A small sample knowledge graph completion method based on multi-view semantic enhancement

Abstract

A small sample knowledge graph completion method based on multi-view semantic enhancement. By combining the structure view and the text view to generate the embedding representation of entities and relationships, the influence of the common sense knowledge implied by the entities on the triplets is solved through an adaptive attention mechanism, and then the neighborhood information of the entities is aggregated by considering that the entities have different common sense information under different relationships; the internal and external optimization complex relationship processing positions the embeddings with the same / different common sense knowledge into the same / different semantic spaces through common sense semantic constraints, then in the internal optimization process, the characteristics of the common sense view and the complex relationship are utilized to realize the supervision of the negative sampling process to generate effective and high-quality negative samples, and to learn the representation and hyperplane parameters of the complex relationship, and then in the external optimization process, all parameters are updated to make the model obtain better results; the entities that do not conform to the common sense knowledge among the candidate entities are filtered through the common sense view, and the final candidate entity ranking is output.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of knowledge graph completion, and particularly relates to a small-sample knowledge graph completion method based on multi-view semantic enhancement. BACKGROUND

[0002] In recent years, knowledge graphs such as Freebase, NELL and Wikidata have been widely used in various natural language processing tasks, including question answering systems, relation extraction and semantic search. However, manually or automatically constructed knowledge graphs are inevitably incomplete, and knowledge graph completion is needed to infer new facts. Knowledge graph completion aims to automatically infer missing facts by examining existing facts, and therefore has received extensive attention.

[0003] A method called knowledge graph embedding has been successfully applied to knowledge graph completion. Specifically, knowledge graph completion based on knowledge graph embedding can be divided into two stages: learning knowledge graph embedding in training, and learning link prediction in validation. Many knowledge graph embedding techniques have been proposed for the knowledge graph completion task, such as TransE proposed in Translating embeddings for modeling multi-relational data, which treats relations as a conversion operation between entity pairs, i.e. the head entity embedding plus the relation embedding can obtain the tail entity embedding. TransH proposed in Knowledge graph embedding by translating on hyperplanes, which embeds entities into specific hyperplanes for different relations to give entities various representations; DistMult proposed in Embedding Entities and Relations for Learning and Inference in Knowledge Bases, which embeds entities and relations into a complex vector space to capture the characteristics of relations and thus handle knowledge graph completion problems. However, in particular, all these methods assume that there are sufficient instances of entities and relations in triples.

[0004] However, small sample problems are widespread in real-world knowledge graphs. For example, about 10% of the relations in Wikidata have no more than 10 triples. In terms of small sample relations in real-world knowledge graphs, their representations cannot be effectively learned through limited training instances, which leads to unsatisfactory performance of such relations in knowledge graph completion.

[0005] To solve the problem of small sample knowledge graph completion, one of the tasks called few-shot link prediction emerged. To solve this problem, existing small sample knowledge graph completion models can be divided into three categories: metric learning based, meta learning based and additional information learning based.

[0006] In the metric learning based method, the first attempt was made in One-shot relational learning for knowledge graphs and GMatching was proposed, which matched entity pairs by utilizing local graph structure of entities and learning distinguishable metrics. FSRL proposed in Few-shot knowledge graph completion extended GMatching from single sample to few-shot case, and further used attention mechanism to capture local graph structure. Adaptive attentional network for few-shot knowledge graph completion considered the dynamic properties of entities and proposed FAAN to learn dynamic entity and reference representation.

[0007] In the meta learning based method, Meta relational learning for few-shot link prediction in knowledge graphs proposed a meta relational learning framework MetaR, which passed information using specific meta and accelerated learning for small sample knowledge graph completion using gradient meta. Relational learning with gated and attentive neighbor aggregator for few-shot knowledge graph completion learned the general representation of few-shot relations through gated neighbor aggregator, considered 1-N, N-1 and N-N complex relations, combined meta learning with TransH for small sample knowledge graph completion.

