Time sequence knowledge graph reasoning method based on super-relation subgraph

Abstract

The invention discloses a time sequence knowledge graph reasoning method based on a super-relation subgraph, which comprises the following steps of: firstly, constructing a time sequence knowledge graph reasoning model based on the super-relation subgraph and training, and then realizing time sequence knowledge graph reasoning by utilizing the trained model. According to the method, the constraint of the relative position of the relation is modeled through the double-layer graph attention network time sequence knowledge reasoning model, the problem of'information isolated island 'existing in relation evolution is solved, and the accuracy of an entity prediction task and a relation prediction task in future time is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of deep learning, and more specifically relates to a time sequence knowledge graph reasoning method based on a super-relation subgraph. BACKGROUND

[0002] Under the background of rapid development of artificial intelligence, as a key technology for knowledge structured representation, knowledge graph has played a core role in semantic understanding, intelligent interaction and other scenarios. Although knowledge graph has made significant progress in scale in recent years, it usually covers tens of millions of entities and hundreds of millions of triple information, but there are still many implicit knowledge that cannot be directly embodied in it. This limitation leads to the two major problems of sparsity and incompleteness of knowledge graph. Taking Freebase knowledge base as an example, although the number of person entities is more than 3 million, about 75% of the person entities lack nationality information, and 71% lack birthplace, a key detail. The above-mentioned sparsity and incompleteness problems seriously limit the effectiveness of knowledge graph in practical applications. In order to solve this challenge, knowledge graph completion (KGC) technology is proposed and developed. Knowledge graph completion uses known explicit knowledge to infer implicit knowledge not explicitly recorded in the graph, thereby making up for the incompleteness of the knowledge graph.

[0003] Traditional knowledge graph describes static knowledge in the form of triple, i.e. the case where entity relationship does not change with time, and is usually used to describe static graphs whose facts do not change with time. However, in the real world, a large amount of knowledge has time characteristics, such as the occurrence of events and the change of entity attributes, which are closely related to time.

[0004] In order to more accurately describe and process these dynamically changing knowledge, time sequence knowledge graph is proposed. In order to capture the dynamically evolving knowledge, researchers innovatively embed time constraints into triple structure to form four-tuple structure (head entity, relationship, tail entity, timestamp) or (entity, attribute, attribute value, timestamp), thereby supporting time sequence reasoning tasks. Similar to traditional static knowledge graph, time sequence knowledge graph also has the problems of incompleteness and missing. In addition, time sequence knowledge graph also involves the prediction of four-tuple under future timestamp. Compared with static knowledge graph reasoning task, time sequence knowledge graph reasoning task needs to consider additional time sequence information and integrate it into knowledge representation process, which makes time sequence knowledge graph reasoning task more challenging than static knowledge graph reasoning. Although there is relatively mature research on static knowledge graph reasoning, although time sequence knowledge graph also has wide application prospect, but the related research is relatively scarce, and it is meaningful to carry out further research.

[0005] Currently, temporal knowledge graph reasoning tasks can generally be divided into three methods: translation-based, tensor-based, and neural network-based. With the development of deep learning, neural network-based methods are becoming the mainstream.

[0006] Know-Evolve[2] first proposed in 2017 to combine deep learning methods to learn nonlinear entity representations that evolve over time. It models the occurrence of facts as a multivariate point process, defines a bilinear relationship score that captures the multi-relational interactions between entities, adjusts the intensity function of the point process, and dynamically updates the entity representation using the intensity function under different facts. This model, for the first time, achieves the ability to predict the time when a fact may occur while predicting its occurrence.

