ProjFE: Prediction of fuzzy entity and relation for knowledge graph completion

Abstract

A growing number of tasks about knowledge graph completion have been studied and improved recently, but most of them use translation matrices or reflect known entity to other space, always focusing on improving the method of translating known entities and relations. Differing from current works, our paper employs a combination operator instead of the translation matrix to avoid massive calculations, and takes fuzzy membership degree into consideration in the predicting process to enhance accuracy of projection. Hence, we propose a method called ProjFE to predict the missing parts of triplets for knowledge graph completion. This model uses fuzzy combination operators to combine the fuzzy known entities and relations. Score function is employed to access to a descending order of the correct candidates after combination, where the target entity is the top one. What is more, we use sigmoid and ReLU activation functions for evaluations, which could alleviate some undesirable gradient problems in the training process. It is worth noting that our method ProjFE tends to have a relatively smaller parameter size than some existing models. Besides, our model is proved to perform better in terms of Mean Rank.