Knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework

Abstract

The application discloses a knowledge graph embedding method based on a global information semantic reconstruction and multi-specification feature sharpening framework, and belongs to the technical field of knowledge graph embedding.The application aims at the problem that the existing knowledge graph embedding method has poor overall attention to feature maps, which influences the prediction result.The application comprises an entity relationship semantic reconstruction method based on global information capture, global information reconstruction is performed on a semantic information matrix composed of entities and relationships through multi-layer dilated convolution, semantic information of the entities and the relationships with context logical association information is acquired, then a multi-specification feature capture mechanism oriented to the semantic reconstruction information is used, different-specification convolution kernels are extracted from relationship vectors and are convolved with the reconstruction information, feature capture is performed on the entity relationship semantic information from different angles, finally, an information sharpening method based on group normalization is used to standardize and adjust a threshold value of a multi-specification feature set.The application improves the accuracy and robustness of entity and relationship representation.

Description

TECHNICAL FIELD

[0001] The application relates to a knowledge graph embedding method based on a global information semantic reconstruction and multi-specification feature sharpening framework, and belongs to the technical field of knowledge graph embedding. BACKGROUND

[0002] A knowledge graph (KG) is a knowledge representation method that represents entities and their relationships in the real world through a graph structure and is applied in search engines, question-answering systems and recommendation systems. Existing knowledge graph databases, such as WordNet, Freebase and YAGO, all have a common problem: as information grows exponentially, it becomes increasingly challenging to maintain an up-to-date knowledge graph, resulting in incomplete knowledge graphs in most real-world applications. The incompleteness of the graph will affect the above-mentioned application methods, leading to inaccurate predictions and incorrect recommendations. Therefore, more and more research is focused on knowledge graph completion (KGC). Link prediction (LP) is an important method for knowledge graph completion, which uses known facts in the knowledge graph to infer missing triple information. For example, Figure 1 An example of a simple incomplete knowledge graph is shown, where the text represents existing facts and? represents information that must be inferred. For example, we need to predict the missing entity in (Zhang San, born,?) or the missing relationship in (Li Si,?, red rice).

[0003] Knowledge graph embedding (KGE) is a widely used link prediction method that converts triple information in a knowledge graph into a continuous vector space, preserving the logical relationship information in the knowledge graph, and then uses a certain model to infer missing triple information (facts). The advantage of using a knowledge graph embedding model is that it can better utilize the structural and semantic information in the knowledge graph and achieve semantic vectorization of the elements of the knowledge graph. Generally, knowledge graph embedding models can be divided into three categories: translation-based embedding models, semantic matching-based embedding models and neural network-based embedding models.

[0004] Translation-based embedding models are an important class of methods in the field of knowledge graph embedding. The core idea of this class of models is to treat the relationship as a translation from the head entity to the tail entity, and then evaluate the validity of the candidate triplets by defining a distance-based scoring function. The most classic translation-based embedding model is TransE, which maps entities and relationships to a low-dimensional vector space, thereby realizing the representation of triples (entity-relation-entity). This simple assumption can achieve good results in processing large-scale simple relationship knowledge graphs, but it often performs poorly in capturing complex relationships and semantic information. To solve this problem, TransH introduces a relationship hyperplane to distinguish different types of relationships based on TransE, allowing the semantic representation of entities to be adjusted according to the relationship, enhancing the ability to express complex relationships such as one-to-many and many-to-one. TransR further separates the entity space and the relationship space based on TransH, introducing a mapping matrix for each relationship, projecting entities from the entity space to the relationship-specific space for operation, and adapting to complex and diverse relationship types such as many-to-many. TransC addresses the problem of parameter redundancy and poor performance on sparse data by introducing the concept of clustering, clustering relationships with similar semantics to further enhance the model's representation and generalization capabilities. Similar types of KGE models include TransD, TransG, TransA, TranSparse, and others.

[0005] Semantic matching-based embedding models are another important method in the field of knowledge graph embedding. This class of models measures the similarity between entities and relationships by designing appropriate similarity functions, thereby capturing semantic information in knowledge graphs. For example, RESCAL treats relationships as a second-order tensor through rank-three decomposition, then represents entities and relationships as the product of matrices and vectors, thereby capturing the representation and prediction between entities and relationships. DistMult calculates the score of a triplet through dot product operation, thereby predicting entity embeddings in a knowledge graph. ComplEx represents entities and relationships using the real and imaginary parts of complex numbers, and uses dot product to calculate the similarity between entities and relationships. HolE combines the advantages of RESCAL and DisMult, utilizing the complementary advantages of both models. SimplE learns two independent embedding vectors for each entity, which can incorporate certain types of background knowledge into these embeddings through weight binding. RotatE embeds entities and relationships into complex number space, and operates on entities in the complex plane to represent triples in a knowledge graph. Compared to translation-based embedding models that use addition operations, semantic matching-based embedding models typically rely on similarity measures such as dot product, cosine similarity, or bilinear functions. Due to the need to calculate complex similarity functions, the computational complexity is relatively high, resulting in better performance than translation distance models, but their computational complexity is generally higher.

[0006] Neural network-based embedding models have wide applications in the field of knowledge graph entity embedding. By designing complex network structures and feature extraction methods, they can capture complex relationships and semantic information in knowledge graphs, improving the accuracy and predictive power of embedding representations. For example, NTN uses a neural network layer and a tensor layer to map entities and relationships into a high-dimensional space, thereby learning representations of entities and relationships; SLM uses the nonlinear operation of a single fully connected layer to implicitly link entity and relationship embeddings. Currently, due to the powerful ability of convolutional networks to extract semantic information implicit between entities and relationships, they have been widely applied in knowledge embedding. Figure 2 Figure (a) shows a general flowchart of the embedding model based on convolutional networks. First, vectors h and r are concatenated and reconstructed into a two-dimensional image; then, the newly generated two-dimensional image is input into a module based on a convolutional neural network to obtain a new feature map as output; finally, the obtained output feature map is used as a score corresponding to the candidate triples, where the score of the correct triples is higher than the score of the incorrect triples. The most representative example is ConvE. Specifically, ConvE reconstructs entity and relation embeddings into two-dimensional matrices and uses filters to extract higher-level non-linear features, thereby capturing complex interactions between entities and relations. CoveKB uses a one-dimensional convolutional neural network as the feature extraction network, taking the embeddings of head entities, tail entities, and relations as input, thus capturing global relationships and transitional features between entities and relations in the knowledge base. ConvR adaptively constructs convolutional filters from relation representations, uses a two-dimensional convolutional neural network as the feature extraction network, and applies these filters to entity representations to generate convolutional features. JointE uses a combination of one-dimensional and two-dimensional convolutional neural networks to improve the model's embedding performance, where the one-dimensional convolutional neural network is used to extract surface and explicit knowledge, and the two-dimensional convolutional neural network is used to facilitate interactions between entities and relations. M-DCN, based on ConvE, uses a method of stacking h and r at intervals during the reconstruction process to increase the interaction between entities and relations. InteractE uses three integer types (i.e., stack, alternation, and chessboard) to extract convolutional kernels from entities as input to a two-dimensional convolutional neural network to increase the degree of interaction between entities and relations. Figure 2 As shown in (b). In addition, some advanced graph learning methods have also received widespread attention. R-GCN extends graph convolutional neural networks to process knowledge graphs, and can efficiently learn node representations in graphs with various relation types. CLGAT-KGC introduces a graph attention mechanism, which adds different entity representations to different relations under the same entity, thereby enhancing the entity-relation interaction function.

