A medical knowledge graph error detection method based on implicit label information enhancement

Abstract

The application discloses a medical knowledge graph error detection method based on implicit label information enhancement. The application comprises the following steps: firstly, reading a knowledge graph G and saving it in a triple (h, r, t). Then, constructing a knowledge graph G1 based on a hypergraph perspective and a knowledge graph G2 reconstructed with triples as nodes. Next, inputting G, G1 and G2 into a model based on label information enhancement for training. Finally, evaluating by using a confidence score. The application can mine potential entity label information based on a hypergraph perspective, and use the information for knowledge graph error detection, so that the medical knowledge graph error detection result is more accurate, and the possibility of medical knowledge graph construction and improvement is provided.

Description

Technical Field

[0001] This invention belongs to the field of error detection in knowledge graphs, specifically relating to an error detection method for medical knowledge graphs based on implicit label information enhancement. Background Technology

[0002] Knowledge graphs (KGs) are semantic networks that can formally describe real-world entities and their relationships, revealing the connections between entities. In particular, domain knowledge graphs, especially medical knowledge graphs, have numerous applications in the medical field, including medical question answering and decision support systems.

[0003] However, errors are often introduced during the construction of medical knowledge graphs. To reduce these errors, existing knowledge graph error detection methods remove noisy triples during representation learning, such as CKRL (a novel confidence-aware knowledge representation learning framework) published by Xie et al. Specifically, the CAGED method (Contrastive Knowledge Graph Error Detection) proposed by Zhang et al. introduces contrastive learning, providing confidence scores for triples based on multiple views. Building on this, Zhang et al. proposed AKAE (Integrating Entity Attributes for Error-Aware Knowledge Graph Embedding), which additionally introduces entity attributes for representation learning and filters erroneous triples using confidence scores. However, AKAE has limitations because the constructed medical knowledge graph may not contain complete entity label information.

[0004] Based on real-world scenarios, this method will mine implicit entity label information from a hypergraph perspective and use it for error detection in medical knowledge graphs. Summary of the Invention

[0005] In view of this, the present invention proposes an error detection method for medical knowledge graphs based on implicit label information enhancement. Specific steps include:

[0006] S1. Obtain the medical knowledge graph to be tested. Arrange triples in the form of head entity-relation-tail entity. save;

[0007] S2, based on The hypergraph is reconstructed from the hypergraph perspective. , G1 nodes are used to mine entity label information. and Composition; at the same time, will Using triples as nodes, relationships are established between triple nodes of shared entities, and the knowledge graph is reconstructed. , Used to obtain neighbor information;

[0008] S3, will , , The input is a model based on label information enhancement for training. The model based on label information enhancement consists of a Bi-LSTM and a graph attention network with implicit label information.

[0009] S4. Construct confidence scores and use the confidence score ranking to detect erroneous triples.

[0010] Further, step S2 includes:

[0011] S21, For all shared... triples Then let At this time, it is called As a set of superedges, Let r be the set of all vertices of the hyperedge; similarly, let t be the set of all triples sharing (r,t). Then let ,say As a set of superedges, The set of all vertices of the hyperedge, obtained by reconstructing the hypergraph from its perspective. ;

[0012] S22, find all triples in G. As If the node, , X1 and X2 are nodes in G2. In G2, if h1=h2 or t1=t2 or h1=t2 or h2=t1, establish... and The relationships, after this reconstruction, result in a knowledge graph as follows: .

