An entity alignment method based on Chinese electronic medical record knowledge graph

Abstract

The present invention proposes an entity alignment method based on the Chinese electronic medical record knowledge graph. The method includes: constructing a training set and a test set for entity alignment on a heterogeneous medical knowledge graph; using rules to reason about the medical knowledge graph, completing missing relationships, and alleviating the structural heterogeneity between medical knowledge graphs; constructing a dual-view graph neural network model based on a gating mechanism, aiming to align and fuse heterogeneous medical knowledge graphs. On the basis of taking into account both accuracy and labor costs, the accuracy of Hits@5 is as high as 85.4%, effectively integrating existing medical resources and promoting the development of smart medicine.

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing, and relates to an entity alignment method in the construction of a Chinese electronic medical record (EMR) knowledge graph (KG), and specifically refers to a medical knowledge graph alignment method based on a neural network and a gating mechanism. Background Art

[0002] Chinese electronic medical records are one of the products of information-based medical and health services. They contain a large number of medical facts. With the accumulation of a large number of electronic medical records in China, it is of great significance to use natural language processing technology to automatically obtain and integrate effective medical information from a large number of electronic medical records. Building a knowledge graph related to electronic medical records is one of the most effective ways to display and utilize medical information in electronic medical records. Knowledge graph is a form of knowledge representation that can structure and standardize information elements and clearly display them in the form of a graph. The knowledge graph based on electronic medical records is a vertical field knowledge graph, which is conducive to the establishment of a scientific intelligent medical knowledge base and knowledge network. However, due to different sources of knowledge and construction purposes, knowledge overlap and complementarity occur between different knowledge graphs, especially in medical knowledge graphs. This phenomenon is common. Building a large-scale medical knowledge graph through knowledge fusion technology helps to improve data quality and promote the development of intelligent medicine. The most critical technology in knowledge fusion is entity alignment, whose goal is to determine whether two entities in different knowledge graphs point to the same thing in the real world.

[0003] In general domains, early entity alignment used similarity metrics to determine whether two entities are aligned by calculating their string similarity and structural similarity. However, such methods usually align entities according to manually designed rules, making them very complex and difficult to implement. Later, with the development of knowledge graph representation learning, more and more researchers began to use knowledge graph representation learning instead of symbolic formalism to solve problems. The most famous knowledge graph representation learning model is a set of translation-based models, one of the most classic of which is TransE. Although translation-based methods do not rely on manually constructed rules, however, translation-based models can only obtain the first-order neighborhood information of entities, so the learned entity embedding can only represent the local features of the entity and cannot fully utilize the structural information of the knowledge graph.

[0004] With the development of graph neural networks in recent years, some studies have used graph neural networks to model the structure of knowledge graphs and enhance entity embedding with neighbor entities, that is, using graph convolution to recursively aggregate the representations of neighbor entities to learn the central entity representation. Therefore, the more similar the entity neighborhood structure is, the closer the representation learned by the graph neural network-based model for the central entity is. However, due to different knowledge sources and construction purposes, there is a widespread structural heterogeneity between knowledge graphs, which seriously affects the development of such models. Wang et al. (Cross-lingual knowledge graph alignment via graph convolutional networks. In: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp. 349–357 (2018)) first attempted to use graph convolutional networks for entity alignment. Ye et al. (A vectorized relational graph convolutional network for multi-relational network alignment. In: IJCAI, pp. 4135–4141 (2019)) introduced the translation characteristics of the TransE model in the convolution process, adding different neighbors of the entity to the corresponding relation vector for merged representation. However, none of them considers the impact of knowledge graph heterogeneity on entity alignment.

