Construction method of geographic knowledge graph alignment model
By constructing a geographic knowledge graph alignment model based on a spatially aware attention network, and utilizing backlinks, spatially aware attention mechanisms, and adversarial network generation strategies, the problem of low alignment accuracy of geographic entities in geographic knowledge graphs is solved, achieving higher alignment accuracy and robustness.
Patent Information
- Application Number
- CN202510770402.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-10
- Publication Date
- 2025-11-07
AI Technical Summary
Existing geographic knowledge graph alignment methods have low accuracy in aligning geographic entities when targeting specific professional fields, and the simple negative sample generation strategy leads to large differences, making it difficult to simulate complex alignment tasks in real-world scenarios.
We construct a geographic knowledge graph alignment model based on a spatially aware attention network. High-quality negative samples are generated through a backlink generation strategy, a spatially aware attention mechanism, and an adversarial network. The model is then trained using vector loss and geographic loss to enhance the accuracy of geographic entity alignment.
It improves the accuracy of geographic entity alignment in geographic knowledge graphs, enhances the robustness and alignment effect of the model, and performs particularly well in highly specialized geographic knowledge graph application scenarios.
Smart Images
Figure CN120911577A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of natural language processing, and more particularly relates to a construction method of a geographical knowledge graph alignment model. BACKGROUND
[0002] Geographical knowledge graph alignment is an important task in the field of natural language processing, and its purpose is to unify data from different sources into the same spatio-temporal framework, solve the inconsistency problems of data format, semantics and spatial information, thereby eliminating redundancy, filling in missing information, and improving the completeness and accuracy of knowledge. Current geographical knowledge graph alignment methods are mainly divided into two categories: rule-based methods, machine learning-based methods and deep learning-based methods.
[0003] Among them, the rule-based method relies on manually designed matching rules and performs alignment through string similarity, logical reasoning and other means; the machine learning-based method uses statistical learning and data-driven methods to automatically learn matching patterns; the deep learning-based method models the structure and semantic information of the knowledge graph through neural networks.
[0004] At the present stage, the research on geographical knowledge graph alignment only provides partial solutions. Most of the existing geographical knowledge graph alignment methods are not suitable for the alignment of geographical knowledge graphs in specific professional fields, as they are mostly designed for general knowledge graphs. First, the limitations of geographical knowledge graph datasets make existing alignment models' experiments often focus on general triple datasets, ignoring professional geographical knowledge graph datasets. Second, geographical entities usually have significant spatial location characteristics, but most alignment methods fail to fully consider these characteristics. Finally, the existing negative sample generation strategies are too simple, and most models use random negative sample generation strategies, resulting in a large difference between the generated negative samples and the real entities, making it difficult to simulate complex alignment tasks in actual scenarios. The above problems result in low accuracy of geographical entity alignment when completing the geographical knowledge graph alignment task.
[0005] Therefore, how to improve the accuracy of geographical entity alignment in geographical knowledge graphs is a problem that needs to be solved. SUMMARY
[0006] In view of the defects of the prior art, the purpose of the present application is to provide a construction method of a geographical knowledge graph alignment model, which can improve the accuracy of geographical entity alignment in geographical knowledge graphs.
[0007] To achieve the above purpose, in a first aspect, the present application provides a construction method of a geographical knowledge graph alignment model, comprising the following steps: S10, combine geographical entities with their spatial location information to construct a geographical knowledge graph structure containing multi-dimensional spatial information; S20, using the strategy generated by the reverse link, creating a corresponding reverse link for each relationship link between geographical entities to represent the bidirectional relationship between entities; S30, on the basis of step S20, using a spatial perception attention mechanism to assign corresponding weights to different neighbor entities of each geographical entity, and combining the orthogonal transformation matrix of the relationship and the spatial distance to calculate the output features of each geographical entity; S40, on the basis of step S20, generating negative samples through an adversarial network, and training the pre-constructed deep learning network using the negative samples, while in the training process, on the basis of step S30, using the combination of vector loss and geographical loss as the total loss to calculate the error of the output features of each geographical entity, and then updating the corresponding weight parameters according to the error result until the network converges to the expected target, completing the training.