[0007] The three knowledge graph embedding methods described above have advantages, the embedding model based on translation and the embedding model based on semantic matching utilize explicit relationship modeling methods such as addition and multiplication, embed entities and relationships in a knowledge graph into a low-dimensional vector space to extract shallow semantic information, and can quickly obtain the final score. However, the simple structure of these two methods cannot exhibit good performance in obtaining deep knowledge. The embedding model based on neural network can learn both shallow semantic information and deep knowledge, so most methods at the present stage focus on the use of neural network methods. The model based on full connection in the embedding model based on neural network usually focuses on linear combination of layer-by-layer information, and lacks modeling ability of multi-scale context. The semantic information of entities and relationships in the knowledge graph often involves different scales and levels, and needs to capture the interaction of global and local features. At the same time, the model based on graph neural network learns features by aggregating neighbors, which may ignore the inherent directionality of the knowledge graph. In recent years, neural networks based on convolution kernels have gradually become the mainstream model for knowledge graph embedding tasks, which use convolution kernels to extract interactive features of entities and relationships and have achieved good performance. At present, the mainstream convolutional neural networks applied to the field of knowledge graph prediction are: ConvE, ConvR, JointE and InteractE. However, these methods focus on local information in the feature map, and the attention to the entire feature map is often insufficient. In fact, the features in the distant positions of the feature map have actual influence on the final prediction result, and the combination of these distant information and the current information plays an important role in the model understanding the semantic information of entities and relationships. SUMMARY

[0008] In view of the problem that the existing knowledge graph embedding method has poor overall attention to the feature map and affects the prediction result, the application provides a knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework.

[0009] The knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework provided by the application comprises,

[0010] Randomly initializing the triples in the knowledge graph to obtain D-dimensional triple embeddings, corresponding splicing the head entity embedding and the relationship embedding in the D-dimensional triple embedding by dimension, and reshaping to obtain initial features Figure X 0;

[0011] adopting a semantic information reconstruction module based on global information capture to extract features from the initial features Figure X 0, combining the initial features Figure X 0 to perform residual processing to obtain residual features Figure X 1;

[0012] Three relational convolution kernels of different sizes were selected to process the residual features respectively. Figure X 1. Feature extraction is performed to obtain three sets of feature mapping maps.

[0013] Then, a feature sharpening module based on group normalization is used to sharpen the feature map. Perform feature sharpening to obtain the sharpened feature map.

[0014] Three sharpened feature maps The hidden layer vectors are flattened and connected along the channel dimension, and a fully connected layer is used to obtain the hidden layer vectors. The dimensions of the hidden layer vectors are the same as those of the tail entity embeddings in the triple embeddings. Then, the hidden layer vectors and the tail entity embeddings in the triple embeddings are multiplied by a matrix to predict multiple scores. The tail entity embedding corresponding to the highest score is taken as the tail entity prediction result to achieve information reconstruction.

[0015] According to the knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework of the present invention, the semantic information reconstruction module based on global information capture sequentially includes depthwise convolution (DW-Conv), two depthwise dilated convolutions (DW-DConv) with different dilation rates, and a regular convolution (Conv); initial features Figure X 0 is sequentially input into the depthwise convolution DW-Conv and the depthwise dilated convolution DW-DConv, and the feature extraction result is output by the ordinary convolution Conv.

[0016] According to the knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework of the present invention, the residual features Figure X 1 is represented as:

[0017] X1 = A f +X0,

[0018] In the formula A f Feature maps output by the semantic information reconstruction module based on globally captured information:

[0019]

[0020] In the formula, Conv1 is a regular convolution Conv with a kernel size of 1×1, and DW-Conv3 is a depthwise convolution DW-Conv with a kernel size of 3×3. This is a depthwise dilated convolution (DW-Dconv) with a kernel size of 3×3 and a dilation rate of 3. It is a depth dilated convolution DW-Dconv with a kernel size of 3×3 and a dilation rate of 9.

[0021] According to the knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework of the present invention, the three relation convolution kernels of different sizes are respectively selected as 1×2. 2×2 and 3×1

[0022] According to the knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework of the present invention, three sets of feature mapping graphs for:

[0023]

[0024] In the formula Let C be the set of real numbers, C be the number of channels, and H be the number of channels. The height, W is The width.

[0025] According to the knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework of the present invention, the feature sharpening module based on group normalization sharpens the feature mapping graph. The process of feature sharpening includes processing the feature map. Standardization, feature segmentation, feature enhancement and weakening, and recombination are performed.

[0026] According to the knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework of the present invention, the standardization method is to evaluate the importance of the feature map using the scaling factor in the group normalization layer:

[0027]

[0028] In the formula To calculate the normalization result obtained from group normalization, μ i for The mean, σ i for standard deviation, ε i for Adjustment parameters, λ i and β i Two training adjustment parameters for the group normalization layer GN;

[0029] Based on normalization results Determine the training adjustment parameter λ i Then, based on training, adjust the parameter λ. i Calculate normalized weights for:

[0030]

[0031] In the formula The normalized weight for the kth channel, For λ i The training adjustment parameter corresponding to the kth channel, For λ i The training adjustment parameter corresponding to the jth channel;

[0032] Feature segmentation is performed again: the normalized weight Weighted feature map The adjusted weight K is calculated by mapping to the range (0, 1) through the sigmoid function i :

[0033]

[0034] In the formula, GN represents the group normalization processing of to obtain the training adjustment parameter λ i ; W() represents the calculation of the normalized weight Sigmoid() is a scaling function that scales the value to the range (0, 1); Threshold() is a threshold adjustable function that sets the value greater than the set weight threshold to 1 and the value less than or equal to the set weight threshold to 0;

[0035] According to the adjusted weight K i , the adjusted weight K i greater than the set weight threshold is set to 1 to obtain the weight K 1i , and the adjusted weight K i less than or equal to the set weight threshold is set to 0 to obtain the weight K 2i .

[0036] According to the knowledge graph embedding method based on the global information semantic reconstruction and multi-specification feature sharpening framework of the present application, the weighted features 1i and the weighted features 2i are obtained by feature enhancement and weakening based on the weights K and the weights K

[0037] The weighted features and the weighted features are recombined to obtain the sharpened feature map

[0038]

[0039] According to the knowledge graph embedding method based on the global information semantic reconstruction and multi-specification feature sharpening framework of the present application, the method for calculating the score is:

[0040]

[0041] wherein Scores(h,r,t) represents a score function of D-dimensional triple embedding, wherein h represents a head entity embedding, r represents a relation embedding, t represents a tail entity embedding, f(·) represents projecting the hidden layer vector into a space with the same dimension as the tail entity embedding, vec(·) represents flattening, GSR(·) represents a global feature strengthening operation, MSC(·) represents a neighbor feature strengthening operation, GNS(·) represents a feature reconstruction operation, R() represents a reshaping operation, W0 is a randomly initialized parameter matrix.

