ETFRE: Entity–Type Fusing for Relation Extraction

This paper proposes a relational extraction framework based on entity–type information fusion by Transformer model. Relational extraction, as an important part of knowledge graph construction, has been paid much attention in recent years. The existing relational extraction and joint triple extraction models rarely use the existing entity–type information, so the semantic features of the entity–type are lost, resulting in limited model performance and difficulty in solving the ambiguity problem. In order to improve this situation, this paper proposes a framework of entity–type information fusing based on a Transformer, which can generate word vector representation with entity–type information for a specific domain. There may be different entity categories for the same word, and the corresponding relationship categories are different at that time. Through deep self-attention, word vector representation is rich in entity–type information, which benefits relationship extraction and ambiguity removal. The multi-layer Transformer is used to realize the interaction between text features and generate a deep word vector representation with entity–type information, thus effectively avoiding ambiguity. Experimental results show that our model outperforms existing methods and performs well in ambiguous contexts relative to other models. We highlight the importance of entity–types in relation extraction.