KR20180092194A - Method and system for embedding knowledge gragh reflecting logical property of relations, recording medium for performing the method - Google Patents

Abstract

A method and system for embedding a knowledge graph reflecting a logical attribute, and a recording medium for performing the same are disclosed. The method for embedding a knowledge graph reflecting a logical attribute according to the present invention includes the steps of: collecting a knowledge triple and assigning at least one role of a head role and a tail role to an entity constituting the knowledge triple; mapping a head entity which is the entity to which the head role is assigned, in a first mapping matrix, and mapping a tail entity which is the entity to which the tail role is assigned, in a second mapping matrix; calculating a score function using the first mapping matrix and the second mapping matrix; and learning the knowledge triple using the calculated score function. Accordingly, the present invention can provide a translation-based knowledge graph in which meaning between the entities is effectively reflected.

Description

TECHNICAL FIELD The present invention relates to a knowledge graph embedding method and system that reflects a logical attribute, and a recording medium for performing the same, and a recording medium for performing the same. BACKGROUND ART [0002]

The present invention relates to a knowledge graph embedding method and system reflecting a logical attribute, and a recording medium for performing the same. More particularly, the present invention relates to a knowledge graph embedding method and system for embedding a knowledge- A graph embedding method and system, and a recording medium for performing the same.

Graphing and expressing knowledge is one of the most effective ways for a computer device to utilize human knowledge. However, due to the scarcity of knowledge graphs, there is a limit to using knowledge graphs in real applications. Therefore, as a solution to overcome the scarcity, studies are being conducted on techniques for completing a knowledge graph.

Among them, the most promising way to complete the knowledge graph is to embed the graph in a contiguous vector space of low dimension. This method learns the vector representation of the knowledge graph and measures the probabilities or associations of specific knowledge (entities) in the graph by logarithmic computation of the vector space. That is, entities and relations are represented as vectors, and relationships are treated as operators that translate entities to different locations in the embedding space. Therefore, new knowledge can be found from the vector space by finding knowledge instances with high probability.

Recently, a technique for completing a knowledge graph using translation based models among various conventional knowledge embedding models has been studied. TransE (Lin et al., 2015) and TransD (Ji et al., 2015), TransH (Wang et al., 2014) ).

TransE gives h, t, r (t) by forcing a vector of t equal to the sum of the vectors of h and r, given a knowledge triple (h, r, t) consisting of a relation r and two entities . TransH includes all relations in a single vector space, while TransH and TransR are techniques that each relationship regards as having its own embedded space. On the other hand, Ji et al. (2015), we found that a single relationship or a single entity generally has several types, thus suggesting TransD that allows a mapping matrix of entities and relationships.

However, conventional techniques related to translation based models ignore the logical properties of the relationship. This means that the transitional and symmetric relationships between the entities lose transitivity and symmetricity in the vector space created by the translation based model, And the new knowledge that is affected can not be completed.

In most knowledge graphs, fulfillment relationships or symmetry relationships are common. In a database in which conventional knowledge triples are stored, about 20% of the triples have a transitive or symmetric relationship. Therefore, there is a need for a technique to complete the knowledge graph by preserving the logical properties of the relationship.

Korean Patent Publication No. 10-2016-0064826 Korean Patent Publication No. 10-2015-0095577

One aspect of the present invention is a knowledge graph embedding method and system that reflects a logical attribute capable of embedding a translation-based knowledge graph preserving logical attributes such as transitivity and symmetry between entities by assigning one or more roles to an entity, And a recording medium.

The technical problem of the present invention is not limited to the technical problems mentioned above, and other technical problems which are not mentioned can be understood by those skilled in the art from the following description.

The knowledge graph embedding method according to an embodiment of the present invention includes collecting knowledge triples and assigning at least one of a head role and a tail role to an entity constituting the knowledge triple, Mapping a head entity, which is a granted entity, to a first mapping matrix, mapping a tail entity, which is an entity granted a tail role, to a second mapping matrix, calculating a score function using the first mapping matrix and the second mapping matrix, And learning the knowledge triple using the calculated score function.

The knowledge triple may consist of two entities and a relation indicating the association of the two entities.

