RSN:Learning to Exploit Long-term Relational Dependencies in Knowledge Graphs

Learning to Exploit Long-term Relational Dependencies in Knowledge Graphs

1 problem

At present, only the contents of the atlas supplement are mainly based on triple-level, So-called triple-level It refers to information that only focuses on triples , Only use the information of triples , Without adding any information , This will cause a problem , Long term dependencies that are difficult to capture , all triple-level Unable to convey a wealth of information , Based on only map completion or entity alignment . The model selects lstm Model , Expand , Achieve long-term dependency , But in order to enrich the information , Entities and relationships have been distinguished , Not just like NLP Data processing in , Think of a path as sequence, This makes entities and relationships indistinguishable . Therefore, the author takes skip Way to distinguish entities and relationships ,skip The method is similar to the residual network , But slightly different .
Because the research method is to complete the field of knowledge map , For entity alignment content , Just skip , No research .

2 Model

2.1 Premise

To enhance the connectivity of the knowledge map , Increase inverse relationship ,$(UnitedKingdom,countr{y}^{-},TimBernersLee,employer,W3C)$

2.2 Model diagram

Transfer the whole model to LSTM in , So-called skip Add something , For the relationship LSTM Later results , Plus and current r Relevant entity information . Then go to predict and r Related tail entities .
Its skip The formula is as follows ：

2.3 Loss function

The loss function is as follows ：

The loss function is the loss function that calculates the predicted result ,$x_1,x_2,....x_n$ Include forecast $Turn\; offsystem{x}_2,x_4.....$, And prediction entities $x_3,x_5......$ And so on .

3 summary

Path based knowledge map completion , Can enhance their ability to express .