A knowledge graph node expansion method and device

Abstract

The specification provides a knowledge graph node expansion method and device, the method comprising: acquiring a knowledge graph node expansion strategy, the strategy containing strategy parameters for controlling the expansion direction of the knowledge graph; determining a candidate node set for the knowledge graph to be expanded, and acquiring current state information of the knowledge graph to be expanded; performing multi-dimensional evaluation on each candidate node in the candidate node set according to the current state information and the strategy parameters, to obtain a preference degree of each candidate node; selecting a target expansion node from the candidate node set according to the preference degree and an exploration control parameter in the strategy parameters; and adding the target expansion node and its associated relationship to the knowledge graph to be expanded.

Description

Technical Field

[0001] One or more embodiments of this specification relate to the field of knowledge graph technology, and in particular to a method and apparatus for expanding knowledge graph nodes. Background Technology

[0002] With the rapid development of artificial intelligence technology, the demand for high-quality and diverse training data for large-scale language models is increasing. Knowledge graphs, as the core carrier of structured knowledge representation, directly impact the performance of downstream tasks due to their construction quality. During knowledge graph expansion, node selection strategies determine the evolution of the graph's topology and the breadth of its domain coverage. Existing technologies mainly rely on random sampling or heuristic rules based on local topological features. For example, while weighted random walk methods introduce edge weight mechanisms, the weight parameters are usually statically preset, lacking dynamic perception of the global graph state. This local-perspective selection mechanism struggles to achieve macro-level goal-oriented control, causing graph growth to tend towards popular topics with high connectivity, resulting in a severe imbalance in domain distribution. Furthermore, some sampling algorithms based on competitive selection, focusing on the survival of the fittest among statistical variables, often ignore long-tail nodes with low connectivity but high potential value, further exacerbating the unevenness of knowledge coverage. In summary, existing node selection technologies generally suffer from rigid strategies and a lack of goal orientation. Traditional methods often employ fixed logic or purely random strategies, failing to adapt to specific construction goals. This results in significant randomness and uncontrollability in the constructed knowledge graphs. Particularly in data generation scenarios targeting specific domain distributions or structural characteristics, existing technologies struggle to balance scalability, domain balance, and structural diversity, easily leading to knowledge blind spots or local overfitting. This limits the effectiveness of knowledge graphs in scenarios requiring refined data provision. Summary of the Invention

[0003] The embodiments in this specification aim to provide a method and apparatus for expanding knowledge graph nodes, which can solve the technical problems in the prior art where the knowledge graph expansion process lacks a goal-oriented macro-control mechanism, resulting in uneven domain coverage and random and uncontrollable structural evolution paths in the constructed graph.

[0004] According to the first aspect, a method for expanding knowledge graph nodes is provided, characterized by comprising:

[0005] Obtain a knowledge graph node expansion strategy, the strategy including strategy parameters for controlling the expansion direction of the knowledge graph;

[0006] Determine the set of candidate nodes for the knowledge graph to be expanded, and obtain the current state information of the knowledge graph to be expanded;

[0007] Based on the current state information and the strategy parameters, each candidate node in the candidate node set is evaluated in multiple dimensions to obtain the preference of each candidate node;

[0008] Based on the preferredness and the exploration control parameters in the strategy parameters, target expansion nodes are selected from the candidate node set; the target expansion nodes and their associations with the knowledge graph to be expanded are added to the knowledge graph to be expanded.

[0009] In one possible implementation, the strategy parameters include domain distribution control parameters and depth and breadth exploration coefficients;

[0010] Each candidate node in the candidate node set is evaluated in multiple dimensions to obtain the optimality of each candidate node, including:

[0011] Based on the domain distribution information in the current state information and the domain distribution control parameters, calculate the domain balance score for each candidate node;

[0012] Based on the depth and breadth exploration coefficients, a structural exploration preference score is determined for each candidate node;

[0013] Based on the domain balance score and the structure exploration preference score, a comprehensive score is determined for each candidate node, and the comprehensive score is used as the preference degree.

[0014] In one possible implementation, calculating the domain balance score for each candidate node based on the domain distribution information in the current state information and the domain distribution control parameters includes:

[0015] Statistical analysis of the current node percentage in each domain of the knowledge graph to be expanded;

[0016] Calculate the difference between the current node percentage and the preset target percentage;

[0017] Based on the differences and the domain distribution control parameters, a domain balance score is determined for each candidate node.

[0018] In one possible implementation, determining the structure exploration preference score for each candidate node in the candidate node set based on the depth and breadth exploration coefficients includes:

[0019] Obtain the graph structure features of the candidate nodes, including betweenness centrality and clustering coefficient;

[0020] An exploration strategy is determined based on the depth and breadth exploration coefficients, and the exploration strategy includes a global exploration strategy or a local mining strategy.

[0021] If the exploration strategy is a global exploration strategy, the structure exploration preference score is calculated based on the betweenness centrality of the candidate nodes;

[0022] The exploration strategy is a local mining strategy, and the structural exploration preference score is calculated based on the clustering coefficient of the candidate nodes.

[0023] In one possible implementation, the multidimensional evaluation further includes determining a connectivity potential score for each candidate node in the candidate node set for the knowledge graph to be expanded;

[0024] Based on the domain balance score and the structure exploration preference score, a comprehensive score is determined for each candidate node, including:

[0025] A comprehensive score for each candidate node is determined by a weighted sum of the domain balance score, the structure exploration preference score, and the connectivity potential score.