[0008] Compared with the traditional geographical knowledge graph alignment method, the present application focuses on the knowledge graph alignment research in the geographical field, and proposes a construction method of a geographical knowledge graph alignment model based on a spatial perception attention network. In addition, the present application generates high-quality negative samples through adversarial training, and uses the generator and the discriminator to alternate optimization, solving the shortcomings of the traditional negative sample generation method. Furthermore, in the loss calculation, a unique geographical loss is introduced to spatially constrain the geographical entity alignment, and a training and test data set required by deep learning is constructed. The experimental results show that the present application effectively utilizes the spatial information in the geographical data, thereby improving the performance of the geographical knowledge graph alignment task (the accuracy of the geographical entity alignment).
[0009] As a further preferred, in step S10, the basic construction unit of the geographical knowledge graph is a geographical space quadruple, and each geographical space quadruple contains four main elements: a geographical entity, a relationship type, an entity attribute, and a spatial position.
[0010] As a further preferred, in step S30, the step of assigning corresponding weights to different neighbor entities of each geographical entity using a spatial perception attention mechanism is specifically: S31, defining the spatial-specific weighted importance and the relationship-specific weighted importance from the neighbor entity to the geographical entity as follows:
[0011]
[0012] wherein || represents the concatenation operation; is a shared spatial and relationship attention weight vector; is and input features of entities, in the first layer network, the input features of entities are their original embeddings; and are the spatial distance and the relation embedding of the incoming edge, respectively, is an orthogonal matrix; S32, the normalized attention weight from entity to is calculated using the softmore function and represent the connectivity of the space and the relation, respectively:
[0013]
[0014] where, is the neighbor set of entity ; and represent the relation set and the coordinate point set from to , respectively; S33, the output feature of entity is calculated by an aggregation operation, which is: first, weighted sum of neighbor entity input features, then combined with the orthogonal transformation of space and relation, and finally passed through a nonlinear activation function :
[0015] where, represents the set of all edges from to .
[0016] As a further preferred, in step S40, the generation of negative samples of the adversarial network is an adversarial training through a generator and a discriminator, the generator uses a probability-based log loss model, and the discriminator uses a distance-based margin loss model.
[0017] As a further preferred, the generator models the probability distribution of negative samples through a softmax function, and adopts the negative log likelihood of the softmax function as its loss function.
[0018] As a further preferred, the gradient of the generator optimization target is estimated using the policy gradient theorem, and the expectation of the generator optimization target is estimated using sampling approximation.
[0019] As a further preferred, the discriminator uses a margin loss function to calculate the distance between vectors, a smaller distance represents a higher similarity, which is calculated as follows:
[0020] wherein, represents the margin loss function; is a margin hyper-parameter, used to control the minimum separation between positive and negative samples; is a hinge function, to ensure the loss value is not negative; is a negative quadruple; The goal of the discriminator is to minimize the following margin loss function:
[0021] wherein, represents the loss function of the discriminator; negative samples from the probability distribution of the generator: That is, the negative samples are not randomly generated, but are sampled by the generator G, making the training more effective.
[0022] As a further preferred, in step S40, the vector loss is used to ensure that the correctly matched entity pairs are closer, while the incorrectly matched entity pairs are farther apart, which is calculated as follows:
[0023] wherein, is a positive sample entity pair; , is an entity vector in a negative sample entity pair; is a hyper-parameter, representing the boundary threshold of the loss; the vector loss is L 1 distance; The geographic loss is calculated by the Haversine formula to calculate the shortest distance between two points on the sphere, which is calculated as follows:
[0024] wherein, and represent the longitude and latitude of entity A and entity B respectively; R is the radius of the earth; For the geographic loss between the positive sample and the negative sample is calculated as follows:
[0025] The vector loss and the geographic loss are combined to form the final loss function optimization goal:
[0026] wherein, is a hyper-parameter used to adjust the weight of the geo-loss. In a second aspect, the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method for constructing a geo-knowledge graph alignment model as described above when executing the program.
[0027] In a third aspect, the present application provides a computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions, when executed by a processor, implement the steps of the method for constructing a geo-knowledge graph alignment model according to any one of the above aspects.
