Learning Knowledge Graph Embedding with Entity Descriptions based on LSTM Networks

Knowledge inference and knowledge prediction is widely used in the intelligent fault diagnosis that is very important to the product safety. Most learning knowledge graph embedding methods represent entities and relations only with fact triples of knowledge graphs (KGs) through translating embedding models without integrating the rich semantic information in entity descriptions. However, entity description-based method DKRL only takes high frequency words of entity descriptions as input data in training using CNN encoder model, which loss the word order feature in entity descriptions and the relevance in context. In this paper, we propose a novel learning knowledge graph embedding method with entity descriptions named as Learning Knowledge Graph Embedding with Entity Descriptions based on LSTM Networks(KGDL), which can integrate word order features of each sentence in entity descriptions and the semantic information in fact triples of KGs, to enrich the semantic representations of KGs for promoting knowledge acquisition and inference. More specifically, we explore LSTM encoder model to encode all semantic information of entity descriptions based on each sentence without losing the independent feature of each sentence and the semantic associations between sentences of entities descriptions, then encode these sentence embeddings into the entity descriptions embeddings, and further learn knowledge graph embeddings from fact triples with entity descriptions embeddings. The experiment results show that KGDL gets better performance than state-of-the-art method DKRL, in terms of mean rank value and HITS@K with highly accurate, fast and robust. Moreover, KGDL based on two-steps relational path of KGs with entity descriptions has promising abilities for relation prediction and entity prediction, which gets better performance than state-of-the-art method Path-based TransE in knowledge inference.