[0005] At present, entity alignment methods in the medical field are very scarce, and there is no systematic and efficient solution. Due to the complex and diverse manifestations of entities in the medical field, such as the disease entity "chronic hepatitis B" may exist in the knowledge graph as the entity "hepatitis B" or "HBV infection", so the general field based on similarity measurement methods are not applicable to the medical field. In addition, due to the obvious structural differences in medical knowledge graphs, the translation-based model only integrates the first-order neighborhood of the entity and cannot fully utilize the structural information of the knowledge graph. At the same time, the ability to alleviate the heterogeneity of the knowledge graph is limited, so this method is not suitable for entity alignment in the medical field. Many researchers have tried to use graph neural networks to solve the entity alignment problem in general fields, but due to the complexity of knowledge in the medical field, there are significant differences in the graph structures generated by different electronic medical records, and the structural heterogeneity is more obvious in the medical knowledge graph. At present, there is almost no research on using graph neural networks to solve entity alignment in the medical field. Summary of the invention

[0006] The present invention proposes a dual-perspective graph neural network entity alignment model to solve the problem of medical knowledge graph fusion based on Chinese electronic medical records, which helps to improve the data quality of existing medical knowledge graphs and promote the development of smart medicine.

[0007] The invention mainly consists of three parts: 1. Constructing training sets and test sets for entity alignment on heterogeneous medical knowledge graphs; 2. Using rules to reason about medical knowledge graphs, complete missing relationships, and alleviate the structural heterogeneity between medical knowledge graphs; 3. Constructing a dual-view graph neural network model based on a gating mechanism, aiming to align and fuse heterogeneous medical knowledge graphs.

[0008] The technical solution adopted by the present invention is as follows:

[0009] An entity alignment method based on the Chinese electronic medical record knowledge graph obtains entity sets from two heterogeneous medical knowledge graphs and annotates the entities as seed pairs. For example, the entities "AIDS" and "AIDS" from the knowledge graphs MeKG_1 and MeKG_2 are a set of seed pairs. The successfully annotated entity seed pairs are used as training sets and test sets; the following steps are also included:

[0010] Step 1: Use rules to complete the knowledge graph;

[0011] The present invention completes the medical knowledge graph by filling in the missing relationships to alleviate the differences in the structure of heterogeneous medical knowledge graphs. According to the preset rules in each heterogeneous medical knowledge graph, all rules are merged to obtain a rule set K; the rule set K is applied to each heterogeneous medical knowledge graph, a rule κ∈K is given, all premise triples that satisfy the rule are found, and the conclusion heterogeneous triples are inferred according to the rule. If it does not exist in the original knowledge graph, it is added to the original knowledge graph to achieve the purpose of knowledge graph completion; two completed heterogeneous medical knowledge graphs G=(E,R,T) and G′=(E′,R′,T′) are obtained, where E and E′ represent the first entity set and the second entity set, respectively, R and R′ represent the first relationship set and the second relationship set, respectively, T and T′ represent the first triple set and the second triple set, respectively, e∈E, r∈R, t∈T represent any entity, relationship, triple in G, respectively; the present invention completes the medical knowledge graph by filling in the missing relationships to alleviate the differences in the structure of heterogeneous medical knowledge graphs.

[0012] Step 2: construct a dual-view graph neural network model based on a gating mechanism;

[0013] The ultimate goal of entity alignment is to find all pairs of entities that point to the same entity in E and E'. At present, most neural network-based entity alignment models cannot solve the problems caused by the heterogeneity of medical knowledge graphs. Based on existing research, this paper proposes a novel dual-view graph neural network model DvGNet based on a gating mechanism to solve the entity alignment problem. This model alleviates the heterogeneity of medical knowledge graphs from two perspectives. The entity representation of each perspective is obtained by aggregating multi-hop neighborhood information through a gating mechanism.

[0014] The model takes two medical knowledge graphs G and G' as input and the embedding representation of each entity as output. It determines whether they should be aligned by measuring the distance between the embedding representations of the entities. The model comprehensively considers the structural heterogeneity of the knowledge graph from the perspective of entity interaction and relationship interaction. The model details are as follows:

[0015] Step 2.1, construct entity interaction perspective network;

[0016] Ideally, if the central entities to be aligned have the same neighborhood structure, the graph neural network-based method can learn very similar entity representations for the central entities. However, due to different construction purposes and sources, medical knowledge graphs often have pattern heterogeneity and incompleteness, which leads to different neighborhoods of central entities. This is an important reason for structural heterogeneity, which is referred to as entity neighborhood heterogeneity in this paper.