[0028] It can be understood that the beneficial effects of the above-mentioned second and third aspects can be referred to the related description in the first aspect above, which will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS
[0029] Figure 1 is a flowchart of the method for constructing a geo-knowledge graph alignment model provided by the embodiments of the present application; Figure 2 is a core module diagram of the geo-knowledge graph alignment model provided by the embodiments of the present application; Figure 3 is a diagram showing the influence of coordinate bias on the performance of a spatial perception model provided by the embodiments of the present application; Note: GeoAlign dataset: coordinate values with coordinate offset, GeoAlign-NP dataset: GeoAlign dataset without coordinate offset; Figure 4 is a diagram showing the influence of the number of aligned entity pairs on the model accuracy provided by the embodiments of the present application.
[0030] It should be noted that in Table 1, GL: geo-loss, AN: adversarial network, SA: spatial perception model; in Table 3, GeoAlign-5k: complete dataset, containing all entities and their spatial coordinate information; GeoAlign-5k-Geo: only contains spatial-related entities (such as geographic location, building, etc.), used to evaluate the performance of the model in processing entities with spatial features; GeoAlign-5k-NonGeo: only contains entities unrelated to spatial information (such as organizations, individuals, etc.), used to evaluate the performance of the model in the absence of spatial features; GeoAlign-5k-NoCoord: the spatial coordinate values of all entities are set to invalid values (-inf, -inf), used to analyze the alignment ability of the model in the absence of spatial information. DETAILED DESCRIPTION
[0031] To make the purposes, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application and are not all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present application.
[0032] The present application is directed to the geographical knowledge graph alignment task, and provides a construction method of a geographical knowledge graph alignment model (SGNN) based on a spatial perception graph neural network, aiming to solve the problem of low geographical entity alignment accuracy in traditional geographical knowledge graph construction, such as the problems of "same name different place" and "same place different name", to adapt to more professional geographical knowledge graph application scenarios.
[0033] As shown in Figure 1 The construction method of the geographical knowledge graph alignment model provided by the present application mainly includes steps S10-S40, which are described in detail as follows: Step S10, combine the geographical entity with its spatial position information to construct a geographical knowledge graph structure containing multi-dimensional spatial information.
[0034] In the present embodiment, the present embodiment first defines a geographical space quadruple as a basic construction unit in the geographical knowledge graph. Each geographical space quadruple contains four main elements: a geographical entity, a relationship type, an entity attribute, and a spatial position. For example, the geographical entity is the city of "Wuhan"; the relationship type is "provincial capital city"; the entity attribute includes characteristics such as large population, developed economy, and transportation hub; and the spatial position is north latitude 30°35', east longitude 114°17'. By combining the entity with its spatial position information, the present application constructs a geographical knowledge graph structure containing multi-dimensional spatial information, which can effectively represent the complex relationships and spatial characteristics of geographical entities. Step S20, using the strategy of reverse link generation, creates a corresponding reverse link for each relationship link between geographical entities to represent the bidirectional relationship (direction of relationship) between entities, enhancing the bidirectional expression capability of relationships in the knowledge graph.
[0035] Step S30, on the basis of step S20, uses a spatial perception attention mechanism to assign weights to different neighbor entities of each geographical entity, and combines a relationship and spatial distance orthogonal transformation matrix to calculate the output features of each geographical entity.
[0036] In the embodiment, a spatial perception attention network is designed, which adjusts attention weights by combining spatial position information, and gives higher matching weights to geographical positions, so as to more accurately capture the spatial similarity between geographical entities during modeling.
[0037] In step S40, based on step S20, a negative sample is generated by an adversarial network, and the pre-constructed deep learning network is trained using the negative sample. Meanwhile, based on step S30, the combination of vector loss and geographical loss is used as the overall loss for error calculation during the training process, and then the weight parameters are updated according to the error result until the network converges to the expected target, and the training is completed.
[0038] In the embodiment, step S40 specifically includes the following sub-steps: Step S41, adversarial network negative sample generation. The application introduces a method of generating negative samples by an adversarial network, uses a probability-based logarithmic loss model as a generator, and uses a distance-based marginal loss model as a discriminator, and generates higher-quality negative samples through an adversarial training strategy.
