Transc knowledge representation model

Catalog

One 、 Research background

Two 、TransC The proposed

3、 ... and 、 Model thinking

Four 、 model training

5、 ... and 、 Thesis and code

---

One 、 Research background

TransE Then it was born again TransD、TransH、TransR Wait for the model , But the basic idea of these models is the same , It is using head Vector and relation The sum of vectors to predict tail vector . The sum of vectors is the accumulation of information . But entities and relationships can be more complex than that . Finding a more effective way to express knowledge may be a challenging research topic .

 

 

Two 、TransC The proposed

《Differentiating Concepts and Instances for Knowledge Graph Embedding》（2018） This paper presents a new learning method of knowledge map representation that distinguishes concepts and instances , Distinguish the superior inferior relationship from the ordinary relationship , It can solve the problem of transitivity of superior inferior relationship , And it can express the hierarchy and inclusion relationship of concepts in space . 

Pictured above ,Alice Is an example of a professor (InstanceOf), The professor belongs to (SubClassOf) Scientific researchers , Researchers belong to （SubClassOf) Staff , Transitivity through relationships (IsA Transitivity), The professor establishes contact with the staff .

The problem solved by the model ：

1. Insufficient concept representation
Can't distinguish concept and instance
2. Lack transitivity of both isA relations
isA This relationship lacks transitivity  

3、 ... and 、 Model thinking

The author puts forward a logical reasoning problem , Be similar to isA This relationship , Chinese understanding is “ yes ” Meaning , Such as (zhangsan, instanceof, teacher)+(teacher, subClassof, people) Introduction (zhangsan, is, teacher), This is it. isA The logic of . 

This paper divides relational triples into three types ：

1. (Instance, Relation, Instance)
namely （Zhangsan, Friendof, Lisi） This is the conventional TransE Model can
2. (Instance, Relation, Concept)
namely （Zhangsan, Is, Teacher）
Examples should be similar to the concepts they belong to , Each one concept Code as a domain , Each one instance Encoded as vectors in the same semantic space , And use the relative position to concept and instance Modeling the relationship between .
3. (Concept, Relation, Concept)
namely （Teacher, Is, People）
Divide this relationship into four different relationships , namely （A, Relation, B）,A and B The relationship between is B Included in A, They have nothing to do with , The two intersect , A Included in B. 

This paper divides reasoning into two types ：

1. concept To concept Of reasoning subClassOf-subClassOf

2. instance To concept Of reasoning instanceOf-subClassOf

Illustrate with examples ：

1. subClassOf-subClassOf 

( professor , yes , scholars ) ∩（ scholars , yes , people ）⇒ ( professor , yes , people )

2. instanceOf-subClassOf 

( Zhang San , yes , teach Delegate ）∩（ teach Delegate , yes , learn person ) ⇒ ( Zhang San , yes , scholars ) 

Four 、 model training

For presence "subClassOf" The triplet of relations (ci,rc,cj), Two concept balls si(pi,mi) and sj(pj,mj) , The distance between the two ball centers is ：d=∥pi−pj∥2. If the relationship holds , The positional relationship between the two balls should look like figure (a) equally , It is the training target for positive sample triples . The other three cases , Define scoring functions respectively , Make its relative position close to the ideal situation . 

 

As shown in the picture above , Two concepts Cj contain Ci, Entities are vectors in concept space , chart a It's an ideal model . chart b Is the separation of two spheres , chart c Is the intersection of two spheres , chart d Is that two spheres are contained in a relationship that deviates from reality . be ：

1.（b） and （c） By shortening the distance between the center of the ball , Add the ball j The radius of , Decrement ball i The radius of , In order to achieve （a） The effect of ;

2. （d） By adding balls j The radius of , Decrement ball i The radius of , In order to achieve （a） The effect of .

5、 ... and 、 Thesis and code

The paper ：https://arxiv.org/abs/1811.04588 

Code ：GitHub - davidlvxin/TransC: Source code and datasets of EMNLP2018 paper: "Differentiating Concepts and Instances for Knowledge Graph Embedding".