a large language model is used to perform semantic analysis on each entity in the knowledge graph to further filter out entities irrelevant to the application domain from the semantic level; and then, by pruning and structural optimization of the knowledge graph, redundant entities are filtered out, and similar connections and different entities with graph structure breaks are optimized. Through multiple noise entity filtering and graph structure break optimization, the accuracy and reliability of the knowledge graph can be effectively improved, thereby enhancing its reliability for intelligent operation and maintenance in the applied domain.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to knowledge graph optimization methods, apparatus and electronic devices. Background Technology

[0002] Knowledge graphs, as a structured semantic knowledge base, can effectively represent entities and the relationships between them, providing powerful knowledge support. They are commonly used in intelligent operation and maintenance scenarios in fields such as ICT. ICT is an abbreviation for Information and Communications Technology.

[0003] Currently, in practical applications, for a specific field such as ICT, an initial knowledge graph is typically constructed directly from raw knowledge data within that field, such as device logs, configuration scripts, and technical manuals, for intelligent operation and maintenance within that field. However, the accuracy and reliability of such an initial knowledge graph are relatively low, thus affecting the reliability of intelligent operation and maintenance in that field. Summary of the Invention

[0004] In view of this, this application provides a knowledge graph optimization method, apparatus, and electronic device to improve the accuracy and reliability of knowledge graphs.

[0005] This application provides a knowledge graph optimization method, the method comprising: The entity embedding vectors corresponding to the knowledge graph are analyzed to identify and delete outlier entities in the knowledge graph that meet the outlier requirements; the entity embedding vectors are obtained by mapping the entities and relations in the knowledge graph to a continuous vector space. The structural features of each entity in the knowledge graph are analyzed to identify and delete meaningless entities; the structural features of the meaningless entities meet the set feature requirements. Using a large language model, semantic relevance analysis is performed on each entity in the knowledge graph and the keywords in the domain where the knowledge graph is set for application, so as to identify and delete entities that meet the requirement of semantic irrelevance. The knowledge graph is pruned and its structure optimized to obtain a target knowledge graph. The pruning is used to remove redundant entities in the knowledge graph. The structure optimization is used to suppress graph structure breaks caused by the deletion of entities or connections in the knowledge graph. The structure optimization includes at least: merging connections in the knowledge graph that meet similarity requirements and supplementing connections in the knowledge graph by introducing new credible connections.

[0006] This application embodiment also provides a knowledge graph optimization device, the device comprising: The first optimization module is used to analyze the entity embedding vectors corresponding to the knowledge graph in order to identify and delete outlier entities in the knowledge graph that meet the outlier requirements; the entity embedding vectors are obtained by mapping the entities and relations in the knowledge graph to a continuous vector space. The second optimization module is used to analyze the structural features of each entity in the knowledge graph in order to identify and delete meaningless entities in the knowledge graph; the structural features of the meaningless entities meet the set feature requirements; The third optimization module is used to perform semantic relevance analysis on each entity in the knowledge graph and the keywords in the domain where the knowledge graph is set for application using a large language model, so as to obtain and delete entities that meet the requirement of semantic irrelevance. The fourth optimization module is used to prune and optimize the structure of the knowledge graph to obtain the target knowledge graph. The pruning is used to remove redundant entities in the knowledge graph. The structure optimization is used to suppress graph structure breakage caused by the deletion of entities or connections in the knowledge graph. The structure optimization includes at least: merging connections in the knowledge graph that meet similarity requirements and supplementing connections by introducing new credible connections in the knowledge graph.

[0007] This application also provides an electronic device, which includes: Processor; and A computer-readable storage medium storing computer program instructions that, when executed by a processor, cause the processor to perform the steps of the method described above.

[0008] This application also provides a computer-readable storage medium storing computer program instructions that, when executed by a processor, cause the processor to perform the steps in the above method.

[0009] As can be seen from the above technical solutions, in this embodiment, by analyzing the entity embedding vectors corresponding to the knowledge graph, outlier entities in the knowledge graph are identified and filtered out from the semantic level represented by the entity embedding vectors; furthermore, by analyzing the structural features of each entity in the knowledge graph, meaningless entities in the knowledge graph that meet the set feature requirements, such as sparse connection structures, are identified and filtered out from the structural feature level; a large language model is used to perform semantic analysis on each entity in the knowledge graph, to identify and filter out entities in the knowledge graph that are irrelevant to the application domain from the semantic level again; and further, by pruning and optimizing the knowledge graph structure, redundant entities in the knowledge graph are identified and filtered out, and similar connection relationships and different entities with graph structure breaks are optimized. Thus, through filtering multiple noisy entities such as outliers and meaningless entities, and optimizing the graph structure with breaks, the accuracy and reliability of the knowledge graph can be effectively improved, thereby enhancing its reliability in intelligent operation and maintenance in the application domain. Attached Figure Description

[0010] The accompanying drawings, which are incorporated in and form part of this application, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0011] Figure 1 This is a schematic diagram of a method provided in an embodiment of this application.

[0012] Figure 2 This is a schematic diagram of another method flow provided for an embodiment of this application.

[0013] Figure 3 This is a schematic diagram of another method flow provided in an embodiment of this application.

[0014] Figure 4 This is a schematic diagram of another method flow provided in an embodiment of this application.

[0015] Figure 5 This is a schematic diagram of the device structure provided in the embodiments of this application.

[0016] Figure 6 This is a schematic diagram of the electronic device structure provided in an embodiment of this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solutions provided in the embodiments of this application, and to make the above-mentioned objectives, features and advantages of the embodiments of this application more apparent and understandable, the technical solutions in the embodiments of this application will be further described in detail below with reference to the accompanying drawings.

[0018] See Figure 1 , Figure 1This is a flowchart illustrating a method provided in an embodiment of this application. In this embodiment, the method can be applied to an electronic device, optionally such as a laptop, desktop computer, or server, etc., and is not specifically limited thereto. Figure 1 As shown, the process may include the following steps: Step 101: Analyze the entity embedding vectors corresponding to the knowledge graph to identify and delete outlier entities in the knowledge graph that meet the outlier requirements.

