Abstract: Aiming at the problem of multiple mentions of entities and the noise of entity pairs in the document-level relation extraction task,  we  proposed a relation extraction model (EUT model) based on entity type information. The model  improved the relation extraction results through two sub-tasks:  entity type judgment and  a priori of the relation types produced by the type pairs. 
After the entity type judgment task labelled entities by type, then categorized all mentions of the entity by type, so that multiple mentions of the entity produced richer and similar feature representations. The relation category prior task enabled the model to obtain a prior of the  relation distribution  generated by the head and tail types of entity pairs, and reduced erroneous entity pair noise through the categories of entity pairs. In order to verify the effectiveness of the EUT model,  the  experiments were conducted on two document-level datasets, DocRED and Re-DocRED. The experimental results show that the model effectively utilizes the entity type information and achieves better relation extraction results compared to the base model, indicating that entity type information has an important impact on document-level relation extraction.

Key words: document-level relation extraction, knowledge graph, structured prior, natural language processing