[0026] In one possible implementation, determining the connectivity potential score of each candidate node in the candidate node set for the knowledge graph to be expanded includes:

[0027] Calculate the semantic similarity between the candidate node and its neighboring nodes in the knowledge graph to be expanded;

[0028] And / or, predict the contribution of the candidate node to the graph connectivity after its addition;

[0029] The connectivity potential score is calculated based on the semantic similarity and / or the contribution.

[0030] In one possible implementation, determining the target expansion node from the candidate node set based on the preference level and the exploration control parameters in the strategy parameters includes:

[0031] The preference degree is converted into an initial probability distribution using a normalization function;

[0032] By introducing a random factor from the strategy parameters as a temperature coefficient, the initial probability distribution is smoothed to obtain the final sampling probability distribution;

[0033] Based on the final sampling probability distribution, the target expansion node is determined from the candidate node set.

[0034] In one possible implementation, determining the set of candidate nodes for the knowledge graph to be expanded includes:

[0035] Select the boundary nodes in the knowledge graph to be expanded as seed nodes;

[0036] Retrieve external entities that have a potential association with the seed node;

[0037] The external entity is used as a candidate node, and a potential association between the candidate node and the seed node is established.

[0038] In one possible implementation, the domains to which the nodes in the knowledge graph to be expanded belong include multiple fields such as medical, financial, and technological fields.

[0039] According to the second aspect, a knowledge graph node expansion device is provided, characterized in that it includes:

[0040] The acquisition module is used to acquire knowledge graph node expansion strategies, wherein the strategies include strategy parameters for controlling the expansion direction of the knowledge graph;

[0041] The determination module is used to determine the set of candidate nodes for the knowledge graph to be expanded, and to obtain the current state information of the knowledge graph to be expanded.

[0042] The evaluation module is used to perform multi-dimensional evaluation on each candidate node in the candidate node set based on the current state information and the strategy parameters, so as to obtain the preference of each candidate node.

[0043] The filtering module is used to filter out target expansion nodes from the candidate node set based on the preference degree and the exploration control parameters in the strategy parameters;

[0044] An execution module is used to add the target expansion node and its association with the knowledge graph to be expanded to the knowledge graph.

[0045] According to a third aspect, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method described in the first aspect.

[0046] According to a fourth aspect, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method described in the first aspect.

[0047] By utilizing one or more of the methods, apparatuses, computing devices, and storage media mentioned above, goal-oriented control over the knowledge graph construction process can be achieved through the introduction of policy configuration files. Using domain distribution weights, the system can identify under-covered domains in the current graph and provide corresponding compensation during node selection, thereby ensuring a balanced distribution of the graph across multiple knowledge domains. Through depth and breadth exploration coefficients, the system can dynamically adjust the ratio of breadth to depth exploration, ensuring broad knowledge coverage while enabling in-depth mining of specific paths. Furthermore, by combining connectivity potential scoring and random factors, this method retains necessary exploration randomness while maintaining goal orientation, avoiding getting trapped in local optima and enhancing the diversity of knowledge discovery. Ultimately, this method increases the controllability of the knowledge graph's growth direction and final form, enabling the generation of high-quality, diverse knowledge graphs that meet specific training data requirements. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This illustration shows a flowchart of a knowledge graph node expansion method provided in one embodiment of this application;

[0050] Figure 2 This illustration shows a detailed flowchart of identifying a set of candidate nodes in one embodiment of this application;

[0051] Figure 3 This illustration shows a schematic diagram of establishing potential associations in one embodiment of this application;

[0052] Figure 4 This illustration shows a schematic diagram of the logical structure for calculating the comprehensive score of candidate nodes in one embodiment of this application;

[0053] Figure 5 This illustration shows a comparison of node selection tendencies under a global exploration strategy and a local mining strategy in one embodiment of this application;

[0054] Figure 6 This illustration shows a flowchart of determining the target expansion node based on weighted probability sampling in one embodiment of this application;

[0055] Figure 7 A schematic diagram of the structure of a knowledge graph node expansion device provided in one embodiment of this application is shown. Detailed Implementation

[0056] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0057] Figure 1 This diagram illustrates a flowchart of a knowledge graph node expansion method according to an embodiment of this application. This method can be applied to servers, cloud computing platforms, or local terminal devices with strong computing power to incrementally expand an initial knowledge graph to generate a target knowledge graph that meets the distribution and structural characteristics of a specific domain. Figure 1 As shown, the method mainly includes steps S110 to S150.

[0058] In step S110, a knowledge graph node expansion strategy is obtained, the strategy including strategy parameters for controlling the expansion direction of the knowledge graph.