[0017] The entity interaction perspective iteratively learns accurate self-attention scores for each neighbor of an entity through the self-attention mechanism. During the training process, higher weights are given to important neighbors to alleviate the heterogeneity of the entity neighborhood. i The self-attention score is calculated by the embedding representation of its neighbor entities, and the neighbor feature vector of the entity is aggregated using the self-attention score to obtain the entity e i The representation of layer l in the entity interaction perspective;

[0018] Step 2.2, construct the relationship interaction perspective network;

[0019] Most current studies aim to solve entity neighborhood heterogeneity. However, due to the relational independence between medical knowledge graphs, there is relational heterogeneity between knowledge graphs. For example, a relationship that exists in knowledge graph G may not exist in knowledge graph G', which is also an important reason for the structural heterogeneity of medical knowledge graphs. The scope of the entity interaction perspective is limited to a single knowledge graph, ignoring the relational heterogeneity between medical knowledge graphs. Therefore, the present invention proposes to use relational interaction to solve this problem.

[0020] The relationship feature matrices of the completed heterogeneous medical knowledge graphs G and G' interact with each other to obtain the relationship similarity matrix, and then perform the maximum pooling operation to obtain the relationship matching vector; finally, the cross-graph matching score obtained from the relationship matching vector is used to aggregate the information from the neighbors; any entity e in the completed heterogeneous medical knowledge graph is obtained. i The representation of level l in the relational interaction perspective;

[0021] Step 2.3, gated aggregation;

[0022] Current research believes that the output layer representation of a multi-layer network aggregates the multi-hop neighborhood information of the entity, so researchers regard the output layer representation of the network as the final embedding of the entity. However, as the number of network layers increases, the number of neighbors aggregated by the central entity increases exponentially, so the representation of the output layer brings a lot of noise to the central entity.

[0023] In order to obtain more accurate entity representation, this paper proposes a new gating mechanism to aggregate the embeddings of hidden layers and output layers, and applies it to the above two perspectives. The gating mechanism can capture the multi-hop neighborhood information of the entity to enhance the embedding of the entity while removing the redundant noise of each layer, thereby solving the noise problem caused by multi-layer convolution.

[0024] The entity interaction perspective network constructed in step 2.1 and the relationship interaction perspective network constructed in step 2.2 are gated and aggregated to obtain the output h of the entity interaction perspective network. i,1 and the output h of the relational interaction perspective network i,2 ;

[0025] The outputs of the two perspective gated aggregations are aggregated again through the gating mechanism to obtain the entity e i The final representation h i ;

[0026] Step 3, calculate the embedding distance;

[0027] Based on the final representation of the entity obtained in step 2.3, d(·) is used to calculate the embedding distance between the entities of two heterogeneous medical knowledge graphs. The smaller the distance, the more similar the entities are. d(·) is the L2 norm. The seeds in the training set are used to train the supervised model so that the distance of aligned entities gradually decreases and the distance of non-aligned entities gradually increases.

[0028] Furthermore, the above step 2.1 constructs the self-attention score in the entity interaction perspective network The calculation formula is as follows:

[0029]

[0030]

[0031] in is the self-attention coefficient, representing the entity e j For i the importance of represents the neighbors of the entity including itself, “||” represents vector concatenation, LeakyReLU(·) is the activation function; W 1 ,W 2 and p are trainable parameters.

[0032] In graph neural networks, the representation of a node is learned by recursively aggregating the feature vectors of its neighbors. Different aggregation strategies lead to different variants of graph neural networks. Graph Attention Network is a very popular variant of graph neural networks.