[0042] According to the knowledge graph embedding method based on the global information semantic reconstruction and multi-specification feature sharpening framework of the application, the calculation of the score function Scores(h,r,t) is realized by selecting 1-N scoring; the cross-entropy loss function ψ(p,y) is calculated according to the score value, and the adjustment parameters ε i , the training adjustment parameters λ i and β i ; the cross-entropy loss function ψ(p,y) is:

[0043]

[0044] wherein p represents a predicted triple embedding score, y represents a binary label vector, N is the total number of triples, p n is the nth predicted triple embedding score, y n is the binary label vector corresponding to the nth prediction.

[0045] The method of the application is based on the multi-specification feature sharpening framework-GIRMSF for global information reconstruction, which firstly reconstructs the global information of the matrix composed of entities and relations through multi-layer dilated convolution, fuses the context information from a larger receptive field to reconstruct the semantic information of entities and relations; secondly, a multi-specification feature capturing mechanism for reconstructed information is proposed, which extracts different specification convolution kernels from the relation vector to perform convolution operation on the reconstructed information, extracts different specification features, and maximally obtains a complete feature set to provide a background feature set for subsequent feature sharpening; finally, a feature sharpening method based on group normalization is adopted to standardize the multi-specification features, and the standardized features are subjected to a segmentation-enhancement / attenuation-reorganization process to realize feature sharpening, thereby further improving the learning efficiency and prediction accuracy of the network model.

[0046] The method of the present application reconstructs the semantic information based on global information capture: existing translation models randomly generate a vector of length N to express the semantic information of the initial entity and relationship, and it can be understood that each component (or several components) corresponds to the semantic information of each field of the entity (relationship) name. However, in order to obtain the semantic information of the character or word, the current semantic learning method is to obtain the final semantic embedding from the context of the character or word in the sentence. In order to obtain accurate semantic information of the entity and the relationship, the information reconstruction method based on global information capture adopted by the present application uses a multi-layer dilated convolution kernel to capture global information of the matrix composed of the entity and the relationship, which expands the information receptive field while reducing the complexity of information reconstruction.

[0047] The method of the present application realizes multi-specification feature capture of reconstructed semantic information: the role of the relationship vector in the triple of the knowledge graph is to map and align the head entity to the tail entity, and the mapping of different relationship vectors to the same head entity vector is different, which contains more information of semantic understanding of the entity. In order to accurately realize the mapping of the head entity vector to the tail entity, the present application extracts different specifications of components from the relationship vector to form a convolution kernel, and captures multi-specification features of the reconstructed entity relationship semantic vector, thereby obtaining features from the entity relationship semantic in different angles and providing a background feature set for subsequent feature sharpening.

[0048] The method of the present application is based on group normalization feature sharpening: after obtaining the multi-specification feature set of the entity relationship, in order to further "sharpen" the features: strengthen the contribution of strong features, and weaken the interference of weak features. The method of the present application standardizes the feature set based on group normalization, sets a "sharpening" threshold, divides the standardized feature set, enlarges the feature values of one group which are higher than the threshold, and reduces the feature values of one group which are lower than the threshold, and then adds the two groups of feature sets to recombine a "sharpened" feature set, thereby further improving the saliency of the semantic features of the entity relationship. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is an example graph of a knowledge graph;

[0050] Figure 2 is an execution process diagram of a knowledge graph embedding model based on a convolutional neural network; the black dashed box in the figure represents a 2x2 convolution kernel; i) represents a feature reshaping mode used in ConvE; ii) represents a feature reshaping mode used in M-DCN; iii) represents a feature reshaping mode used in InteractE;

[0051] Figure 3 is a knowledge graph embedding method flowchart of the global information semantic reconstruction and multi-specification feature sharpening framework according to the present application;

[0052] Figure 4 is a structural diagram of an entity relationship semantic reconstruction module GFE based on global information capture;

[0053] Figure 5 is a feature extraction schematic diagram of a semantic information reconstruction module GSR based on global information capture; wherein S, T, U represent entity information, a i represents the features corresponding to the entity information, and it can be seen that the a1-a6 features correspond to the entity S;

[0054] Figure 6 is a specific convolution process schematic diagram of three different size relationship convolution kernels;

[0055] Figure 7 is a structural schematic diagram of a feature sharpening module based on group normalization;

[0056] Figure 8 is a comparison diagram of the influence of different learning rates on FB15K-237 and WN18RR;

[0057] Figure 9 is a comparison diagram of the influence of different L2 regularizations on FB15K-237 and WN18RR;

[0058] Figure 10 is a comparison diagram of the influence of different convolution kernel sizes on FB15K-237 and WN18RR. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0060] The present application will be further described below with reference to the drawings, but not as a limitation of the present application.

[0061] DETAILED DESCRIPTION Figures 1 to 7 As shown in the drawings, the present application provides a knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework, comprising,

[0062] Randomly initializing the triples in the knowledge graph to obtain D-dimensional triple embeddings, corresponding splicing of the head entity embedding and the relationship embedding in the D-dimensional triple embedding is performed according to the dimensions, and the initial feature Figure X 0 is obtained after reshaping;

[0063] The semantic information reconstruction module based on global information capture is used for feature extraction on the initial feature Figure X 0, and the initial featureFigure X 0 performs residual processing to obtain residual features Figure X 1;

[0064] Three relational convolution kernels of different sizes were selected to process the residual features respectively. Figure X 1. Feature extraction is performed to obtain three sets of feature mapping maps.

[0065] Then, a feature sharpening module based on group normalization is used to sharpen the feature map. Perform feature sharpening to obtain the sharpened feature map.

[0066] Three sharpened feature maps The hidden layer vectors are flattened and connected along the channel dimension, and a fully connected layer is used to obtain the hidden layer vectors. The dimensions of the hidden layer vectors are the same as those of the tail entity embeddings in the triple embeddings. Then, the hidden layer vectors and the tail entity embeddings in the triple embeddings are multiplied by a matrix to predict multiple scores. The tail entity embedding corresponding to the highest score is taken as the tail entity prediction result to achieve information reconstruction.

[0067] This implementation includes a semantic information reconstruction module based on global information capture (GSR), a multi-specification feature capture module for reconstructing semantic information (MSC), a feature sharpening module based on group normalization (GNS), and a scoring module. First, multi-layer dilated convolutions are used to reconstruct the semantic information of entity relationships, fusing contextual information to enhance the completeness of semantic expression and reduce reconstruction complexity. Then, by extracting convolution kernels of different specifications from the relationship vectors, multi-dimensional feature capture is performed on the reconstructed semantic information, obtaining feature sets from entity relationship semantics from different perspectives. Next, a feature sharpening method based on group normalization is used, achieving feature sharpening through a standardization-segmentation-enhancement / weakening-reconstruction process. Finally, in the scoring module, the score of the triples is calculated.