[0013] Further, step S3 includes:

[0014] S31. Randomly initialize the triples in G to obtain the vector. ;

[0015] S32, will Inputting into a Bi-LSTM layer yields triple vectors. ;

[0016] S33. Based on the attention mechanism, we will mine the label information in the hypergraph view, and for the obtained triples and its vector representation ,according to and The importance score of triples from a hypergraph perspective is obtained through an attention mechanism. , ,in, It is a similarity function;

[0017] in, , , ,in For the tail entity t relative to The coefficients of other entities can be obtained similarly. ;

[0018] S34, For a , obtain Neighboring Triple Group and the obtained The input graph is fed into an attention network, which obtains the vector representation Z of the reconstructed triplet based on the neighbor triplet based on the attention mechanism. The specific formula is as follows:

[0019]

[0020] in, Neighboring Triple Group Compared to triples relative coefficient Where A is an attention scoring function that converts a vector into a scalar, and W represents a trainable matrix that maps triple vectors X to the same space;

[0021] S35. Perform negative sampling training using a joint loss function.

[0022] Here, G* represents the negative sample set, which is constructed by randomly replacing the head and tail entity relationships. Let G* be a triple, [ ] + Indicates positive, >0, It is a distance hyperparameter. ,in For the TransE scoring function, This represents the distance between the original triple X and the reconstructed triple Z. A hyperparameter, which is balanced by the hyperparameter λ1. and Its function.

[0023] Further, step S4 includes:

[0024] S41. After training, obtain the confidence scores of the triplet. . The similarity function has a confidence score range of [0,1]. The closer the score is to 0, the more likely it is to be wrong; the closer the score is to 1, the more likely it is to be right.

[0025] S42. Sort the triples from highest to lowest confidence score, and the bottom 5% of triples are incorrect triples.

[0026] use and As an indicator for model evaluation.

[0027] After adopting the above strategy, the positive effects of the present invention are:

[0028] (1) It can mine potential entity label information based on the hypergraph perspective without the need for additional manual annotation;

[0029] (2) The above-mentioned tag information can be used to enhance the error detection of knowledge graphs, providing an effective method for improving medical knowledge graphs. Attached Figure Description

[0030] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0031] Figure 1 This is a framework diagram of a medical knowledge graph error detection method based on implicit label information enhancement according to the present invention;

[0032] Figure 2 This invention relates to a method for error detection in medical knowledge graphs based on implicit label information enhancement. Example diagram of the construction;

[0033] Figure 3 This invention relates to a knowledge graph error detection method for medical knowledge graphs based on implicit label information enhancement. Construct an example graph. Detailed Implementation

[0034] To enable those skilled in the art to better understand the present invention and to make the above-mentioned objectives, technical solutions and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0035] like Figure 1 As shown, the main process of this case consists of four steps: 1. Obtain the medical knowledge graph to be tested. 2. Constructing a knowledge graph based on a hypergraph perspective. Reconstructing the knowledge graph using triples as nodes 3. , , 4. Train the model using the label-based augmentation input. Evaluate using confidence scores.

[0036] S1. Obtain the medical knowledge graph to be tested. Arrange triples in the form of head entity-relation-tail entity. save;

[0037] S2, based on The hypergraph is reconstructed from the hypergraph perspective. , Used to mine entity label information, G1's nodes consist of... and Composition; at the same time, will Using triples as nodes, relationships are established between triple nodes of shared entities, and the knowledge graph is reconstructed. , Used to obtain neighbor information;

[0038] S3, will , , The input is a model based on label information enhancement for training. The model based on label information enhancement consists of a Bi-LSTM and a graph attention network with implicit label information.

[0039] S4. Construct confidence scores and use the confidence score ranking to detect erroneous triples.

[0040] h represents the head entity, and r represents the relation. This indicates the tail entity.

[0041] Even better, since existing knowledge graphs do not have labels for incorrect triples,

[0042] By artificially introducing 5% noisy triples through random head-to-tail entity relationships, the noisy knowledge graph is used as the medical knowledge graph to be tested. Introducing a noisy knowledge graph as input can solve the problem of insufficient negative samples in existing knowledge graphs, and experiments show that the model performs better.