[0007] DyRep[3] proposed a dual-time-scale deep time point process model to model the global and local evolution of dynamic graph networks, improving the ability to capture the intertwined dynamics of the observation process. The RE-GCN[4] model uses a stacked multiple graph convolutional network layers to effectively model the correlation features between synchronous facts, and introduces a gated recurrent structure to efficiently extract long-term temporal features from historical facts through a temporal information transmission mechanism. At the same time, it uses residual connection design to alleviate the gradient decay problem in deep networks. Gao et al. proposed SToKE[5] based on the pre-trained language model BERT, constructing a hierarchical event evolution tree, transforming the entity relationships and their temporal evolution paths in the knowledge graph into serialized event representations, thereby integrating structural associations with temporal dynamic features and achieving joint modeling of structural and temporal features.

[0008] Although the above model has greatly improved the accuracy of temporal knowledge graph reasoning tasks by combining deep learning technology, it still has insufficient modeling of relative position constraints of relationships and "information islands" in relationship evolution.

[0009] The inadequate modeling of relative position constraints in relationships refers to the fact that, in a knowledge graph, two adjacent relationships can have four different relative positions relative to a central shared entity, depending on the orientation of the head and tail entities. Previous models have rarely modeled this relative position of relationships. Furthermore, the "information island" problem of relationship evolution refers to the fact that information about the dynamic evolution of relationships is not transmitted across time, creating "information islands." Summary of the Invention

[0010] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a temporal knowledge graph reasoning method based on a super-relationship subgraph. The constraints of the relative position of the relationship are modeled through a double-layer graph attention network temporal knowledge reasoning model, and the "information island" problem existing in the evolution of the relationship is solved, thereby improving the accuracy of future time entity prediction tasks and relationship prediction tasks.

[0011] To achieve the above-mentioned purpose of the invention, the present invention provides a temporal knowledge graph reasoning method based on a super-relationship subgraph, which is characterized by comprising the following steps:

[0012] (1) Download and initialize the dynamic knowledge graph dataset;

[0013] (2) Constructing a super-relationship subgraph;

[0014] (3) Construct a temporal knowledge graph reasoning model based on super-relationship subgraph;

[0015] (4) Training a temporal knowledge graph reasoning model based on a super-relationship subgraph;

[0016] (5) Temporal knowledge graph reasoning based on super-relationship subgraphs;

[0017] The object of the invention of the present invention is achieved like this:

[0018] The present invention is based on a temporal knowledge graph reasoning method based on a super-relationship subgraph. First, a temporal knowledge graph reasoning model based on a super-relationship subgraph is constructed and trained, and then the trained model is used to realize temporal knowledge graph reasoning. This temporal knowledge reasoning model uses a double-layer graph attention network to model the constraints on the relative position of relationships, solves the "information island" problem existing in relationship evolution, and improves the accuracy of future time entity prediction tasks and relationship prediction tasks.

[0019] At the same time, the temporal knowledge graph reasoning method based on the super-relationship subgraph of the present invention also has the following beneficial effects:

[0020] (1) By adding the definition and training of super-relationship subgraphs, the relative position constraints between relations are effectively learned, and information capture based on the graph attention mechanism is achieved at the historical subgraph level and the super-relationship subgraph level.

[0021] (2) By adding an LSTM-based relationship evolution update mechanism, the LSTM layer that captures relationship evolution information is designed into the relationship update mechanism, which successfully alleviates the temporal redundancy problem of the GRU-LSTM cascade structure and the "information island" problem of relationship evolution, making the MRR index and Hits@k index of the knowledge reasoning task significantly improved compared with baseline models such as RE-GCN and Trend. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of the temporal knowledge graph reasoning method based on the super-relationship subgraph of the present invention;

[0023] Figure 2 It is a schematic diagram of mapping the knowledge graph into a super-relationship subgraph;

[0024] Figure 3 is a super-relation sub-graph based temporal knowledge graph reasoning model graph;

[0025] Figure 4 is a historical sub-graph level attention module structure diagram;

[0026] Figure 5 is a relation updating mechanism module structure diagram;

[0027] Figure 6 is a super-relation sub-graph level attention module structure diagram. DETAILED DESCRIPTION

[0028] The specific embodiments of the present application will be described below with reference to the accompanying drawings so that those skilled in the art can better understand the present application. It should be particularly noted that in the following description, when detailed description of known functions and designs may obscure the main content of the present application, these descriptions will be omitted here.