The score function may be a function that calculates the sum vector of the head entity vector and the relation vector and represents the absolute value of the difference value between the sum vector and the tail entity vector.

The step of learning the knowledge triple comprises: a link prediction step of predicting a missing entity if there is an entity missing in the collected knowledge triple; and a triple classifying step of determining whether the collected knowledge triple is represented without error by the relation Step < / RTI >

The link prediction step predicts a missing entity using the score function, and the triple classification step may determine whether the knowledge triple is represented without error by comparing the result value of the score function with a reference value.

The mapping may include mapping the specific entity to both the first mapping matrix and the second mapping matrix if the particular entity is granted both the head role and the tail role.

The knowledge graph embedding system in which the logical attribute according to an embodiment of the present invention is reflected includes a role assigning unit that acquires a knowledge triple and assigns at least one role of a head role or a tail role to an entity constituting the knowledge triple, A mapping unit for mapping a head entity, which is an entity to which a role is assigned, to a first mapping matrix, and a tail entity, which is an entity granted a tail role, to a second mapping matrix; And a learning unit for learning the knowledge triple using the calculated score function.

The knowledge triple may consist of two entities and a relation indicating the association of the two entities.

The score function may be a function that calculates the sum vector of the head entity vector and the relation vector and represents the absolute value of the difference value between the sum vector and the tail entity vector.

The learning unit may predict a missing entity or determine whether the collected knowledge triple is represented without error by the relation, if there is a missing entity in the collected knowledge triple.

The mapping unit may map the specific entity to both the first mapping matrix and the second mapping matrix when the specific entity is given both the head role and the tail role.

Further, it may be a computer-readable recording medium on which a computer program is recorded, for providing a knowledge graph embedding method and system in which logical attributes are reflected.

According to one aspect of the present invention, it is possible to embed at least one role to an entity, thereby embedding a knowledge graph preserving logical attributes such as transitivity and symmetry between entities, Knowledge graphs can be provided.

1 is a diagram showing a schematic configuration of a knowledge graph embedding system in which logical attributes are reflected according to an embodiment of the present invention.
2 is a diagram illustrating an example of translation-based embedding according to the prior art.
FIGS. 3 to 4 are diagrams showing an example of a vector expression in which logical attributes are preserved by the mapping unit of FIG. 1. FIG.
5 to 11 are diagrams showing the performance results of the system of FIG.
FIG. 12 is a flowchart illustrating a method of embedding a knowledge graph reflecting a logical attribute according to an embodiment of the present invention. Referring to FIG.

The following detailed description of the invention refers to the accompanying drawings, which illustrate, by way of illustration, specific embodiments in which the invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention. It should be understood that the various embodiments of the present invention are different, but need not be mutually exclusive. For example, certain features, structures, and characteristics described herein may be implemented in other embodiments without departing from the spirit and scope of the invention in connection with an embodiment. It is also to be understood that the position or arrangement of the individual components within each disclosed embodiment may be varied without departing from the spirit and scope of the invention. The following detailed description is, therefore, not to be taken in a limiting sense, and the scope of the present invention is to be limited only by the appended claims, along with the full scope of equivalents to which such claims are entitled, if properly explained. In the drawings, like reference numerals refer to the same or similar functions throughout the several views.

Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the drawings.

FIG. 1 is a block diagram showing a schematic configuration of a knowledge graph embedding system 1000 in which logical attributes are reflected according to an embodiment of the present invention.

Specifically, the knowledge graph embedding system 1000 incorporating the logical attribute according to the present embodiment includes a role assigning unit 10, a mapping unit 20, a calculating unit 30, and a learning unit 40.

The role assignment (10) can assign a role to an entity.

An entity is a unit of meaningful information. In a file processing system, a record or information constituting a single piece of data corresponds to a single entity. The relationship between entities is called a relationship (relationship), and a map representation of a relationship between a plurality of entities can be defined as a knowledge graph.

The role assignment (10) can assign an entity a head role or a tail role. Hereinafter, an entity to which a head role is assigned is defined as a head entity, and an entity to which a tail role is assigned is defined as a tail entity.

