REKER: Relation Extraction with Knowledge of Entity and Relation

Relation Extraction (RE) is an important task to mine knowledge from massive text corpus. Existing relation extraction methods usually purely rely on the textual information of sentences to predict the relations between entities. The useful knowledge of entity and relation is not fully exploited. In fact, off-the-shelf knowledge bases can provide rich information of entities and relations, such as the concepts of entities and the semantic descriptions of relations, which have the potential to enhance the performance of relation extraction. In this paper, we propose a neural relation extraction approach with the knowledge of entity and relation (REKER) which can incorporate the useful knowledge of entity and relation into relation extraction. Specifically, we propose to learn the concept embeddings of entities and use them to enhance the representation of sentences. In addition, instead of treating relation labels as meaningless one-hot vectors, we propose to learn the semantic embeddings of relations from the textual descriptions of relations and apply them to regularize the learning of relation classification model in our neural relation extraction approach. Extensive experiments are conducted and the results validate that our approach can effectively improve the performance of relation extraction and outperform many competitive baseline methods.