A graph convolutional network entity alignment method based on multi-information aggregation
By introducing a multi-information aggregation mechanism in the graph convolution network, learning entity relationships and structural features respectively and performing gated aggregation, the entity alignment errors caused by insufficient feature information and structural differences are solved, and the entities between large-scale knowledge graphs are significantly improved. Alignment effect.
Patent Information
- Application Number
- CN202210137390.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-15
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2042-02-15
AI Technical Summary
The entity alignment method based on graph convolutional networks has representation learning errors caused by insufficient feature information and structural differences in large-scale knowledge graphs, resulting in poor alignment results.
The graph convolution network method of multi-information aggregation is adopted to extract and learn feature information through the entity relationship feature learning module and the structural feature learning module, and aggregation of feature vectors is used to generate more accurate entity representation vectors.
Improve the accuracy and effect of entity alignment, ensuring the richness and accuracy of feature information when aligning entities between heterogeneous knowledge graphs.
Smart Images

Figure CN114676261B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of entity alignment, and mainly to a graph convolutional network entity alignment method based on multi-information aggregation. Background Art
[0002] Knowledge graph is actually a semantic network that reveals the relationship between entities. It can describe the real things and the relationship between things in the form of a graph. With the development of knowledge graph related technologies, knowledge graph has been widely used in knowledge question and answer, intelligent translation, recommendation system and other fields. At present, the construction of knowledge graph has begun to develop towards scale and specialization. Entity alignment technology is the key to realize large-scale knowledge graph construction through knowledge fusion. Entity alignment refers to finding entities in two knowledge graphs that point to the same real-world things, aligning these two entities existing in different graphs, realizing the fusion of two knowledge graphs, and thus constructing a larger-scale knowledge graph.
[0003] In view of the current development trend of knowledge graphs expanding in scale and specializing in fields, how to accurately realize the knowledge fusion of knowledge graphs has become the key to building large-scale knowledge graphs. Therefore, the research on entity alignment methods has become a hot field. The entity alignment methods that are currently used more are mainly based on entity representation learning. The entity alignment method based on entity representation learning uses machine learning to learn entities as representations of feature vectors, and calculates the vector similarity between entities as the basis for entity alignment. This type of method requires that the entity data be initially vectorized and then placed in the entity representation learning model for learning based on the existing aligned entity pairs, so that the vectors between aligned entities are more similar, and the vectors between entities that should not be aligned are more different. After obtaining the feature representation vector of the entity through learning, the vector distance between the entity vectors is calculated as the basis for entity alignment.
[0004] The entity representation learning models used in entity alignment methods based on entity representation learning can be mainly divided into methods based on translation models and methods based on graph convolutional networks. In the entity representation learning process, the entity alignment method based on the translation model uses the relationship entity triples (head entity, relationship, tail entity) in the graph to learn the representation of entities in the graph according to the mapping method of head entity vector + relationship vector = tail entity vector. The entity representation method based on the graph convolutional network model mainly performs entity representation learning operations on the entity data of the graph structure based on the neighbor structure of the entity in the graph. Due to the powerful learning ability of the graph convolutional network for the structural features of the entity in the graph, it has become the basis of more and more entity alignment methods based on entity representation learning. However, in the current entity alignment methods based on graph convolutional networks for entity representation learning, although the coping ability has been greatly improved when facing large-scale knowledge graphs, when learning entities, the feature information of the entity cannot be fully learned. Usually, the learning is targeted at a specific aspect of the features, resulting in the loss of key feature information contained in the entity representation vector and the lack of accuracy of the entity representation vector. In addition, in real graphs, since different knowledge graph structures are usually very different, the neighbor structures of aligned entities are also not similar. During the representation learning process, the graph convolutional network model continuously learns based on the neighbor structure of the central entity, resulting in aligned entities with different neighbor structures. It is very likely that different entity representation vectors will be learned, resulting in errors in alignment.