The role assignment 10 may collect knowledge triples from an external database and assign at least one role to each entity included in the collected knowledge triple. That is, one entity may be granted two roles. For example, the role assignment (10) may grant a first entity a role of a head, a second entity a role of a tail, and a third entity may assign a role of both a head and a tail. Accordingly, the first entity may be operated as a head entity, and the second entity may be operated as a tail entity. While the first entity and the second entity perform a single role, the third entity may perform both the roles of both the head entity and the tail entity depending on the situation.

The mapping unit 20 can map the knowledge triple to the vector space. That is, the mapping unit 20 can express the association between the head entity h and the tail entity t using the relation r. The mapping unit 20 maps (h, r, t) constituting the knowledge triple to the vector space, and thus the knowledge triple can be expressed in the form of a vector.

At this time, the mapping unit 20 may map entity vectors to different mapping matrices according to roles assigned to the entity vectors. Specifically, the mapping unit 20 may map the head entity to the first mapping matrix, and map the tail entity to the second mapping matrix. That is, the head entity and the tail entity are mapped to different mapping matrices, and the same mapping matrix can be projected only to entities assigned the same role.

The mapping unit 20 may embed a knowledge graph in which logical attributes are stored in a vector space using an entity to which at least one role is assigned. That is, according to the present invention using an entity to which a role is assigned, the entities constituting the knowledge graph can be represented by a transitive relation or a symmetric relation. The specific function of the mapping unit 20 will be described later with reference to FIG. 2 to FIG.

The calculating unit 30 may calculate a score function for an entity vector mapped to different mapping matrices. The score function, also called an error function, is a function that is a measure for determining whether a knowledge triple projected in a vector space is correctly represented.

For example, the score function disclosed in the prior art TransE (Bordes et al., 2013) is as follows.

,

이다. 여기서,

은 n개의 엔티티들로 구성된 엔티티 집합을 의미한다.In a single embedding space, h denotes a head entity vector, t denotes a tail entity vector, r denotes a relation,

,

to be. here,

Is an entity set consisting of n entities.

As described above, TransE is obtained by forcing a vector of t equal to the sum of the vectors of h and r, given a knowledge triple (h, r, t) composed of a relation r and two entities h, t, and r. That is, the score function means an absolute value obtained by subtracting the vector t from the sum of the vector h and the relation vector r, and the closer the value is to 0, the more correctly the two entities h and t are represented by the relation r.

The calculating unit 30 can calculate the score function in which the logical attribute is preserved using the score function disclosed in the prior art. According to the present invention, the role assignment (10) assigns at least one role (head or tail) to the entity, and the mapping unit (20) maps the entity assigned another role to another mapping matrix, transitivity and symmetricity of a knowledge graph with preserved logical properties.

)와 테일 공간 매핑 행렬(M_t,

)을 이용할 수 있다. 이에 따라, 산출부(30)는 lppTransE의 스코어 함수를 아래와 같이 산출할 수 있다.Thus, on logical property preserving TransE (lppTransE) where logical attributes are preserved, h and t can be mapped to different spaces. To this end, a head space mapping matrix ( _Mh ,

) And a tail space mapping matrix (M _t ,

) Can be used. Accordingly, the calculating unit 30 can calculate the score function of lppTransE as follows.

The score function of IppTransE may be similar to the score function of TransR, described below. However, TransR is mapped by a single mapping matrix, h and t are mapped by a single mapping matrix, whereas IppTransE according to the present invention is mapped to h by M _h and t can be mapped by M _t .

As another example, the calculating unit 30 may calculate the score function of the TransR (lppTransR) whose logical attribute is preserved using the score function of the conventional TransR (Lin et al., 2015).

In TransR, entities can be mapped to vectors of different relationship spaces according to the relationship. This can be expressed as follows.

,

)과 테일 매핑 행렬(

,

)로 분할될 수 있다. 그리고, 산출부(30)는 논리적 속성이 보존된 TransR(lppTransR)의 스코어 함수를 다음의 수학식을 이용하여 산출할 수 있다.Where M _r is a mapping matrix for the relation r. Therefore, in the TransR (lppTransR) in which the logical attribute is preserved, the mapping matrix of each relation is a head mapping matrix

,

) And a tail mapping matrix (

,

). ≪ / RTI > Then, the calculating unit 30 can calculate the score function of the TransR (lppTransR) in which the logical attribute is stored, using the following equation.