[0068] Entity Relationship Semantic Reconstruction Module Based on Global Information Capture—GSR:

[0069] Combination Figure 3 As shown,

[0070] First, the entities and relations in the knowledge graph are randomly initialized with a random D-dimensional embedding, where the embedding of the triple (e1, rel, e2) consisting of the head entity e1, relation rel, and tail entity e2 is as follows: To ensure that subsequent feature enhancement is not affected by the initial operation, only h and r are concatenated and their shapes are changed, resulting in the following new feature matrix:

[0071]

[0072] in, represents a concatenation operation; R() represents a feature reshaping operation after the concatenation operation. It can be seen that the initial dimension is changed to a feature belonging to by the concatenation and reshaping operation, where m x n = 2D, m is the number of rows of X0, and n is the number of columns of X0.

[0073] Further, as shown in Figure 4 and Figure 5 , the semantic information reconstruction module based on global information capture sequentially includes a deep convolution DW-Conv, two deep dilated convolutions DW-DConv with different dilated rates, and a normal convolution Conv; the initial feature Figure X 0 is sequentially input to the deep convolution DW-Conv and the deep dilated convolution DW-DConv, and the feature extraction result is output by the normal convolution Conv.

[0074] For the concatenation information composed of the head entity and the relation, in the process of using the convolutional neural network CCN to perform semantic extraction, the convolution operation can capture the logical relationship between the head entity and the relation only when the convolution kernel moves to a specific position in the information matrix. In other words, the inherent locality of the convolutional neural network CNN gives a limited degree of interaction between the information of the head entity and the relation. In order to cope with this challenge, the embodiment designs a new semantic information reconstruction module (GSR) based on global information capture, which uses a combination of multiple convolutions to focus on features from different neighboring positions, as shown in Figure 4 It can be seen that the GSR includes three parts: a deep convolution (DW-Conv), two deep dilated convolutions (DW-DConv) with different dilated rates, and a normal convolution (Conv). Through the combination of convolution and different dilated rates, some of the most distant features can be captured to the greatest extent, as shown in Figure 5 .

[0075] After the operation of the input information matrix by the GSR, a feature map with long-range dependencies can be obtained, and finally a residual is introduced to prevent overfitting.

[0076] In the embodiment, the feature Figure X 1 after the residual is represented as:

[0077] X1=A f +X0(3),

[0078] wherein A f is a feature map output by the semantic information reconstruction module based on global information capture:

[0079]

[0080] wherein Conv1 is a normal convolution Conv with a kernel size of 1x1, DW-Conv3 is a depthwise convolution DW-Conv with a kernel size of 3x3, is a depthwise dilated convolution DW-Dconv with a kernel size of 3x3 and a dilation rate of 3, is a depthwise dilated convolution DW-Dconv with a kernel size of 3x3 and a dilation rate of 9.

[0081] Based on the design of global information capture, GSR can capture each position in the information matrix during the calculation process of the matrix, ensuring that the information of each position can obtain the global semantic context through the dilated convolution. Through multi-layer dilated convolution, the model expands the receptive field, so that the generated feature map can fuse the global information in the receptive field. This feature makes it possible to maintain the integrity of global information during feature accumulation, and provides sufficient support for subsequent multi-scale feature capture.

[0082] The multi-specification feature capture module MSC reconstructs semantic information:

[0083] In combination with FIG. 3, three different size relationship convolution kernels are selected as 1x2 Figure 6 2x2 and 3x1

[0084] To extract the feature of the tail entity alignment from the entity relationship semantic based on global information capture, the embodiment adopts the multi-specification feature capture method of reconstructing semantic information. By extracting different specifications of components from the relationship vector to form a convolution kernel, the reconstructed entity relationship semantic is captured from multiple angles and multiple scales, providing diversified and sufficient background information support for subsequent feature sharpening. This mechanism not only improves the efficiency of the head entity vector and the tail entity mapping alignment, but also lays a solid foundation for the global and local feature interaction of the knowledge graph embedding model.

[0085] Based on the relationship embedding r, the embodiment constructs three different size convolution kernels and respectively.Taking the construction process of the convolution kernel as an example, the relationship embedding r is divided into blocks of the same size, wherein each block is reorganized into the same convolution kernel wherein n is the number of convolution kernels, h l and w l represent the height and width of each convolution kernel, and n, h l and w l maintain the following relationship: D=nh l w l .Figure 6 A simple example of MSC module is shown, in which the relation vector of length 9 is split and shaped into three different convolution kernels and The reason for using relation embedding as convolution kernel is that relation can understand the semantics of head entity and project it to the correct tail entity as a bridge of connection, so the decoding information of entity semantics is contained in the relation embedding, and different convolution kernels extracted from it can accurately obtain the local features of the correct alignment of tail entities from entity semantics.

[0086] Subsequently, the three convolution kernels are used as the convolution kernels of different convolutions to extract features from the entity semantic matrix from different scales, obtaining three different feature maps.

[0087] Three groups of feature maps are:

[0088]

[0089] wherein is a real set, C is the number of channels, H is the height of , and W is the width of .

[0090] Through three kinds of convolution kernels of n specifications, the entity semantic matrix in the embodiment obtains a feature set containing nxCxHxW elements. Each convolution kernel performs convolution operation on the entity relation semantic matrix from different scales to extract multi-angle and multi-scale feature maps, thereby obtaining more rich and comprehensive semantic information. The feature set corresponding to each group of convolution kernels can effectively capture the relationship information between entities and provide diversified and sufficient background information support for the subsequent feature sharpening step.

[0091] Feature sharpening module based on group normalization-GNS:

[0092] Further, in order to further reduce irrelevant feature interference, combined with Figure 7 As shown, the feature sharpening process of the feature sharpening module based on group normalization includes standardization, feature segmentation, feature enhancement and weakening, and recombination of the feature map

[0093] ​​The standardization step is to unify the order of magnitude of the feature set and constrain the importance of the features in the set with a scaling factor; the purpose of the segmentation step is to separate the features associated with the alignment target from the weak features, thereby providing a clear processing target for subsequent operations; the enhancement / attenuation step separately processes the two groups of features after segmentation: amplifying strong features above a certain threshold, strengthening their contribution in subsequent feature learning; attenuating weak features below the threshold, reducing their interference with model performance; the reorganization step integrates the enhanced and attenuated features into a feature set, generating a "sharpened" feature set. The final generated feature set can highlight key semantic features and suppress redundant interference features, thereby significantly improving the significance and embedding efficiency of entity relationship semantic features.

[0094] The standardization method is to evaluate the importance of the feature map using the scaling factor in the group normalization layer:

[0095]

[0096] wherein is the normalization result obtained according to the group normalization, μ i is the mean of , σ i is the standard deviation of , ε i is the adjustment parameter of , λ i and β i are two training adjustment parameters of the group normalization layer GN;

[0097] It can be seen that by using the value of the trainable parameter in the GN layer, the spatial feature importance of each batch and channel can be measured. Based on the normalization result determine the training adjustment parameter λ i ; then based on the training adjustment parameter λ i calculate the normalization weight is:

[0098]

[0099] wherein is the normalization weight of the kth channel, is the training adjustment parameter λ i corresponding to the kth channel, is the training adjustment parameter λ i corresponding to the jth channel;

[0100] The normalization weight can reflect the importance of different features.