[0029] EMBODIMENT

[0030] In this embodiment, as shown in the figure, Figure 1 a super-relation sub-graph based temporal knowledge graph reasoning method of the present application, characterized in that it comprises the following steps:

[0031] (1) Download and initialize the dynamic knowledge graph dataset;

[0032] Thus, the dynamic knowledge graph dataset G = {G1, G2,..., G T} is downloaded from the ICEWS18 dataset, T represents the number of discrete time samples, G t represents the knowledge graph at time t;

[0033] Each knowledge graph at each time consists of a number of triples, each triple is represented as (head entity s, relation r, tail entity o), for example: triple (Mr. Li, award, Nobel Peace Prize);

[0034] For each knowledge graph G t at each time, word2vec-based word embedding processing is performed, so that each entity and relation in G t is mapped into a multi-dimensional dense vector, wherein the head entity and the tail entity are uniformly denoted as the relation is denoted as N is the number of entities, d e is set to 200, M is the number of relations, d r is set to 200;

[0035] (2) Construct a super-relation sub-graph;

[0036] Iterate Gt The relationship between all triples in Figure 2 As shown, each relationship is mapped into a twin super-relationship subgraph HG t A node in , and define four super relations hr as super relation subgraph HG t edge;

[0037] According to the above definition, the super-relation subgraph at time t is also composed of several triples, each of which is represented by (head relation r o , super relation hr, tail relation r s ), the four super relations are hr = {oo, ss, os, so}, where oo represents two relations r o and r s The intermediate entity is the head entity in the triples corresponding to the two relations; ss represents the two relations r o and r s The intermediate entity is the tail entity in both triples; os represents the intermediate entity in the head relation r o The corresponding triple is the head entity, and the tail relation r s The corresponding triple is the tail entity; so the middle entity is in the head relation r o The triple is the tail entity, and the tail relation r s The triple is the head entity;

[0038] In this embodiment, taking the triples (Mr. Li, award, Nobel Peace Prize) and (Mr. Li, position, doctor) as examples, the corresponding super-relationship triples in the super-relationship subgraph are (award, oo, position).

[0039] Therefore, this embodiment constructs a one-to-one corresponding super-relationship subgraph set based on the historical subgraph set of the ICEWS18 dataset.

[0040] (3) Construct a temporal knowledge graph reasoning model based on super-relationship subgraph;

[0041] In this embodiment, if Figure 3 As shown, the temporal knowledge graph reasoning model based on the super-relation subgraph includes an encoder module and a decoder module;

[0042] Among them, Figure 3 As shown in the left half of , the encoder module includes a historical subgraph level attention module, a relation update mechanism module, and a super-relation subgraph level attention module;

[0043] like Figure 4 As shown, in the attention module at the historical subgraph level, the entity at the previous moment is embedded into E t-1 and relational embedding R t-1And the multi-dimensional dense vector at time t is input into the two-layer GAT network with improved attention weight calculation method, and the aggregate embedding of historical subgraph structure information is obtained The improved attention weight is where e i and e j are entity i and entity j, r ij is the relationship between entity i and entity j, W0 and W1 are weight matrices, α0 is the learning function; then the aggregation is embedded and the entity embedding E at the previous moment t-1 Enter a GRU network as input and get the entity embedding E at time t t ;

[0044] like Figure 5 As shown, in the relationship update mechanism module, the aggregation at the current moment is embedded in All entities directly connected to relation r are embedded Perform the average pooling operation together, and then splice it with the initial value R0 of the relation embedding to obtain the relation evolution embedding R′ at time t for updating t ; Then a long short-term memory network LSTM module is used to embed the relationship evolution at time t into R′ t The temporal information of Finally, by embedding the relationship at time t-1 and the relationship evolution with temporal information Concatenate to get the updated relation embedding