[0008] In the method based on additional information learning, Zero or few shot knowledge graph completions by text enhancement with multi-grained attention obtains the text representation of the entity by describing the relevance between the text knowledge and the given entity at the word and sentence levels, and combines it with the triple to complete the small sample knowledge graph. Challenging the Assumption of Structure-based embeddings in Few-and Zero-shot Knowledge Graph Completion enriches two knowledge graph benchmark datasets by using the text related to the entity, and combines the text and the structure of the graph on the FAAN to infer the missing relationship.

[0009] However, as pointed out in GMatching, structured knowledge graph embeddings often lack the ability to predict small sample relationships, ignoring the semantic knowledge of entities and relationships. Both the first and second types of methods only consider the structure view, and only use the graph structure of the knowledge graph to complete the small sample knowledge graph. In contrast, multi-view learning can combine more views that use internal and external information for link prediction. The third type of method only combines text information and structure information, without considering other additional information such as common sense information for entities and relationships. And their work only combines structure embeddings with text embeddings, without considering other factors such as complex relationships that will affect the completion results. SUMMARY

[0010] In view of the deficiencies of the prior art, the present application provides a small sample knowledge graph completion method based on multi-view semantic enhancement, which extracts semantic information of entities and relationships from multiple views by using a multi-view joint encoder, and combines common sense views to generate high-quality negative triples with the characteristics of common sense and complex relationships, solve the problem of invalid negative sampling of the model, and introduce common sense semantic constraints to limit the embedding of entities in the semantic space and improve the generalization ability and prediction quality of the model, thereby improving the overall performance of the model.

[0011] To achieve the above purpose, the present application provides a small sample knowledge graph completion method based on multi-view semantic enhancement, comprising:

[0012] Through the multi-view joint encoder, the structure view and text view embeddings are generated, and the participation degree of different views is adaptively adjusted, specifically:

[0013] The reverse relationship and complex relationship existing in the structure view are processed using TransH, and structural embedding of the structure view entity and relationship is generated.

[0014] The description information of each entity and relationship in the text view is processed using uSIF, and text embedding of the entity and relationship is generated.

[0015] The embedding representation of the entity and relationship containing rich semantic information is obtained by combining the structure view and the text view through the adaptive attention mechanism.

[0016] Neighborhood aggregation is performed on the small sample knowledge graph, and the neighborhood information of the entity is aggregated by considering that the entity has different common sense information under different relationships.

[0017] By setting common sense semantic constraints, the potential semantic relationship between entities is mined using the common sense view, specifically:

[0018] The constraint of limiting the embedding of entities with the same common sense knowledge in the similar semantic space is set;

[0019] For entities with the same common sense knowledge, the embedding of the tail entity is limited by setting the similar semantic constraint, so that it is positioned in the similar semantic space;

[0020] Constraints are set for entities with different semantics to ensure that their semantics do not diverge;

[0021] For entities with different common sense knowledge, the embedding of the tail entity is limited by setting the different semantic constraint, so that it is positioned in the different semantic space;

[0022] Further, by using the characteristics of the common sense view and the complex relationship, the negative sampling process is supervised in the internal optimization process to generate effective and high-quality negative samples, and the representation of the complex relationship and the hyperplane parameter are learned, specifically:

[0023] The representation of the entity pair and the corresponding relationship is encoded using Transformer to update the general representation of the current small sample relationship;

[0024] Considering the complex relationship, the complex relationship is classified according to whether the tail entity is unique;

[0025] According to the complex relationship, the common sense of the candidate entity is selected, and the common sense is converted into a specific entity;

[0026] The specific entity converted from the common sense and the semantic constraint are used to design the scoring function of the internal optimization process;

[0027] The loss function of the support set in the internal optimization process is defined to update the final representation of the relationship and the hyperplane normal vector;

[0028] updating the representation of the relation and the hyperplane normal vector;

[0029] updating all parameters through an external optimization process to make the model obtain better results, specifically:

[0030] testing the small sample relations of the query set, designing a scoring function and defining the loss function of the external optimization process;

[0031] evaluating the hyperplane parameters in the small sample task of the query set, updating the well-initialized hyperplane parameters and returning to the internal optimization for continuous training;

[0032] Further, in the external optimization process, the common sense view is used to filter out entities that do not conform to common sense knowledge in the candidate entities;

[0033] obtaining the total query loss of all tasks in the meta-training task to minimize the loss function.