[0101] Then perform feature segmentation: use the normalization weight Weighted feature map The weights are mapped to the (0, 1) range using the sigmoid function, and then the adjusted weights K are calculated. i :

[0102]

[0103] In the formula, GN represents the pair of pairs ... The training adjustment parameter λ is obtained by performing group normalization. i The process; W() represents the calculation of normalized weights. Sigmoid() is a scaling function that scales values ​​to the range (0,1); Threshold() is an adjustable threshold function that sets values ​​greater than the set weight threshold to 1 and values ​​less than or equal to the set weight threshold to 0.

[0104] According to the adjusted weight K i The adjusted weight K is greater than the set weight threshold. i Set it to 1 to get the weight K. 1i The adjusted weight K, which is less than or equal to the set weight threshold, will be... i Set it to 0 to get the weight K. 2i .

[0105] Feature enhancement / weakening: transforming input features Multiply by K respectively 1i and K 2i This yields two weighted features: features with higher information content. and features with less information At this point, the input elements are divided into two parts: Spatial content with strong informational and expressive power, and Having little or no information is considered redundant.

[0106] Based on weight K 1i and weight K 2i right Weighted features are obtained by performing feature enhancement and weakening. and weighted features

[0107] Feature recombination: and By performing positional addition, we obtain the "sharpened" feature set. The important features become more prominent, while the weaker features become weaker.

[0108] Then weighted features and weighted features The sharpened feature map is obtained by recombination

[0109]

[0110] wherein, represents an element-wise addition. After the model is applied to the intermediate input feature , not only the features with large information amount are separated from the features with small information amount, but also the features are reconstructed by "sharpening" to suppress the redundant features in the global dimension by the representative features.

[0111] The scoring module:

[0112] In the scoring module, firstly, the "sharpened" feature set is connected along the channel dimension, and a fully connected layer is applied to obtain a hidden layer vector with the same embedding dimension as the embedding dimension of the tail entity. Secondly, the vector is matrix-multiplied with the tail entity embedding to obtain a series of scores, each score reflecting the possibility that the head entity and the relationship embedding obtained by the model can correctly predict the triple. Finally, the maximum score in the score is selected as the prediction result.

[0113] Further, the method for calculating the score is:

[0114]

[0115] wherein Scores(h,r,t) represents a score function of the D-dimensional triple embedding, wherein h represents the head entity embedding, r represents the relationship embedding, t represents the tail entity embedding, f(·) represents projecting the hidden layer vector into a space with the same dimension as the tail entity embedding, vec(·) represents flattening the feature map, GSR(·) represents a global feature strengthening operation, MSC(·) represents a neighbor feature strengthening operation, GNS(·) represents a feature reconstruction operation, and R() represents a reshaping operation, W0 is a randomly initialized parameter matrix, and after the feature map is flattened, a linear transformation parameterized by the matrix W0 is used to project into a k-dimensional space.

[0116] In the embodiment, the calculation of the score function Scores(h,r,t) is implemented by selecting the 1-N score; the cross-entropy loss function ψ(p,y) is calculated according to the score value, and the adjustment parameters ε i , the training adjustment parameters λ i and β i ; the cross-entropy loss function ψ(p,y) is:

[0117]

[0118] wherein p represents the predicted triple embedding score, y represents a binary label vector, N is the total number of triples, and p nThe triplet embedding score for the nth prediction, y n The binary label vector corresponding to the nth prediction.

[0119] In order to more efficiently calculate the score, the 1-N score is selected in the embodiment to evaluate and calculate the correctness of the triplet. The 1-N score refers to that the model calculates the score of the embedding generated by (e1, rel) with all tail entities e2 (that is, all possible candidate entities) at one time, and selects the highest or lowest entity among all calculated scores as the prediction result to generate a score vector. In addition, the parameters are trained using the ADAM optimizer according to the score, and batch normalization is used in the training process to stabilize and accelerate convergence, and dropout and L2 regularization are used to alleviate overfitting and improve generalization. Finally, the model is trained through the cross-entropy loss function.

[0120] The GIRMSF model program of the knowledge graph embedding is as follows:

[0121]

[0122] Verification experiment:

[0123] Firstly, the commonly used seven data sets are introduced below to facilitate the understanding and query of related data, and to ensure the transparency, repeatability and credibility of the results of the experiment; secondly, the evaluation index and parameter setting of the experiment are introduced, the experiment evaluation index is used to evaluate the results of the experiment, which is helpful to evaluate the performance of the model comprehensively, the experiment parameter setting can optimize the performance of the model, ensure the reliability and repeatability of the experimental results, and is helpful to the subsequent reproduction of the experiment; then, the comparative baseline method adopted is introduced; then the results of the model corresponding to the application on five data sets are introduced, and the results are analyzed, and the advancement and superiority of the model of the application method are reasonably analyzed; the effectiveness of the model is verified, and the ablation experiment is carried out for different modules; finally, the influence of different hyperparameters on the performance of GIRMSF is discussed.

[0124] (I) Data set

[0125] Seven standard knowledge graph data sets, FB15K, WN18, FB15K-237, WN18RR, YAGO3-10, KINSHIP and UMLS, are used to evaluate the model of the application. The specific information in these data sets is shown in Table 1.

[0126] The FB15K dataset is derived from the Freebase knowledge base, which is a large open-source knowledge base led by Google and contains a large amount of entity and relationship information. The WN18 dataset is extracted from the WordNet knowledge base, which is a large, hand-built lexical database. FB15K-237 is a subset of FB15K, and FB15K obtains the triple data in the test set by reversing the triple data in the training set, which means that the relationship in the training set and the relationship in the test set are the same (the case of relationship leakage), and in order to solve the problem that simple models can also achieve good prediction results due to relationship leakage, the original relationship in the training set is deleted from the test set to obtain FB15K-237. WN18RR is a subset of WN18, and WN18 is also obtained by deleting the relationship in the training set from the test set to obtain WN18RR.

[0127] YAGO3-10 is a subset of YAGO3, and YAGO (Yet Another Great Ontology) is developed by the Max Planck Institute for Computer Science, aiming to integrate information from different data sources such as Wikipedia, WordNet and GeoNames, covering multilingual entries and descriptions, and also containing rich type information (type hierarchy).

[0128] The KINSHIP dataset was first collected by Denham and Warlpiri, etc. for studying the kinship structure of Australian aborigines. After being organized and expanded, it is widely used in knowledge graph, relationship learning, logical reasoning and other fields. It is usually used by academia and researchers to test the performance of algorithms in dealing with complex relationship data.

[0129] The UMLS dataset is created by the National Library of Medicine (NLM) in the United States, aiming to unify the representation of different medical terms and coding systems. UMLS is a medical knowledge graph dataset that integrates multiple medical terminology systems and classification systems to promote the standardization and interoperability of medical information.

[0130] Table 1 Detailed description of seven datasets

[0131]

[0132] (ii) Evaluation indicators

[0133] In order to evaluate the prediction performance of the model on different datasets, we use four different evaluation indicators, including MRR, Hit@1, Hit@5 and Hit@10.