[0045] In the super-relation subgraph level attention module, And, the relational dense vector is input into the two-layer GAT network with improved attention weights to obtain the structural information aggregation embedding of the super-relational subgraph The improved attention weight is Among them, e nm Head relation r n , tail relation r m The common central entity, W0 and W1 are weight matrices, α0 is the learning function; then the aggregate is embedded The relationship with the previous moment is embedded in R t-1 Input into the GRU network to get the relationship embedding R at time t t .

[0046] (4) Training a temporal knowledge graph reasoning model based on a super-relationship subgraph;

[0047] (4.1),capturing historical subgraph level information;

[0048] (4.1.1), the knowledge graph dataset G = {G1, G2, ..., G T Knowledge graph G at each moment t As a historical subgraph, select k historical subgraphs from the dataset G and input them into the temporal knowledge graph reasoning model based on the super-relationship subgraph;

[0049] Using historical subgraphs to improve the graph attention network module to obtain the aggregate embedding of historical subgraph structural information at time t

[0050]

[0051] Among them, E t-1 and R t-1 are the entity embedding and relation embedding of the previous timestamp respectively, and is a multi-dimensional dense vector;

[0052] (4.1.2) Use the gated recurrent unit GRU to capture the temporal information of the historical subgraph and obtain the entity embedding E at time t t ;

[0053]

[0054] (4.2),relation embedding update;

[0055] (4.2.1), computing relational evolution embedding;

[0056] Embed the aggregation of the current moment and the embeddings of all entities directly connected to relation r at the current moment Perform average pooling operation together with the initial value of relation embedding Splice and get the relationship evolution embedding R′ at time t for updating t ;

[0057]

[0058] Among them, MP() is the average pooling operation;

[0059] (4.2.2), capturing the temporal information of relational evolution embedding;

[0060] Capturing relational evolution using long short-term memory (LSTM) embeddings in R t ′’s temporal information, and obtains the relational evolution embedding with temporal information

[0061]

[0062] in, is the relation evolution embedding with temporal information at time t-1, C t and C t-1 are the cell states of the LSTM network at time t and time t-1 respectively;

[0063] (4.3), Update relation embedding;

[0064] Embed the relationship at time t-1 into R t-1 and the relational evolution embedding at time t Perform splicing to obtain the updated relation embedding;

[0065]

[0066] Among them, || represents the splicing operation, Embedding for the updated relation;

[0067] (4.4), information capture at the super-relationship subgraph level;

[0068] (4.4.1) Using the hyper-relational subgraph to improve the graph attention network Hyper-GAT to obtain the aggregate embedding of the hyper-relational subgraph structural information at time t

[0069]

[0070] (4.4.2) Use the gated recurrent unit GRU to capture the temporal information of the super-relationship subgraph and obtain the relationship embedding R at time t t ;

[0071]

[0072] Among them, R t-1 The relation embedding for the moment of the previous timestamp;

[0073] (4.5) Embed the entity into E t and relational embedding R t Input Conv-TransE network to calculate the probability vector of each candidate entity in the entity prediction task And the probability vector of each candidate relationship in the relationship prediction task

[0074] (4.6) Calculate the total loss after this round of training;

[0075] (4.6.1) Calculate the loss of entity prediction task and relationship prediction task;

[0076]

[0077] in, and The loss for entity prediction tasks and relation prediction tasks, represents all fact triples at timestamp t+1, and They represent the label vector of the entity prediction task and the label vector of the relationship prediction task at the t+1th time, respectively. In this embodiment, k is 6;

[0078] (4.6.2), calculate the overall loss;

[0079]

[0080] Among them, λ1 and λ2 are weighting coefficients, which are 0.7 and 0.3 respectively;