[0059] The strategy is the core set of rules controlling the direction of graph expansion, containing static strategy parameters for controlling the expansion direction. In one embodiment, the strategy parameters can be stored in a structured data format such as JSON, XML, or YAML, containing parameters with three key dimensions: the target proportion of each domain, the depth and breadth exploration coefficients (also known as structure exploration coefficients), and a random factor. The target proportion of each domain is used to characterize the expected proportion of different knowledge domains in the target graph, and its value is usually normalized so that the sum of the proportions of all domains is 1. The structure exploration coefficient is used to balance the breadth-first search (BFS) and depth-first search (DFS) tendencies in the graph expansion process. The value of this coefficient is usually between 0 and 1, with a closer value to 1 indicating a greater tendency towards global breadth exploration and a closer value to 0 indicating a greater tendency towards local depth mining. The random factor is used to introduce uncertainty in the sampling process to avoid the algorithm getting trapped in local optima, and its role is similar to the temperature parameter in the simulated annealing algorithm or the smoothing coefficient in the probabilistic model. In one embodiment, the strategy parameters may also include domain distribution control parameters, used to characterize the expected importance or basic weight of different knowledge domains in the target graph. This parameter remains constant throughout the iteration process and is used to differentiate the weighting of nodes in different domains during the scoring phase.

[0060] In step S120, a set of candidate nodes for the knowledge graph to be expanded is determined, and the current state information of the knowledge graph to be expanded is obtained.

[0061] A knowledge graph to be expanded refers to graph structure data that is currently under construction and needs further addition of nodes and edges. The process of identifying the candidate node set is essentially determining the "search boundary" for the next round of expansion. Specifically, in an instantiation, for example, the knowledge graph to be expanded can be traversed first, selecting nodes located at the graph edge or with unclosed potential relationships as seed nodes. The current state information of the knowledge graph to be expanded can include the current node percentage in each domain. In one embodiment, the number of existing nodes in each knowledge domain (such as medical, financial, and technological) in the knowledge graph to be expanded can be determined through external knowledge base retrieval, entity linking algorithms, or the reasoning ability of a pre-trained language model. Then, the current node percentage in each domain can be calculated based on the existing nodes in each domain, thereby obtaining a snapshot of the global distribution state of the graph. This distribution will serve as a benchmark for subsequent scoring calculations. In different embodiments, the domain to which a node belongs in the knowledge graph to be expanded can be determined in different specific ways. In one embodiment, the domain to which a node belongs can be determined, for example, based on node labels or node attributes. In another embodiment, the domain to which a node belongs can be determined, for example, based on the domains to which its neighboring nodes belong.

[0062] Furthermore, in step S130, each candidate node in the candidate node set can be evaluated in multiple dimensions based on the current state information and strategy parameters to obtain the preference degree (also known as the comprehensive score) of each candidate node.

[0063] Specifically, in one embodiment, the preference of each candidate node can be obtained by determining multiple scores for each candidate node. The domain balance score aims to guide the knowledge graph towards a preset domain distribution (i.e., a preset proportion of each domain), correcting the "Matthew effect" in the graph growth process, that is, preventing the overexpansion of nodes in popular domains while nodes in less popular domains are ignored. When calculating the domain balance score, the knowledge domain to which the candidate node belongs can be identified first. If the candidate node belongs to a domain whose current proportion in the graph is lower than a preset target proportion, the node receives a higher positive incentive score; conversely, if it belongs to a domain with an excessively high proportion, it receives a lower score or even a negative penalty. The specific calculation logic can be implemented by comparing the difference between the current proportion and the target proportion. Let... This represents the percentage of nodes in region d within the current graph. The target percentage for domain d as set in the strategy configuration file. Let be the domain distribution control parameter for domain d. For a candidate node i belonging to domain d, its domain balance score is... It can be represented as:

[0064] (1)

[0065] Among them, when > When the difference is positive, it indicates that the area needs supplementation, and the score is positive; when < When the difference is negative, it indicates that the domain is saturated, and the score is negative. In this way, the algorithm can dynamically guide the expansion process towards scarce domains.

[0066] The structural exploration preference score is used to control the topological structure of the graph. Based on the depth and breadth exploration coefficients α in the strategy configuration file, the system determines the current exploration strategy. If α is greater than a preset threshold (e.g., 0.5), it is determined to be a global exploration strategy (or breadth-first exploration strategy), in which candidate nodes that can connect distant communities in the graph and have high betweenness centrality are prioritized. If α is less than or equal to the preset threshold, it is determined to be a local mining strategy (or depth-first mining strategy), in which candidate nodes that can enhance the density of local communities and have high clustering coefficients are prioritized.

[0067] Under the global exploration strategy, structural exploration preference scoring It mainly depends on the betweenness centrality of candidate node i :

[0068] (2)

[0069] Betweenness centrality reflects the frequency with which a node appears in the shortest path of other node pairs. Nodes with high betweenness centrality often act as bridges, helping to expand the coverage of the graph.

[0070] Structural exploration preference scoring under a local mining strategy It mainly depends on the clustering coefficient of candidate node i :

[0071] (3)

[0072] Clustering coefficients reflect the tightness of connections between a node's neighbors. Nodes with high clustering coefficients help form close knowledge communities and deepen the knowledge density in a specific domain.

[0073] After determining the domain balance score and the structure exploration preference score, a comprehensive score for each candidate node can be determined based on these two scores. In one embodiment, the comprehensive score for each candidate node can be determined based on the weighted sum of the domain balance score and the structure exploration preference score. In different specific embodiments, the weighting weights of the domain balance score and the structure exploration preference score may be different, and this specification does not impose any restrictions on this.

[0074] In one embodiment, a connection potential score for each candidate node in the candidate node set for the knowledge graph to be expanded can also be determined.