[0019] In this embodiment, the knowledge graph can be a knowledge graph constructed from original knowledge data (such as device logs, configuration scripts, technical manuals, etc.) in the domain of the application being defined. The specific method for constructing this knowledge graph will be illustrated below and will not be elaborated upon here.

[0020] In practical applications, when automatically extracting entities and their relationships from raw knowledge data (such as unstructured or semi-structured data like device logs, configuration scripts, and technical manuals) within a specific domain to construct a corresponding knowledge graph, a large number of noisy entities are often introduced. These noisy entities may include random characters, parsing errors, or words unrelated to the core concepts of the domain. The presence of these noisy entities not only increases the size of the knowledge graph and reduces storage and computational efficiency but also misleads subsequent knowledge reasoning and applications, thus affecting the accuracy and reliability of the knowledge graph.

[0021] Based on this, this embodiment can optimize the knowledge graph and improve its accuracy and reliability by identifying and filtering noisy entities in the knowledge graph. For example, in this embodiment, the entity embedding vectors corresponding to the knowledge graph can be analyzed first to identify and delete outlier entities in the knowledge graph that meet the outlier requirements.

[0022] In this embodiment, the entity embedding vector can be obtained by mapping entities and relations in the knowledge graph to a continuous vector space. Optionally, for example, a graph neural network (GNN) can be used to map entities and relations in the knowledge graph to a continuous vector space to obtain the entity embedding vector of each entity in the knowledge graph; wherein, the entity embedding vector of any entity at least indicates the entity attribute information of the entity, the entity and the association relationship between the entity and other entities in the knowledge graph, and the entity attribute information of the other entities.

[0023] In this embodiment, the outlier entity that meets the outlier requirement can refer to an entity that significantly deviates from the behavior or characteristics of most other entities in the knowledge graph. For example, the distance between the entity embedding vector of the entity and most other entities is greater than a set distance threshold, or the entity density is located in a preset low-density region, or the entity embedding vector is abnormal, etc. There is no specific limitation here.

[0024] As for how to analyze the entity embedding vectors corresponding to the knowledge graph in order to identify outlier entities that meet the outlier requirements, examples will be given below, and will not be elaborated here.

[0025] Step 102: Analyze the structural features of each entity in the knowledge graph to identify and delete meaningless entities; the structural features of meaningless entities meet the set feature requirements.

[0026] In this embodiment, the above structural features are not specifically limited, but may include at least: degree centrality, betweenness centrality, entity importance, and entity group attributes.

[0027] Degree centrality can refer to the number of edges directly connected to an entity in a knowledge graph. For example, the lower the degree centrality of an entity, that is, the fewer edges directly connected to that entity in the knowledge graph, the lower the importance of that entity.

[0028] Betweenness centrality can be defined as the frequency with which an entity appears on the shortest path between all other entity pairs in a knowledge graph. Here, "other entity pairs" can refer to any pair of entities in the knowledge graph other than the entity in question. For example, the lower the betweenness centrality of an entity, that is, the lower the frequency with which the entity appears on the shortest path between all other entity pairs in the knowledge graph, the more likely that the entity is located at the edge of the knowledge graph, has weaker information transmission capabilities, and therefore has lower importance.

[0029] Entity importance refers to the significance of an entity within a knowledge graph, such as its PageRank value. Here, the PageRank value can be calculated using the PageRank algorithm. The basic principle of the PageRank algorithm is to define a random walk model, i.e., a first-order Markov chain, on a directed graph such as a knowledge graph. This model describes the behavior of a random walker randomly visiting nodes along the directed graph, eventually converging to a stationary distribution. In this distribution, the probability of each node being visited can be considered its PageRank value, representing the node's importance. For example, a lower PageRank value indicates a lower node importance.

[0030] The "belonging to an entity group" attribute refers to the attribute of the entity group to which an entity belongs. Entity groups can be obtained by subdividing the knowledge graph into subdomains, with entities within the same entity group belonging to the same subdomain. Here, a subdomain can refer to a subdomain within the domain to which the knowledge graph is applied. The specific method for subdividing the knowledge graph to obtain at least one entity group will be described with examples below and will not be elaborated upon here.

[0031] Attributes of an entity group may include its size (i.e., the number of entities it contains) and connectivity. The size of an entity group is represented by the number of entities it contains. The connectivity of an entity group is represented by the number of connecting edges between its entities; the greater the number of connecting edges, the greater the connectivity of the entity group. Therefore, if an entity belongs to a smaller entity group and / or has lower connectivity, it indicates that the entity is less important.

[0032] Based on the above description, in this embodiment, as an example, satisfying the set feature requirement may mean that at least one structural feature of the entity is less than or equal to the set structural feature threshold of that structural feature.

[0033] Step 103: Using a large language model, perform semantic relevance analysis on each entity in the knowledge graph and the keywords in the domain where the knowledge graph is applied, so as to obtain entities that meet the requirement of semantic irrelevance and delete them.

[0034] In this embodiment, as an example, for the domain where the knowledge graph is to be applied, a keyword dictionary for that domain can be pre-constructed based on actual application needs. This keyword dictionary can contain multiple keywords related to that domain. Based on this, a large language model can be used to perform semantic relevance analysis on each entity in the knowledge graph and the keywords in the domain where the knowledge graph is to be applied, so as to obtain and delete entities that meet the semantic irrelevance requirement. Optionally, the semantic relevance analysis here can be, for example, semantic similarity calculation; if the semantic similarity between an entity in the knowledge graph and each keyword in the keyword dictionary is less than a preset similarity threshold, then the entity can be determined to be an entity that meets the semantic irrelevance requirement.

[0035] As another embodiment, a large language model can be used directly to perform semantic analysis on each entity in the knowledge graph. Based on the semantic analysis results, it can be determined whether each entity is meaningful in the domain context in which the knowledge graph is applied (e.g., the original knowledge data in that domain). If it is meaningless, the entity can be determined to be an entity that meets the requirement of semantic irrelevance. Here, "meaningful" can be understood as the entity having a clear technical meaning in the domain context, corresponding to an actual object (e.g., device, interface, protocol, etc.), and being able to assume an effective semantic role in the domain context, rather than random characters, invalid symbols, semantically ambiguous content, or irrelevant domain terms without actual technical meaning.