[0043] Further, step S2 includes:

[0044] S21, For all shared... triples Then let At this time, it is called For a set of superedges, Let r be the set of all vertices of the hyperedge; similarly, let t be the set of all triples sharing (r,t). Then let ,say For a set of superedges, The set of all vertices of the hyperedge, obtained by reconstructing the hypergraph from its perspective. ;

[0045] in, , ,… This represents the tail entity, where n is the number. , …, This represents the header entity, where k is the ID. For example... Figure 2 As shown, for (the department, neurology) such a Formal correctness, existence ={Neuro-Lyme disease, syringomyelia, cerebral infarction, Alzheimer's disease}. The head entity in the text contains hidden tag information, such as disease. Similarly, for a condition like (Alzheimer's disease, recommended foods),... Formal correctness, existence ={chestnut, ginkgo}. These two entities share the same implicit tag information: food.

[0046] S22, find all triples in G. As If the node, , Let X1 and X2 be nodes in G2. If h1=h2 or t1=t2 or h1=t2 or h2=t1, then establish the node in G2. and The relationships, after this reconstruction, result in a knowledge graph as follows: .

[0047] like Figure 3 As shown, (Alzheimer's disease, related diseases, Alzheimer's disease) is a triad. (Recommended food for Alzheimer's disease, chestnuts) is a triplet. (Alzheimer's disease, relevant department: neurology) is a triad. .in, and , They share the entities "Alzheimer's disease" and "Alzheimer's disease" respectively, therefore it is believed that... and , There are links between them, that is, relationships or edges in a graph.

[0048] Further, step S3 includes:

[0049] S31. Randomly initialize the triples in G to obtain the vector. ;

[0050] S32, will Inputting into a Bi-LSTM layer yields triple vectors. ;

[0051] S33. Based on the attention mechanism, we will mine the label information in the hypergraph view, and for the obtained triples and its vector representation According to G2 and The importance score of triples from a hypergraph perspective is obtained through an attention mechanism. , ,in, It is a similarity function;

[0052] in, , , ,in For the tail entity t relative to The coefficients of other entities can be obtained similarly. ;

[0053] S34, For a , obtain Neighboring Triple Group and the obtained The input graph is fed into an attention network, which obtains the vector representation Z of the reconstructed triplet based on the neighbor triplet based on the attention mechanism. The specific formula is as follows:

[0054]

[0055] in, Neighboring Triple Group Compared to triples relative coefficient Where A is an attention scoring function that converts a vector into a scalar, and W represents a trainable matrix that maps triple vectors X to the same space;

[0056] S35. Perform negative sampling training using a joint loss function.

[0057] Here, G* represents the negative sample set, which is constructed by randomly replacing the head and tail entity relationships. Let G* be a triple, [ ] + Indicates positive, >0, It is a distance hyperparameter. ,in For the TransE scoring function, This represents the distance between the original triple X and the reconstructed triple Z. A hyperparameter, which is balanced by the hyperparameter λ1. and Its function.

[0058] The graph attention network with implicit label information consists of steps S33, S34, and S35.

[0059] Further, step S4 includes:

[0060] S41. After training, obtain the confidence scores of the triplet. . The similarity function has a confidence score range of [0,1]. The closer the score is to 0, the more likely it is to be wrong; the closer the score is to 1, the more likely it is to be right.

[0061] S42. Sort the triples from highest to lowest confidence score, and the bottom 5% of triples are incorrect triples.

[0062] use and As an evaluation indicator.

[0063] This invention uses the open-source dataset DiseaseKG for testing. DiseaseKG is a medical knowledge graph built on authoritative medical websites, focusing on common diseases, and contains approximately 44,000 entities and 310,000 relationships. The experimental results are shown in Table 1.

[0064] Table 1 Experimental Results

[0065] Precision@K Recall@K TransE 0.331 0.331 Dismult 0.402 0.402 RotatE 0.374 0.374 CAGED <![CDATA[ 0.611 ]]> <![CDATA[ 0.611 ]]> Our 0.648 0.648

[0066] As shown in Table 1, when K=5, Precision@K and Recall@K are exactly equal, and the performance of the present invention is significantly better than other methods.