[0134] MRR (Mean Reciprocal Rank) is an internationally recognized mechanism for evaluating search algorithms. That is, if the score of the first matching result is 1, the score of the second matching result is 0.5, the score of the nth matching result is 1 / n, and if there is no match, the score is 0, and the final score is the average of all scores. The value range of MRR is MRR ∈ (0, 1], and the calculation of MRR for a triple (h, r, t) is shown in equation (14):

[0135]

[0136] where T test represents the test set combination; rank r,t (h) represents the ranking position of the correct head entity in the sorted prediction list in the test set, rank h,r (t) represents the ranking position of the correct tail entity in the sorted prediction list in the test set. When predicting the hidden head entity, the knowledge graph embedding model calculates the scores of all entities in the triple (?, r, t) of the head entity. Similarly, when predicting the hidden tail entity, the knowledge graph embedding model calculates the scores of all entities in the triple (h, r,?) of the tail entity.

[0137] Hit@h is a measure of whether the model contains at least one relevant item in the first h returned results, that is, to evaluate how many of the first k entities predicted by the model are correct. The commonly used h values are 1 and 10. For a triple (h, r, t), the specific calculation of Hit@h is shown in equations (15) and (16).

[0138]

[0139] where rank r,t (h) ≤ h indicates whether the model contains the correct target head entity in the first h returned results. Similarly, rank h,r (t) ≤ h indicates whether the model contains the correct target tail entity in the first h returned results.

[0140] It can be seen that the larger the values of MRR and Hit@h, the better the prediction results obtained, indicating that the performance of the model is more superior.

[0141] (III) Experimental setup

[0142] GIRMSF is implemented using PyTorch 1.21.1 and experiments are conducted on a personal server equipped with an Intel Core i9-1490K central processing unit (CPU), an NVIDIA GeForce RTX 4090 graphics processing unit (GPU), and 96 GB of random access memory (RAM). All parameters are initialized with a Haar normal distribution. If the training loss does not decrease for five epochs, the learning rate will be reduced by a factor of 0.9. For all datasets, training will stop if MRR and Hit@10 do not improve for 20 epochs, and the maximum epoch is predetermined to be 500. The specific hyperparameter selection is shown in Table 2.

[0143] Table 2 Parameter settings for each dataset

[0144]

[0145]

[0146] (Four) Comparative method: A large number of knowledge graph embedding models are selected as comparative methods. The classic translation-based embedding model methods are TransE, TransR, the semantic matching-based embedding model is DisMult, ComplEx, RotatE, TorusE, the neural network-based embedding model is HypER, ConvE, R-GCN, InteractE, ConvR, CTKGC, etc. The current more advanced methods ConvHL, MSHE, SDFormer, etc. By comparing with existing models, the superiority of the model of the application is verified.

[0147] (Five) Main results: FB15K-237 and WN18RR are two of the most widely used datasets for embedding prediction. A large number of research work has been proposed based on these two datasets. Table 3 shows the comparison results of GIRMSF and benchmark models between FB15K-237 and WN18RR, the best result is highlighted in bold, and the second best result is shown in underlined.

[0148] Table 3 Embedding results on FB15k-237 and WN18RR

[0149]

[0150] The experimental results show that the performance of GIRMSF in the method is better than that of the baseline method on the two data sets. Among all the methods, the neural network-based method is generally better than other types of methods. This is because the nonlinear fitting capability of the neural network is very powerful. Even compared with the most advanced neural network model, GIRMSF still shows a significant performance advantage, especially on the WN18RR data set. For example, GIRMSF exceeds SDFormer by 6.0%, 4.5%, 6.4% and 7.7% on the MRR, Hits@1, Hits@3, Hits@10 indicators, respectively. However, the models based on translation or semantic matching have simple assumptions. The only way to improve the expression capability of these models is to increase the embedding dimension, which in turn imposes a limitation on the expression capability of the model. InteractE, ConvR, CTKGC and M-DCN are all methods to enhance the performance of the model by increasing the interaction between entities and relations. Among them, InteractE is the first to define the important influence of interaction on the performance of knowledge graph embedding models, and increases the interaction by reshaping the feature matrix. These methods have relatively excellent performance among all the compared methods, which is sufficient to prove the key role of increasing the interaction between entities and relations.

[0151] Detailed comparison of GIRMSF and convolutional neural network-based knowledge graph embedding models: For the newly proposed M-DCN, GIRMSF exceeds the MRR, Hits@1, Hits@3, Hits@10 indicators of FB15k-237 by 3.6%, 4.5%, 4.5% and 4.3% respectively, and the MRR, Hits@1, Hits@3, Hits@10 indicators of WN18RR by 1.6%, 1.1%, 3.6% and 5.6% respectively. These are significant performance improvements. The reason for the performance advantage of the method is not only the use of the information reconstruction module based on global information capture and the reconstructed information multi-specification feature capture module to maximize the interaction between entities and relations, but also the use of the information sharpening module based on group normalization to reduce the influence of irrelevant features. The combination of these advanced components promotes the performance improvement of GIRMSF.

[0152] Secondly, the GIRMSF of the present application is applied to large-scale knowledge graph data sets FB15K and WN18. The FB15K and WN18 data sets have been widely used in early knowledge graph embedding research, although the information leakage problem in the training set of the two data sets greatly reduces the complexity of link prediction, but the experiment on the two data sets also plays an important role in proving the effectiveness of the model. Table 4 shows the comparison results of GIRMSF and the benchmark model between FB15K and WN18.

[0153] Table 4 Embedding results on FB15K and WN18

[0154]

[0155]

[0156] For GIRMSF, the performance on FB15K and WN18 is significantly better than other comparative models. It can be seen that even if GIRMSF fails to achieve the best result in some indicators, the difference between it and the best experimental result is very small. It is speculated that the reason is that the information reconstruction module based on global information capture of GIRMSF can extract far information, and the reconstruction information multi-specification feature capture module can dynamically adapt and adjust the utilization rate of high-dimensional features, and the information sharpening module based on group normalization can adaptively propose redundant features, thereby obtaining better performance.

[0157] Then, GIRMSF is applied to the large-scale knowledge graph dataset YAGO3-10 to further verify the performance of its embedding. The training set of YAGO3-10 has more than 100 million triples. The performance on large datasets can test the fitting ability of the knowledge graph embedding model. Table 5 shows the comparison results between GIRMSF and the benchmark model on YAGO3-10.

[0158] Table 5 Embedding results on YAGO3-10

[0159]

[0160] It can be seen that GIRMSF is better than other methods in all indicators on large datasets, indicating the progress of the model of the present application and its applicability to large knowledge graphs. Specifically, compared with the latest neural network-based model MSHE, GIRMSF can show very significant improvement, with an increase of 2.0%, 2.1%, 2.3% and 2.6% in MRR, Hits@1, Hits@3 and Hits@10, respectively. Large knowledge graphs generally contain more knowledge information, and more knowledge information tests the learning ability and noise reduction ability of the model. It can be seen that GIRMSF has relatively excellent results in large, multi-relation datasets.

[0161] In order to verify the universality of the model of the present application on small datasets, GIRMSF is applied to small knowledge graph datasets KINSHIP and UMLS, which have 8544 and 5216 triples in the training set, respectively. Table 6 shows the comparison results between GIRMSF and the benchmark model on KINSHIP and UMLS.