[0036] In this embodiment, the entity filtering method, which integrates multiple dimensions such as entity semantics and entity structural features, can comprehensively and accurately identify and filter out noisy entities in the knowledge graph, thereby effectively improving the quality of the knowledge graph and the reliability of subsequent applications.

[0037] Step 104: Prune and optimize the knowledge graph to obtain the target knowledge graph. Pruning is used to remove redundant entities in the knowledge graph. Optimization is used to suppress graph structure breaks caused by the deletion of entities or connections in the knowledge graph. Optimization includes at least: merging connections in the knowledge graph that meet similarity requirements and supplementing connections by introducing new trustworthy connections in the knowledge graph.

[0038] In practical applications, the sources of raw knowledge data in a given domain are wide-ranging and diverse in format. For example, raw knowledge data in the IC field may include device configuration information, network traffic data, security event logs, and protocol specification documents. This data is typically scattered across different systems and business departments, lacking a unified semantic standard and structure. Therefore, in the process of integrating this data into a knowledge graph, without effective global planning and guidance from specialized domain knowledge, the constructed knowledge graph is prone to problems such as loose structure, ambiguous hierarchy, and unclear topic division. Such loosely structured knowledge graphs struggle to support efficient knowledge navigation, querying, and logical reasoning, thus limiting the effectiveness of knowledge graphs in scenarios such as complex network analysis and intelligent decision-making.

[0039] Furthermore, the knowledge graphs constructed as described above often contain many isolated entities. Due to inherent limitations in information extraction and incomplete data sources, some entities cannot establish effective relationships with other entities, thus forming isolated entities in the knowledge graph. Because these isolated entities lack contextual information within the knowledge graph, they struggle to demonstrate their practical value in knowledge reasoning and application. This also significantly reduces the overall connectivity and usability of the knowledge graph, making it difficult to effectively perform analytical tasks such as path discovery and community detection based on the knowledge graph. For example, for an entity representing a specific network device, if it lacks effective relationships with other entities in the knowledge graph, such as being connected to, configured as, or running services, its application value within the knowledge graph will be greatly limited.

[0040] Based on this, as an example, after identifying and filtering noisy entities in the knowledge graph, this embodiment can further prune and optimize the knowledge graph to remove redundant entities, merge connections that meet similarity requirements, and complete connections by introducing new trustworthy connections, thereby reducing the number of isolated entities and enhancing the overall structure and relevance of the knowledge graph to solve the problems of loose structure and isolated entities in the knowledge graph.

[0041] As for how to prune and optimize the structure of the knowledge graph to obtain the target knowledge graph, examples will be given below, and will not be elaborated here.

[0042] This concludes the process. Figure 1 The process is shown below.

[0043] pass Figure 1 As shown in the flowchart, in this embodiment, by analyzing the entity embedding vectors corresponding to the knowledge graph, outliers in the knowledge graph are identified and filtered out at the semantic level represented by the entity embedding vectors. Furthermore, by analyzing the structural features of each entity in the knowledge graph, meaningless entities that meet set feature requirements, such as sparse connection structures, are identified and filtered out at the structural feature level. A large language model is also used to perform semantic analysis on each entity in the knowledge graph, again identifying and filtering out entities irrelevant to the application domain at the semantic level. Finally, by pruning and optimizing the knowledge graph, redundant entities are identified and filtered out, and similar connections and different entities with graph structure breaks are optimized. Through filtering multiple noisy entities, such as outliers and meaningless entities, and optimizing graph structure breaks, the accuracy and reliability of the knowledge graph can be effectively improved, thereby enhancing its reliability in intelligent operation and maintenance within the application domain.

[0044] The following is an example of step 101 above: In this embodiment, as an example, see [example]. Figure 2 As shown, the knowledge graph in step 101 above can be constructed, for example, through the following steps: Step 201: Perform data cleaning on the original knowledge data in the above-mentioned domains to remove abnormal data, inconsistent data, and redundant data from the original knowledge data.

[0045] Optionally, in this embodiment, there is no specific limitation on how to perform data cleaning. It can be flexibly set according to actual application needs. For example, abnormal data, inconsistent data and redundant data in the original knowledge data can be identified and removed based on preset data quality rules, regular expression matching, data type verification and simple statistical analysis methods.

[0046] Abnormal data refers to erroneous data that does not conform to preset formats, data types, syntax specifications, or domain rules, and cannot be parsed or used normally. For example, records with empty, garbled, missing, or illegally formatted device IDs; records with missing or invalid key configuration parameters; records where the number of unique values ​​in each field exceeds a set threshold (e.g., 1); and so on. Here, a unique value can refer to a specific numerical value or string that does not appear repeatedly in a field, such as the device ID in the device ID field, like "SW-1001-01" or "RT-2002-02".

[0047] Inconsistent data refers to data where the same entity, attribute, or configuration appears contradictory, inconsistently described, or with conflicting values ​​in different data sources and records, resulting in non-unique data semantics. For example, the same router might be labeled as different device models in different logs, or the same interface IP address might appear inconsistently in different configuration snippets; these are examples of inconsistent data.

[0048] Redundant data can refer to data that is completely duplicated, highly duplicated, or has no actual new information. For example, multiple identical log entries in a device log, configuration items that appear repeatedly in a configuration file with completely identical content, and multiple records with identical core entities, attributes, and relationships except for irrelevant fields (with no actual new information) are all considered redundant data.

[0049] Step 202: Denoise the original knowledge data to remove noisy data that does not meet the semantic features of the above domain, and obtain the target knowledge data.

[0050] Optionally, in this embodiment, word frequency can be counted to filter out words whose frequency is below a preset frequency threshold and does not conform to the aforementioned common terminology patterns in the field. The aforementioned common terminology patterns can be flexibly set based on actual application needs, such as whether they fall under the categories of technical terms like device, protocol, interface, IP address, service, or policy; this is not specifically limited here.

[0051] For example, in this embodiment, a stop word list for the aforementioned domain can be pre-constructed, which may include words that have no practical meaning in the domain but may appear in the original knowledge data of the domain, such as HIM, log, file, record, note, etc.; then, words in the original knowledge data that match the aforementioned stop word list can be regarded as noise data and filtered.