Using two separate mapping matrices, entities can have two different vector representations in the same relationship space.

As another example, the calculating unit 30 may calculate the score function of TransD (lppTransD) in which the logical attribute is preserved using the score function of the conventional TransD (Ji et al., 2015).

TransD can map entity vectors to different vectors contained in the relationship space depending on the entity type and relationship type. That is, the entity vector can be mapped by an entity-relationship specific mapping matrix, which can be expressed as follows.

과

은 엔티티-관계 특정 매핑 행렬을 의미한다. 상술한 매핑 행렬은 엔티티의 투영 벡터와 관계를 다음의 수학식에 따라 연산하여 산출될 수 있다.here,

and

Means an entity-relationship specific mapping matrix. The above-described mapping matrix can be calculated by calculating the relationship between the projection vector of the entity and the following equation.

,

)로 나눌 수 있다. 매핑 행렬은 하기의 수학식에 의해 산출될 수 있다.Where h _p , t _p, and r _p refer to the projection vector for the head, tail, and relationship, respectively. In addition, TransD (lppTransD) with preserved logical properties can transform r _p into two projection vectors

,

). The mapping matrix can be calculated by the following equation.

And the score function of TransD (lppTransD) with logical attribute preservation is as follows.

In this way, the calculating unit 30 can calculate the score function reflecting the logical attribute using the score function disclosed in the prior art.

The learning unit 40 can learn the knowledge graph in which the logical attribute is reflected by using the score function reflecting the logical attribute. This can be expressed by the following equation.

은 논리적 속성이 보존된 임베딩에 대응하는 스코어 함수를 의미하고, P는 올바른 트리플 세트, N은 잘못된 트리플 세트이고,

은 마진(margin)을 의미한다. 지식 그래프는 올바른 트리플만 존재하기 때문에, N은 지식 기존 트리플에서 헤드 또는 테일 엔티티를 대체하여 생성될 수 있다. here,

P denotes a correct triple set, N denotes an invalid triple set,

Means a margin. Since the knowledge graph has only the correct triple, N can be generated by replacing head or tail entities in existing knowledge triples.

That is, as shown in the above-described equation, the learning unit 40 can learn a correct triple and an incorrect triple set using margin-based ranking loss.

The learning unit 40 can learn a knowledge triple with a stochastic gradient descent algorithm. The learning unit 40 can learn the knowledge triple in such a manner that the score function reaches the low point from any one of the highest points on the three-dimensional space expressed in three dimensions. According to the above equation, the correct knowledge triple set and the incorrect knowledge triple set can be separately distributed in the three-dimensional space. The learning unit 40 can learn the correct knowledge triple stepwise during the path from the high point of the correct knowledge triple distribution to the low point. The learning unit 40 can learn the incorrect knowledge triple in a similar manner. Therefore, the knowledge graph embedding system reflecting the logical attributes according to the present invention can be optimized by the stochastic gradient descent algorithm.

Referring to Figs. 2 to 4, the specific function of the mapping unit 20 of Fig. 1 is shown.

FIG. 2 is a diagram illustrating an example of conventional translation-based embedding without considering the role of an entity.

Translation based embedding aims at finding the vector representation of knowledge graph entities in the embedding space by taking into account the relationships of the entities in the space. Conventionally, since an entity is mapped to a vector space without considering the role of the entity, there is a problem in that it can not express the logical attribute of the relationship such as transitivity and symmetry. That is, the vector of the transition relation or the symmetry relation loses the transitivity or symmetry in the embedding space.

For example, if is the subject of three triple _{(e 1, r 1, e} 2), (e 2, r 1, e 3), (e 1, r 1, e 3), r 1 is the implementation parties, If r ₁ is not a zero vector, there can be three kinds of entity vectors as shown. In FIG. 2 (a), e ₁ , e ₂ , and e ₃ may be arranged linearly. In this case, the techniques disclosed in the prior art can not represent triples of (e ₁ , r ₁ , e ₃ ). Furthermore, as shown in 2 (b), the e ₁ and e ₂ when placed on the same point (e _1, r _1, e ₂₎ it is not be expressed in the triple. And, as shown in FIG.'S 2 (c), the e ₂ and e ₃ when located at the same point (e _2, r _1, e ₃₎ this can not be expressed in the triple. Similarly, the conventional translation based embedding technique has a problem that the symmetric relation can not be expressed perfectly.