[0162] Table 6 Embedding results on KINSHIP and UMLS

[0163]

[0164] From the experimental results, it can be found that HypER, which performs well on medium and large datasets, does not have obvious performance advantages on small datasets. This may be because its complex architecture is more likely to cause overfitting on small datasets. While GIRMSF uses the information reconstruction module (GIRM) based on global information capture, the multi-specification feature capture module (MSC), and the information sharpening module (GNS) based on group normalization to reduce the fitting difficulty of the model, thereby improving the generalization ability of the model. Therefore, GIRMSF has achieved the best performance on almost all indicators on both datasets. Experiments on small datasets can also prove the competitiveness of traditional methods. Even the ComplEx model, which performs poorly on medium and large datasets, has shown good performance on these two datasets. This is because the size of these datasets is small, and semantic matching models can achieve effective fitting.

[0165] (VI) Running efficiency of GIRMSF

[0166] In addition to the performance of embedding prediction, time efficiency is also an important indicator of knowledge graph embedding models. A model that is too complex will make the model difficult to run and difficult to apply to large knowledge graphs. Many works have been devoted to improving the efficiency of the model. For example, the parameter efficiency of ConvE is 17 times that of R-GCN and 8 times that of DisMult. In order to evaluate the efficiency of GIRMSF, the running time of GIRMSF at each epoch is compared with the classic convolutional neural network-based knowledge graph embedding model ConvE and the newly proposed CTKGC and M-DCN. Experiments were conducted on the medium-sized dataset WN18RR and the large-scale dataset YAGO3-10. The experimental results are shown in Table 7.

[0167] Table 7 Comparison of running time of five models on two datasets

[0168]

[0169] From the experimental results, it can be seen that the model training time of GIRMSF on the WN18RR dataset is much smaller than that of the other three models, and on the YAGO3-10 dataset, the running time of CTKGC is the shortest, but the running time of GIRMSF is not much different from that of the other two methods. Analysis shows that because GIRMSF increases the information reconstruction module based on global information capture, the multi-specification feature capture module of reconstructed information, and the information sharpening module based on group normalization, for complex datasets, more time is needed for learning and reconstruction. For relatively simple datasets, even if the data volume is large, GIRMSF can quickly learn important features and quickly reconstruct features, thereby speeding up the training. In general, GIRMSF shows high model efficiency and is very competitive compared with the current most advanced CNN-based model. Therefore, it has good application prospects and can be widely applied to medium and large-scale datasets.

[0170] (VII) Ablation experiment

[0171] An ablation experiment is performed to verify the effectiveness of GSR, MSC and GNS. Specifically, taking GIRMSF as the baseline model, GSR, MSC and GNS are deleted from the complete model respectively, and three comparison methods are set, including GIRMSF-GSR (without using GSR), GIRMSF-MSC (without using MSC) and GIRMSF-GNS (without using GNS). The FB15K-237 and WN18RR datasets are selected for experiments, and each experiment is repeated 10 times under the best hyperparameters, and the average value is taken as the final experimental result. Finally, the results are shown in Table 8.

[0172] Table 8 Ablation experiment performed on WN18RR and FB15k-237 datasets

[0173]

[0174] The experimental results show that during the training process of GIRMSF, the neglect of any one of GSR, MSC and GNS will lead to a significant decrease in model performance, and the complete GIRMSF is superior to all ablation models. It can be seen that each key module of the method of the present application can play a certain role, and the complete model has strong generalization ability and performs better than the comparison methods on multiple datasets.

[0175] (VIII) Discussion on the influence of different hyperparameters on the performance of GIRMSF

[0176] Next, the impact of some key hyperparameters on the performance of GIRMSF is evaluated to determine its hyperparameter sensitivity. Specifically, the performance of GIRMSF is experimented with different hyperparameters of convolution kernel size, L2 regularization, and learning rate, with the FB15K-237 and WN18RR datasets selected.

[0177] Impact of different learning rates:

[0178] Learning rate is a key parameter of neural networks. It directly affects the convergence speed of the model and the performance of the model. Selecting an appropriate learning rate is crucial for neural network-based knowledge graph embedding models. A grid search method is adopted. This strategy helps to fine-tune the weights better, thus fitting the combined data better. At this time, the initial learning rate needs to be set. Experiments are conducted based on four different initial learning rates (0.0001, 0.0005, 0.001, and 0.005) to explore the impact of learning rate on the performance of GIRMSF. The experimental results are shown in Tables 9 and Figure 8 From the results, it can be seen that for different datasets, it is important to choose the corresponding learning rate, which will significantly affect the performance of the model. On the FB15K-237 dataset, the best learning rate is 0.0005, while on the WN18RR dataset, the best learning rate is 0.001.

[0179] Table 9 Impact of different learning rates on FB15K-237 and WN18RR

[0180]

[0181] Impact of different L2 regularizations: L2 regularization adds a regularization term to the loss function, which penalizes large weight values and encourages the model to use smaller weights for prediction, and is an effective means to reduce the risk of model overfitting. L2 regularization introduces a hyperparameter, by adjusting this hyperparameter, the influence of regularization can be controlled. The performance of GIRMSF is studied when the regularization coefficient is 5e-3, 5e-4, 5e-5, and 5e-6. The experimental results are shown in Tables 10 and Figure 9 It can be found that when a too large L2 regularization coefficient is selected, such as 5e-3, the model does not fit, and the results will be very poor. On the FB15K-237 dataset, the best L2 regularization coefficient is 5e-4, while on the WN18RR dataset, the best L2 regularization coefficient is also 5e-4.

[0182] Table 10 Impact of different L2 regularizations on FB15K-237 and WN18RR

[0183]

[0184] Impact of different convolution kernel sizes

[0185] Convolution kernel size represents the different receptive fields used for feature extraction within the model. By conducting sufficient experiments, the optimal combination of multi-scale convolution kernel sizes is selected.

[0186] In the experiment, various combinations of different convolution kernel sizes are selected, including [[1*3], [1*3], [1*3]], [[3*3], [1*3], [1*3]], [[1*3], [3*3], [1*3]] and [[1*3], [3*3], [3*3]], among which the last two kernel combinations are single-scale. The experimental results are shown in Table 11 and Figure 10 As can be seen from the experimental results, different convolution kernel sizes have little effect on the overall performance of the model, so the [[1*3], [3*3], [1*3]] multi-scale convolution kernel group used in the test is selected.

[0187] Table 11 Effect of different convolution kernel sizes on FB15K-237 and WN18RR

[0188]

[0189] In summary, the method of the present application realizes the embedding representation of the knowledge graph based on the global information semantic reconstruction and multi-specification feature sharpening framework-GIRMSF. First, based on the entity relationship semantic reconstruction method of global information capture, the semantic information of entities and relationships with context logical association information is obtained by reconstructing the global information of the semantic information matrix composed of entities and relationships through multi-layer dilated convolution; second, in order to improve the efficiency of entity relationship alignment in the knowledge graph, a multi-specification feature capture mechanism for semantic reconstruction information is proposed, which extracts different specification convolution kernels from the relationship vector and convolves with the reconstruction information to capture the entity relationship semantic information from different angles. Finally, an information sharpening method based on group normalization is used to standardize and threshold adjust the set of multi-specification features, enhance the key information and suppress the redundant features, and further improve the learning efficiency and prediction accuracy of the model. Through this multi-level information processing strategy, GIRMSF can perform well in the knowledge graph embedding task, and improve the accuracy and robustness of entity and relationship representation.