[0052] For example, in this embodiment, a terminology dictionary for the aforementioned field can be pre-constructed, which may contain a large number of standardized professional terms in that field. For example, network equipment terms (such as routers, switches, firewalls, etc.) and network protocol terms (such as BGP, TCP, IP, etc.). Then, Named-entity recognition (NER) technology can be used, combined with the terminology dictionary for the aforementioned field, to initially filter out words that do not match the terminology dictionary, treating them as noise data and filtering them out.

[0053] Step 203: Based on the preset entity types and relationship types related to the above-mentioned domains, construct a knowledge graph for the target knowledge data to obtain the knowledge graph.

[0054] In this embodiment, entity types and relationship types (which can also be referred to as domain rules) related to the above-mentioned domains can be pre-defined based on industry standards (such as communication protocol standards), technical documents (such as equipment manuals, technical white papers, etc.), expert experience, and existing network models (such as network topology models containing standardized entities and their relationships, etc.) for the construction of knowledge graphs.

[0055] For example, in the ICT field, it can be stipulated that the "router" entity and the "interface" entity are connected through an "owner" relationship, while the "interface" entity and the "IP address" entity can be connected through a "configured as" relationship, and so on. These domain rules can serve as constraints to guide entity identification and relationship extraction models during the entity and relationship extraction process, or for verification and correction after extraction. For instance, suppose a triple "router"-"run"-"user" is extracted, which obviously does not conform to common sense in ICT networks. In this case, it can be marked as an error or corrected according to the domain rules.

[0056] In this embodiment, through the above-mentioned data cleaning, denoising, and domain rule constraints, the amount of noisy data input to the knowledge graph can be significantly reduced, and the types and structures of entities and relationships can be initially standardized, providing cleaner and higher-quality knowledge graph input for subsequent optimization steps such as knowledge graph embedding, community detection, and noisy entity filtering.

[0057] In this embodiment, as an example, see [example]. Figure 3 As shown, step 101 above analyzes the entity embedding vectors corresponding to the knowledge graph to identify outlier entities in the knowledge graph that meet the outlier requirements. In specific implementations, this may include the following steps: Step 301: Perform cluster analysis on the entity embedding vectors corresponding to the knowledge graph to identify entities corresponding to entity embedding vectors that do not belong to any cluster as outliers.

[0058] Optionally, in this embodiment, clustering algorithms such as K-means or DBSCAN can be used to perform clustering analysis on the entity embedding vectors corresponding to the knowledge graph, so as to identify entities corresponding to entity embedding vectors that do not belong to any cluster as outliers. DBSCAN is an abbreviation for Density-Based Spatial Clustering of Applications with Noise. For details on how to perform clustering analysis, please refer to the descriptions of related technologies; specific limitations are not limited here.

[0059] Step 302: Perform density analysis on the entity embedding vectors corresponding to the knowledge graph, and identify entities corresponding to entity embedding vectors whose density analysis results meet the set low density requirements as outlier entities.

[0060] Optionally, in this embodiment, a density algorithm such as the Local Outlier Factor (LOF) algorithm can be used to calculate the entity embedding vector of each entity in the knowledge graph to obtain the local outlier factor corresponding to each entity. The local outlier factor represents the degree of outlier of the entity (which can also be regarded as local density, i.e., the number of entities in the preset neighborhood of the entity). The larger the factor value, the higher the degree of outlier, and the smaller the factor value, the lower the degree of outlier. Then, if the local outlier factor of an entity is greater than the preset density threshold, it means that the entity is in a low-density area and the entity is an outlier entity.

[0061] Step 303: Perform anomaly analysis on the magnitude and vector value of the entity embedding vectors corresponding to the knowledge graph, so as to identify the entities corresponding to the entity embedding vectors that meet the anomaly conditions as outlier entities.

[0062] Optionally, in this embodiment, for each entity in the knowledge graph, if the magnitude of the entity's entity embedding vector is less than a preset magnitude threshold (such as 0.1), then the entity is determined to be an outlier entity; or, if the vector value in a specified dimension of the entity's entity embedding vector is less than a set vector threshold (such as 0.1), then the entity is determined to be an outlier entity. Here, the specified dimension can be, for example, the dimension in the x or y direction.

[0063] The following example illustrates how to obtain the entity embedding vectors corresponding to the aforementioned knowledge graph: In this embodiment, as an example, the first step is to select a suitable graph neural network algorithm. For instance, considering that knowledge graphs in a particular domain typically possess complex topological structures and rich node and edge attributes, this embodiment preferably employs advanced graph neural network models, including but not limited to Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT). GCN can effectively capture the local structural features of knowledge graphs by aggregating feature information from the neighborhood of nodes. GAT introduces an attention mechanism, allowing the model to assign differentiated importance weights to different neighboring nodes during the learning process, thus more flexibly adapting to complex graph structures. Furthermore, this embodiment can also employ other more advanced Knowledge Graph Embedding (KGE) models, or models combining textual description information such as KG-BERT. These models learn vector representations of entities and relations, making semantically similar entities or relations appear closer in the vector space, thereby improving the semantic representation effect of entities and relations.

[0064] Next, embedding training and representation learning are performed on the knowledge graph. Specifically, a pre-selected model, such as GNN or KGE, is used to train the knowledge graph to learn the vector representations of entities and relations in the knowledge graph, i.e., entity embedding vectors. This training process typically aims to optimize an objective function so that existing true triples (i.e., head entity, relation, and tail entity) in the knowledge graph receive higher scores, while non-existent false triples receive lower scores. For example, in translation-based knowledge graph embedding models (such as the TransE model), the training objective is to minimize the distance between the sum of the head entity vector and relation vector and the tail entity vector. During training, domain characteristics can be further incorporated to enhance the representation learning effect. For example, entity attribute information (such as device model, manufacturer, IP address, etc.) can be used as initial features of nodes, or relation description information can be introduced into the model to enhance the vector representation capability of relations.