[0033] Furthermore, the self-attention score of formula (1) is used in the above step 2.1 to construct the entity interaction perspective network Aggregate Entity i Neighbor feature vector of entity e i The representation of the lth layer in the entity interaction perspective is calculated as follows:

[0034]

[0035] in, is the weight parameter of the lth layer in the perspective network, σ(·) is the activation function, and ReLU(·) is selected.

[0036] Furthermore, in the above step 2.2 to construct the relationship interaction perspective network, the calculation formula of the relationship matching vector is:

[0037] M=f M (f S (R,R′)) (4)

[0038] Among them, f S (·) represents the relationship similarity calculation function, defined as f S (R,R′)=R T R′, R and R′ represent the relationship feature matrices of the two medical knowledge graphs to be aligned; f M (·) represents the maximum pooling operation function; the cross-graph matching score is calculated by the relationship matching vector M The formula is as follows:

[0039]

[0040]

[0041] Where M[·] represents the relationship matching index operation; T represents the triple set of the knowledge graph;

[0042] Using cross-graph matching scores Computational Entity i The representation of the lth layer in the relational interaction perspective is calculated as:

[0043]

[0044] in is the weight parameter of the lth layer in the perspective network, σ(·) is the activation function, and ReLU(·) is selected.

[0045] Furthermore, in the above step 2.2 to construct the relational interaction perspective network, the initialization representation of relation r is obtained by the embedding representation of the head and tail entities of all triples with r as the relation, and the calculation formula is as follows:

[0046]

[0047] Where T r represents a set of triples with r as the relation, e h and e t Represent the corresponding head and tail entity embeddings respectively.

[0048] Furthermore, in the above step 2.3 gated aggregation, the gated aggregation formula of the ε layer network is

[0049]

[0050] where ρ ξ (α,β)=g ξ ·α+(1-g ξ )·β,g ξ is a trainable parameter used to control the output of each layer of each network; τ is the perspective category, i.e., the relationship interaction perspective or the entity interaction perspective;

[0051] Formula (9) is used to gate and aggregate the entity interaction perspective network constructed in step 2.1 and the relationship interaction perspective network constructed in step 2.2, respectively, to obtain the output h of the entity interaction perspective network. i,1 and the output h of the relational interaction perspective network i,2 ;

[0052] Entity i The final representation of is obtained by re-aggregating the outputs of the two perspective gating aggregations through the gating mechanism. The calculation formula is as follows:

[0053] h i =g 1 ·h i,1 +(1-g 1 )·h i,2 (10)

[0054] where g1 is a trainable parameter used to control the aggregation of the two views.

[0055] The beneficial effect of the present invention is that the present invention first uses the constructed training set to train the proposed dual-view graph neural network model based on the gating mechanism, and uses the trained model to mine unknown aligned entity pairs in the medical knowledge graph, which can realize the fusion of heterogeneous medical knowledge graphs. On the basis of taking into account the accuracy and labor cost, the accuracy of Hits@5 is as high as 85.4%, which effectively integrates the existing medical resources and promotes the development of smart medicine. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Diagram of the medical knowledge graph alignment framework.

[0057] Figure 2 Dual-view graph neural network model based on gating mechanism. DETAILED DESCRIPTION

[0058] The technical solution of the present invention is described in detail with reference to the accompanying drawings.

[0059] like Figure 1 Figure 1 is a diagram of the medical knowledge graph alignment framework. Table 1 is the detailed data statistics of the heterogeneous medical knowledge graph.

[0060] 1. Corpus Preprocessing

[0061] The present invention manually annotates the heterogeneous medical knowledge graphs MeKG_1 and MeKG_2 constructed by electronic medical records, and uses the manually annotated seed pairs as training sets and test sets. The number of seed pairs in the training set and the test set are 227 and 136, respectively.