[0075] Connectivity potential scoring can be used to evaluate the semantic fit and structural contribution of candidate nodes to the existing graph. This score is obtained by calculating the semantic similarity between the candidate node and its neighboring seed nodes in the graph. Using a pre-trained word vector model or graph neural network encoder, candidate nodes and seed nodes are mapped to vector representations, and cosine similarity is calculated as the semantic similarity score. Furthermore, it can predict the contribution of adding a candidate node to the graph's connectivity, such as whether it reduces the average path length or increases the size of the largest connected subgraph. Connectivity potential scoring It can be semantic similarity and connectivity contribution Weighted sum:

[0076] (4)

[0077] in, and This is a preset balance coefficient used to adjust the weight between semantic consistency and structural importance.

[0078] In one embodiment, the overall score It can be a linear combination of the scores from the three dimensions mentioned above, aiming to comprehensively reflect the optimality of candidate nodes:

[0079] (5)

[0080] in, This represents the overall score of the i-th candidate node; , , These represent the node's domain balance score, structural exploration preference score, and connectivity potential score, respectively. , , The weighting coefficients for each dimension's score typically satisfy the following conditions: This can be adjusted according to specific application scenarios. For example, in the early stages of development, more emphasis may be placed on domain balance, while in later stages, structural stability may be prioritized. This overall score... This reflects the value of candidate nodes in the current stage of knowledge graph construction. The higher the score, the greater the probability of being selected, thereby achieving high-quality and controllable expansion of the knowledge graph.

[0081] Subsequently, in step S140, the target expansion node can be selected from the candidate node set based on the preference degree and the exploration control parameters in the strategy parameters.

[0082] Traditional greedy algorithms directly select the node with the highest overall score, which can easily lead to a rigid graph structure and a lack of diversity. This embodiment uses a random factor τ in the strategy parameters as a temperature coefficient (exploration control parameter) to soften the overall score, converting it into a probability distribution. First, the overall score of each node in the candidate node set is normalized using the Softmax function to obtain an initial probability distribution. The final sampling probability after introducing the random factor τ is then calculated. for:

[0083] (6)

[0084] Where τ > 0. When τ approaches 0, the probability distribution tends towards a single-point distribution, meaning the node with the highest score is almost always selected, exhibiting deterministic selection; when τ is large, the probability distribution tends towards uniformity, exhibiting random selection. By adjusting τ, a balance can be achieved between "exploitation" (selecting high-scoring nodes) and "exploration" (trying low-scoring but potentially valuable nodes). Furthermore, based on the calculated probability distribution Pi, one or more nodes can be sampled from the candidate node set as target expansion nodes. In different embodiments, the specific method of sampling nodes from the candidate node set can vary. In one embodiment, for example, a multinomial distribution sampling method can be used to extract one or more nodes from the candidate node set as target expansion nodes.

[0085] After determining the target expansion node, in step S150, the target expansion node and its association with the knowledge graph to be expanded can be added to the knowledge graph to be expanded.

[0086] This step performs the actual data writing operation. The system instantiates the selected target expansion node and stores its attribute information (such as name, type, description, etc.) in the graph database. Simultaneously, it establishes edge relationships between the target node and the previously determined seed nodes; the edge type is determined by the entity link or relationship prediction module. At this point, the knowledge graph to be expanded has completed one iteration update. The number of nodes and edges increases accordingly, and the graph's topology and domain distribution change, providing a new foundational state for the next iteration.

[0087] In actual production scenarios, steps S110-S150 can be iterated multiple times. In one embodiment, the termination condition for multiple iterations may be one or more of the following: a preset upper limit on the total number of nodes, reaching the expected domain distribution ratio, the graph diameter reaching a specified value, or the quality score of newly added nodes in consecutive rounds of expansion falling below a certain threshold. Through this closed-loop iterative mechanism, the graph can gradually grow to a predetermined size and quality standard.

[0088] Figure 2 This diagram illustrates a specific flowchart for identifying a set of candidate nodes in one embodiment of this application. This flowchart corresponds to a detailed implementation of step S110, "identifying the set of candidate nodes determined based on the knowledge graph to be expanded." Figure 2 As shown, the process mainly includes steps S210 to S230.

[0089] In step S210, the boundary nodes in the knowledge graph to be expanded are selected as seed nodes.

[0090] Boundary nodes are nodes located at the edge of the current knowledge graph that have the potential to expand outwards. In one embodiment, the selection of boundary nodes is not random but can be based on the nodes' topological characteristics and semantic activity. Specifically, in one example, the out-degree and in-degree of each node in the knowledge graph to be expanded can be calculated. If the out-degree of a node is less than a preset threshold (e.g., 1 or 2), and the node has unresolved semantic slots or potential links to external knowledge bases, it is marked as a candidate seed node. Furthermore, node centrality indicators can be used to prioritize nodes located at the edges of connected components of the graph, with high betweenness centrality but low local clustering coefficients. These nodes typically act as bridges between different knowledge subgraphs; using them as seeds for expansion can more effectively connect isolated knowledge fragments and improve the overall connectivity of the graph.