There are two problems caused by misrepresentation of migration relations and symmetry relationships.

First, relations with logical attributes are common in knowledge-based file processing systems. For example, FB15K, one of the reference datasets, consists of 483,142 triples, of which 84,172 triples have a transitional or symmetric relationship. Thus, conventional transformation-based embedding techniques may not accurately represent about 17% triples. One of the other reference datasets, WN18, has a 22.4% triple transitional or symmetric relationship.

Second, a fulfillment or symmetry relationship does not affect entities directly connected by a relation, but affects other entities that are shared by non-fulfillment or asymmetric relationships through entities. Therefore, it is important to demonstrate a fulfillment relationship and a symmetric relationship in a translation-based embedding technique.

Referring to Figs. 3 to 4, an example of a vector expression in which logical attributes are preserved by the mapping unit 20 of Fig. 1 is shown.

3 is a diagram showing an example of a vector expression in which the transition relation is reflected by the mapping unit 20.

As described above, the conventional translation-based embedding techniques ignore the role of the entity when the entity is embedded in the vector space, so that the transition relation or the symmetric relation is not accurately represented.

On the other hand, the mapping unit 20 according to the present invention may represent an entity according to at least one role assigned to the entity. That is,

,

)을 이용하여 r₁의 공간으로 매핑시킬 수 있다. 즉, 매핑부(20)는 헤드 엔티티를

매핑 행렬에 매핑시키고, 테일 엔티티를

매핑 행렬에 투영시킬 수 있다.In the figure, the solid line means to be mapped to the head role, and the dotted line to be mapped to the tail role. For example, if is the subject of three triple _{(e 1, r 1, e} 2), (e 2, r 1, e 3), (e 1, r 1, e 3), r 1 is the implementation parties, whether the role section 10 e ₁ performs only the head and serve, e _3, while performing only the tail acts, e ₂ can be given a role to perform both roles. The mapping unit 20 transforms the entity vector existing in the entity space into two different mapping matrices

,

) Can be used to map to the space of r ₁ . That is, the mapping unit 20 stores the head entity

Mapping to the mapping matrix, and the tail entity

Can be projected onto a mapping matrix.

은 제1 매핑 행렬

에 투영된 제1 엔티티(e1)의 벡터를 나타내고,

은 제1 매핑 행렬

에 투영된 제2 엔티티(e2)의 벡터를 나타내고,

은 제2 매핑 행렬

에 투영된 제2 엔티티(e2)의 벡터를 나타내고,

은 제2 매핑 행렬

에 투영된 제3 엔티티(e3)의 벡터를 의미한다. 즉,

과

는 (e₂, r₁, e₃)과 (e₁, r₁, e₃)의 두 지식 트리플의 r₁의 공간에서 동일한 위치에 배치될 수 있다. 이와 유사하게,

와

은 (e1, r1, e2)와 (e1, r1, e3)로부터 동일한 위치에 배치될 수 있다. 여기서, 제2 엔티티(e₂)는 역할 부여부(10)에 의해 헤드와 테일 모두의 역할을 부여받았기 때문에,

와

로 다르게 매핑될 수 있다. 결과적으로, 도시된 바와 같이, 역할 부여부(10)가 엔티티에 적어도 하나의 역할을 부여함으로써 매핑부(20)는 세 개의 트리플을 공간에 모두 표현할 수 있다.In the drawing,

A first mapping matrix

Gt; e1 < / RTI >

A first mapping matrix

Represents the vector of the second entity e2,

A second mapping matrix

Represents the vector of the second entity e2,

A second mapping matrix

(E3) of the third entity projected on the third entity e3. In other words,

and

Can be placed at the same position in the space of r ₁ of the two knowledge triples (e ₂ , r ₁ , e ₃ ) and (e ₁ , r ₁ , e ₃ ). Similarly,

Wow

Can be arranged at the same position from (e1, r1, e2) and (e1, r1, e3). Here, since the second entity e ₂ is given the role of both the head and the tail by the role assignment unit 10,

Wow

. ≪ / RTI > As a result, as shown, the role assigning unit 10 assigns at least one role to the entity, so that the mapping unit 20 can represent all three triples in space.