[0062] 2. Use rules to complete the knowledge graph

[0063] First, experts derive rules from each medical knowledge graph, and then convert these rules between medical knowledge graphs based on the knowledge invariance assumption to form a unified set of rules. After that, rule matching is performed in each medical knowledge graph. Finally, to ensure the quality of the triples obtained by rule reasoning, experts are asked to screen the inferred triples.

[0064] (1) Reasoning and transfer of rules

[0065] Experts design the rules that exist in each medical knowledge graph and merge all the rules to get the rule set K. For example,

[0066] (2) Implementation of rules

[0067] Apply the rule set K obtained in step (1) to each medical knowledge graph. Given a rule κ∈K, find all premise triples that satisfy the rule and infer the conclusion triples based on the rule. If they do not exist in the original knowledge graph, add them to the original knowledge graph, thus achieving the purpose of knowledge graph completion. For example, according to the above rules, we get:

[0068] ①

[0069] Assumptions Figure 1 The entity e in the equation is "imbalance syndrome", the entity g is "hemodialysis", and the entity k is "coma". Then the relationship r5 is "disease shows symptoms", and the relationship r6 is "treatment causes disease". Since "hemodialysis" treatment may cause the complication "imbalance syndrome", the relationship between e and k can be inferred that the relationship between g and k, r7, is "treatment causes symptoms".

[0070] ②

[0071] Assumptions Figure 1 The entity e in the table is "nephrotic syndrome", the entity g is "prednisone", and the entity k is "proteinuria". Then the relationship r5 is "disease shows symptoms", and the relationship r6 is "treatment improves the disease". Since "prednisone" can treat "nephrotic syndrome", the relationship between e and k can be inferred that the relationship between g and k, r7, is "symptomatic treatment".

[0072] (2) Expert appraisal

[0073] Experts screened the triples inferred in step (2) to ensure their quality. After expert screening, the number of triples inferred from the two medical knowledge graphs was 11668 and 11820 respectively. The detailed data of the completed knowledge graphs are shown in Table 1.

[0074] Table 1: Statistics of heterogeneous knowledge graph data

[0075]

[0076] 3. Model training

[0077] (1) Obtaining entity interaction perspective embedding

[0078] For any entity e in G i , which uses the self-attention score calculated by formula (1) Iteratively aggregate neighbor information for the central entity e i The entity interaction perspective is embedded and represented, and the gating mechanism is used to control the output aggregation of each layer of the network. In this embodiment, the number of network layers is set to 2, and the entity ei The calculation formula for the embedding representation in the τth perspective is as follows:

[0079]

[0080] where g 0 It is a trainable parameter used to control the aggregation of the hidden layer and output layer representation of the network. When the number of network layers is set to 2, is the hidden layer embedding of the network, is the output layer embedding of the network. Finally, we get entity e i Embedding h from the entity interaction perspective i,1 .

[0081] (2) Obtaining relational interaction perspective embedding

[0082] Similar to (1), for any entity e in G i , which uses the cross-graph matching score calculated by formula (5) Iteratively aggregate neighbor information for the central entity e i The relationship interaction perspective embedding representation is used, and the gating mechanism is used to control the output aggregation of each layer of the network. According to formula (11), the entity e is finally obtained i Embedding h in the relational interaction perspective i,2 .

[0083] (3) Gated aggregation

[0084] The entity interaction perspective embedding h obtained by steps (1) and (2) is aggregated using formula (10) i,1 and the embedding of the relational interaction perspective i,2 , get entity e i The final embedding representation h i .

[0085] Similarly, the same processing as steps (1), (2), and (3) is performed on all entities in G'.

[0086] (4) Calculate the embedding distance

[0087] Based on the entity embedding representation obtained in step (3), d(·) is used to calculate the embedding distance between entities. The smaller the distance, the more similar the entities are. d(·) is defined as the L2 norm. The supervised model is trained using the seeds in the training set so that the distance between aligned entities gradually decreases and the distance between non-aligned entities gradually increases.