[0065] In this embodiment, through knowledge graph embedding and representation learning, discrete graph structure data can be transformed into continuous numerical vector representations. These entity embedding vectors can effectively represent the semantics of entities and relationships. For example, in vector space, the embedding vectors of network devices with similar functions (such as different models of routers) may cluster together, while the embedding vectors of entity pairs with specific dependencies (such as a service depending on a port) may have specific geometric relationships. These learned embedding vectors characterize the potential semantic similarity and association between entities and relationships, and can serve as important input features for subsequent steps (such as community detection, noisy entity recognition, etc.), enabling a deeper understanding and optimization of the knowledge graph.

[0066] The following is an example of the entity group in step 102 above: In this embodiment, after obtaining the entity embedding vectors corresponding to the knowledge graph, community detection and knowledge clustering techniques can be used to cluster the entities in the knowledge graph to identify sets of entities with tight internal connections and sparse external connections, i.e., "communities" (i.e., the aforementioned entity groups). This can effectively improve the problems of loose structure, ambiguous hierarchy, and unclear topic division in knowledge graphs. For example, by dividing a large-scale, complex knowledge graph into multiple relatively independent communities, the overall structure of the graph can be simplified, facilitating local optimization, knowledge navigation, and focused analysis.

[0067] Specifically, firstly, based on practical application requirements such as the size, sparsity, and preset community granularity of the knowledge graph, a suitable community detection algorithm is flexibly selected. For example, this may include, but is not limited to, the Leiden algorithm and the Louvain algorithm. These algorithms are all based on the principle of modularity optimization to divide entities in the knowledge graph into different communities, making the connection density between entities within a community significantly higher than the connection density between communities. The Leiden algorithm is an improved version of the Louvain algorithm, which can obtain higher quality and more stable community partitioning results, and can avoid the problem of small, scattered, and arbitrarily connected communities that the Louvain algorithm easily produces. In addition, this embodiment may also use spectral clustering algorithms, label propagation algorithms, etc., without specific limitations.

[0068] Then, community partitioning and knowledge clustering are performed. Specifically, the selected community detection algorithm is used to partition the entities in the knowledge graph into communities. The input of this algorithm can be, for example, the adjacency matrix of the knowledge graph, or an augmented graph representation combining entity embedding vectors. The output of this algorithm, i.e., the community detection result, is the allocation of entities in the knowledge graph to different communities, with each community corresponding to a relatively independent knowledge topic or functional module. For example, in the knowledge graph of the ICT network domain, communities can be partitioned according to topics such as "network security," "routing protocols," "data center networks," and "wireless access" (the topic here can be understood as the subdomain to which the above entity groups belong). Entities within the same community usually have high semantic relevance and structural tightness; for example, taking the network security community as an example, this community can contain entities such as firewalls, intrusion detection systems, vulnerabilities, and attack patterns, and these entities can be closely linked through relationships such as "protection, detection, and exploitation."

[0069] Finally, community summaries and applications are performed. For each community identified above, a community summary or representative description can be generated. This can be achieved by extracting core entities and high-frequency relationships within the community, or by using a large language model to semantically summarize the entity and relationship descriptions within the community. These community summaries or representative descriptions can be used to enhance knowledge retrieval, such as improving the global retrieval performance of GraphRAG (a graph-based retrieval enhancement generation technique), and can also guide subsequent noisy entity filtering and knowledge graph pruning. This embodiment, through community detection, can reconstruct a loosely structured knowledge graph into a modular, clear, and hierarchical structured knowledge system, significantly improving the understandability and manageability of the knowledge graph and providing reliable support for subsequent knowledge graph optimization.

[0070] The following example illustrates how to prune and optimize the structure of the knowledge graph in step 103 above to obtain the target knowledge graph: In this embodiment, as an example, see [example]. Figure 4 As shown, the above-mentioned pruning and structural optimization of the knowledge graph to obtain the target knowledge graph may include the following steps in its specific implementation: Step 401: Entities in the knowledge graph whose importance is less than or equal to a preset importance threshold are identified as redundant entities and removed.

[0071] Optionally, in this embodiment, as described above, entity importance can be represented by the entity's PageRank value. Based on this, for example, entities in the knowledge graph with PageRank values ​​less than or equal to a preset importance threshold can be considered redundant entities and removed. It should be noted that the preset importance threshold here can be the same as or different from the structural feature threshold corresponding to the entity importance structural feature in step 102 above.

[0072] Step 402: Based on the original knowledge data in the above-mentioned domain and the connection topology of each entity in the current knowledge graph, new reliable connection relationships are obtained through reasoning analysis to connect different entities in the knowledge graph where graph structure breaks.

[0073] Optionally, in this embodiment, for example, a large model can be used to compare the current knowledge graph with the initially constructed knowledge graph to infer and analyze the entities in the current knowledge graph that have graph structure breaks. Then, for any two entities in the current knowledge graph that have graph structure breaks, the large model can be used to infer and analyze whether there is a relationship between the two entities based on the original knowledge data in the above-mentioned domain and the connection topology of each entity in the current knowledge graph. If there is a relationship, the two entities are connected based on the relationship obtained by inference.

[0074] For example, suppose entity A in a knowledge graph is connected to entity B, and entity B is connected to entity C, but there is no direct connection between entity A and entity C. If entity B is deleted, then entity A and entity C can be understood as entities with a broken graph structure. In this case, we can infer that there may be some indirect connection between entity A and entity C based on the transitivity of the network topology, and connect entity A and entity C based on the inferred connection relationship.

[0075] Furthermore, in this embodiment, after obtaining new reliable connection relationships through reasoning analysis to connect different entities in the knowledge graph that have experienced graph structure breaks, if there are still isolated entities in the knowledge graph that have not established effective connections with any other entities, then the isolated entities can be directly deleted.

[0076] Step 403: Modify the relation attributes of the connections in the knowledge graph that meet the similarity requirements to the same relation attributes.

[0077] Optionally, in this embodiment, for example, the semantic similarity between the connections can be calculated. If the semantic similarity is greater than a set similarity threshold, it means that the connections meet the similarity requirements. Then, the relationship attributes of the connections in the knowledge graph that meet the similarity requirements can be uniformly modified to the relationship attributes of one of the connections that meet the similarity requirements.

[0078] Based on the above description, the knowledge graph optimization method described above will be further described below: In this embodiment, after optimizing the knowledge graph as described above, the GraphRAG-based retrieval and generation mechanism can be further optimized to fully utilize the optimized knowledge graph and improve the performance of question answering and information generation.