FIG. 4 is a diagram showing an example of a vector expression in which a symmetric relationship is reflected by the mapping unit 20. FIG.

에 의해 매핑되는 것을 나타내고, 점선으로 표시된 부분은 엔티티가 매핑 행렬

로 매핑되었음을 나타낸다. 매핑부(20)는

와

를 동일한 위치에 배치시키고,

와

또한 동일한 위치에 배치시킬 수 있다. 이에 따라, r₂는 임베딩 공간에서 대칭 관계를 정확하게 표현할 수 있다.Similar to the embodiment described above with reference to Fig. 3, the mapping section 20 can also clearly express the symmetry relationship. 4 is a diagram showing a case where two triples (e ₄ , r ₂ , e ₅ ) and (e ₅ , r ₂ , e ₄ ) having a symmetrical relation (relation) r ₂ are given. In the illustrated embodiment, the solid lines indicate that the entity is a mapping matrix

And the dotted line indicates that the entity is mapped by the mapping matrix < RTI ID = 0.0 >

Lt; / RTI > The mapping unit 20

Wow

Are arranged at the same position,

Wow

They can be arranged at the same position. Thus, r ₂ can accurately represent the symmetry relationship in the embedding space.

5 to 10 are diagrams illustrating performance results of a knowledge graph embedding method and system in which logical attributes according to the present invention are reflected.

Specifically, FIG. 5 is a chart showing the complexity of the knowledge graph embedding models, FIG. 6 is a chart showing characteristics of a data set used in a performance result, FIG. 7 is a table showing parameter values used in link prediction, 8 is a chart comparing the performance results of the link prediction with the prior art, FIG. 9 is a table showing the parameter values used in the triple classification, and FIG. 10 is a chart comparing the performance results of the triple classification with the prior art.

Figure 5 shows the complexity of knowledge graph embedding models.

In the figure, Ne and Nr represent the number of relations with an entity, and n and m denote entities and dimensions of the embedding space, respectively. The complexity tends to depend mainly on the number of relations. Therefore, when compared with the prior art TransE, TransR and TransD, the increase in complexity of the embedding reflecting the logical attribute according to the present invention is not important.

Figure 6 is a chart illustrating the characteristics of the dataset used for performance results.

In order to evaluate the performance of the knowledge graph embedding system 1000 reflecting the logical attributes according to the present invention and the knowledge graph embedding method reflecting the logical attributes described later, two well-known knowledge graphs of WordNet (Miller, 1995 ) And Freebase (Bollacker et al., 2008).

WordNet provides semantic relations between words, and there are two subdatasets consisting of WN11 (Socher et al., 2013) and WN18 (Bordes et al., 2014). Freebase is a knowledge graph representing general facts about the world and has two subdatasets: FB13 (Socher et al., 2013) and FB15K (Bordes et al., 2014). FB15K is used for both link prediction operation and triple classification operation, whereas FB13 can only be used for triple classification operation. A detailed description of the link prediction operation and the triple classification operation will be described later.

In the diagram shown, #Triples means the number of learning triples of each benchmark dataset, #Transit is the number of knowledge triples having a transition relation (relationship), #Symmetric is a knowledge triple having a symmetric relationship . ≪ / RTI > In addition, Ratio represents a ratio of knowledge triples, which is a transition relation or a symmetric relation. As shown, it can be seen that knowledge triples having a fulfillment relationship or a symmetric relationship take up a significant portion of the data set.

FIGS. 7 to 11 are views showing performance results of an embedding system (or method) in which logical attributes according to the present invention are reflected.

The performance of the embedding system (or method) that reflects the logical attributes according to the present invention may occur through two tasks. Both tasks can be categorized into a link prediction task and a triple classification task.

First, the link prediction task is a task of predicting a missing relation when there is a missing relation (relation) in the collected knowledge triple.