[0067] Specific embodiments of the present invention have been described above with reference to the accompanying drawings. However, those skilled in the art will understand that various modifications and substitutions can be made to the specific embodiments of the present invention without departing from the spirit and scope of the invention. All such modifications and substitutions fall within the scope defined by the claims of the present invention.

Claims

1. A method for error detection in medical knowledge graphs based on implicit label information enhancement, characterized in that, The specific steps are as follows: S1. Obtain the medical knowledge graph to be tested. Arrange triples in the form of head entity-relation-tail entity. save; S2, based on The hypergraph is reconstructed from the hypergraph perspective. , Used to mine entity label information, G1's nodes consist of... and Composition; at the same time, will Using triples as nodes, relationships are established between triple nodes of shared entities, and the knowledge graph is reconstructed. , Used to obtain neighbor information; S21, For all shared... triples Then let At this time, it is called As a set of superedges, Let r be the set of all vertices of the hyperedge; similarly, let t be the set of all triples sharing (r,t). Then let ,say For a set of superedges, The set of all vertices of the hyperedge, obtained by reconstructing the hypergraph from its perspective. ; S22, find all triples in G. As If the node, , Let X1 and X2 be nodes in G2. If h1=h2 or t1=t2 or h1=t2 or h2=t1, then establish the node in G2. and The relationships, after this reconstruction, result in a knowledge graph as follows: ; S3, will , , The input is a model based on label information enhancement for training. The model based on label information enhancement consists of a Bi-LSTM and a graph attention network with implicit label information. S4. Construct confidence scores and use the confidence score ranking to detect erroneous triples.

2. The method for error detection in medical knowledge graphs based on implicit labeling information enhancement according to claim 1, characterized in that, Step S3 specifically includes: S31. Randomly initialize the triples in G to obtain the vector. ; S32, will Inputting into a Bi-LSTM layer yields triple vectors. ; S33. Based on the attention mechanism, we will mine the label information in the hypergraph view, and for the obtained triples and its vector representation ,according to and The importance score of triples from a hypergraph perspective is obtained through an attention mechanism. , ,in, For similarity functions; in, , , ,in For the tail entity t relative to The coefficients of other entities can be obtained similarly. ; S34, For a , obtain Neighboring Triple Group and the obtained The input graph is fed into an attention network, which obtains the vector representation Z of the reconstructed triplet based on the neighbor triplet based on the attention mechanism. The specific formula is as follows: in, Neighboring Triple Group Compared to triples relative coefficient Where A is an attention scoring function that converts a vector into a scalar, and W represents a trainable matrix that maps triple vectors X to the same space; S35. Perform negative sampling training using a joint loss function. Here, G* represents the negative sample set, which is constructed by randomly replacing the head and tail entity relationships. Let G* be a triple, [ ] + Indicates positive, >0, It is a distance hyperparameter. ,in For the TransE scoring function, This represents the distance between the original triple X and the reconstructed triple Z. A hyperparameter, which is balanced by the hyperparameter λ1. and Its function.

3. The method for error detection in medical knowledge graphs based on implicit labeling information enhancement according to claim 1, characterized in that, Step S4 specifically includes: S41. After training, obtain the confidence scores of the triplet. ; The similarity function has a confidence score range of [0,1]. The closer the score is to 0, the more likely it is to be wrong; the closer the score is to 1, the more likely it is to be right. S42. Sort the triples from highest to lowest confidence score, and the bottom 5% of triples are incorrect triples.

4. The method for error detection in medical knowledge graphs based on implicit labeling information enhancement according to claim 1, characterized in that, Step S1 includes: In the initial medical knowledge graph to be tested, 5% noisy triples are introduced by randomly assigning head-to-tail entity relationships. This noisy knowledge graph is then used as the medical knowledge graph to be tested. .

Images