[0079] For example, as an implementation, when a user question in the aforementioned domain is received (such as "What could be the reason for the frequent downtime of the core router RT-100 port?"), the target entity group matching the user question can be determined based on the group description information (i.e., the aforementioned community summary or representative description) of the entity group to which each entity belongs in the target knowledge graph. The entity group is obtained by dividing the knowledge graph into subdomains, and entities in the same entity group belong to the same subdomain. Then, the target answer to the user question can be determined based on the entities in the target entity group and the connection relationships between entities.

[0080] Specifically, during the retrieval phase, instead of relying solely on traditional entity matching or keyword matching methods, the process further integrates the structural and semantic information unique to knowledge graphs. First, utilizing the community detection results obtained above, upon receiving a user question, the topic related to the question is identified, and searches are prioritized within communities related to that topic. This narrows the search scope, improves search efficiency, and reduces interference from irrelevant information. Second, combining the entity importance obtained through algorithms such as PageRank, when retrieving subgraphs related to the user question, priority is given to retrieving entities with higher importance and their relationships within the knowledge graph to ensure the density of key information in the search content. Furthermore, the retrieval process can consider the semantic coherence of relation types and connection paths, avoiding the return of discrete, fragmented entities or relationships, and instead outputting knowledge fragments or related subgraphs with complete semantics as much as possible.

[0081] Furthermore, in this embodiment, during the retrieval phase, a preset scoring mechanism can be used to score the association between entities and relationships within the community that match the user's question and the user's question. Entities and relationships with scores below a preset threshold are then removed to further narrow the retrieval scope and improve retrieval efficiency. The specific scoring method is not limited in this embodiment; for example, scoring can be based on semantic similarity to the user's question, with higher semantic similarity resulting in a higher score.

[0082] During the generation phase, when the large language model receives information retrieved from the knowledge graph (i.e., the subgraphs related to the user's question, such as those containing entities, relationships, and community summaries), it can convert the retrieved information into a structured text description adapted to the large language model's understanding and processing, and then input it into the large language model. Alternatively, the retrieved information can be input into the large language model in a preset format, and a prompt engineering strategy can be employed. Preset prompts guide the large language model to focus on key entities and key relationships in the input retrieval information for reasoning and analysis, thereby generating a more comprehensive and accurate answer. The specific setting of the prompts is not specifically limited in this embodiment and can be flexibly set based on actual needs.

[0083] In this embodiment, the aforementioned knowledge graph construction, optimization, and retrieval can be uniformly constructed into a graph-based retrieval enhancement generation architecture, namely the GraphRAG architecture. To ensure that the quality of the knowledge graph and the performance of the GraphRAG architecture can continuously improve over time and with changes in application scenarios, this embodiment introduces a continuous learning and model optimization mechanism. This mechanism uses user feedback data, expert-annotated data, and newly added knowledge data as iterative basis to iteratively update and optimize the machine learning models related to the GraphRAG architecture (such as entity recognition models, relation extraction models, GNN embedding models, and large language models).

[0084] Specifically, as an example, in the actual application of the GraphRAG architecture, user questions and corresponding output answers are logged and stored. Users can evaluate the output answers (including but not limited to evaluation labels such as "valid," "invalid," and "incorrect") or directly provide correction suggestions. This user feedback data can be used to construct a model optimization training sample set to evaluate the performance of the current architecture and provide data support for subsequent knowledge correction and model fine-tuning. For example, if users frequently point out errors in entity recognition or relation extraction, adjustments to the recognition and extraction rules in the corresponding entity recognition and relation extraction models can be triggered, such as manual adjustments or model fine-tuning training based on user feedback data. The specific methods for model fine-tuning training are not specifically limited here.

[0085] In this embodiment, for knowledge data in key domains or knowledge data for which accurate knowledge graphs are difficult to generate automatically, experts in the field can be introduced to perform manual annotation and knowledge verification. For example, experts can review information such as entities, relationships, and relationship attributes in the knowledge graph to correct erroneous entities and relationships, and supplement missing entities and relationships. They can also provide optimization suggestions for the entity recognition and relationship extraction rules of the graph, such as adding, deleting, or modifying entities and connections. The data obtained from the above expert annotation and knowledge verification can be directly used as retraining for the knowledge graph construction model, or as prior constraints to guide the graph generation of the current knowledge graph construction model.

[0086] In this embodiment, taking the ICT field as an example, due to the dynamic nature of the ICT network environment, new network devices, communication protocols, security vulnerabilities, configuration specifications, and other information are constantly emerging, resulting in a large amount of new knowledge data (such as new technical documents, security announcements, network logs, etc.). This embodiment can use the knowledge graph construction and optimization methods provided in the above embodiments, employing techniques such as incremental entity and relation extraction, entity linking, and conflict resolution, to update and optimize the current knowledge graph based on the new knowledge data, ensuring that the new knowledge is compatible with the original knowledge and free from contradictions. Simultaneously, the aforementioned entity recognition model, relation extraction model, embedding representation model, and other related machine learning models can also be periodically fine-tuned using the new knowledge data to adapt the models to the dynamic changes in knowledge distribution, improving the model's generalization ability and accuracy. The specific methods for fine-tuning training are not specifically limited here.

[0087] In this embodiment, pre-prepared test data (such as multiple user questions and answers for testing, and multiple technical documents for testing) can be periodically used to quantitatively evaluate the knowledge graph generation quality and question-answering performance of the GraphRAG architecture. The knowledge graph generation quality evaluation may include dimensions such as accuracy and completeness, while the question-answering performance evaluation may include dimensions such as response accuracy, response latency, and user satisfaction. Afterwards, based on the evaluation results, model performance degradation trends or potential problems can be identified, automatically triggering corresponding optimization processes. These processes may include adjustments to entity recognition and extraction rules, retraining of the overall model, etc., which are not specifically limited here.

[0088] This embodiment, through this closed-loop mechanism of continuous learning and model optimization, ensures that the knowledge graph always maintains high quality and timeliness, and the GraphRAG architecture can continuously adapt to new needs and challenges, thereby playing a greater role in long-term operation and application.