TransE (Lin et al., 2015) and TransD (Ji et al., 2013), TransH (Wang et al., 2014) , 2015) and the knowledge graph preserving the logical attributes according to the present invention will be compared.

8 is a diagram of a performance result comparison table for a link prediction task.

In FIG. 8, Mean Rank is the result of measuring the average rank of all the correct entities, and Hits @ 10 is the result of representing the percentage of correct triples ranked in the top 10. unif and bern are the results of two sampling methods. Two sampling methods, unif and bern, are disclosed in the prior art Wang et al., 2014, and a detailed description thereof will be omitted.

은 마진, n, m은 엔티티와 릴레이션에 대한 임베딩 차원, 그리고 D.S는 임베딩 스코어 함수의 비유사도 측정이다. 그리고 확률적 기울기 하강(stochastic gradient descent)의 반복수는 1000회이다.Referring to FIG. 7, the knowledge graph embedding system in which the logical attribute according to the present invention is reflected has five parameters. α is the learning rate, B is the number of learning triples in each group,

Is the margin, n, m is the embedding dimension for the entity and relation, and DS is the univariate measure of the embedding score function. And the number of iterations of the stochastic gradient descent is 1000 times.

In FIG. 8, Hits @ 10 of the TransD (lppTransD) preserving the logical properties according to the present invention are 92.5% (unif sampling) and 92.2% (bern sampling), respectively, 93.6% and 94.3%, respectively. That is, it can be seen that the performance of TransD (lppTransD) in which the logical attribute according to the present invention is preserved is 1.8% higher than that of TransD in the prior art. In a similar manner, knowledge graphs (lppTransE, lppTransR, lppTransD) in which logical attributes according to the present invention are preserved can achieve higher performance than conventional knowledge graphs.

FIG. 9 is a chart showing Hits @ 10 according to the mapping characteristics of relations of FB15K. 9, the embedding system incorporating the logical attributes according to the present invention exhibits a higher Hits @ 10 than the base model of N-to-1 and N-to-N, and is similar to the basic model of 1: 1 and 1: 10.

In the diagram shown, it can be seen that TransD (lppTransD) reflecting the logical attributes improves 7.3% performance in N vs. 1 and 3.7% in N vs. N, compared to the prior art TransD. Since all transitive relations and some symmetric relations are generally N to N, it is important to achieve good performance in N to N. That is, according to FIG. 9, it can be shown that the knowledge graph embedding system in which the logical attributes according to the present invention are reflected in N to N can solve the problem of the transitivity and symmetry of the conventional embedding technology.

FIGS. 10 to 11 are diagrams for a performance result comparison table for a triple classification operation. FIG.

The triple classification task is to determine whether a given knowledge triple is correct. FIG. 10 is a table of parameter types used in the triple classification operation. The parameters shown in FIG. 10 are the same as those shown in FIG. 7, and a repeated description thereof will be omitted.

As shown in FIG. 11, the knowledge graph embedding system 1000 in which the logical attribute according to the present invention is preserved can exhibit improved performance in comparison with the prior art in the triple classification operation. These results indicate that the knowledge graph embedding system 1000 preserving logical attributes according to the present invention effectively solves the problem of conventional translation based embedding.

12 is a flowchart showing a schematic flow of a knowledge graph embedding method in which logical attributes according to the present invention are reflected.

First, knowledge triples can be collected to assign at least one role to an entity that constitutes a knowledge triple (110). A knowledge triple can be collected from an external database and can assign at least one of a head role or a tail role to an entity that constitutes the collected knowledge triple. Also, when collecting knowledge triples from a standard data set, the process of assigning roles to entities may be omitted.

Next, the role-assigned entities may be mapped to different mapping matrices according to roles (120). That is, a head entity, which is an entity to which a head role is assigned, may be distinguished from a tail entity to which a tail role is assigned, so that head entities may be projected onto a first mapping matrix and tail entities may be projected onto a second mapping matrix. In this process, entities that can perform both a head role and a tail role can be simultaneously mapped to two different mapping matrices.