[0039] Step S42, loss calculation. The application proposes a spatial perception-based loss function for quantifying the error of geographical entity alignment. The loss function combines traditional distance measurement methods and spatial position correlation measurement, and through the introduction of spatial constraints, the geographical loss directly punishes the embedding representation with too large spatial distance, and finally applies the obtained distance function to the multi-view representation of the entity to predict entity alignment.
[0040] Step S43, experimental training process and analysis. The application selects GeoAlign and DBP100K datasets to perform a series of experiments to evaluate the performance of the SGNN model. The coordinate deviation experiment shows (see Figure 3 ), the coordinate deviation significantly reduces the alignment accuracy of the model, and the model performs better when the coordinates are consistent. The ablation experiment shows (see Table 1), the addition of geographical loss and adversarial network can significantly improve the performance of the model, and the adversarial network plays a key role in improving the accuracy.
[0041] Table 1 Ablation experiment results
[0042] The generalization experiment shows (see Table 2) that the SGNN model can also maintain high alignment accuracy on non-geographical knowledge graphs, and exhibits strong generalization ability.
[0043] Table 2 Test results of SGNN model on GeoAlign and DBP100K datasets
[0044] Quantitative experiments show (see Figure 4 ), sufficient alignment entities are essential for the number of seeds to model performance, and the number of seeds decreases significantly. Spatial sensitivity experiments show (see Table 3) that spatial information has a significant impact on the alignment of geographic entities, but even if the spatial information is missing, the model can still effectively align through semantic information.
[0045] Table 3 Spatial sensitivity experiment results
[0046] The working principle of the construction method of the geographic knowledge graph alignment model provided in this embodiment is as follows: The embodiment provides a method of a geographic knowledge graph alignment model based on a spatial perception graph neural network, which mainly includes four parts of reverse link generation, spatial perception attention mechanism, adversarial network negative sample generation and loss calculation, as shown in Figure 2 The specific content is as follows: (1) Reverse link refers to adding a corresponding reverse relationship for each relationship in the knowledge graph in addition to the original forward relationship, so as to be able to represent the bidirectional relationship between entities. That is, the original relationship set is expanded to . At the same time, for each geographic spatial quadruple (h, r, t, s) , its corresponding reverse quadruple (t, r, h, s) is constructed.
[0047] (2) The spatial perception attention mechanism assigns different weights to different neighbor nodes, and combines the orthogonal transformation matrix of the relationship and the spatial distance to enhance the information expression ability.
[0048] The spatial-specific weighted importance and the relationship-specific weighted importance from the neighbor entity to the geographic entity are defined as follows:
[0049]
[0050] Where, || represents the splicing operation; is the shared spatial and relationship attention weight vector; is the input feature of and In the first layer network, the input feature of the entity is its original embedding; and are the spatial distance and relationship embedding of the incoming edge, respectively, It is an orthogonal matrix.
[0051] Use the softmore function to calculate from the entity arrive Normalized attention weights and , respectively representing spatial and relational connectivity:
[0052]
[0053] in, It is a physical entity The set of neighbors; and They represent from arrive The set of relationships and the set of coordinate points.
[0054] Finally, the entity Output characteristics The calculation is performed through an aggregation operation, which weights and sums the input features of neighboring entities, combines spatial and relational orthogonal transformations, and finally passes through a nonlinear activation function. :
[0055] in, Indicates from arrive The set of all edges.
[0056] (3) Adversarial network negative sample generation is a method to generate high-quality negative samples through adversarial training of generator and discriminator, so as to enhance the model’s discrimination and generalization ability.
[0057] The generator models the probability distribution of negative samples using the softmax function:
[0058] in, The generator generates negative samples. The probability of; It is the generator's scoring function for negative samples. These are candidate negative samples after randomly replacing the head and tail entities.
[0059] To train the generator, the negative log-likelihood of the softmax function is used as the loss function:
[0060] Where T is the set of positive samples in the training set; denotes the set of negative samples; is the scoring function for positive samples; is the scoring function for negative samples.
[0061] The optimization objective of the generator is to maximize the expected score of negative samples, i.e., the generator should tend to generate negative samples that the discriminator considers more realistic:
[0062] where, denotes the optimization objective of the generator; is the sampling probability of negative samples by the generator; E represents expectation, which is used to calculate the average effect of negative samples sampled from the generator distribution is the scoring function of negative samples by the discriminator.