[0091] In practical engineering implementations, a hierarchical sampling strategy can be adopted to balance computational efficiency and expansion quality. First, all nodes in the graph are grouped according to their domain labels. Then, within each domain group, the nodes are sorted according to the aforementioned topological features, and the top N nodes are selected as representative seed nodes for that domain. This hierarchical selection method ensures a uniform start-up of the subsequent expansion process across all domains, avoiding subsequent domain coverage deviations caused by uneven initial seed distribution.

[0092] In step S220, external entities that have a potential association with the seed node are retrieved.

[0093] The core of this step lies in discovering new knowledge entities using external data sources. In one embodiment, external data sources may include structured knowledge bases (such as Wikidata, DBpedia), semi-structured web page data, or unstructured text corpora. In one embodiment, the retrieval process may be divided into two stages: a candidate generation stage and a coarse ranking stage.

[0094] In one example, during the candidate generation stage, the name, alias, or key attributes of the seed node can be used as query terms to call a pre-defined search engine interface or knowledge base API to obtain a batch of relevant candidate entities. In a specific example, to improve recall, the query terms can be expanded, for example, by adding synonyms or hypernyms of the seed node's domain. For instance, if the seed node is "deep learning," the query terms can be expanded to "neural network," "machine learning algorithm," etc.

[0095] In the coarse-ranking stage, preliminary filtering of the recalled candidate entities can be performed. In one example, preliminary filtering of the recalled candidate entities can be performed using preset filtering rules. Filtering rules may include, for example, entity type matching (ensuring that candidate entities are semantically compatible with seed nodes, such as a "person" node should not link to a "location" entity), deduplication (removing entities that already exist in the knowledge graph to be expanded), and confidence threshold filtering (removing low-quality candidates with retrieval scores below a preset threshold). After coarse-ranking, the remaining entities can constitute a preliminary external entity set.

[0096] Figure 3 A schematic diagram illustrating the establishment of potential relationships is shown in one embodiment of this application. For example... Figure 3 As shown, seed node 301 has potential semantic connections with multiple external entities 302, 303, and 304. These connections are referred to as potential associations before they are formally added to the graph.

[0097] In step S230, external entities are selected as candidate nodes, and potential associations between candidate nodes and seed nodes are established.

[0098] Establishing potential relationships involves more than just recording the connection between two entities; it can also include a quantitative description of the nature and strength of the connection. The system uses relationship prediction models or rule-matching methods to infer the possible relationship types between the seed node and each external entity. For example, if the seed node is "Apple Inc." and the external entity is "iPhone," the potential relationship type could be "manufacturing products" or "owning a brand."

[0099] For each potential association, an association confidence score can be calculated. This score can be derived based on a combination of factors: entity co-occurrence frequency (the number of times the two entities appear together in a large corpus), semantic similarity (the distance between their vector representations), and the matching degree of the relationship template. The association confidence score will serve as one of the important input parameters for subsequent calculation of the connectivity potential score.

[0100] It's worth noting that potential relationships can be represented in the data structure as temporary edge objects, containing information such as source node ID, target node ID, relationship type, confidence score, and timestamp. In one example, these temporary edge objects may not be immediately persisted to the main graph database, but rather stored in an in-memory candidate pool for use by subsequent scoring and sampling modules. This design allows the system to evaluate and filter a large number of candidate expansion paths without polluting the main graph data.

[0101] In a different embodiment, if the external data source is unstructured text, steps S220 and S230 can be implemented using Named Entity Recognition (NER) and Relation Extraction (RE) models. The system first scans the text fragment containing the context of the seed node, identifies new entities within it, and extracts their syntactic dependencies or semantic roles with the seed node, thereby constructing potential associations. This approach is particularly suitable for mining newly emerging knowledge entities from news articles, papers, or technical documents, compensating for the lag in updates to structured knowledge bases.

[0102] Through the above steps, the system can accurately filter out the candidate nodes most relevant to the current graph state from massive amounts of external information and establish quantified potential connectivity indicators for each candidate node. This not only expands the search space of the graph but also provides a rich data foundation for subsequent refined scoring. Compared to traditional blind expansion, this identification mechanism based on seed nodes and potential associations can significantly reduce the probability of introducing invalid nodes and improve the signal-to-noise ratio of graph expansion.

[0103] Figure 4 This diagram illustrates the logical structure for calculating the comprehensive score of candidate nodes in one embodiment of this application. This process corresponds to step S120, which aims to select nodes from the candidate node set that meet the current graph construction goals through multi-dimensional quantitative evaluation. Figure 4 As shown, the scoring module 400 receives domain distribution data from the distribution statistics module 401, parameter settings from the strategy configuration module 402, and node feature information from the candidate pool 403. It calculates the domain balance score, structure exploration preference score, and connection potential score respectively. Finally, the comprehensive scoring calculation unit 404 outputs the comprehensive score of each candidate node.

[0104] When calculating the domain balance score, it is necessary to dynamically perceive the global state of the knowledge graph to be expanded. The specific method for calculating the domain balance score can vary in different embodiments. In one embodiment, the current node proportion of each domain in the graph can be statistically analyzed and compared with the preset target proportion in the strategy configuration file to obtain a domain deviation value. To reflect the importance of different domains and smooth the score distribution, the system introduces domain distribution weights and nonlinear transformation functions (such as the hyperbolic tangent function) to process the deviation value. When the current proportion of a domain is lower than the target proportion, candidate nodes in that domain receive a positive incentive score; conversely, they receive a negative inhibition score. This mechanism ensures that the graph can gradually converge to the preset domain distribution pattern at the macro level, avoiding excessive expansion of some domains while other domains are lacking.