[0005] Therefore, in order to address the two bottleneck problems that limit the entity alignment effect, namely insufficient key feature information in the entity representation vector and entity representation learning errors caused by structural differences, the new entity alignment method that uses graph convolutional networks for representation learning must study the above two problems during the representation learning process and propose corresponding improvements to alleviate the impact of the problems. Summary of the invention
[0006] Purpose of the invention: In view of the problems existing in the above-mentioned background technology, the present invention provides a graph convolutional network entity alignment method based on multi-information aggregation. Based on the excellent ability of graph convolutional networks to learn features of entity data, different improved graph convolutional network models are used to learn entities from different feature information separately, and then aggregate them using a gating mechanism. This ensures that the feature information in the final entity representation vector is sufficient, the entity representation vector representation is more accurate, and the effect of entity alignment between large-scale heterogeneous knowledge graphs is improved.
[0007] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is:
[0008] A graph convolutional network entity alignment method based on multi-information aggregation includes the following steps:
[0009] Step S1, generating entity semantic initialization vectors for knowledge graph data;
[0010] Generate the corresponding entity semantic initialization vector based on the entity semantics from the input knowledge graph data;
[0011] Step S2, inputting the entity semantic initialization vector into the entity relationship feature learning module and the entity structure feature learning module for learning respectively;
[0012] The entity relationship feature learning module extracts and learns the relationship features of entities in the knowledge graph; constructs an entity relationship dual graph to achieve relationship feature modeling; and obtains an entity relationship feature representation vector by interacting with the constructed relationship graph and the original entity graph;
[0013] The entity structure feature learning module extracts and learns the adjacent structure features and the distant neighbor structure of the entity in the knowledge graph respectively; obtains the entity adjacent structure features and the entity distant structure features respectively, aggregates the feature vectors based on the Resgate gating mechanism, and obtains the entity structure feature representation vector;
[0014] Step S3, entity multi-feature aggregation;
[0015] Based on the entity structure feature representation vector and the entity relationship feature representation vector obtained in step S2, the feature vectors are aggregated through the highway gating mechanism to obtain the entity feature representation vector;
[0016] Step S4, performing entity alignment based on entity feature representation vectors;
[0017] Calculate the vector distance between entities as the similarity between entities; select the entity with the closest vector distance to the entity to be aligned for alignment.
[0018] Furthermore, the specific method of generating the entity semantic initialization vector in step S1 includes:
[0019] Step S1.1, translating the name language of the entity into the corresponding English word;
[0020] Step S1.2, according to the entity of the converted language, find the corresponding trained semantic vector in the glove.840B.300d word vector library;
[0021] Step S1.3: All corresponding semantic vectors are constructed into a matrix as an initialization input containing entity semantic feature information.
[0022] Furthermore, the extraction and learning of entity structure features in step S2 mainly includes:
[0023] Step S2.1, generate the knowledge graph entity adjacency matrix and the second-order neighbor matrix;
[0024] Step S2.2: In the process of learning each layer of structural features, the entity representation vector and the adjacency matrix of the graph entity are input into the GCN model to extract the adjacency structural feature information and learn the structural information of the first-order neighbor nodes of the central entity;
[0025] Step S2.3, input the entity representation vector and the second-order neighbor matrix of the graph entity into the GAT model, extract the second-order neighbor structural feature information, and perform structural feature learning on the second-order neighbor nodes of the central entity;
[0026] Step S2.4: Based on the Res gate mechanism, the feature vectors of the GCN model learning results and the GAR model learning results are aggregated to obtain the entity structure feature representation vector.
[0027] Furthermore, the formula for fusion feature information of the Res gate gating mechanism is as follows:
[0028] h i =σ(σW2(σ(W1h i1 +b))+h i2 )
[0029] Where W1 and W2 represent the learnable parameter matrices; σ represents the activation function; h i represents the feature representation vector of entity i after integration; h i1 、h i2 Respectively represent the entity feature representations learned by the two types of models.
[0030] Furthermore, the entity relationship features are extracted and learned in step S2, mainly including:
[0031] The entity representation vector is placed in the RDGCN model to model and learn the relational features of the entity. An entity-relationship dual graph is constructed based on the existing RDGCN model to achieve the modeling of relational features. During the learning process, the constructed entity-relationship dual graph interacts with the original entity graph to obtain the entity-relationship feature representation vector.