[0034] The beneficial technical effects of the present application are: the method proposed in the present application is experimented on two data sets FB15k237-One and NELL-One, in the small sample knowledge graph completion, the present method compared with the existing most advanced baseline method has significant improvement, which is summarized in that the present method can learn the semantic information of entities and relations from multiple views, generate more rich semantic embedding, and more importantly, the present method can effectively utilize the semantic information of different views, especially the semantic brought by the common sense view, thereby mining the potential semantic relationship between entities, solving the overfitting problem of complex relations, and generating high-quality negative triples by using the characteristics of complex relations, and then realizing the improvement of the overall performance of the model. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 A flowchart of a small sample knowledge graph completion method based on multi-view semantic enhancement in the embodiment;

[0036] Figure 2 A common sense semantic constraint example graph in the embodiment. DETAILED DESCRIPTION

[0037] The present application will be further described below in conjunction with the drawings and embodiments;

[0038] The overall flowchart of the present application is shown in Figure 1 as shown:

[0039] The present application comprises three components: a multi-view joint encoding module, an internal and external optimization complex relation processing module, and a multi-view link prediction module.

[0040] The first, a multi-view joint encoding module, combines a structural view and a text view to generate embedded representations of entities and relationships, solves the influence of common sense knowledge implied by entities on triples through an adaptive attention mechanism, and finally aggregates neighborhood information of entities by considering that entities have different common sense information under different relationships.

[0041] The second, an internal-external optimization complex relationship processing module, limits the embedding of the tail entity of a link by common sense semantic constraints, positions embeddings with the same (different) common sense knowledge into the same (different) semantic space (as shown in Figure 2 Then, in the internal optimization process, the characteristics of the common sense view and the complex relationship are utilized to supervise the negative sampling process to generate effective and high-quality negative samples, and to learn the representation of the complex relationship and the hyperplane parameter. In the external optimization process, all parameters are updated to make the model obtain better results.

[0042] The third, a multi-view link prediction module, filters out entities that do not conform to common sense knowledge among candidate entities through a common sense view, and outputs a final candidate entity ranking.

[0043] A small sample knowledge graph completion method based on multi-view semantic enhancement, specifically comprising the following steps:

[0044] Step 1: Use a multi-view joint encoder on a small sample knowledge graph to generate embedded representations of entities and relationships by combining a structural view and a text view;

[0045] Step 1.1: Structural view embedding

[0046] Pre-train the small sample knowledge graph using TransH to map the small sample knowledge graph into a relationship space to obtain initialized embedded representations that can handle inverse relationships and complex relationships. The structural embedding of entities and relationships in the structural view is represented as:

[0047] x es ∈R ds , x rs ∈R ds

[0048] where ds is the vector dimension set by the structural view; x es is the structural embedding of an entity; and x rs is the structural embedding of a relationship.

[0049] Step 1.2: Text view embedding

[0050] Apply uSIF to generate embeddings for the description information of each entity and relationship in the small sample knowledge graph. The text embedding of entities and relationships in the text view is represented as:

[0051] xet ∈R dt , x rt ∈R dt

[0052] where dtis the vector dimension of the text view setting; x et is the text embedding of the entity; x rt is the text embedding of the relation;

[0053] Step 1.3: Information aggregation

[0054] The entity embedding is mapped into the same representation space using the structure view projection matrix M s and the text view projection matrix M t :

[0055] x′ es = M s x es

[0056] x′ et = M t x et

[0057] An adaptive attention mechanism is introduced to enable the model to adaptively adjust the participation degree of the structure view and the text view; at the same time, the representation x e ∈R d of the entity is obtained by combining the information of the structure view and the text view, and the specific process is as follows:

[0058] M e = tanh(W e [x′ es , x′ et ])

[0059] α e = softmax(ω e T M e +b e )

[0060]

[0061] where W e , ω e , b e are learnable parameters, and x e represents the entity embedding combined with the structure view and the text view; is the adaptive attention mechanism coefficient of the structure view; is the adaptive attention mechanism coefficient of the text view;

[0062] The representation of the relation is similar to that of the entity, that is:

[0063]

[0064] Step 1.4: Aggregating neighborhood

[0065] To enhance the embedding representation of each entity in the small sample knowledge graph, the neighborhood information of the entity is aggregated to optimize the entity embedding; specifically, the common sense view is introduced in the first-order neighbor aggregation process of the entity, and it is coupled with x e to obtain the embedding representation of the entity:

[0066] s = W[r; r n ]

[0067]

[0068] x′ e = LeakyReLU(W1x e +W2x c )

[0069] where W is a linear transformation matrix, W1 and W2 are learnable parameters, LeakyReLU is an activation function; r is a small sample relationship; r n is the first-order neighborhood relationship of r; N h is the first-order neighborhood of the head entity h; x en is the neighbor entity contained in the neighborhood of the entity;

[0070] Step 2: Handle complex relationships through internal and external optimization

[0071] Step 2.1: Common sense semantic constraints

[0072] To limit the embedding of entities in similar semantic spaces, entities with the same common sense knowledge x i and x j are constrained:

[0073] R1 = ||x i || + ||x j ||

[0074] A similar semantic constraint is set to limit the embedding of the tail entity, so that it is located in a similar semantic space:

[0075]

[0076]

[0077] where and represent the common sense concepts of entities x i and x j ; x ran embedding representing the relation r;

[0078] The constraint function g1 of the similar semantic constraint is obtained:

[0079] g1=R1+R2

[0080] To ensure that entities x i with different common sense knowledge k The semantic does not diverge, set the constraint D1:

[0081] D1=||x i ||+||x k ||

[0082] The embedding of the tail entity is restricted using the dissimilar semantic constraint, so that it is located in a different semantic space, and the constraint function g2 of the dissimilar semantic constraint is obtained:

[0083]

[0084]

[0085] g2=D1+D2

[0086] Step 2.2: Internal optimization process

[0087] The common sense semantic information is encoded using a Transformer module; specifically, the entity pair and the corresponding relation are regarded as a sequence X=(h, r, t), and for the elements in the sequence, the input is constructed as:

[0088] x 0 =x ele +x pos

[0089] where x el e represents the embedding representation obtained by the multi-view joint encoder, x pos represents the position embedding in the sequence X; x 0 is input into an L-layer Transformer module to encode the sequence X:

[0090] x l =Transformer(x l-1 ), l=1, 2,..., L

[0091] The general representation of each small sample relation r i in the support set S r in the internal optimization process is obtained:

[0092] r′ i =W l x l +bl

[0093] For multiple entity pairs, the specific relation meta is generated by averaging the current small sample relation r i General representation to generate the final relation representation in the current internal optimization process:

[0094]

[0095] Where W l and b l are the weights and biases of the lth layer of the Transformer;

[0096] From the common sense concept set C2 containing all the head entity h and tail entity t common sense of the relation r, the common sense of the candidate entity is selected by converting the common sense into specific entities; The conversion process follows the following distribution:

[0097]

[0098] Where t′ j is the contaminated tail entity obtained by the negative sampling process, t′ i is the contaminated tail entity used in the current conversion process, τ is the sampling temperature that controls the discrimination degree of the model negative sampling, E(·) represents the score function of the model, which is defined as follows:

[0099] h pi = h i -h i P r h i

[0100] t pi = t i -t i P r t i

[0101] E(h i , r, t i ) = ||h pi +r′-t pi || + λ1g1 + λ2g2

[0102] Where h i and t i represent the head entity and tail entity embedding obtained by the multi-view joint encoder, P r is the hyperplane normal vector about the relation r, λ1, λ2 represent the regularization coefficients;

