《KG-BERT: BERT for Knowledge Graph Completion》

Article address ：https://arxiv.org/pdf/1909.03193.pdf?ref=https://githubhelp.com

Article content

   The judgment task for the integrity of knowledge map , The author regards the knowledge graph as a triple < Head entity , Relationship 、 Tailstock > Text sequence , And a new bi-directional knowledge graph is proposed Transformer Coding framework （KG-BERT） To model these triples . This is also the first time that the pre training language model is used in the knowledge map .

Article model

The rationality of predicting triples

   For the rationality judgment of triples , The model framework proposed by the author is shown in the figure above , The way to deal with it is simple , That is, the head entity 、 Between relation and tail entity BERT special token"[SEP]" Separate , The specific example is shown in the figure below ：

Relationship prediction

   The author's prediction of the relationship between the two entities is also relatively simple , That is to use two entities "[SEP]" Just separate them , The details are shown in the following figure ：

Article summary

   At present, the idea of this article is relatively simple , That is to use the advantages of the pre training language model , Use triples with “[SEP]” Separate them and input them into the model for training , Of course, this can also transform knowledge map into text training, which can be more suitable for the input of pre training language model .