[0032] Furthermore, the obtained entity structure feature representation vector and entity relationship feature representation vector are aggregated through the highway gating mechanism. Specifically, the fusion feature information formula of the highway gating mechanism is as follows:
[0033] h i =gate(h i2 )·h i1 +(1-gate(h i2 ))hi2
[0034] gate(h)=sigmoid(Wh+b)
[0035] where h i represents the representation vector of entity i obtained after the two types of feature representations are integrated in the gating mechanism; h i1 、h i2 They respectively represent the entity feature representations learned by model 1 and model 2 from different feature aspects; gate(h) is the gating coefficient, which is used to control the aggregation of two types of feature information vectors.
[0036] Furthermore, in step S4, based on the obtained entity feature representation vectors, the Manhattan distance between the entity feature representation vectors is calculated as the similarity between the entities. The distance calculation formula between the vector representations of entities i and j is expressed as:
[0037] d ij =||h i -h j ||
[0038] Select the entity that is closest to the entity vector to be aligned for alignment.
[0039] Beneficial effects:
[0040] The graph convolutional network entity alignment method based on multi-information aggregation provided by the present invention performs representation learning on entities from a multi-feature perspective and aggregates feature information through a gating mechanism. This makes it possible to enrich the feature information in the learned entity representation vector and to achieve high precision in the vector representation of entities. This makes it possible to ensure that when performing entity alignment, the entity feature vector is sufficient to identify heterogeneous alignment entities, and the entity vectors that should not be aligned are clearly distinguished, thereby significantly improving the effect of entity alignment. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 It is a flow chart of the graph convolutional network entity alignment method based on multi-information aggregation provided by the present invention;
[0042] Figure 2 It is a flow chart of entity structure feature extraction and learning provided by the present invention. DETAILED DESCRIPTION
[0043] The present invention is further described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0044] The graph convolutional network entity alignment method based on multi-information aggregation provided by the present invention is as follows Figure 1 As shown, the specific steps include:
[0045] Step S1: Generate entity semantic initialization vectors for knowledge graph data. The entity semantic initialization vector is the initialized entity representation vector used in the first training process, and the entity representation vectors obtained in the previous training are used in subsequent training processes.
[0046] The input knowledge graph data is used to generate the corresponding entity semantic initialization vector according to the entity semantics. Specifically,
[0047] Step S1.1: Translate the name language of the entity into the corresponding English word.
[0048] Step S1.2: According to the entity of the converted language, find the corresponding trained semantic vector in the glove.840B.300d word vector library.
[0049] Step S1.3: All corresponding semantic vectors are constructed into a matrix as an initialization input containing entity semantic feature information.
[0050] Step S2: input the entity semantic initialization vector into the entity relationship feature learning module and the entity structure feature learning module for learning respectively.
[0051] The entity relationship feature learning module extracts and learns the relationship features of entities in the knowledge graph. The entity representation vector is placed in the RDGCN model to model and learn the relationship features of the entity. Based on the existing RDGCN model, an entity relationship dual graph is constructed to achieve the modeling of relationship features. During the learning process, the constructed entity relationship dual graph interacts with the original entity graph to obtain the entity relationship feature representation vector.
[0052] The entity structure feature learning module extracts and learns the adjacent structure features and long-distance neighbor structures of entities in the knowledge graph. The entity adjacent structure features and entity long-distance structure features are obtained respectively, and the feature vectors are aggregated based on the Res gate gating mechanism to obtain the entity structure feature representation vector. The objective function used in entity structure feature representation learning and relationship feature representation learning is set based on the purpose of making the vector distance between entities that should be aligned as small as possible and the vector distance between entities that should not be aligned as large as possible. Specifically, Figure 2 As shown,
[0053] Step S2.1, generate the knowledge graph entity adjacency matrix and the second-order neighbor matrix.
[0054] Step S2.2: In the process of learning each layer of structural features, the entity representation vector and the adjacency matrix of the graph entity are input into the GCN model to extract the adjacency structural feature information and learn the structural information of the first-order neighbor nodes of the central entity.
