Entity alignment method and system based on deentanglement graph neural network

Abstract

The invention discloses an entity alignment method and system based on a disentanglement graph neural network, and the method comprises the steps: constructing a relation reflection neighborhood routing mechanism, dividing the neighborhood of an entity into a plurality of parts with different semantics, and learning corresponding representations for different aspects of the entity; mutual information regularization constraint is applied to the representations output in the micro-disentanglement stage so as to reduce the correlation between the representations of the entities in different aspects; aligning perception self-attention is introduced, appropriate weights are distributed to different aspects of the entity in a self-adaptive mode, and the similarity expressed in the multiple aspects is weighted and summed to obtain the final entity similarity; a plurality of seed generators having different aspect values are trained, and entity pair seeds having a particular sparseness are iteratively selected and added to the initial training data. According to the method, disentangled entity representation is learned from the microcosmic level and the macroscopic level, the neighborhood heterogeneity problem of equivalent entities in different knowledge maps is relieved, and the accuracy and robustness of entity alignment are improved.

Description

Technical Field

[0001] The present invention belongs to the field of knowledge graph fusion in natural language processing, and in particular to an entity alignment method and system based on a disentangled graph neural network. Background Art

[0002] Knowledge graphs (KG) are a knowledge representation model that stores facts in the real world in a graphical way. It is a directed graph consisting of nodes (representing entities) and edges (representing the relationship between entities) to capture and organize rich and diverse information. Knowledge graphs are widely used in fields such as natural language processing, machine learning, search engine optimization, and recommendation systems. Its core value lies in integrating cross-source and cross-domain data, and being able to provide semantically rich association information to help understand and reason about information. Entity alignment can automatically discover equivalent entities in different knowledge graphs, and has become a key technology for knowledge graph fusion and promoting knowledge coverage.

[0003] At present, embedding-based methods are the mainstream methods for entity alignment tasks. The key idea is to represent elements in the knowledge graph (such as entities and relations) as low-dimensional vectors (called embeddings) and judge whether the entities are aligned by the distance between the entity embeddings. According to the different embedding models, embedding-based entity alignment can be divided into two categories, namely translation-based entity alignment and graph neural network-based entity alignment. The first type of methods uses TransE and its variants to model the structure of the knowledge graph. They regard the relationship vector as a translation vector from its head entity to its tail entity. The second type of methods uses graph neural networks (GNNs) to model the knowledge graph. They can encode neighborhood information into entity embeddings in order to capture the similarity of equivalent entities. Because graph neural networks have a natural advantage in modeling graph structured data, under the same conditions, the performance of graph neural network-based methods is usually better than that of translation-based methods.

[0004] However, most of the previous GNN-based methods are based on an important assumption: equivalent entities in different knowledge graphs have similar neighborhood structures, and after GNN encoding, they will produce similar embeddings. In practice, different knowledge graphs are built by different institutions using different data sources for different purposes. They usually focus on different aspects of entities, some of which are consistent, while others are different. This leads to serious heterogeneity in the neighborhoods of some equivalent entities. Statistics show that in the benchmark dataset DBP15K for entity alignment, most of the aligned entity pairs have different neighboring entities. Heterogeneous neighborhood structures induce GNNs to produce different representations, even for equivalent entities. Although some previous studies have attempted to alleviate the neighborhood heterogeneity between equivalent entities, they are still constrained by the inherent limitations of traditional GNNs. They treat the neighborhood of an entity as a whole and aggregate all neighbor features into a single, static representation. This encoding strategy cannot clearly distinguish the contributions of different aspects of an entity, and inevitably introduces noise, leading to incorrect alignment results. In addition, some methods use semi-supervised learning to iteratively select new alignment seeds to expand the training data, which may produce erroneous seeds and lead to further error propagation. Summary of the invention

[0005] Purpose of the invention: In view of the problems existing in the above-mentioned existing methods, the present invention provides an entity alignment method and system based on a disentangled graph neural network, which effectively alleviates the neighborhood heterogeneity problem of equivalent entities in different knowledge graphs, thereby improving the accuracy and robustness of entity alignment.