[0105] In different embodiments, the specific methods for determining the structure exploration preference score can also differ. In one embodiment, the calculation of the structure exploration preference score can rely on real-time analysis of the graph topology and balance the "global exploration" and "local mining" strategies through depth and breadth exploration coefficients. In global exploration mode, the algorithm tends to select bridge nodes located between loose communities. In this case, the score is mainly based on the betweenness centrality of the node, which measures the frequency of the node's appearance in all shortest paths. The higher the betweenness centrality, the greater its value as a hub. In local mining mode, the algorithm tends to select nodes that can form close triangular relationships. In this case, the score is mainly based on the clustering coefficient of the node, which measures the closeness between the node's neighbors. Figure 5 This diagram illustrates a comparison of node selection tendencies under global exploration and local mining strategies. (For example...) Figure 5 As shown, in global exploration mode, the algorithm tends to select bridge nodes 501 located between two loose communities to enhance the overall connectivity of the graph; while in local mining mode, the algorithm tends to select nodes 502 that can form close triangular relationships with existing neighbors to deepen the density of local knowledge. The scoring strategy can be dynamically adjusted according to preset depth and breadth exploration coefficients α: if α is greater than a set threshold, the global exploration path is activated, and a score is calculated based on the node's betweenness centrality to measure its value as a hub; if α is less than or equal to the set threshold, the local mining path is activated, and a score is calculated based on the node's clustering coefficient to measure the closeness between its neighbors. In one embodiment, to avoid structural oscillations caused by abrupt strategy changes, the depth and breadth exploration coefficients can also be used as soft weights to weight and fuse the normalized betweenness centrality and clustering coefficient, allowing the graph expansion process to smoothly transition between breadth and depth.

[0106] The connectivity potential score aims to evaluate the semantic compatibility and structural contribution of candidate nodes to the existing graph. In different embodiments, the specific methods for determining the connectivity potential score can also differ. For example, in one embodiment, at the semantic level, a semantic similarity score can be obtained by calculating the vector cosine similarity between the candidate node and its associated seed nodes, reflecting the proximity of the candidate node to existing knowledge in the semantic space. At the structural level, the system evaluates the degree to which the addition of a candidate node improves the connectivity of the graph, such as detecting whether the node connects previously disconnected subgraphs or shortens the path of long-distance node pairs. If a candidate node helps merge isolated subgraphs, it is assigned a higher connectivity contribution score. Finally, the connectivity potential score is obtained by linearly weighting semantic similarity and connectivity contribution, and the weight parameters can be adjusted according to the emphasis on semantic consistency or structural integrity in the business scenario.

[0107] Finally, the comprehensive scoring calculation unit 404 performs a weighted sum of the sub-scores of the above three dimensions to obtain the comprehensive score of each candidate node.

[0108] Figure 6 This diagram illustrates a flowchart of a method for determining target expansion nodes based on weighted probability sampling in one embodiment of this application. This process corresponds to step S130 and aims to maintain the diversity of the graph expansion while preserving target orientation by introducing a randomness mechanism. Figure 6 As shown, the process mainly includes three sub-steps: score normalization, probability distribution construction, and random sampling.

[0109] In this embodiment, after obtaining the comprehensive score for each candidate node, score normalization is first performed. Since the units and ranges of scores may differ across dimensions, directly sampling using the original scores could lead to the larger dimensions dominating the selection result. Therefore, the system uses the Softmax function to convert the comprehensive score into an initial probability distribution. During this process, the comprehensive score of all candidate nodes is mapped to a range of 0 to 1, and the sum of the probabilities of all nodes is 1. This transformation ensures that nodes with higher scores have a higher probability of being selected, but does not completely exclude nodes with lower scores, thus preserving the possibility of exploring low-scoring but potentially valuable nodes.

[0110] Subsequently, the system introduces a random factor as a temperature coefficient to smooth the initial probability distribution, resulting in the final sampled probability distribution. The random factor plays a role in balancing "greedyness" and "randomness" in the algorithm. When the random factor is small, the probability distribution tends to be sharper, with high-scoring nodes having a significantly higher probability of being selected than other nodes. In this case, the algorithm exhibits strong determinism and tends to select the current optimal solution. When the random factor is large, the probability distribution tends to be more uniform, with the difference in the probability of each node being selected decreasing. In this case, the algorithm exhibits strong randomness, which helps to escape local optima and discover long-tail knowledge or related entities in niche areas. In practical applications, the value of the random factor can be dynamically adjusted according to the stage of graph construction. For example, in the early stages of construction, a smaller random factor can be set to quickly cover the main areas; while in the later stages, the random factor can be gradually increased to enrich the details and diversity of the graph.

[0111] Furthermore, the target expansion node can be determined from the candidate node set based on the final sampling probability distribution. In one example, the specific sampling method can be Cumulative Probability Distribution Selection or Multinomial Distribution Sampling. Cumulative Probability Distribution Selection maps the probability distribution to a cumulative distribution function, generating a random number between 0 and 1; the node is selected based on the interval in which the number falls. This method is simple to implement, computationally efficient, and suitable for scenarios with a moderate number of candidate nodes. If the number of candidate nodes is huge, the Alias ​​Method can be used to improve sampling efficiency. This method reduces the sampling time complexity to constant level through preprocessing, making it suitable for the real-time expansion needs of large-scale graphs. In one iteration, the system can sample a single node or sample multiple nodes according to a preset batch size to accelerate the growth rate of the graph.

