Entity alignment method based on multi-relation fusion and potential relation mining graph convolutional network

Abstract

The invention discloses a multi-relation fusion and potential relation mining graph convolutional network for entity alignment, which is used for improving the calculation efficiency and reducing the memory consumption while ensuring the accuracy of entity alignment, and mainly comprises the following steps of: capturing potential important related entities by using a graph random walk method; fusing the plurality of relationships in a local and global manner to obtain a preliminary representation of the entity; a multi-head attention mechanism is adopted to generate a plurality of complete association diagrams, and the correlation between entities is represented; a dense connection layer is constructed to capture potential relationships to obtain a fused representation of entities. The method aims to provide a method for enhancing entity representation for a user on the premise of ensuring better use experience of the user, and a better entity alignment effect is achieved.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing, and in particular to an entity alignment method based on graph convolution. Background Art

[0002] Knowledge fusion aims to consolidate entities that represent the same concept, integrate knowledge from different sources into a cohesive and concise knowledge base, and promote interoperability between applications using different knowledge graphs. Entity alignment is a mechanism to achieve knowledge fusion. This task is widely used in many fields such as link prediction, recommendation systems, and question answering.

[0003] Entity alignment is a mechanism for achieving knowledge fusion. Entity alignment aims to correspond entities in different knowledge graphs, thereby measuring the "distance" in the shared embedding space and achieving structural alignment between knowledge graphs. Entity alignment uses seed entities for model training and ultimately achieves alignment of unknown entity pairs, so entity similarity plays an important role in this field.

[0004] Translation-based representation methods mainly project entities and relations into low-dimensional vector spaces. Although these methods have achieved certain success, these models cannot express the complex relationships between entities and cannot utilize the rich topological semantic information between entities, resulting in insufficient entity representation.

[0005] Thanks to the rapid development of graph neural networks, especially in the representation of graph structured data, GNN-based methods can use topological information to show good performance in entity alignment. Compared with translation-based models, GNN-based models contain topological information of entities and train GNN to integrate entities of each KG. This method can aggregate neighborhood structure information and enrich entity vector representation.

[0006] Most existing GNN-based entity alignment methods directly treat multi-relation knowledge graphs as a single relationship, and rarely consider the type of relationship between entities, resulting in insufficient and inaccurate entity representation. Although some methods consider complex multiple relationships between entities, they are mainly oriented towards explicit relationships and ignore the potential relationships between entities. Obviously, it is not enough to only consider explicit relationships. How to fuse and mine multiple relationships and potential relationships between entities to obtain better entity representation for entity alignment is the focus of this field.

[0007] To solve the above problems, we propose a graph convolutional neural network (MFLM-GCN) based on multi-relation fusion and latent relation mining, which makes full use of the topological semantic information of entities. It is worth noting that MFLM-GCN is an extension of our previous conference paper. In this work, we have made significant extensions, mainly in the following aspects: (i) We added a new preprocessing module to enhance the connectivity between entities and promote information propagation within and between graphs through local structural isomorphism between knowledge edge graphs and graph random walks. (ii) A new multi-relation fusion module was added to locally and globally fuse the multi-relation information between entities and enhance the representation ability of entities. (iii) The latent relation mining module was optimized using the graph random walk results, and all fully associated graphs were pruned to improve the mining performance of latent relations between entities. (iii) We added many new experiments to verify the effectiveness of MFLM-GCN, such as effect and performance comparison, parameter analysis, ablation verification, etc. Experimental results show that our method can obtain better entity node representation, thereby improving performance. Specifically, we first connect two knowledge graphs through seed entities to promote information propagation between graphs. Second, we exploit the local isomorphism between seed entities and graph random walks to capture meaningful latent relations. Third, we propose an explicit and implicit multi-relation fusion method for the preliminary representation of entities, which can aggregate relations in a local and global manner. Fourth, to further explore the latent relations between entities, we adopt a multi-head attention mechanism to generate multiple complete association graphs to represent the latent associations between entities, and construct a dense connection layer to obtain a multi-branch representation of entities. In addition, to improve the aggregation performance of the dense connection layer, we prune all graphs related to the previous graph random walk results before information aggregation. Finally, a linear layer is used to fuse entity features to obtain the final entity embedding and complete entity alignment. Summary of the invention

[0008] The main purpose of the present invention is a new multi-relation fusion and potential relationship mining graph convolutional network for entity alignment, including:

[0009] Step S10: preprocessing the original knowledge graph, including steps S11-S12;

[0010] Step S11: Fusion and isomorphism of two knowledge graphs

[0011] T=T∪{(u,align,v)|(u,v)∈S,align∈R},

[0012]

[0013] Based on the graph fusion and isomorphism process, the original graph G(o) and its corresponding adjacency matrix A(o) are obtained;

[0014] Step S12: Expand neighbor nodes by random walk of the graph

[0015]

[0016] where e i represents the input query node, f(x) represents the algorithm for calculating the proximity between nodes, Indicates e i The new neighbor sequence formed, Represents node e i 1-hop neighbor of Indicates nodes with higher intimacy except 1-hop neighbors. ;

[0017] Step S20: extracting multi-relationship information of the entity node, including steps S21-S22;

[0018] Step S21: Aggregate local relationships of entity nodes

[0019]

[0020] Among them, N r (e i ) indicates that through the relationship r and entity node e i A collection of connected adjacent entity nodes. Represents the adjacent entity node e under the first-level relationship r i The hidden vector of The weight matrix corresponding to the linear transformation of the input;