[0006] Technical solution: The entity alignment method based on a disentangled graph neural network described in the present invention specifically includes the following steps:

[0007] (1) Construct a relation-reflective neighborhood routing mechanism to divide the neighborhood of an entity into multiple semantically distinct parts and learn corresponding representations for different aspects of the entity;

[0008] (2) Apply mutual information regularization constraints to the representation output by the micro-disentanglement stage to reduce the correlation between the representations of different aspects of the entity;

[0009] (3) Construct alignment-aware self-attention to adaptively assign appropriate weights to different aspects of an entity and weightedly sum the similarities represented by multiple aspects to obtain the final entity similarity;

[0010] (4) In a semi-supervised learning framework, multiple seed generators with different aspect values ​​are jointly trained, and entity pair seeds with specific sparsity are iteratively selected and added to the initial training data.

[0011] Furthermore, the implementation process of step (1) is as follows:

[0012] Use K independent embeddings to jointly represent entities and relations:

[0013]

[0014] Among them, K represents the total number of aspects of an entity or relationship; It is used to describe the entity e i The disentangled representation of the k-th aspect feature, It is used to describe the relationship m The disentangled representation of the k-th aspect feature; d is the dimension of the representation of each aspect;

[0015] For the relationship m The following relational reflection transformation matrix is ​​established to project the entity embedding into the subspace associated with the kth aspect:

[0016]

[0017] Where I is an identity matrix; when embedded Normalized to hour, is an orthogonal matrix;

[0018] Given a central entity e i and all containing e i The following Q rounds of clustering are iteratively performed to obtain the triples of e in different aspects. i Attention weights for neighbor entities:

[0019]

[0020] in, is the center of a subspace cluster, is a weight coefficient, indicating that entity e j To what extent can we gather information about entity e i The information of the kth aspect; q is the current iteration number, for e i The set of adjacent entity-relationship pairs, and It is entity e i and e j The hidden representation of the lth layer in the kth aspect; τ is a hyperparameter; the L1 distance of the vector is used when calculating the similarity; at the beginning of the first clustering, Initialized to 1 / K;

[0021] After Q rounds of clustering are completed, the embeddings of neighbor entities in different aspects are aggregated to obtain the disentangled entity representation:

[0022]

[0023] in, It is entity e i The new representation of the k-th aspect is used as the output of the l+1-th layer;

[0024] The representations of different layers in the disentangled graph neural network contain different levels of semantic information. The entity embeddings output by different layers are concatenated to generate the following global entity representation:

[0025]

[0026] Furthermore, the implementation process of step (2) is as follows:

[0027] A mutual information regularization constraint is imposed on the global representation of the entity. Specifically, a contrastive estimator is used to calculate the upper bound of the mutual information. A simple neural network with parameter θ is used to construct a variational distribution To approximate the true conditional distribution, the upper bound of the mutual information is calculated as follows:

[0028]

[0029] in, and Respectively represent the log-likelihood of positive sample pairs and negative sample pairs, and minimizing the upper bound can achieve the purpose of minimizing mutual information;

[0030] Minimize the KL divergence between the two distributions to ensure that the calculation result of the above formula is the upper bound of the mutual information; therefore, the final optimization goal of the mutual information regularization constraint is as follows:

[0031]

[0032] in, is the true conditional probability distribution between various aspects. The regularization loss and the entity alignment loss will be optimized alternately during training.

[0033] Furthermore, the implementation process of step (3) is as follows:

[0034] Calculate the distance of the entity in each aspect. The distance of the entity pair (e1, e2) in the kth aspect is defined as follows:

[0035]

[0036] An alignment-aware self-attention strategy is proposed to adaptively assign appropriate weights to different aspects; the distances within multiple aspects are weighted summed to obtain the final distance of (e1, e2):

[0037]

[0038] Among them, b is a trainable attention weight vector;

[0039] For the entity alignment problem, the following margin-based entity alignment objective function is defined to make the embeddings of aligned entity pairs close to each other and the embeddings of misaligned entity pairs far away from each other:

[0040]

[0041] Among them, P is the set of positive samples, that is, some pre-aligned entity pairs; P′ is the set of negative samples, and γ>0 is a marginal hyperparameter used to separate positive samples and negative samples.

[0042] Furthermore, the mutual information regularization L reg Align with entity L ea The loss is optimized in an alternating manner throughout the training process.