[0055] Step S2.3: Input the entity representation vector and the second-order neighbor matrix of the graph entity into the GAT model, extract the second-order neighbor structural feature information, and perform structural feature learning on the second-order neighbor nodes of the central entity.
[0056] Step S2.4: Based on the Res gate mechanism, the feature vectors of the GCN model learning results and the GAR model learning results are aggregated to obtain the entity structure feature representation vector.
[0057] The formula for fusion feature information of the Res gate gating mechanism is as follows:
[0058] h i =σ(σW2(σ(W1h i1 +b))+h i2 )
[0059] Where W1 and W2 represent the learnable parameter matrices. σ represents the activation function. h i It represents the feature representation vector of entity i after integration. i1 、h i2 Respectively represent the entity feature representations obtained by learning two types of models. The setting of Res gate gating allows a certain type of feature information to be directly passed during the information integration process, so that the feature information on this aspect is retained as completely as possible, while the feature information on the other hand is reduced in weight during integration by participating in the learning of the gating coefficient, thereby achieving the purpose of highlighting the feature information on a certain aspect. Therefore, the Res gate gating mechanism is used in the present invention to aggregate the feature information of adjacent structures and distant neighbor structures in the structural feature learning module.
[0060] Step S3: Aggregation of entity multi-features.
[0061] Based on the entity structure feature representation vector and entity relationship feature representation vector obtained in step S2, the highway gating mechanism is used to aggregate the feature vectors and obtain the entity feature representation vector. The formula for the fusion feature information of the highway gating mechanism is as follows:
[0062] h i =gate(h i2 )·h i1 +(1-gate(h i2 ))h i2
[0063] gate(h)=sigmoid(Wh+b)
[0064] where h i It represents the representation vector of entity i obtained after the two types of feature representations are integrated by the gating mechanism. i1 、h i2 They respectively represent the entity feature representations learned by model 1 and model 2 from different feature aspects. gate(h) is the gating coefficient, which is used to control the aggregation of two types of feature information vectors. When highway gating performs multi-feature aggregation, the feature representations of the two inputs are assigned corresponding gating coefficient weights. During the learning process, the gating coefficient enables the two types of input information to be integrated with similar coefficient weights, so that there is no offset in the features contained in the feature information after aggregation. Therefore, the highway gating mechanism is used in the present invention to aggregate the feature information of the relational feature vector and the structural feature vector.
[0065] Before the end of training, the entity representation vector is used as the initial input for the next training and the process of step S1 is continued again.
[0066] Step S4: aligning entities based on entity feature representation vectors.
[0067] The vector distance between entities is calculated as the similarity between entities. The distance calculation formula between the vector representations of entities i and j is expressed as:
[0068] d ij =||h i -h j ||
[0069] Select the entity that is closest to the entity vector to be aligned for alignment.
[0070] The present invention performs representation learning on entity data based on a graph convolutional network with multi-information aggregation, and then uses the learned entity representation as the basis for similarity calculation of entity alignment. First, in the process of generating entity feature vectors in the entity representation learning part, the input graph entity data is searched for the corresponding pre-learned entity vector in glove.840B.300d as the initial input vector of the entity representation learning module, so that the initial input and subsequent learning can contain the semantic feature information of the entity. Then the initialization vector is input into the improved RDGCN model and the structural feature learning module respectively to capture and learn the relational information and structural information respectively. In the capture and learning of structural feature information, the method proposed in this paper constructs a two-layer structural feature learning network based on the improved GCN model and the GAT model. The results obtained in each layer during the structural information learning process are subjected to feature aggregation using the Res gate gating mechanism to obtain the overall entity structure information, which is then integrated with the relational information learned by the RDGCN model through highway gating to obtain the final entity representation. In the entity alignment part, after obtaining the entity representation vector, the entity alignment is performed by calculating the Manhattan distance between the vectors as the basis of entity similarity to obtain the result required for the final entity alignment task. The specific process is shown in Table 1 below.
