Document-Level Relation Extraction with Additional Evidence and Entity Type Information

Document-level relation extraction faces the challenges of longer text and more complex context than sentence-level relation extraction. In document-level relation extraction, the relation information of an entity pair is usually contained within one or several sentences. However, excessively long document text may lead the model to focus on irrelevant sentences containing wrong information. On the other hand, using only textual information for relation extraction may not be sufficient, some previous models used only text information for relation extraction, ignoring some features of entities themselves, such as entity types, which can be guidance to relation extraction. To address these issues, a Sentence-Token Attention (STA) layer is developed to integrate sentence-level information into tokens. With a supervised attention optimization, the STA layer enables entities to focus more on relevant sentences. After that, we use an evidence fusion method to fuse the sentence information with context embedding. In addition, we indirectly incorporate the entity type information into the entity embedding as guidance to relation classification. Compared with different models, it is found that our model performs better in both relation extraction and evidence retrieval tasks than previous works.