[0043] Furthermore, the implementation process of step (4) is as follows:

[0044] A semi-supervised learning framework with multi-aspect seed generators is constructed. The seed generators use a bidirectional iterative algorithm to identify new alignment seeds with specific sparsity according to their own aspect values; specifically, the sparsity of the following entities is defined:

[0045]

[0046] Among them, e* is the entity with the most neighbors in the knowledge graph; all entities are sorted in descending order of sparsity and divided into M equal parts; if entity e i In the mth part, its sparsity ranking is:

[0047]

[0048] Accordingly, M seed generators are trained and sorted in ascending order according to their respective aspect values; the mth seed generator can only select entity pairs with sparsity ranking equal to m, and seeds of other sparsities will be filtered out; the outputs of all seed generators are merged and added to the initial training set.

[0049] The entity alignment system based on a disentangled graph neural network described in the present invention comprises:

[0050] The micro-disentangled representation module designs a relation-reflective neighborhood routing mechanism, divides the neighborhood of an entity into multiple semantically distinct parts, and learns corresponding representations for different aspects of the entity;

[0051] The macro disentanglement constraint module imposes mutual information regularization constraints on the representations output by the micro disentanglement stage to reduce the correlation between representations of different aspects of the entity;

[0052] The multi-aspect entity similarity calculation module builds alignment-aware self-attention, adaptively assigns appropriate weights to different aspects of the entity, and weighted sums the similarities represented by multiple aspects to obtain the final entity similarity;

[0053] The semi-supervised learning module jointly trains multiple seed generators with different aspect values, iteratively selects entity pair seeds with specific sparsity, and adds them to the initial training data.

[0054] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention utilizes the relation-reflective neighborhood routing mechanism to learn disentangled entity representations from both micro and macro levels, thereby alleviating the neighborhood heterogeneity phenomenon that is prevalent in entity alignment; 2. The present invention designs an alignment-aware self-attention strategy, which can adaptively assign appropriate weights to different aspects of entities when calculating the similarity of entity pairs; 3. The present invention constructs a semi-supervised learning framework to jointly train multiple seed generators with different aspect values, which can iteratively select new entity pair seeds with specific sparsity. In this way, the data enhancement process can be refined to further improve the accuracy of entity alignment. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flowchart of the entity alignment method based on the disentangled graph neural network;

[0056] Figure 2 It is a framework diagram of the entity alignment system based on the disentangled graph neural network. DETAILED DESCRIPTION

[0057] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0058] like Figure 1 As shown, the present invention provides an entity alignment method based on a disentangled graph neural network, which specifically includes the following steps:

[0059] Step 1: Design a relation-reflective neighborhood routing mechanism to divide the neighborhood of an entity into multiple semantically distinct parts and learn corresponding representations for different aspects of the entity.

[0060] Unlike previous entity alignment methods, which all use a complete vector to represent entities and relations in the knowledge graph, this paper uses K independent embeddings to jointly represent entities and relations:

[0061]

[0062] Among them, K represents the total number of aspects of an entity or relationship. It is used to describe the entity e i The disentangled representation of the k-th aspect feature, It is used to describe the relationship m The disentangled representation of the k-th aspect feature. d is the dimension of the representation of each aspect.

[0063] For the relationship m The following relational reflection transformation matrix is ​​established to project the entity embedding into the subspace associated with the kth aspect:

[0064]

[0065] Where I is an identity matrix. When embedded Normalized to hour, is an orthogonal matrix.

[0066] Given a central entity e i and all containing e i The following Q rounds of clustering are iteratively performed to obtain the triples of e in different aspects. i Attention weights for neighbor entities:

[0067]

[0068] Among them, q is the current iteration number, for e i The set of adjacent entity-relationship pairs, and It is entity e i and e j The hidden representation of layer 1 in the k-th aspect. τ is a hyperparameter. can be viewed as the center of a subspace cluster. is a weight coefficient, which represents the entity e j To what extent can we gather information about entity e i The kth aspect of information. When calculating the similarity, we used the L1 distance of the vector. At the beginning of the first clustering, Initialized to 1 / K.

[0069] After Q rounds of clustering are completed, the embeddings of neighbor entities in different aspects are aggregated to obtain the disentangled entity representation:

[0070]

[0071] in, It is entity e iThe new representation of the k-th aspect can be used as the input of the l+1-th layer.