[0112] Next, the target expansion node and its associations with the knowledge graph to be expanded are added to the knowledge graph. This step marks the formal transformation of the candidate node into part of the graph. The system first instantiates the target expansion node and writes its attribute information (such as entity name, type, description text, etc.) into the node storage area of ​​the graph database. Then, based on the previously established potential associations, the system creates edge objects between the target node and the seed node. The edge attributes include relationship type, confidence score, and creation timestamp.

[0113] It is worth noting that after adding a node to the graph, the system needs to synchronously update the global state information of the graph. This includes updating the node counts in each domain, recalculating the topological indices of the affected regions (such as betweenness centrality, clustering coefficients, etc.), and refreshing the identifiers of connected components. As mentioned earlier, steps S110~S30 can be iterated multiple times in actual production scenarios. Therefore, these update operations provide the latest basic data for the score calculation in the next iteration. To ensure data consistency, the above update operations are usually executed under a transaction mechanism to ensure the atomicity of node insertion and edge creation.

[0114] In one different implementation, considering the performance bottleneck of large-scale graph updates, an asynchronous update strategy can be adopted. This involves first writing the target nodes and their relationships to a temporary buffer or message queue, and then having a background process batch-merge them into the main graph database. While this introduces a slight latency, it significantly reduces the system load from frequent I / O operations and improves overall scalability. Furthermore, for high-concurrency scenarios, a sharded storage architecture can be used, distributing nodes from different domains across different storage shards to achieve parallel scaling and load balancing.

[0115] Through the steps described above, the system completes a full knowledge graph expansion iteration. This process not only achieves incremental knowledge accumulation but also ensures the controllability of the graph's domain distribution and structural features through refined scoring and sampling mechanisms. Compared to traditional random or rule-based expansion, the method provided in this embodiment can more flexibly adapt to diverse construction goals and generate high-quality knowledge graph data.

[0116] Figure 7 A schematic diagram of a knowledge graph node expansion device according to an embodiment of this application is shown. This device 700 can be integrated into a server, cloud computing platform, or high-performance computing terminal to execute the steps in the aforementioned method embodiments. Figure 7 As shown, the device mainly includes an acquisition module 701, a determination module 702, an evaluation module 703, a screening module 704, and an execution module 705, wherein...

[0117] The acquisition module 701 is used to acquire knowledge graph node expansion strategies, the strategies including strategy parameters for controlling the expansion direction of the knowledge graph;

[0118] The determining module 702 is used to determine the set of candidate nodes for the knowledge graph to be expanded, and to obtain the current state information of the knowledge graph to be expanded;

[0119] Evaluation module 703 is used to perform multi-dimensional evaluation on each candidate node in the candidate node set based on the current state information and the strategy parameters, so as to obtain the preference of each candidate node.

[0120] The filtering module 704 is used to filter out target expansion nodes from the candidate node set based on the preference degree and the exploration control parameters in the strategy parameters;

[0121] The execution module 705 is used to add the target expansion node and its association with the knowledge graph to be expanded to the knowledge graph.

[0122] It should be noted that the modules in the above device embodiments are not necessarily physically independent hardware components, but can also be software functional modules running on a general-purpose processor. For example, the functions of the acquisition module, scoring module, and update module can be implemented by the same multi-core processor by running different threads or processes. In addition, the device may also include a feedback control module (not shown in the figure) for executing the deviation calculation and dynamic weight adjustment logic described in the foregoing embodiments, thereby achieving closed-loop optimization.

[0123] In another aspect, embodiments of this specification provide a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform any of the methods described above.

[0124] In another aspect, embodiments of this specification provide a computing device, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement any of the methods described above.

[0125] It should be understood that the descriptions such as "first" and "second" in this article are merely for the sake of simplicity in description and to distinguish similar concepts, and do not have any other limiting function.

[0126] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must also be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0127] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0128] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. A typical implementation device is a server system. Of course, this application does not exclude the possibility that, with the future development of computer technology, the computer implementing the functions of the above embodiments can be, for example, a personal computer, a laptop computer, an in-vehicle human-machine interaction device, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0129] While one or more embodiments of this specification provide the operational steps of the methods described in the embodiments or flowcharts, more or fewer operational steps may be included based on conventional or non-inventive means. The order of steps listed in the embodiments is merely one possible order of execution among many steps and does not represent the only possible order. In actual device or end product execution, the methods shown in the embodiments or drawings may be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment, or even a distributed data processing environment). The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitations, the presence of other identical or equivalent elements in the process, method, product, or apparatus that includes said elements is not excluded. For example, the use of terms such as "first," "second," etc., is to denote names and does not indicate any particular order.