[0081] (4.7), the overall loss Send it to the Adam optimizer to optimize the vector initialization method in step (1); then repeat step (4) until the overall loss function of the model is Convergence means completing the model training task;

[0082] (5) Temporal knowledge graph reasoning based on super-relationship subgraphs;

[0083] (5.1) Input the query triple to be completed at time t+1 into the dynamic knowledge graph reasoning model to obtain the corresponding embedding vector;

[0084] (5.2) If the query triple to be completed is an entity prediction, the embedding vector is input into the Conv-TransE network, and then the probability of the candidate entities is calculated according to step (4.5), and the entity with the highest probability is selected as the output;

[0085] (5.3) If the query triple to be completed is a relation prediction, the embedding vector is input into the Conv-TransE network, and then the probability of the candidate relations is calculated according to step (4.5), and the relation with the highest probability is selected as the output.

[0086] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

Claims

1. A temporal knowledge graph reasoning method based on a super-relationship subgraph, characterized in that: The following steps are included: (1) Download and initialize the dynamic knowledge graph dataset; Download dynamic knowledge graph dataset G={G1,G2,...,G T }, T represents the number of discrete time samples, G t Represents the knowledge graph at time t; The knowledge graph at each moment consists of several triples, each of which is represented by (head entity s, relation r, tail entity o); For each moment of knowledge graph G t Based on word2vec word embedding processing, G t Each entity and relationship in is mapped into a multi-dimensional dense vector, where the head entity and the tail entity are uniformly recorded as The relationship is recorded as N is the number of entities, d e is the entity vector dimension, M is the number of relations, d r is the dimension of the relationship vector; (2) Constructing a super-relationship subgraph; Traverse G t The relationship of all triples in , mapping each relationship into a twin super-relationship subgraph HG t A node in , and define four super relations hr as super relation subgraph HG t edge; According to the above definition, the super-relation subgraph at time t is also composed of several triples, each of which is represented by (head relation r o , super relation hr, tail relation r s ), the four super relations are hr = {oo, ss, os, so}, where oo represents two relations r o and r s The intermediate entity is the head entity in the triples corresponding to the two relations; ss represents the two relations r o and r s The intermediate entity is the tail entity in both triples; os represents the intermediate entity in the head relation r o The corresponding triple is the head entity, and the tail relation r s The corresponding triple is the tail entity; so the middle entity is in the head relation r o The triple is the tail entity, and the tail relation r s The triple is the head entity; (3) Construct a temporal knowledge graph reasoning model based on super-relationship subgraph; (4) Training a temporal knowledge graph reasoning model based on a super-relationship subgraph; (4.1),capturing historical subgraph level information; (4.1.1), the knowledge graph dataset G = {G1, G2, ..., G T Knowledge graph G at each moment t As a historical subgraph, select k historical subgraphs from the dataset G and input them into the temporal knowledge graph reasoning model based on the super-relationship subgraph; Using historical subgraphs to improve the graph attention network module to obtain the aggregate embedding of historical subgraph structural information at time t Among them, E t-1 and R t-1 are the entity embedding and relation embedding of the previous timestamp respectively, and is a multi-dimensional dense vector; (4.1.2) Use the gated recurrent unit GRU to capture the temporal information of the historical subgraph and obtain the entity embedding E at time t t ; (4.2),relation embedding update; (4.2.1), computing relational evolution embedding; Embed the aggregation of the current moment and the embeddings of all entities directly connected to relation r at the current moment Perform average pooling operation together with the initial value of relation embedding Splice and get the relationship evolution embedding R′ at time t for updating t ; Among them, MP() is the average pooling operation; (4.2.2), capturing the temporal information of relational evolution embedding; Capturing relational evolution using long short-term memory (LSTM) embeddings in R t ′’s temporal information, and obtains the relational evolution embedding with temporal information in, is the relation evolution embedding with temporal information at time t-1, C t and C t-1 are the cell states of the LSTM network at time t and time t-1 respectively; (4.3), Update relation embedding; Embed the relationship at time t-1 into R t-1 and the relational evolution embedding at time t Perform splicing to obtain the updated relation embedding; Among them, || represents the splicing operation, Embedding for the updated relation; (4.4), information capture at the super-relationship subgraph level; (4.4.1) Using the hyper-relational subgraph to improve the graph attention network Hyper-GAT to obtain the aggregate embedding of the hyper-relational subgraph structural information at time t (4.4.2) Use the gated recurrent unit GRU to capture the temporal information of the super-relationship subgraph and obtain the relationship embedding R at time t t ; Among them, R t-1 The relation embedding for the moment of the previous timestamp; (4.5) Embed the entity into E t and relational embedding R t Input Conv-TransE network to calculate the probability vector of each candidate entity in the entity prediction task And the probability vector of each candidate relationship in the relationship prediction task (4.6) Calculate the total loss after this round of training; (4.6.1) Calculate the loss of entity prediction task and relationship prediction task; in, and The loss for entity prediction tasks and relation prediction tasks, represents all fact triples at timestamp t+1, and They represent the label vector of the entity prediction task and the label vector of the relationship prediction task at time t+1 respectively; (4.6.2), calculate the overall loss; Among them, λ1,λ2 are weighting coefficients; (4.7), the overall loss Send it to the Adam optimizer to optimize the vector initialization method in step (1); then repeat step (4) until the overall loss function of the model is Convergence means completing the model training task; (5) Temporal knowledge graph reasoning based on super-relationship subgraphs; (5.1) Input the query triple to be completed at time t+1 into the dynamic knowledge graph reasoning model to obtain the corresponding embedding vector; (5.2) If the query triple to be completed is an entity prediction, the embedding vector is input into the Conv-TransE network, and then the probability of the candidate entities is calculated according to step (4.5), and the entity with the highest probability is selected as the output; (5.3) If the query triple to be completed is a relation prediction, the embedding vector is input into the Conv-TransE network, and then the probability of the candidate relations is calculated according to step (4.5), and the relation with the highest probability is selected as the output.