[0088] 4. Model Quality Assessment

[0089] This embodiment uses Hits@k and MRR as evaluation indicators of the model. Among them, Hits@1 represents the percentage of correctly aligned entities among the first k candidate entities, k is set to 1, 5, 10 and 50 respectively, and MRR represents the average of the last ranked correctly aligned entities. The higher the values ​​of Hits@k and MRR, the better the model performance. The results of the model on the test set are shown in Table 2.

[0090] Table 2: Model DvGNet results table

[0091] Hits@1 Hits@5 Hits@10 Hits@50 MRR DvGNet 75.9％ 85.4％ 86.9％ 95.6％ 80.6％

[0092] 5. Using DvGNet to mine unknown aligned entity pairs

[0093] Since only a small number of seed pairs are manually labeled as training sets and test sets, there are still a large number of entity pairs to be aligned in the knowledge graph. The present invention uses DvGNet to mine the remaining entity pairs to be aligned in the medical knowledge graphs MeKG_1 and MeKG_2. As can be seen from the results in Table 2, the accuracy of Hits@5 is as high as 85.4%. On the basis of taking into account both accuracy and labor costs, DvGNet recommends the top 5 candidate entities with the highest similarity for each entity to be aligned, and experts annotate the candidate entities to obtain the final alignment results. As shown in Table 3, the alignment entity of the entity "acute lymphocytic leukemia" is "ALL", the alignment entity of the entity "Alzheimer's disease" is "Alzheimer's disease", and the alignment entity of "Bartter syndrome" is "Bartter syndrome". The number of entity pairs to be aligned finally identified by the model is 546.

[0094] Table 3: Entity alignment candidate example table

[0095] Entities to be aligned Top1 Top2 Top3 Top4 Top 5 Acute lymphoblastic leukemia ALL leukemia Acute leukemia Central nervous system leukemia CNSL Alzheimer's disease Alzheimer's disease Memory impairment dementia ADS Cognitive dysfunction Bartter syndrome Hypokalemic alkalosis Bartter syndrome Hypokalemia Gitleman Synthesis Juxtaglomerular hyperplasia

Claims

1. An entity alignment method based on the Chinese electronic medical record knowledge graph, characterized in that: The entity sets of two heterogeneous medical knowledge graphs are obtained respectively, and the seed pairs of entities are annotated, and the successfully annotated entity seed pairs are used as training sets and test sets; the following steps are also included: Step 1: Use rules to complete the knowledge graph; According to the preset rules in each heterogeneous medical knowledge graph, all rules are merged to obtain the rule set K; the rule set K is applied to each heterogeneous medical knowledge graph, given a rule κ∈K, all premise triples that satisfy the rule are found, and the conclusion heterogeneous triples are inferred according to the rule. If they do not exist in the original knowledge graph, they are added to the original knowledge graph to achieve the purpose of knowledge graph completion; two completed heterogeneous medical knowledge graphs G=(E,R,T) and G′=(E′,R′,T′) are obtained, where E and E′ represent the first entity set and the second entity set respectively, R and R′ represent the first relationship set and the second relationship set respectively, T and T′ represent the first triple set and the second triple set respectively, and e∈E, r∈R, t∈T represent any entity, relationship, and triple in G respectively; Step 2: construct a dual-view graph neural network model based on a gating mechanism; Step 2.1, construct entity interaction perspective network; According to any entity e in the completed heterogeneous medical knowledge graph i The self-attention score is calculated by the embedding representation of its neighbor entities, and the neighbor feature vector of the entity is aggregated using the self-attention score to obtain the entity e i The representation of layer l in the entity interaction perspective; Step 2.2, construct the relationship interaction perspective network; The relationship feature matrices of the completed heterogeneous medical knowledge graphs G and G' interact with each other to obtain a relationship similarity matrix, and then perform a maximum pooling operation to obtain a relationship matching vector; finally, the cross-graph matching score obtained from the relationship matching vector is used to aggregate information from neighbors; Get any entity e in the completed heterogeneous medical knowledge graph i The representation of level l in the relational interaction perspective; Step 2.3, gated aggregation; The entity interaction perspective network constructed in step 2.1 and the relationship interaction perspective network constructed in step 2.2 are gated and aggregated to obtain the output h of the entity interaction perspective network. i,1 and the output h of the relational interaction perspective network i,2 ; The outputs of the two perspective gated aggregations are aggregated again through the gating mechanism to obtain the entity e i The final representation h i ; Step 3, calculate the embedding distance; Based on the final representation of the entity obtained in step 2.3, d(·) is used to calculate the embedding distance between the entities of two heterogeneous medical knowledge graphs. The smaller the distance, the more similar the entities are. where d(·) is the L2 norm; the supervised model is trained using seeds in the training set so that the distance of aligned entities gradually decreases and the distance of non-aligned entities gradually increases.