[0072] Because representations at different layers in the disentangled graph neural network contain different levels of semantic information, we concatenate entity embeddings output from different layers to generate the following global entity representation:

[0073]

[0074] Step 2: A mutual information regularization constraint is imposed on the representation output by the micro-disentanglement stage to reduce the correlation between the representations of different aspects of the entity.

[0075] Step 1 only completes the disentangled representation of the entity at the micro level, but micro disentanglement alone is not enough. It is also necessary to make the representations of different aspects of the same entity separable at the macro level. The present invention chooses to impose mutual information regularization constraints on the global representation of the entity, because mutual information can measure the linear and nonlinear correlations between random variables, which is suitable for complex multi-relationship scenarios in knowledge graphs.

[0076] Specifically, a contrastive estimator is used to calculate the upper bound of the mutual information. Since the true conditional probability distribution between each aspect is unknown, and a simple neural network with parameters θ is used to construct a variational distribution To approximate the true conditional distribution, the upper bound of the mutual information is calculated as follows:

[0077]

[0078] in, and Represent the log-likelihood of positive sample pairs and negative sample pairs respectively. Minimizing this upper bound can achieve the goal of minimizing mutual information.

[0079] In addition, the KL divergence between the two distributions must be minimized to ensure that the calculation result of the above formula is the upper bound of the mutual information. Therefore, the final optimization goal of the mutual information regularization constraint in step 2 is as follows:

[0080]

[0081] The regularization loss and entity alignment loss will be optimized alternately during training.

[0082] Step 3: Design an alignment-aware self-attention that can adaptively assign appropriate weights to different aspects of an entity and weighted sum the similarities represented by multiple aspects to obtain the final entity similarity.

[0083] In the present invention, the representation of an entity consists of the embedding of multiple disentangled aspects, and the contributions of different aspects are different. Therefore, it is necessary to first calculate the distance of the entity in each aspect. For an entity pair (e1, e2), its distance in the kth aspect is defined as follows:

[0084]

[0085] We propose an alignment-aware self-attention strategy that can adaptively assign appropriate weights to different aspects. By weighted summing the distances within multiple aspects, we can get the final distance of (e1, e2):

[0086]

[0087] Where b is a trainable attention weight vector.

[0088] During training, it is expected that the embeddings of aligned entity pairs are close to each other, while the embeddings of misaligned entity pairs are far away from each other. Therefore, for the entity alignment problem, the following margin-based objective function is defined:

[0089]

[0090] Where P is the positive sample set, i.e., some pre-aligned entity pairs; P′ is the negative sample set, which randomly selects other entities to replace the entities in the positive sample pairs to obtain negative samples; γ>0 is a marginal hyperparameter used to separate positive samples from negative samples. Mutual Information Regularization (L reg ) and entity alignment (L ea ) loss is optimized in an alternating manner throughout the training process.

[0091] Step 4: In a semi-supervised learning framework, multiple seed generators with different aspect values ​​are jointly trained, which can iteratively select entity pair seeds with specific sparsity and add them to the initial training data.

[0092] The lack of sufficient training data limits the performance of embedding-based entity alignment methods. To this end, existing methods propose a semi-supervised learning strategy called bidirectional iteration: if the entity e in two knowledge graphs i and e j The embeddings of are closest to each other, then (e i , e j) will be identified as potential alignment seeds and added to the training dataset. However, this may produce erroneous seeds and lead to further error propagation. In addition, the information content of different entities in the knowledge graph is different. It is also inappropriate for the present invention to use a single value (i.e., K) to determine the number of aspects of all entities. Intuitively, an entity node with a dense neighborhood requires a larger K value, while an entity node with a sparse neighborhood requires a smaller K value.

[0093] Considering the above two points, a semi-supervised learning framework with multiple aspect seed generators is designed. These seed generators have different aspect values, and they each use a bidirectional iterative algorithm to identify new alignment seeds with specific sparsity according to their own aspect values. This strategy helps to generate high-quality alignment seeds and enhance entity alignment results. Specifically, the sparsity of the following entities is defined:

[0094]

[0095] Among them, e * is the entity with the most neighbors in the knowledge graph. All entities are sorted in descending order of sparsity and divided into M equal parts. If entity e i In the mth part, its sparsity ranking is:

[0096]

[0097] Accordingly, M seed generators are trained and sorted in ascending order according to their respective aspect values. The mth seed generator can only select entity pairs with sparsity ranking (srank) equal to m, and seeds with other sparsities will be filtered out. The outputs of all seed generators are merged and added to the initial training set.

