TransD 论文笔记

- Abstract

TransR没有考虑关系的类型问题，而CtransR是仅考虑了关系的类型问题，而没有考虑实体的类型问题。由于在通用领域中，不一样的实体是属于不一样的类型的。TransD不仅考虑了关系的多样性，还考虑了实体的多样性。

- Background

TransR/CTransR several flaws:
(1) For a typical relation r, all entities share the same mapping matrix Mr. How- ever, the entities linked by a relation always con- tains various types and attributes;
(2) The projection oper- ation is an interactive process between an entity and a relation, it is unreasonable that the map- ping matrices are determined only by relations;
(3) Matrix-vector multiplication makes it has large amount of calculation, and when relation number is large, it also has much more param- eters than TransE and TransH. As the complex- ity, TransR/CTransR is difficult to apply on large- scale knowledge graphs.

- Algorithm

In TransD, we de- fine two vectors for each entity and relation. The first vector represents the meaning of an entity or a relation, the other one (called projection vector) represents the way that how to project a entity em- bedding into a relation vector space and it will be used to construct mapping matrices. There- fore, every entity-relation pair has an unique map- ping matrix.

Im×n to denote the identity matrix of size m × n.



损失函数：

TransE is a special case of TransD when the dimension of vectors satisfies m = n and all projec- tion vectors are set zero. 即M是单位矩阵。

when m = n, the difference between TransD and TransH is that projection vectors are determinded only by relations in TransH, but TransD’s projection vectors are determinded by both entities and relations.

（加速）TransD has no matrix-vector multiplication operation which can be replaced by vector operations.

- Experiment

Tiplets Classification

(1) On WN11, TransD outperforms any other previous models including TransE, TransH and TransR/CTransR, especially NTN(+E);
(2) On FB13, the classification accuracy of TransD achieves 89.1%, which is significantly higher than that of TransE, TransH and TransR/CTransR and is near to the performance of NTN(+E) (90.0%);
(3) Under most circumstances, the ”bern” sampling method works better than ”unif”.

Hence, the ability of reasoning new facts based on knowledge graphs is under a certain limitation, and a complementary approach is to extract facts from plain texts.

LInk Prediction

(1) TransD outperforms other baseline embedding models (TransE, TransH and TransR/CTransR), especially on sparse dataset, i.e., FB15k;
(2) Compared with CTransR, TransD is a more fine-grained model which considers the multiple types of entities and relations simultane- ously, and it achieves a better performance. It in- dicates that TransD handles complicated internal correlations of entities and relations in knowledge graphs better than CTransR;
(3) The “bern” sam- pling trick can reduce false negative labels than “unif”.

Therefore, the diversity of entities and relations in knowl- edge grahps is an important factor and the dynamic mapping matrix is suitable for modeling knowl- edge graphs.


We seek the similar objects (entities and rela- tions) for a given object (entities and relations) by projection vectors.
show that the same category objects have similar projection vectors.
对给定的三元组所找到相似三元组是满足实际的，说明TransD所设置的投射向量是合理的。

- Conclusion

Our contributions in this paper are:
(1)We pro- pose a novel model TransD, which constructs a dynamic mapping matrix for each entity-relation pair by considering the diversity of entities and re- lations simultaneously. It provides a flexible style to project entity representations to relation vec- tor space;
(2) Compared with TransR/CTransR, TransD has fewer parameters and has no matrix- vector multiplication. It is easy to be applied on large-scale knowledge graphs like TransE and TransH;
(3) In experiments, our approach outperforms previous models including TransE, TransH and TransR/CTransR in link prediction and triplets classification tasks.

These relations are difficult to infer from all other information, One possible way to obtain these new triplets is to extract facts from plain texts.We will seek methods to complete knowledge graphs with new triplets whose entities and relations come from plain texts.