[0089] This concludes the description of the method provided in the embodiments of this application. The apparatus provided in the embodiments of this application will now be described: As an example, this embodiment also provides a knowledge graph optimization device. For example, see... Figure 5 , Figure 5 This is a schematic diagram of a knowledge graph optimization device provided in an embodiment of this application. The device corresponds to… Figure 1 The process is shown below. Figure 5 As shown, the knowledge graph optimization device 500 includes: The first optimization module 501 is used to analyze the entity embedding vectors corresponding to the knowledge graph in order to identify and delete outlier entities in the knowledge graph that meet the outlier requirements; the entity embedding vectors are obtained by mapping the entities and relations in the knowledge graph to a continuous vector space. The second optimization module 502 is used to analyze the structural features of each entity in the knowledge graph in order to identify and delete meaningless entities in the knowledge graph; the structural features of the meaningless entities meet the set feature requirements; The third optimization module 503 is used to perform semantic relevance analysis on each entity in the knowledge graph and the keywords in the domain where the knowledge graph is set for application using a large language model, so as to obtain and delete entities that meet the semantic irrelevance requirement. The fourth optimization module 504 is used to prune and optimize the structure of the knowledge graph to obtain the target knowledge graph. The pruning is used to remove redundant entities in the knowledge graph. The structure optimization is used to suppress graph structure breakage caused by the deletion of entities or connections in the knowledge graph. The structure optimization includes at least: merging connections in the knowledge graph that meet similarity requirements and supplementing connections in the knowledge graph by introducing new credible connections.

[0090] As an example, the knowledge graph is constructed through the following steps: Data cleaning is performed on the original knowledge data in the aforementioned domain to remove abnormal, inconsistent, and redundant data. The original knowledge data is denoised to remove noisy data that does not meet the semantic features of the domain, thereby obtaining the target knowledge data. Based on preset entity types and relationship types related to the domain, a knowledge graph is constructed from the target knowledge data to obtain the knowledge graph.

[0091] As an example, the analysis of the entity embedding vectors corresponding to the knowledge graph to identify outlier entities in the knowledge graph that meet the outlier requirements includes: Cluster analysis is performed on the entity embedding vectors corresponding to the knowledge graph to identify entities whose entity embedding vectors do not belong to any cluster as outliers. Density analysis is performed on the entity embedding vectors corresponding to the knowledge graph, and entities corresponding to entity embedding vectors whose density analysis results meet the set low density requirements are identified as outlier entities. Anomaly analysis is performed on the magnitude and vector value of the entity embedding vectors corresponding to the knowledge graph to identify the entities corresponding to the entity embedding vectors that meet the anomaly conditions as the outlier entities.

[0092] As an example, the structural features include at least: degree centrality, betweenness centrality, entity importance, and entity group attribute; the entity group is obtained by sub-domain partitioning of the knowledge graph, and entities in the same entity group belong to the same sub-domain; The requirement of satisfying the set feature means that at least one structural feature of the entity is less than or equal to the set structural feature threshold of that structural feature.

[0093] As an example, the pruning and structural optimization of the knowledge graph to obtain the target knowledge graph includes: Entities in the knowledge graph whose importance is less than or equal to a preset importance threshold are considered redundant entities and removed. Based on the original knowledge data in the domain and the connection topology of each entity in the current knowledge graph, new reliable connection relationships are obtained through reasoning analysis to connect different entities in the knowledge graph that have experienced graph structure breaks. The relationship attributes of connections in the knowledge graph that meet the similarity requirements will be uniformly modified to the same relationship attributes.

[0094] As one embodiment, the device further includes: The receiving module is configured to, upon receiving a user question in the domain, determine a target entity group that matches the user question based on the group description information of the entity group to which each entity in the target knowledge graph belongs; the entity group is obtained by dividing the knowledge graph into subdomains, and entities in the same entity group belong to the same subdomain; Based on the entities in the target entity group and the connections between them, the target answer to the user question is determined.

[0095] This concludes the process. Figure 5 Structural description of the device shown.

[0096] The specific implementation process of the functions and roles of each module in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0097] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0098] Please see Figure 6 This is a schematic diagram of the hardware structure of an electronic device provided as an exemplary embodiment of this application. The electronic device includes a processor and a computer-readable storage medium; the computer-readable storage medium stores a plurality of computer program instructions, which, when executed by the processor, implement the method disclosed in the above example of this application. Depending on the actual function of the electronic device, other hardware may also be included, which will not be elaborated further.

[0099] Based on the same concept as the above method, this application also provides a computer-readable storage medium storing a plurality of computer program instructions, which, when executed by a processor, can implement the method disclosed in the above examples of this application.

[0100] For example, the aforementioned computer-readable storage medium can be any electronic, magnetic, optical, or other physical storage device that can contain or store information such as executable instructions, messages, etc. For instance, computer-readable storage media can be: RAM (Random Access Memory), volatile memory, non-volatile memory, flash memory, storage drives (such as hard disk drives), solid-state drives, any type of storage disk (such as optical discs, DVDs, etc.), or similar storage media, or combinations thereof.

[0101] The above are merely preferred embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A knowledge graph optimization method, characterized in that, The method includes: The entity embedding vectors corresponding to the knowledge graph are analyzed to identify and delete outlier entities in the knowledge graph that meet the outlier requirements; the entity embedding vectors are obtained by mapping the entities and relations in the knowledge graph to a continuous vector space. The structural features of each entity in the knowledge graph are analyzed to identify and delete meaningless entities; the structural features of the meaningless entities meet the set feature requirements. Using a large language model, semantic relevance analysis is performed on each entity in the knowledge graph and the keywords in the domain where the knowledge graph is set for application, so as to identify and delete entities that meet the requirement of semantic irrelevance. The knowledge graph is pruned and its structure optimized to obtain a target knowledge graph. The pruning is used to remove redundant entities in the knowledge graph. The structure optimization is used to suppress graph structure breaks caused by the deletion of entities or connections in the knowledge graph. The structure optimization includes at least: merging connections in the knowledge graph that meet similarity requirements and supplementing connections in the knowledge graph by introducing new credible connections.