[0098] The present invention also provides a medical entity alignment system based on a disentangled graph neural network, such as Figure 2 As shown in the figure, two commonly used medical knowledge graphs, CMeKG and PrimeKG, are aligned, including:

[0099] The micro-disentangled representation module includes a micro-disentangled encoder. By designing a relation-reflective neighborhood routing mechanism, the neighborhood of an entity is automatically divided into multiple semantically different parts, and corresponding representations are learned for different aspects of the entity.

[0100] The macro disentanglement constraint module imposes a mutual information regularization constraint on the global representation output by the micro disentanglement stage to reduce the correlation between the representations of different aspects of an entity.

[0101] The multi-aspect entity similarity calculation module adaptively assigns appropriate weights to different aspects of the entity and weighted sums the similarities represented by multiple aspects to obtain the final entity similarity.

[0102] Semi-supervised learning module, in a semi-supervised learning framework, trains multiple seed generators with different aspect values ​​simultaneously; the generators iteratively select new aligned medical entity pairs with specific sparsity. In this way, the data augmentation process can be refined to produce high-quality alignment seeds, thereby alleviating the problem of insufficient training data. Therefore, we only need to manually annotate a small number of aligned entity pairs to achieve satisfactory alignment results.

[0103] In this embodiment, the present invention is used to align two commonly used medical knowledge graphs, CMeKG and PrimeKG. CMeKG is the most well-known Chinese medical knowledge graph, which is constructed with reference to authoritative international medical standard terminology sets such as ICD-10, ATC, MeSH, and large-scale, multi-source heterogeneous clinical pathway guidelines, clinical practices, medical encyclopedias and other resources, covering examples of medical concept relationships such as diseases, drugs, and diagnosis and treatment technologies. PrimeKG is an English knowledge graph for precision medicine, which provides a holistic view of the disease by integrating up to 20 high-quality biomedical databases. First, a certain number of aligned entity pairs are manually marked in these two medical knowledge graphs as training sets. Figure 2 The double-arrow dashed line connecting entities in different knowledge graphs represents the alignment relationship between entities. Then, these two knowledge graphs and a certain number of pre-aligned entity pairs are input into the medical field entity alignment system based on disentangled graph neural network, and more aligned entity pairs between the two graphs can be automatically obtained.

[0104] There are many methods and approaches to implement the technical solution of the present invention, and the above are only preferred embodiments 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, and these improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. An entity alignment method based on a disentangled graph neural network, characterized in that: The following steps are involved: (1) Construct a relation-reflective neighborhood routing mechanism to divide the neighborhood of an entity into multiple semantically distinct parts and learn corresponding representations for different aspects of the entity; (2) Apply mutual information regularization constraints to the representation output by the micro-disentanglement stage to reduce the correlation between the representations of different aspects of the entity; (3) Construct alignment-aware self-attention to adaptively assign appropriate weights to different aspects of an entity and weightedly sum the similarities represented by multiple aspects to obtain the final entity similarity; (4) In a semi-supervised learning framework, multiple seed generators with different aspect values ​​are jointly trained, and entity pair seeds with specific sparsity are iteratively selected and added to the initial training data.