[0021] Step S22: Globally aggregate all relationship information of all previous layers

[0022]

[0023] in represents the hidden vector of the entity node in the lth layer. We propose a global relationship aggregator It uses the relationship weight matrix Learn the local relation information of layer l and fuse the hidden entity node vectors of layer l-1 through splicing aggregation operators such as splicing or multi-layer perceptron (MLP);

[0024] Step S30: extracting potential structural information between entities, including steps 31-33;

[0025] Step S31: Obtain a fully connected knowledge graph through multi-head attention and generate multiple attention adjacency matrices

[0026]

[0027] Where z is the multi-head attention index and the full association graph, z∈{1,…,m}. A(z) represents the adjacency weight matrix of the zth full association graph, and A(0) represents the weight matrix of a ij The initial adjacency weight matrix is ​​composed of Q∈Rz×d, K∈Rz×d is the feature vector of the entity node in the 1-1 layer. As the linear transfer matrices of Q and K respectively. z is the dimension of the feature output;

[0028] Step S32: Introduce dense connection layers and data enhancement mechanisms to promote deeper information propagation. The node feature representation generated by the lth layer of the zth attention head is defined as

[0029]

[0030] The sub-layer calculation formula is as follows:

[0031]

[0032] Step S33: Obtain the final relationship perception vector of the entity node through the linear connection layer

[0033]

[0034] Among them, b line is the bias term of the linear fusion formula, W line is the linear weight matrix. β is a hyperparameter;

[0035] Step S40: align entities according to the distance between their vectors. The distance formula between two entities is specifically expressed as

[0036]

[0037] Among them, e i From G1, e j From G2;

[0038] Compared with the prior art, the above technical solution conceived by the present invention has at least the following beneficial effects:

[0039] (1) The entity alignment method based on graph convolution of the present invention is applicable to various types of knowledge graphs, which helps to implement different entity alignment models according to different needs of users, helps users obtain the required entities more accurately, and meets users' needs for entity retrieval.

[0040] (2) The improved method for aligning multiple relationship entities used in the present invention can meet the needs of users more accurately and conveniently, has low requirements on the user's computer hardware configuration, occupies less user computer resources, and provides a better user experience.

[0041] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0043] Figure 1 It is a flowchart of a graph convolution-based entity alignment method according to an exemplary embodiment.

[0044] Figure 2 is a baseline comparison diagram of a graph convolution-based entity alignment method according to an exemplary embodiment.

[0045] Figure 3 It is a large-scale data representation diagram of a graph convolution-based entity alignment method according to an exemplary embodiment.

[0046] Figure 4 It is a relationship utilization effect diagram of a graph convolution-based entity alignment method according to an exemplary embodiment.

[0047] Figure 5 It is a module effect diagram of a graph convolution-based entity alignment method according to an exemplary embodiment.

[0048] Figure 6 is a time and space saving diagram of a graph convolution-based entity alignment method according to an exemplary embodiment.

Claims

1. A new graph convolutional network for multi-relation fusion and latent relation mining for entity alignment, include: Step S10: preprocessing the original knowledge graph, including steps S11-S12; Step S11: Fusion and isomorphism of two knowledge graphs T=TU{(u,align,v)|(u,v)∈S,align∈R}, Based on the graph fusion and isomorphism process, the original graph G(o) and its corresponding adjacency matrix A(o) are obtained; Step S12: Expand neighbor nodes by random walk of the graph where e i represents the input query node, f(x) represents the algorithm for calculating the proximity between nodes, Indicates e i The new neighbor sequence formed, Represents node e i 1-hop neighbor of Indicates nodes with higher intimacy except 1-hop neighbors. ; Step S20: extracting multi-relationship information of the entity node, including steps S21-S22; Step S21: Aggregate local relationships of entity nodes Among them, N r (e i ) indicates that through the relationship r and entity node e i A collection of connected adjacent entity nodes. Represents the adjacent entity node e under the first-level relationship r i The hidden vector of The weight matrix corresponding to the linear transformation of the input; Step S22: Globally aggregate all relationship information of all previous layers in represents the hidden vector of the entity node in the lth layer. We propose a global relationship aggregator It uses the relationship weight matrix Learn the local relation information of layer l and fuse the hidden entity node vectors of layer l-1 through splicing aggregation operators such as splicing or multi-layer perceptron (MLP); Step S30: extracting potential structural information between entities, including steps 31-33; Step S31: Obtain a fully connected knowledge graph through multi-head attention and generate multiple attention adjacency matrices Where z is the multi-head attention index and the full association graph, z∈{1,…,m}. A(z) represents the adjacency weight matrix of the zth full association graph, and A(0) represents the weight matrix of a ij The initial adjacency weight matrix is ​​composed of Q∈Rz×d, K∈Rz×d is the feature vector of the entity node in the 1-1 layer. As the linear transfer matrices of Q and K respectively. z is the dimension of the feature output; Step S32: Introduce dense connection layers and data enhancement mechanisms to promote deeper information propagation. The node feature representation generated by the lth layer of the zth attention head is defined as The sub-layer calculation formula is as follows: Step S33: Obtain the final relationship perception vector of the entity node through the linear connection layer Among them, b line is the bias term of the linear fusion formula, W line is the linear weight matrix. β is a hyperparameter; Step S40: aligning entities according to the distance between their vectors. The distance formula between two entities is specifically expressed as Among them, e i From G 1 , e j From G 2。

Images