[0103] The loss function of the support set S r in the internal optimization process is defined to update the final representation of the relation r′ and the hyperplane normal vector P r:

[0104]

[0105] where [x] + = max(0, x) is the standard hinge loss function, and γ is the margin; E(h i , r, t i ) denotes the score of the contaminated triplets;

[0106] Updating the representation of the relation and the hyperplane normal vector P r :

[0107]

[0108]

[0109] where l r and l p are the learning rates for updating the representation of the relation and the hyperplane hyperparameters, respectively; is the gradient for updating the representation of the relation; is the gradient for updating the hyperplane normal vector;

[0110] Step 2.3: External optimization procedure

[0111] Test the small sample relations in the query set Q r ; specifically, their scores and losses are:

[0112] h pj = h j - h j P rm h j

[0113] t pj = t j - t j P rm t j

[0114]

[0115]

[0116] where P rm denotes the hyperplane parameters updated by the internal optimization procedure;

[0117] Evaluate the hyperplane parameters P r in the task r , update the well-initialized hyperplane parameters and return to the internal optimization to continue training:

[0118]

[0119] wherein, is the gradient when updating the hyperplane parameters by the external optimization procedure;

[0120] Step 3: Multi-view link prediction

[0121] By filtering out the entities in the candidate entities that do not conform to common sense knowledge under the common sense view; specifically, a tail entity concept set C conforming to common sense knowledge is defined t is:

[0122] C t ={c ti |(c hi , r, c ti )∈C1}

[0123] wherein, for the entity predicted in the external optimization procedure, it is judged whether its common sense knowledge belongs to the set C t only the entity whose common sense satisfies C t is more likely to be a real tail entity;

[0124] Step 4: Model optimization

[0125] The training target is to minimize the sum of query losses of all tasks in the meta-training task, and the loss L is defined as follows:

[0126]

[0127] The method of the application has carried out a large number of experiments based on two real world data sets FB15k237-One and NELL-One. In the small sample completion scene, two widely used evaluation indicators MRR and Hits@N are used to evaluate the performance of the model. The performance of all models on FB15k237-One and NELL-One is shown in Table 2. From the experimental results, it can be seen that the performance of the method of the application is always better than the most advanced baseline on the two data sets. More specifically, compared with the best results of other methods, the performance of the method of the application is improved by 10.0%, 7.9%, 8.0%, and 9.6% in MRR, Hits@10, Hits@5, and Hits@1, respectively, on the 5-shot test data of FB15k237-One, and by 9.2%, 6.6%, 7.6%, and 8.8% on NELL-One, respectively.

[0128] Table 1 Evaluation results on FB15k237-One and NELL-One.

[0129]

[0130]

Claims

1. A small sample knowledge graph completion method based on multi-view semantic enhancement, characterized in that, Specifically comprising the following steps: Step 1: using a multi-view joint encoder on a small sample knowledge graph to generate embedding representations of entities and relationships by combining a structure view and a text view; Step 1.1: structure view embedding Applying TransH to pre-train the small sample knowledge graph, mapping the small sample knowledge graph into a relationship space to obtain initialized embedding representations that can handle reverse relationships and complex relationships; wherein the structure embedding of the entity and the relationship of the structure view is respectively represented as: ； wherein, a vector dimension for a structural view; a structural embedding for an entity; a structural embedding for a relationship; Step 1.2: text view embedding Applying uSIF to generate embedding for the description information of each entity and relationship in the small sample knowledge graph; the text embedding of the entity and the relationship of the text view is respectively represented as: ； wherein, a vector dimension for the text view; a text embedding for the entity; a text embedding for the relationship; Step 1.3: information aggregation Using a structural view projection matrix and a textual view projection matrix mapping the structural embedding of an entity and the textual embedding of an entity into the same representation space: ； ； An adaptive attention mechanism is introduced, so that the model can adaptively adjust the participation degree of the structural view and the text view; meanwhile, the information of the structural view and the text view is combined to obtain the representation of the entity The specific process is as follows: ； ； ； wherein, are learnable parameters, respectively, denotes an entity embedding combining the structural view and the textual view; are adaptive attention mechanism coefficients for the structural view; are adaptive attention mechanism coefficients for the textual view; Representation of the relationship: ； Step 1.4: aggregated neighborhood To enhance the embedding representation of each entity in the small sample knowledge graph, the neighborhood information of the entity is aggregated to optimize the entity embedding; specifically, the common sense view is introduced in the first-order neighbor aggregation process of the entity, and it is combined with the embedding representation of the entity: ； ； ； wherein, is a linear transformation matrix, and are learnable parameters, is an activation function; is an entity first-order neighborhood contained in the entity; Step 2: processing complex relationships through internal and external optimization; Step 3: multi-view link prediction; By common sense view to filter out the candidate entity in the entity does not conform to common sense knowledge; Specifically, define the tail entity concept set conforming to common sense knowledge Is: ； wherein, for an entity predicted in the outer optimization process, it is determined whether the common sense knowledge of the entity belongs to the set ; Step 4: model optimization The training objective is to minimize the sum of the query losses of all tasks in the meta-training tasks, loss is defined as follows:

2. The method of claim 1, wherein the multi-view semantic enhancement based small sample knowledge graph completion method is characterized by, The specific implementation process of step 2 is as follows: Step 2.1: common sense semantic constraint To limit the embedding of entities in a similar semantic space, entities having the same common sense knowledge are constrained with the same common sense knowledge. ； A similar semantic constraint is set to limit the embedding of the tail entity, so that it is located in a similar semantic space: ； ； wherein, with representing common sense concepts of entities and embeddings of . Constraint functions that obtain similar semantic constraints : ； To ensure entities with different common sense knowledge With semantics do not diverge, set constraints : ； The embedding of the tail entity of the link is restricted using the distinct semantic constraint, so as to be positioned in different semantic spaces, to obtain a constraint function of the distinct semantic constraint : ； ； ； Step 2.2: internal optimization process The common sense semantic information is encoded using a Transformer module; specifically, entity pairs and corresponding relations are treated as sequences For elements in the sequence, the input is structured as: ； in, This represents the embedded representation obtained by the multi-view joint encoder. Represents a sequence Position embedding in; Input to In the Transformer module of the layer, for the sequence Encode: ； Obtaining a support set for an internal optimization process of each subsample relationship representation: ； For multiple entity pairs, a specific relation meta is generated by averaging the current small sample relation representation to generate the final relation representation in the current internal optimization process: ； wherein, and are the weights and biases of the layer. From the containment relationship To all head entities And tail entities Commonsense concepts set of common sense Commonsense to select candidate entities by converting common sense to specific entities; the process of conversion follows the following distribution: ； wherein, is the contaminated tail entity resulting from the negative sampling process, is the contaminated tail entity used in the current conversion process, is the sampling temperature controlling the model negative sampling discrimination, denotes the score function of the model, defined as follows: ； ； ； wherein, and denote head and tail entity embeddings obtained by a multi-view joint encoder, is a hyperplane normal vector with respect to the relation denotes a regularization coefficient;​ Defining a support set for an internal optimization process Loss function for updating a final representation of a relationship and hyperplane normal vector : ； wherein, is the standard hinge loss function, is the outer margin; denotes the score of the contaminated triple. Representations of relationships and hyperplane normal vectors Update: ； ； where, and are the learning rates for updating the hyperplane parameters and the bias term, respectively; is the gradient for updating the representation of the relation; is the gradient for updating the hyperplane normal vector; Step 2.3: external optimization process Test on small sample relations in query set ; specifically, their scores and losses are: ； ； ； ； wherein, represents the hyperplane parameters obtained after updating by the internal optimization process; In the task of evaluating hyperplane parameters , updating well-initialized hyperplane parameters and returning to the inner optimization to continue training: ； wherein, is the gradient for updating the hyperplane parameters by the external optimization process.

Images