2. The entity alignment method based on disentangled graph neural network according to claim 1, characterized in that: The implementation process of step (1) is as follows: Use K independent embeddings to jointly represent entities and relations: Among them, K represents the total number of aspects of an entity or relationship; It is used to describe the entity e i The disentangled representation of the k-th aspect feature, It is used to describe the relationship m The disentangled representation of the k-th aspect feature; d is the dimension of the representation of each aspect; For the relationship m The following relational reflection transformation matrix is ​​established to project the entity embedding into the subspace associated with the kth aspect: Where I is an identity matrix; when embedded Normalized to hour, is an orthogonal matrix; Given a central entity e i and all containing e i The following Q rounds of clustering are iteratively performed to obtain the triples of e in different aspects. i Attention weights for neighbor entities: in, is the center of a subspace cluster, is a weight coefficient, indicating that entity e j To what extent can we gather information about entity e i The information of the kth aspect; q is the current iteration number, for e i The set of adjacent entity-relationship pairs, and It is entity e i and e j The hidden representation of the lth layer in the kth aspect; τ is a hyperparameter; the L1 distance of the vector is used when calculating the similarity; at the beginning of the first clustering, Initialized to 1 / K; After Q rounds of clustering are completed, the embeddings of neighbor entities in different aspects are aggregated to obtain the disentangled entity representation: in, It is entity e i The new representation of the k-th aspect is used as the output of the l+1-th layer; The representations of different layers in the disentangled graph neural network contain different levels of semantic information. The entity embeddings output by different layers are concatenated to generate the following global entity representation:

3. The entity alignment method based on disentangled graph neural network according to claim 1, characterized in that: The implementation process of step (2) is as follows: A mutual information regularization constraint is imposed on the global representation of the entity. Specifically, a contrastive estimator is used to calculate the upper bound of the mutual information. A simple neural network with parameter θ is used to construct a variational distribution To approximate the true conditional distribution, the upper bound of the mutual information is calculated as follows: in, and Respectively represent the log-likelihood of positive sample pairs and negative sample pairs, and minimizing the upper bound can achieve the purpose of minimizing mutual information; Minimize the KL divergence between the two distributions to ensure that the calculation result of the above formula is the upper bound of the mutual information; therefore, the final optimization goal of the mutual information regularization constraint is as follows: in, is the true conditional probability distribution between various aspects. The regularization loss and the entity alignment loss will be optimized alternately during training.

4. The entity alignment method based on disentangled graph neural network according to claim 1, characterized in that: The implementation process of step (3) is as follows: Calculate the distance of the entity in each aspect. The distance of the entity pair (e1, e2) in the kth aspect is defined as follows: An alignment-aware self-attention strategy is proposed to adaptively assign appropriate weights to different aspects; the distances within multiple aspects are weighted summed to obtain the final distance of (e1, e2): Among them, b is a trainable attention weight vector; For the entity alignment problem, the following margin-based entity alignment objective function is defined to make the embeddings of aligned entity pairs close to each other and the embeddings of misaligned entity pairs far away from each other: Among them, P is the set of positive samples, that is, some pre-aligned entity pairs; P' is the set of negative samples, and γ>0 is a marginal hyperparameter used to separate positive samples and negative samples.

5. The entity alignment method based on disentangled graph neural network according to claims 3 and 4, characterized in that: The mutual information regularization L reg Align with entity L ea The loss is optimized in an alternating manner throughout the training process.

6. The entity alignment method based on disentangled graph neural network according to claim 1, characterized in that: The implementation process of step (4) is as follows: A semi-supervised learning framework with multi-aspect seed generators is constructed. The seed generators use a bidirectional iterative algorithm to identify new alignment seeds with specific sparsity according to their own aspect values; specifically, the sparsity of the following entities is defined: Among them, e * is the entity with the most neighbors in the knowledge graph; all entities are sorted in descending order of sparsity and divided into M equal parts; if entity e i In the mth part, its sparsity ranking is: Accordingly, M seed generators are trained and sorted in ascending order according to their respective aspect values; the mth seed generator can only select entity pairs with sparsity ranking equal to m, and seeds of other sparsities will be filtered out; the outputs of all seed generators are merged and added to the initial training set.

7. An entity alignment system based on a disentangled graph neural network as described in claims 1 to 6, characterized in that: include: The micro-disentangled representation module designs a relation-reflective neighborhood routing mechanism, divides the neighborhood of an entity into multiple semantically distinct parts, and learns corresponding representations for different aspects of the entity; The macro disentanglement constraint module imposes mutual information regularization constraints on the representations output by the micro disentanglement stage to reduce the correlation between representations of different aspects of the entity; The multi-aspect entity similarity calculation module builds alignment-aware self-attention, adaptively assigns appropriate weights to different aspects of the entity, and weighted sums the similarities represented by multiple aspects to obtain the final entity similarity; The semi-supervised learning module jointly trains multiple seed generators with different aspect values, iteratively selects entity pair seeds with specific sparsity, and adds them to the initial training data.

Images