2. A temporal knowledge graph reasoning method based on a super-relationship subgraph according to claim 1, characterized in that: The temporal knowledge graph reasoning model based on the super-relationship subgraph includes an encoder module and a decoder module; The encoder module includes a history subgraph level attention module, a relationship update mechanism module and a super-relationship subgraph level attention module; In the historical subgraph level attention module, the entity of the previous moment is embedded into E t-1 and relational embedding R t-1 The multi-dimensional dense vector at time t is input into the GAT network with improved attention weight calculation method, and the aggregate embedding of historical subgraph structure information is obtained. The improved attention weight is where e i and e j are entity i and entity j, r ij is the relationship between entity i and entity j, W0 and W1 are weight matrices, α0 is the learning function; then the aggregation is embedded and the entity embedding E at the previous moment t-1 Enter a GRU network as input and get the entity embedding E at time t t ; In the relationship update mechanism module, the aggregation at the current moment is embedded All entities directly connected to relation r are embedded Perform the average pooling operation together, and then splice it with the initial value R0 of the relation embedding to obtain the relation evolution embedding R′ at time t for updating t ; Then a long short-term memory network LSTM module is used to embed the relationship evolution captured at time t into R′ t The temporal information of Finally, by embedding the relationship at time t-1 and the relationship evolution with temporal information Concatenate to get the updated relation embedding In the super-relation subgraph level attention module, And, the relational dense vector is input into the GAT network with improved attention weight to obtain the structural information aggregation embedding of the super relational subgraph The improved attention weight is Among them, e nm Head relation r n , tail relation r m The common central entity, W0 and W1 are weight matrices, α0 is the learning function; then the aggregate is embedded The relationship with the previous moment is embedded in R t-1 Input into the GRU network to get the relationship embedding R at time t t .

Images