[0190] Extensive experiments on seven different scale benchmark datasets show that GIRMSF has significantly improved performance, especially in medium and large datasets. These experimental results prove that GIRMSF can consistently provide high-quality embedding representation in knowledge graph datasets of different scales and complexities, improving the accuracy and efficiency of the completion task. Through these experimental verifications, GIRMSF fully demonstrates its potential as a powerful and versatile knowledge graph embedding method, laying a solid foundation for further application of knowledge graphs.

[0191] While the application has been described with reference to particular embodiments, it is to be understood that the application is not limited to the particulars disclosed. Rather, it is a continuation of the principles and application of the present application. It is therefore to be understood that many modifications, omissions, and additions can be made to the examples and embodiments described herein without departing from the spirit and scope of the present application, which is limited only by the claims that follow.

Claims

1. A knowledge graph embedding method based on a global information semantic reconstruction and multi-specification feature sharpening framework, characterized in that, include, Randomly initialize the triples in the knowledge graph to obtain D-dimensional triple embeddings. Then, concatenate the head entity embeddings and relation embeddings in the D-dimensional triple embeddings according to their dimensions, and reshape them to obtain the initial feature map X0. A semantic information reconstruction module based on global information capture is used to extract features from the initial feature map X0, and then residual processing is performed on the initial feature map X0 to obtain the residual feature map X1. Three relational convolution kernels of different sizes were selected to extract features from the residual feature map X1, resulting in three sets of feature maps. i = 1, 2, 3; the relational convolution kernel is a convolution kernel composed of components of different specifications extracted from the relational vector; Then, a feature sharpening module based on group normalization is used to sharpen the feature map. Perform feature sharpening to obtain the sharpened feature map. ; Three sharpened feature maps The hidden layer vectors are flattened and connected along the channel dimension, and a fully connected layer is used to obtain the hidden layer vectors. The dimensions of the hidden layer vectors are the same as those of the tail entity embeddings in the triple embeddings. Then, the hidden layer vectors and the tail entity embeddings in the triple embeddings are multiplied by a matrix to predict multiple scores. The tail entity embedding corresponding to the highest score is taken as the tail entity prediction result to achieve information reconstruction. The semantic information reconstruction module based on global information capture includes, in sequence, a deep convolution DW-Conv, two deep dilated convolutions DW-DConv with different dilation rates, and a regular convolution Conv; the initial feature map X0 is input to the deep convolution DW-Conv and the deep dilated convolution DW-DConv in sequence, and the feature extraction result is output by the regular convolution Conv. The feature sharpening module based on group normalization sharpens the feature map. The process of feature sharpening includes processing the feature map. Standardization, feature segmentation, feature enhancement and weakening, and reorganization are performed. The standardization step involves unifying the size of the feature set and using a scaling factor to constrain the importance of each feature within the set. The feature segmentation step separates features that are strongly associated with the alignment target from those that are weak. The feature enhancement and weakening steps process the two sets of features after segmentation differently: strong features above a set threshold are amplified to enhance their contribution to subsequent feature learning; weak features below the threshold are attenuated to reduce their interference with model performance. The recombination step integrates the enhanced and weakened features into a single feature set to generate a sharpened feature set.

2. The knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework according to claim 1, characterized in that, The residual feature map X1 is represented as follows: X1=A f +X0, In the formula A f Feature maps output by the semantic information reconstruction module based on globally captured information: In the formula, Conv1 is a regular convolution Conv with a kernel size of 1×1, and DW-Conv3 is a depthwise convolution DW-Conv with a kernel size of 3×3. This is a depthwise dilated convolution (DW-Dconv) with a kernel size of 3×3 and a dilation rate of 3. It is a depth dilated convolution DW-Dconv with a kernel size of 3×3 and a dilation rate of 9.

3. The knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework according to claim 2, characterized in that, The three relational convolution kernels of different sizes are each selected as 1×2. 2×2 and 3×1 4. The knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework according to claim 3, characterized in that, Three sets of feature mapping maps for: In the formula Let C be the set of real numbers, C be the number of channels, and H be the number of channels. The height, W is The width.

5. The knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework according to claim 4, characterized in that, The standardization method involves evaluating the importance of feature maps using a scaling factor in the group normalization layer. In the formula To calculate the normalization result obtained from group normalization, μ i for The mean, σ i for standard deviation, ε i for Adjustment parameters, λ i and β i Two training adjustment parameters for the group normalization layer GN; Based on normalization results Determine the training adjustment parameter λ i Then, based on training, adjust the parameter λ. i Calculate normalized weights for: In the formula The normalized weight for the k-th channel. For λ i The training adjustment parameters corresponding to the k-th channel, For λ i The training adjustment parameters corresponding to the j-th channel; Next, feature segmentation will be performed: normalized weights will be used. Weighted feature map The weights are mapped to the (0, 1) range using the sigmoid function, and then the adjusted weights K are calculated. i : In the formula, GN represents the pair of pairs ... The training adjustment parameter λ is obtained by performing group normalization. i The process; W() represents the calculation of normalized weights. Sigmoid() is a scaling function that scales values ​​to the range (0,1); Threshold() is an adjustable threshold function that sets values ​​greater than the set weight threshold to 1 and values ​​less than or equal to the set weight threshold to 0. According to the adjusted weight K i The adjusted weight K is greater than the set weight threshold. i Set it to 1 to get the weight K. 1i The adjusted weight K, which is less than or equal to the set weight threshold, will be... i Set it to 0 to get the weight K. 2i .

6. The knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework according to claim 5, characterized in that, Based on weight K 1i and weight K 2i right Weighted features are obtained by performing feature enhancement and weakening. and weighted features Then weighted features and weighted features Reconstruction yields a sharpened feature map.

7. The knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework according to claim 6, characterized in that, The method for calculating the score is as follows: In the formula, Scores(h,r,t) represents the scoring function of the D-dimensional triplet embedding, where h represents the head entity embedding, r represents the relation embedding, t represents the tail entity embedding, f(·) represents projecting the hidden layer vector into a space of the same dimension as the tail entity embedding, vec(·) represents flattening, GSR(·) represents global feature enhancement operation, MSC(·) represents nearest neighbor feature enhancement operation, GNS(·) represents feature reconstruction operation, and R() represents shaping operation. W0 is a randomly initialized parameter matrix.

8. The knowledge graph embedding method based on global information semantic reconstruction and multi-specification feature sharpening framework according to claim 7, characterized in that, The scoring function Scores(h,r,t) is calculated using a 1-N rating scale; the cross-entropy loss function ψ(p,y) is calculated based on the score values, and the adjustment parameter ε is adjusted accordingly. i Training adjustment parameter λ i and β i The cross-entropy loss function ψ(p,y) is: In the formula, p represents the predicted triple embedding score, y represents the binary label vector, N is the total number of triples, and p n Let y be the triple embedding score of the nth prediction. n This is the binary label vector corresponding to the nth prediction.

Images