[0063] However, this optimization objective involves discrete sampling, which makes it impossible to directly calculate its gradient using gradient descent method. Therefore, the Policy Gradient Theorem is used to estimate the gradient:
[0064] In actual calculation, a sampling approximation is used to estimate the expectation:
[0065] where, denotes the gradient of the generator; is the number of samples; is the negative sample sampled from the generator.
[0066] The discriminator uses a translation-based model to calculate the distance between vectors using a margin loss function, where a smaller distance represents a higher similarity, and the calculation is as follows:
[0067] where, denotes the margin loss function; is the margin hyperparameter, which is used to control the minimum interval between positive and negative samples; is the hinge function, which ensures that the loss value is not negative; is the negative quadruple.
[0068] The objective of the discriminator is to minimize the following margin loss function to ensure that the score of the real quadruple is higher than that of the negative quadruple:
[0069] where, Loss function of discriminator; negative sample Probability distribution from generator: That is, the negative sample is not randomly generated, but is sampled by the generator G, making the training more effective.
[0070] (4) The loss calculation part adopts the combination of vector loss and geographic loss as the overall loss to enhance the geographic spatial perception ability of the model.
[0071] Vector loss (L1 distance) is used to ensure that the correct matching entity pair is close in distance, while the wrong matching entity pair is far in distance. The specific calculation is as follows:
[0072] Where, is the positive sample entity pair; , is the entity vector in the negative sample entity pair; is a hyperparameter, representing the boundary threshold of the loss.
[0073] Geographic loss calculates the shortest distance between two points on the sphere through the Haversine formula, which is calculated as follows:
[0074] Where, and represent the latitude and longitude of entity A and entity B respectively; R is the radius of the earth.
[0075] For the geographic loss between the positive sample and the negative sample , the calculation is as follows:
[0076] Combine the vector loss and geographic loss to form the final loss function optimization target:
[0077] Where, is a hyperparameter used to adjust the weight of the geographic loss.
[0078] Compared with the traditional geographical knowledge graph alignment method, the embodiment focuses on the knowledge graph alignment research in the geographical field, and proposes a construction method of a geographical knowledge graph alignment model based on a space perception attention network. In addition, the embodiment generates high-quality negative samples through adversarial training, and uses the generator and the discriminator to alternately optimize, solving the shortcomings of the traditional negative sample generation method. In addition, a unique geographical loss is introduced in the loss calculation to constrain the spatial entity alignment, and the training and test datasets required by deep learning are constructed. The experimental results show that the application effectively utilizes the spatial information in the geographical data, thereby improving the performance of the geographical knowledge graph alignment task (the accuracy of the geographical entity alignment).
[0079] The beneficial effects of the embodiment are: the embodiment proposes a construction method of a geographical knowledge graph alignment model (SGNN) based on a space perception attention network, which can model geographical relationships through a space perception attention mechanism, and combine an adversarial negative sample generation strategy to improve the robustness of entity alignment. In addition, a geographical loss term is introduced to constrain the spatial characteristics of the entity, effectively enhancing the alignment effect. In terms of datasets, the embodiment proposes a new geographical knowledge graph representation method, i.e. a (h, r, t, coords) four-tuple, where coords is used to store the spatial coordinates (latitude and longitude). Based on this representation method, the GeoAlign dataset is constructed, and multiple subsets of different scales are designed to simulate the coordinate deviation problem in real application scenarios. Experiments show that SGNN can achieve good alignment performance on datasets of different scales, and the addition of the adversarial network and the geographical loss can improve the effect of the model. In addition, the applicability experiment of SGNN on non-geographical knowledge graphs shows that the method can still effectively align entities without coordinate information, verifying its generalization ability.
[0080] Those skilled in the art will readily understand that the above description is only a preferred embodiment of the present application, and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application.