2. The entity alignment method based on the Chinese electronic medical record knowledge graph according to claim 1 is characterized in that: The step 2.1 constructs the self-attention score in the entity interaction perspective network The calculation formula is as follows: in is the self-attention coefficient, representing the entity e j For i the importance of represents the neighbors of an entity including itself, "||" represents vector concatenation, LeakyReLU(·) is the activation function; W1, W2 and p are trainable parameters.

3. The entity alignment method based on the Chinese electronic medical record knowledge graph according to claim 2 is characterized in that: The step 2.1 constructs the entity interaction perspective network using the self-attention score of formula (1) Aggregate Entity i Neighbor feature vector of entity e i The representation of the lth layer in the entity interaction perspective is calculated as follows: in, is the weight parameter of the lth layer in the perspective network, σ(·) is the activation function, and ReLU(·) is selected.

4. The entity alignment method based on the Chinese electronic medical record knowledge graph according to claim 1 is characterized in that: In the step 2.2 of constructing the relationship interaction perspective network, the calculation formula of the relationship matching vector is: M=f M (f S (R,R′)) (4) Among them, f S (·) represents the relationship similarity calculation function, defined as f S (R,R′)=R T R′, R and R′ represent the relationship feature matrices of the two medical knowledge graphs to be aligned; f M (·) represents the maximum pooling operation function; the cross-graph matching score is calculated by the relationship matching vector M The formula is as follows: Where M[·] represents the relationship matching index operation; T represents the triple set of the knowledge graph; Using cross-graph matching scores Computational Entity i The representation of the lth layer in the relational interaction perspective is calculated as: in is the weight parameter of the lth layer in the perspective network, σ(·) is the activation function, and ReLU(·) is selected.

5. The entity alignment method based on the Chinese electronic medical record knowledge graph according to claim 1 is characterized in that: In the step 2.2 of constructing the relational interaction perspective network, the initialization representation of relation r is obtained by the embedding representation of the head and tail entities of all triples with r as the relation, and the calculation formula is as follows: Where T r represents a set of triples with r as the relation, e h and e t Represent the corresponding head and tail entity embeddings respectively.

6. The entity alignment method based on the Chinese electronic medical record knowledge graph according to claim 1 is characterized in that: In the step 2.3 gated aggregation, the gated aggregation formula of the ε layer network is: where ρ ξ (α,β)=g ξ ·α+(1-g ξ )·β,g ξ is a trainable parameter used to control the output of each layer of each network; τ is the perspective category, i.e., the relationship interaction perspective or the entity interaction perspective; Formula (9) is used to gate and aggregate the entity interaction perspective network constructed in step 2.1 and the relationship interaction perspective network constructed in step 2.2, respectively, to obtain the output h of the entity interaction perspective network. i,1 and the output h of the relational interaction perspective network i,2 ; Entity i The final representation of is obtained by re-aggregating the outputs of the two perspective gating aggregations through the gating mechanism. The calculation formula is as follows: h i =g1·h i,1 +(1-g1)·h i,2 (10) Where g1 is a trainable parameter used to control the aggregation of two views.

Images