[0071] Table 1 Graph convolutional network entity alignment method based on multi-information aggregation
[0072]
[0073] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A graph convolutional network entity alignment method based on multi-information aggregation, It is characterized in that The following steps are involved: Step S1, generating entity semantic initialization vectors for knowledge graph data; Generate the corresponding entity semantic initialization vector based on the entity semantics from the input knowledge graph data; Step S2, inputting the entity semantic initialization vector into the entity relationship feature learning module and the entity structure feature learning module for learning respectively; The entity relationship feature learning module extracts and learns the relationship features of entities in the knowledge graph; Construct entity-relationship dual graph to realize relational feature modeling; By interacting the constructed relationship graph with the original entity graph, an entity relationship feature representation vector is obtained; The entity structure feature learning module extracts and learns the adjacent structure features and the distant neighbor structure of the entity in the knowledge graph respectively; obtains the entity adjacent structure features and the entity distant structure features respectively, aggregates the feature vectors based on the Res gate gating mechanism, and obtains the entity structure feature representation vector; Step S3, entity multi-feature aggregation; Based on the entity structure feature representation vector and the entity relationship feature representation vector obtained in step S2, the feature vectors are aggregated through the highway gating mechanism to obtain the entity feature representation vector; Step S4, performing entity alignment based on entity feature representation vectors; Calculate the vector distance between entities as the similarity between entities; Select the entity closest to the vector of the entity to be aligned for alignment; The step S2 extracts and learns the entity structure features, which mainly includes: Step S2.1, generate the knowledge graph entity adjacency matrix and the second-order neighbor matrix; Step S2.2: In the process of learning each layer of structural features, the entity representation vector and the adjacency matrix of the graph entity are input into the GCN model to extract the adjacency structural feature information and learn the structural information of the first-order neighbor nodes of the central entity; Step S2.3, input the entity representation vector and the second-order neighbor matrix of the graph entity into the GAT model, extract the second-order neighbor structural feature information, and perform structural feature learning on the second-order neighbor nodes of the central entity; Step S2.4, based on the Res gate mechanism, the learning results of the GCN model and the GAR model are aggregated into feature vectors to obtain the entity structure feature representation vector; The formula for fusion feature information of the Res gate gating mechanism is as follows: h i =σ(σW 2 (σ(W 1 h i1 +b))+h i2 ) Where W1 and W2 represent the learnable parameter matrices; σ represents the activation function; h i represents the feature representation vector of entity i after integration; h i1 、h i2 Respectively represent the entity feature representations learned by the two types of models; The entity relationship features are extracted and learned in step S2, mainly including: The entity representation vector is put into the RDGCN model to model and learn the relational features of the entity. An entity-relationship dual graph is constructed based on the existing RDGCN model to achieve the modeling of relational features. In the learning process, the constructed entity-relationship dual graph interacts with the original entity graph to obtain the entity-relationship feature representation vector. The obtained entity structure feature representation vector and entity relationship feature representation vector are aggregated through the highway gating mechanism. Specifically, the fusion feature information formula of the highway gating mechanism is as follows: h i =gate(h i2 )·h i1 +(1-gate(h i2 ))h i2 gate(h)=sigmoid(Wh+b) where h i represents the representation vector of entity i obtained after the two types of feature representations are integrated in the gating mechanism; h i1 、h i2 They represent the entity feature representations learned by model 1 and model 2 from different feature aspects respectively; gate(h) is the gating coefficient, which is used to control the aggregation of two types of feature information vectors; In step S4, based on the obtained entity feature representation vectors, the Manhattan distance between the entity feature representation vectors is calculated as the similarity between the entities. The distance calculation formula between the vector representations of entities i and j is expressed as: d ij =||h i -h j || Select the entity that is closest to the entity vector to be aligned for alignment.
2. According to the graph convolutional network entity alignment method based on multi-information aggregation according to claim 1, It is characterized in that The specific method of generating the entity semantic initialization vector in step S1 includes: Step S1.1, translating the name language of the entity into the corresponding English word; Step S1.2, according to the entity of the converted language, find the corresponding trained semantic vector in the glove.840B.300d word vector library; Step S1.3: All corresponding semantic vectors are constructed into a matrix as an initialization input containing entity semantic feature information.