Subsequently, the score function may be calculated using different mapping matrices (130). The score function can calculate the sum vector of the head entity vector and the relation (relation) vector, and the difference value between the sum vector and the tail entity vector as an absolute value. The closer the result value using the score function is to 0, the more accurate the knowledge triple is represented by the relation.

Thereafter, the knowledge triple can be learned using the calculated score function (140). The step of learning the knowledge triple comprises: a link prediction step of predicting a missing entity if there is an entity missing in the collected knowledge triple; a triple classification step of determining whether the collected knowledge triple is represented without error by a relation . ≪ / RTI >

Such a technique for providing a knowledge graph embedding method and system reflecting logical attributes can be implemented in an application or can be implemented in the form of program instructions that can be executed through various computer components and recorded in a computer-readable recording medium. The computer-readable recording medium may include program commands, data files, data structures, and the like, alone or in combination.

The program instructions recorded on the computer-readable recording medium may be ones that are specially designed and configured for the present invention and are known and available to those skilled in the art of computer software.

Examples of computer-readable media include magnetic media such as hard disks, floppy disks and magnetic tape, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, media, and hardware devices specifically configured to store and execute program instructions such as ROM, RAM, flash memory, and the like.

Examples of program instructions include machine language code such as those generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules for performing the processing according to the present invention, and vice versa.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it will be understood by those skilled in the art that various changes and modifications may be made therein without departing from the spirit and scope of the invention as defined in the appended claims. It will be possible.

10: Role availability
20:
30:
40:

Claims (12)

Collecting knowledge triples and assigning at least one of a head role or a tail role to an entity constituting the knowledge triple;
Mapping a head entity, which is an entity to which a head role is assigned, to a first mapping matrix, and mapping a tail entity, which is an entity granted a tail role, to a second mapping matrix;
Calculating a score function using the first mapping matrix and the second mapping matrix; And
And learning the knowledge triple using the calculated score function.
The method according to claim 1,
Wherein the knowledge triple is composed of two entities and a relation indicating a relation between the two entities, wherein the logical attribute is reflected. 3. The method of claim 2,
The score function may include:
Wherein the logical attribute is a function representing a sum of the head entity vector and the relation vector and representing a difference value between the sum vector and the tail entity vector.
3. The method of claim 2,
Wherein learning the knowledge triple comprises:
A link predicting step of predicting a missing entity when a missing entity is present in the collected knowledge triple; And
And a triple classification step of determining whether the collected knowledge triples are represented without error by the relation.
The method of claim 3,
Wherein, in the link prediction step,
Predicting a missing entity using the score function,
The triple classification step includes:
And comparing the result value of the score function with a reference value to determine whether the knowledge triple is expressed without error, wherein the logical attribute is reflected.
The method according to claim 1,
Wherein the mapping step comprises:
Mapping the specific entity to both the first mapping matrix and the second mapping matrix when a specific entity is granted both the head role and the tail role.
Acquiring knowledge triples and assigning at least one of a head role or a tail role to an entity constituting the knowledge triple;
A mapping unit that maps a head entity, which is an entity to which a head role is assigned, to a first mapping matrix, and a tail entity that is an entity to which a tail role is assigned, to a second mapping matrix;
A calculating unit for calculating a score function using the first mapping matrix and the second mapping matrix; And
And a learning unit for learning the knowledge triple using the calculated score function, wherein the logical attribute is reflected.
8. The method of claim 7,
Wherein the knowledge triple comprises two entities and a relation indicating the association of the two entities. 9. The method of claim 8,
The score function may include:
Wherein the logical attribute is a function representing a sum of a head entity vector and an relation vector and representing an absolute value of a difference value between the sum vector and the tail entity vector.
9. The method of claim 8,
Wherein,
If there are missing entities in the collected knowledge triples,
Wherein the logic attribute is reflected to determine whether the collected knowledge triple is represented without error by the relation.
8. The method of claim 7,
Wherein the mapping unit comprises:
And mapping the specific entity to both the first mapping matrix and the second mapping matrix when a specific entity is granted both the head role and the tail role.
12. A computer-readable recording medium storing a computer program for providing a method and system for embedding knowledge graphs in which logical attributes are reflected according to any one of claims 1 to 11.

Images