2. The method according to claim 1, characterized in that, The knowledge graph is constructed through the following steps: Data cleaning is performed on the original knowledge data in the aforementioned domain to remove abnormal, inconsistent, and redundant data. The original knowledge data is denoised to remove noisy data that does not meet the semantic features of the domain, thereby obtaining the target knowledge data. Based on preset entity types and relationship types related to the domain, a knowledge graph is constructed from the target knowledge data to obtain the knowledge graph.

3. The method according to claim 1, characterized in that, The analysis of the entity embedding vectors corresponding to the knowledge graph to identify outlier entities in the knowledge graph that meet the outlier requirements includes: Cluster analysis is performed on the entity embedding vectors corresponding to the knowledge graph to identify entities whose entity embedding vectors do not belong to any cluster as outliers. Density analysis is performed on the entity embedding vectors corresponding to the knowledge graph, and entities corresponding to entity embedding vectors whose density analysis results meet the set low density requirements are identified as outlier entities. Anomaly analysis is performed on the magnitude and vector value of the entity embedding vectors corresponding to the knowledge graph to identify the entities corresponding to the entity embedding vectors that meet the anomaly conditions as the outlier entities.

4. The method according to claim 1, characterized in that, The structural features include at least: degree centrality, betweenness centrality, entity importance, and entity group attribute; the entity group is obtained by sub-domain division of the knowledge graph, and entities in the same entity group belong to the same sub-domain; The requirement of satisfying the set feature means that at least one structural feature of the entity is less than or equal to the set structural feature threshold of that structural feature.

5. The method according to claim 1, characterized in that, The process of pruning and structurally optimizing the knowledge graph to obtain the target knowledge graph includes: Entities in the knowledge graph whose importance is less than or equal to a preset importance threshold are considered redundant entities and removed. Based on the original knowledge data in the domain and the connection topology of each entity in the current knowledge graph, new reliable connection relationships are obtained through reasoning analysis to connect different entities in the knowledge graph that have experienced graph structure breaks. The relationship attributes of connections in the knowledge graph that meet the similarity requirements will be uniformly modified to the same relationship attributes.

6. The method according to claim 1, characterized in that, The method further includes: Upon receiving a user question in the aforementioned domain, a target entity group matching the user question is determined based on the group description information of the entity groups to which each entity belongs in the target knowledge graph. The entity group is obtained by dividing the knowledge graph into subdomains, and entities in the same entity group belong to the same subdomain. Based on the entities in the target entity group and the connections between them, the target answer to the user question is determined.

7. A knowledge graph optimization device, characterized in that, The device includes: The first optimization module is used to analyze the entity embedding vectors corresponding to the knowledge graph in order to identify and delete outlier entities in the knowledge graph that meet the outlier requirements; the entity embedding vectors are obtained by mapping the entities and relations in the knowledge graph to a continuous vector space. The second optimization module is used to analyze the structural features of each entity in the knowledge graph in order to identify and delete meaningless entities in the knowledge graph; the structural features of the meaningless entities meet the set feature requirements; The third optimization module is used to perform semantic relevance analysis on each entity in the knowledge graph and the keywords in the domain where the knowledge graph is set for application using a large language model, so as to obtain and delete entities that meet the requirement of semantic irrelevance. The fourth optimization module is used to prune and optimize the structure of the knowledge graph to obtain the target knowledge graph. The pruning is used to remove redundant entities in the knowledge graph. The structure optimization is used to suppress graph structure breakage caused by the deletion of entities or connections in the knowledge graph. The structure optimization includes at least: merging connections in the knowledge graph that meet similarity requirements and supplementing connections by introducing new credible connections in the knowledge graph.

8. The apparatus according to claim 7, characterized in that, The knowledge graph is constructed through the following steps: Data cleaning is performed on the original knowledge data in the aforementioned domain to remove abnormal, inconsistent, and redundant data. The original knowledge data is denoised to remove noisy data that does not meet the semantic features of the domain, thereby obtaining the target knowledge data. Based on preset entity types and relationship types related to the domain, a knowledge graph is constructed from the target knowledge data to obtain the knowledge graph; And / or, The analysis of the entity embedding vectors corresponding to the knowledge graph to identify outlier entities in the knowledge graph that meet the outlier requirements includes: Cluster analysis is performed on the entity embedding vectors corresponding to the knowledge graph to identify entities whose entity embedding vectors do not belong to any cluster as outliers. Density analysis is performed on the entity embedding vectors corresponding to the knowledge graph, and entities corresponding to entity embedding vectors whose density analysis results meet the set low density requirements are identified as outlier entities. Anomaly analysis is performed on the magnitude and vector value of the entity embedding vectors corresponding to the knowledge graph, so as to identify the entities corresponding to the entity embedding vectors that meet the anomaly conditions as the outlier entities. And / or, The structural features include at least: degree centrality, betweenness centrality, entity importance, and entity group attribute; the entity group is obtained by sub-domain division of the knowledge graph, and entities in the same entity group belong to the same sub-domain; The requirement of satisfying the set feature means that at least one structural feature of the entity is less than or equal to the set structural feature threshold of that structural feature. And / or, The process of pruning and structurally optimizing the knowledge graph to obtain the target knowledge graph includes: Entities in the knowledge graph whose importance is less than or equal to a preset importance threshold are considered redundant entities and removed. Based on the original knowledge data in the domain and the connection topology of each entity in the current knowledge graph, new reliable connection relationships are obtained through reasoning analysis to connect different entities in the knowledge graph that have experienced graph structure breaks. The relationship attributes of connections in the knowledge graph that meet the similarity requirements are uniformly modified to be the same relationship attributes. And / or, The device further includes: The receiving module is configured to, upon receiving a user question in the domain, determine a target entity group that matches the user question based on the group description information of the entity group to which each entity in the target knowledge graph belongs; the entity group is obtained by dividing the knowledge graph into subdomains, and entities in the same entity group belong to the same subdomain; Based on the entities in the target entity group and the connections between them, the target answer to the user question is determined.

9. An electronic device, characterized in that, The electronic device includes: Processor; and A computer-readable storage medium storing computer program instructions that, when executed by the processor, cause the processor to perform the steps of the method according to any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, cause the processor to perform the steps of the method as described in any one of claims 1 to 6.

Images