Claims
1. A method for constructing a geographical knowledge graph alignment model, characterized in that, Comprising the following steps: S10, combine the geographical entity with its spatial location information to construct a geographical knowledge graph structure containing multi-dimensional spatial information; S20, using the strategy generated by the reverse link, create a corresponding reverse link for each relationship link between geographical entities to represent the bidirectional relationship between entities; S30, on the basis of step S20, use the spatial perception attention mechanism to assign corresponding weights to different neighbor entities of each geographical entity, and combine the orthogonal transformation matrix of the relationship and the spatial distance to calculate the output features of each geographical entity; S40, on the basis of step S20, generate negative samples through the adversarial network, and train the pre-constructed deep learning network using negative samples, while in the training process, on the basis of step S30, use the combination of vector loss and geographical loss as the total loss to calculate the error of the output features of each geographical entity, then update the corresponding weight parameters according to the error result until the network converges to the expected target, complete the training.
2. The method for constructing a geographic knowledge graph alignment model as described in claim 1, characterized in that, In step S10, the basic construction unit of the geographical knowledge graph is a geographical space quadruple, and each geographical space quadruple contains four main elements: geographical entity, relationship type, entity attribute, and spatial location. 3.The method of claim 1, wherein, In step S30, the step of assigning corresponding weights to different neighbor entities of each geographical entity using the spatial perception attention mechanism is as follows: S31, defining spatially specific weighted importance of geographical entities from neighbor entities as follows: where || denotes concatenation operation; is the shared spatial and relational attention weight vector; is and input features of and, in the first layer network, the input features of entities are their original embeddings; and are the spatial distance of the incoming edge and the embedding of the relation, respectively, is an orthogonal matrix; S32, compute normalized attention weights from entities to using softmore function and representing connectivity of space and relation, respectively: wherein, is a neighbor set of the entity ; and represent a relation set and a coordinate point set from to respectively; S33, the output feature of the entity of the output feature is calculated by an aggregation operation, which is to first perform weighted summation on the input features of the neighbor entities, then combine the orthogonal transformations of the space and the relationship, and finally pass through a nonlinear activation function : wherein, represents a set of all edges from to .
4. The method for constructing a geographic knowledge graph alignment model as described in claim 1, characterized in that, In step S40, the generation of negative samples by the adversarial network is an adversarial training through a generator and a discriminator, the generator uses a probability-based logarithmic loss model, and the discriminator uses a distance-based marginal loss model. 5.The method of claim 4, wherein, The generator models the probability distribution of negative samples through a softmax function, and uses the negative logarithmic likelihood of the softmax function as its loss function.
6. The construction method of the geographical knowledge graph alignment model according to claim 4, characterized in that, The gradient of the generator optimization target is estimated using the policy gradient theorem, and the expectation of the generator optimization target is estimated using sampling approximation. 7.The method of claim 4, wherein, The discriminator uses a marginal loss function to calculate the distance between vectors, and a smaller distance represents a higher similarity, which is calculated as follows: wherein, represents a margin loss function; is a margin hyper-parameter, used to control the minimum separation between positive and negative samples; is a hinge function, ensuring that the loss value is not negative; is a negative quadruple; The goal of the discriminator is to minimize the following marginal loss function: where, Ldis the loss function of the discriminator; negative samples The probability distribution from the generator: That is, the negative samples are not randomly generated, but are sampled by the generator G, making the training more effective. 8.The method of claim 1, wherein, In step S40, the vector loss is used to ensure that correctly matched entity pairs are close in distance, while incorrectly matched entity pairs are far apart, which is calculated as follows: wherein, is a positive sample entity pair; , is an entity vector in a negative sample entity pair; is a hyperparameter representing a boundary threshold for the loss; the vector loss is L 1 distance; The geographical loss calculates the shortest distance between two points on a sphere through the Haversine formula, which is calculated as follows: wherein, and respectively represent the latitude and longitude of entity A and entity B; R is the earth radius; For positive samples and negative samples the geo-loss is calculated as follows: Combine the vector loss and the geographical loss to form the final loss function optimization target: wherein, is a hyperparameter used to adjust the weight of the geo-loss.
9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the steps of the construction method of the geographical knowledge graph alignment model according to any one of claims 1-8 when executing the program.
10. A computer readable storage medium having stored thereon computer instructions, wherein, The computer instructions implement the steps of the construction method of the geographical knowledge graph alignment model according to any one of claims 1-8 when executed by the processor.