[0130] For ease of description, the above devices are described in terms of function, divided into various modules. Of course, when implementing one or more of these specifications, the functions of each module can be implemented in one or more software and / or hardware components, or a module that performs the same function can be implemented by a combination of multiple sub-modules or sub-units. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0131] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0132] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0133] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0134] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0135] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0136] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage, graphene storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0137] Those skilled in the art will understand that one or more embodiments of this specification can be provided as a method, system, or computer program product. Therefore, one or more embodiments of this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, one or more embodiments of this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0138] One or more embodiments of this specification can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a particular task or implement a particular abstract data type. One or more embodiments of this specification can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0139] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, system embodiments are basically similar to method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. In the description of this specification, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this specification. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0140] The above description is merely an embodiment of one or more embodiments of this specification and is not intended to limit the scope of these embodiments. Various modifications and variations can be made to these embodiments by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims.

Claims

1. A method for expanding knowledge graph nodes, characterized in that, include: Obtain a knowledge graph node expansion strategy, the strategy including strategy parameters for controlling the expansion direction of the knowledge graph; Determine the set of candidate nodes for the knowledge graph to be expanded, and obtain the current state information of the knowledge graph to be expanded; Based on the current state information and the strategy parameters, each candidate node in the candidate node set is evaluated in multiple dimensions to obtain the preference of each candidate node; Based on the preferredness and the exploration control parameters in the strategy parameters, target expansion nodes are selected from the candidate node set; The target expansion node and its association with the knowledge graph to be expanded are added to the knowledge graph to be expanded.

2. The method as described in claim 1, characterized in that, The strategy parameters include domain distribution control parameters, as well as depth and breadth exploration coefficients; Each candidate node in the candidate node set is evaluated in multiple dimensions to obtain the optimality of each candidate node, including: Based on the domain distribution information in the current state information and the domain distribution control parameters, calculate the domain balance score for each candidate node; Based on the depth and breadth exploration coefficients, a structural exploration preference score is determined for each candidate node; Based on the domain balance score and the structure exploration preference score, a comprehensive score is determined for each candidate node, and the comprehensive score is used as the preference degree.

3. The method as described in claim 2, characterized in that, The calculation of the domain balance score for each candidate node based on the domain distribution information in the current state information and the domain distribution control parameters includes: Statistical analysis of the current node percentage in each domain of the knowledge graph to be expanded; Calculate the difference between the current node percentage and the preset target percentage; Based on the differences and the domain distribution control parameters, a domain balance score is determined for each candidate node.

4. The method as described in claim 2, characterized in that, The determination of the structure exploration preference score for each candidate node in the candidate node set based on the depth and breadth exploration coefficients includes: Obtain the graph structure features of the candidate nodes, including betweenness centrality and clustering coefficient; An exploration strategy is determined based on the depth and breadth exploration coefficients, and the exploration strategy includes a global exploration strategy or a local mining strategy. If the exploration strategy is a global exploration strategy, the structure exploration preference score is calculated based on the betweenness centrality of the candidate nodes; The exploration strategy is a local mining strategy, and the structural exploration preference score is calculated based on the clustering coefficient of the candidate nodes.

5. The method of claim 2, wherein, The multi-dimensional evaluation also includes determining a connection potential score for each candidate node in the candidate node set for the knowledge graph to be expanded. Based on the domain balance score and the structure exploration preference score, a comprehensive score is determined for each candidate node, including: A comprehensive score for each candidate node is determined by a weighted sum of the domain balance score, the structure exploration preference score, and the connectivity potential score.

6. The method as described in claim 5, characterized in that, Determining the connectivity potential score of each candidate node in the candidate node set for the knowledge graph to be expanded includes: Calculate the semantic similarity between the candidate node and its neighboring nodes in the knowledge graph to be expanded; And / or, predict the contribution of the candidate node to the graph connectivity after its addition; The connectivity potential score is calculated based on the semantic similarity and / or the contribution.

7. The method as described in claim 1, characterized in that, The step of determining the target expansion node from the candidate node set based on the preference degree and the exploration control parameters in the strategy parameters includes: The preference degree is converted into an initial probability distribution using a normalization function; By introducing a random factor from the strategy parameters as a temperature coefficient, the initial probability distribution is smoothed to obtain the final sampling probability distribution; Based on the final sampling probability distribution, the target expansion node is determined from the candidate node set.

8. The method as described in claim 1, characterized in that, The determination of the candidate node set for the knowledge graph to be expanded includes: Select the boundary nodes in the knowledge graph to be expanded as seed nodes; Retrieve external entities that have a potential association with the seed node; The external entity is used as a candidate node, and a potential association between the candidate node and the seed node is established.

9. The method of claim 1, wherein, The knowledge graph to be expanded includes nodes belonging to various fields such as medical, financial, and technological fields.

10. A knowledge graph node expansion device, characterized in that, include: The acquisition module is used to acquire knowledge graph node expansion strategies, wherein the strategies include strategy parameters for controlling the expansion direction of the knowledge graph; The determination module is used to determine the set of candidate nodes for the knowledge graph to be expanded, and to obtain the current state information of the knowledge graph to be expanded. The evaluation module is used to perform multi-dimensional evaluation on each candidate node in the candidate node set based on the current state information and the strategy parameters, so as to obtain the preference of each candidate node. The filtering module is used to filter out target expansion nodes from the candidate node set based on the preference degree and the exploration control parameters in the strategy parameters; An execution module is used to add the target expansion node and its association with the knowledge graph to be expanded to the knowledge graph.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the knowledge graph node expansion method as described in any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the knowledge graph node expansion method as described in any one of claims 1 to 9.

Images